text
stringlengths 1
2.56M
| id
stringlengths 40
40
| metadata
dict |
|---|---|---|
\section{Introduction} \label{sec:intro}
Titan is the second largest moon in the Solar System (radius $R_T$ = 2576 km) and possesses a dense extended atmosphere principally composed of $\sim$96\% molecular nitrogen, $<$4\% methane, and $<$1\% hydrogen \citep{Vervack04,Niemann05,Waite05}. Aerosol-type particles envelop the moon in a thick organic photochemical haze \citep{Danielson73}, a phenomenon also present at Pluto \citep{Gladstone16}, Triton \citep{Broadfoot89}, the Archean Earth \citep{Miller59} and likely also methane rich extra-solar planets. The production mechanisms and composition of these naturally occurring organic compounds are however far from understood.
The Cassini spacecraft has sampled the ionised regions of Titan's upper atmosphere down to altitudes of $<$900~km in-situ and observed positively charged ions (cations) extending up to nearly 1000~u/q \citep{Crary09,Coates10} and, surprisingly, negatively charged ions (anions) and aerosol precursors extending up to $13,800$~u/q \citep{Waite07,Coates07,Coates09}. The cations were detected at nearly all masses up to 100~u with over 50 species identified in this range \citep{Cravens06,Vuitton07}. At $>$100 u, evidence for carbon-based aromatic compounds has been reported, although unique identifications were not possible \citep{Crary09, Westlake14, Wahlund09}. The anions and larger negatively charged molecules were obtained at a lower resolution and classified into broad mass groupings of 12-30, 30-55, 55-90, 90-125, 125-195, 195-625, and 625$+$ u/q with the higher masses observed at lower altitudes and higher latitudes \citep{Coates07,Wellbrock13}. In the deep ionosphere below $\sim$1000~km the anion/aerosol precursor charge density was observed to exceed that of the electrons resulting in an ion-ion (dusty) plasma \citep{Shebanits13,Shebanits16}.
\begin{figure*}
\includegraphics[width=0.85\textwidth]{Figure_1.pdf}
\centering
\caption{Histogram of the CAPS-ELS anion ion mass/charge spectrum at various altitudes within Titan's ionosphere during the T40 encounter. The fitting routine (blue) and errorbars (red) are calculated as described in Equations \ref{eq1}-\ref{eq7}. The nominal centre of each group is marked (dotted black line) and the fitting parameters are given in Table \ref{table1}.
\label{fig2}}
\end{figure*}
A number of studies have modelled the cation and neutral chemistry occurring in Titan's ionosphere at $<$100~u/q, but only two studies have attempted to model the anion chemistry \citep{Vuitton09,Dobrijevic16}. These focussed on low mass species of $<$75~$\textrm{u}$ and in particular
inferred the presence of the C$_{n-1}$N$^-$ and C$_{n}$H$^-$ anions, where n=2-6. These carbon chain anions have all also been detected in dark molecular clouds, prestellar
cores or protostellar envelopes \citep{Cordiner13,Millar17} where their high reactivity acts as a catalyst for the formation of larger organic molecules \citep{Millar00,Walsh09}. Chemical models of these environments also predict even larger anions containing up to 23 carbon atoms \citep{Bettens96}.
At Titan, chemical schemes are only beginning to provide theories as to how the larger species can be produced. Stochastic charging models provide some explanation for how species of $\sim$100~u could be ionised and aggregate to form $>$10,000~u molecules \citep{Michael11, Lavvas13, Lindgren16} but only a few studies have looked at precise chemical routes for producing molecules $>$100 u.
\citet{Westlake14} demonstrate how cations of $>$100~u are likely formed from smaller hydrocarbon compounds through ion-molecule growth processes and \citet{Ali15} provided a mechanistic analysis of possible routes from small to large cations of $<$250~u/q. The latter was based upon Olah's three-membered H\"uckel aromatic rings \citep{Olah72, Olah16}, and suggest the presence of several carbocations and their corresponding carbanions.
This letter provides analysis of the anion and aerosol precuror dataset using a forward model of the Cassini Plasma Spectrometer (CAPS) Electron Spectrometer (ELS) instrument response function to these species. Statistical evidence for the low mass carbon chain anions, CN$^-$/C$_2H^-$ and C$_3$N$^-$/C$_4$H$^-$ is presented as well as constraints on intermediary anions in the range of 50-200~u/q. The role of these species is then examined with respect to the growth of larger molecules with decreasing altitude.
\section{Methodology} \label{sec:ps1_data}
The results presented in this letter are derived from CAPS-ELS observations obtained during the T16, T18, T32, T40 and T48 encounters. These include measurements across Titan's sunlit (T16 ingress, T18, T32 ingress, T40, T48) and anti-sunlit (T16 egress \& T32 egress) hemispheres, a variety of latitudes and also when Titan was immersed directly within the solar wind (T32).
Further information on the geometry and ambient conditions of the Titan encounters can be found in \citet{Coates09}.
The CAPS-ELS is a top-hat electrostatic-analyser sensitive to negatively charged particles in the 0.6 - 28,000 eV range \citep{Young04}. Anions can be identified within the ELS three-dimensional velocity distribution due to being highly super-sonic in the spacecraft frame and preferentially registering in anodes aligned with the spacecraft velocity vector. Thus as Cassini travels through Titan's ionosphere, the ELS observes an anion mass/charge spectrum,
\begin{equation}
\label{eq1}
\frac{m}{q}=\frac{2}{qv_{sc}^2}\left(E_{ELS}-\phi_{sc}\right)
\end{equation}
where $E_{ELS}$ is the nominal acceptance energy, $v_{sc}$ is the spacecraft velocity relative to Titan and $\phi_{sc}$ is the ELS spacecraft potential shift applied in accordance with Liouville's theorem \citep{Lewis08}.
To isolate the anion detections, the count rates are taken from each scan across the ram direction and the isotropically observed electrons on non-ram pointing anodes averaged and subtracted.
The count rate $R_C$, can then be related to the number density, $n_{ni}$, using the ion current approximation \citep{Waite07,Coates07},
\begin{equation}
\label{eq2}
n_{ni}=\frac{R_c}{ v_{sc} \hspace{0.5mm} A_F \hspace{0.5mm} \varepsilon_{} },
\end{equation}
where $A_F = 0.33~\textrm{cm}^2$ is the effective area of acceptance, and $\varepsilon_{}$ is the Microchannel Plate (MCP) anion detection efficiency function which is energy dependent. A value of $\varepsilon_{}=0.05$ was used in previous studies based upon the extensive study by \citet{Fraser02} which remains the best estimate for the larger species. Studies have however shown that at lower energies this could be significantly larger for negatively charged molecules \citep[e.g.][]{Peko00} where electron multiplication is increasingly dependent on potential as well as kinetic emission processes \citep{Hagstrum88}. Further analysis of the density uncertainties will be addressed by Wellbrock et al. (manuscript in preparation) and in this study we do not implement a new value.
The ELS energy bins are quasi-logarithmically spaced to match the energy resolution and overlap at FWHM which results in a nominal electron or anion distribution registering counts across multiple energy bins. To take this into account we forward model the ELS response to negative ions. The ELS $\Delta E/E$ resolution can be represented using a normalised Gaussian of the form
\begin{equation}
\label{eq3}
f(E) = R_{nc} \exp \left( -\frac{1}{2}\left( \frac{E-E_0}{E_W} \right )^2 \right )
\end{equation}
where $E_0$ is the center of the distribution, $E_W$ is the width and $R_{nc}$ is the normalised count rate. Here, $E_W$ corresponds to $\Delta E/E=16.7\%$ in the case of a single anion distribution but can be larger in the case of multiple overlapping distributions. The thermal energy spread of the anions is approximated by a drifting Maxwellian expressed in count rates,
\begin{equation}
\label{eq4}
g(E)=\frac{2nE^2G}{\sqrt{m}}\left(\frac{1}{2 \pi kT} \right)^{\frac{3}{2}} \exp \left(-\frac{(E-E_0)}{kT}\right),
\end{equation}
as adapted from \citet{Rymer01}, where $T$ is the ion temperature in Titan's ionosphere and $G$ the geometric factor,
\begin{equation}
\label{eq5}
G=A_F \frac{\Delta E}{E} \varepsilon,
\end{equation}
which is derived under the assumption that the negative ion current fills the ELS aperture. The ion temperature in Titan's ionosphere has been determined to be significantly less than the $\Delta E/E$ instrument sensitivity and tests with or without this thermal contribution produced similar results. In this study it is therefore held constant at $kT = 0.02$~eV ($\sim$150K) \citep{Crary09} and included for completeness.
The ELS response function and the anion distribution can then be convolved,
\begin{equation}
\label{eq6}
h(E)= f(E) \ast g(E),
\end{equation}
and the resulting function modelled to fit the observed data using a ${\chi}^2$ minimisation routine.
Of further mention is that the spacecraft surfaces charge to negative values in Titan's relatively dense ionosphere and the various surfaces will also charge to different potentials based upon variations in material conductivities and incident electron and ion currents \citep{Crary09}. This results in the exact potential correction, $\phi_{sc}$, also being unknown and the centers of the fitted distributions are therefore established relative to one another. Errors in the observed count rates are taken as
\begin{deluxetable*}{ccccccccccccc}
\tablenum{1}
\tablecaption{Fitting results for peaks 1-5 in the CAPS-ELS anion mass/charge spectrum at $<$200 u/q as marked in Figure \ref{fig2} and Figure \ref{fig3}. The $\chi^2_{red}$ values are for 3 DOF fits within the 20-60 u/q range and the super- and sub-scripts correspond to a 2$\sigma$ deviation. Here, notice that for $p=0.05$ we obtain a critical $\chi^2_{red} =2.60$. The ELS FWHM for a single distribution function is $\sim$16.7$\%$ and the spacecraft potentials correspond to the nearest RPWS-LP measurement of this parameter to closest approach.
\label{table1}}
\tablewidth{0.1pt}
\tablehead{
\colhead{} & \colhead{} & \colhead{} &
\multicolumn{2}{c}{\textbf{Peak 1}} & \multicolumn{2}{c}{\textbf{Peak 2}}
& \colhead{\textbf{Peak 3}} & \colhead{\textbf{Peak 4}} & \colhead{\textbf{Peak 5}} & \multicolumn{2}{c}{\textbf{Potential ($\mathbf{\phi_{sc}}$})} \\
\colhead{\textbf{Flyby}} & \colhead{\textbf{Alt.}} & \colhead{\boldmath$\mathbf{\chi^2_{red}}$} &
\colhead{\textbf{Center}} & \colhead{\textbf{FWHM}} & \colhead{\textbf{Center}} & \colhead{\textbf{FWHM}} &
\colhead{\textbf{Range}} & \colhead{\textbf{Range}} & \colhead{\textbf{Range}} &
\colhead{\textbf{RPWS-LP}} & \colhead{\textbf{ELS}}
\\
\colhead{$\mathbf{(\#)}$} & \colhead{$\mathbf{(km)}$} & \colhead{$\mathbf{(-)}$} &
\colhead{$\mathbf{(u/q)}$} & \colhead{\boldmath${(\%)}$} & \colhead{$\mathbf{(u/q)}$} & \colhead{\boldmath${(\%)}$} &
\colhead{$\mathbf{(u/q)}$} & \colhead{$\mathbf{(u/q)}$} & \colhead{$\mathbf{(u/q)}$} &
\colhead{$\mathbf{(V)}$} & \colhead{$\mathbf{(V)}$}
}
\startdata
\textbf{T40} & 1244 & $\rightarrow$0 & $25.8_{-0.6}^{+0.8}$ & $20.1_{-4.6}^{+2.6}$ & $49.6_{-1.5}^{+1.8}$ & $20.9+$ & 72-78 & - & - & -0.63 & -1.05
\\
\textbf{-} & 1148 & 1.50 & $25.8_{-0.5}^{+0.9}$ & $21.7_{-3.3}^{+2.1}$ & $γ50.1γ_{-0.8}^{+1.1}$ & $γ17.9γ_{-2.7}^{+1.5}$ & 71-89 & 109-120 & - & -0.59 & -1.20
\\
\textbf{-} & 1078 & 0.67 & $26.0_{-0.6}^{+0.4}$ & $γ21.6γ_{-2.8}^{+0.6}$ &$γ49.9γ_{-1.1}^{+2.0}$ & $γ17.0γ_{-3.5}^{+2.2}$ & 73-91 & 108-122 & 146-166 & -0.58 & -1.20
\\
\textbf{-} & 1032 & 0.72 & $25.7_{-0.5}^{+0.5}$ & $γ22.7γ_{-2.0}^{+0.6}$ & $γ49.0γ_{-0.5}^{+0.8}$ & $γ19.6γ_{-2.0}^{+1.4}$ & 73-91 & 109-123 & - & -0.57 & -1.20
\\
\textbf{-} & 1013 & 2.46 & $26.0_{-0.6}^{+0.8}$ & $γ22.7γ_{-2.9}^{+1.9}$ & $γ49.3γ_{-0.6}^{+0.6}$ & $γ18.1γ_{-1.6}^{+1.0}$ & 72-94 & 107-128 & - & -0.59 & -1.20
\\
\textbf{T16} & 1031 & 0.88 & $25.9_{-0.6}^{+0.4}$ & $γ22.8γ_{-6.0}^{+2.5}$ & $γ48.9γ_{-0.7}^{+0.7}$ & $γ19.6γ_{-2.2}^{+1.6}$ & 72-77 & 108-118 & 141-160 & -0.66 & -1.40
\\
\textbf{T18} & 969 & 0.73 & $26.0_{-1.0}^{+0.9}$ & $γ22.8γ_{-6.2}^{+3.7}$ & $γ49.0γ_{-0.7}^{+1.5}$ & $γ17.0γ_{-2.1}^{+2.0}$ & 74-82 & 113-123 & 148-166 & -1.65 & -2.50
\\
\textbf{T32} & 1036 & 0.92 & $25.8_{-1.1}^{+0.6}$ & $γ18.5γ_{-2.4}^{+3.7}$ & $γ49.5γ_{-1.2}^{+0.8}$ & $γ18.5γ_{-3.5}^{+1.8}$ & 74-94 & 114-125 & - & -0.80 & -1.55
\\
\textbf{T48} & 1082 & 2.18 & $26.0_{-0.2}^{+0.2}$ & $γ18.3γ_{-0.4}^{+0.7}$ & $γ49.7γ_{-0.6}^{+0.7}$ & $γ17.7γ_{-1.8}^{+1.1}$ & 74-88 & 110-121 & 147-164 & -0.80 & -1.40
\\
\enddata
\end{deluxetable*}
\begin{equation}
\label{eq7}
\sigma = \sigma_p + \sigma_{std}
\end{equation}
where $\sigma_{p}$ corresponds to the Poisson counting statistics and $\sigma_{std}$ corresponds to the standard deviation of counts on non-ram oriented anodes. This is used as a measure of electron anisotropies and inter-anode scaling uncertainties, introduced when isolating the anion detections.
\section{Species identification} \label{sec:Carbon chain identification}
Figure \ref{fig2} shows the anion mass/charge spectrum measured at various altitudes during T40. At higher altitudes ($>1300$ km) the larger $>$100 u/q species are absent and at the highest altitudes it is difficult to identify anions due to decreased densities. As Cassini descends, clear detections appear at $<$200 u/q and the larger $>$200 u/q distribution starts to grow below $\sim$1250 km. The five resolved clustered detections in the spectra are hereafter referred to as peaks 1-5 and fall within the range of mass groups 1-5 as described by \citet{Wellbrock13}. Figure \ref{fig3} also shows anion spectra obtained during encounters T16, T18, T32 and T48.
The fitting procedure applied to peaks 1 and 2 finds the center of the primary two peaks to be separated by $23-24.3$~u/q in all encounters, see Table \ref{table1}. Chemical models for Titanβs atmosphere predict efficient production of C$_n$H$^-$ and C$_{n-1}$N$^-$ to result from dissociative electron attachment to, or de-protonation of, parent neutral species C$_n$H$_2$ and HC$_{n-1}$N \citep{Vuitton09,Dobrijevic16}. For example, CN$^-$ and C$_3$N$^-$ are produced by
\begin{equation}
\label{eq8}
{HCN + e^- \rightarrow H + CN^-},
\end{equation}
\begin{equation}
\label{eq9}
HC_3N + CN^- \rightarrow HCN + C_3N^-,
\end{equation}
which proceeds rapidly due to abundant HCN and HC$_3$N. A similar reaction sequence exists for hydrocarbons where C$_2$H$^-$ and C$_4$H$^-$ are produced by
\begin{equation}
\label{eq10}
{C_2H_2 + e^- \rightarrow H + C_2H^-,}
\end{equation}
\begin{equation}
\label{eq11}
{C_4H_2 + C_2H^- \rightarrow C_2H_2 + C_4H^-.}
\end{equation}
%
due to abundance C$_2$H$_2$ and C$_4$H$_2$. This $23-24.3$~u/q separation in the ELS mass/charge spectrum is indicative of these processes and the CN$^-$/C$_2$H$^-$ and C$_3$N$^-$/C$_4$H$^-$ carbon chain anions as the dominant constituents within the primary and secondary peaks respectively. It is not possible to further resolve the 1~u difference between these nitrile and hydrocarbon compounds but CN$^-$ is estimated to be two orders of magnitude more abundant than C$_2$H$^-$, and C$_3$N$^-$ and C$_4$H$^-$ are predicted in comparable abundances \citep{Vuitton09}. The main anion loss process considered is associative detachment with neutral radicals.
\begin{figure*}
\centering
\includegraphics[width=0.85\textwidth]{Figure_2.pdf}
\caption{Histogram of the CAPS-ELS anion mass/charge spectrum during the T16, T18, T32 and T48 encounters. The fitting routine (blue) and errorbars (red) are calculated as described in Equations \ref{eq1}-\ref{eq7}. The nominal centre of each group is marked (dotted black line) and the fitting parameters provided in Table \ref{table1}.
\label{fig3}}
\end{figure*}
The width of the primary peak is however often larger than the ELS $\Delta E/E\approx16.7\%$. This can be explained by a multi-species composition, with further possible anion species such as C$_2^-$, CH$_{1,2,3}^-$, NH$_{1,2,3}^-$ and CO$^-$ possibly contributing, the latter due to the introduction of water-group ions (O$^+$, OH$^+$, H$_2$O$^+$, H$_3$O$^+$) from Enceladus \citep{Hartle06,Cravens08}. It is however possible that the spacecraft potential also acts to spread a given distribution's energy relative to the spacecraft, an effect which would be more pronounced for lower mass ions due to their lower inertia. The width of the secondary C$_3$N$^-$/C$_4$H$^-$ peak falls across a range which encompasses the FWHM of the ELS, indicating this is likely composed of a single distribution function with only a minor contribution from further species possible.
This analysis indicates the spacecraft potential experienced by the ELS is $0.4-0.9$~V more negative than that measured by the Radio and Plasma Wave Science (RPWS) Langmuir Probe (LP) although within the $-3.5$~V absolute range observed by the instrument in Titan's ionosphere \citep{Crary09}. An $\sim$-0.3~V discrepancy was found between conjugate CAPS Ion Beam Spectrometer (IBS) and Cassini's Ion and Neutral Mass Spectrometer (INMS) observations \citep{Crary09} and a more negative spacecraft potential correction is expected for anion detections \citep{Jones11}. This is due to focussing effects where the spacecraft-generated potential field acts to deflect incident ions such that the anions arrive from a direction closer to the spacecraft surface.
\begin{figure*}
\centering
\includegraphics[width=1.0\textwidth]{Figure_3.pdf}
\caption{Altitude profiles with ingress ($\triangle$) and egress ($\triangledown$) marked. (a-d) Shows the T40 relative anion densities, (e) the electron density, (f) the maximum mass detected and (g) the correlation between the depletion of the low mass ($<150$~u/q) anions below the altitude of peak density and corresponding increase in the maximum mass. A linear trend line (black) is fitted as well as individually to the ingress and egress of each encounter, the maximum and minimum of which encompass the region shaded grey. Relative errors are assumed small compared with the overall spread.
\label{fig4}}
\end{figure*}
At 1244 km (Figure \ref{fig2}b) the third peak is also visible where at higher altitudes it appears compatible with C$_5$N$^-$/C$_6$H$^-$, although this cannot be statistically verified as at $>$50~u/q there are fewer measurements than free parameters. At lower altitudes this peak widens, extending as high as $\sim$94~u/q in some instances.
This range does not include any previously observed anions and indicates the presence of anionic structures other than linear chains.
INMS and CAPS-IBS measurements at these altitudes also show a grouping of neutrals and cations over a similar range \citep{Waite07,Crary09}, the most abundant of which was inferred to be benzene (C$_6$H$_6$) \citep{Vuitton08}. This is particularly relevant as benzene and benzene products are thought to be the seeds for larger aromatic compounds \citep{Vuitton16}. The lack of reaction rates and known chemical pathways at these high masses severely restricts the analysis of the anion chemistry but several candidate species can be suggested. For example, the cyclic radical C$_6$H$_5^+$ is observed in appreciable quantities \citep{Vuitton16} and can be a source for C$_6$H$_5^-$ anion radical production \citep{Fenzlaff84}. The stable C$_6$H$_7^-$ cyclic anion can also be produced through the interaction of benzene with H$^-$ \citep{Coletti12}. H$^-$ cannot be measured however due to constraints imposed by the spacecraft velocity (see Equation \ref{eq1}) but H$^-$ fluxes can result from the interaction of methane with ionospheric electrons \citep{Dobrijevic16}.
The four further encounters in Figure \ref{fig3} also show this third peak to be the most variable between encounters, possibly indicating an enhanced sensitivity to the ambient conditions. A number of azine anions resulting from benzene, pyridine, pyridazine, pyrazine, and s-triazine have been explored by \citet{Wang15} with application to Titan, and determined to be highly reactive with nitrogen and oxygen. These include C$_6$H$_3^-$, C$_5$H$_2$N$^-$, C$_5$H$_3$N$^-$, C$_5$H$_4$N$^-$, C$_4$H$_3$N$_2^-$ and C$_3$H$_2$N$_3^-$ and are of high astrobiological interest. The higher end of this third mass range is however less explored and further anions such as the benzyl anion C$_7$H$_7^-$ and anilinide anion C$_6$NH$_6^-$ \citep{Wang16}, could be derived from the $>20$ mostly Polyaromatic Hydrocarbons and nitrated heterocyclic species suggested in this range.
At altitudes $<$1200~km a fourth distinct peak, with consistent detections at $\sim$117$\pm$3~u/q, is identified during every encounter. This appears at count rates comparable to the primary peaks and sometimes double that of the neighboring peaks. This is most evident during T16 where the highest mass species at 13,800~u/q were observed \citep{Coates09}. A smaller fifth peak, with detections consistently at $\sim$154$\pm$8~u/q, is also sometimes present. The $>$100~u/q regime is however even more unconstrained. It should be noted though that \citet{Trainer13} detected ring structures near 117~u in laboratory simulations of aerosol-tholin production which could represent growth processes involving aromatic rings at Titan.
Multiply charged anionic states are not considered here due to inter-electron repulsive forces making this phenomena increasingly unlikely for smaller molecules. This is evident as the smallest known multiply charged anions, C$_n^{2-}$ (n=7-28), have lifetimes of tens of micro-seconds in the gas phase whereas the larger C$_{60,70}^{2-}$ molecules can persist for milliseconds \citep[][]{Wang09}. It therefore appears that multiple charges are much more likely on molecules larger than a few hundred amu in Titan's ionosphere, as indeed was reported by \citet{Shebanits16}.
\section{Molecular growth}
Figure \ref{fig4} (a-d) shows the altitude profiles of peak 1 (25.8-26.0~u/q, associated with CN$^-$/C$_2$H$^-$), peak 2 (49.0-50.1~u/q, associated with C$_3$N$^-$/C$_4$H$^-$), peak 3 (71-94~u/q), and peak 4 (107-123~u/q) during the T40 encounter. Figure \ref{fig4} (e-f) shows the electron densities and maximum detected mass of the larger molecules as obtained in \citet{Coates09}. The altitude profiles in Figure \ref{fig4} show the carbon chain anions to peak in density above the region where the highest mass aerosol monomer are observed and to extend several scale heights above this to where the larger species are not present in measurable quantities. While the precise relative densities are not known, the profiles show the negative charge to be increasingly carried by the larger species at lower altitudes. This trend is also observed in further encounters, see \citet{Wellbrock13}, Figure 3.
Below $\sim$1100~km the depletion of the $<$150~u/q anions can be seen to be related to the size increase of the larger molecules. Figure \ref{fig4} (g) shows this correlation for all encounters except T18 which was omitted as the altitude of peak density of the low mass ions was not definitively surpassed (indeed during T48 this was only surpassed for a brief instance). The two parameters can be seen to be linearly proportional although there is some spread to the data at the higher masses. This overall proportional decrease in the low mass species with the increase of the larger species points to dependencies between these.
The data also point to a possible diurnal variation with the day-side measurements grouped together at and above the fitted trend line (black) and the night-side measurements appearing below this and with T16 and T32 crossing Titan's solar terminator. \citet{Coates09} previously determined significant spatial variations of the high mass monomers and these data point to this being echoed in the smaller species. Further statistical analyses are however required to fully disentangle such influences.
\section{Summary \& Conclusions} \label{sec:cite}
This study used a forward model of the CAPS-ELS response function to achieve an increased mass resolving capability for anions in Titan ionosphere, the results of which are as follows: The first peak ($25.8-26.0$~u/q) and second peak ($49.0-50.1$~u/q) in the anion spectrum were statistically shown to be compatible with the CN$^-$/C$_2$H$^-$ and C$_3$N$^-$/C$_4$H$^-$ respectively, although it was not possible to differentiate between these nitrile or hydrocarbon compounds. At altitudes above $\sim$1200~km the third peak is consistent with the further chain anion C$_5$N$^-$/C$_6$H$^-$ but at lower altitudes becomes dominated by higher mass species not consistent with carbon chain anions. Notably, this is evidence for the presence of more complex structures which may well come to represent the first astrophysical detection of anions not composed of linear chains, see recent review by \citet{Millar17}. A number of species were suggested to account for this intermediary peak based upon the relatively better understand cation and neutral chemistry. Further persistent detections were also constrained at 117$\pm3$~u/q and 154$\pm$6~u/q which appear to also add to the growing body of evidence for growth processes involving ring structures although the current lack of known reaction pathways impedes their definitive interpretation at present.
The evolution of the low mass ($<$150~u/q) anions was then examined with respect to the impending growth of the larger organic molecules. Deep within the ionosphere the lower mass anions were observed to become depleted as the larger aerosol precursors coincidentally underwent rapid growth. This trend contributes to the idea that smaller species form the seeds for the larger species via a series of reactions and processes which the chain anions and further intermediary anions appear to be tightly coupled to. These results demonstrate the importance of tracing a route from small to large species in order to fundamentally understand how complex organic molecules can be produced within a planetary atmosphere.
\acknowledgments
RTD acknowledges STFC Studentship 1429777. AJC, AW and GHJ acknowledge support from the STFC consolidated grants to UCL-MSSL ST/K000977/1 and ST/N000722/1. DGC acknowledges Becas-Chile CONICYT Fellowship (No. 72150555). OS acknowledges funding from SNSB, Dnr 130/11:2
| d9df315bf87fad18a109dbca68aa8cdda4a3d027 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
One of the central tenet of statistical mechanics is the notion of statistical ensembles. In thermal equilibrium, a system can be described by different statistical ensembles: the microcanonical, the canonical, and the grand-canonical ensemble. In the thermodynamic limit, and in the presence of short-ranged interactions, bulk properties do not generically depend on the choice of the ensemble. Such a property is known as ensemble equivalence.\footnote{Nevertheless, systems with long-ranged interactions exhibit violation of the ensemble equivalence \cite{BMR-01,LR-02}.} In particular, a textbook argument for the equivalence between the canonical and the grand-canonical ensembles consists in the following observation. In the grand-canonical ensemble the particle number as well as the energy are \textit{sharp} in the thermodynamic limit, i.e., their relative fluctuation vanishes in the thermodynamic limit. This stems from the fact that the specific heat,
\begin{equation}
C_V = \frac{d \langle \hat{H} \rangle}{d T} = k_B\beta^2 \left( \langle \hat{H}^2 \rangle - \langle \hat{H} \rangle^2 \right),
\label{specific_heat}
\end{equation}
and the charge susceptibility
\begin{equation}
\Xi_c = \frac{d \langle \hat{N} \rangle}{d \mu } = \beta \left( \langle \hat{N}^2 \rangle - \langle \hat{N} \rangle^2 \right),
\label{charge_chi_extensive}
\end{equation}
are extensive quantities that measure energy and particle-number fluctuations.
In Eqs.~(\ref{specific_heat}) and (\ref{charge_chi_extensive}) $\hat{H}$ is the Hamiltonian of the system, $\hat{N}$ the particle-number operator, $\beta=1/k_BT$ the inverse temperature in units of the Boltzmann constant $k_B$, and $\mu$ the chemical potential.
Thus,
\begin{equation}
\lim_{N \rightarrow \infty} \frac{ \sqrt{ \left( \langle \hat{H}^2 \rangle - \langle \hat{H} \rangle^2 \right) }} { \langle \hat{H} \rangle } =
\lim_{N \rightarrow \infty} \frac{ \sqrt{ \left( \langle \hat{N}^2 \rangle - \langle \hat{N} \rangle^2 \right) }} { \langle \hat{N} \rangle } = 0
\end{equation}
and the selection of the ensemble is merely a matter of convenience.
Nonetheless, in many cases the choice of ensemble is dictated by the physical properties of the system under study. In fact, while the canonical ensemble requires the presence of a heat bath which fixes the temperature, the grand-canonical ensemble additionally needs a particle reservoir which allows to fix the chemical potential. Systems which lack such a particle bath, like those found in nuclear physics or in cold atoms, require a description in terms of the canonical ensemble. Moreover, in the case of mesoscopic systems with a finite particle number, a reliable comparison with experimental data needs a theoretical computation based on the canonical ensemble. In this context, we mention that, unlike the finite-temperature auxiliary field quantum Monte Carlo (QMC) method considered here, the so-called projective auxiliary field QMC, which targets the ground state of fermionic models, is a method which is intrinsically formulated within the canonical ensemble \cite{Assaad08_rev}.
The aim of this paper is twofold. On one hand we will introduce a QMC method for fermionic models in the canonical ensemble, consisting in a simple formulation of the auxiliary field QMC method which enforces the conservation of the particle number. Our approach differs from that adopted in Ref.~\cite{KDK-97} and supplements the Hamiltonian that we simulate in the grand-canonical ensemble by the long-ranged interaction term
\begin{equation}
\lambda \left( \hat{N} - N \right)^2,
\end{equation}
such that in the infinite-$\lambda$ limit charge fluctuations are suppressed and the canonical ensemble is recovered. This type of interaction is easily incorporated in the auxiliary field QMC, especially in the formulation provided in Ref.~\cite{ALF}. The advantage of such an approach is that $ \lambda$ can be dynamically chosen. For instance, at low temperatures the charge susceptibility can vanish due to finite size or correlation-induced charge gaps. In this case $\lambda$ can be set to a very small number, or even to zero since both canonical and grand-canonical ensembles yield identical results.
At high temperatures, where the grand-canonical ensemble exhibits significantly large charge fluctuations, bigger values of $\lambda $ are required to impose the constraint.
The second motivation of the paper is to look into finite-size corrections both in the canonical and grand-canonical ensembles, which we study in quantum and classical lattice models.
Concerning classical models on a lattice, it should be noted that in the literature the canonical ensemble is often defined by the usual partition function sum, where one considers all the configurations without any constraint. In the case of the standard Ising model, this corresponds to the usual partition function:
\begin{equation}
Z_{\rm gc}(h) = \sum_{\{S_k=\pm 1\}}\exp\left\{\beta J\sum_{<i j>}S_iS_j+h\sum_i S_i\right\}.
\label{ZIsing_gc}
\end{equation}
However, through the mapping to the lattice gas, the magnetization of the model corresponds to the particle number, which in the ensemble of Eq.~(\ref{ZIsing_gc}) is allowed to fluctuate. In order to provide a more meaningful comparison to quantum models, we refer to the lattice gas language and define the grand-canonical ensemble as the one where the magnetization is not fixed; in Eq.~(\ref{ZIsing_gc}) we have anticipated this definition, such that the subscript gc refers the grand-canonical ensemble. Conversely, we define the canonical ensemble as the ensemble where the magnetization is fixed, so that the corresponding partition function of the Ising model is
\begin{equation}
\begin{split}
Z_{\rm can}(h,m) = \sum_{\{S_k=\pm 1\}}&\exp\left\{\beta J\sum_{<i j>}S_iS_j+h\sum_i S_i\right\}\\
&\ \ \cdot\delta\left(m, \frac{1}{V}\sum_i S_i\right),
\end{split}
\label{ZIsing_can}
\end{equation}
where the constraint is enforced by employing the Kronecker delta function $\delta(m,n)$. In the literature, the ensemble of Eq.~(\ref{ZIsing_can}) is often referred to as the {\it fixed-magnetization} ensemble. In three dimensions, the Ising model at fixed magnetization has been investigated by means of Monte Carlo (MC) simulations in Ref.~\cite{BHT-00}, using the geometric cluster algorithm \cite{HB-98,HB-98B}.
In this work, we study the approach to the thermodynamic limit in the presence of a finite mass gap or, in the language of statistical physics, with a finite exponential correlation length. Generically for short-ranged Hamiltonians, on a finite volume with periodic boundary conditions and in the grand-canonical ensemble, the various observables are expected to show a finite-size correction which is proportional to $\exp(-L/\xi)$, where $L$ is the linear size of the system and $\xi$ is the exponential correlation length (or inverse mass gap). This expectation has been confirmed by explicit field theory calculations, both in the continuum \cite{Luescher-86,Neuberger-89} and on a lattice \cite{Muenster-85}; early numerical studies confirmed these prediction \cite{MW-87}. An exponential approach to the thermodynamic limit is also verified, e.g., in the well-known solution of the one-dimensional Ising model, as well as in generic one-dimensional $O(N)-$invariant spin models \cite{CMPS-97}. Nevertheless, it should be noted that, in the grand-canonical ensemble, some specific observables can exhibit a leading finite-size correction proportional to a power law of the system size. This is the case of the most common definition of the second-moment correlation length on a lattice, where finite-size corrections $\propto 1/L^2$ are due to the discretization of momenta on a finite lattice; see, e.g., the corresponding discussion in Ref.~\cite{CP-98} and Appendix A of Ref.~\cite{PTHAH-14}. We also remark that, in the presence of nontranslationally invariant boundary conditions, finite-size corrections polynomial in the inverse lattice size $1/L$ arise naturally, being related to subleading terms in the free energy; for instance, open boundary conditions result in the presence of a surface free energy which is depressed by a factor $1/L$ with respect to the bulk one and gives rise to finite-size corrections $\propto 1/L$ for bulk observables.
Conversely, in the canonical ensemble the prediction of exponentially decaying finite-size corrections fails, since the constraint introduces a long-ranged interaction, such that fluctuations in spatially separated regions (as measured by the correlation length) are not independent. Such a long-ranged (weak) correlation modifies also the high-temperature expansion of a model \cite{ISR-15} and results in a slower approach to the thermodynamic limit of various observables, so that the leading finite-size correction is proportional to the inverse volume $V$. Several important properties of the free energy in the canonical ensemble have been, in fact, discussed in the literature, although under a different perspective and notation. In quantum field theory, the so-called constrained effective potential $U_{\rm eff}$, introduced in the context of scalar field theories in Ref.~\cite{FK-75}, is defined as
\begin{equation}
e^{-VU_{\rm eff}(m,V)} = \int [{\cal D}\varphi] e^{-S[\varphi]}\delta\left(m - \frac{1}{V}\int d^dx\varphi(x)\right),
\label{Ueff_def}
\end{equation}
where $S[\varphi]$ is the action of the theory and the right-hand side of Eq.~(\ref{Ueff_def}) is a constrained path-integral over the field configurations where the volume-average value of $\varphi$ is fixed to $m$. In the language of statistical physics, the right-hand side of Eq.~(\ref{Ueff_def}) is precisely a constrained partition function sum at fixed magnetization, i.e., the partition function in the canonical ensemble. Hence, $U_{\rm eff}(m,V)$ is the free energy per volume and $k_BT$ in the canonical ensemble. A detailed analysis of the constrained effective potential has shown that it admits an infinite-volume limit $U_{\rm eff}(m,V\rightarrow\infty)$ which coincides with the usual effective potential $\Gamma(m)$ of the theory \cite{ORWY-86}. Moreover, as argued in Ref.~\cite{Neuberger-89}, $U_{\rm eff}(m,V)$ exhibits finite-size corrections which are polynomial in $1/V$. This is because, as a consequence of the definition in Eq.~(\ref{Ueff_def}), the grand-canonical average of any function of the magnetization $m$ is equivalent to an average over an effective probability measure $\propto \exp\{-VU_{\rm eff}(m,V)\}$, which for $V\rightarrow\infty$ can be evaluated by a saddle-point expansion, resulting in a series in $1/V$. On the other hand, the grand-canonical average converges exponentially to the limit $V\rightarrow\infty$. This is possible only if $U_{\rm eff}(m,V)$ displays finite-size corrections polynomial in $1/V$, which exactly cancel the expansion in $1/V$ originating from the saddle-point evaluation \cite{Neuberger-89}. A renormalized loop expansion for a $\phi^4$ theory on the lattice has confirmed the existence of finite-size corrections $\propto 1/V$ \cite{Palma-92}.
In this context, a recent study verified the existence of finite-size corrections $\propto 1/V$ in the canonical ensemble, and, conversely, of exponentially decaying finite-size corrections in the grand-canonical ensemble \cite{ISR-15}.
In this paper we provide an exact formula for the leading finite-size corrections in the canonical ensemble of the free energy density, energy density and other observables. While our analysis is restricted to the case of a finite correlation length, we mention that the introduction of a constraint to a nonordering parameter results in the so-called Fisher renormalization, leading to a modification of the singularities associated with a critical point, such that the critical exponents differ from those observed in the corresponding unconstrained system \cite{EG-67,Fisher-68}. The choice of ensemble is also relevant to the so-called critical Casimir force \cite{FG-78}, whose behavior in the canonical ensemble has been recently investigated within mean-field theory and MC simulations~\cite{GVGD-16}.
This paper is organized as follows. In Sec.~\ref{sec:method} we illustrate the QMC method that we use to generate numerical data for fermionic models in the canonical ensemble. In Sec.~\ref{sec:fscanonical} we provide an exact determination of the leading finite-size corrections in the canonical ensemble. In Sec.~\ref{sec:qmc} we study the finite-size corrections of the Hubbard model in one and two dimensions. In Sec.~\ref{sec:conclusions} we summarize our results. In the Appendix we provide an exact solution of the one-dimensional classical Ising model in the canonical ensemble to the leading order in $1/V$, which confirms the general result of Sec.~\ref{sec:fscanonical}.
\section{Canonical auxiliary field methods}
\label{sec:method}
\subsection{General formulation}
\label{sec:method:general}
In this section we review various methods to achieve canonical auxiliary field QMC simulations at finite temperature. We will consider a Hamiltonian of the form
\begin{equation}
\begin{split}
&\hat{H} = \hat{T} + \hat{V},\\
&\hat{T} \equiv \sum_{x,y} \hat{c}^{\dagger}_{x} T_{x,y} \hat{c}^{\phantom\dagger}_{y},\\
&\hat{V} \equiv \sum_{k} U_{k} \left(\hat{V}^{(k)} + \alpha_k \right)^2,\qquad \hat{V}^{(k)}\equiv \sum_{x,y} \hat{c}^{\dagger}_{x} V^{(k)}_{x,y} \hat{c}^{\phantom\dagger}_{y},
\end{split}
\end{equation}
that can be readily implemented in the ALF package \cite{ALF}. Here $x$ is a super-index encoding orbital and spin degrees of freedom, $\hat{c}^{\dagger}_{x}$ are fermion creation operators, $V^{(k)}$ and $T$ are Hermitian matrices, and $U_k$, $\alpha_k$ real numbers.
To simplify the notation, in the following we assume that the chemical potential term $\mu\hat{N}$, with $\hat{N} \equiv \sum_{x} \hat{c}^{\dagger}_{x} \hat{c}^{\phantom\dagger}_{x}$, has been adsorbed into the Hamiltonian $\hat{H}$.
Using the Trotter decomposition with $L_\tau \Delta \tau = \beta$, and a discrete version of the Hubbard-Stratonovich (HS) transformation,
\begin{equation}
\label{HS_squares}
e^{\Delta \tau \lambda \hat{A}^2 } = \frac{1}{4}
\sum_{ l = \pm 1, \pm 2} \gamma(l)
e^{ \sqrt{\Delta \tau \lambda }
\eta(l) \hat{A} }
+ {\cal O} (\Delta \tau ^4)\;,
\end{equation}
with $ \gamma(\pm 1) = 1 + \sqrt{6}/3$, $\eta(\pm 1 ) = \pm \sqrt{\smash[b]{2 (3 - \sqrt{6} )}} $, and
$ \gamma(\pm 2) = 1 - \sqrt{6}/3$, $\eta(\pm 2 ) = \pm \sqrt{\smash[b]{2 (3 + \sqrt{6} )}}$,
one can approximate the imaginary time propagator $e^{-\beta \hat{H}}$ as
\begin{equation}
\begin{split}
e^{-\beta \hat{H} } = &\sum_{ \left\{ l_{k,\tau} \right\} } e^{S_0 \left\{ l_{k,\tau} \right\} }\prod_{\tau=1}^{L_{\tau}} e^{-\Delta \tau \hat{T}} \prod_{k}e^{ \sqrt{- \Delta \tau U_k } \eta(l_{k,\tau} )\hat{V}^{(k)} }\\
& + i\Delta\tau \hat{R} + O(\Delta\tau^2).
\end{split}
\label{trotter}
\end{equation}
Here $S_0 \left\{ l_{k,\tau} \right\} = \sum_{l_{k,\tau} } \ln \left( \gamma(l_{k,\tau} ) \right) + \sqrt{- \Delta \tau U_k } \eta(l_{k,\tau} ) \alpha_k $ . It is easy to show that the contribution of the anti-Hermitian operator $i\hat{R}$ to the expectation value of an Hermitian oberservable is purely imaginary, so that the discretization error $\propto \Delta\tau$ can be filtered out, leading to a Trotter error $\propto \Delta\tau^2$.
The systematic error involved in this discrete HS transformation is of a higher order than the one encountered in the Trotter decomposition so that it can be regarded as good as exact.
At this point, one can integrate out the fermions so as to obtain the grand-canonical partition function:
\begin{equation}
Z_{\rm gc} = \Tr\left\{e^{-\beta \hat{H}} \right\}= \sum_{ \left\{ l_{k,\tau} \right\} } e^{S_0 \left\{ l_{k,\tau} \right\} } \det( 1 + U(l_{k,\tau} ) )
\end{equation}
with
\begin{equation}
U(l_{k,\tau} ) = \prod_{\tau=1}^{L_{\tau}} e^{-\Delta \tau T} \prod_{k}e^{ \sqrt{- \Delta \tau U_k } \eta(l_{k,\tau} )V^{(k)} }.
\end{equation}
Using the Leibniz formula for determinants, one can show that:
\begin{equation}
\begin{split}
\det &( 1 + U) \\
&= 1 + \sum_{N=1}^{N_s}\ \sum_{x_N > x_{N-1} > \cdots > x_1} \det
\begin{bmatrix} U_{x_1,x_1} & \dots & U_{x_1,x_N} \\
\vdots & \ddots & \vdots \\
U_{x_N,x_1} & \cdots & U_{x_N,x_N} \end{bmatrix}\\
&= 1 + \sum_x U_{x,x} + \sum_{x_2 > x_1} \det
\begin{bmatrix} U_{x_1,x_1} & U_{x_1,x_2} \\
U_{x_2,x_1} & U_{x_2,x_2} \end{bmatrix} + \cdots
\end{split}
\end{equation}
Here $N_s$ corresponds to the number of single-particle states, and one can readily see that each term of the sum corresponds to the canonical trace of $N$ single-particle states. Thereby, the canonical partition function $Z_{\rm can}(N)$ is given by:
\begin{equation}
Z_{\rm can}(N) = \frac{d^N}{dz^N} \sum_{ \left\{ l_{k,\tau} \right\} }\left. e^{S_0 \left\{ l_{k,\tau} \right\} } \det( 1 + z U(l_{k,\tau} ) ) \right|_{z=0}.
\end{equation}
A numerical implementation of the above equation reads:
\begin{equation}
\begin{split}
&Z_{\rm can}(N)\\
&= \frac{1}{N_s} \sum_{m=1}^{N_s} \sum_{ \left\{ l_{k,\tau} \right\} } e^{S_0 \left\{ l_{k,\tau} \right\} } e^{-i \phi_m N } \det( 1 + e^{i \phi_m} U(l_{k,\tau} ) ),
\end{split}
\label{Z_can1.eq}
\end{equation}
where $\phi_m = 2 \pi m/ N_s $.
An equivalent way to show the above result is to note that the total particle number $\hat{N}$ commutes with the Hamiltonian such that:
\begin{equation}
\begin{split}
Z_{\rm can}(N) &= {\rm Tr} \left[ \delta_{\hat{N}, N } e^{-\beta \hat{H} }\right]\\
&=\frac{1}{N_s} \sum_{m=1}^{N_s} e^{-i \phi_m N } {\rm Tr} \left[ e^{i \phi_m \hat{N}}e^{-\beta \hat{H} } \right].
\end{split}
\label{Zcan_alt}
\end{equation}
By applying a Trotter decomposition and HS transformation to the right-hand side of Eq.~(\ref{Zcan_alt}), one can reproduce Eq.~(\ref{Z_can1.eq}). Implementations of canonical simulations using the above results have been proposed in Refs.~\cite{Ormand94,Gilbreth15}. In these approaches, the discrete Fourier transformation is computed exactly at each MC step. For the method to be successful, the chemical potential has to be chosen such that the average particle number is peaked around the desired value.
\subsection{Constraint of the particle-number fluctuations}
\label{sec:method:constraint}
Here we follow a slightly different approach and modify the Hamiltonian as
\begin{equation}
\begin{split}
&\hat{H}(\lambda) = \hat{H} + \hat{H}_\lambda,\\
&\hat{H}_\lambda \equiv \lambda \left( \hat{N} - N_0 \right)^2,
\end{split}
\label{Ham_CE_QMC}
\end{equation}
such that
\begin{equation}
Z_{\rm can}(N_0) = \lim_{\lambda \rightarrow \infty } {\rm Tr} \left[ e^{-\beta \hat{H}(\lambda) } \right].
\label{Z_CE_QMC}
\end{equation}
As discussed above, in Eqs.~(\ref{Ham_CE_QMC}) and (\ref{Z_CE_QMC}) the Hamiltonian $\hat{H}$ implicitly depends on the chemical potential $\mu$, which needs to be tuned such that $\<\hat{N}\>=N_0$. In practice, this is done by computing $\<\hat{N}\>$ as a function of $\mu$, for a suitable interval in $\mu$, by means of auxiliary field QMC and then fixing $\mu$ such that the equation $\<\hat{N}\>=N_0$ is satisfied within the desired statistical accuracy; at half-filling one has exactly $\mu=0$.
Since $\hat{H}$ conserves the particle number, one can foresee rapid convergence in $\lambda$ because particle-number sectors with $\hat{N} = N \neq N_0$ have a statistical weight suppressed by a factor $ e^{-\lambda \beta \left( N - N_0 \right)^2 }$. The latter also shows that the relevant parameter for the convergence is $\beta \lambda$ rather than $\lambda $ itself. The additional term is a perfect square term which is easily implemented within the ALF code \cite{ALF}. Since $ \left( \hat{N} - N_0 \right)^{2} $ effectively corresponds to a long-ranged interaction, one may face the issue that the acceptance rate of a single HS flip becomes excessively small on large lattices. To circumvent this problem we have used the following decomposition:
\begin{equation}
e^{-\beta \hat{H}} = \prod_{\tau = 1}^{L_{\tau}} \left[ e^{-\Delta \tau \hat{T}} e^{-\Delta \tau \hat{V}}
\underbrace{e^{-\frac{\Delta \tau}{n_{\lambda}} \hat{H}_{\lambda} } \cdots e^{-\frac{\Delta \tau}{n_{\lambda}} \hat{H}_{\lambda} } }_{n_\lambda \text{-times } }\right].
\label{trotter_nlambda}
\end{equation}
Thereby, we need $n_\lambda $ fields per time slice to impose the constraint. For each field, the coupling constant is effectively suppressed by a factor $n_{\lambda}$, thus allowing to control the acceptance of the QMC algorithm.
In order to test the efficiency of our QMC method in the canonical ensemble, we computed the uniform intensive charge susceptibility $\chi_c$, defined as
\begin{equation}
\chi_c \equiv \frac{\beta}{V}\left( \langle \hat{N}^2 \rangle - \langle \hat{N} \rangle^2 \right).
\label{charge_chi}
\end{equation}
Note that compared with the extensive definition in Eq.~(\ref{charge_chi_extensive}), here the susceptibility is divided by the system volume $V$.
In Fig.~\ref{chi_c_1D} we show $\chi_c$ for the one-dimensional (1D) Hubbard model as a function of $\beta \lambda$ and $n_{\lambda}$.
As shown in Fig.~\ref{chi_c_1D}(a),
$\chi_c$ decays gradually from a finite value to zero on increasing $\beta\lambda$.
The threshold in $\lambda$ for which $\chi_c$ converges to zero corresponds to the canonical ensemble.
A comparison of the results for lattice sizes $L=4$, $8$, and $16$ suggests that
the charge fluctuations are easier to suppress for larger system sizes. The dependence of $\chi_c$ on $n_{\lambda}$ defined in Eq.~(\ref{trotter_nlambda})
is shown in Fig.~\ref{chi_c_1D}(b), which
illustrates the increased Trotter error for larger values of $\beta\lambda$.
\begin{figure}
\subfigure{\label{chi_c_1D_L}
\includegraphics[width=8cm, height=6cm]{chi_c_1D_L}
}
\subfigure{
\includegraphics[width=8cm, height=6cm]{chi_c_1D_L8}
\label{chi_c_1D_L8}
}
\centering
\caption{$\beta \lambda$ and $n_{\lambda}$ dependence of $\chi_c$ for the $1D$ Hubbard model at $U=4.0$ and $\beta=0.5$.
(a) $\chi_c$ as a function of $\beta \lambda$ for $L=4$, $8$, and $16$. For each $\beta \lambda$ we have taken the
parameter $n_{\lambda}$ large enough as to effectively suppress the discretization error in the decomposition of the constraint.
(b) $\chi_c$ as a function of $n_{\lambda}$ for $L=8$ and two values of $\lambda$. }
\label{chi_c_1D}
\end{figure}
Figure \ref{chi_c_2D_beta} shows the decay of charge susceptibility $\chi_c$ as a function of $\lambda$
in the two-dimensional (2D) Hubbard model, for $U=4.0$, $L=4$ and
several inverse temperatures $\beta=0.5$, $2.0$, and $5.0$.
Inspection of Fig.~\ref{chi_c_2D_beta} reveals that in the
grand-canonical ensemble the $\beta=2.0$ case exhibits charge fluctuations larger than the $\beta=0.5$ case, thereby requiring a larger value of $\beta\lambda$
to realize the canonical ensemble.
\begin{figure}
\includegraphics[width=8cm, height=6cm]{chi_c_2D_beta}
\centering
\caption{ $\lambda$ dependence of $\chi_c$ for the $2D$ Hubbard model at $U=4.0$, $L=4$, and $\beta=0.5$, $2.0$, and $5.0$.}
\label{chi_c_2D_beta}
\end{figure}
\section{Finite-size corrections in the canonical ensemble: exact results}
\label{sec:fscanonical}
In this section, by exploiting the relation between the canonical and the grand-canonical free energy, we determine the leading finite-size correction of the free energy in the canonical ensemble and relate it to a susceptibility. To be concrete, we consider a quantum model on a lattice, where in the canonical ensemble the number of particles is fixed, and we prove that on a finite volume $V$
\begin{equation}
\begin{split}
F_{\rm can}&(n_0,V) - F_{\rm gc} (V)\\
&= \frac{1}{2V}\ln\left(2\pi V\right) + \frac{1}{2V}\ln\left(\frac{\chi_c}{\beta}\right) + O\left(\frac{1}{V^2}\right),
\end{split}
\label{F_can_diff_gc}
\end{equation}
where $F_{\rm can}(n_0,V)$ and $F_{\rm gc} (V)$ are the free energies per volume $V$ and in units of $k_BT$ in the canonical and grand-canonical ensembles, respectively, and $\chi_c$ is the charge susceptibility (in the grand-canonical ensemble), defined in Eq.~(\ref{charge_chi}); the filling fraction $n_0$ in $F_{\rm can}(n_0,V)$ is fixed to the corresponding expectation value in the grand-canonical ensemble. Equation (\ref{F_can_diff_gc}) provides the leading additional contribution to the free energy density due to the particle-number constraint. As discussed towards the end of this section, Eq.~(\ref{F_can_diff_gc}) allows also to determine the leading finite-size correction of observables in the canonical ensemble if, as expected, finite-size corrections in the grand-canonical ensemble decay faster than $1/V$.
In order to prove Eq.~(\ref{F_can_diff_gc}), we observe that the free energy density $F_{\rm can}(n,V)$ can be related to a path-integral formulation of the canonical partition function as
\begin{equation}
e^{-VF_{\rm can}(n,V)} = \int [{\cal D}\Psi] e^{-S[\Psi]}\delta\left(n,\frac{1}{V}\hat{N}(\Psi)\right),
\label{Fcan}
\end{equation}
where $\Psi$ indicates collectively the fields entering in the path integral, $S[\Psi]$ is the action of the model, $\hat{N}(\Psi)$ is the expression of the total number operator $\hat{N}$ in terms of the fields $\Psi$, and $n$ is the intensive filling fraction, which is fixed in the canonical ensemble. In Eq.~(\ref{Fcan}), $S[\Psi]$, as well as $F_{\rm can}(n,V)$, additionally depend on the temperature and coupling constants, inessential for the present discussion.
On a lattice, $\hat{N}$ is the sum of single-site and single-species number operators $\hat{N}(x)$, $\hat{N}=\sum_x \hat{N}(x)$, therefore $n$ can only take discrete values, separated by an interval of $1/V$. By summing over the allowed values of $n$, we obtain the grand-canonical free energy density $F_{\rm gc}(V)$
\begin{equation}
e^{-VF_{\rm gc}(V)} = \sum_{n=n_{\rm min}}^{n_{\rm max}} e^{-VF_{\rm can}(n,V)},
\label{F_gc_discrete}
\end{equation}
where, as before, we have ignored the dependence of $F_{\rm gc}(V)$ on the various coupling constants, and $n_{\rm min}$, $n_{\rm max}$ indicate the minimum and maximum number of particles per volume that the model can host; usually $n_{\rm min}=0$, while $n_{\rm max}$ depend on the number and type of particle species. For $V\rightarrow\infty$ the sum in Eq.~(\ref{F_gc_discrete}) can be approximated by the Euler-Maclaurin formula as
\begin{equation}
\begin{split}
&e^{-VF_{\rm gc}(V)} = V\Bigg[\int_{n_{\rm min}}^{n_{\rm max}} dn\ e^{-VF_{\rm can}(n,V)}\\
&+ \frac{e^{-VF_{\rm can}(n_{\rm min},V)}+e^{-VF_{\rm can}(n_{\rm max},V)}}{2V}+O\left(\frac{e^{-cV}}{V}\right)\Bigg],
\end{split}
\label{F_gc_continous}
\end{equation}
where the next-to-leading term in the Euler-Maclaurin formula is $\propto (1/V^2)\partial (e^{-VF_{\rm can}})/\partial n$ computed at the end points, hence it is of order $e^{-cV}/V$.
In the limit $V\rightarrow\infty$, the integral on the right-hand side of Eq.~(\ref{F_gc_continous}) is dominated by the minimum $n_0$ of $F_{\rm can}(n,V)$. If $n_0$ is an interior point\footnote{The case of multiple saddle points, or a saddle point at an end point requires a separate analysis.} of the integration interval $[n_{\rm min},n_{\rm max}]$, by using the saddle-point method we obtain
\begin{equation}
\begin{split}
&e^{-VF_{\rm gc}(V)}/V = e^{-VF_{\rm can}(n_0,V)}\cdot\\
&\left[\frac{2\pi}{V(\partial^2F_{\rm can}/\partial n^2)(n_0,V)}\right]^{1/2}\left[1+O\left(\frac{1}{V}\right)\right]\\
&+ \frac{e^{-VF_{\rm can}(n_{\rm min},V)}+e^{-VF_{\rm can}(n_{\rm max},V)}}{2V}+O\left(\frac{e^{-cV}}{V}\right),
\end{split}
\label{F_gc_continous_saddle}
\end{equation}
where the factor $1+O(1/V)$ represents the next-to-leading term in the saddle-point expansion.
The second term on the right-hand side of Eq.~(\ref{F_gc_continous}) is depressed by a factor $\propto \exp\{-V[F_{\rm can}(n_{\rm min},V)-F_{\rm can}(n_0,V)]\}/V^{1/2} + \exp\{-V[F_{\rm can}(n_{\rm max},V)-F_{\rm can}(n_0,V)]\}/V^{1/2}$ with respect to the first term, therefore, since $n_0$ is the minimum of $F_{\rm can}(n,V)$, it is subleading with respect to the first factor. Moreover, the convergence of the integral in Eq.~(\ref{F_gc_continous}) requires the last term on the right-hand side of Eq.~(\ref{F_gc_continous_saddle}) to be subleading with respect to the first factor. Thus, by factorizing the first term on the right-hand side of Eq.~(\ref{F_gc_continous_saddle}) and taking the logarithm, the last two terms give a contribution of order $\ln(1+\exp\{-cV\}/V^{1/2})\sim \exp\{-cV\}/V^{1/2}$, which is negligible with respect to the correction of order $1/V$ originating from the next-to-leading term of the saddle-point expansion.
On taking the logarithm on both sides of Eq.~(\ref{F_gc_continous_saddle}) we find
\begin{equation}
\begin{split}
F_{\rm gc}(V) &= F_{\rm can}(n_0,V) - \frac{1}{V}\ln V - \frac{1}{2V}\ln\left(\frac{2\pi}{V}\right) \\
&+ \frac{1}{2V}\ln\left[\frac{\partial^2F_{\rm can}}{\partial n^2}(n_0,V)\right] + O\left(\frac{1}{V^2}\right),
\end{split}
\label{F_gc_vs_can}
\end{equation}
where subleading exponential corrections have been neglected. The second and third terms $\propto \ln V$ on the right-hand side of Eq.~(\ref{F_gc_vs_can}) represent an entropic contribution which is due to the larger configurational space of the grand-canonical ensemble as compared to the canonical one. In particular, the first constant originates from the discretization of the allowed values of $n$ [see the discussion after Eq.~(\ref{Fcan})] and is absent in continuous models.
The saddle-point position $n_0$ appearing in the previous equations corresponds precisely to the grand-canonical expectation value of $\<\hat{N}/V\>_{\rm gc}$. This is because, using Eq.~(\ref{Fcan}) and Eq.~(\ref{F_gc_discrete}), one can write $\<\hat{N}/V\>_{\rm gc}e^{-VF_{\rm gc}(V)} = \sum_{n=n_{\rm min}}^{n_{\rm max}} ne^{-VF_{\rm can}(n,V)}$. Along the same line of reasoning as above, one finds that, as expected also from thermodynamic considerations, $\lim_{V\rightarrow\infty}\<\hat{N}/V\>_{\rm gc} = n_0$. Thus, the quantity $F_{\rm can}(n_0,V)$ on the right-hand side of Eq.~(\ref{F_gc_vs_can}) is precisely the free energy density with a particle number fixed to its expectation value in the grand-canonical ensemble, i.e., the thermodynamic quantity which is meaningful to compare with the grand-canonical free energy density. The fluctuation of the particle number, which determines the charge susceptibility $\chi_c$ defined in Eq.~(\ref{charge_chi}), can be related to the finite-size correction on the right-hand side of Eq.~(\ref{F_gc_vs_can}). By using Eq.~(\ref{Fcan}), Eq.~(\ref{F_gc_discrete}), and the definition of Eq.~(\ref{charge_chi}), one obtains
\begin{equation}
\chi_c=\frac{\sum\limits_{n=n_{\rm min}}^{n_{\rm max}}\frac{\beta}{V}(nV-n_0V)^2 e^{-VF_{\rm can}(n,V)}}{\sum\limits_{n=n_{\rm min}}^{n_{\rm max}} e^{-VF_{\rm can}(n,V)}}.
\label{n_fluctuation}
\end{equation}
The right-hand side of Eq.~(\ref{n_fluctuation}) can be evaluated for $V\rightarrow\infty$ using a saddle-point expansion as above, resulting in
\begin{equation}
\chi_c \underset{V\rightarrow\infty}{=} \frac{\beta}{\left(\partial^2F_{\rm can}/\partial n^2\right)(n_0,V)}.
\label{n_fluctuation_result}
\end{equation}
Finally, inserting Eq.~(\ref{n_fluctuation_result}) into Eq.~(\ref{F_gc_vs_can}), we obtain Eq.~(\ref{F_can_diff_gc}).
If finite-size corrections of $F_{\rm gc}(V)$ decay faster than $1/V$ (indeed, as discussed in Sec.~\ref{sec:intro}, we expect exponentially decaying finite-size corrections), we can replace $F_{\rm gc}(V)$ on the left-hand side of Eq.~(\ref{F_can_diff_gc}) with its ensemble-independent thermodynamic limit $F(V=\infty) = F_{\rm gc}(V=\infty) = F_{\rm can}(n_0,V=\infty)$, such that the leading finite-size corrections in $F_{\rm can}(n_0,V)$ are
\begin{equation}
\begin{split}
F_{\rm can}&(n_0,V) - F(V=\infty)\\
&= \frac{1}{2V}\ln\left(2\pi V\right) + \frac{1}{2V}\ln\left(\frac{\chi_c}{\beta}\right) + O\left(\frac{1}{V^2}\right).
\end{split}
\label{F_can_finitesize}
\end{equation}
From Eq.~(\ref{F_can_finitesize}) we can, e.g., determine the leading finite-size correction of the energy density in the canonical ensemble by taking the derivative with respect to $\beta$:
\begin{equation}
E_{\rm can}(V) - E(V=\infty) = \frac{\partial\left(\chi_c/\beta\right)/\partial\beta}{2V\left(\chi_c/\beta\right)}.
\label{E_can_finitesize}
\end{equation}
It is useful to remark that the charge susceptibility $\chi_c$ appearing in Eqs.~(\ref{F_can_diff_gc}) and (\ref{n_fluctuation})-(\ref{E_can_finitesize}) is computed in the grand-canonical ensemble. Since $\chi_c$ has a finite thermodynamic limit and exponentially decaying finite-size corrections, it does not give rise to a further algebraic volume dependence.
Equation ~(\ref{E_can_finitesize}) can be generalized to other local observables and correlations thereof. To this end, one can supplement the action of the model with an external source term
\begin{equation}
S[\Psi] \rightarrow S[\Psi] - h\sum_{x}\int d\tau O(x,\tau),
\label{action_with_h}
\end{equation}
where the sum extends to the lattice sites and $O(x,\tau)$ is a local observable, to be expressed in terms of the fields $\Psi$ entering in the path integral of Eq.~(\ref{Fcan}). Such an addition corresponds to the insertion of external lines in the Feynman diagram expansion, and hence one expects, in line with the analysis of Ref.~\cite{Neuberger-89}, that in the presence of a finite mass gap correlations including $O(x,\tau)$ are characterized by exponentially decaying finite-size corrections. Under the substitution of Eq.~(\ref{action_with_h}), the charge susceptibility $\chi_c$ entering in Eq.~(\ref{F_can_diff_gc}) and Eq.~(\ref{F_can_finitesize}) acquires a dependence on the external field $h$. Differentations of the free energy density with respect to $h$ provide the analogous of Eq.~(\ref{E_can_finitesize}) for the finite-size corrections of the volume-average and susceptibility of $O$:
\begin{equation}
\begin{split}
&O \equiv \frac{1}{\beta V} \sum_{x}\int d\tau \<O(x,\tau)\>_{h=0},\\
&O_{\rm can}(V) - O(V=\infty) = -\frac{\partial\chi_c(h)/\partial h|_{h=0}}{2\beta V\chi_c(h=0)},
\end{split}
\label{O_can_finitesize}
\end{equation}
\begin{equation}
\begin{split}
&\chi_O \equiv \frac{1}{\beta V} \sum_{x, x'}\int d\tau d\tau' \Big[\<O(x,\tau)O(x',\tau')\>_{h=0}\\
&\qquad\qquad\qquad\qquad\qquad - \<O(x,\tau)\>\<O(x',\tau')\>_{h=0}\Big],\\
&\chi_{O,{\rm can}}(V) - \chi_O(V=\infty)\\
&= \frac{\left(\partial\chi_c(h)/\partial h|_{h=0}\right)^2-\chi_c(h=0)\partial^2\chi_c(h)/\partial h^2|_{h=0}}{2\beta V\chi_c(h=0)^2},
\end{split}
\label{chiO_can_finitesize}
\end{equation}
where we emphasize that the expectation values of $O(x,\tau)$ are computed in absence of the external field $h$.
In particular, in a spinful model Eq.~(\ref{chiO_can_finitesize}) implies a leading finite-size correction $\propto 1/V$ of the spin susceptibility in the canonical ensemble.
We remark that the derivatives of $\chi_c(h)$ appearing in Eqs.~(\ref{O_can_finitesize}) and (\ref{chiO_can_finitesize}) can be in principle directly computed by sampling a suitable observable, thus avoiding a numerical differentiation.
The results of Eqs.~(\ref{F_can_diff_gc}), (\ref{F_can_finitesize}), (\ref{E_can_finitesize}), (\ref{O_can_finitesize}), and (\ref{chiO_can_finitesize}) can be easily generalized to other correlations by considering a considering a space- and imaginary time-dependent source $h(x,\tau)$ in Eq.~(\ref{action_with_h}), or to other types of constrained models, along the same line of reasoning.
\section{Fermionic simulations in the canonical ensemble}
\label{sec:qmc}
We performed QMC simulation of the $SU(2)$ Hubbard model in both the grand-canonical and canonical ensemble. The Hamiltonian of the Hubbard model is defined as:
\begin{align}
\hat{H} = &-t \sum_{< \bm i, \bm j>, \sigma} \hat{c}^{\dagger}_{\bm i, \sigma} \hat{c}^{\phantom\dagger}_{\bm j, \sigma}
+ U \sum_{\bm i} \left(\hat{n}_{\bm i, \uparrow} -\frac{1}{2} \right) \left(\hat{n}_{\bm i, \downarrow} -\frac{1}{2} \right) \nonumber\\
&-\mu\sum_{\bm i}\left(\hat{n}_{\bm i, \uparrow}+\hat{n}_{\bm i, \downarrow}\right),
\label{Hubbard_Hamiltonian}
\end{align}
where $\hat{n}_{\bm i, \sigma}\equiv \hat{c}^{\dagger}_{\bm i, \sigma} \hat{c}_{\bm i, \sigma }$. The canonical ensemble is realized by adding the constraint given in Eq.~(\ref{Ham_CE_QMC}). For such a modified Hamiltonian, the total number of particles
converges quickly to $N_0$ on increasing $\beta \lambda$.
Here we simulated both ensembles on a 1D lattice, as well as on the 2D square lattice at finite temperature, both of which are known to be disordered.
We mainly considered the models at half filling ($N_0=N_s/2$, with $N_s=2L^d$)
with zero chemical potential $\mu=0$ and carried out some test calculations for the two-dimensional doped Hubbard model. In all simulations we fixed $t=1$ and $U=4.0$.
Our basic MC observables are as follows:
\begin{enumerate}
\item Energy density\footnote{Up to an inessential, filling-dependent, additive constant, $E$ corresponds to the energy part on the right-hand side of Eq~(\ref{Hubbard_Hamiltonian}).}:
\begin{equation}
E=\frac{1}{L^d} \Bigg\langle -t \sum_{< \bm i, \bm j>, \sigma } \hat{c}^{\dagger}_{\bm i, \sigma} \hat{c}^{\phantom\dagger}_{\bm j, \sigma }
+ U \sum_{\bm i} \hat{n}_{\bm i, \uparrow} \hat{n}_{\bm i, \downarrow} \Bigg\rangle
\end{equation}
\item Uniform spin susceptibility:
\begin{equation}
\chi_s = \frac{\beta}{L^d} \sum_{\bm i, \bm j} \<\hat{S}_{\bm i} \hat{S}_{\bm j}\>
\end{equation}
\end{enumerate}
\begin{figure}
\includegraphics[width=8cm, height=6cm]{Energy_1D}
\centering
\caption{Finite-size data of the energy density $E$ for the 1D Hubbard model in the grand-canonical and canonical ensembles, at $\beta=0.5$ and half-filling. The red line is a linear fit of
the canonical ensemble data to $E_{\rm can}(L)=E(L \rightarrow \infty)+a/L$, with $E(L \rightarrow \infty)=0.1771(2)$ and $a=-0.738(4)$,
where the minimum lattice size taken into account is $L_{\rm min}=16$;
the dashed green line linking the grand-canonical data is a guide to the eye.}
\label{Energy_1D}
\end{figure}
\begin{figure}[b]
\includegraphics[width=8cm, height=6cm]{Spinsus_1D}
\centering
\caption{Same as Fig.~\ref{Energy_1D} for the spin susceptibility $\chi_s$. The red line is the linear fit of the
canonical ensemble data to $\chi_{s, {\rm can}}(L)= \chi_s (L \rightarrow \infty) + a/L$, with $\chi_s(L \rightarrow \infty)= 0.3177(1)$ and $a=0.135(1)$, where the minimum lattice size taken into account is $L_{\rm min}=12$.}
\label{Spinsus_1D}
\end{figure}
\subsection{1D model}
The QMC simulations of the one-dimensional Hubbard model are performed in both the grand-canonical and canonical ensembles at inverse temperature $\beta=0.5$,
system sizes $L=4$, $8$, $12$, $16$, $20$, $24$, $28$, $32$, $40$, $48$, $64$, $72$, $80$, and at half-filling.
A comparison of the size effect for the energy density $E(L)$ and for the uniform spin susceptibility $\chi_s(L)$
in the two ensembles is shown in Fig.~\ref{Energy_1D} and Fig.~\ref{Spinsus_1D}, respectively.
We observe that in the grand-canonical ensemble both
$E$ and $\chi_s$ converge quickly to the thermodynamic limit for
small system sizes.
This indicates a small correlation length $\xi$ at this temperature.
On the other hand, except for the smallest system sizes, in the canonical ensemble
both observables exhibit a linear-like behavior as a function of $1/L$.
A fit of energy density in the canonical ensemble $E_{\rm can}(L)$ to
$E_{\rm can}(L)= E(L \rightarrow \infty) + a L^{-1} $
exhibits a good $\chi^2/\text{DOF}$ ($\text{DOF}$ denotes the number of degrees of freedom),
when the data for the small sizes are discarded;
the extrapolated value $E(L \rightarrow \infty)$ matches the grand-canonical result.
Similar considerations hold for a fit of the spin susceptibility in the canonical ensemble $\chi_{s, {\rm can}}(L)$ to $\chi_{s,{\rm can}}(L)= \chi_s (L \rightarrow \infty) + a L^{-1}$.
Moreover, a fit of $E_{\rm can}(L)$ to $E(L\rightarrow \infty) + a L^{-d}$, leaving $d$ as a free parameter,
gives $d=1.05(2)$ when the smallest lattice size taken into account for the fit is
$L_{\rm min}=16$. An equivalent fit for $\chi_s(L)$ gives $d=1.04(2)$, when $L_{\rm min}=12$.
This confirms that finite-size corrections of observables in the canonical ensemble are $\propto 1/L$.
\begin{figure}
\includegraphics[width=8cm, height=6cm]{Energy_2Db05}
\centering
\caption{ Same as Fig.~\ref{Energy_1D} for the 2D Hubbard model.
The red line is a linear fit of the canonical ensemble data to $E_{\rm can}(L)=E(L \rightarrow \infty)+a/L^2$, with $E(L \rightarrow \infty)=-0.1387(1)$ and $a=-0.714(2)$
, where the minimum lattice size taken into account is $L_{\rm min}=6$. }
\label{Energy_2Db05}
\end{figure}
\begin{figure}[b]
\includegraphics[width=8cm, height=6cm]{Spinsus_2Db05}
\centering
\caption{Same as Fig.~\ref{Energy_2Db05} for the spin susceptibility $\chi_s$.
The red line is a linear fit of the canonical ensemble data to $\chi_{s, {\rm can}}(L)= \chi_s (L \rightarrow \infty) + a/L^2$, with $\chi_s(L \rightarrow \infty)= 0.2867(1)$ and $a=0.104(1)$
, where the minimum lattice size taken into account is $L_{\rm min}=6$. }
\label{Spinsus_2Db05}
\end{figure}
\subsection{2D model}
We simulated the Hubbard model on the square lattice for both ensembles at $\beta=0.5$ and $\beta=2.0$, lattice sizes $L=4$, $6$, $8$, $10$, $12$, $14$, $16$, and at half-filling.
Figure \ref{Energy_2Db05} and Fig.~\ref{Spinsus_2Db05}
show the size behavior of $E$ and $\chi_s$ for the two ensembles at $\beta=0.5$.
The observed tiny size dependence of the observables in the grand-canonical ensemble suggests that the correlation length $\xi$
is smaller than the minimum lattice size $L=4$.
On the other hand, in the canonical ensemble
the energy density $E$ and the spin susceptibility $\chi_s$ show a linear-like behavior as function of $1/L^2$.
For a more quantitative check of the finite-size
correction in the canonical ensemble, we fitted $E_{\rm can}(L)$ to
$E_{\rm can}(L)= E(L\rightarrow\infty) + aL^{-1} +bL^{-2} + cL^{-3} $
and $\chi_{s, {\rm can}}(L)$ to an equivalent Ansatz, leaving $a,b$ and $c$ as free parameters.
Fit results for both observables show a good $\chi^2/\text{DOF}$ when $L_{\rm min}=6$, and the coefficient $a$ vanishes within error bars, whereas $b$ acquires a finite value.
On the other hand, a fit of $E_{\rm can}(L)$ to $E_{\rm can}(L)=E(L\rightarrow \infty) + bL^{-d}$, leaving $b$ and $d$ as free parameters, and of $\chi_s(L)$ to an equivalent Ansatz,
gives $d=2.05(3)$ and $d=1.9(1)$ for $E$ and $\chi_s$, respectively, when $L_{\rm min}=6$.
In line with the discussions of Sec.~\ref{sec:fscanonical},
these fit results confirm that the leading finite-size correction in the
canonical ensemble is $\propto 1/L^2$.
\begin{figure}[b]
\centering
\includegraphics[width=8cm, height=6cm]{Energy_2Db2}
\caption{Same as Fig.~\ref{Energy_2Db05} for $\beta=2.0$. The green line is the exponential fit of
the grand-canonical ensemble data with minimum size $L_{\rm min}=6$ and parameters $E(L\rightarrow\infty)=-0.717075$, $b=36$ and $c=0.67$ (see main text). }
\label{Energy_2Db2}
\end{figure}
We also simulated the 2D Hubbard model at a lower temperature $\beta=2.0$. A corresponding comparison of the finite-size energy density for the grand-canonical and canonical ensemble is shown in Fig.~\ref{Energy_2Db2}.
Generically, finite-size corrections in the canonical ensemble are expected to be temperature dependent.
On the other hand, the exponential correction characterized by the correlation length in the grand-canonical ensemble may start to be relevant at a lower temperature, because of an increased correlation length.
The data shown in Fig.~\ref{Energy_2Db2} exhibit a visible decay of the energy density in the grand-canonical ensemble $E_{\rm gc}$, on increasing the system size.
As a guide to the eye, we fitted $E_{\rm gc}$ to $E_{\rm gc}(L)= E(L\rightarrow \infty)+ b\cdot e^{-L/c}$.
The finite-size values of $E$ in the canonical ensemble show a nonmonotonic behavior between $L=4$ and $6$, which might be due
to a combination of various sources of finite-size corrections, such as the one $\propto 1/V$ originating from the particle-number constraint, the one related to the correlation length, and the residual correction term due to the regular part of the free energy.
Nevertheless, a finite-size dependence $\propto 1/L^2$ can be clearly observed in Fig.~\ref{Energy_2Db2} for $L>6$, with a smaller slope compared to the $\beta=0.5$ case (compare with Fig.~\ref{Energy_2Db05}).
\begin{figure}
\includegraphics[width=8cm, height=6cm]{Sign_2Db2_doping}
\centering
\caption{Finite-size data of MC average sign $\<\text{sign}\>$ for the 2D Hubbard model at $\beta=2.0$ in the grand-canonical and canonical ensemble, with a $\frac{3}{8}$-filling fraction and for lattice sizes up to $L=20$.}
\label{Sign_2Db2_doping}
\end{figure}
\begin{figure}[b]
\includegraphics[width=8cm, height=6cm]{Energy_2Db2_doping}
\centering
\caption{Finite-size data of energy density $E$ for the same case as in Fig.~\ref{Sign_2Db2_doping}.}
\label{Energy_2Db2_doping}
\end{figure}
We note that the auxiliary field QMC for the grand-canonical ensemble has a mild sign-problem under doping the system away from half filling, provided that the temperature is high enough.
Here we also tested the efficiency of the canonical ensemble QMC method under doping.
To this end, for every lattice size we tuned the chemical potential $\mu$ such that the expectation value of the number of particles in the grand-canonical ensemble
matches the desired number $N_0$ of particles in the canonical ensemble.
Subsequently, the canonical ensemble is realized by introducing a Lagrange multiplier, as discussed in Sec.~\ref{sec:method:constraint}.
In order to test the sign-problem, we also calculated the average sign during the MC simulation:
\begin{equation}\label{def_sign}
\langle \text{sign} \rangle = \frac{\sum_C Re[e^{-S(C)}]}{\sum_C |Re[e^{-S(C)}]|}
\end{equation}
where $S(C)$ is the action for the MC configuration $C$, so that the corresponding statistical weight is $\propto e^{-S(C)}$.
The sign is not necessarily
a real number (when the MC is sign-problem free, $S(C)$ is real and $\langle \text{sign} \rangle =1$).
On the other hand, the expectation value of observables can be computed via a reweighting scheme only when $\langle \text{sign} \rangle $
is not too small.
Figure \ref{Sign_2Db2_doping} shows the average sign during MC of a doped 2D Hubbard model at $\frac{3}{8}$ filling, which in the grand-canonical ensemble system does not exhibit a significant sign-problem at an inverse temperature of $\beta=2.0$.
Quite remarkably, when the number of particles is fixed in the canonical ensemble,
the code still exhibits an average sign higher than $0.97$, for system sizes up to $L=20$.
This confirms the feasibility of our QMC method for the canonical ensemble, even under doping.
Figure \ref{Energy_2Db2_doping} shows the finite-size behavior of energy density of the two ensembles at $\frac{3}{8}$ filling. Similar to the results at half-filling, and in line with the analysis of Sec.~\ref{sec:fscanonical}, the energy density in the grand-canonical ensemble exhibits very small finite-size corrections when $L\ge 10$, whereas in the canonical ensemble we observe a finite-size correction approximately linear in $1/L^2$ for $L\ge 12$.
\section{Summary}
\label{sec:conclusions}
In this paper we have introduced a method to simulate fermionic models in the canonical ensemble. It consists in an auxiliary field QMC simulation, where the Hamiltonian is supplemented by an additional Lagrange multiplier, which constraints the particle number. The method can implemented using the ALF package for fermionic simulations \cite{ALF}.
In general, we find that canonical simulations are more computationally demanding than the corresponding ones in the grand-canonical ensemble.
Although in the presence of short-ranged interactions the grand-canonical and the canonical ensemble are equivalent in the thermodynamic limit, their approach to the infinite-volume limit is distinctively different. In the canonical ensemble the observables are generically found to display a finite-size correction which is proportional to the inverse volume. In Sec.~\ref{sec:fscanonical} we prove an exact formula for the leading finite-size correction of the free energy density, the energy density, and other observables. Such a correction is controlled by the charge susceptibility and is found to be proportional to the inverse volume. This result is further substantiated by an exact calculation for the one-dimensional Ising model reported in the Appendix. Our numerical simulations of the Hubbard model reported in Sec.~\ref{sec:qmc} confirm the presence of finite-size corrections proportional to the inverse volume in the canonical ensemble. In line with previous theoretical results, in the presence of a finite correlation length and for periodic boundary conditions, observables computed in the grand-canonical ensemble display a faster approach to the thermodynamic limit, such that the leading finite-size correction is exponential in the ratio of the linear size over the correlation length.
\noindent
{\bf Note added}: After completing this paper we became aware of related research presented in Ref.~\cite{GGD-17}, which investigates the effect of a constraint within statistical field theory.
\acknowledgments
FFA acknowledges useful conversations with A. Sandvik. FPT acknowledges useful communications with M. Gross. This work was supported by the German Research Foundation (DFG) through Grant No. SFB 1170 ToCoTronics and Grant No. FOR 1807. We acknowledge the computing time granted by the John von Neumann Institute for Computing (NIC) and provided on the supercomputer JURECA \cite{Jureca16} at the J\"ulich Supercomputing Centre.
| b331e37c2834059c1b45e2b5f1a47509f88620cb | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
In the past two decades, a variety of cosmological data \cite{DWeinberg:2013} has been pointing to the conclusion that the expansion of the Universe is accelerating at present.
The favored explanation, the $\Lambda$CDM model, constitutes the standard cosmological paradigm. In this model, $\Lambda$ is the cosmological constant, which drives the accelerated expansion, and cold dark matter (CDM) forms the large-scale structures in the Universe. However, the model suffers from theoretical and conceptual issues, such as the cosmological constant and coincidence problems, as well as observational challenges, such as the description of small-scale cosmological structures, see \cite{Bull:2016} for a review. For these reasons, a long list of alternatives have been explored, mainly in the form of dynamical dark energy (DE) or modifications of general relativity (see \cite{Copeland:2006wr,Clifton:2011jh} for reviews). Very different alternatives, also compliant with some of the cosmological data, have been proposed such as models where the acceleration effects are explained by quantum effects \cite{GonzalezDiaz:2006tr,GonzalezDiaz:2008ci,GonzalezDiaz:2008ba,RozasFernandez2017}, or also the averaging approach to cosmology \cite {Smale:2011}.
Here we will consider a well-studied variation of $\Lambda$CDM: the unified dark fluid approach, also known as quartessence, sometimes as unified dark energy, but more usually known as unified dark matter (UDM)\footnote{The name ``unified dark matter'' that prevails in the literature is misleading. To our knowledge, it is a colloquial simplification of ``unified dark matter-energy'', which was the original meaning of the acronym UDM first proposed in \cite{Makler:2002jv}. We will follow the original proposal, reinstating the naming ``unified dark matter-energy'' as the meaning of the well-established acronym UDM.}. A plethora of UDM models have been proposed (see \cite{Bertacca:2010ct} for a review) after the pioneering introduction of the Chaplygin gas \cite{Kamenshchik:2001cp,Bilic:2001cg,Bento:2002ps}.
The unification of dark matter and dark energy is an interesting approach that assumes the existence of a single fluid capable of accounting for both the accelerated expansion at late times and the large-scale structure formation at early times, due to the evolution of its equation of state (EOS) and speed of sound. In principle this is more efficient than postulating two different fluids and equally valid, since the nature of the fluids is still elusive. It also has the advantage of evading, by definition, the coincidence problem \cite{cp}. These models substantially alleviate as well the tension between some recent high and low redshift measurements \cite{Camera:2017tws}.
A serious issue in most UDM models is the presence of an effective speed of sound that can be very different from zero during the cosmological evolution. This prevents the dark fluid to cluster below a thresholding scale (the Jeans scale) \cite{Hu:1998kj,Garriga:1999vw,Pietrobon:2008js}. In addition, the evolution of the gravitational potential may also give rise to a strong signature in the integrated Sachs Wolfe (ISW) effect \cite{Bertacca:2007cv}. It is therefore crucial to make sure that the single dark fluid is able to cluster and create the observed cosmic structures as well as reproducing the well-known pattern of cosmic microwave background (CMB) temperature anisotropies \cite{Carturan:2002si}. However, for the majority of UDM models in the literature, these requirements, together with the necessity of having a background evolution that complies with observations, lead to a severe fine-tuning of the parameters, to the point that the models become almost indistinguishable from $\Lambda$CDM and are thus less interesting \cite{Sandvik:2002jz, Scherrer:2004au, Giannakis:2005kr, Piattella:2009da}.
The problem of the lack of clustering, or production of oscillations, can be avoided with a technique introduced in \cite{Bertacca:2008uf}. In particular, the dark fluid is a scalar field, $\varphi$, with a noncanonical kinetic term, i.e., a term $f(\dot\varphi^2)$ instead of the standard $\dot\varphi^2/2$. In this way it was possible to build a UDM model with a small effective sound speed that allows structure formation and has a weak ISW effect, being compliant with weak lensing data \cite{Camera:2009uz, Camera:2010wm}. This model has, however, the same background evolution as $\Lambda$CDM. A more recent alternative are the so-called UDM models with fast transition where, during a short period, the effective speed of sound can be large, but is otherwise zero. This produces a fast transition between a CDM-like era, with an Einstein-de Sitter evolution, and an accelerated DE-like era, and allows for structure formation. In addition, these models are not forced by construction to have the same background evolution as $\Lambda$CDM and are free from the problem of fine-tuning of the parameters that plagues many UDM models. The thermodynamics of a UDM model with fast transition was explored in \cite{Radicella:2014nka}.
The dynamics of UDM models with fast transition can be prescribed in three different ways: starting from either the EOS $w$, the pressure $p$ or the energy density $\rho$. The first UDM model with fast transition was introduced in \cite{Piattella:2009kt} and prescribed the evolution of $p$. The pressure and energy density were related by a barotropic EOS, $p=p(\rho)$ and the perturbations were adiabatic. A second UDM model with fast transition was presented in \cite{Bertacca:2010mt} and was built from a k-essence \cite{Chiba:1999ka, ArmendarizPicon:2000ah} scalar field Lagrangian (see also \cite{Kang:2007vs, Cruz:2008cwa, RozasFernandez:2011je, Rozas-Fernandez:2014tsa}). This model also prescribed $p$ but, differently from the first one, since it is based on a scalar field the perturbations are naturally nonadiabatic \cite{DiezTejedor:2005fz, Bilic:2008zk}, allowing for a small Jeans length even when the speed of sound is non-negligible. The model also contains a future attractor that acts as an effective cosmological constant\footnote{A scalar field with a potential that admits a minimum $V_{0}=V(\phi_{0})\not =0$ is equivalent to a cosmological constant $\rho_{\Lambda}=V_{0}$ and a scalar field in a potential $\tilde{V}=V-V_{0}$.}, $\rho_{\infty}$; i.e., an asymptotic limit $w = -1$ is built in. A third UDM model with fast transition was proposed in \cite{Bruni:2012sn}. This is a phenomenological model, with the dynamics prescribed through the fluid density $\rho$, and it has adiabatic perturbations.
Models with a fast transition might also be a step towards a unified description of dark matter, dark energy and inflation \cite{Liddle:2006qz} but, regardless of that possibility, they are considered among the most promising UDM models \cite{Amendola:2016saw}. Even though they are built with the goal of enabling structure formation, it is also important to test them at the background level since they may have a background evolution quite distinct from $\Lambda$CDM. In particular, such tests will constrain the rapidity of the transition and may already give an indication whether the allowed rapidity range favors structure formation. The phenomenological UDM model, and variations of it, were recently constrained at background level in \cite{Lazkoz:2016hmh}. In the present work, we apply supernova, galaxy clustering and CMB data to test the scalar field UDM model of \cite{Bertacca:2010mt}, constraining its parameters and making a statistical model comparison with both $\Lambda$CDM and the phenomenological UDM model of \cite{Bruni:2012sn} tested in \cite{Lazkoz:2016hmh}.
In the rest of the paper, we present in Sec.~\ref{sec:UDMmodelfast} the UDM model that will be tested in Sec.~\ref{sec:results} using the data and methods described in Sec.~\ref{sec:constraints}. We conclude with a summary and some remarks in Sec.~\ref{sec:concl}.
\section{The UDM model}
\label{sec:UDMmodelfast}
We consider the scalar field UDM model proposed in \cite{Bertacca:2010mt}, where the evolution of the pressure has the following form:
\begin{equation}
\label{ptanh}
p(a) = -\rho_\infty \left\{\frac{1}{2} + \frac{1}{2}\tanh\left[\frac{\beta}{3} \left(a^3 - a_{\rm t}^3\right)\right]\right\}\,.
\end{equation}
This model allows for a fast transition in the pressure evolution, since for large values of $\beta$ the $\tanh$ function tends to a step function. The transition occurs at a scale factor $a_t$, with rapidity parameterized by $\beta$, while $\rho_\infty$ parameterizes the pressure amplitude. The fluid goes from an Einstein-de Sitter DM era ($p=0$ at early times), through $p(a_t)=-\rho_\infty/2$ at transition, to a DE era at late times (with $p$ reaching $-\rho_\infty$ the sooner for faster transitions).
Considering a Friedmann-Lema\^{i}tre-Robertson-Walker (FLRW) background metric (and a frame with proper time coinciding with the cosmic time), the density can be derived from the pressure using the energy conservation equation
\begin{equation}
\label{coneq}
\dot{\rho} = -3H\left(\rho + p\right)=-3H\rho\left(1 + w\right)\,,
\end{equation}
where $w = p/\rho$ is the EOS and the dot means differentiation with respect to time. The density is obtained from the pressure by integrating Eq.~(\ref{coneq}):
\begin{equation}
\label{tanh}
\rho(a) = \rho_\infty \left\{\frac{1}{2}+\frac{3}{2\beta}a^{-3}\ln\left(\cosh\left[\frac{\beta}{3} \left(a^3 - a_{\rm t}^3\right)\right]\right)\right\} + \rho_{c0} a^{-3}\,.
\end{equation}
The integration introduces another constant. It is usual to choose it as the amplitude of a "CDM sector of the UDM": $\rho_{c0}$, defined at $a=1$. Note that the density does not have a fast transition, since the $\ln(\cosh)$ function is not a step function. The density decreases smoothly from its maximum amplitude at $a=0$, through $\rho(a_t) =(\rho_{c0}a_t^{-3}+\rho_\infty)$ at transition, to $\rho_\infty$ when $a \rightarrow \infty$. Note also that for fast transitions (large $\beta$) and after the transition, $\tanh \sim 1$ and $\ln[\cosh(x)]/x \sim 1$, and thus $p \sim -\rho_\infty$ and $w \sim -\rho_\infty / (\rho_\infty + \rho_{c0})$. This means that fastest models become degenerate and are more similar to $\Lambda$CDM than the slower ones (with the exception of the singular case $\beta=0$).
The UDM model contain thus four parameters: $\rho_{c0}$, $\rho_\infty$, $\beta$ and $a_t$. With this choice of parameters, the density is written as the sum of three parts: the CDM-like term $\rho_c(a)=\rho_{c0} a^{-3}$, a constant term $\rho(a)=\rho_\infty / 2$ and the $\ln\cosh$ term, with the latter two defining a ``dark energy sector''. To compare UDM models with $\Lambda$CDM, it is useful to define today's densities for these two sectors. Introducing the critical density today, $\rho_{\rm cr}=3H_0^{2}$, we define the two dimensionless density parameters:
\begin{equation}
\Omega_c = \frac{\rho_{c0}}{3H_0^2}
\end{equation}
and
\begin{equation}
\label{omegaDEdef}
\Omega_{DE}=\frac{\rho_\infty}{3H_0^2} \left\{\frac{1}{2}+\frac{3}{2\beta}\ln\left(\cosh\left[\frac{\beta}{3} \left(1 - a_{\rm t}^3\right)\right] \right) \right\}\,.
\end{equation}
All the background probes we will use in the likelihood analysis depend on the Hubble function
\begin{equation}\label{friedman-2}
E^2(a)=H^2/H_0^2=\Omega_{r} a^{-4}+ \Omega_{b} a^{-3}+ \Omega_{\rm UDM}(a)\,,
\end{equation}
where $\Omega_b$ and $\Omega_r=2.49\times10^{-5}h^{-2}$ are the baryonic matter and radiation densities, respectively, and
\begin{equation}\label{omegaUDM}
\Omega_{\rm UDM}(a)=\Omega_{c}a^{-3}+\Omega_{DE}\left\{ \frac{1}{2}+\frac{3}{2\beta}a^{-3}\ln\left\{\cosh\left[\frac{\beta}{3} \left(a^3 - a_{\rm t}^3\right)\right] \right\} \right\}\;.
\end{equation}
The four parameters $\Omega_{DE}$, $\Omega_c$, $\beta$ and $a_t$ are not all independent. Indeed, applying Friedmann's equation, $\sum_i\Omega_i =1$, we can write
\begin{equation}
\label{omegaDE}
\Omega_{DE}=\frac{1-\Omega_{r}-\Omega_{b}-\Omega_{c}}{ \frac{1}{2}+\frac{3}{2\beta}\ln\left\{\cosh\left[\frac{\beta}{3} \left(1 - a_{\rm t}^3\right)\right] \right\} }.
\end{equation}
The definition of two sectors allows the introduction of an EOS of the dark energy sector,
\begin{equation}
\label{wDE}
w_{\rm DE}(a)=\frac{p(a)}{\rho(a)-\rho_{\rm c0}a^{-3}}\;,
\end{equation}
in addition to the EOS $w(a)=p(a)/\rho(a)$.
We finally note that an explicit analytical Lagrangian can be written for this model, since the general Lagrangian for a UDM scalar field $\varphi$, within the framework of k-essence, is
\begin{equation}
L= L_G+ L_\varphi=\frac{1}{16\pi G}R+ L_\varphi(\varphi,X),
\end{equation}
where $X$ is the kinetic term and the pressure can be identified with the term $p=L_\varphi(\varphi,X)$.
\section{Methodology}
\label{sec:constraints}
We test the model with a Markov chain Monte Carlo (MCMC) exploration of the parameter space \cite{Christensen:2001gj,Lewis:2002ah}, combining various probes of the expansion history of the Universe: luminosity distances to type Ia supernovae, baryon acoustic oscillation scale parameter, Alcock-Paczynski distortion parameter, and include CMB distance priors. The various data sets are uncorrelated and thus the total $\chi^2$ used in the analysis is simply the sum
\begin{equation}
\chi^2 = \chi^{2}_{CMB} + \chi^{2}_{BAO} + \chi^{2}_{SN}\;.
\end{equation}
\subsection{SNe Ia data}
As in the previous analysis \cite{Lazkoz:2016hmh}, we use the Union2.1 compilation \cite{Suzuki:2011hu}, which provides not only the distance modulus $\mu(z_i)$ for each SN, but also the full statistical plus systematics covariance matrix. The data set consists of $580$ type Ia supernovae with redshifts in the interval $0.015 < z < 1.414$. The cosmological model is tested through the dimensionless luminosity distance
\begin{equation}
d_L(z) = (1+z) \int_0^z \frac{dz'}{E(z')}\, ,
\end{equation}
which depends on the dimensionless Hubble function $E(z)$ and is directly related to the observable: the distance modulus
\begin{equation}
\mu(z) = 5 \log_{10} d_L(z) + \mu_{0} \;.
\end{equation}
This relation includes an additive nuisance parameter, $\mu_{0}$, involving the values of the speed of light $c$, Hubble constant $H_{0}$, and SNe Ia absolute magnitudes.
The likelihood is assumed to be Gaussian and is defined as
\begin{equation}
\chi^2_{SN}=\textbf{X}_{SN}^T \cdot \textbf{C}_{SN}^{-1} \cdot \textbf{X}_{SN} \,,
\end{equation}
where $\textbf{X}_{SN}$ is the difference vector with elements $X_{iSN}=\mu_{\rm model}(z_{i}) - \mu_{\rm obs}(z_{i})$ and $\textbf{C}_{SN}$ is the data covariance matrix. We analytically marginalize over the additive parameter $\mu_0$, as an alternative to including it in the MCMC parameter space. The resulting $\chi^2$ is given by \cite{Conley:2011ku}
\begin{equation}
\chi^2_{SN} = a + \log \frac{d}{2 \pi } - \frac{b^2}{d} \, ;
\end{equation}
where $a \equiv \textbf{X}_{SN}^{T} \cdot \textbf{C}_{SN}^{-1} \cdot \textbf{X}_{SN}$, $b \equiv \textbf{X}_{SN}^{T} \cdot \textbf{C}_{SN} ^{-1} \cdot \bf{1}$, and $d \equiv {\bf 1}^{T} \cdot \textbf{C}_{SN}^{-1} \cdot \bf{1}$, with $\bf{1}$ being the identity matrix.
\subsection{Baryon acoustic oscillation (BAO) data}
Unlike the previous analysis \cite{Lazkoz:2016hmh}, we will now use the baryon acoustic oscillation scale parameter $A(z)$ and the Alcock-Paczynski distortion parameter $F(z)$ provided by the WiggleZ Dark Energy Survey \cite{Blake:2012pj}, as the BAO observables. They are defined as
\begin{eqnarray}
A(z) & \equiv & 100 D_V(z) \sqrt{\Omega_m h^2} /cz \, , \\
F (z) & \equiv & (1+z) D_A(z) H(z)/c \, ,
\end{eqnarray}
probing the angular-diameter distance
\begin{equation}
D_A(z) = \frac{c}{(1+z)} \int^z_0 \frac{dz'}{H(z')}
\end{equation}
and the volume-averaged distance
\begin{equation}
D_V (z) = \left[ (1+z)^2 D_A(z)^2 \frac{c z}{H(z)} \right]^{1/3} \, .
\end{equation}
WiggleZ measured these observables in three overlapping redshift bins, with effective redshifts $(z_{1}, \, z_{2}, \, z_{3}) = (0.44, \, 0.60, \, 0.73)$. The data values are
\begin{eqnarray}
\textbf{X}_{obs} &=& (A_1, A_2, A_3, F_1, F_2, F_3) \\
&=& (0.474, 0.442, 0.424, 0.482, 0.650, 0.865)
\end{eqnarray}
with correlated errors described by the covariance matrix
\begin{eqnarray}
\textbf{C}_{BAO}=10^{-3} \times
\left( \begin{array}{cccccc}
1.156 & 0.211 & 0.0 & 0.400 & 0.234 & 0.0 \\
0.211 & 0.400 & 0.189 & 0.118 & 0.276 & 0.336 \\
0.0 & 0.189 & 0.441 & 0.0 & 0.167 & 0.399 \\
0.400 & 0.118 & 0.0 & 2.401 & 1.350 & 0.0 \\
0.234 & 0.276 & 0.167 & 1.350 & 2.809 & 1.934 \\
0.0 & 0.336 & 0.399 & 0.0 & 1.934 & 5.329 \\
\end{array} \right)
\end{eqnarray}
The BAO contribution to the total $\chi^2$ is
\begin{equation}
\chi^2_{BAO} = \textbf{X}_{BAO}^T \cdot \textbf{C}_{BAO}^{-1} \cdot \textbf{X}_{BAO}
\end{equation}
where $\textbf{X}_{BAO} = (\textbf{X}_{obs} - \textbf{X}_{mod} )$ is the difference vector.
\subsection{Priors}
In order to reduce the volume of the parameter space in the MCMC analysis, it is useful to include the so-called distance priors \cite{Wang2013} in our analysis. These are priors on the CMB shift parameters, geometrical quantities that effectively summarize the CMB data, since they capture the degeneracies between the parameters that determine the CMB power spectrum \cite{WangMukherjee2007}.
The first shift parameter is a dimensionless distance to the photon-decoupling surface
\begin{equation}
R \equiv \sqrt{\Omega_m H^2_0} \,\frac{r(z_*)}{c} \;,
\end{equation}
defined from the comoving distance to the photon-decoupling surface
\begin{equation}
r(z_*) = c\,\int_0^{z_*} \frac{dz'}{H(z')}\,.
\end{equation}
The redshift of the photon-decoupling surface may be computed from a fitting formula
\cite{Hu:1995en}
\begin{equation}
z_{*}= 1048 \left[ 1+0.00124(\Omega_b h^2)^{-0.738} \right] \left[ 1 + g_1 (\Omega_m h^2)^{g_2} \right] ,
\end{equation}
where $g_1$ and $g_2$ are functions of the physical baryon density.
The second shift parameter is a dimensionless size of the sound horizon at the photon-decoupling epoch, i.e., the angular scale of the sound horizon at the photon-decoupling epoch
\begin{equation}
l_a \equiv \pi \frac{r(z_*)}{r_s(z_*)} \;,
\end{equation}
defined from the comoving sound horizon
\begin{eqnarray}
r_s(z_*) &=& \int_0^{a_*} \frac{da'}{a'^2} \frac{c_s}{H(a')} ,
\end{eqnarray}
where the sound speed $c_s$ depends on the physical baryon density and the temperature of the CMB \cite{Wang2013}. We take $T_{CMB}=2.725$ K \cite{Fixsen:2009ug}.
The distance prior we use is the Gaussian fit to the joint probability density function of $R$ and $l_a$ presented in \cite{Wang2013} and derived from \textit{Planck} first release data \cite{Ade:2013zuv} and \textit{WMAP} 7 \cite{wmap7} and \textit{WMAP9} \cite{wmap9} temperature and polarization data. Since the shift parameters correlate with the physical baryon density $\Omega_b h^2$ the prior also includes the baryon density and it is a three-dimensional Gaussian with mean
\begin{equation}
\left(\langle l_a \rangle , \langle R \rangle, \langle \Omega_b h^2 \rangle\right) =
(301.57 , 1.7407 , 0.02228)\,
\end{equation}
and covariance
\begin{eqnarray}\label{Vcmb}
\textbf{C}_{CMB} = 10^{-2} \times
\left( \begin{array}{ccc}
3.24 & 0.08883 & -0.0022869\\
0.08883 & 0.008836 & -1.953\times 10^{-4}\\
-0.0022869 & -1.953 \times 10^{-4} & 9.\times 10^{-6}\\
\end{array} \right) \; .
\end{eqnarray}
The CMB data contribution to the total $\chi^2$ is thus
\begin{equation}
\chi^2_{CMB} = \textbf{X}_{CMB}^T \cdot \textbf{C}_{CMB}^{-1} \cdot \textbf{X}_{CMB} \; ,
\end{equation}
where the three-dimensional difference vector between model and observations is
\begin{eqnarray} \label{data-vector}
\textbf{X}_{CMB} = \left( \begin{array}{c}
l_a - \langle l_a \rangle \\
R - \langle R \rangle \\
\Omega_b h^2 - \langle \Omega_b h^2 \rangle \end{array} \right) \; .
\end{eqnarray}
Besides the distance priors, we also include some broad and flat conditions: the dark matter density must be positive $0<\Omega_{c}<1$; the baryonic matter density must be positive and smaller than the dark matter density $0 < \Omega_b < \Omega_{c}$; the Hubble function must be positive for all values of the scale factor $a$, $ E(a) > 0 $ ; and $0 < a_t < 1 $ because we want the transition to actually have happened. Finally, the Hubble constant $H_0$ is analytically marginalized in the SN likelihood and is left as a free parameter, with a broad flat prior, in the BAO and CMB likelihoods.
\section{Analysis and Results}
\label{sec:results}
We ran a set of Markov chains on the five-dimensional parameter space $(h, \Omega_c, \Omega_b, a_t, \beta)$, using the following three-step procedure. We start by running a short preliminary chain of around 20 000 iterations in order to find the region of maximum probability density. Then we make a second run for around 50 000 iterations to find a tentative covariance matrix. Finally, we start the final chain, using the previously found covariance matrix as a proposal step. The final chain has around 200 000 points and we assess its convergence using the ratio of variances proposed in \cite{Dunkley:2005}. Differently from the more standard Gelman and Rubin ratio of variances that compare parallel chains \cite{Tereno:2005}, the convergence ratio of \cite{Dunkley:2005} uses only one chain. It is based on a spectral analysis of the single MCMC chain and in order to perform the test we compute the power spectrum of the chain on 1000 Fourier modes.
Besides the UDM scenario, we also ran an MCMC for the $\Lambda$CDM scenario. Figure~\ref{fig:UDM} shows the posterior probabilities for each parameter of the UDM and $\Lambda$CDM models, along with 1- and 2-$\sigma$ two-dimensional confidence regions. Table~\ref{table-UDM} gives the corresponding median and marginalized 1-$\sigma$ interval for each chain parameter and some derived parameters.
\begin{figure*}
\begin{minipage}{1.0\textwidth}
\centering
\includegraphics[width=1.0\textwidth]{Figure1-combined.pdf}~~~~~
\caption{Posterior distribution from the full-range MCMC chains.
Diagonal panels: One-dimensional marginalized posterior distributions for UDM (red, solid lines) and $\Lambda$CDM (blue, dashed lines) parameters. Off-diagonal panels: 1- and 2-$\sigma$ two-dimensional marginalized contours for UDM (red) and $\Lambda$CDM (blue) parameters.}
\label{fig:UDM}
\end{minipage}
\end{figure*}
\begin{table}[h]
\centering
\caption{Median and 1-$\sigma$ uncertainty for the UDM and $\Lambda$CDM model parameters, from the full-range MCMC chains.
}\label{table-UDM}
\vspace{5pt}
\begin{tabular} {lcc}
\hline
Parameter & UDM & $\Lambda$CDM\\
\hline
\hline
$h$ & $0.685\pm 0.012$ & $0.684\pm 0.011$ \\
$\Omega_c$ & $0.259^{+0.012}_{-0.018}$ & $0.253\pm 0.013$
\vspace{3pt}
\\
$\Omega_b$ & $0.0476^{+0.0011}_{-0.0014}$ & $0.0476\pm 0.0011$
\vspace{3pt}
\\
$a_t$ & $0.22^{+0.13}_{-0.15}$ &
\vspace{3pt}
\\
$\beta$ & $227500^{+200000}_{-200000}$ &
\vspace{3pt}
\\
\hline
$\Omega_{DE} $ & $0.693^{+0.019}_{-0.013} $&
\vspace{3pt}
\\
$w_{DE} $ & $-1.011^{+0.011}_{-0.0039} $&
\vspace{3pt}
\\
\hline
$w $ & $-0.735^{+0.013}_{-0.015} $&\\
\hline
\end{tabular}
\end{table}
The constraints on the three standard parameters are similar in the two models. The probability contours of the Hubble parameter vs densities show the usual anti-correlations that arise because Hubble function and distance measurements probe physical densities $\Omega_ih^2$. The main new feature is a slight correlation between the scale factor of transition $a_t$ and $\Omega_c$, especially for higher values of $a_t$ (and a corresponding anti-correlation with $h$). This degeneracy broadens the $\Omega_c$ contours, being responsible for the decrease of precision in the $\Omega_c$ estimate quoted in Table~\ref{table-UDM}. This differs from the behavior found in the analysis of the phenomenological UDM models \cite{Lazkoz:2016hmh}, where the constraint on $\Omega_c$ was found to be stronger than in the $\Lambda$CDM model, even though the evidence was not conclusive in favor of that UDM model.
For model comparison purposes, we start by noticing in Table~\ref{table-comp} that the UDM best fit has a lower $\chi^2$ value than the one found in the $\Lambda$CDM analysis. This may be due to overfitting, and indeed the best-fit reduced $\chi^2$ is larger than for the $\Lambda$CDM case. A more robust way to compare the models is through the ratio of the model evidences, i.e., the Bayes factor \cite{Trotta:2005ar}. We compute the evidence with an implementation of the nested sampling algorithm of \cite{Mukherjee:2005wg}. In particular, we use $10^3$ sample points, chosen randomly, and compute the evidence in up to $10^4$ steps. We repeat the procedure 100 times, varying the sample points, and quote the average evidence from the 100 realizations. We obtain a Bayes factor very close to 0, and thus the model comparison is highly inconclusive, according to Jeffreys' scale \cite{Gordon:2007xm}.
\begin{table}[h!]
\centering
\caption{Values from five methods to perform model comparison between UDM, $\Lambda$CDM and the phenomenological UDM.}\label{table-comp}
\vspace{5pt}
\begin{tabular} {cccc}
\hline
& UDM & $\Lambda$CDM & $\rm{UDM_{ph}}$\\
\hline
{$\chi^2_{\rm min}$} & $552.59$& $552.77$ & 552.75\\
{$\chi^2_{\rm red}$}& $0.9478$ & $0.9449$ & 0.9481 \\
{$\rm \ln B_{U\Lambda}$} & $-0.0196$ & $0$ & 0.6850\\
BIC & 584.485 & 571.902 & 584.644\\
DIC & 553.250 & 552.770 & 552.814\\
\hline
\end{tabular}
\end{table}
With a Bayes factor so close to 0, we decided to investigate if the behavior would be any different when using approximate evidence measures, namely information criteria. The Bayesian information criterion (BIC) is defined as \cite{Liddle:2004}
\begin{equation}
{\rm BIC} = -2\ln{\rm L_{max}} + k \ln{N}.
\end{equation}
Since the number of data points used, $N$, was the same for the two models, BIC directly penalizes the lower minimum $\chi^2$ of UDM with the higher number of free parameters $k$. For the deviance information criterion (DIC), we followed \cite{SaezGomez:2016} and computed
\begin{equation}
{\rm DIC} = 2\left<\chi^2\right> - \chi^2_{\rm \min},
\end{equation}
where the average $\chi^2$ were computed from the chains and not with the nested sampling code. The results are consistent with the comparison of evidence in that both information criteria assign a weak but inconclusive preference to $\Lambda$CDM.
We are also interested in comparing the stronger motivated scalar field UDM model with the phenomenological one. For that purpose we made a new analysis of the latter, testing it with the same set of data used in our present analysis. The results from this second model comparison analysis are also summarized in Table~\ref{table-comp}. Model comparison between the two UDM models is more direct, since both have the same number of parameters and data points. Therefore, BIC reduces to a measure of the best fit, which is slightly in favor of the scalar field model. It is interesting to note that even though the scalar field model shows a better best fit $\chi^2$, it has a worse $\chi^2$ behavior on average and consequently a lower DIC value and evidence. Again, the analysis does not favor one model over the other, with a weak but inconclusive preference for the phenomenological model.
We can also look at the dark energy sector of the UDM model. A DE density may be defined as in Eq.~(\ref{omegaDEdef}) and its value computed from Eq.~(\ref{omegaDE}) as a function of all the other parameters. The corresponding EOS is dynamical and can be computed from Eq.~(\ref{wDE}). The constraints on $\Omega_{\rm DE}$ and $w_{\rm DE}(a=1)$, derived from the MCMC chains, are shown in Table~\ref{table-UDM}. The evolution of $w_{\rm DE}$ for the best-fit parameter values is shown in Fig.~\ref{fig:w}, together with its 1-$\sigma$ variation. Notice that even though $w_{DE}$ is phantom after the fast transition, approaching $w_{DE} \sim -1$ today from the negative side, the UDM fluid does not violate the null energy condition because its EOS, also shown in Fig. \ref{fig:w}, does not cross the phantom divide.
\begin{figure*}[h!]
\begin{minipage}{1.0\textwidth}
\centering
\includegraphics[width=0.5\textwidth]{Figure2-w.pdf}~~~~~
\caption{Evolution of the EOS of the UDM fluid (green) for the best-fit model parameters, including derived uncertainty, and EOS of the DE section of the fluid (blue).}
\label{fig:w}
\end{minipage}
\end{figure*}
We have thus a UDM model with fast transition that is viable given background data. Let us analyze now the behavior of its core parameters: the scale factor at the transition, $a_t$, and the rapidity of transition, $\beta$. Their constraints, also shown in Fig.~\ref{fig:UDM} and Table~\ref{table-UDM}, are weak. The 1-$\sigma$ interval for the transition redshift ranges from $z\sim 2$ to $z\sim 13$, while $\beta$ does not show a correlation with the other parameters. The posterior probability of $\beta$ shows a peaked structure. Looking in more detail into the likelihood values, we see the likelihood is essentially flat for $\beta > 1000$. The peaks in the $\beta$ posterior indicate the chain is not yet converged for this parameter, meaning there was not enough time to sample the unbound flat distribution and the chain remained occasionally stuck in some positions of the flat distribution. We have thus found that $\beta$ is unbound from above, which reflects the fact that for $\beta > \sim 1000$ the Hubble function is essentially identical for all $\beta$ values. On the other hand, $\beta$ is bound from below, we do not impose a $\beta > 0$ prior in the analysis.
These considerations led us to probe the low $\beta$ limit with better resolution. For this, we ran new chains considering only the range $\beta<100$. Given the low level of correlation with other parameters, we keep the density parameters fixed at the best-fit values, varying only $\beta$ and $a_t$. The scale factor at the transition must be kept free, since it is coupled with $\beta$ in the evolution of pressure and density, Eqs.~(\ref{ptanh}) and (\ref{tanh}), even though a degeneracy with $\beta$ does not show in Fig.~\ref{fig:UDM}. Notice also that this setup will artificially tighten the $a_t$ constraint due to its correlation with $\Omega_c$. We also ran separate chains for each data set and show the results in Fig.~\ref{fig:UDM-bgap}. We see now a sharp peak in the posterior of $\beta$ at $\beta=0$ that had not been picked up before. This point is basically a singularity in the space of UDM parameters. Indeed, in the $\beta=0$ limit, Eq.~(\ref{omegaUDM}) no longer presents a transition and the model reduces to $\Lambda$CDM, which explains its high likelihood. No transition, also means that the value of $a_t$ is meaningless, which explains the very narrow horizontal contour seen in the contour plot at $\beta=0$. As $\beta$ increases, the Hubble function starts to deviate from $\Lambda$CDM, until $\beta \sim 15$, and afterwards it approaches it again. This explains the dip in the $\beta$ posterior seen in all data sets. This effect is especially dramatic for the CMB shift parameters, which are able to reject the range $\beta < 40$.
\begin{figure}[h]
\centering
\begin{minipage}{0.85\textwidth}
\centering
\includegraphics[width=0.6\textwidth]{Figure3-100}
\caption{Posterior distribution from the slow transition MCMC chains.
Diagonal panels: One-dimensional marginalized posterior distributions for the UDM model parameters, for different data sets (SN, BAO, CMB, all combined). Off-diagonal panel: 1- and 2-$\sigma$ two-dimensional $(a_t,\beta)$ contours for the same data sets.} \label{fig:UDM-bgap}
\end{minipage}
\end{figure}
Regarding $a_t$, the noisy structure seen in its posterior corresponds to the $\beta=0$ solution, while the rest of the probability volume lies along a well-defined degeneracy in the $(\beta, a_t)$ plane. Indeed, in this regime of low $\beta$ the data is able to pick up the degeneracy that arises from the fact that a slower transition needs to occur earlier in order to be able to reach today's density ratio. We fit the degeneracy direction with a cubic polynomial $\beta-\beta_0 = (a_t/0.22)^3$ to capture the $(\beta,a_t)$ dependence in the Hubble function, Eqs.~(\ref{friedman-2}) and (\ref{omegaUDM}). Here $\beta_0=54.6$ is the average chain value of $\beta$ for $a_t=0$, while $a_t=0.22$ is the median $a_t$ value quoted in Table~\ref{table-UDM}. With these assumptions, we find the following 1-$\sigma$ constraint:
\begin{equation}
(\beta-\beta_0)\,\left(\frac{a_t}{0.22}\right)^{-3} = 24.8 \pm 5.9\,.
\end{equation}
We also need to look with higher resolution to the intermediate regime of $\beta$, to compare the likelihoods of the slow transition models with the fast transition ones. This is the regime of $\beta$ of a few hundreds, where the $\tanh$ function is not yet a step function. We thus ran a new $(a_t,\beta)$ chain restricted to $\beta < 2000$. The results of this analysis are shown in Fig.~\ref{fig:UDM-bmin}. The distribution of $a_t$ is now well constrained, showing a tight peak with a low-likelihood tail for low $a_t$ values. The tail corresponds to the slow transition regime studied in Fig.~\ref{fig:UDM-bgap}. This result then strongly favors intermediate and fast transitions over slow ones. This is confirmed by the posterior of $\beta$ that shows a strong increase from slow to fast transition, peaking around $\beta=600$. After the peak, the distribution falls down slowly with a long tail, which is just an effect of the strong prior $\beta < 2000$ imposed in this analysis, since the likelihood is essentially flat. We see then that the $\beta$ distribution is far from Gaussian and we can only find a lower limit for this parameter. From the $\Delta \chi^2$ values, we find a 1-$\sigma$ lower bound of $\beta > 300$.
\begin{figure}[h]
\centering
\begin{minipage}{0.85\textwidth}
\centering
\includegraphics[width=0.6\textwidth]{Figure4-2000}~~~~~
\caption{Posterior distribution from the fast transition MCMC chains.
Diagonal panels: One-dimensional marginalized posterior distributions for the UDM model parameters. Off-diagonal panel: 1- and 2-$\sigma$ two-dimensional $(a_t,\beta)$ contours.} \label{fig:UDM-bmin}
\end{minipage}
\end{figure}
\section{Conclusions}
\label{sec:concl}
In recent years, UDM models, for which DM and DE are described by a single dark fluid, have become increasingly popular and drawn a considerable amount of attention. These models are undoubtedly promising candidates as effective theories. In this work, we have constrained a UDM scalar field model with a fast transition. The scalar field used has a noncanonical kinetic term in its Lagrangian and accounts for both the accelerated expansion of the Universe at late times and the clustering properties of the large-scale structure of the Universe at early times. The fast transition occurs between an Einstein-de Sitter CDM-like epoch and a late accelerated DE-like epoch and allows one to have a sufficiently small Jeans length, even if the speed of sound is large during the transition, because this happens so quickly that its effect is negligible.
In this study we investigated the regimes of slow and fast transition and assessed if they were distinguishable at background level. For this analysis we tested the models using supernovae Ia, baryon acoustic oscillations and CMB distance data. We have found a lower bound constraint for the rapidity of the transition $\beta > 300$, independent of the transition redshift. Slow transition models $\beta < 40$ were ruled out, while low-likelihood intermediate rapidity models featured a correlation between the transition redshift and rapidity.
The evidence of this model was compared to the evidence of $\Lambda$CDM and a phenomenological fast transition UDM model that was previously shown to be a good fit to background data. In both comparisons the model fared well, with no conclusive evidence against it.
The preference found for the fast transition regime, which is the condition required for enabling structure formation, together with the fact that the model has a similar evidence to $\Lambda$CDM and is a k-essence type physically motivated model with a well-defined Lagrangian, makes it an interesting and viable fundamental cosmological model.
For completeness, it is worth mentioning that k-essence-type models suffer in general from the development of caustics in the nonlinear regime \cite{Babichev:2016hys} (see however \cite{Mukohyama:2016ipl}). That is to say, characteristics of equations of motion cross at some finite time rendering the k-essence scalar field no longer single valued and consequently second derivatives of the field diverge. That would suggest that k-essence models cannot be considered as fundamental. However, this issue could possibly be solved by making the metric dynamical, such that gravitational backreaction would prevent the formation of caustics \cite{Babichev:2016hys}. Another possible way out of this problem was recently proposed in \cite{Babichev:2017lrx}, by introducing a complex scalar field such that the singularity does not develop in the real time and the real time evolution always remains smooth.
\acknowledgments
We thank Vincenzo Salzano for the use of his nested sampling code and Ruth Lazkoz and Diogo Castel\~ao for discussions. We also thank the anonymous referee for having raised the point of the inconsistency of the name 'unified dark matter'.
This work was supported by Funda\c{c}\~ao para a Ci\^encia e a Tecnologia (FCT) through the research Grant No. UID/FIS/04434/2013. I.T. acknowledges support from FCT through the Investigador FCT Contract No. IF/01518/2014 and POPH/FSE (EC) by FEDER funding through the program COMPETE.
A.R.F. gratefully acknowledges support from FCT through Fellowship No. SFRH/BPD/96981/2103 (Portugal) and from Ministerio de Econom\'ia y Competitividad (Spain) through Project No. FIS2012-38816.
I.L. acknowledges financial support through research Projects No. FIS2014-57956-P (comprising FEDER funds) from Ministerio de Econom\'ia y Competitividad and No. GIC17/116-IT956-16 from the Basque Government. I.L. further acknowledges financial support from the University of the Basque Country (UPV/EHU) through PhD Grant No. 750/2014, and from FCT through the Exploratory Project No. IF/01518/2014 (GLUE) during his stay at Instituto de Astrof\'isica e Ci\^encias do Espa\c{c}o, Faculdade de Ci\^encias da Universidade de Lisboa where part of this work was carried out.
This article is based upon work from COST Action CA15117 (CANTATA), supported by COST (European Cooperation in Science and Technology).
\bibliographystyle{apsrev}
| 95f71242465657016aff267700d4a8724b5828c1 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{intro}
Multibody system dynamics is a well established discipline in the context of railway vehicle design.
It is used for new concept performance evaluation, stability, lifetime, wear prediction, etc. In general it is desirable to be able to do these analyses as fast as possible. In particular, due to the huge number of computations required, computational performance can be very important when dealing with design optimization. Nevertheless these tasks do not demand strict real-time performance.
The computational power available on today's off-the-shelf computers is getting closer to allow real-time direct numerical simulation of complex railway vehicle models. This in turn opens up new possibilities
that can greatly benefit the design, safety, and model based predictive maintenance in the railway field.
Most important applications can be considered to be HiL (Hardware in the Loop) on the design side, and on-line filtering techniques (Kalman filter alike) in the context of safety, and model based predictive maintenance.
These developments usually run on the heels of previous work done in the context of vehicle dynamics.
Symbolic multibody models have demonstrated to be an effective tool for the modeling of general multibody systems. In particular they have been shown to be very fast when using recursive $O(n^3)$ formulations \cite{samin2003}, that in turn require a parametrization based on relative coordinates. Main challenges are related to the enormous size of the expressions that the symbolic processor needs to deal with as this can limit the size of the problem to be analyzed. Recently, in \cite{Plaza2015} the authors presented a symbolic multibody library in which the concept of recursivity is extended so that it is no longer based on the formulation but, instead, on the parametrization level. This is achieved by the definition of an algebra that includes the typical mechanics operators (position vector, velocity,... ) and that deals with the recursivity that might be embedded into the parametrization. The typical tree-shaped body structure is replaced by a tree structure for points and another one for bases. This gives a fine grained control of the recursivity that, in this way, can be different for both tree structures. No limitation is imposed on the parametrization of the system. As a consequence the library allows to implement arbitrary dynamics formulations. Atomization (optimization of symbolic expression representation) is embedded into the library from the very bottom upwards. This alleviates the symbolic manipulation of expressions and lowers their complexity to a minimum. This in turn allows to obtain optimal atomizations that minimize the computational complexity and increases the size of the problems that is possible to analyze.
This article pursues to evaluate the feasibility of real-time numerical simulation of a complex locomotive multibody model using state of the art symbolic modeling techniques referred above. For our study we use the FEVE 3000 locomotive. A generic (spline based) definition for the contact surfaces of the wheels and rails, including irregularities, is used. Based on these, creep forces are modeled using a direct symbolic implementation of the standard linear Kalker model without simplifications of any kind. Bodies and rail are considered rigid with three-dimensional kinematics. No further simplifications as contact coordinate removal \cite{Shabana2005}, pre-calculated tables \cite{malvezzi2008determination}, partial linearization \cite{Escalona2015} or base parameter reduction \cite{Iriarte2015} are presented.
To that end the modeling is done based on the multibody system symbolic library \verb!lib_3D_MEC_GiNaC! \cite{Plaza2015}, using a relative parametrization with respect to the inertial reference.
The paper is structured as follows: In section \ref{sec1} the symbolic methods used in this work are briefly described. In section \ref{sec2} the description of the modeled system is presented. In section \ref{sec3} the most interesting details of the multibody modeling are presented. In section \ref{sec4} the results of the simulations are shown and discussed. Finally in section \ref{sec5} the main conclusions of this work are presented.
\section{Symbolic modeling procedures}
\label{sec1}
Simply stated, the main goal of the symbolic modeling of multibody systems can be defined as:
``\textit{to obtain a set of functions that allow for the determination of the position, velocities and accelerations of all the bodies of the system}''.
Special symbolic procedures are required if a real-time-capable fast multibody model is desired. The main features of these procedures, as
proposed in \cite{PlazaThesis2015}, are summarized below.
\subsection{Parametrization and system topology.}
In order to model the multibody system a set of geometric parameters $\mathbf{p}$ and generalized coordinates $\mathbf{q}$, along with their associated velocity $\dot{\mathbf{q}}$ and generalized accelerations $\ddot{\mathbf{q}}$ are defined. We propose to split up the classical tree-shaped body structure into two different tree-shaped structures: 1) the bases structure and 2) the points structure, see Fig.~\ref{fig:points_bases}.
In this approach, bases $\B_{j}$ and points $P_{j}$ are defined in terms of other bases $\B_{i}$ and points $P_{i}$ by the way of relative base-change or rotation matrices $\mathbf{R}_{\B_{i}}^{\B_{j}}$ and positions vectors $\mathbf{r}_{P_{i}}^{P_{j}}$. The functions used by the symbolic library \cite{ros2007lib3d} can be schematically represented as
\begin{eqnarray}
\label{eq:new_base}
\B_{i} &\xrightarrow{\mathbf{R}_{\B_{i}}^{\B_{j}}(exp_x(\mathbf{q},~t,\mathbf{p}),exp_y(\mathbf{q},t,\mathbf{p}),exp_z(\mathbf{q},t,\mathbf{p}),~exp_\phi(\mathbf{q},t,\mathbf{p}))} &\B_{j}~~~~~~~ \\
\label{eq:new_point}
P_{i} &\xrightarrow{\mathbf{r}_{P_{i}}^{P_{j}}(exp_x(\mathbf{q},t,\mathbf{p}),~exp_y(\mathbf{q},t,\mathbf{p}),~exp_z(\mathbf{q},t,\mathbf{p}),~\B_k)}& P_{j},
\end{eqnarray}
where $exp_*(\mathbf{q},t,\mathbf{p})$ represent arbitrary symbolic expressions in terms of which vectors and base-change matrices are defined. This in turn confers physical meaning to the defined coordinates and parameters. Note that, in order to illustrate the procedure, the rotation matrix appearing in Eq.~(\ref{eq:new_base}) is parametrized using Euler parameters.
\begin{figure}
\centering
\begin{minipage}{.33\textwidth}
\centering
\includegraphics[width=\textwidth]{./bases_tree.pdf}
\end{minipage}%
\begin{minipage}{.05\textwidth}
\phantom{A}
\end{minipage}%
\begin{minipage}{.6\textwidth}
\centering
\includegraphics[width=\textwidth]{./points_tree.pdf}
\end{minipage}
\caption{Illustrative examples of bases (left) and points (right) structures }
\label{fig:points_bases}
\end{figure}
This splitting of the body structure into the bases and points structures confers complete flexibility to the choice of the parametrization. A body position and orientation
no longer needs to be defined with respect to the preceding body in the tree-shaped bodies structure.
Instead, the body position is given by a point in the points structure and an orientation by a base in the bases structure.
It should be noted that the bases structure is independent of the points structure. Conversely, the points structure is dependent on the bases structure as the relative position vector components are given using arbitrary bases (note the $\B_k$ parameter in Eq.~(\ref{eq:new_point})). Finally, it should be understood that the nature of the parameterization depends on the definition of the rotation matrices and position vectors. For example, if they are defined with respect to another point or orientation the coordinates will be ``relative'', but if they are defined with respect to an absolute point or orientation they will be ``absolute''.
\subsection{``On-the-way'' atomization}
The symbolic expressions that we need to deal with can be huge. The successive multiplication of symbolic expressions leads to an explosive growth in the size of the expressions that can limit the maximum size of the multibody systems that can be analyzed. In order to deal with this problem we use a standard technique in the context of symbolic computations that we call \textit{atomization}.
Atomization is a technique that condenses a symbolic expression set by splitting their expressions into several elemental sub-expressions. We call ``atoms'' to these elemental sub-expressions. They can be defined in terms of binary operations between symbols, numbers and/or other atoms. Or as transcendental functions of atoms.
This technique is beneficial when repeated sub-expressions appear and the same atom is used to represent them. Symbolically, this means less memory -as the sub-expression is allocated in memory once- and faster
symbolic manipulations. Numerically, this implies that the repeated sub-expression is only computed once leading to computational cost savings.
In the present context, we deal with the sets of expressions related to
the functions used to computationally implement a given MSD formalism. This atomized representation leads directly to the exportation of these functions in a way that benefits from the referred computational cost savings. To get a less abstract idea, Fig.~\ref{fig_atomization} shows the exported C code for the mass matrix of a simple four-bar linkage mechanism.
\begin{figure}
\hrule
\begin{multicols}{2}
{\tiny
\begin{verbatim}
atom27 = sin(theta2);
atom0 = cos(theta3);
atom1 = sin(theta3);
atom26 = cos(theta2);
atom49 = atom1*atom26+atom27*atom0;
atom46 = -atom27*atom1+atom26*atom0;
atom237 = m3*l2*( cg3x*atom0+atom1*cg3z);
atom253 = m3*l1*( cg3x*atom46+atom49*cg3z)+atom237;
atom200 = -l1*atom26;
atom214 = -l2*atom200;
atom197 = (l1*l1);
atom215 = (l2*l2);
atom270 = atom27*l1*cg2z*m2+m3*( atom214+atom215)
-m2*cg2x*atom200+I3yy+atom237+I2yy+atom253;
atom229 = m3*l2*cg3x*atom0+m3*l2*atom1*cg3z;
atom271 = m3*l1*cg3x*atom46+I3yy+m3*atom49*l1*cg3z
+atom229;
atom273 = I3yy+atom229;
_M[0] = -m3*( atom197+2.0*atom214+atom215)-atom197*m2
-I3yy-I2yy+-2.0*atom253-I1yy
+-2.0*( atom27*cg2z+atom26*cg2x)*l1*m2;
_M[1] = -atom270;
_M[2] = -atom271;
_M[3] = -atom270;
_M[4] = -m3*atom215-I3yy+-2.0*atom237-I2yy;
_M[5] = -atom273;
_M[6] = -atom271;
_M[7] = -atom273;
_M[8] = -I3yy;
\end{verbatim}
}
\end{multicols}
\hrule
\caption{Exported C code for atomized mass matrix.}
\label{fig_atomization}
\end{figure}
The atomization process should ideally be done ``on-the-way'', meaning that every time a new algebraic operation is performed a new atom is created or replaced by an existing matching atom.
Thus, the symbolic method takes advantage of the memory savings and the associated complexity reduction as soon as possible in the problem setup.
This means that the symbolic algebra system works internally with atomized expressions. A feature that is not obvious for the standard user but that is widespread in computer algebra systems. See Fig.~\ref{fig:on_the_way} for an elemental example.
\begin{figure}
\hrule
The addition of vectors $\vect{u}$ and $\vect{v}$ given their components represented in bases $\B_1$ and $\B_2$
\begin{equation*}
\begin{aligned}
\begin{Bmatrix} \vect{u} \end{Bmatrix}_{\B_1}= \begin{Bmatrix} u_x\\ u_y \\ u_z \end{Bmatrix}_{\B_1}
~~\text{and }~~
\begin{Bmatrix} \vect{v} \end{Bmatrix}_{\B_2}= \begin{Bmatrix} v_x\\ v_y \\ v_z \end{Bmatrix}_{\B_2}
\end{aligned}
\end{equation*}
is sought. Let the rotation matrix be
\begin{equation*}
\begin{aligned}
\mR{\B_1}{\B_2} =
\begin{bmatrix}
1 & 0&0\\
0 &\cos(\theta) & -\sin(\theta)\\
0 & \sin(\theta) & \cos(\theta)
\end{bmatrix}
\end{aligned}.
\end{equation*}
The addition of the two vectors represented in $\B_1$ base is performed as follows:
\begin{equation*}
\begin{aligned}
\begin{Bmatrix}\vect{u}+ \vect{u}\end{Bmatrix}_{\B_1} &=\begin{Bmatrix} \vect{u} \end{Bmatrix}_{\B_1} +\mR{\B_1}{\B_2} \begin{Bmatrix} \vect{v} \end{Bmatrix}_{\B_2} =
\begin{Bmatrix} u_x\\ u_y \\ u_z \end{Bmatrix}_{\B_1} +
\begin{bmatrix} 1 & 0&0\\0 &\cos(\theta) & -\sin(\theta)\\ 0 & \sin(\theta) & \cos(\theta)\end{bmatrix}
\begin{Bmatrix} v_x\\ v_y \\ v_z \end{Bmatrix}_{\B_2}
=\\
&\begin{Bmatrix} u_x\\ u_y \\ u_z \end{Bmatrix}_{\B_1} +
\begin{bmatrix} 1 & 0&0\\0 &\alpha_1 & -\alpha_2 \\ 0 & \alpha_2 & \alpha_1 \end{bmatrix}
\begin{Bmatrix} v_x\\ v_y \\ v_z \end{Bmatrix}_{\B_2}
=
\begin{Bmatrix} u_x\\ u_y \\ u_z \end{Bmatrix}_{\B_1} +
\begin{Bmatrix} v_x\\ \alpha_1 v_y -\alpha_2 v_z \\ \alpha_2 v_y +\alpha_1 v_z \end{Bmatrix}_{\B_1}
=\\
&\begin{Bmatrix} u_x\\ u_y \\ u_z \end{Bmatrix}_{\B_1} +
\begin{Bmatrix} v_x\\ \alpha_3 - \alpha_4 \\ \alpha_5 + \alpha_6 \end{Bmatrix}_{\B_1}
=
\begin{Bmatrix} u_x\\ u_y \\ u_z \end{Bmatrix}_{\B_1} +
\begin{Bmatrix} v_x\\ \alpha_7 \\ \alpha_8 \end{Bmatrix}_{\B_1}
=
\begin{Bmatrix} \alpha_9\\ \alpha_{10} \\\alpha_{11} \end{Bmatrix}_{\B_1}
\end{aligned}
\end{equation*}
where $\alpha_i$ are the atoms:
\begin{equation*}
\begin{aligned}
\alpha_1 &= \cos(\theta) & \alpha_2 &= \sin(\theta) & \alpha_3 &=\alpha_1 v_y & \alpha_4 &= \alpha_2 v_z & \alpha_5 &=\alpha_2 v_y &\alpha_6 =\alpha_1 v_z \\
\alpha_7 &=\alpha_3 - \alpha_4 & \alpha_8 &= \alpha_5 - \alpha_6 & \alpha_9 &= u_x+v_x & \alpha_{10} &=u_y + \alpha_7 & \alpha_{11} &= u_z + \alpha_8\\
\end{aligned}
\end{equation*}
\hrule
\caption{``On-the-way'' atomization example.}
\label{fig:on_the_way}
\end{figure}
In the same line, it is important to remember that the fundamental symbolic differentiation and substitution operations should be implemented to work directly on atomized expressions. This maximizes atom recycling and limits enormously the time and memory requirements of the algorithms.
In this context, to take advantage of the atomization, care should be taken when choosing the way and order in which the required operations
are performed.
The operation number should be minimized and atom recycling maximized.
A general purpose algorithm
aiming at finding an absolute minimum number of operations would require an exhaustive search that is beyond the reach of reasonable computational resources.
Therefore, it is required to define appropriate heuristics. Recursive dynamics formulations are usually taken as the starting point to define such heuristics. In our work, these heuristics are partly implemented by the way of mechanics operators, as will be explained in the next section.
\subsection{Recursive kinematic operators}
Recursive formulations represent the state-of-the-art on symbolic MSD \cite{samin2003}. These formulations use relative coordinates to parametrize the system leading to a tree-shaped body structure\footnote{Closed loops are opened to parametrize and closed through constraint equations.}.
This allows the recursive determination positions, velocities and accelerations of points as well as orientations, angular velocities and accelerations of bodies, by the way of the well known ``motion composition laws''. When different elements -points and orientations- share a common path towards the tree root, this implies the sharing of common sub-expressions. If applied symbolically, this recursive computation produces nearly good optimal ``on-the-way'' atomizations. This sharing of expressions is the main feature on which the kinematic forward recursion step, found in recursive formulations, is based.
For example, for a serial multibody system the angular velocity of body $S_{i+1}$ with respect to $S_{i-1}$ could be expressed as follows:
\begin{eqnarray}
\boldsymbol{\omega}^{S_{i+1}}_{S_{i-1}} = \boldsymbol{\omega}^{S_{i}}_{S_{i-1}} + \boldsymbol{\omega}^{S_{i+1}}_{S_{i}}
\end{eqnarray}
In the same way the angular velocity of body $S_{i+2}$ with respect to $S_{i-1}$ is expressed as:
\begin{eqnarray}
\boldsymbol{\omega}^{S_{i+2}}_{S_{i-1}} =
\boldsymbol{\omega}^{S_{i}}_{S_{i-1}} + \boldsymbol{\omega}^{S_{i+1}}_{S_{i}} + \boldsymbol{\omega}^{S_{i+2}}_{S_{i+1}}=
\boldsymbol{\omega}^{S_{i+1}}_{S_{i-1}} + \boldsymbol{\omega}^{S_{i+2}}_{S_{i+1}}
\end{eqnarray}
So, when computing magnitudes related to a given element it can be appreciated how computations related to elements down in the same chain can be reused. Other kinematic entities like position vectors, base-change matrices, linear velocities, accelerations and angular accelerations can be dealt with analogously.
In correspondence with the substitution of the bodies structure by the bases and points structures proposed in our work, the recursivity at the level of bodies is now dealt with at the bases and points structure levels.
This allows not only to use arbitrary parametrizations, as commented before, but also a better use of the any degree of recursivity that may be implicit when using arbitrary parametrizations.
To that end, kinematic operators that take advantage of any recursivity present in the parameterization are defined: Position vector between two points, velocity of a point with respect to a given frame (point plus orientation), angular velocity, base-change matrix, and so on. Basically the typical \textit{recursivity} found in recursive algorithms is translated to the operator algebra.
To support this an algebra of 3D vectors and tensors is defined. This algebra relieves the user of dealing with base-changes that are internally dealt with.
The full system works using ``on-the-way'' atomization and the operators are implemented taking advantage of the aforementioned recursivity. In this way, the number of operations is minimized and the reuse of atoms is maximized. As a consequence an optimal implementation of the given formalism for any parameterization chosen by the user is obtained.
The backward recursion of $O(n^3)$ algorithms can be considered a particular implementation of the principle of the virtual power.
The inertia forces and moments of the bodies affected by a given virtual movement appear added together in the contribution of this virtual movement to the system dynamic equations. Recursive formulations take advantage of this grouping so that they minimize the required operation count.
Taking advantage of this when applying symbolically the principle of virtual power produces atomizations as efficient as state-of-the-art $O(n^3)$ formulations. This is the approach followed by the symbolic implementation of the virtual power principle used in this work.
As an illustration of the achievements of our symbolic methods, we can obtain nearly optimal atomized equations for standard multibody systems using for example the principle of the virtual power and relative coordinates. Some authors \cite{samin2003} claim to be unable to do the same unless a direct symbolic implementation of a recursive formulation is used\footnote{In comparison, our method presents the overhead of having to deal with the common atom search. Even if using hash tables to do the search our symbolic processing phase seems to take longer.}.
\subsection{Other symbolic methods}
There are other symbolic methods that can be applied to reduce even further the complexity of the resulting model: \textit{``trigonometricaly simplifiable expression removal''} \cite{PlazaThesis2015}, \textit{``base parameter formulation of the system inertias''}\cite{khalil1987,Ros2012,Ros2015}, \textit{``base parameter elimination''}\cite{Iriarte2015}, etc... This methods can be applied directly on top of the presented modeling techniques. However, they are not considered in this work.
\section{Multibody model description}
\label{sec2}
The \textit{FEVE 3000} \cite{feve3000} locomotive multibody model developed in this work is depicted in Fig.~\ref{fig:locomotive} with an expanded view of the main parts shown in Fig.~\ref{fig:expanded_bodies}.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{./locomotive.png}
\caption{Multibody model}
\label{fig:locomotive}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=\textwidth]{./expanded_bodies.png}
\caption{Expanded View}
\label{fig:expanded_bodies}
\end{figure}
The $Vehicle\ Body$ (dark grey) is attached to the front and rear bogies thought two $Slider$ (green) bodies. The front bogie consists of a $Slider$ that rests on a couple of
$Suspender$ (orange) bodies hanging from the $Bogie\ Frame$ (dark blue), and two $Wheelset$ bodies (light blue) each of them with two $Axle\ Box$ (red) bodies. There is a couple of anti-yaw links (grey) between each $Suspender$ and the bogie frame. The rear bogie is identical to the front one but it includes two motors (one per each $Wheelset$). The motor $Housing$ (mauve) rotates around the relative $Wheelset$ and is attached to the $Bogie\ Frame$ using a bushing. The motor includes a $Rotor$ (pink). The transmission of motion from the $Rotor$ to the $Wheelset$ is done by the way of a gear pair.
The $Slider$ is connected by four identical spring-dampers to the Suspender part of the $Bogie\ Frame$. In a similar way, each $Axle\ Box$ is connected to the $Bogie\ Frame$ by two identical spring-dampers. The $Housing$ is also attached to the $Bogie\ Frame$ using a bushing. Compliance is considered in the gearing contacts. Linear stiffness and damping is assumed for spring-dampers, bushings and gear compliance.
Braking on the wheels and traction on the rotors is modeled considering externally applied torques.
The wheel-rail interaction model considers a fully three-dimensional rolling contact considering a single contact point per wheel. Normal contact is enforced through the use of constraints, while the tangential forces are determined based on the standard Kalker linear constitutive model.
Note that we consider generic wheel and rail profiles. The rails can present general irregularities along the track.
\subsection{Parametrization}
\subsubsection*{Multibody}
The $Vehicle\ Body$ is positioned relative to the track using absolute coordinates (3 translations followed by 3 Euler rotations).
Each $Slider$ is attached to the $Vehicle$ $Body$ by a revolute joint. A rotation relative to the $Vehicle\ Body$, in the vertical direction, is used to position the $Slider$.
To simplify the modeling, the effect of the anti-yaw bar is accounted for by removing the relative yaw motion between the $Bogie\ Frame$ and the $Slider$. With the same purpose, the $Suspender$ is considered fixed to the $Bogie\ Frame$. We use a vertical translation followed by two successive horizontal rotations (roll and tilt) to position the $Bogie\ Frame$ relative to the $Slider$. Each $Axle\ Box$ is positioned fixed to the ``non-spinning wheelset'' frame (NSWHS), a frame that follows the relative $Wheelset$ but that does not spin with it. Each $Wheelset$ is positioned relative to the $Bogie\ Frame$ using a vertical translation and two horizontal rotations (roll and spin). Other relative degrees of freedom between these bodies are removed by the particular configuration of the spring-dampers.
A rotation around the $Wheelset$ axis, relative to the NSWHS frame, is used to place each motor $Housing$. A rotation in the same direction, also relative to the NSWHS frame, is introduced to give the angular position of the $Rotor$ of each motor.
A total number of $60$ generalized coordinates, ${\mathbf{q}}$, is used in this parametrization.
\subsubsection*{Contact}
The rails and wheel surfaces are described using cubic splines \cite{pombo2003general},
\begin{equation}\begin{aligned}
f^*(u^*) = (((a^* (u^*-u^*_{bp}) + b^* ) (u^*-u^*_{bp}) + c^*) (u^*-u^*_{bp}) +d^*),
\label{eq:spline}
\end{aligned} \end{equation}
defined based on a set of control points that approximate their geometry.
\begin{figure}
\centering
\includegraphics[width=\textwidth]{./wheel_rail_profiles.pdf}
\caption{Surfaces and railway parametrization}
\label{fig:surfaces}
\end{figure}
Figure \ref{fig:surfaces} schematically shows the parametrization for a wheel-rail pair. The wheel and rail profiles are given respectively by $f^w(u^w)$ and $f^r(u^r)$, while the shape of the center line of the base of the rail along the track is given by $f_x^t(s^r)$, $f_y^t(s^r)$ and $f_z^t(s^r)$. Wheel surfaces are supposed to have cylindrical symmetry. Not represented in the figure is the spline used to represent the camber of the rail $f_\theta^t(s^r)$.
For the contact point at each wheel-rail pair the parameters $\theta^w, u^w$ and $s^r, u^r$ play the role of ``generalized coordinates'' used to position the contact points, $P^w$ and $P^r$, respectively onto the wheel and rail surfaces. When the contact is materialized, the point $P^w$
and the point $P^r$ -defined as an arbitrary points in the surfaces of the wheel and rail respectively- are coincident.
The position of $P^w$ is given relative to a reference point in the wheel axis, $O^w$, as
\begin{equation}\begin{aligned}
\mathbf{r}_{O^w}^{P^w}(u^w,\theta^w) = f^w(u^w) \cos(\theta^w) \mathbf{e}^w_x + u^w \mathbf{e}^w_y - f^w(u^w)\sin(\theta^w) \mathbf{e}^w_z.
\label{eq:OwPw}
\end{aligned} \end{equation}
For numerical reasons, the position of this point is given relative to a NSWHS base, $\mathbf{e}^w_x,\mathbf{e}^w_y ,\mathbf{e}^w_z$.
Analogously, the position of $P^r$ is given relative to a reference point in ground or track reference, $O^r$
\begin{equation}\begin{aligned}
\mathbf{r}_{O^r}^{P^r}(u^r,s^r) = f_x^t(s^r)\mathbf{e}^r_x + f_y^t(s^r) \mathbf{e}^r_y + f_z^t(s^r) \mathbf{e}^r_z
+ u^r \hat{\mathbf{c}} + f^r(u^r) \hat{\mathbf{n}}
\end{aligned} \end{equation}
The base $\mathbf{e}^r_x,\mathbf{e}^r_y ,\mathbf{e}^r_z$ is fixed at the ground. Defining $\hat{\mathbf{t}}$ as a unit vector tangent to the center line of the rail base, $\hat{\mathbf{c}}$ is defined as a unit vector perpendicular to $\hat{\mathbf{t}}$ with an angle $f_\theta^t(s^r)$ with the ground measured in the positive direction of $\hat{\mathbf{t}}$. $\hat{\mathbf{n}}$ completes the base so that it is dexterous.
The whole set of $8\times 4=32$ generalized coordinates required to position the $i=1,...,8$ contact points used in the analyzed example is referred as $\mathbf{s}=[...,\mathbf{s}_i^\mathsf{T},...]^\mathsf{T}$. Where $\mathbf{s}_i=[\theta^w_i, u^w_i,s^r_i, u^r_i]^\mathsf{T}$ is set of coordinates required to position points $P^w_i$ and $P^r_i$ at the $i$-th contact point.
From the symbolic modeling point of view, functions $f^*(u^*)$ are the are modeled as if a single $3^{rd}$ order polynomial completely represents the whole profiles. At ``show-time'', the coefficients $a^*$, $b^*$, $c^*$, $d^*$ and break-points $u^*_{bp}$ are updated depending on the position of the contact point.
As commented before, for the purposes of this paper we only consider a single point of contact. Note, that
flange contact rarely occurs when the train runs along straight tracks or huge radii curved tracks unless train velocity is close to its critical speed \cite{andersson1999rail}. Nevertheless the parametrization proposed here is compatible
some multiple-point-of-contact approaches \cite{andersson1999rail}.
\subsection{Constraint equations}\label{subsec:ConstrEqus}
The only constrains that are present in the analyzed problem are those related to the contact points between wheel an rail.
At a given contact, it should be enforced that the points $P^w$ and $P^r$ are coincident and that the surfaces at these points are tangent.
Defining the tangent and normal vectors to the wheel at point $P^w$ as $\mathbf{t}_x^w$, $\mathbf{t}_y^w$, $\mathbf{n}^w$, and
the tangent and normal vectors to the rail at point $P^r$ as $\mathbf{t}_x^r$, $\mathbf{t}_y^r$, $\mathbf{n}^r$. These conditions can be written
\cite{Shabana2005} as:
%
\begin{equation}
{\boldsymbol{\phi}^n} ({\mathbf{q}},{\mathbf{s}})= \mathbf{n}^r \cdot \mathbf{r}^{P^w}_{P^r} = \mathbf{0}
\end{equation}
and
\begin{equation}
{\boldsymbol{\phi}^d} ({\mathbf{q}},{\mathbf{s}})=
\left[
\begin{array}{c}
\mathbf{t}_x^r\cdot\mathbf{r}^{P^w}_{P^r}\\
\mathbf{t}_y^r\cdot\mathbf{r}^{P^w}_{P^r}\\
\mathbf{t}_x^w\cdot\mathbf{n}^r\\
\mathbf{t}_y^w\cdot\mathbf{n}^r
\end{array}
\right]
=
\mathbf{0},
\end{equation}
where ${\boldsymbol{\phi}^n}$ is the so called normal constraint, and ${\boldsymbol{\phi}^d}$ are the so called tangent constraints.
For each contact point, the tangent and normal vectors can be defined as:
\begin{eqnarray}
\mathbf{t}^r_x&=\dfrac{{\partial\mathbf{r}_{O^r}^{P^r}}/{\partial s^r}}{\left|{\partial\mathbf{r}_{O^r}^{P^r}}/{\partial s^r}\right|}
, ~ ~
\mathbf{t}^r_y=\dfrac{{\partial\mathbf{r}_{O^r}^{P^r}}/{\partial u^r}}{\left|{\partial\mathbf{r}_{O^r}^{P^r}}/{\partial u^r}\right|}
\text{, and } ~ ~
\mathbf{n}^r= \mathbf{t}^r_x \times \mathbf{t}^r_y\\
\mathbf{t}^w_x&=\dfrac{{\partial\mathbf{r}_{O^w}^{P^wr}}/{\partial \theta^w}}{\left|{\partial\mathbf{r}_{O^w}^{P^w}}/{\partial \theta^w}\right|}
, ~ ~
\mathbf{t}^w_y=\dfrac{{\partial\mathbf{r}_{O^w}^{P^w}}/{\partial u^w}}{\left|{\partial\mathbf{r}_{O^w}^{P^w}}/{\partial u^w}\right|}
\text{, and } ~ ~
\mathbf{n}^w= \mathbf{t}^w_x \times \mathbf{t}^w_y.
\end{eqnarray}
These vectors and constraint equations can easily be defined using the symbolic procedures previously discussed. It will be seen that
this symbolic implementation will be very efficient as well.
Now, we use the subindex $i=1,...,8$ to refer to the constraint equations relative to each of the $8$ contact points. The set of all the normal constraints is referred as ${\boldsymbol{\phi}^n} ({\mathbf{q}},{\mathbf{s}})=[...,{\boldsymbol{\phi}^n}_i ({\mathbf{q}},{\mathbf{s}}),...]^\mathsf{T}$. Analogously,
the set of all the tangent constraints is referred as ${\boldsymbol{\phi}^d} ({\mathbf{q}},{\mathbf{s}})=[...,{\boldsymbol{\phi}^d}_i ({\mathbf{q}},{\mathbf{s}}),...]^\mathsf{T}$.
\subsection{Dynamic equations}
As commented previously the dynamic equations are obtained based on the direct application of the principle of virtual power. Using the vector $[{\mathbf{q}},{\mathbf{s}}]$ as the set of generalized coordinates, the mass matrix $\mathbf{M}$ is obtained by differentiation of the equations motion with respect to the the generalized accelerations and the generalized force $\boldsymbol{\delta}$ vector is obtained by substitution of the generalized accelerations by zero in the equations of motion. These equations should be complemented by the second derivative of the constraint equations to have a determined set equations
\begin{eqnarray}
\ddot{{\boldsymbol{\phi}^n}}({\mathbf{q}},{\mathbf{s}}) &= \mathbf{0} \label{eq:phin}\\
\ddot{{\boldsymbol{\phi}^d}}({\mathbf{q}},{\mathbf{s}}) &= \mathbf{0} \label{eq:phit},
\end{eqnarray}
This set of equations shows the following structure:
%
\begin{align}
\begin{bmatrix}
{\mathbf{M}}_{{\mathbf{q}}\vq}({\mathbf{q}},{\mathbf{\dot{q}}}) & \mathbf{0} & \dot{\boldsymbol{\phi}}_{\vdq}^{n}{\T}({\mathbf{q}},{\mathbf{s}}) & \dot{\boldsymbol{\phi}}_{\vdq}^{d}{\T}({\mathbf{q}},{\mathbf{s}}) \\
\mathbf{0} &\mathbf{0} & \mathbf{0} & \dot{\boldsymbol{\phi}}_{\vds}^{d}{\T}({\mathbf{q}},{\mathbf{s}}) \\
\dot{\boldsymbol{\phi}}_{\vdq}^{n}({\mathbf{q}},{\mathbf{s}}) & \mathbf{0} & \mathbf{0} &\mathbf{0} \\
\dot{\boldsymbol{\phi}}_{\vdq}^{d}({\mathbf{q}},{\mathbf{s}}) & \dot{\boldsymbol{\phi}}_{\vds}^{d}({\mathbf{q}},{\mathbf{s}}) & \mathbf{0} &\mathbf{0}
\end{bmatrix}
\begin{bmatrix} {\mathbf{\ddot{q}}} \\ {\mathbf{\ddot{s}}} \\ {\boldsymbol{\lambda}}^n \\ {\boldsymbol{\lambda}}^d \end{bmatrix}
=
\begin{bmatrix} {\boldsymbol{\delta}}_{{\mathbf{q}}}({\mathbf{q}},{\mathbf{\dot{q}}}) \\ \mathbf{0} \\ {\boldsymbol{\gamma}^n}({\mathbf{q}},{\mathbf{s}},{\mathbf{\dot{q}}},{\mathbf{\dot{s}}}) \\ {\boldsymbol{\gamma}}^d({\mathbf{q}},{\mathbf{s}},{\mathbf{\dot{q}}},{\mathbf{\dot{s}}}) \end{bmatrix},
\end{align}
where ${\mathbf{M}}_{{\mathbf{q}}\vq}$ and ${\boldsymbol{\delta}}_{{\mathbf{q}}}$ are the blocks of the mass matrix $\mathbf{M}$ and vector ${\boldsymbol{\delta}}$ related to the set of coordinates ${\mathbf{q}}$.
In reference \cite{Shabana2005}, the authors refer this formulations as the Augmented Contact Constraint Formulation (ACCF).
The particular structure of the dynamic equations for the problem analyzed can appreciated on Fig.~\ref{fig1}. There, the nonzero entries for matrix $\left[[ {\mathbf{M}},\dot{\boldsymbol{\phi}}_{\vdq}{\T} ;\dot{\boldsymbol{\phi}}_{\vdq},\mathbf{0}],[{\boldsymbol{\delta}};{\boldsymbol{\gamma}} ]\right]$ are shown as dots. It should be noted that, $\dot{\boldsymbol{\phi}}_{\vds}^{n}$ is zero numerically, even if symbolically there can be seen few nonzero expressions.
\begin{figure}[h]
\centering
\includegraphics[height=80mm]{./M_Phi_dqT_Phi_dq_delta_gamma.pdf}
\caption{Dynamic model structure $\left[[ {\mathbf{M}},\dot{\boldsymbol{\phi}}_{\vdq}{\T} ;\dot{\boldsymbol{\phi}}_{\vdq},\mathbf{0}],[{\boldsymbol{\delta}};{\boldsymbol{\gamma}} ]\right]$}
\label{fig1}
\end{figure}
As accelerations ${\mathbf{\ddot{s}}}$ are not needed, the previous system of equations can be reduced to
\begin{equation}
\begin{bmatrix}
{\mathbf{M}}_{{\mathbf{q}}\vq}({\mathbf{q}},{\mathbf{\dot{q}}}) & \dot{\boldsymbol{\phi}}_{\vdq}^{n}{\T}({\mathbf{q}},{\mathbf{s}}) \\
\dot{\boldsymbol{\phi}}_{\vdq}^{n}({\mathbf{q}},{\mathbf{s}}) & \mathbf{0}
\end{bmatrix}
\begin{bmatrix} {\mathbf{\ddot{q}}} \\ {\boldsymbol{\lambda}}^n \end{bmatrix} \label{eq:MphiTPhiZero4}
=
\begin{bmatrix} {\boldsymbol{\delta}}_{{\mathbf{q}}}({\mathbf{q}},{\mathbf{\dot{q}}}) \\ {\boldsymbol{\gamma}^n}({\mathbf{q}},{\mathbf{s}},{\mathbf{\dot{q}}},{\mathbf{\dot{s}}}) \end{bmatrix},
\end{equation}
The reduced structure of the dynamic equations can appreciated on Fig.~\ref{fig2}. Obviously, standard linear solution procedures are going to perform much more efficiently in this case.
There, the nonzero entries for matrix $\left[[ {\mathbf{M}}_{{\mathbf{q}}\vq},\dot{\boldsymbol{\phi}}_{\vdq}^{n}{\T} ;\dot{\boldsymbol{\phi}}_{\vdq}^{n},\mathbf{0}],[{\boldsymbol{\delta}}_{{\mathbf{q}}};{\boldsymbol{\gamma}^n}]\right]$ are shown as dots.
\begin{figure}[h]
\centering
\includegraphics[height=32mm]{./M_Phi_dqT_Phi_dq_delta_gamma-reduced.pdf}
\caption{Dynamic model structure $\left[[ {\mathbf{M}}_{{\mathbf{q}}\vq},\dot{\boldsymbol{\phi}}_{\vdq}^{n}{\T} ;\dot{\boldsymbol{\phi}}_{\vdq}^{n},\mathbf{0}],[{\boldsymbol{\delta}}_{{\mathbf{q}}};{\boldsymbol{\gamma}^n}]\right]$}
\label{fig2}
\end{figure}
In reference \cite{Shabana2005}, the authors refer this formulation as the Embedded Contact Constraint Formulation (ECCF) \cite{Shabana2005}.
Constraint stabilization is performed based on projection on the coordinate and velocity manifolds. That requires the solution of the following equations
\begin{eqnarray}
{\boldsymbol{\phi}^n}({\mathbf{q}},{\mathbf{s}}) &= \mathbf{0} \label{eq:phin}\\
{\boldsymbol{\phi}^d}({\mathbf{q}},{\mathbf{s}}) &= \mathbf{0} \label{eq:phit},
\end{eqnarray}
at the coordinate level, and of
\begin{eqnarray}
\dot{\boldsymbol{\phi}}_{\vdq}^{n}({\mathbf{q}},{\mathbf{s}}) {\mathbf{\dot{q}}} &= {\boldsymbol{\beta}}^n({\mathbf{q}},{\mathbf{s}}) = \mathbf{0} \label{eq:dqproj} \\
\dot{\boldsymbol{\phi}}_{\vdq}^{d}({\mathbf{q}},{\mathbf{s}}) {\mathbf{\dot{q}}} + \dot{\boldsymbol{\phi}}_{\vds}^{d}({\mathbf{q}},{\mathbf{s}}) {\mathbf{\dot{s}}} &= {\boldsymbol{\beta}}^d({\mathbf{q}},{\mathbf{s}}) = \mathbf{0} \label{eq:dsproj}
\end{eqnarray}
at the velocity level. Note that there are no rheonomous equations in the problem analyzed, an therefore ${\boldsymbol{\beta}}^n({\mathbf{q}},{\mathbf{s}}) $ and ${\boldsymbol{\beta}}^d({\mathbf{q}},{\mathbf{s}})$ are $\mathbf{0}$. For example, we frequently use coordinate partitioning \cite{Wehage1982,Haug1989} into dependent and independent. Dependent coordinates and velocities are obtained in terms of the independent ones.
It should be remarked that in the ECCF context, ${\mathbf{s}}$ and ${\mathbf{\dot{s}}}$ are better not considered standard generalized coordinates, but a set of auxiliary variables that must be known in order to solve equation system (\ref{eq:MphiTPhiZero4}). After the integration, Eq.~(\ref{eq:phit}) and Eq.~(\ref{eq:dsproj}) can be used to
obtain the auxiliar variables ${\mathbf{s}}$ and ${\mathbf{\dot{s}}}$ in terms of ${\mathbf{q}}$ and ${\mathbf{\dot{q}}}$. In this context the position and velocity projection would be performed after this step, based
on Eq.~(\ref{eq:phin}) and Eq.~(\ref{eq:dqproj}), respectively. These inertia-less coordinates find different names in the literature, such as ``surface parameters'' \cite{shabana2004development,Shabana2005,pombo2007new,auciello2009dynamic,arnold2004simulation}, ``non-generalized coordinates'' \cite{shabana2001augmented} or ``auxiliary variables'' \cite{fisette2000multibody}, but their main feature is that they do not participate in the system dynamics.
\subsection{Contact model.}
In the context of railway dynamic simulation is very important to correctly determine the values of the creep forces between the wheel and the rail. To that end, in this work the well known linear contact theory of Kalker \cite{kalker1967rolling} is used. This theory requires the determination of several data: location of the contact point, the creepages, normal loads at this contact point, wheel and rail surface curvatures at the contact point, tangent and normal vectors at the contact patch. This computations are
big and must be done for every contact. To speed up computations, several authors \cite{bozzone2011lookup,sugiyama2011wheel,escalona2015modeling} propose the use of pre-calculated look-up tables to determine the required data. This procedure is tedious and usually requires to introduce some modeling simplifications. We propose to compute these quantities without simplifications, on line, based on functions exported using the proposed symbolic methods. This is a more simple and general procedure to apply. The results will confirm that this is very fast procedure.
\subsubsection*{Contact patch geometry determination.}
Based on classical Hertzian contact theory, the contact patch is a flat ellipse \cite{Iwnicki2006}. The semi-axes of this ellipse in the longitudinal and transversal directions, $a$ and $b$ respectively, are determined as follows:
\begin{equation}
a=\left(\frac{3}{2} \frac{1-\nu^2}{E}\frac{1}{A+B} N \right)^{\frac{1}{3}}m(\theta)~~~\text{and}~~~
b=\left(\frac{3}{2} \frac{1-\nu^2}{E}\frac{1}{A+B} N \right)^{\frac{1}{3}}n(\theta).
\end{equation}
In these expressions, $N$ is the normal contact force acting on the wheel, $E$ is the Youngβs modulus, $\nu$ the Poissonβs ratio. $m(\theta)$ and $n(\theta)$ are adimensional functions proposed by Hertz. We use on-line interpolation in table $4.1$ in Ref.~\cite{Iwnicki2006} to evaluate these functions. $\theta=\cos^{-1}{\left(\frac{\left|A-B\right|}{A+B}\right)}$, where $A$ and $B$ are determined as
\begin{equation}
A=\frac{1}{2}\left(\frac{1}{R^{w}_{x}}+\frac{1}{R^{r}_{x}}\right)~~~\text{and}~~~B=\frac{1}{2}\left(\frac{1}{R^{w}_{y}}+\frac{1}{R^{r}_{y}}\right).
\end{equation}
$R^{w}_{x}$, $R^{w}_{y}$, $R^{r}_{x}$, $R^{r}_{y}$ are the curvature radii of wheel and rail surfaces at the contact point. For the case studied these are computed as:
\begin{equation}
\begin{array}{cc}
R^{w}_{x}= \dfrac{\sqrt{(\mathbf{r}_{O^w}^{P^w}\mathbf{e}_{x}^{w})^2+(\mathbf{r}_{O^w}^{P^w}\mathbf{e}_{y}^{w})^2}}{\sqrt{1-\left(\mathbf{n}^r \mathbf{e}^w_y\right)^2}}~~~~~\footnotemark
&
~~~~~R^{w}_{y}=\left|\dfrac{(1+\frac{\partial f^w}{\partial{u_w}})^{3/2}}{\frac{\partial^2 f^w}{{\partial u_w}^2}}\right|
\\
R^{r}_{y}=\left|\dfrac{(1+\frac{\partial f^r}{\partial{u_r}})^{3/2}}{\frac{\partial^2 f^r}{{\partial u_r}^2}}\right|
&
R^{r}_{x}=\infty
\end{array}
\end{equation}
\footnotetext{Distance from the rotation axis to the contact point along the normal at contact point.}
These curvature radii are obtained and exported based on the symbolic methods presented preciously in this paper. The normal force $N$ is the Lagrange multiplier associated with the normal constraint of the relative contact point. It is obtained directly from the solution of the dynamic system of
equations. To keep the dynamic problem linear, avoiding a nonlinear iteration, the normal force used is the one obtained in the previous integration step.
\subsubsection*{Creep forces and moments}
From the modeling perspective creep forces and moments are considered external actions, so their effect is contained in vector $\boldsymbol{\delta}_{{\mathbf{q}}}$, along with all the contributions due to other inertial constitutive and external forces.
For each wheel the these forces and moments are symbolically expressed as
\begin{equation}
\mathbf{f}=f_x~\mathbf{t}^r_x+f_y~\mathbf{t}^r_y ~~~~\text{and}~~~~ \mathbf{m}=m_z~\mathbf{n}^r,
\end{equation}
where $f_x$, $f_y$ and $m_z$ are symbols. These represents the tangent contact force and spin contact moment acting on the wheel at the contact point $P^w$. They are defined at the contact base $\mathbf{t}^r_x$,$\mathbf{t}^r_y$, $\mathbf{n}^r$.
The components of these vectors are numerically computed as
\begin{equation}
\left[
\begin{array}{c}
f_x\\
f_y\\
m_z
\end{array}
\right]=-G
\left[
\begin{array}{ccc}
a b ~c_{11}&0&0\\
0& a b ~c_{22}& \sqrt{ab}~c_{23}\\
0&-\sqrt{ab}~c_{23}& (ab)^2~c_{33}
\end{array}
\right]
\left[
\begin{array}{c}
\xi_x\\
\xi_y\\
\varphi_z
\end{array}
\right]
\label{eq:kalker1}
\end{equation}
\cite{Iwnicki2006}. Parameter $G$ is the material shear modulus and $c_{ij}(a/b,\nu)$ are the coefficients determined by Kalker, tabulated in Ref.~\cite{kalker1968tangential}. We use on-line interpolation in these tables. The creepages $\xi_x$, $\xi_y$ and $\varphi$ are defined as:
\begin{eqnarray}
\xi_x = \frac{\mathbf{v}_{Gr.}^{P^w}}{\frac{1}{2}(\left|\mathbf{v}_{Gr.}^{O^w}\right|+\left|\boldsymbol{\omega}_{Gr.}^{w} \wedge \mathbf{r}_{O^w}^{P^w}\right|)}~\mathbf{t}^r_x \\
\xi_y = \frac{\mathbf{v}_{Gr.}^{P^w}}{\frac{1}{2}(\left|\mathbf{v}_{Gr.}^{O^w}\right|+\left|\boldsymbol{\omega}_{Gr.}^{w} \wedge \mathbf{r}_{O^w}^{P^w}\right|)}~\mathbf{t}^r_y \\
\varphi_z = \frac{\boldsymbol{\omega}_{Gr.}^{w}}{\frac{1}{2}(\left|\mathbf{v}_{Gr.}^{O^w}\right|+\left|\boldsymbol{\omega}_{Gr.}^{w} \wedge \mathbf{r}_{O^w}^{P^w}\right|)}~\mathbf{n}^r
\end{eqnarray}
where $\mathbf{v}_{Gr.}^{P^w}$ in the velocity with respect to the ground ($Gr.$) of the $P^w$ contact point when moves ``attached'' to the wheel and $\boldsymbol{\omega}_{Gr.}^{w}$ is the wheel-set angular velocity with respect to the ground. $\mathbf{v}_{Gr.}^{O^w}$ in the velocity with respect to the ground of the center of the wheel-set and $\mathbf{r}_{O^w}^{P^w}$ is the position vector form $O^w$ to $P^w$ as shown in Eq.~(\ref{eq:OwPw}). The creepages are determined numerically based on exported functions for the numerators and denominators of these expressions. In this way division by zero can be dealt within the numerical solver.
All the symbolic functions required for the implementation of the Kalker model for all the different contact points, referred previously, are computed in a single function call. In this way recycling of atoms is maximized.
\section{Numerical integration}
\label{sec3}
For our case study the Linear Kalker model used to model the contact forces has been a mayor source of problems.
To some extent, these forces can be considered viscous friction forces with a
huge value for the equivalent viscous constant, making the system dynamics stiff. In this context, the use of an explicit integration scheme
is going to require a very small time step, spoiling real-time performance.
In order to solve this problem we have devised an Implicit-Explicit (IMEX) \cite{ascher1995implicit} integration schema, and adjusted it to overcome the problem associated with the contact forces without penalizing the computational cost.
The use of these schemes is not new. It has appeared in the bibliography under other names: semi-implicit \cite{arnold2007linearly}, additive or combined methods \cite{fulton2004semi}, etc. These methods use different types of discretization for the different terms in the dynamic equations. Those terms that are not related to the stiff behavior of the equations are discretized using a low-cost explicit scheme while, the stiff terms are discretized using an implicit scheme.
As commented before creep forces and moments are introduced in the model as external actions, and their contribution is embedded in vector ${\boldsymbol{\delta}}_{{\mathbf{q}}}$. Lets make this contribution explicit, by splitting ${\boldsymbol{\delta}}$ into two Kalker ($K$) and non-Kalker ($NK$) contributions.
In this way, first equation in Eq.~(\ref{eq:MphiTPhiZero4}) can be rewritten as:
\begin{equation}
\begin{bmatrix}
{\mathbf{M}}_{{\mathbf{q}}\vq} & \dot{\boldsymbol{\phi}}_{\vdq}^{n}{\T} \\
\dot{\boldsymbol{\phi}}_{\vdq}^{n} & \mathbf{0}
\end{bmatrix}
\begin{bmatrix} {\mathbf{\ddot{q}}} \\ {\boldsymbol{\lambda}}^n \end{bmatrix}
=
\begin{bmatrix} {\boldsymbol{\delta}}^{K}_{{\mathbf{q}}} + {\boldsymbol{\delta}}^{NK}_{{\mathbf{q}}} \\ {\boldsymbol{\gamma}^n} \end{bmatrix}
\end{equation}
This new set of equations can be integrated using an IMEX method. The terms related to creep forces will be integrated using an Implicit scheme and the rest using an Explicit scheme.
Eq.~(\ref{eq:kalker1}) can be expressed as a typical viscous contribution
\begin{equation}
\mathbf{f}^K= -\frac{1}{V}
\mathbf{C}^K
\boldsymbol{\nu}
\label{eq:kalker_viscous}
\end{equation}
where $\mathbf{f}^K=[f_x,f_y, m_z]\T1$, $V=\frac{1}{2}(\left|\mathbf{v}_{Gr.}^{O^w}\right|+\left|\boldsymbol{\omega}_{Gr.}^{w} \wedge \mathbf{r}_{O^w}^{P^w}\right|)$, $\boldsymbol{\nu} = [\mathbf{v}_{Gr.}^{P^w} \mathbf{t}^r_x ~~ \mathbf{v}_{Gr.}^{P^w}\mathbf{t}^r_y ~~ \boldsymbol{\omega}_{Gr.}^{w}\mathbf{n}^r]^\mathsf{T}$ and
\begin{equation}
\mathbf{C}^K=G
\left[
\begin{array}{ccc}
a b ~c_{11}&0&0\\
0& a b ~c_{22}& \sqrt{ab}~c_{23}\\
0&-\sqrt{ab}~c_{23}& (ab)^2~c_{33}
\end{array}
\right]
\end{equation}
Adding a subindex $i$ to refer to a particular contact point, the contribution ${\boldsymbol{\delta}}^{K}$ can be obtained as:
\begin{align}
{\boldsymbol{\delta}}^{K}_{{\mathbf{q}}} = \sum_{i=1}^8 \frac{\partial {\boldsymbol{\delta}}^{K}_{{\mathbf{q}}}}{\partial \mathbf{f_i}^{K}} \mathbf{f_i}^{K} =
- \sum_{i=1}^8 \frac{\partial {\boldsymbol{\delta}}^{K}_{{\mathbf{q}}}}{\partial \mathbf{f_i}^{K}} \frac{1}{V_i} \mathbf{C}^K_i
\frac{\partial \boldsymbol{\nu}_i}{\partial {\mathbf{\dot{q}}}} {\mathbf{\dot{q}}} = - \mathbf{C}^{K}_{{\mathbf{q}}\vq} {\mathbf{\dot{q}}}
\end{align}
In order to determine matrix $ \mathbf{C}^{K}_{{\mathbf{q}}\vq}$, we symbolically export matrices
\begin{equation}
\frac{\partial {\boldsymbol{\delta}}^{K}_{{\mathbf{q}}}}{\partial \mathbf{f_i}^{K}} \mathbf{C}_i^{K}\frac{\partial \boldsymbol{\nu}_i}{\partial {\mathbf{\dot{q}}}} ~,~~~i=1,\ldots,8,
\end{equation}
and numerically assemble matrix as
\begin{equation}
\mathbf{C}^{K}_{{\mathbf{q}}\vq}= \sum_{i=1}^8 \frac{1}{V_i} ~~\frac{\partial {\boldsymbol{\delta}}^{K}_{{\mathbf{q}}}}{\partial \mathbf{f_i}^{K}} \mathbf{C}^K_i
\frac{\partial \boldsymbol{\nu}_i}{\partial {\mathbf{\dot{q}}}},
\end{equation}
where $V_i$ and $\mathbf{C}^K_i$ are determined using the same procedures described in the previous section.
The contribution ${\boldsymbol{\delta}}^{NK}$, can be obtained symbolically substituting by zero en ${\boldsymbol{\delta}}_{{\mathbf{q}}}$ the symbols associated to the external forces
$f_x,f_y, m_z$ for every contact point.
Now write the dynamic equation set can be expressed as follows:
\begin{align}
\begin{bmatrix}
{\mathbf{M}}_{{\mathbf{q}}\vq} & \dot{\boldsymbol{\phi}}_{\vdq}^{n}{\T} \\
\dot{\boldsymbol{\phi}}_{\vdq}^{n} & \mathbf{0}
\end{bmatrix}
\begin{bmatrix} {\mathbf{\ddot{q}}} \\ {\boldsymbol{\lambda}}^n \end{bmatrix}
&=
\begin{bmatrix} {\boldsymbol{\delta}}^{NK}_{{\mathbf{q}}}-\mathbf{C}^{K}_{{\mathbf{q}}\vq}{\mathbf{\dot{q}}} \\ {\boldsymbol{\gamma}^n} \end{bmatrix}
\label{eq:MphiTPhiZero5}
\end{align}
The IMEX integration procedure proposed follows directly from this equation.
Must be observed that in Kalker's Linear Theory when saturation occurs this method is also valid, because in this case force can be also written as the product of a constant matrix and the creepages. The numerical solver must handle with which matrix use at each moment.
\subsubsection*{Discretization}
The contribution $\mathbf{C^{K}}_{{\mathbf{q}}\vq}{\mathbf{\dot{q}}}$ is discretized using an implicit Euler. To that end it is evaluated at the next time step $t+\Delta t$,
\begin{equation}
\mathbf{C}^{K}_{{\mathbf{q}}\vq}{\mathbf{\dot{q}}}_{t+\Delta t}
\end{equation}
An explicit Euler scheme for the remaining terms requires acceleration to be discretized as
\begin{equation}
{\mathbf{\ddot{q}}}_{t+\Delta t}=\frac{{\mathbf{\dot{q}}}_{t+\Delta t}-{\mathbf{\dot{q}}}_{t}}{\Delta t}
\end{equation}
and ${\boldsymbol{\delta}}^{K}_{{\mathbf{q}}}$ to be evaluated at $t$.
Substituting this into Eq.~(\ref{eq:MphiTPhiZero5}) the final discretization of the system
takes the form:
\begin{align}
\begin{bmatrix}
{\mathbf{M}}_{{\mathbf{q}}\vq} + \mathbf{C}^{K}_{{\mathbf{q}}\vq} \Delta t &~ &\dot{\boldsymbol{\phi}}_{\vdq}^{n}{\T} \Delta t\\
\dot{\boldsymbol{\phi}}_{\vdq}^{n} &~& \mathbf{0}
\end{bmatrix}
\begin{bmatrix} {\mathbf{\dot{q}}}_{t+\Delta t} \\ {\boldsymbol{\lambda}}^n \end{bmatrix}
&=
\begin{bmatrix} {\boldsymbol{\delta}}^{NK}_{{\mathbf{q}}} \Delta t + {\mathbf{M}}_{{\mathbf{q}}\vq} {\mathbf{\dot{q}}}_{t} \\ {\boldsymbol{\gamma}^n} \Delta t + \dot{\boldsymbol{\phi}}_{\vdq}^{n} {\mathbf{\dot{q}}}_{t} \end{bmatrix}
\label{eq:MphiTPhiZero6}
\end{align}
where all the functions are computed at time $t$.
Note that, to keep the equation solution linear, $\mathbf{C}^{K}$ is evaluated at $t$ instead of $t+\Delta t$.
The structure of this system of equations can be observed in Fig.~ \ref{fig3}. It is noticeable that the sparsity structure is very similar to the one seen in Fig.~\ref{fig2}.
\begin{figure}[h]
\centering
\includegraphics[height=32mm]{./M_Phi_dqT_Phi_dq_delta_gamma-emb.pdf}
\caption{Dynamic model structure $\left[[ {\mathbf{M}}_{{\mathbf{q}}\vq} + \mathbf{C}^{K}_{{\mathbf{q}}\vq} \Delta t,\dot{\boldsymbol{\phi}}_{\vdq}^{n}{\T} \Delta t;\dot{\boldsymbol{\phi}}_{\vdq}^{n},\mathbf{0}],[ {\boldsymbol{\delta}}^{NK}_{{\mathbf{q}}} \Delta t + {\mathbf{M}}_{{\mathbf{q}}\vq} {\mathbf{\dot{q}}}_{t};{\boldsymbol{\gamma}^n} \Delta t + \dot{\boldsymbol{\phi}}_{\vdq}^{n} {\mathbf{\dot{q}}}_{t}]\right]$}
\label{fig3}
\end{figure}
This problem has the same mathematical structure that the standard full dynamic set, so it can be solved using the same procedures. We use coordinate partitioning \cite{Wehage1982,Haug1989}. This is a good performing strategy that is also used by other practitioners in the symbolic multibody field.
We use a $LU$ procedure with full pivoting on the non tangent constraint Jacobian in the generalized velocities ${\mathbf{\dot{q}}}$, so we can choose the set of independent coordinates at each iteration step. This way, no conditions are enforced on the parameterization ${\mathbf{q}}$ used.
Thus using this IMEX scheme comes for free, as the evaluation of the functions appearing in Eq.~ (\ref{eq:MphiTPhiZero6}) has the same complexity as the functions in Eqs.~ (\ref{eq:MphiTPhiZero5}) or (\ref{eq:MphiTPhiZero4}).
It should be noted that the matrix ${\mathbf{M}}_{{\mathbf{q}}\vq} + \mathbf{C}_{{\mathbf{q}}\vq}^{K} \Delta t$ is not symmetric, so $LU$ decomposition should be used in place of $LDL^{^\mathsf{T}}$ incurring a small penalty in performance.
The solution of this system will give the value of the generalized velocities at $t+\Delta t$, ${\mathbf{\dot{q}}}_{t+\Delta t}$. To obtain the coordinates
at ${\mathbf{q}}_{t+\Delta t}$ the following explicit mid-point rule is used:
\begin{equation}
{\mathbf{q}}_{t+\Delta t}={\mathbf{q}}_{t}+\frac{{\mathbf{\dot{q}}}_{t+\Delta t}+{\mathbf{\dot{q}}}_{t}}{2}{\Delta t}.
\end{equation}
Note that it is second order and comes at no cost.
Next coordinate projection is performed. First Eq.~(\ref{eq:phit}) is used to obtain the contact coordinates ${\mathbf{s}}_{t+\Delta t}$ in terms of the ${\mathbf{q}}_{t+\Delta t}$. As ${\mathbf{q}}_{t+\Delta t}$ is accurate to second order, this procedure gives a error of the same order. To this end the following iterative Newton-Raphson procedure is used:
\begin{equation}
\dot{\boldsymbol{\phi}}_{\vds}^{d}({\mathbf{q}},{\mathbf{s}}) ({\mathbf{s}}_{k+1}-{\mathbf{s}}_k) = -{\boldsymbol{\phi}^d}({\mathbf{q}},{\mathbf{s}}) \label{eq:NRs}
\end{equation}
this usually involves a single iteration\footnote{This is related to the NSWHS frame used to define the contact point in the wheel.}. After the update of ${\mathbf{s}}$, Eq.~(\ref{eq:phin}) is solved for ${\mathbf{q}}$ using the same iterative procedure:
\begin{equation}
\dot{\boldsymbol{\phi}}_{\vdq}^{n}({\mathbf{q}},{\mathbf{s}}) ({\mathbf{q}}_{k+1}-{\mathbf{q}}_k) = -{\boldsymbol{\phi}^n}({\mathbf{q}},{\mathbf{s}}) \label{eq:NRq}
\end{equation}
This procedure usually converges in a single iteration. Note that ${\mathbf{q}}_{t+\Delta t}$ is accurate to second order after the integration step. Note that the LU decomposition of the previous Jacobians, $\dot{\boldsymbol{\phi}}_{\vdq}^{n}({\mathbf{q}},{\mathbf{s}})$ and $\dot{\boldsymbol{\phi}}_{\vdq}^{d}({\mathbf{q}},{\mathbf{s}})$, is known as they have computed at the previous velocity projection step (described latter). So the Jacobian and its decomposition is not updated in this step.
In the velocity projection step, first Eq.~(\ref{eq:dqproj}) is solved for ${\mathbf{\dot{q}}}$. To that end, the Jacobian $\dot{\boldsymbol{\phi}}_{\vdq}^{n}({\mathbf{q}},{\mathbf{s}})$ and its $LU$ decomposition are updated. Then, Eq.~(\ref{eq:dsproj}) is solved for ${\mathbf{\dot{s}}}$. To that end, the Jacobians $\dot{\boldsymbol{\phi}}_{\vdq}^{d}({\mathbf{q}},{\mathbf{s}})$ and $\dot{\boldsymbol{\phi}}_{\vds}^{d}({\mathbf{q}},{\mathbf{s}})$ are updated
and the $LU$ decomposition of $\dot{\boldsymbol{\phi}}_{\vds}^{d}({\mathbf{q}},{\mathbf{s}})$ is computed.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{algoritmo.pdf}
\caption{Integration procedure}
\label{fig:integration}
\end{figure}
In Fig.~\ref{fig:integration}, a schematic representation of the integration procedure described here is presented.
To get a more clear picture, the steps related to the determination of the creep forces has been represented.
\section{Results}
\label{sec4}
\subsubsection*{Simulation description}
The track used in the simulation starts and ends with two straight and
parallel segments running in the $x$ direction and separated $50 ~\text{m}$.
Both stretches are joined by a symmetric and smooth double transition curve $270 ~\text{m}$ long in direction $x$.
On top of the defined geometry, two harmonic vertical irregularities with an amplitude of $10~\text{mm}$ are added. These irregularities are defined
using a sine wave that runs in direction of $x$ with a wave length of $10~\text{m}$. Right and left rail
irregularities present a phase difference of $\pi/2$. As commented earlier, third order splines are used to discretize the whole track, including the irregularity.
The simulation starts with an initial forward speed of $23.7~\text{m/s}$ with the $Vehicle$ $Body$ centered at $x=0 ~\text{m}$ and with a lateral misalignment of $5~\text{mm}$ with respect to to the track center.
Vehicle motors are actuated with a constant $200~\text{Nm}$ torque.
\subsubsection*{Computational results}
Fig.~\ref{fig:xzmainbody2} shows the trajectory followed by the $Vehicle\ Body$ center.
Note that the the given initial state is not in dynamic equilibrium and therefore, the oscillations at the beginning
of the simulation are in part due to this. This is related to the sudden application of torque at the simulation start.
By the time that the vehicle center enters the track, the oscillations seen are no longer related to the initial condition.
\begin{figure}[h]
\centering
\includegraphics[width=0.99\textwidth]{trainpos2.pdf}
\caption{Trajectories followed by the main body}
\label{fig:xzmainbody2}
\end{figure}
Fig.~\ref{fig:hunting} shows a zoom of the first graph in Fig.~\ref{fig:xzmainbody2}. This is done to make the oscillations
in that plane visible. The zone in which the vehicle exits the second curve is shown.
Two different oscillations can be seen. Two oscillations are clearly distinguishable: The hunting oscillation is the one with the largest wave length, while the shorter one is related
to the irregularities of the track.
\begin{figure}[h]
\centering
\includegraphics[width=0.99\textwidth]{zoom.pdf}
\caption{Hunting oscillation in detail.}
\label{fig:hunting}
\end{figure}
Creep velocities and creep forces and moments are presented in Fig.~\ref{fig:kalkerforces}.
It can be observed that creep velocities are higher when the vehicle is at the middle of the curved tracks
($t\approx 5~s$ and $t\approx 10~s$). The same behavior is seen for the forces and moments. Small fluctuations
on the creepages and forces in the second straight track ($t> 15~s$) are due to the vertical irregularities.
\begin{figure}[h]
\centering
\includegraphics[width=0.99\textwidth]{kalker26.pdf}
\caption{Creep velocities and forces the Rear Bogie, Front-Right wheelset }
\label{fig:kalkerforces}
\end{figure}
Using the proposed IMEX integrator with a $\Delta t = 1~\text{ms}$ a stable integration is achieved by a generous margin.
In the same conditions, using an explicit Euler for velocities and the explicit mid-point rule for accelerations time steps smaller than
$10^{-4}~\text{ms}$, not compatible with a real-time performance, are required. Implicit trapezoidal rule has also been used with showing a stable
behavior around $\Delta t \approx 1~\text{ms}$ with not such a generous margin.
Using the IMEX Euler method a fine grained analysis of the computation times required by the different steps of the proposed algorithm is done.
The results are given in Table~ \ref{table:times}. A seven years old \textit{Intel Core vPro i5 @ 3500}{MHz} has been used for the test.
From this data it can be seen that it takes $256 ~\mu \text{s}$ of CPU time to complete one integration step. That is, \textit{hard} real-time performance
is achieved by a wide margin using the proposed procedures. In comparison, using the trapezoidal rule \textit{soft} real-time performance can be achieved by a short margin.
\begin{table}[]
\centering
\caption{Results per time step ($1~\text{ms}$)}
\label{table:times}
\begin{tabular}{lrr}
\hline
Task & CPU Time $\mu \text{s}$ \\
\hline
Contact Area, Look up Kalker coeffs. and evaluate $\frac{\partial {\boldsymbol{\delta}}^{K}_{{\mathbf{q}}}}{\partial \mathbf{f_i}^{K}} \mathbf{C}_i^{K}\frac{\partial \boldsymbol{\nu}_i}{\partial {\mathbf{\dot{q}}}}$ ($\times 8$) & 30 \\
Look up Wheel profile, rail profile and railway spline coeffs. ($\times 8$) & 1 \\
Evaluate ${\mathbf{M}}$,${\boldsymbol{\delta}}$ and ${\boldsymbol{\gamma}^n}$ & 55 \\
Dynamics Solution $\Rightarrow {\mathbf{\dot{q}}}, {\boldsymbol{\lambda}}$ & 103 \\
Time Integration $\Rightarrow {\mathbf{q}}$ & 1 \\
Evaluate ${\boldsymbol{\phi}^n}$, ${\boldsymbol{\phi}^d}$, $\dot{\boldsymbol{\phi}}_{\vdq}^{n}$, $\dot{\boldsymbol{\phi}}_{\vdq}^{d}$, $\dot{\boldsymbol{\phi}}_{\vds}^{d}$ and ${\boldsymbol{\beta}}^n$ & 29 \\
Projection Solution $\Rightarrow {\mathbf{q}}, {\mathbf{\dot{q}}}, {\mathbf{s}}, {\mathbf{\dot{s}}}$ & 37 \\
\hline
Total Time & 256 \\
\hline
\end{tabular}
\end{table}
In Fig.~\ref{fig:nriter} the number of iterations required by the ${\mathbf{q}}$-projection and ${\mathbf{s}}$-projection steps are shown.
It is noticeable that the ${\mathbf{q}}$-projection only requires a single Newton-Raphson\footnote{the tolerance used is $10^{-6}$ amounting to a negligible error of $\approx 10^{-3}~\text{mm}$ for lengths} iteration.
The same is true for the ${\mathbf{s}}$-projection. This has required to integrate ${\mathbf{s}}$ after the integration step using an explicit Euler procedure ${\mathbf{s}}_{t+\Delta t}={\mathbf{s}}_t + {\mathbf{\dot{s}}}_t ~\Delta t$ leading to an to a smaller error ($O(\Delta t^2)$)
at the start of the Newton-Raphson iteration. Clearly, the increased number of iterations is coincident with
the curved stretches. This result justify the approach adopted in which, the ${\mathbf{s}}$-projection is performed before the ${\mathbf{q}}$-projection.
\begin{figure}[h]
\centering
\includegraphics[width=0.99\textwidth]{nr_iter2.pdf}
\caption{Iterative steps needed by the Newton-Rapshon algorithm}
\label{fig:nriter}
\end{figure}
In Table~\ref{table:functions} the number of operations required for the evaluation of the different functions used by the proposed dynamic formalism are presented.
\begin{table}[h]
\centering
\caption{Atoms and operation for the evaluation of the model functions}
\label{table:functions}
\begin{tabular}{lrr}
\hline
Function & Atoms & Operations \\ \hline
${\mathbf{M}}_{{\mathbf{q}}\vq}$ & 1795 & 10910 \\
${\boldsymbol{\delta}}_{{\mathbf{q}}}$ & 3648 & 19015 \\
${\boldsymbol{\phi}^n}$ & 425 & 1709 \\
${\boldsymbol{\phi}^d}$ & 541 & 2150 \\
$\dot{\boldsymbol{\phi}}_{\vdq}^{n}$ & 784 & 4116 \\
$\dot{\boldsymbol{\phi}}_{\vds}^{n}$ & 846 & 6668 \\
$\dot{\boldsymbol{\phi}}_{\vdq}^{d}$ & 964 & 6574 \\
$\dot{\boldsymbol{\phi}}_{\vds}^{d}$ & 1113 & 8437 \\
${\boldsymbol{\beta}}^n$ & 0 & 0 \\
${\boldsymbol{\gamma}^n}$ & 3779 & 25870 \\
$\frac{\partial {\boldsymbol{\delta}}^{K}_{{\mathbf{q}}}}{\partial \mathbf{f_i}^{K}} \mathbf{C}_i^{K}\frac{\partial \boldsymbol{\nu}_i}{\partial {\mathbf{\dot{q}}}}$ & 350 & 1917 \\ \hline
\end{tabular}
\end{table}
The results show a correlation between the time for function evaluations and the number of operations.
As a major result of this study, it can be seen that using the symbolic procedures proposed the penalties incurred for
using an exact treatment of the linear Kalker contact model are barely noticeable. Note that this is a fair comparison,
as the operation count related to other dynamic and kinematic computations are very optimized, showing numbers compatible with state-of-the-art
recursive formulations. This
puts into perspective the relevance of symbolic methods proposed in achieving hard real-time performance in the railway dynamics simulation context.
Still, there are still some possibilities to further improve the results given in this article.
1.-The dynamic system structure shown in Fig.~\ref{fig2} shows a decent amount of sparsity. This sparsity is shared with the IMEX discretized dynamic matrix.
Important savings can therefore be obtained using a sparse $LU$ algorithm.
2.- In the ${\mathbf{s}}$-projection and ${\mathbf{\dot{s}}}$-projection problems $\dot{\boldsymbol{\phi}}_{\vds}^{d}$ is a maximum rank block-diagonal matrix with
$4\times4$ blocks \cite{samin2003}. Therefore, its computation can be speeded up by big integer factor. The solution could be easily
implemented symbolically or even in parallel.
3.- Removal of the repeated evaluation of constant atoms from the symbolic functions and reuse of atoms common to different exported functions.
As commented in the introduction, at the expense of some accuracy, partial linearization \cite{Escalona2015} or
base parameter reduction \cite{Iriarte2015}, can be used to further improve the computational performance of the model.
\section{Conclusions}
\label{sec5}
The purpose of the article was to test state-of-the-art methods for the symbolic modeling
in the railway context. A complex locomotive running on a track with a complex and general surface geometry has been modeled and tested.
Main aspects of the symbolic methods proposed are summarized: atomization, recursive operators, points and bases structures, general parameterization, etc.
Based on this methods the model is obtained using a direct implementation of the principle of virtual work.
Creep forces and moments are modeled using a direct symbolic implementation of the linear Kalker model without simplifications.
An Implicit-Explicit (IMEX) integrator has been proposed to cope with the contact model while attaining real-time performance.
The resulting equations are solved using coordinate partitioning MSD procedures.
A very stable hard-real-time-compatible performance with a time step of $1~\text{ms}$ is obtained. A CPU time of $256~\mu\text{s}$ per time step
is required in a seven year old \textit{Intel Core vPro i5 @ 3500} MHz.
It is noticeable the small time required for the determination of the creep forces
when an exact implementation of the linear Kalker model is used. Also, the compromise efficiency/robustness of the IMEX integrator proposed is remarkable.
The results obtained show the relevance of the
methods proposed for the real-time simulation of railway vehicles.
There are still obvious possibilities to improve on the results presented in this work: better sharing of atoms,
constant atom revaluation, sparse linear solver implementation and parallelization, are the most obvious.
On top of this, with a small accuracy
penalty, techniques such as partial-linearization and parameter reduction can be used
to improve even further the results presented.
\section{Acnowledgements}
This work was partially supported by the ``Plan Nacional de InvestigaciΓ³n CientΓfica, Desarrollo e InnovaciΓ³n TecnolΓ³gica'', ``Ministerio de EconomΓa y Competitividad'' [grant numbers IPT-2011-1149-370000 and TRA2014\_57609\_R].
\bibliographystyle{unsrt}
| aef569ce659e77ea5011a8209c6470e2577bb397 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The spiral flow in the annulus between concentric circular cylinders is an example
of a flow that is susceptible to transition through combined action of several
physical mechanisms. In the particular case when the helical pattern of the basic flow streamlines that characterises a spiral flow results from an imposed axial pressure gradient
in conjunction with rotation of the inner cylinder alone, there are two well-known distinct physical mechanisms of disturbance propagation that can trigger transition. These are:
1) the {\it Tollmien-Schlichting mechanism} and, 2) the {\it Taylor mechanism}. The features of these mechanisms acting on their own come convincingly to light when transition of
the spiral flow in question in its limiting cases are contrasted with each other. The limiting cases are when the pitch of the helix, measured through the angle the helix makes
with the common axis of the cylinders, assumes values of either $ 0 $ or $\frac{\pi}{2}$. Differences between the {\it Tollmien-Schlichting mechanism} and the {\it Taylor mechanism}
are indeed striking and readily visible in flow visualization studies, see eg. \cite{albumvandyke}. In theory the differences manifest themselves most convincingly when dynamical
characteristics of transition inducing disturbances in the limiting cases of the spiral flow are contrasted with each other. An outstanding difference between the two limiting cases
is the appearance of a {\it critical layer} within the flow when the {\it Tollmien-Schlichting mechanism} alone is responsible for transition, which is realised when the pitch of
the helix is $0$. In contrast, no {\it critical layer} appears when the {\it Taylor mechanism} alone causes transition as it is when the pitch is $\frac{\pi}{2}$. From physical
reasoning based on affinity of disturbance propagation characteristics, one would expect that a {\it critical layer} might arise in the spiral flow undergoing transition when the
pitch of the helix of its basic flow streamline is close to $0$. It would vanish on the pitch increasing from $0$, to disappear entirely when the pitch reaches $\frac{\pi}{2}$. It
is then of scientific interest to examine when and/or under what conditions a {\it critical layer} may arise or disappear during transition of the spiral flow. The aim of the
present work is to derive these conditions from first principles, {\it i.e.} from the linearised equations governing propagation of small-amplitide disturbances to the spiral flow.
The significance of the {\it critical layer} in transition is related to the generation of {\it Reynolds stresses}, see eg. \cite{stuart}, \cite{cclin}, \cite{drazinreid},
\cite{maslowe1}, \cite{maslowe2}, \cite{schmidhenningson}, \cite{ramagovindarajan}, \cite{criminalejacksonjoslin}. Besides the differences in the mechanisms referred to earlier
some further pronounced differences may be outlined. These are as follows. In the {\it Tollmien-Schlichting mechanism} transition is induced by travelling waves of the
disturbance. These are effective even in a basic flow with straight and parallel streamlines. Curvature of the basic flow streamlines is therefore not essential for onset of
transition through this mechanism. In contrast, in the {\it Taylor mechanism} transition is not possible in the absence of curvature of the basic flow streamlines. When the
{\it critical layer} appears, it is located at a position in the flow where flow velocity and the phase velocity of the neutrally stable travelling disturbance wave are equal to
each other. Its location coincides with a singularity occuring in the inviscid form of the set of linearized equations for disturbance propagation. The singularity is most readily
evident when the set of linearised equations for disturbances are cast in a form of the {\it Orr-Sommerfeld equation}. It is then recognisable through the coefficient of the second
derivative of the wall-normal component of the disturbance velocity changing its sign.
\subsection*{Outline of the present paper}
In Section 2 the present state of understanding of the {\it critical layer} during transition of spiral Poiseuille flow is subject to a brief critical review. The purpose of this
Section is to prepare ground for the reformulation of the disturbance propagatioon problem in a spiral Poiseulle flow undertaken in the present work.
It will be seen in the course of this work that a {\it reformulation} of disturbance propagation is essential for proper imbedding of the phenomenon in the flow of current interest
in the extensively studied past work of disturbance propagation in its two limiting cases.
In Section 3 the reformulation of the problem is introduced, which is essentially the derivation of the
{\it Generalised Orr-Sommerfeld and Squire equations}. In Section 4 the relations of the {\it Generalised Orr-Sommerfeld and Squire equations} to the limiting cases of flows
with mild and strong swirl are illuminated. The equations for the case of mild swirl bring out the outstanding characteristics of the {\it critical layer in spiral flow} in an
analytical form. In Section 5, the meaning of the analytical results derived are discussed and summarised.
\section{A brief critical review of the present state of understanding of the {\it critical layer} during transition of the spiral Poiseuille flow}
When both the {\it Tollmien-Schlichting mechanism} and the {\it Taylor mechanism} are acting simultaneously, as in the {\it annular spiral flow}, questions regarding the dominance
of one over the other in a certain region of the parameter space arise. Forms of manifestation of this dominance are, as expected, of fundamental scientific interest. Physical reasoning would lead one
to expect that the {\it Tollmien-Schlichting mechanism} would dominate over the other when the pitch of the helix is close to $0$, whereas it would be the {\it Taylor mechanism}
when the pitch is close to $\frac{\pi}{2}$. This expectation would in turn lead to the possibility of a {\it critical layer} arising in the spiral flow
when its helical angle lies close to $0$. Besides the mechanism as such, knowledge of the character of the dominant transition inducing disturbance, in particular whether it is
toroidal or helical in nature, would shed light on the changes that might be expected in the observable flow
pattern as the flow is taken through the parameter space, thus forming a major step towards understanding of the complexity of transition in this apparently simple flow.
It is therefore only but natural that a large number of research papers are devoted to studies of disturbance propagation and transition in spiral flows. A
cursory search through published literature shows them to have been appearing from time to time, starting from the early days of transition research, see eg. the classical book of
Chandrasekhar \cite{chandrasekhar},
and continuing right into the present day. An insight into the highly ramified nature of the problems arising in this and related flows is provided by the
papers of Hasoon and Martin \cite{hasoonmartin}, DiPrima and Pridor \cite{diprimapridor}, Takeuchi and Jankowski \cite{takeuchijankowski}, Ng and Turner \cite{ngturner},
papers authored or co-authored by Cotrell and Pearlstein \cite{cotrellpearlstein}, by Meseguer and Marques \cite{meseguermarques1, meseguermarques2, meseguermarques3}, by Avila,
Meseguer and Marques \cite{avilameseguermarques}, \cite{meseguermellibovskyavilamarques}, \cite{marquesmeseguerlopezetal}, by Leclerq, Pier and Scott \cite{leclercqpierscott},
Deguchi and Altmeyer \cite{deguchialtmeyer}, \cite{altmeyer}, and by the authors of further references cited in these papers.
An examination of the equations for the study of disturbance propagation which are employed in the cited references shows that the approaches taken by the authors may be grouped
under two broad headings depending upon the manner in which the pressure disturbance is treated. In one, the pressure disturbance is treated explicitly as an unknown, which leaves
four unknowns to be determined from the four equations that are available, {\it viz.} the momentum equations in the three directions and the continuity equation. In the other, the
pressure disturbance does not appear as an unknown, and there are only three unknowns which are the components of the velocity disturbance. The task then faced is to formulate an
appropriate set of three equations for determining the three components of the velocity disturbance, without loss of substance of the subject of investigation. A study of
published literature shows that this task is carried out through procedures tailored to the basic flow. For the basic flow with straight and parallel streamlines it is the
procedure for derivation of the {\it Orr-Sommerfeld equation}. For the basic flow with concentric circular streamlines it is the one worked out by Taylor for the
{\it Taylor stability problem}, see Taylor \cite{taylorscientificpapers}. Both the procedures are described in well known published literature, eg. Stuart \cite{stuart},
Schmid and Henningson \cite{schmidhenningson}, Criminale, Jackson and Joslin \cite{criminalejacksonjoslin}, Chossat and Iooss \cite{chossatiooss}, Meyer-Spasche \cite{meyerspasche}. The relation between the two
apparently diverse derivation procedures for elimination of the pressure disturbance in these two basic
flows does not seem to have been a subject for discussion in published literature. A pressure disturbance elimination procedure for the basic flow of present interest which is
characterized by helically wound streamlines, that is worked out along lines analogous to those in the other kinds of basic flow referred to, is also, to the knowledge of the
authors, not known from published literature. One of the objectives of the present work is to present such a derivation
procedure for the {\it spiral Poiseuille flow} that bridges the procedures for elimination of the pressure disturbance as an unknown in basic flows with
{\it straight and parallel streamlines} on one hand and {\it concentrically circular streamlines} on the other.
At this juncture it is in order to draw attention to some salient differences between the equations for the disturbance formulated in the two approaches just mentioned, when they
are applied to the two limiting cases considered. In the approach with the pressure disturbance retained as an unknown, derivatives of the velocity disturbance appear up to the second order,
whereas for the pressure disturbance they are present only to the first order. In the approach with the pressure disturbance eliminated, fourth order derivatives of the
velocity disturbance appear for both the limiting cases, however with a subtle structural difference between the two. For the limiting case with the basic flow described by
straight and parallel streamlines, the coefficient of the term with a second-order derivative of the velocity disturbance may pass through a zero, whereas there is no such
zero-crossing in the equation for the velocity disturbance in the case with concentrically circular streamlines in the basic flow. It is now well known that the zero-crossing in the coefficient of the second derivative is decisive for the
occurence of a {\it critical layer}, {\it viz. Tollmien's critical layer}, its location being given by this zero-crossing as an asymptotic approximation. As a matter of fact, it is
the zero-crossing of the coefficient of the second derivative of the velocity disturbance that renders the occurence of the critical layer transparent from the governing equations
themselves. As against this, no such critical layer arises in the corresponding equations for the basic flow with concentric circular streamlines. Neither is the critical layer
discernible from the governing equations themselves if the pressure disturbance is retained as an unknown in their formulation.
In {\it spiral Poiseuille flow}, the {\it strength} of swirl may be characterised through the ratio of the {\it two characteristic velocities} of the basic flow, which are the
{\it axial} and {\it azimuthal characteristic velocities}, denoted $U_{refx}$ and $U_{ref \varphi}$ respectively. When the two mechanisms are acting simultaneously, as in the flow
in question, intuition would suggest that the disturbance propagation mechanism of the transition inducing disturbance would lie closer to either the {\it Tollmien-Schlichting} or
the {\it Taylor} mechanism, depending upon the ratio of the two characteristic velocities. One would then expect that, for an arbitrarily given value of this ratio, which may lie
between between zero and infinity,
the set of linearised equations themselves that describe the velocity disturbances in the swirling
basic flow in the annulus would lie closer to the corresponding set of equations for the velocity disturbances when the ratio of the two characteristic
velocities attains values of zero or infinity, as the case may be.
These are the Orr-Sommerfeld and Squire equations for the plane channel flow, possibly corrected for the effects
of transverse curvature, when the azimuthal characteristic velocity is zero, and, Taylor's equations for disturbance propagation in the annular gap between concentric cylinders
when the axial characteristic velocity is zero.
A close surveillance of the published work in this area, however, shows that, when the approach with elimination of the pressure
disturbance is followed, the equations employed in the cited literature to describe disturbance propagation in swirling flows do not degenerate to the Orr-Sommerfeld and Squire
equations in the absence of swirl in the basic flow.
\section{The proposed alternative formulation}
The basic flow of our problem is the fully developed spiral flow in an annulus. The proposed alternative approach views the phenomenon of disturbance propagation in this
flow in terms of a set of revised parameters which are going to be shortly introduced. Conventionally, the axial and azimuthal velocity components in the
basic flow of our problem, $V_{Gx}$ and $V_{G \varphi}$, are written as products of the two characteristic velocities in the two directions, {\it viz.} $U_{refx}$ and $U_{ref\varphi}$
respectively, with the shape functions in that order, $U_x (r; \epsilon_R)$ and $U_{\varphi} (r; \epsilon_R)$. Here $r$ is the radius, and $\epsilon_R = \frac{R_o - R_i}{R_o + R_i}$ is a
measure of the transverse curvature with $R_o$ and $R_i$ denoting the
radii of the outer and the inner cylindrical surfaces of the annulus. In the particular case of the spiral flow in question, in which the flow is maintained in the
wall-bounded annulus between the concentric cylinders by an axial pressure gradient, $\frac{dP_G}{dx}$, together with the inner cylinder rotating with
an angular velocity, $\Omega_i$, a natural choice for the characteristic velocities would be: $U_{refx}= - \frac{H^2}{2 \mu} \frac{dP_G}{dx}$ and $U_{ref\varphi} = R_i \Omega_i$.
Here, $2H = R_o - R_i$ denotes the annular gap between the cylinders, and $\mu$ the dynamic viscosity. We may note that the profiles of the axial and the azimuthal velocity
components of the basic flow, $U_{Gx}$ and $U_{G \varphi}$, depend upon the transverse curvature parameter $\epsilon_R$. They may also depend upon further parameters entering the
problem, which might be the case when e.g. a cylinder wall is prescribed to move with a certain axial translational velocity.
The revised parameters for our problem are the set constituting a characteristic velocity, $U_{ref}$, defined through $U_{ref}^2 = U_{refx}^2 + U_{ref \varphi}^2$, and a swirl
parameter, $S$, defined through $S = \frac{U_{ref \varphi}}{U_{refx}}$. The velocity components of the basic flow are then given in terms of these revised parameters
through the following expressions:
\begin{eqnarray}
V_{Gx} = U_{ref} (1+S^2)^{- \frac{1}{2}} U_{Gx}(r; \epsilon_R); ~ ~ ~ V_{G \varphi} = U_{ref} S (1+S^2)^{- \frac{1}{2}} U_{G \varphi}(r; \epsilon_R).
\label{basicflowrevisedparam}
\end{eqnarray}
The starting point for our work is the set of equations of motion in cylindrical co-ordinates which may be found in standard text books on fluid dynamics, ({\it vide} e.g.
\cite{schlichting}, \cite{batchelor}). Non-dimensionalisation of these equations with $U_{ref}$ and $H$ as reference quantities, setting the velocity and the pressure fields as sums of
the basic flow and a disturbance, followed by linearisation for disturbances lead to equations for small-amplitude disturbances to the velocity and pressure fields of the basic flow.
From these linearised equations for disturbances, the pressure disturbance may be eliminated through two different procedures which are described for the basic flow with straight
and parallel streamlines in standard works on hydrodynamic stability {\it vide} e.g. \cite{schmidhenningson}. We refer to these procedures as the {\it Orr-Sommerfeld} and
{\it Squire} procedures respectively. Carrying out these procedures in the cylindrical co-ordinate system adopted for the present problem leads in that order to the
{\it generalised Orr-Sommerfeld} and {\it Squire equations}. They are, together with the {\it Continuity equation}, the governing equations for our problem. The unknowns in this
set of equations are only the three components of the velocity disturbance, since the disturbance to the pressure does not appear any longer as an unknown. Profiles of the axial
and azimuthal components of the basic flow, $U_{Gx}(r)$ and $U_{G \varphi}(r)$, and their first and second derivatives occur as coefficients in the governing equations. These
equations are given to the order $O(r^{-2})$ in the {\bf Appendix}. For the purposes of the present work it is convenient to write the set in a compact matrix form as follows:
\begin{equation}
D \mathbf u = \mathbf 0.
\label{generalisedossqcont}
\end{equation}
We refer to (\ref{generalisedossqcont}) as the set of {\it generalised Orr-Sommerfeld, Squire and Continuity equations}.
Here the linear differential operator, $ D$, depends upon the Reynolds number $Re$ based on the characteristic velocity $ U_{ref}$ and the semi-gap width $ H $ as reference quantities,
on the swirl parameter $S = \frac{U_{ref \varphi}}{U_{ref x}}$, and also upon the ratio of the annular gap-width to the mean cylinder radius, $\epsilon_R$.
The differential operator $ D$ itself may be
written in a matrix form as follows:
\begin{eqnarray}
D = \left(
\begin{array}{ccc}
D_{OSr} & D_{OS \varphi} & D_{OSx}\\
D_{Sqr} & D_{Sq \varphi} & D_{Sqx}\\
D_{Cor} & D_{Co \varphi} & D_{Cox}
\end{array}
\right),
\label{breakdownofossqco}
\end{eqnarray}
where the subscripts $OS$, $Sq$ and $Co$ indicate in that order the matrix partial differential operators arising from the {\it generalised Orr-Sommerfeld}, {\it Squire} and Continuity
equations, and the subscripts $r, \varphi$ and $x$ stand for partial derivatives operating on $u_r, u_{\varphi}$ and $u_x$ respectively. Expressions for the operators in
(\ref{breakdownofossqco}) follow from a comparison of (\ref{breakdownofossqco}) with (\ref{generalisedorrsommerfeld}), (\ref{generalisedsquire}) and the continuity equation given in the {\bf Appendix}. We may draw the reader's
attention at this juncture to a salient difference between the {\it generalised Orr-Sommerfeld equation}, and {\it Orr-Sommerfeld equation} for the conventional problem. It is that,
in contrast to the conventional case in which disturbance propagation in a basic flow with wall-parallel streamlines is described, in the {\it generalised Orr-Sommerfeld equation}, the
wall-normal component of the velocity disturbance $u_r$ does not stand all by itself. The transverse curvature inherent in the geometry causes $u_r$ to occur coupled with $u_{\varphi}$.
\section{The relation between the {\it generalised Orr-Sommerfeld and Squire equations} and the equations for the classical cases.}
It will be shown in this section that the {\it generalised Orr-Sommerfeld and Squire equations}, (\ref{generalisedossqcont}), which are written in terms of the revised parameters,
degenerate into the known classical cases when the swirl parameter, $S$, approaches the limits of $0$ or $\infty$. These are the well-established equations
describing the transition mechanisms of {\it Tollmien-Schlichting} and {\it Taylor} respectively. Besides
establishing these relationships, the equations for the limit $S \rightarrow 0$ also show in analytical form the nature of the dependence of the location of the critical layer
on the parameter $S$. For illustrative purposes we will demonstrate this degeneration process for the example of the fully developed spiral flow in an annulus whose values of
the ratio of the gap width to the mean radius is small, and in conjunction with rotation of the inner cylinder alone. To this end, it is meaningful to transform the variable from the radius $r$ to $y$, where $y$
is the radial co-ordinate measured from the mean radius $\frac{R_o + R_i}{2}$, non-dimensionalised with $H$, and convert the partial differential equation (\ref{generalisedossqcont})
into an ordinary differential equation. The conversion may be carried out through substitution of the standard modal ansatz
\begin{equation}
\mathbf u = \mathbf A(y) \exp \{\imath (\lambda_x x + n_{\varphi} \varphi) - \omega t \} + c.c.,
\label{modalansatzforu}
\end{equation}
in (\ref{generalisedossqcont}). In (\ref{modalansatzforu}), $\bf A$ is a column vector whose components are, in a self-explanatory notation, $(A_r, A_{\varphi}, A_x)^T$. Using
(\ref{modalansatzforu}) in conjunction with asymptotic expansions for $U_{Gx}$ and $U_{G \varphi}$ for small values of the geometrical parameter
$\epsilon_R = \frac{R_o - R_i}{R_o + R_i} = \frac{2 H}{R_o + R_i} \rightarrow ~ 0$, which may be verified to be $ \left(U_{Gx} \right)_{\epsilon_R \rightarrow 0} \simeq \left((1-y^2) - \epsilon_R \frac{y(1-y^2)}{3} + O(\epsilon_R^2) \right) $
and
$ \left(U_{G \varphi}\right)_{\epsilon_R \rightarrow 0} \simeq \left(\frac{(1-y)}{2} - \epsilon_R \frac{(1-y^2)}{2} + O(\epsilon_R^2) \right)$,
casts (\ref{generalisedossqcont}) into the set of ordinary differential equations for $(A_r, A_{\varphi}, A_x)^T$.
In order to facilitate comparison of the different derivatives of $\mathbf A$ in the {\it generalised Orr-Sommerfeld, Squire and Continuity equations} with corresponding terms in
plane channel flow it is meaningful to rewrite the converted set of ordinary differential equations ordered according to the coefficients of the vectorial velocity disturbance and
its derivatives according to the following scheme:
\begin{equation}
\mathbf K_4 \frac{d^4 \bf A}{dy^4} + \mathbf K_3 \frac{d^3 \bf A}{dy^3} + \mathbf K_2 \frac{d^2 \bf A}{dy^2}+ \mathbf K_1 \frac{d \bf A}{dy} + \mathbf K_0 \bf A = 0.
\label{reorderedgeneralisedossqcont}
\end{equation}
In (\ref{reorderedgeneralisedossqcont}), $\mathbf K_4, . . \mathbf K_0 $ are matrices that are derivable from the matrix differential operator $ D$, together with the
ansatz (\ref{modalansatzforu}). They may themselves be grouped as follows:
\begin{eqnarray}
\mathbf K_k = \left(
\begin{array}{ccc}
K_{kOSr} & K_{kOS \varphi} & K_{OSx}\\
K_{kSqr} & K_{Sq \varphi} & K_{Sqx}\\
K_{Cor} & K_{Co \varphi} & K_{Cox}
\end{array}
\right),
\label{defnmathbfk}
\end{eqnarray}
where the subscript $k$ stands for $k = 0, 1, 2, 3$ or $4$. Many elements of the coefficient matrices $\mathbf K_0, . . . \mathbf K_4 $ are zero.
From the {\it Orr-Sommerfeld} row of the equations (\ref{reorderedgeneralisedossqcont}, \ref{defnmathbfk}) we may obtain, on division through $(\imath \omega)$,
the ordinary {\it generalised Orr-Sommerfeld} equation that can be written in a form that clearly brings out the similarities and differences between the classical
case of the plane channel flow and the annular spiral flow presently under consideration. We get, including terms to the order $O(\epsilon_R)$:
\begin{eqnarray}
\frac{1}{\imath \omega} \frac{1}{Re} \left(\frac{d^4 A_r}{dy^4} \right) + \left(1- \frac{\lambda_x}{\omega} \frac{(1-y^2)}{\sqrt{(1+S^2)}} - \imath \frac{2}{Re}\frac{\lambda_x^2}{\omega} \right) \frac{d^2 A_r}{dy^2} \nonumber \\
- \left(1 + \frac{\lambda_x}{\omega} \frac{2}{\sqrt{(1+S^2)}} - \frac{\lambda_x^3}{\omega} \frac{(1-y^2)}{\sqrt{(1+S^2)}} \right) A_r \nonumber \\
- \epsilon_R \left[\frac{\imath}{ \omega} \frac{2}{Re} \frac{d^3A_r}{dy^3} + \left( \frac{\lambda_x}{\omega} \frac{1}{\sqrt{(1+S^2)}} \frac{y(1-y^2)}{3} + \frac{n_{\varphi}}{\omega} \frac{S}{\sqrt{(1+S^2)}} \frac{(1-y)}{2} \right) \frac{d^2 A_r}{dy^2} \right] \nonumber \\
+ \epsilon_R \left[1 - \frac{\lambda_x}{\omega} \frac{(1-y^2)}{\sqrt{(1+S^2)}} + \imath \frac{2 \lambda_x^2}{\omega Re} \right] \frac{dA_r}{dy} \nonumber \\
+ \epsilon_R \left[\frac{\lambda_x}{\omega} \frac{1}{\sqrt{(1+S^2)}}\left(4y + \frac{y(1-y^2) \lambda_x^2}{3} \right) + \frac{n_{\varphi}}{\omega} \frac{S}{\sqrt{(1+S^2)}} \frac{(1-y)}{2} \lambda_x^2 \right] A_r = 0.
\label{oseqnforar}
\end{eqnarray}
A cursory examination of the above ordinary differential equation shows that for $\epsilon_R = 0$, (\ref{oseqnforar}) reduces to the Orr-Sommerfeld equation for plane channel flow
that is known from literature, eg. \cite{schlichting}, \cite{stuart},
\cite{schmidhenningson}, \cite{criminalejacksonjoslin}. For $\epsilon_R \ne 0 $ corrections to the classical Orr-Sommerfeld equation arise and a surveillance of the correction terms shows that among these, two different
forms of dependence on the swirl parameter, $S$, are distinguishable. In one of these $\frac{1}{\sqrt{(1+S^2)}}$ appears as a coefficient, and this contribution is ascribable to the
transverse curvature that is inherent in the geometry considered. The other exhibits $\frac{S}{\sqrt{(1+S^2)}}$ as a coefficient, and this describes the effect of swirl. Terms
belonging to the former remain small over the entire range of values of $S$ from $0$ to $\infty$, whereas those belonging to the latter need not remain small when
$S \rightarrow \infty$. We will refer to these two as cases of small and large swirl respectively, and examine them more closely.
\subsection{The case of small swirl and the critical layer}
When the corrections to the equations are small, which is the case as long as the product $\epsilon_R S n_{\varphi}$ remains small, we may expect the departure of the solutions
from the case $S = 0$ also to remain small, as long as there are no changes in the other parameters, $Re, \epsilon_R$ and in the mode $n_{\varphi}$. This is tantamount to the expectation
that the neutral stability curves in the $(Re,\lambda_x)$-plane in the $(Re, S, \epsilon_R; \lambda_x, n_{\varphi})$-space bear similarity to each other, a feature that would
facilitate computational determination of the neutrally stable surface and the critical parameters in the multi-parameter space of $Re, S$ and $\epsilon_R$. We may note that the
equation (\ref{oseqnforar}) that is necessary to be solved to reach this objective, is an eigenvalue problem in which $A_r$ is the only unknown, without any coupling with
$A_{\varphi}$ or $A_x$. Numerical methods for solving the Orr-Sommerfeld equation in the more classical problems of hydrodynamic stability that have been developed, tested, and
are available in published literature, see eg. \cite{schmidhenningson}, \cite{criminalejacksonjoslin}, may then be expected to be applicable for solving (\ref{oseqnforar}).
\subsubsection{The critical layer, its location and scaling properties}
The {\it Orr-Sommerfeld equation} for the case of mild swirl, (\ref{oseqnforar}), on comparison with the corresponding equation for the case of the basic flow with straight and
parallel streamlines, see eg. \cite{criminalejacksonjoslin}, exhibits a singularity of the same nature as in the classical case. It arises for neutrally stable disturbances in the
inviscid counterpart of the problem presently under study when the Reynolds number is large. This observation leads to the expression for the location of the {\it critical layer}
in the flow in question. Analogous to the classical case, this singularity is removable in the complete problem through introduction of a thin {\it critical layer} of appropriate
thickness in which viscous effects are retained. In the classical case of a basic flow with straight and parallel streamlines this is the well-known {\it Tollmien's critical layer},
and its thickness is $O(Re^{-\frac{1}{3}})$. Application of the same procedure yields the scaling behaviour of the {\it critical layer} for the flow in question.
\subsubsection*{Location of the critical layer}
Inspection of (\ref{oseqnforar}) shows that the coefficient of the second derivative of $A_r$ goes through zero for neutrally stable
disturbances when the Reynolds number $Re$ is not small, as it does in the classical case. The criterion for the {\it location of the critical layer}, $y = y_c$, may therefore be
written as follows:
\begin{equation}
1 - \frac{\lambda_x}{\omega} \frac{(1-y_c^2)}{\sqrt{(1+S^2)}}
- \epsilon_R \left( \frac{\lambda_x}{\omega} \frac{1}{\sqrt{(1+S^2)}} \frac{y_c(1-y_c^2)}{3} + \frac{n_{\varphi}}{\omega} \frac{S}{\sqrt{(1+S^2)}} \frac{(1-y_c)}{2} \right) = 0.
\label{critlayerloc}
\end{equation}
For $\epsilon_R = 0$ the above expression is seen to be identical with the known result, i. e. the critical layer is located at a wall distance where the basic flow velocity and
the phase speed of neutrally stable disturbances equal each other, see eg. \cite{criminalejacksonjoslin}. For $\epsilon_R \ne 0 $ the location
of the critical layer depends, besides on $S$, also upon the mode of the disturbance $n_{\varphi}$. It is only for toroidal modes, for which $n_{\varphi} = 0$, that (\ref{critlayerloc})
fulfills the condition of the local flow velocity and the phase velocity of the wave equalling each other, cf. expression for $U_{Gx}$ in the text following
(\ref{modalansatzforu}). For the
modes $n_{\varphi} \ne 0$ the terms subtracted from $1$ in the expression (\ref{critlayerloc}) may be interpreted as the {\it inner product} of the {\it basic flow velocity vector},
$(U_{Gx}, U_{G \varphi})$, and the {\it slowness vector} of the wave of the particular mode in question, $\frac{(\lambda_x, n_{\varphi})}{\omega}$, a concept introduced by
Whitham in \cite{whitham}.
\subsubsection*{The order of magnitude of the critical layer thickness}
An estimate of the order of magnitude of the critical layer thickness is obtainable by application of the same procedure as described in literature, eg. \cite{criminalejacksonjoslin}), to the
present problem. To this end we introduce in the critical layer the independent variable $\eta_c$ defined through $\eta_c = (y-y_c) Re^m_c$, write (\ref{oseqnforar}) in the
neighbourhood of the critical layer and determine $m_c$ such that the coefficients of $\frac{d^4 A_r}{dy^4}$ and of $\frac{d^2 A_r}{dy^2}$ in (\ref{oseqnforar}) can stand in balance
with each other. Carrying out this
procedure for the present problem shows that a distinction has to be made between the {\bf modes of the disturbance}, i.e. between disturbances that are {\it toroidal} for which
$n_{\varphi}= 0$, and those that are {\it helical} for which $n_{\varphi} \ne 0$. For {\it toroidal} disturbances one gets the same result as in the conventional case which is
$m_c = -\frac{1}{3}$. This result for $m_c$ remains applicable to {\it helical disturbances}, $n_{\varphi} \ne 0$, too, only as long as the product $\epsilon_R S n_{\varphi}$
remains numerically small. The equation governing $A_r$ in the {\it critical layer} is then as follows:
\begin{eqnarray}
\frac{1}{\imath \omega} \left(\frac{d^4 A_r}{d\eta_c^4} \right) + \left(1+ \frac{\lambda_x}{\omega} \frac{1}{\sqrt{(1+S^2)}}(2y_c \eta_c) \right) \frac{d^2 A_r}{d\eta_c^2} \nonumber \\
+ \epsilon_R \left[ \left( \frac{\lambda_x}{\omega} \frac{1}{\sqrt{(1+S^2)}} \frac{(2y_c \eta_c)}{3} + \frac{n_{\varphi}}{\omega} \frac{S}{\sqrt{(1+S^2)}} \frac{\eta_c}{2} \right) \frac{d^2 A_r}{d\eta_c^2} \right]
\label{oseqnforarincritlayer}
\end{eqnarray}
\subsection{The case of large swirl}
It was seen in the previous subsection that the {\it generalised Orr-Sommerfeld and Squire equations}, (\ref{generalisedossqcont}) or (\ref{generalisedorrsommerfeld}, \ref{generalisedsquire}),
degenerate to their classical counterparts in plane channel flow, with corrections due to transverse curvature, as the {\it swirl parameter}, $S$, goes to $0$. In contrast,
in the other limit $S \rightarrow \infty $, a close examination of the same set of equations would show that the equations ( \ref{generalisedorrsommerfeld}, \ref{generalisedsquire})
then do not degenerate to those derived by Taylor for his stability problem. The noteworthy aspect of this failure is that proceeding to the limit $S \rightarrow \infty$ in a
straightforward manner does not result in an eigenvalue problem with a mutual coupling of $A_r$ and $A_{\varphi}$ deciding its stability, as it is the case in Taylor's treatment of this
stability problem. However, this shortcoming may be overcome through rewriting the equations (\ref{generalisedorrsommerfeld}, \ref{generalisedsquire}) in {\it Taylor variables},
$(\hat t, \hat y, \hat \varphi, \hat u_r, \hat u_{\varphi}, \hat u_x)$, defined through
\begin{equation}
\hat t = \frac{t}{Re}; \hat y = y; \hat \varphi = \varphi; \hat x = x; \hat u_r = u_r \frac{Re}{\epsilon_R}; \hat u_{\varphi} = u_{\varphi}; \hat u_x = u_x \frac{Re}{\epsilon_R} ,
\label{defntaylorvariables}
\end{equation}
before proceeding to the limit $S \rightarrow \infty$. One then gets the {\it generalised Orr-Sommerfeld, Squire and Continuity equations in Taylor variables} which we write as
\begin{equation}
\hat D \mathbf {\hat u} = \mathbf 0,
\label{generalisedossqcontintaylorvariables}
\end{equation}
where the expressions for the elements of the operator $ \hat D$ in (\ref{generalisedossqcontintaylorvariables}) follow from a comparison of this equation with its detailed breakdown
given below.
\\
{\bf The generalised Orr-Sommerfeld equation for large swirl}
\begin{eqnarray}
\frac{\partial}{\partial \hat t} \left[ \left(-\frac{\partial^2 \hat u_r}{\partial \hat y^2} - \frac{\partial^2 \hat u_r}{\partial \hat x^2} \right) - \epsilon_R \frac{\partial \hat u_r}{\partial \hat y} + Re \left(\frac{\partial^2 \hat u_{\varphi}}{\partial \hat y \partial \hat \varphi} - \frac{\partial \hat u_{\varphi}}{\partial \hat \varphi}\right) \right] \nonumber \\
+ \frac{U_{Gx}}{\sqrt{(1+S^2)}} Re \left[ \left(- \frac{\partial^3 \hat u_r}{\partial \hat y^2 \partial \hat x} - \frac{\partial^3 \hat u_r}{\partial \hat x^3} \right) + \epsilon_R \left( \frac{\partial ^3 \hat u_r}{\partial \hat y^2 \partial \hat \varphi} - \frac{\partial ^2 \hat u_r}{\partial \hat y \partial \hat x}\right) + Re \left( \frac{\partial ^2 \hat u_{\varphi}}{\partial \hat y \partial \hat \varphi} \right) \right] \nonumber \\
+ \frac{1}{\sqrt{(1+S^2)}} \frac{d U_{Gx}}{dy} \left[ \epsilon_R Re \frac{\partial \hat u_r}{\partial \hat x} + Re^2 \frac{\partial ^2 \hat u_{\varphi}}{\partial \hat \varphi \partial \hat x} \right] + \frac{1}{\sqrt{(1+S^2)}} \frac{d^2 U_{Gx}}{dy^2} Re \frac{\partial \hat u_r}{\partial \hat x} \nonumber \\
+ \frac{S }{\sqrt{(1+S^2)}} U_{G \varphi} \left[\epsilon_R Re \left(- \frac{\partial ^3 \hat u_r}{\partial \hat y^2 \partial \hat \varphi} - \frac{\partial ^3 \hat u_r}{\partial \hat x^2 \partial \varphi} \right) + 2 Re^2 \frac{\partial ^2 \hat u_{\varphi}}{\partial \hat x^2} \right] \nonumber \\
+ \frac{S \epsilon_R}{\sqrt{(1+S^2)}} \frac{d U_{G \varphi}}{dy} Re \left[ \left( \frac{\partial ^2 \hat u_r}{\partial \hat y \partial \hat \varphi} + \frac{\partial ^2 \hat u_x}{\partial \hat x \partial \hat \varphi} \right) \right] + \frac{S \epsilon_R}{\sqrt{(1+S^2)}} \frac{d^2 U_{G \varphi}}{d y^2} \left[Re \frac{\partial \hat u_r}{\partial \hat \varphi} \right] \nonumber \\
+ \left[ \left( \frac{\partial^4 \hat u_r }{\partial \hat x^4} + \frac{\partial^4 \hat u_r}{\partial \hat y^4} + 2 \frac{\partial^4 \hat u_r}{\partial \hat x^2 \partial \hat y^2} \right) + \epsilon_R Re \cdot 2 \frac{\partial^3 \hat u_r}{\partial \hat y \partial \hat x^2} + Re \cdot 2 \frac{\partial ^3 \hat u_{\varphi}}{\partial \hat y^3} \right] = 0 .
\label{generalisedorrsommerfeldintaylorvariables3}
\end{eqnarray}
\\
{\bf The generalised Squire equation for large swirl}
\begin{eqnarray}
\frac{\partial^2 \hat u_{\varphi}}{ \partial \hat t \partial \hat x} + \frac{1}{\sqrt{(1+S^2)}} U_{Gx} \left[Re \frac{\partial^2 \hat u_{\varphi}}{\partial \hat x^2} \right] + \frac{1}{\sqrt{(1+S^2)}} \frac{dU_{Gx}}{dy} \left[ \epsilon_R \frac{\partial \hat u_r}{\partial \hat x} \right] \nonumber \\
+ \frac{S}{\sqrt{(1+S^2)}} U_{G \varphi} \left[ \epsilon_R Re \frac{\partial ^2 \hat u_{\varphi}}{\partial \hat \varphi \partial \hat x} \right]
+ \left[\frac{\partial^3 \hat u_{\varphi}}{\partial \hat x^3} + \frac{\partial^3 \hat u_{\varphi}}{\partial \hat x \partial \hat y^2} \right] + \epsilon_R \frac{\partial ^2 \hat u_{\varphi}}{\partial \hat x \partial \hat y} = 0.
\label{generalisedsquireintaylorvariables3}
\end{eqnarray}
{\bf The Continuity equation for large swirl}
\begin{eqnarray}
\frac{\partial \hat u_r}{\partial \hat y}+ \epsilon_R \hat u_r + Re \frac{\partial \hat u_{\varphi}}{\partial \hat \varphi} + \frac{\partial \hat u_x}{\partial \hat x} = 0.
\label{generalisedcontinuityintaylorvariables3}
\end{eqnarray}
It may be verified that, proceeding now to the limit $S \rightarrow \infty$, the above set of equations,
(\ref{generalisedorrsommerfeldintaylorvariables3}, \ref{generalisedsquireintaylorvariables3}, \ref{generalisedcontinuityintaylorvariables3})
get reduced for toroidal disturbances, $\frac{\partial}{\partial \varphi} = 0$, to the equations set up by \cite{taylorscientificpapers} for the classical problem of stability of the
flow in the gap between rotating cylinders. It may be noted in this context that $Re^2$ may be replaced by the Taylor number $Ta$, and $Re$ by $\sqrt{Ta}$ since $Re^2 = Ta$, see eg.
\cite{criminalejacksonjoslin}.
\section{Discussion, Conclusions and Outlook}
The set of {\it generalised Orr-Sommerfeld, Squire and Continuity} equations derived, (\ref{generalisedossqcont}) or (\ref{generalisedorrsommerfeld}, \ref{generalisedsquire}, \ref{generalisedcontinuity})
with homogeneous boundary conditions for $\mathbf A$, is an eigenvalue problem for the complex frequency $\omega = \omega_r + \imath \omega_i$ that depends upon the three parameters, $(Re, S, \epsilon_R)$ and the mode of the
disturbance, $n_{\varphi}$. In the present state of the art this eigenvalue problem is not tractable through analytical means that extract the nature of the dependence of the
eigenvalues on the parameters, making it inevitable to obtain this information only numerically. The task then remaining is to extract the {\it globally critical Reynolds number}
out of the {\it neutrally stable surfaces} defined through $\omega _i (Re, S, \epsilon_R) = 0$ over the entire range of possible modes $n_{\varphi}$. The computational task involved
is too voluminous to be accomodated within the present short paper. It is therefore meaningful to
focus attention in this paper on such inferences on the characteristics of transition that can be drawn directly from
the equations themselves.
A point that seems to have been overlooked in published literature up to now is that in both the limiting cases, which belong to classical problems of flow transition,
the set of known and well established equations governing small-amplitude disturbance propagation may be understood as approximations of one and the same set of equations.
These are obtainable from the set of {\it generalised Orr-Sommerfeld, Squire and Continuity} for basic flows with helical streamlines viewed in the revised parameter space. The
well known equations for the particular cases of transition in plane channel flow and for {\it Taylor instability} in the flow in the gap between concentric circular cylinders
with rotation of the inner cylinder are both recoverable from the {\it generalised Orr-Sommerfeld, Squire and Continuity equations} on setting the {\it Swirl parameter}, $S$, to
either $0$ or $\infty$ respectively.
A feature of interest brought out by the set of {\it generalised Orr-Sommerfeld, Squire and Continuity equations} themselves follows from a cursory inspection of the coefficient
of the second derivative of $A_r$ in (\ref{reorderedgeneralisedossqcont}), analogous to the case of the basic flow with straight and parallel streamlines. A {\it critical layer}
may then arise when transition occurs at a sufficiently high Reynolds number and the coefficient of the second derivative of $A_r$ passes through zero. It is worth noting in this
context that it is the Reynolds number based on $U_{ref}$ that is decisive, and this depends upon both the characteristic velocities of the problem on hand, {\it viz.} $U_{refx}$
and $U_{ref \varphi}$. It follows herefrom that the {\it location of the critical layer} is governed not merely by the swirl and geometric parameters, $S$ and $\epsilon_R$, but
also by the {\it mode of the disturbance}, describable through $n_{\varphi}$. Furthermore, the {\it thickness of the critical layer} scales with $Re^{-\frac{1}{3}}$ a result that,
although formally is the same as in the classical case of no swirl, expresses dependence on swirl since $Re$ is formed with $U_{ref}$ which in turn depends upon both the
characteristic velocities $U_{refx}$ and $U_{ref \varphi}$.
The question that then arises, as to whether the {\it transition initiating disturbance of the spiral flow at given values of $S$ and $\epsilon_R$} is {\it toroidal} or
{\it helical}, remains open at this stage. Answering this question calls for a comparison of the {\it local critical Reynolds numbers}, which are neutrally stable points in the
($S, \epsilon_R$)-space at which $\frac{dRe}{d \lambda_x} = 0$, over the range of interest of $S$ and $\epsilon_R$ for different modes of disturbance $n_{\varphi}$. Such a
comparison of the {\it local critical Reynolds numbers} calls first for obtaining values for the {\it local critical Reynolds numbers} over a wide range of values of
($S, \epsilon_R$) for different modes, $n_\varphi$, which, in the present state of the art, can be done only numerically.
The finding from the present analysis may be summaised as follows: \\ The conclusion that the {\it critical layer}, when it arises, is located at a wall distance where the axial
flow velocity and the disturbance wave speed equal each other holds only when the transition initiating disturbance is toroidal in nature, not helical. Whether the
{\it transition initiating disturbance} is toroidal $(n_{\varphi}=0)$ or helical and with which value of $n_{\varphi} \ne 0$, would depend upon the {\it Swirl parameter}, $S$, and
the ratio of the gap width to the mean radius, $\epsilon_R$, and this requires numerically solving the eigenvalue problem and analysing the solutions obtained.
The author gratefully acknowledges the benefit of discussions with Professors Roddam Narasimha, FRS, K. R. Sreenivasan and Jeanette Hussong on the subject of this paper.
\section*{Appendix: The generalised Orr-Sommerfeld, Squire and Continuity equations in fully developed annular spiral flows}
{\bf The generalised Orr-Sommerfeld equation}
\begin{eqnarray}
\frac{\partial}{\partial t} \left[-\frac{\partial^2 u_r}{\partial r^2} - \frac{\partial^2 u_r}{\partial x^2} - \frac{1}{r} \frac{\partial u_r}{\partial r} + \frac{1}{r} \frac{\partial^2 u_{\varphi}}{\partial r \partial \varphi} - \frac{1}{r} \frac{\partial u_{\varphi}}{\partial \varphi} + O(r^{-2}) \right] \nonumber \\
+ \frac{1}{\sqrt{(1+S^2)}} U_{Gx}\left[\frac{1}{r}\frac{\partial^3 u_r}{\partial r^2 \partial \varphi } - \frac{\partial^3 u_r}{\partial r^2 \partial x} - \frac{1}{r} \frac{\partial^2 u_r}{\partial r \partial x} - \frac{1}{r} \frac{\partial^2 u_{\varphi}}{\partial r \partial \varphi} - \frac{\partial^3 u_r}{\partial x^3} + O(r^{-2}) \right] \nonumber \\
+ \frac{1}{\sqrt{(1+S^2)}} \frac{d U_{Gx}}{dr} \left[ \frac{1}{r} \frac{\partial u_r}{\partial x} + \frac{1}{r} \frac{\partial^2 u_{\varphi}}{\partial \varphi \partial x} + O(r^{-2}) \right] + \frac{1}{\sqrt{(1+S^2)}} \frac{d^2 U_{Gx}}{dr^2} \frac{\partial u_r}{\partial x} \nonumber \\
+ \frac{S}{\sqrt{(1+S^2)}} U_{G \varphi} \left[-\frac{1}{r} \frac{\partial ^3 u_r}{\partial r^2 \partial \varphi} - \frac{1}{r} \frac{\partial ^3 u_r}{\partial x^2 \partial \varphi} + \frac{2}{r} \frac{\partial ^2 u_\varphi}{\partial x^2} + O(r^{-2}) \right] \nonumber \\
+ \frac{S}{\sqrt{(1+S^2)}} \frac{d U_{G \varphi}}{dr} \left[\frac{1}{r} \frac{\partial^2 u_r}{\partial r \partial \varphi} + \frac{1}{r} \frac{\partial^2 u_x}{\partial x \partial \varphi} + O(r^{-2}) \right] + \frac{S}{\sqrt{(1+S^2)}} \frac{d^2 U_{G \varphi}}{d r^2} \left[ \frac{1}{r} \frac{\partial u_r}{\partial \varphi} \right] \nonumber \\
+ \frac{1}{Re} \left[\frac{\partial^4 u_r }{\partial x^4} + \frac{\partial^4 u_r}{\partial r^4} + 2 \frac{\partial^4 u_r}{\partial x^2 \partial r^2} + \frac{2}{r} \frac{\partial^3 u_r}{\partial r \partial x^2} + \frac{2}{r}\frac{\partial^3 u_{\varphi} }{\partial r^3} \right] = 0 .
\label{generalisedorrsommerfeld}
\end{eqnarray}
{\bf The generalised Squire equation}
\begin{eqnarray}
\frac{\partial}{\partial t} \left[\frac{\partial u_{\varphi}}{\partial x} - \frac{1}{r} \frac{\partial u_x}{\partial \varphi} \right] + \frac{1}{\sqrt{(1+S^2)}} U_{Gx} \left[\frac{\partial^2 u_{\varphi}}{\partial x^2} + \frac{1}{r} \frac{\partial u_r}{\partial r} + O(r^{-2}) \right] \nonumber \\
+\frac{1}{\sqrt{(1+S^2)}} \frac{dU_{Gx}}{dr} \left[\frac{\partial u_r}{\partial x} \right] \nonumber
+ \frac{S}{\sqrt{(1+S^2)}} U_{G \varphi} \left[ \frac{1}{r} \frac{\partial ^2 u_{\varphi}}{\partial \varphi \partial x} - \frac{1}{r} \frac{\partial ^2 u_x}{\partial \varphi ^2} + \frac{1}{r} \frac{\partial u_r}{\partial x} \right] \nonumber \\
- \frac{S}{\sqrt{(1+S^2)}} \frac{d U_{G \varphi}}{dr} \left[\frac{1}{r} \frac{\partial u_r}{\partial \varphi} \right] \nonumber
+\frac{1}{Re} \left[\frac{\partial^3 u_{\varphi}}{\partial x^3} + \frac{\partial^3 u_{\varphi}}{\partial x \partial r^2} + \frac{1}{r}\frac{\partial^2 u_{\varphi}}{\partial x \partial r} - \frac{1}{r} \frac{\partial^3 u_x}{\partial x^2 \partial \varphi}-\frac{1}{r}\frac{\partial^3 u_x}{\partial r^2 \partial \varphi} \right]=0. \nonumber \\
\label{generalisedsquire}
\end{eqnarray}
{\bf The Continuity equation}
\begin{eqnarray}
\frac{1}{r} \frac{\partial (r u_r)}{\partial r} + \frac{1}{r} \frac{\partial u_{\varphi}}{\partial \varphi} + \frac{\partial u_x}{\partial x} = 0.
\label{generalisedcontinuity}
\end{eqnarray}
| 26dbe822f3944cb0813042bf385cf1b49c52f62d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
It is well-known from the fundamental paper of It\^o \cite{Ito42} on the real-valued case, and several works \cite{AlRud05,App07,RieGaa,Dett82,Baum15} on the vector-valued case, that for any Banach space~$X$, any centered $X$-valued L\'evy process has a unique decomposition \mbox{$L = W +\widetilde N$}, where $W$ is an $X$-valued Wiener process, and $\widetilde N$ is an $X$-valued weak integral with respect to a certain compensated Poisson random measure. Moreover, $W$ and $\widetilde N$ are independent, and therefore since $W$ is symmetric, for each $1<p<\infty$ and $t\geq 0$,
\begin{equation}\label{eq:LevyItop-norm}
\mathbb E \|\widetilde N_t\|^p \leq \mathbb E \|L_t\|^p.
\end{equation}
The natural generalization of this result to general martingales in the real-valued setting was provided by Meyer in \cite{Mey76} and Yoeurp in \cite{Yoe76}. Namely, it was shown that any real-valued martingale $M$ can be uniquely decomposed into a sum of two martingales $M^d$ and $M^c$ such that $M^d$ is purely discontinuous (i.e.\ the quadratic variation $[M^d]$ has a pure jump version), and $M^c$ is continuous with $M^c_0=0$. The reason why they needed such a decomposition is a further decomposition of a semimartingale, and finding an exponent of a semimartingale (we refer the reader to \cite{Kal} and \cite{Yoe76} for the details on this approach). In the present article we extend Meyer-Yoeurp theorem to the vector-valued setting, and provide extension of \eqref{eq:LevyItop-norm} for a~general martingale (see Subsection \ref{subsec:MeyYoudec}). Namely, we prove that for any UMD Banach space $X$ and any $1<p<\infty$, an $X$-valued $L^p$-martingale $M$ can be uniquely decomposed into a~sum of two martingales $M^d$ and $M^c$ such that $M^d$ is purely discontinuous (i.e.\ $\langle M^d, x^*\rangle$ is purely discontinuous for each $x^* \in X^*$), and $M^c$ is continuous with $M^c_0=0$. Moreover, then for each $t\geq 0$,
\begin{equation}\label{eq:MeyYoeIntro}
(\mathbb E \|M^d_t\|^p)^{\frac 1p} \leq \beta_{p, X}(\mathbb E \|M_t\|^p)^{\frac 1p},\;\;\;\;\;
(\mathbb E \|M^c_t\|^p)^{\frac 1p} \leq \beta_{p, X}(\mathbb E \|M_t\|^p)^{\frac 1p},
\end{equation}
where $\beta_{p, X}$ is the UMD$_p$ constant of $X$ (see Subsection \ref{subsec:UMD}). Theorem \ref{thm:exampleforlowboundofA^qA^a} shows that such a decomposition together with $L^p$-estimates of type \eqref{eq:MeyYoeIntro} is possible if and only if $X$ has the UMD property.
The purely discontinuous part can be further decomposed: in \cite{Yoe76} Yoeurp proved that any real-valued purely discontinuous $M^d$ can be uniquely decomposed into a sum of a purely discontinuous quasi-left continuous martingale~$M^q$ (analogous to the ``compensated Poisson part'', which does not jump at predictable stopping times), and a purely discontinuous martingale with accessible jumps $M^a$ (analogous to the ``discrete part'', which jumps only at certain predictable stopping times). In Subsection \ref{subsec:Youdec} we extend this result to a UMD space-valued setting with appropriate estimates. Namely, we prove that for each $1<p<\infty$ the same type of decomposition is possible and unique for an $X$-valued purely discontinuous $L^p$-martingale $M^d$, and then for each $t\geq 0$,
\begin{equation}\label{eq:YouIntro}
(\mathbb E \|M^q_t\|^p)^{\frac 1p} \leq \beta_{p, X}(\mathbb E \|M^d_t\|^p)^{\frac 1p},\;\;\;\;\;
(\mathbb E \|M^a_t\|^p)^{\frac 1p}\leq \beta_{p, X}(\mathbb E \|M^d_t\|^p)^{\frac 1p}.
\end{equation}
Again as Theorem \ref{thm:exampleforlowboundofA^qA^a} shows, the \eqref{eq:YouIntro}-type estimates are a possible only in UMD Banach spaces.
\smallskip
Even though the Meyer-Yoeurp and Yoeurp decompositions can be easily extended from the real-valued case to a Hilbert space case, the author could not find the corresponding estimates of type \eqref{eq:MeyYoeIntro}-\eqref{eq:YouIntro} in the literature, so we wish to present this special issue here. If $H$ is a Hilbert space, $M:\mathbb R_+ \times \Omega \to H$ is a martingale, then there exists a unique decomposition of $M$ into a continuous part $M^c$, a purely discontinuous quasi-left continuous part $M^q$, and a purely discontinuous part $M^a$ with accessible jumps. Moreover, then for each $1<p<\infty$, and for $i=c,q,a$,
\begin{equation}\label{eq:Hvaluedcasintro}
(\mathbb E \|M^i_t\|^p)^{\frac 1p} \leq (p^*-1)(\mathbb E \|M_t\|^p)^{\frac 1p},
\end{equation}
where $p^* = \max\{p, \frac{p}{p-1}\}$. Notice that though \eqref{eq:Hvaluedcasintro} follows from \eqref{eq:MeyYoeIntro}-\eqref{eq:YouIntro} since $\beta_{p,H}=p^*-1$, it can be easily derived from the differential subordination estimates for Hilbert space-valued martingales obtained by Wang in~\cite{Wang}.
\smallskip
Both the Meyer-Yoeurp and Yoeurp decompositions play a significant r\^ole in stochastic integration: if $M=M^c + M^q + M^a$ is a decomposition of an $H$-valued martingale $M$ into continuous, purely discontinuous quasi-left continuous and purely discontinuous with accessible jumps parts, and if $\Phi:\mathbb R_+ \times \Omega \to \mathcal L(H, X)$ is elementary predictable for some UMD Banach space $X$, then the decomposition $\Phi \cdot M=\Phi \cdot M^c + \Phi \cdot M^q + \Phi \cdot M^a$ of a stochastic integral $\Phi \cdot M$ is a decomposition of the martingale $\Phi \cdot M$ into continuous, purely discontinuous quasi-left continuous and purely discontinuous with accessible jumps parts, and for any $1<p<\infty$ we have that
\[
\mathbb E \|(\Phi \cdot M)_{\infty}\|^p \eqsim_{p,X}\mathbb E \|(\Phi \cdot M^c)_{\infty}\|^p + \mathbb E \|(\Phi \cdot M^q)_{\infty}\|^p + \mathbb E \|(\Phi \cdot M^a)_{\infty}\|^p.
\]
The corresponding It\^o isomorphism for $\Phi \cdot M^c$ for a general UMD Banach space $X$ was derived by Veraar and the author in \cite{VY2016}, while It\^o isomorphisms for $\Phi \cdot M^q$ and $\Phi \cdot M^a$ have been shown by Dirksen and the author in \cite{DY17} for the case $X = L^r(S)$, $1<r<\infty$.
\smallskip
The major underlying techniques involved in the proofs of \eqref{eq:MeyYoeIntro} and \eqref{eq:YouIntro} are rather different from the original methods of Meyer in \cite{Mey76} and Yoeurp in \cite{Yoe76}. They include the results on the differentiability of the Burkholder function of any finite dimensional Banach space, which have been proven recently in \cite{Y17FourUMD} and which allow us to use It\^o formula in order to show the desired inequalities in the same way as it was demonstrated by Wang in \cite{Wang}.
\smallskip
The main application of the Meyer-Yoeurp decomposition are $L^p$-estimates for weakly differentially subordinated martingales. The weak differential subordination property was introduced by the author in \cite{Y17FourUMD}, and can be described in the following way: an $X$-valued martingale $N$ is weakly differentially subordinated to an $X$-valued martingale $M$ if for each $x^* \in X^*$ a.s.\ $|\langle N_0, x^*\rangle| \leq |\langle M_0, x^*\rangle|$ and for each $t\geq s\geq 0$
\[
[\langle N, x^*\rangle]_t - [\langle N, x^*\rangle]_s \leq [\langle M, x^*\rangle]_t - [\langle M, x^*\rangle]_s.
\]
If both $M$ and $N$ are purely discontinuous, and if $X$ is a UMD Banach space, then by \cite{Y17FourUMD}, for each $1<p<\infty$ we have that $\mathbb E \|N_{\infty}\|^p \leq \beta_{p, X}^p\mathbb E \|M_{\infty}\|^p$. Section \ref{sec:weakdifsubandgenmart} is devoted to the generalization of this result to continuous and general martingales. There we show that if both $M$ and $N$ are continuous, then $\mathbb E \|N_{\infty}\|^p \leq c_{p,X}^p\mathbb E \|M_{\infty}\|^p$, where the least admissible $c_{p,X}$ is within the interval $[\beta_{p, X}, \beta_{p, X}^2]$. Furthermore, using the Meyer-Yoeurp decomposition and estimates \eqref{eq:MeyYoeIntro} we show that for general \mbox{$X$-va}lu\-ed martingales $M$ and $N$ such that $N$ is weakly differentially subordinated to $M$ the following holds
\[
(\mathbb E \|N_{\infty}\|^p)^{\frac 1p} \leq \beta_{p, X}^2(\beta_{p, X}+1) (\mathbb E \|M_{\infty}\|^p)^{\frac 1p}.
\]
\smallskip
The weak differential subordination as a stronger version of the differential subordination is of interest in Harmonic Analysis. For instance, it was shown in \cite{Y17FourUMD} that sharp $L^p$-estimates for weakly differentially subordinated purely discontinuous martingales imply sharp estimates for the norms of a broad class of Fourier multipliers on $L^p(\mathbb R^d;X)$. Also there is a strong connection between the weak differential subordination of continuous martingales and the norm of the~Hilbert transform on $L^p(\mathbb R; X)$ (see \cite{Y17FourUMD} and Remark~\ref{rem:Hiltranfweakdiffsubcont}).
Alternative approaches to Fourier multipliers for functions with values in UMD spaces have been constructed from the differential subordination for purely discontinuous martingales (see Ba\~{n}uelos and Bogdan \cite{BB}, Ba\~{n}uelos, Bogdan and Bie\-la\-szew\-ski \cite{BBB}, and recent work \cite{Y17FourUMD}), and for continuous martingales (see McConnell \cite{McC84} and Geiss, Montgomery-Smith and Saksman \cite{GM-SS}). It remains open whether one can combine these two approaches using the general weak differential subordination theory.
\section{Preliminaries}
In the sequel we will omit proofs of some statements marked with a star (e.g.\ Lemma$^*$, Theorem$^*$, etc.) Please find the corresponding proofs after the references or in the supplement \cite{Y17MDSupp}.
\smallskip
We set the scalar field to be $\mathbb R$. We will use the {\em Kronecker symbol} $\delta_{ij}$, which is defined in the following way: $\delta_{ij} =1$ if $i=j$, and $\delta_{ij} =0$ if $i\neq j$. For each $p\in (1,\infty)$ we set $p'\in (1,\infty)$ and $p^* \in [2,\infty)$ to be such that $\frac 1p + \frac 1{p'}=1$ and $p^* = \max\{p, p'\}$. We set $\mathbb R_+ := [0,\infty)$.
\subsection{UMD Banach spaces}\label{subsec:UMD}\nopagebreak
A Banach space $X$ is called a {\em UMD space} if for some (equivalently, for all)
$p \in (1,\infty)$ there exists a constant $\beta>0$ such that
for every $n \geq 1$, every martingale
difference sequence $(d_j)^n_{j=1}$ in $L^p(\Omega; X)$, and every $\{-1,1\}$-valued sequence
$(\varepsilon_j)^n_{j=1}$
we have
\[
\Bigl(\mathbb E \Bigl\| \sum^n_{j=1} \varepsilon_j d_j\Bigr\|^p\Bigr )^{\frac 1p}
\leq \beta \Bigl(\mathbb E \Bigl \| \sum^n_{j=1}d_j\Bigr\|^p\Bigr )^{\frac 1p}.
\]
The least admissible constant $\beta$ is denoted by $\beta_{p,X}$ and is called the {\em UMD constant}. It is well-known (see \cite[Chapter 4]{HNVW1}) that $\beta_{p, X}\geq p^*-1$ and that $\beta_{p, H} = p^*-1$ for a Hilbert space $H$.
We refer the reader to \cite{Burk01,HNVW1,Rubio86,Pis16} for details.
The following proposition is a vector-valued version of \cite[Theorem 4.1]{Choi92}.
\begin{proposition}\label{prop:UMD01sequence}
Let $X$ be a Banach space, $p\in (1,\infty)$. Then $X$ has the UMD property if and only if there exists $C>0$ such that for each $n\geq 1$, for every martingale
difference sequence $(d_j)^n_{j=1}$ in $L^p(\Omega; X)$, and every sequence
$(\varepsilon_j)^n_{j=1}$ such that $\varepsilon_j\in \{0,1\}$ for each $j=1,\ldots,n$
we have
\begin{align*}
\Bigl(\mathbb E \Bigl\| \sum^n_{j=1} \varepsilon_j d_j\Bigr\|^p\Bigr )^{\frac 1p}
\leq C \Bigl(\mathbb E \Bigl \| \sum^n_{j=1}d_j\Bigr\|^p\Bigr )^{\frac 1p}.
\end{align*}
If this is the case, then the least admissible $C$ is in the interval $[\frac {\beta_{p, X}-1}{2}, \beta_{p, X}]$
\end{proposition}
\subsection{Martingales and stopping times in continuous time}\label{subsec:prelimmart}
Let $(\Omega,\mathcal F, \mathbb P)$ be a probability space with a filtration $\mathbb F = (\mathcal F_t)_{t\geq 0}$ which satisfies the usual conditions. Then $\mathbb F$ is right-continuous, and the following proposition holds (see \cite{Y17FourUMD}):
\begin{proposition}\label{prop:cadlagvers}
Let $X$ be a Banach space. Then any martingale $M:\mathbb R_+ \times \Omega \to X$ has a {\em c\`adl\`ag} version
\end{proposition}
Let $1\leq p\leq \infty$. A martingale $M:\mathbb R_+ \times \Omega \to X$ is called an {\em $L^p$-martingale} if $M_t \in L^p(\Omega; X)$ for each $t\geq 0$, there exists an a.s.\ limit $M_{\infty} := \lim_{t\to \infty} M_t$, $M_{\infty} \in L^p(\Omega; X)$ and $M_t \to M_{\infty}$ in $L^p(\Omega; X)$ as $t\to \infty$. We will denote the space of all $X$-valued $L^p$-martingales on $\Omega$ by $\mathcal M_X^p(\Omega)$. For brevity we will use $\mathcal M_X^p$ instead. Notice that $\mathcal M_X^p$ is a Banach space with the given norm: $\|M\|_{\mathcal M_X^p} := \|M_{\infty}\|_{L^p(\Omega; X)}$ (see \cite{Kal,Jac79} and \cite[Chapter 1]{HNVW1}).
\begin{proposition}\label{prop:dualofMXp}
Let $X$ be a Banach space with the Radon-Nikod\'ym property (e.g.\ reflexive), $1<p<\infty$. Then $(\mathcal M_X^p)^* = \mathcal M_{X^*}^{p'}$, and $\|M\|_{(\mathcal M_X^p)^*} = \|M\|_{\mathcal M_{X^*}^{p'}}$ for each $M \in \mathcal M_{X^*}^{p'}$.
\end{proposition}
A random variable $\tau:\Omega \to \mathbb R_+$ is called an {\em optional stopping time} (or just a {\em stopping time}) if $\{\tau\leq t\} \in \mathcal F_t$ for each $t\geq 0$. With an optional stopping time $\tau$ we associate a $\sigma$-field $\mathcal F_{\tau} = \{A\in \mathcal F_{\infty}: A\cap \{\tau\leq t\}\in \mathcal F_{t}, t\in\mathbb R_+\}$. Note that $M_{\tau}$ is strongly $\mathcal F_{\tau}$-measurable for any local martingale $M$. We refer to \cite[Chapter 7]{Kal} for details.
Due to the existence of a c\`adl\`ag version of a martingale $M:\mathbb R_+ \times \Omega \to X$, we can define an $X$-valued random variables $M_{\tau-}$ and $\Delta M_{\tau}$ for any stopping time $\tau$ in the following way: $M_{\tau-} = \lim_{\varepsilon \to 0}M_{(\tau - \varepsilon)\vee 0}$, $\Delta M_{\tau} = M_{\tau} - M_{\tau-}$.
\subsection{Quadratic variation}
Let $(\Omega, \mathcal F, \mathbb P)$ be a probability space with a filtration $\mathbb F = (\mathcal F_t)_{t\geq 0}$ that
satisfies the usual conditions, $H$ be a Hilbert space. Let $M:\mathbb R_+ \times \Omega \to H$ be a local martingale. We define a {\em quadratic variation} of $M$ in the following way:
\begin{equation}\label{eq:defquadvar}
[M]_t := \mathbb P-\lim_{{\rm mesh}\to 0}\sum_{n=1}^N \|M(t_n)-M(t_{n-1})\|^2,
\end{equation}
where the limit in probability is taken over partitions $0= t_0 < \ldots < t_N = t$. Note that $[M]$ exists and is nondecreasing a.s. The reader can find more on quadratic variations in \cite{MetSemi,MP} for the vector-valued setting, and in \cite{Kal,Prot,MP} for the real-valued setting.
For any martingales $M, N:\mathbb R_+ \times \Omega \to H$ we can define a {\em covariation} $[M,N]:\mathbb R_+ \times \Omega \to \mathbb R$ as $[M,N] := \frac{1}{4}([M+N]-[M-N])$.
Since $M$ and $N$ have c\`adl\`ag versions, $[M,N]$ has a c\`adl\`ag version as well (see \cite[Theorem I.4.47]{JS} and \cite{MetSemi}).
\begin{remark}[\cite{MetSemi}]\label{rem:covariation}
The process $\langle M,N\rangle - [M,N]$ is a local martingale.
\end{remark}
\subsection{Continuous martingales}\label{subsec:prelimcontmart}
Let $X$ be a Banach space. A martingale $M:\mathbb R_+ \times \Omega \to X$ is called {\em continuous} if $M$ has continuous paths.
\begin{remark}[\cite{Kal,MP}]\label{rem:qvcont}
If $X$ is a Hilbert space, $M,N:\mathbb R_+ \times \Omega \to X$ are continuous martingales, then $[M, N]$ has a continuous version.
\end{remark}
Let $1\leq p\leq \infty$. We will denote the linear space of all continuous $X$-valued $L^p$-martingales on $\Omega$ which start at zero by $\mathcal M_X^{p,c}(\Omega)$. For brevity we will write $\mathcal M_X^{p,c}$ instead of $\mathcal M_X^{p,c}(\Omega)$ since $\Omega$ is fixed. Analogously to \cite[Lemma 17.4]{Kal} by applying Doob's maximal inequality \cite[Theorem 3.2.2]{HNVW1} one can show the following proposition.
\begin{proposition}\label{prop:M^p,ciscomplete}
Let $X$ be a Banach space, $p\in(1,\infty)$. Then $\mathcal M_X^{p,c}$ is a Banach space with the following norm: $\|M\|_{\mathcal M_X^{p,c}} := \|M_{\infty}\|_{L^p(\Omega; X)}$.
\end{proposition}
\subsection{Purely discontinuous martingales}\label{subsec:prelimpdmart}
An increasing c\`adl\`ag process $A:\mathbb R_+ \times \Omega \to \mathbb R$ is called {\em pure jump} if a.s.\ for each $t\geq 0$, $A_t = A_0 + \sum_{s=0}^t \Delta A_s$.
A local martingale $M:\mathbb R_+\times \Omega \to \mathbb R$ is called {\em purely discontinuous} if $[M]$ is a pure jump process.
The reader can find more on purely discontinuous martingales in \cite{JS,Kal}. We leave the following evident lemma without proof.
\begin{lemma}\label{lemma:A^dA^c}
Let $A:\mathbb R_+ \times \Omega \to \mathbb R_+$ be an increasing adapted c\`adl\`ag process such that $A_0 =0$. Then there exist unique up to indistinguishability increasing adapted c\`adl\`ag processes $A^c, A^d:\mathbb R_+ \times \Omega \to \mathbb R_+$ such that $A^c$ is continuous a.s., $A^d$ is pure jump a.s., $A^c_0 = A^d_0=0$ and $A= A^c + A^d$.
\end{lemma}
\begin{remark}\label{rem:YoeMeyR}
According to the works \cite{Mey76} by Meyer and \cite{Yoe76} by Yoeurp (see also \cite[Theorem~26.14]{Kal}), any martingale $M:\mathbb R_+ \times \Omega \to \mathbb R$ can be uniquely decomposed into a sum of a purely discontinuous local martingale $M^d$ and a continuous local martingale $M^c$ such that $M^c_0 =0$. Moreover, $[M]^c= [M^c]$ and $[M]^d = [M^d]$, where $[M]^c$ and $[M]^d$ are defined as in Lemma \ref{lemma:A^dA^c}.
\end{remark}
\begin{corollary}\label{cor:cont=purdiscR}
Let $M:\mathbb R_+ \times \Omega \to\mathbb R$ be a martingale which is both continuous and purely discontinuous. Then $M=M_0$ a.s.
\end{corollary}
\begin{proposition*}\label{thm:purdiscorthtoanycont1}
A martingale $M:\mathbb R_+\times \Omega \to \mathbb R$ is purely discontinuous if and only if $M N$ is a martingale for any continuous bounded martingale $N:\mathbb R_+ \times \Omega \to \mathbb R$ with $N_0 = 0$.
\end{proposition*}
Note that some authors take this equivalent condition as the definition of a purely discontinuous martingale, see e.g.\ \cite[Definition I.4.11]{JS} and \cite[Chapter I]{Jac79}.
\begin{definition}\label{def:purelydiscXvalued}
Let $X$ be a Banach space, $M:\mathbb R_+\times \Omega \to X$ be a local martingale. Then $M$ is called {\em purely discontinuous} if for each $x^* \in X^*$ the local martingale $\langle M, x^*\rangle$ is purely discontinuous.
\end{definition}
\begin{remark}\label{rem:YoeMey}
Let $X$ be finite dimensional. Then similarly to Remark \ref{rem:YoeMeyR} any martingale $M:\mathbb R_+ \times \Omega \to X$ can be uniquely decomposed into a sum of a purely discontinuous local martingale $M^d$ and a continuous local martingale $M^c$ such that $M^c_0 =0$.
\end{remark}
\begin{remark}\label{rem:purdiscorthtoanycont2}
Analogously to Proposition \ref{thm:purdiscorthtoanycont1}, a martingale $M:\mathbb R_+\times \Omega \to X$ is purely discontinuous if and only if $\langle M, x^*\rangle N$ is a martingale for any $x^* \in X^*$ and any continuous bounded martingale $N:\mathbb R_+ \times \Omega \to \mathbb R$ such that $N_0 = 0$.
\end{remark}
Let $p\in [1,\infty]$. We will denote the linear space of all purely discontinuous $X$-valued $L^p$-martingales on $\Omega$ by $\mathcal M_X^{p,d}(\Omega)$. Since $\Omega$ is fixed, we will use $\mathcal M_X^{p,d}$ instead.
The scalar case of the next result have been presented in \cite[Lemme I.2.12]{Jac79}.
\begin{proposition}\label{thm:M^p,discomplete}
Let $X$ be a Banach space, $p\in (1,\infty)$. Then $\mathcal M_X^{p, d}$ is a Banach space with a norm defined as follows: $\|M\|_{\mathcal M_X^{p, d}}:= \|M_{\infty}\|_{L^p(\Omega; X)}$.
\end{proposition}
\begin{proof}
Let $(M^n)_{n\geq 1}$ be a sequence of purely discontinuous $X$-valued $L^p$-martingales such that $(M^n_{\infty})_{n\geq 1}$ is a Cauchy sequence in $L^p(\Omega; X)$. Let $\xi\in L^p(\Omega; X)$ be such that $\lim_{n\to \infty}M^n_{\infty} =\xi$. Define a martingale $M:\mathbb R_+ \times \Omega \to X$ as follows: $M = (M_s)_{s\geq 0} = (\mathbb E(\xi|\mathcal F_s))_{s\geq 0}$. Let us show that $M \in \mathcal M_X^{p,d}$. First notice that $\|M_{\infty}\|_{L^p(\Omega; X)} = \|\xi\|_{L^p(\Omega; X)} < \infty$. Further for each $x^* \in X^*$ by \cite[Lemme I.2.12]{Jac79} we have that $\langle M, x^*\rangle$ as a limit of real-valued purely discontinuous martingales $(\langle M^n, x^*\rangle)_{n\geq 1}$ in $\mathcal M_{\mathbb R}^p$ is purely discontinuous. Therefore $M$ is purely discontinuous by the definition.
\end{proof}
\begin{lemma}\label{lemma:contpuredisczero}
Let $X$ be a Banach space, $M:\mathbb R_+ \times \Omega \to X$ be a martingale such that $M$ is both continuous and purely discontinuous. Then $M = M_0$ a.s.
\end{lemma}
\begin{proof}
Follows analogously Corollary \ref{cor:cont=purdiscR}.
\end{proof}
\subsection{Time-change}
A nondecreasing, right-continuous family of stopping times $\tau =(\tau_s)_{s\geq 0}$ is called
a \textit{random time-change}. If $\mathbb F$ is right-continuous, then
according to \cite[Lemma~7.3]{Kal} the same holds true
for the \textit{induced filtration} $\mathbb G = (\mathcal G_s)_{s \geq 0} = (\mathcal F_{\tau_s})_{s\geq 0}$
(see more in \cite[Chapter~7]{Kal}).
Let $X$ be a Banach space. A~martingale $M:\mathbb R_+ \times \Omega \to X$ is said to be \textit{$\tau$-continuous} if $M$ is an a.s.\ constant on every interval $[\tau_{s-}, \tau_s]$, $s \geq 0$, where we let $\tau_{0-} = 0$.
\begin{theorem*}\label{thm:apptimechange}
Let $A:\mathbb R_+ \times \Omega\to \mathbb R_+$ be a strictly increasing continuous predictable process such that $A_0 = 0$ and $A_{t} \to \infty$ as $t\to \infty$ a.s. Let $\tau = (\tau_s)_{s\geq 0}$ be a random time-change defined as
$\tau_s := \{t: A_t=s\}$, $s\geq 0$.
Then $(A\circ \tau)(t) = (\tau \circ A)(t) = t$ a.s.\ for each $t\geq 0$. Let $\mathbb G = (\mathcal G_s)_{s \geq 0} = (\mathcal F_{\tau_s})_{s\geq 0}$ be the induced filtration. Then $(A_t)_{t\geq 0}$ is a random time-change with respect to $\mathbb G$ and for any $\mathbb F$-martingale $M:\mathbb R_+ \times \Omega \to\mathbb R$ the following holds
\begin{itemize}
\item [(i)] $M \circ \tau$ is a continuous $\mathbb G$-martingale if and only if $M$ is continuous, and
\item[(ii)]$M \circ \tau$ is a purely discontinuous $\mathbb G$-martingale if and only if $M$ is purely discontinuous.
\end{itemize}
\end{theorem*}
\subsection{Stochastic integration}
Let $X$ be a Banach space, $H$ be a Hilbert space. For each $h\in H$, $x\in X$ we denote the linear operator $g\mapsto \langle g, h\rangle x$, $g\in H$, by $h\otimes x$. The process $\Phi: \mathbb R_+ \times \Omega \to \mathcal L(H,X)$ is called \textit{elementary progressive}
with respect to the filtration $\mathbb F = (\mathcal F_t)_{t \geq 0}$ if it is of the form
\begin{equation}\label{eq:elprog}
\Phi(t,\omega) = \sum_{k=1}^K\sum_{m=1}^M \mathbf 1_{(t_{k-1},t_k]\times B_{mk}}(t,\omega)
\sum_{n=1}^N h_n \otimes x_{kmn},\;\;\; t\geq 0, \omega \in \Omega,
\end{equation}
where $0 \leq t_0 < \ldots < t_K <\infty$, for each $k = 1,\ldots, K$ the sets
$B_{1k},\ldots,B_{Mk}$ are in $\mathcal F_{t_{k-1}}$ and the vectors $h_1,\ldots,h_N$ are orthogonal.
Let $M:\mathbb R_+ \times \Omega \to H$ be a martingale. Then we define the {\em stochastic integral} $\Phi \cdot M:\mathbb R_+ \times \Omega \to X$ of $\Phi$ with respect to $M$ as follows:
\begin{equation}\label{eq:defofstochintwrtM}
(\Phi \cdot M)_t = \sum_{k=1}^K\sum_{m=1}^M \mathbf 1_{B_{mk}}
\sum_{n=1}^N \langle(M(t_k\wedge t)- M(t_{k-1}\wedge t)), h_n\rangle x_{kmn},\;\; t\geq 0.
\end{equation}
We will need the following lemma on stochastic integration (see \cite{Y17FourUMD}).
\begin{lemma}\label{lemma:stochintmoment}
Let $d$ be a natural number, $H$ be a $d$-dimensional Hilbert space, $p\in (1,\infty)$, $M, N:\mathbb R_+ \times \Omega \to H$ be $L^p$-martingales, $F:H \to H$ be a measurable function such that $\|F(h)\|\leq C \|h\|^{p-1}$ for each $h\in H$ and some $C>0$. Define $N_-:\mathbb R_+ \times \Omega \to H$ by $(N_-)_t = N_{t-}$, $t\geq 0$. Then $F(N_{-})\cdot M$ is a martingale and for each $t\geq 0$
\begin{equation}\label{eq:stochintmoment}
\mathbb E |(F(N_-)\cdot M)_t|\lesssim_{p, d} C(\mathbb E \|N_t\|^p)^{\frac {p-1}p} (\mathbb E \|M_t\|^p)^{\frac {1}p}.
\end{equation}
\end{lemma}
\subsection{Multidimensional Wiener process}
Let $d$ be a natural number. $W:\mathbb R_+ \times \Omega \to \mathbb R^d$ is called a {\em standard $d$-dimensional Wiener process} if $\langle W, h\rangle$ is a standard Wiener process for each $h\in \mathbb R^d$ such that $\|h\|=1$.
The following lemma is a multidimensional variation of \cite[(3.2.19)]{KS}.
\begin{lemma}\label{lemma:covstochintwrtcylbrmo}
Let $X = \mathbb R$, $d\geq 1$, $W$ be a standard $d$-dimensional Wiener process, $\Phi, \Psi:\mathbb R_+ \times \Omega \to \mathcal L(\mathbb R^d, \mathbb R)$ be elementary progressive. Then for all $t\geq 0$ a.s.
\[
[\Phi \cdot W, \Psi \cdot W]_t = \int_0^t \langle \Phi^*(s), \Psi^*(s) \rangle\ud s.
\]
\end{lemma}
The reader can find more on stochastic integration with respect to a Wiener process in the Hilbert space case in \cite{DPZ}, in the case of Banach spaces with a martingale type~2 in \cite{Brz1}, and in the UMD case in \cite{NVW}. Notice that the last mentioned work provides sharp $L^p$-estimates for stochastic integrals for the broadest till now known class of spaces.
\subsection{Brownian representation} The following theorem can be found in \cite[Theorem 3.4.2]{KS} (see also \cite{SV,Yar16Br}).
\begin{theorem}\label{thm:Brrepres}
Let $d\geq 1$, $M:\mathbb R_+ \times \Omega \to \mathbb R^d$ be a continuous martingale such that $[M]$ is a.s.\ absolutely continuous with respect to the Lebesgue measure on $\mathbb R_+$. Then there exist an enlarged probability space $(\widetilde{\Omega}, \widetilde {\mathcal F}, \widetilde{\mathbb P})$ with an enlarged filtration $\widetilde{\mathbb F} = (\widetilde F_t)_{t\geq 0}$, a $d$-dimensional standard Wiener process $W:\mathbb R_+ \times \widetilde{\Omega}\to \mathbb R^d$ which is defined on the filtration $\widetilde {\mathbb F}$, and an $\widetilde {\mathbb F}$-progressively measurable $\Phi :\mathbb R_+\times \widetilde{\Omega} \to \mathcal L(\mathbb R^d)$ such that $M = \Phi \cdot W$.
\end{theorem}
\subsection{Lebesgue measure}
Let $X$ be a finite dimensional Banach space. Then according to Theorem 2.20 and Proposition 2.21 in \cite{FolHarm} there exists a unique translation-invariant measure $\lambda_X$ on $X$ such that $\lambda_X(\mathbb B_X) = 1$ for the unit ball $\mathbb B_X$ of $X$. We will call $\lambda_X$ the {\em Lebesgue measure}.
\section{UMD Banach spaces and martingale decompositions}
Let $X$ be a Banach space, $1<p<\infty$. In this section we will show that the Meyer-Yoeurp and Yoeurp decompositions for $X$-valued $L^p$-martingales take place if and only if $X$~has the UMD property.
\subsection{Meyer-Yoeurp decomposition in UMD case}\label{subsec:MeyYoudec}
This subsection is devoted to the ge\-ne\-ra\-li\-za\-tion of Meyer-Yoeurp decomposition (see Remark \ref{rem:YoeMeyR}) to the UMD Banach space case:
\begin{theorem}[Meyer-Yoeurp decomposition]\label{thm:Meyer-Yoeurp}
Let $X$ be a UMD Banach space, $p\in (1,\infty)$, $M:\mathbb R_+ \times \Omega \to X$ be an $L^p$-martingale. Then there exist unique martingales $M^d, M^c:\mathbb R_+ \times \Omega \to X$ such that $M^d$ is purely discontinuous, $M^c$ is continuous, $M^c_0 = 0$ and $M=M^d + M^c$. Moreover, then for all $t\geq 0$
\begin{equation}\label{eq:Meyer-Yoeurp}
(\mathbb E\|M^d_t\|^p)^{\frac 1p} \leq \beta_{p,X}(\mathbb E \|M_t\|^p)^{\frac 1p},\;\;\;\;\;
(\mathbb E\|M^c_t\|^p)^{\frac 1p} \leq \beta_{p,X}(\mathbb E \|M_t\|^p)^{\frac 1p}.
\end{equation}
\end{theorem}
The proof of the theorem consists of several steps. First we introduce the main tool of our proof -- the Burkholder function.
\begin{definition}
Let $E$ be a linear space with a scalar field $\mathbb R$.
\begin{itemize}
\item[(i)] A function $f:E\times E \to \mathbb R$ is called {\em biconcave} if for each $x, y \in E$ one has that the mappings $e\mapsto f(x, e)$ and $e\mapsto f(e,y)$ are concave.
\item[(ii)] A function $f:E\times E \to \mathbb R$ is called {\em zigzag-concave} if for each $x, y\in E$ and $\varepsilon \in \mathbb R$ such that $|\varepsilon| \leq 1$, the function $z\mapsto f(x+z, y+\varepsilon z)$ is concave.
\end{itemize}
\end{definition}
The following theorem is a small variation of \cite{Burk86} and \cite[Theorem 4.5.6]{HNVW1}, and has been proven in \cite{Y17FourUMD}.
\begin{theorem}[Burkholder]\label{thm:Burkholder}
For a Banach space $X$ the following are equivalent
\begin{enumerate}
\item $X$ is a UMD Banach space;
\item for each $p\in (1,\infty)$ there exists a constant $\beta$ and a zigzag-concave function $U:X\times X \to \mathbb R$ such that
\begin{equation}\label{eq:ineqonU}
U(x,y)\geq \|y\|^p - \beta^p\|x\|^p,\;\;\;x,y\in X.
\end{equation}
\end{enumerate}
The smallest admissible $\beta$ for which such $U$ exists is $\beta_{p, X}$.
\end{theorem}
\begin{remark}\label{rem:propertiesofU}
Fix a UMD space $X$ and $p\in (1,\infty)$. A special zigzag-concave function $U$ from Theorem \ref{thm:Burkholder} have been obtained in \cite[Theorem 4.5.6]{HNVW1}. We will call this function {\em the Burkholder function}. For the convenience of the reader we leave out the construction of the Burkholder function. The following properties of the Burkholder function $U$ were demonstrated in \cite[Section 3]{Y17FourUMD}:
\begin{itemize}
\item [\rm{(A)}] $U(\alpha x, \alpha y) = |\alpha|^p U(x, y)$ for all $x,y\in X$, $\mathbb \alpha \in \mathbb R$.
\item [\rm{(B)}] $U(x,\alpha x)\leq 0$ for all $x\in X$, $\alpha \in [-1,1]$.
\item [\rm{(C)}] $U$ is continuous.
\end{itemize}
\end{remark}
\begin{remark}\label{rem:propertiesofV}
Fix a UMD space $X$ and $p\in (1,\infty)$. Let the Burkholder function $U$ be as in Remark \ref{rem:propertiesofU}. Then there exists a biconcave function $V:X\times X \to \mathbb R$ such that
\begin{equation}\label{eq:VthroughU}
V(x,y) = U\Bigl(\frac{x-y}{2},\frac{x+y}2\Bigr),\;\;\; x,y\in X.
\end{equation}
In \cite[Section 3]{Y17FourUMD} the following properties of $V$ have been explored:
\begin{itemize}
\item [\rm{(A)}] For each $x,y\in X$ and $a,b \in \mathbb R$ such that $|a+b|\leq |a-b|$ one has that the function
$$
z\mapsto V(x+az,y+bz) = U\Bigl(\frac{x-y}{2} + \frac{(a-b)z}2,\frac{x+y}2 + \frac{(a+b)z}2\Bigr)
$$
is concave.
\item [\rm{(B)}] $V$ is continuous.
\item [\rm{(C)}] Let $X$ be finite dimensional. Then $x\mapsto V(x, y)$ and $y\mapsto V(x,y)$ are a.s.\ Fr\'echet-differentiable with respect to the Lebesgue measure $\lambda_X$, and for a.a.\ $(x,y)\in X\times X$ for each $u,v\in X$ there exists the directional derivative $\frac{\partial V(x+tu,y+tv)}{\partial t}$. Moreover,
\begin{equation}\label{eq:dirderivative}
\frac{\partial V(x+tu,y+tv)}{\partial t} = \langle \partial_x V(x,y), u\rangle+\langle \partial_y V(x,y), v\rangle,
\end{equation}
where $\partial_x V$ and $\partial_y V$ are the corresponding Fr\'echet derivatives with respect to the first and the second variable.
\item [\rm{(D)}] Let $X$ be finite dimensional. Then for a.e.\ $(x,y)\in X\times X$, for all $z\in X$ and real-valued $a$ and $b$ such that $|a+b|\leq |a-b|$
\begin{equation}\label{eq:dirderivativeV}
\begin{split}
V(x+az,y+bz)&\leq V(x,y) + \frac{\partial V(x+atz,y+btz)}{\partial t}\\
&= V(x,y)+a\langle \partial_x V(x,y), z\rangle+b\langle \partial_y V(x,y), z\rangle.
\end{split}
\end{equation}
\item [\rm{(E)}] Let $X$ be finite dimensional. Then there exists $C>0$ which depends only on $V$ such that for a.e.\ pair $x, y\in X$, $\|\partial_xV(x, y)\|, \|\partial_yV(x, y)\| \leq C(\|x\|^{p-1} + \|y\|^{p-1})$.
\end{itemize}
\end{remark}
\begin{definition}\label{def:corrbasis}
Let $d$ be a natural number, $E$ be a $d$-dimensional linear space, $(e_n)_{n=1}^d$ be a basis of~$E$. Then $(e_n^*)_{n=1}^d\subset E^*$ is called the {\em corresponding dual basis} of $(e_n)_{n=1}^d$ if $\langle e_n, e_m^*\rangle = \delta_{nm}$ for each $m,n=1,\ldots, d$.
\end{definition}
Note that the corresponding dual basis is uniquely determined. Moreover, if $(e_n^*)_{n=1}^d$ is the corresponding dual basis of $(e_n)_{n=1}^d$, then, the other way around, $(e_n)_{n=1}^d$ is the corresponding dual basis of $(e_n^*)_{n=1}^d$ (here we identify $E^{**}$ with $E$ in the natural way).
\begin{lemma*}\label{lemma:traceindep}
Let $d$ be a natural number, $E$ be a $d$-dimensional linear space. Let $V: E\times E \to \mathbb R$ and $W:E^* \times E^* \to \mathbb R$ be two bilinear functions. Then the expression
\begin{equation}\label{eq:traceindep}
\sum_{n, m=1}^d V(e_n, e_m) W(e_n^*, e_m^*)
\end{equation}
does not depend on the choice of basis $(e_n)_{n=1}^d$ of $E$ (here $(e_n^*)_{n=1}^d$ is the corresponding dual basis of $(e_n)_{n=1}^d$).
\end{lemma*}
The following It\^o formula is a version of \cite[Theorem 26.7]{Kal} that does not use the Euclidean structure of a finite dimensional Banach space. The proof can be found in \cite{Y17FourUMD}.
\begin{theorem}[It\^o formula]\label{thm:itoformula}
Let $d$ be a natural number, $X$ be a $d$-dimensional Banach space, $f\in C^2(X)$, $M:\mathbb R_+ \times \Omega \to X$ be a martingale. Let $(x_n)_{n=1}^d$ be a basis of $X$, $(x_n^*)_{n=1}^d$ be the corresponding dual basis. Then for each $t\geq 0$
\begin{equation}\label{eq:itoformula}
\begin{split}
f(M_t) = f(M_0)&+ \int_0^t \langle \partial_xf(M_{s-}), \ud M_s\rangle\\
&+ \frac 12 \int_0^t \sum_{n,m=1}^d f_{x_n, x_m}(M_{s-})\ud[\langle M, x_n^*\rangle,\langle M, x_m^*\rangle]_s^c\\
&+ \sum_{s\leq t}(\Delta f(M_s) - \langle \partial_xf(M_{s-}), \Delta M_{s}\rangle).
\end{split}
\end{equation}
\end{theorem}
\begin{proposition}\label{prop:approxXbyYvalued}
Let $X$ be a finite dimensional Banach space, $p\in (1,\infty)$. Let $Y = X \oplus \mathbb R$ be a Banach space such that $\|(x, r)\|_Y = (\|x\|^p_X + |r|^p)^{\frac 1p}$. Then $\beta_{p, Y} = \beta_{p, X}$. Moreover, if $M:\mathbb R_+ \times \Omega \to X$ is a martingale on a probability space $(\Omega, \mathcal F, \mathbb P)$ with a filtration $\mathbb F = (\mathcal F_t)_{t\geq 0}$, then there exists a sequence $(M^m)_{m\geq 1}$ of $Y$-valued martingales on an enlarged probability space $(\overline{\Omega}, \overline{\mathcal F}, \overline{\mathbb P})$ with an enlarged filtration $\overline{\mathbb F} = (\overline{\mathcal F}_t)_{t\geq 0}$ such that
\begin{enumerate}
\item $M^m_t$ has absolutely continuous distributions with respect to the Lebesgue measure on $Y$ for each $m\geq 1$ and $t\geq 0$;
\item $M^m_t \to (M_t,0)$ pointwise as $m\to \infty$ for each $t\geq 0$;
\item if for some $t\geq 0$ $\mathbb E\|M_t\|^p<\infty$, then for each $m\geq 1$ one has that $\mathbb E\|M^m_t\|^p<\infty$ and $\mathbb E\|M^m_t-(M_t,0)\|^p\to 0$ as $m\to \infty$;
\item if $M$ is continuous, then $(M^m)_{m\geq 1}$ are continuous as well,
\item if $M$ is purely discontinuous, then $(M^m)_{m\geq 1}$ are purely discontinuous as well.
\end{enumerate}
\end{proposition}
\begin{proof}
The proof of (1)-(3) follows from \cite{Y17FourUMD}, while (4) and (5) follow from the construction of $M^m$ and $N^m$ given in \cite{Y17FourUMD}.
\end{proof}
\begin{remark}\label{rem:sumofapprox}
Notice that the construction in \cite{Y17FourUMD} also allows us to sum these approximations for different martingales. Namely, if $M$ and $N$ are two $X$-valued martingales, then we can construct the corresponding $Y$-valued martingales $(M^m)_{m\geq 1}$ and $(N^m)_{m\geq 1}$ as in Proposition \ref{prop:approxXbyYvalued} in such a way that $M^m_t + N^m_t$ has an absolutely continuous distribution for each $t\geq 0$ and $m\geq 1$.
\end{remark}
\begin{proof}[Proof of Theorem \ref{thm:Meyer-Yoeurp}]
{\em Step 1: finite dimensional case.} Let $X$ be finite dimensional. Then $M^d$ and $M^c$ exist due to Remark \ref{rem:YoeMey}. Without loss of generality $\mathcal F_t = \mathcal F_{\infty}$, $M^d_t = M^d_{\infty}$ and $M^c_t = M^c_{\infty}$. Let $d$ be the dimension of $X$.
Let $\vertiii{\cdot}$ be a Euclidean norm on $X$. Then $(X, \vertiii{\cdot})$ is a Hilbert space, and by Remark \ref{rem:qvcont} the quadratic variation $[M^c]$ exists and has a continuous version. Let us show that without loss of generality we can suppose that $[M^c]$ is a.s.\ absolutely continuous with respect to the Lebesgue measure on $\mathbb R_+$. Let $A:\mathbb R_+ \times \Omega \to \mathbb R_+$ be as follows: $A_t = [M^c]_t + t$. Then $A$ is strictly increasing continuous, $A_0=0$ and $A_{\infty} = \infty$ a.s. Let the time-change $\tau = (\tau_s)_{s\geq 1}$ be defined as in Theorem \ref{thm:apptimechange}. Then by Theorem \ref{thm:apptimechange}, $M^c \circ \tau$ is a continuous martingale, $M^d\circ \tau$ is a purely discontinuous martingale, $(M^c\circ \tau)_0=0$, $(M^d\circ \tau)_0=M^d_0$ and due to the Kazamaki theorem \cite[Theorem 17.24]{Kal}, $[M^c\circ \tau] = [M^c]\circ \tau$. Therefore for all $t>s\geq 0$ by Theorem \ref{thm:apptimechange} and the fact that $\tau_t \geq \tau_s$ a.s.
\begin{align*}
[M^c\circ \tau]_t - [M^c\circ \tau]_s = [M^c]_{\tau_t} - [M^c]_{\tau_s} &\leq [M^c]_{\tau_t} - [M^c]_{\tau_s} + (\tau_t - \tau_s)\\
& = ([M^c]_{\tau_t}+\tau_t) - ([M^c]_{\tau_s} + \tau_s)\\
&= A_{\tau_t} - A_{\tau_s}=t-s.
\end{align*}
Hence $[M^c\circ \tau]$ is a.s.\ absolutely continuous with respect to the Lebesgue measure on~$\mathbb R_+$. Moreover, $(M^i\circ \tau)_{\infty} = M^i_{\infty}$, $i\in\{c,d\}$, so this time-change argument does not affect~\eqref{eq:Meyer-Yoeurp}. Hence we can redefine $M^c := M^c\circ \tau$, $M^d := M^d \circ \tau$, $\mathbb F = (\mathcal F_s)_{s\geq 0} := \mathbb G = (\mathcal F_{\tau_s})_{s\geq 0}$.
Since $[M^c]$ is a.s.\ absolutely continuous with respect to the Lebesgue measure on~$\mathbb R_+$ and thanks to Theorem \ref{thm:Brrepres}, we can extend $\Omega$ and find a $d$-dimensional Wiener process $W:\mathbb R_+\times \Omega \to \mathbb R^d$ and a stochastically integrable progressively measurable function $\Phi: \mathbb R_+ \times\Omega \to \mathcal L(\mathbb R^d, X)$ such that $M^c = \Phi \cdot W$.
Let $U:X \times X \to \mathbb R$ be the Burkholder function that was discussed in Remark~\ref{rem:propertiesofU} and Remark \ref{rem:propertiesofV}. Let us show that $\mathbb E U(M_t, M^d_t)\leq 0$.
Due to Proposition \ref{prop:approxXbyYvalued} and Remark \ref{rem:sumofapprox} we can assume that $M^c_s$, $M^d_s$ and $M_s = M^d_s + M^c_s$ have absolutely continuous distributions with respect to the Lebesgue measure $\lambda_X$ on $X$ for each $s\geq 0$. Let $(x_n)_{n=1}^d$ be a basis of $X$, $(x_n^*)_{n=1}^d$ be the corresponding dual basis of $X^*$ (see Definition \ref{def:corrbasis}). By the It\^o formula \eqref{eq:itoformula},
\begin{equation}\label{eq:EU(M^d+M^c,M^d)}
\begin{split}
\mathbb E U(M_t,M^d_t) = \mathbb E U(M_0,M^d_0) &+ \mathbb E \int_{0}^t \langle \partial_x U(M_{s-},M^d_{s-}),\ud M_s\rangle\\
&+\mathbb E \int_{0}^t \langle \partial_y U(M_{s-},M^d_{s-}),\ud M^d_s\rangle + \mathbb E I_1 + \mathbb E I_2,
\end{split}
\end{equation}
where
\begin{align*}
I_1 &= \sum_{0<s\leq t}[\Delta U(M_s,M^d_s)-\langle \partial_x U(M_{s-},M^d_{s-}),\Delta M_s\rangle - \langle \partial_y U(M_{s-},M^d_{s-}),\Delta M^d_s\rangle],\\
I_2 &= \frac 12 \int_0^t\sum_{i,j=1}^d U_{x_i, x_j}(M_{s-}, M^d_{s-})\ud[\langle M, x_i^*\rangle, \langle M, x_j^*\rangle]_s^c\\
&\quad =\frac 12 \int_0^t\sum_{i,j=1}^d U_{x_i, x_j}(M_{s-}, M^d_{s-})\langle \Phi^*(s) x_i^*, \Phi^*(s) x_j^*\rangle\ud s.
\end{align*}
(Recall that by \eqref{eq:VthroughU} and Remark \ref{rem:propertiesofV}(C), $U$ is Fr\'echet-differentiable a.s.\ on $X\times X$, hence $\partial_x U$ and $\partial_y U$ are well-defined. Moreover, $U$ is zigzag-concave, so $U$ is concave in the first variable, and therefore the second-order derivatives $U_{x_i, x_j}$ in the first variable are well-defined and exist a.s.\ on $X \times X$ by the Alexandrov theorem \cite[Theorem 6.4.1]{EG}.) The last equality holds due to Theorem \ref{thm:itoformula} and the fact that by Lemma \ref{lemma:covstochintwrtcylbrmo} for all $s\geq 0$~a.s.\
\begin{align*}
[\langle M, x_i^*\rangle, \langle M, x_j^*\rangle]_s^c = [\langle \Phi\cdot W, x_i^*\rangle, \langle \Phi\cdot W, x_j^*\rangle]_s &= [(\Phi^*x_i^*)\cdot W, (\Phi^*x_j^*)\cdot W]_s\\
&=\int_0^s \langle\Phi^*(r)x_i^*, \Phi^*(r)x_j^*\rangle\ud r.
\end{align*}
Let us first show that $I_1\leq 0$ a.s. Indeed, since $M^d$ is a purely discontinuous part of $M$, then by Definition \ref{def:purelydiscXvalued} $\langle M^d,x^*\rangle$ is a purely discontinuous part of $\langle M,x^*\rangle$, and due to Remark \ref{rem:YoeMeyR} a.s.\ for each $t\geq 0$
$$
\Delta |\langle M^d,x^*\rangle|_t^2= \Delta [\langle M^d,x^*\rangle]_t= \Delta [\langle M,x^*\rangle]_t=\Delta |\langle M,x^*\rangle|_t^2
$$
for each $x^* \in X^*$. Thus for each $s\geq 0$ by \eqref{eq:dirderivative} and \eqref{eq:dirderivativeV} $\mathbb P$-a.s.
\begin{align*}
&\Delta U(M_s,M^d_s)-\langle \partial_x U(M_{s-},M^d_{s-}),\Delta M_s\rangle - \langle \partial_y U(M_{s-},M^d_{s-}),\Delta M^d_s\rangle\\
&= V(M_{s-}+M^d_{s-}+2\Delta M_s,M^d_{s-}-M_{s-})- V(M_{s-}+M^d_{s-},M^d_{s-}-M_{s-})\\
&\quad-\langle \partial_x V(M_{s-}+M^d_{s-},M^d_{s-}-M_{s-}),2\Delta M_s\rangle \leq 0,
\end{align*}
so $I_1\leq 0$ a.s., and $\mathbb E I_1 \leq 0$.
Now we show that
\[
\mathbb E \Bigl(\int_{0}^t \langle \partial_x U(M_{s-},M^d_{s-}),\ud M_s\rangle + \int_{0}^t \langle \partial_y U(M_{s-},M^d_{s-}),\ud M^d_s\rangle\Bigr) = 0.
\]
Indeed,
\begin{align*}
\int_{0}^t \langle \partial_x U(M_{s-},M^d_{s-}),\ud M_s\rangle &+ \int_{0}^t \langle \partial_y U(M_{s-},M^d_{s-}),\ud M^d_s\rangle\\
&=\int_{0}^t \langle \partial_x V(M_{s-}+M^d_{s-},M^d_{s-}-M_{s-}),\ud (M_s+M^d_s)\rangle \\
&+ \int_{0}^t \langle \partial_y V(M_{s-}+M^d_{s-},M^d_{s-}-M_{s-}),\ud (M^d_s-M_s)\rangle
\end{align*}
so by Lemma \ref{lemma:stochintmoment} and Remark \ref{rem:propertiesofV}(E) it is a martingale which starts at zero, hence its expectation is zero.
Finally let us show that $I_2 \leq 0$ a.s. Fix $s\in [0,t]$ and $\omega \in \Omega$. Then $x^* \mapsto \|\Phi^*(s, \omega) x^*\|^2$ defines a nonnegative definite quadratic form on $X^*$, and since any nonnegative quadratic form defines a Euclidean seminorm, there exists a basis $(\tilde x_n^*)_{n=1}^d$ of $X^*$ and a $\{0, 1\}$-valued sequence $(a_n)_{n=1}^d$ such that
\[
\langle \Phi^*(s, \omega) \tilde x_n^*, \Phi^*(s, \omega) \tilde x_m^* \rangle = a_n\delta_{mn},\;\;\; m,n= 1,\ldots,d.
\]
Let $(\tilde x_n)_{n=1}^d$ be the corresponding dual basis of $X$ as it is defined in Definition \ref{def:corrbasis}. Then due to Lemma~\ref{lemma:traceindep} and the linearity of $\Phi$ and directional derivatives of $U$ (we skip $s$ and $\omega$ for the simplicity of the expressions)
\begin{equation*}
\begin{split}
\sum_{i,j=1}^d U_{x_i, x_j}(M_{s-}, M^d_{s-})\langle \Phi^* x_i^*, \Phi^* x_j^*\rangle &=\sum_{i,j=1}^d U_{\tilde x_i,\tilde x_j}(M_{s-}, M^d_{s-})\langle \Phi^*\tilde x_i^*, \Phi^* \tilde x_j^*\rangle\\
&=\sum_{i=1}^d U_{\tilde x_i,\tilde x_i}(M_{s-}, M^d_{s-})\|\Phi^*\tilde x_i^*\|^2.
\end{split}
\end{equation*}
Recall that $U$ is zigzag-concave, so $t\mapsto U(x+t\tilde x_i, y)$ is concave for each $x,y\in X$, $i=1,\ldots,d$. Therefore $U_{\tilde x_i,\tilde x_i}(M_{s-}, M^d_{s-}) \leq 0$ a.s., and a.s.
\[
\sum_{i=1}^d U_{\tilde x_i,\tilde x_i}(M_{s-}(\omega), M^d_{s-}(\omega))\|\Phi^*(s,\omega)\tilde x_i^*\|^2\leq 0.
\]
Consequently, $I_2\leq 0$ a.s., and by \eqref{eq:EU(M^d+M^c,M^d)}, Remark \ref{rem:propertiesofU}(B) and the fact that $M^d_0=M_0$
$$
\mathbb E U(M_t, M^d_t)\leq \mathbb E U(M_0, M_0)\leq 0.
$$
By \eqref{eq:ineqonU},
$\mathbb E \|M^d_t\|^p - \beta_{p, X}^p\mathbb E\|M_t\|^p \leq \mathbb E U(M_t, M^d_t) \leq 0$,
so the first part of \eqref{eq:Meyer-Yoeurp} holds.
The second part of \eqref{eq:Meyer-Yoeurp} follows from the same machinery applied for $V$. Namely, one can analogously show that
\[
\mathbb E\|M^c_t\|^p - \beta_{p, X}^p \mathbb E \|M_t\|^p \leq \mathbb E U(M_t, M^c_t) = \mathbb E V(M^d + 2M^c, -M^d) \leq 0
\]
by using a $V$-version of \eqref{eq:EU(M^d+M^c,M^d)}, inequality \eqref{eq:dirderivativeV}, and the fact that $V$ is concave in the first variable a.s.\ on $X\times X$.
{\em Step 2: general case.} Without loss of generality we set $\mathcal F_{\infty} = \mathcal F_t$. Let $M_t = \xi$. If $\xi$ is a simple function, then it takes its values in a finite dimensional subspace $X_0$ of $X$, and therefore $(M_s)_{s\geq 0} = (\mathbb E(\xi|\mathcal F_s))_{s\geq 0}$ takes its values in $X_0$ as well, so the theorem and \eqref{eq:Meyer-Yoeurp} follow from Step 1.
Now let $\xi$ be general. Let $(\xi_n)_{n\geq 1}$ be a sequence of simple $\mathcal F_t$-measurable functions in $L^p(\Omega; X)$ such that $\xi_n \to \xi$ as $n\to \infty$ in $L^p(\Omega; X)$. For each $n\geq 1$ define $\mathcal F_t$-measurable $\xi_n^d$ and $\xi_n^c$ such that
\begin{equation}\label{eq:M^dnM^cn}
\begin{split}
M^{d, n} &= (M^{d, n}_s)_{s\geq 0}=(\mathbb E(\xi_n^d|\mathcal F_{s}))_{s\geq 0},\\
M^{c,n} &= (M^{c, n}_s)_{s\geq 0}=(\mathbb E(\xi_n^c|\mathcal F_{s}))_{s\geq 0}
\end{split}
\end{equation}
are the respectively purely discontinuous and continuous parts of martingale $M^n=(\mathbb E(\xi_n|\mathcal F_{s}))_{s\geq 0}$ as in Remark \ref{rem:YoeMey}. Then due to Step 1 and \eqref{eq:Meyer-Yoeurp}, $(\xi_n^d)_{n\geq 1}$ and $(\xi_n^c)_{n\geq 1}$ are Cauchy sequences in $L^p(\Omega; X)$. Let $\xi^c := L^p-\lim_{n\to \infty} \xi^c_n$ and $\xi^d := L^p-\lim_{n\to\infty} \xi^d_n$. Define the $X$-valued $L^p$-martingales $M^d$ and $M^c$ by
\begin{align*}
M^d = (M^d_s)_{s\geq 0} := (\mathbb E(\xi^d|\mathcal F_{s}))_{s\geq 0},\;\;\;
M^c = (M^c_s)_{s\geq 0} := (\mathbb E(\xi^c|\mathcal F_{s}))_{s\geq 0}.
\end{align*}
Thanks to Proposition \ref{thm:M^p,discomplete}, $M^d$ is purely discontinuous, and due to Proposition~\ref{prop:M^p,ciscomplete} $M^c$ is continuous and $M^c_0=0$, so $M = M^d + M^c$ is the desired decomposition.
The uniqueness of the decomposition follows from Lemma \ref{lemma:contpuredisczero}. For estimates \eqref{eq:Meyer-Yoeurp} we note that by Step 1, \eqref{eq:Meyer-Yoeurp} applied for Step 1, and \cite[Proposition 4.2.17]{HNVW1} for each $n\geq 1$
$$
(\mathbb E\|\xi_n^d\|^p)^{\frac 1p}\leq \beta_{p,X}(\mathbb E\|\xi_n\|^p)^{\frac 1p},\;\;\;\;
(\mathbb E\|\xi_n^c\|^p)^{\frac 1p}\leq \beta_{p,X}(\mathbb E\|\xi_n\|^p)^{\frac 1p},
$$
and it remains to let $n\to \infty$.
\end{proof}
\begin{remark}\label{cor:approxofconandpurediscmartbyfd}
Let $X$ be a UMD Banach space, $1<p<\infty$, $M:\mathbb R_+ \times \Omega \to X$ be continuous (resp.\ purely discontinuous) $L^p$-martingale. Then there exists a sequence $(M^n)_{n\geq 1}$ of continuous (resp.\ purely discontinuous) $X$-valued $L^p$-martingales such that $M^n$ takes its values is a finite dimensional subspace of $X$ for each $n \geq 1$ and $M^n_{\infty} \to M_{\infty}$ in $L^p(\Omega; X)$ as $n\to \infty$. Such a sequence can be provided e.g.\ by \eqref{eq:M^dnM^cn}.
\end{remark}
We have proven the Meyer-Yoeurp decomposition in the UMD setting. Next we prove a converse result which shows the necessity of the UMD property.
\begin{theorem}\label{thm:exampleforlowboundofA}
Let $X$ be a finite dimensional Banach space, $p\in(1,\infty)$, $\delta\in(0, (\beta_{p, X}-1)\wedge 1)$. Then there exist a purely discontinuous martingale $M^d:\mathbb R_+ \times \Omega \to X$, a continuous martingale $M^c:\mathbb R_+ \times \Omega \to X$ such that $\mathbb E \|M^d_{\infty}\|^p, \mathbb E \|M^c_{\infty}\|^p<\infty$, $M^d_0 = M^c_0=0$, and for $M = M^d + M^c$ and $i\in\{c,d\}$ the following hold
\begin{equation}\label{eq:exampleforlowboundofA}
(\mathbb E \|M^i_{\infty}\|^p)^{\frac 1p} \geq \Bigl(\frac{\beta_{p,X}-1}{2}-\delta\Bigr) (\mathbb E \|M_{\infty}\|^p)^{\frac 1p}.
\end{equation}
\end{theorem}
Recall that by \cite[Proposition 4.2.17]{HNVW1} $\beta_{p, X} \geq \beta_{p, \mathbb R}= p^*-1 \geq 1$ for any UMD Banach space $X$ and $1<p<\infty$.
\begin{definition}
A random variable $r:\Omega \to \{-1, 1\}$ is called a~{\em Rademacher variable} if $\mathbb P(r=1) = \mathbb P(r=-1) = \frac 12$.
\end{definition}
\begin{lemma*}\label{lem:contapproxradem}
Let $\varepsilon>0$, $p\in (1,\infty)$. Then there exists a continuous martingale $M:[0,1]\times \Omega \to [-1,1]$ with a symmetric distribution such that $\mathrm{sign}\text{ } M_1$ is a~Rademacher random variable and
\begin{equation}\label{eq:lemcontapproxradem}
\|M_1-\mathrm{sign}\text{ } M_1\|_{L^p(\Omega)}<\varepsilon.
\end{equation}
\end{lemma*}
We will need a definition of a Paley-Walsh martingale.
\begin{definition}[Paley-Walsh martingales]\label{def:Paley-Walsh}
Let $X$ be a Banach space. A discrete $X$-valued martingale $(f_n)_{n\geq 0}$ is called a {\em Paley-Walsh martingale} if there exist a sequence of independent Rademacher variables $(r_n)_{n\geq 1}$, a function $\phi_n:\{-1, 1\}^{n-1}\to X$ for each $n\geq 2$ and $\phi_1 \in X$ such that $df_n = r_n \phi_n(r_1,\ldots, r_{n-1})$ for each $n\geq 2$ and $df_1 = r_1 \phi_1$.
\end{definition}
\begin{remark}\label{rem:thesamelowerboundford_j}
Let $X$ be a UMD space, $1<p<\infty$, $\delta>0$. Then using Proposition \ref{prop:UMD01sequence} one can construct a martingale difference sequence $(d_j)^n_{j=1}\in L^p(\Omega; X)$ and a $\{-1,1\}$-valued sequence
$(\varepsilon_j)^n_{j=1}$ such that
\begin{align*}
\Bigl(\mathbb E \Bigl \| \sum^n_{j=1}\frac {\varepsilon_j\pm1}{2}d_j\Bigr\|^p\Bigr )^{\frac 1p}& \geq \frac{\beta_{p,X}-\delta-1}{2} \Bigl(\mathbb E \Bigl \| \sum^n_{j=1}d_j\Bigr\|^p\Bigr )^{\frac 1p}.
\end{align*}
\end{remark}
\begin{proof}[Proof of Theorem \ref{thm:exampleforlowboundofA}]
Denote $\frac{\beta_{p,X}-\delta-1}{2}$ by $\gamma_{p,X}^{\delta}$. By Proposition \ref{prop:UMD01sequence} there exists a natural number $N\geq 1$, a~discrete $X$-valued martingale $(f_n)_{n=0}^N$ such that $f_0 =0$, and a sequence of scalars $(\varepsilon_n)_{n=1}^N$ such that $\varepsilon_n\in \{0,1\}$ for each $n=1,\ldots, N$, such that
\begin{equation}\label{eq:peoofofexampleforlowboundofA}
\Bigl(\mathbb E \Bigl\| \sum^N_{n=1} \varepsilon_n df_n\Bigr\|^p\Bigr )^{\frac 1p}
\geq \gamma_{p,X}^{\delta} (\mathbb E \|f_N\|^p)^{\frac 1p}.
\end{equation}
According to \cite[Theorem 3.6.1]{HNVW1} we can assume that $(f_n)_{n=0}^N$ is a Paley-Walsh martingale. Let $(r_n)_{n=1}^N$ be a sequence of Rademacher variables and $(\phi_n)_{n=1}^N$ be a sequence of functions as in Definition \ref{def:Paley-Walsh}, i.e.\ be such that $f_n = \sum_{k=2}^n r_k\phi_k(r_1,\ldots,r_{k-1}) + r_1\phi_1$ for each $n=1, \ldots,N$. Without loss of generality we assume that
\begin{equation}\label{eq:analogueofpeoofofexampleforlowboundofAprivetvsem}
(\mathbb E\|f_{N}\|^p)^{\frac 1p} \geq 2.
\end{equation}
For each $n=1, \ldots,N$ define a continuous martingale $M^n:[0,1]\times \Omega \to [-1,1]$ as in Lemma \ref{lem:contapproxradem}, i.e.\ a martingale $M^n$ with a symmetric distribution such that $\mathrm{sign}\text{ } M^n_1$ is a Rademacher variable and
\begin{equation}\label{eq:M^n_1-signM_n^1}
\|M^n_1-\mathrm{sign}\text{ } M^n_1\|_{L^p(\Omega)}<\frac{\delta}{KL},
\end{equation}
where $K =\beta_{p, X}N \max\{\|\phi_1\|, \|\phi_2\|_{\infty},\ldots, \|\phi_N\|_{\infty}\}$, and $L =2\beta_{p, X}$. Without loss of generality suppose that $(M^n)_{n=1}^N$ are independent. For each $n=1,\ldots, N$ set $\sigma_n = \mathrm{sign}\text{ } M^n_1$. Define a martingale $M:[0, N+1]\times \Omega \to X$ in the following way:
\[
M_t =
\begin{cases}
0, &\text{if}\;\; 0\leq t<1;\\
M_{n-}+ M^n_{t-n}\phi_{n}(\sigma_1,\ldots,\sigma_{n-1}),&\text{if}\;\; t\in[n,n+1)\;\; \text{and}\;\; \varepsilon_n=0;\\
M_{n-}+ \sigma_n \phi_{n}(\sigma_1,\ldots,\sigma_{n-1}),&\text{if}\;\; t\in[n,n+1)\;\; \text{and}\;\; \varepsilon_n=1.
\end{cases}
\]
Let $M = M^d + M^c$ be the decomposition of Theorem \ref{thm:Meyer-Yoeurp}. Then
\begin{align*}
M^c_{N+1} &= \sum_{n=1}^NM^n_1\phi_{n}(\sigma_1,\ldots,\sigma_{n-1})\mathbf 1_{\varepsilon_n=0},\\
M^d_{N+1} &= \sum_{n=1}^N\sigma_n \phi_{n}(\sigma_1,\ldots,\sigma_{n-1})\mathbf 1_{\varepsilon_n=1} =\sum_{n=1}^N\varepsilon_n\sigma_n \phi_{n}(\sigma_1,\ldots,\sigma_{n-1}).
\end{align*}
Notice that $(\sigma_n)_{n=1}^N$ is a sequence of independent Rademacher variables, so by~\eqref{eq:peoofofexampleforlowboundofA} and the discussion thereafter
\begin{equation}\label{eq:analogueofpeoofofexampleforlowboundofA}
\Bigl(\mathbb E \Bigl\| \sum^N_{n=1} \varepsilon_n \,\sigma_n \phi_{n}(\sigma_1,\ldots,\sigma_{n-1})\Bigr\|^p\Bigr)^{\frac 1p}
\geq \gamma_{p,X}^{\delta}\Bigl(\mathbb E \Bigl\| \sum^N_{n=1} \sigma_n \phi_{n}(\sigma_1,\ldots,\sigma_{n-1})\Bigr\|^p\Bigr)^{\frac 1p}.
\end{equation}
Let us first show \eqref{eq:exampleforlowboundofA} with $i=d$. Note that by the triangle inequality, \eqref{eq:analogueofpeoofofexampleforlowboundofAprivetvsem}~and~\eqref{eq:M^n_1-signM_n^1}
\begin{equation}\label{eq:analogueofpeoofofexampleforlowboundofA12121}
\begin{split}
(\mathbb E\|M_{N+1}\|^p)^{\frac 1p} &\geq (\mathbb E\|f_{N}\|^p)^{\frac 1p} - \sum^N_{n=1} \Bigl(\mathbb E \Bigl\| (M^n_1-\sigma_n) \phi_{n}(\sigma_1,\ldots,\sigma_{n-1})\Bigr\|^p\Bigr)^{\frac 1p}\\
&\geq 2-\frac{\delta}{KL}\cdot N \cdot\max\{\|\phi_1\|, \|\phi_2\|_{\infty},\ldots, \|\phi_N\|_{\infty}\} >1.
\end{split}
\end{equation}
Therefore,
\begin{align*}
(\mathbb E\|M^d_{N+1}\|^p)^{\frac 1p}
&= \Bigl(\mathbb E \Bigl\| \sum^N_{n=1} \varepsilon_n \,\sigma_n \phi_{n}(\sigma_1,\ldots,\sigma_{n-1})\Bigr\|^p\Bigr)^{\frac 1p} \stackrel{(i)}\geq\gamma_{p,X}^{\delta} \Bigl(\mathbb E \Bigl\| \sum^N_{n=1} \sigma_n \phi_{n}(\sigma_1,\ldots,\sigma_{n-1})\Bigr\|^p\Bigr)^{\frac 1p} \\
&\stackrel{(ii)}\geq\gamma_{p,X}^{\delta} \Bigl(\mathbb E \Bigl\| \sum^N_{n=1}\mathbf 1_{\varepsilon_n = 1} \sigma_n \phi_{n}(\sigma_1,\ldots,\sigma_{n-1}) + \sum^N_{n=1}\mathbf 1_{\varepsilon_n = 0} M^n_1 \phi_{n}(\sigma_1,\ldots,\sigma_{n-1})
\Bigr\|^p\Bigr)^{\frac 1p} \\
&\quad -\gamma_{p,X}^{\delta}\sum^N_{n=1} \Bigl(\mathbb E \Bigl\| (M^n_1-\sigma_n) \phi_{n}(\sigma_1,\ldots,\sigma_{n-1})\Bigr\|^p\Bigr)^{\frac 1p}\\
&\stackrel{(iii)}\geq\gamma_{p,X}^{\delta}(\mathbb E\|M_{N+1}\|^p)^{\frac 1p} - \frac{\delta}{L}\stackrel{(iv)}\geq\Bigl(\frac{\beta_{p,X}-1}{2}-\delta\Bigr)(\mathbb E\|M_{N+1}\|^p)^{\frac 1p} ,
\end{align*}
where $(i)$ follows from \eqref{eq:analogueofpeoofofexampleforlowboundofA}, $(ii)$ holds by the triangle inequality, $(iii)$ holds by \eqref{eq:M^n_1-signM_n^1}, and $(iv)$ follows from \eqref{eq:analogueofpeoofofexampleforlowboundofA12121}. By the same reason and Remark \ref{rem:thesamelowerboundford_j}, \eqref{eq:exampleforlowboundofA} holds for $i=c$.
\end{proof}
Let $p\in (1,\infty)$. Recall that $\mathcal M_X^p$ is a space of all $X$-valued $L^p$-martingales, $\mathcal M_X^{p, d}, \mathcal M_X^{p, c}\subset \mathcal M_X^{p}$ are its subspaces of purely discontinuous martingales and continuous martingales that start at zero respectively (see Subsection \ref{subsec:prelimmart}, \ref{subsec:prelimcontmart}, and \ref{subsec:prelimpdmart}).
\begin{theorem*}\label{thm:charofUMDbyMeyYoe}
Let $X$ be a Banach space. Then $X$ is UMD if and only if for some (or, equivalently, for all) $p\in(1,\infty)$, for any probability space $(\Omega, \mathcal F,\mathbb P)$ with any filtration $\mathbb F = (\mathcal F_t)_{t\geq 0}$ that satisfies the usual conditions, $\mathcal M_X^{p} = \mathcal M_X^{p, d} \oplus \mathcal M_X^{p, c}$, and there exist projections $A^d, A^c\in \mathcal L(\mathcal M_X^{p})$ such that $\text{ran } A^d = \mathcal M_X^{p, d}$, $\text{ran } A^c = \mathcal M_X^{p, c}$, and for any $M \in \mathcal M_X^{p}$ the decomposition $M = A^d M + A^cM$
is the Meyer-Yoeurp decomposition from Theorem \ref{thm:Meyer-Yoeurp}. If this is the case, then
\begin{equation}\label{eq:charofUMDbyMeyYoe1}
\|A^d\|\leq \beta_{p,X}\;\; \text{and}\;\;
\|A^c\|\leq \beta_{p,X}.
\end{equation}
Moreover, there exist $(\Omega, \mathcal F,\mathbb P)$ and $\mathbb F = (\mathcal F_t)_{t\geq 0}$ such that
\begin{equation}\label{eq:charofUMDbyMeyYoe2}
\|A^d\|,\|A^c\|\geq \frac {\beta_{p, X}-1}{2}\vee 1.
\end{equation}
\end{theorem*}
\begin{corollary}\label{cor:M_X^pd^*=M_X^*^p'dM_X^pc^*=M_X^*^p'c}
Let $X$ be a UMD Banach space, $p\in (1,\infty)$. Let $i\in \{c,d\}$. Then $(\mathcal M_X^{p, i})^* \simeq \mathcal M_{X^*}^{p', i}$, and for each $M\in \mathcal M_{X^*}^{p', i}$ and $N\in \mathcal M_X^{p, i}$
\begin{equation*}
\langle M, N\rangle:= \mathbb E\langle M_{\infty}, N_{\infty}\rangle,\;\; \;\;
\|M\|_{(\mathcal M_X^{p, i})^*}\eqsim_{p, X}\|M\|_{\mathcal M_{X^*}^{p', i}}.
\end{equation*}
\end{corollary}
To prove the corollary above we will need the following lemma.
\begin{lemma}
Let $X$ be a UMD Banach space, $p\in (1,\infty)$, $M \in \mathcal M_X^{p, d}$, $N \in \mathcal M_{X^*}^{p', c}$. Then $\mathbb E\langle M_{\infty}, N_{\infty}\rangle = 0$.
\end{lemma}
\begin{proof}
First suppose that $N_{\infty}$ takes it values in a finite dimensional subspace $Y$ of $X^*$. Let $d\geq 1$ be the dimension of $Y$, $(y_k)_{k=1}^d$ be the basis of $Y$. Then there exist $N^1,\ldots,N^d \in \mathcal M_{\mathbb R}^{p', c}$ such that $N = \sum_{k=1}^d N^k y_k$. Hence
\begin{equation}\label{eq:EM_inftyN_infty=0}
\begin{split}
E\langle M_{\infty}, N_{\infty}\rangle = E\Bigl\langle M_{\infty}, \sum_{k=1}^d N^k_{\infty} y_k\Bigr\rangle = \sum_{k=1}^d\mathbb E\langle M_{\infty}, y_k\rangle N^k_{\infty} \stackrel{(*)}= 0,
\end{split}
\end{equation}
where $(*)$ holds due to Proposition \ref{thm:purdiscorthtoanycont1}.
Now turn to the general case. By Remark \ref{cor:approxofconandpurediscmartbyfd} for each $N \in \mathcal M_{X^*}^{p', c}$ there exists a se\-qu\-ence $(N^n)_{n\geq 1}$ of continuous martingales such that each of $N^n$ is in $\mathcal M_{X^*}^{p', c}$ and takes its valued in a finite dimensional subspace of $X^*$, and $N^n_{\infty} \to N_{\infty}$ in $L^{p'}(\Omega; X^*)$ as $n\to \infty$. Then due to \eqref{eq:EM_inftyN_infty=0},
$ E\langle M_{\infty}, N_{\infty}\rangle = \lim_{n\to \infty} E\langle M_{\infty}, N^n_{\infty}\rangle =0$,
so the lemma holds.
\end{proof}
\begin{proof}[Proof of Corollary \ref{cor:M_X^pd^*=M_X^*^p'dM_X^pc^*=M_X^*^p'c}]
We will show only the case $i=d$, the case $i=c$ can be shown analogously.
$\mathcal M_{X^*}^{p', d}\subset (\mathcal M_X^{p, d})^*$ and $\|M\|_{(\mathcal M_X^{p, d})^*} \leq \|M\|_{\mathcal M_{X^*}^{p', d}}$ for each $M \in \mathcal M_{X^*}^{p', d}$ thanks to the H\"older inequality. Now let us show the inverse. Let $f\in (\mathcal M_X^{p, d})^*$. Since due to Proposition \ref{thm:M^p,discomplete} $\mathcal M_X^{p, d}$ is a closed subspace of $\mathcal M_X^{p}$, by the Hahn-Banach theorem and Proposition \ref{prop:dualofMXp} there exists $L \in \mathcal M_{X^*}^{p'}$ such that
$\mathbb E\langle L_{\infty}, N_{\infty}\rangle = f(N)$ for any $N \in \mathcal M_X^{p, d}$,
and $\|L\|_{\mathcal M_{X^*}^{p'}} = \|f\|_{ (\mathcal M_X^{p, d})^*}$. Let $L = L^d + L^c$ be the Meyer-Yoeurp decomposition of $L$ as in Theorem \ref{thm:Meyer-Yoeurp}. Then by \eqref{eq:Meyer-Yoeurp}
$$
\|L^d\|_{\mathcal M_{X^*}^{p', d}}\lesssim_{p, X}\|L\|_{\mathcal M_{X^*}^{p'}} = \|f\|_{ (\mathcal M_X^{p, d})^*}
$$
and
$ \mathbb E\langle L^d_{\infty}, N_{\infty}\rangle = \mathbb E\langle L_{\infty}, N_{\infty}\rangle$,
so the theorem holds.
\end{proof}
\subsection{Yoeurp decomposition of purely discontinuous martingales}\label{subsec:Youdec}
As Yoeurp shown in \cite{Yoe76}, one can provide further decomposition of a purely discontinuous martingale into two parts: a martingale with accessible jumps and a quasi-left continuous martingale. This subsection is devoted to the generalization of this result to a UMD case.
\begin{definition}
Let $\tau$ be a stopping time. Then $\tau$ is called a {\em predictable stopping time} if there exists a sequence of stopping times $(\tau_n)_{n\geq 1}$ such that $\tau_n<\tau$ a.s.\ on $\{\tau>0\}$ for each $n\geq 1$ and $\tau_n \nearrow\tau$ a.s.
\end{definition}
\begin{definition}
Let $\tau$ be a stopping time. Then $\tau$ is called a {\em totally inaccessible stopping time} if $\mathbb P\{\tau = \sigma < \infty\} =0$ for each predictable stopping time $\sigma$.
\end{definition}
\begin{definition}\label{def:accjumpsqlc}
Let $A:\mathbb R_+ \times \Omega \to \mathbb R$ be an adapted c\`adl\`ag process. $A$ has {\em accessible jumps} if $\Delta A_{\tau}=0$ a.s.\ for any totally inaccessible stopping time $\tau$. $A$ is called {\em quasi-left continuous} if $\Delta A_{\tau}=0$ a.s.\ for any predictable stopping time $\tau$.
\end{definition}
For the further information on the definitions given we refer the reader to \cite{Kal}.
\begin{remark}\label{rem:incprocdecom}
According to \cite[Proposition 25.17]{Kal} one can show that for any pure jump increasing adapted c\`adl\`ag process $A:\mathbb R_+ \times \Omega \to \mathbb R$ there exist unique increasing adapted c\`adl\`ag processes $A^a, A^q:\mathbb R_+ \times \Omega \to \mathbb R$ such that $A^a$ has accessible jumps, $A^q$ is quasi-left continuous, $A^q_0 =0$ and $A=A^a + A^q$.
\end{remark}
The following decomposition theorem was shown by Yoeurp in \cite{Yoe76} (see also \cite[Corollary 26.16]{Kal}):
\begin{theorem}\label{thm:Yoeurpdec}
Let $M:\mathbb R_+ \times \Omega \to \mathbb R$ be a purely discontinuous martingale. Then there exist unique purely discontinuous martingales $M^a,M^q:\mathbb R_+ \times \Omega \to \mathbb R$ such that $M^a$ is has accessible jumps, $M^q$ is quasi-left continuous, $M^q_0=0$ and $M = M^a + M^q$. Moreover, then $[M^a] = [M]^a$ and $[M^q] = [M]^q$.
\end{theorem}
\begin{corollary}\label{cor:M=M_0ifMbothaccjandqlc}
Let $M:\mathbb R_+ \times \Omega \to \mathbb R$ be a purely discontinuous martingale which is both with accessible jumps and quasi-left continuous. Then $M = M_0$ a.s.
\end{corollary}
\begin{proof}
Without loss of generality we can set $ M_0=0$. Then $M = M+ 0 = 0+M$ are decompositions of $M$ into a sum of a martingale with accessible jumps and a quasi-left continuous martingale. Since by Theorem \ref{thm:Yoeurpdec} this decomposition is unique, $M=0$ a.s.
\end{proof}
\begin{proposition*}\label{prop:limthmforaccjqlc}
Let $1<p<\infty$, $M:\mathbb R_+ \times \Omega \to \mathbb R$ be a purely discontinuous \mbox{$L^p$-mar}\-tin\-gale. Let $(M^n)_{n\geq 1}$ be a sequence of purely discontinuous martingales such that $M^n_{\infty} \to M_{\infty}$ in $L^p(\Omega)$. Then the following assertions hold
\begin{itemize}
\item [(a)] if $(M^n)_{n\geq 1}$ have accessible jumps, then $M$ has accessible jumps as well;
\item [(b)] if $(M^n)_{n\geq 1}$ are quasi-left continuous martingales, then $M$ is quasi-left continuous as well.
\end{itemize}
\end{proposition*}
\begin{definition}
Let $X$ be a Banach space. A martingale $M:\mathbb R_+ \times \Omega \to X$ has {\em accessible jumps} if $\Delta M_{\tau}=0$ a.s.\ for any totally inaccessible stopping time $\tau$. A~martingale $M:\mathbb R_+ \times \Omega \to X$ is called {\em quasi-left continuous} if $\Delta M_{\tau}=0$ a.s.\ for any predictable stopping time $\tau$.
\end{definition}
\begin{lemma*}\label{lem:weakdefforaccjumpsandqlc}
Let $X$ be a reflexive Banach space, $M:\mathbb R_+ \times \Omega \to X$ be a purely discontinuous martingale.
\begin{itemize}
\item [(i)] $M$ has accessible jumps if and only if for each $x^* \in X^*$ the martingale $\langle M, x^*\rangle$ has accessible jumps;
\item [(ii)] $M$ is quasi-left continuous if and only if for each $x^* \in X^*$ the martingale $\langle M, x^*\rangle$ is quasi-left continuous.
\end{itemize}
\end{lemma*}
\begin{definition}
Let $X$ be a Banach space, $p\in (1,\infty)$. Then we define $\mathcal M_X^{p, q}\subset \mathcal M_X^{p,d}$ as a linear space of all $X$-valued purely discontinuous quasi-left continuous $L^p$-martingales which start at $0$. We define $\mathcal M_X^{p, a}\subset \mathcal M_X^{p,d}$ as a linear space of all $X$-valued purely discontinuous $L^p$-martingales with accessible jumps.
\end{definition}
\begin{proposition*}\label{prop:M^pqandM^paareclosed}
Let $X$ be a Banach space, $1<p<\infty$. Then $\mathcal M_X^{p, q}$ and $\mathcal M_X^{p, a}$ are closed subspaces of $\mathcal M_X^{p, d}$.
\end{proposition*}
The following lemma follows from Corollary \ref{cor:M=M_0ifMbothaccjandqlc}.
\begin{lemma*}\label{lem:accjumps+qlc=cons}
Let $X$ be a Banach space, $M:\mathbb R_+ \times\Omega \to X$ be a purely discontinuous martingale. Let $M$ be both with accessible jumps and quasi-left continuous. Then $M = M_0$ a.s. In other words, $\mathcal M_X^{p, q}\cap\mathcal M_X^{p, a} = 0$.
\end{lemma*}
The main theorem of this subsection is the following UMD variant of Theorem~\ref{thm:Yoeurpdec}.
\begin{theorem}\label{thm:YoeurpdecUMD}
Let $X$ be a UMD Banach space, $M:\mathbb R_+ \times \Omega \to X$ be a purely discontinuous $L^p$-martingale. Then there exist unique purely discontinuous martingales $M^a,M^q:\mathbb R_+ \times \Omega \to X$ such that $M^a$ has accessible jumps, $M^q$ is quasi-left continuous, $M^q_0=0$ and $M = M^a + M^q$. Moreover, if this is the case, then for $i\in \{a,q\}$
\begin{equation}\label{eq:thm:YoeurpdecUMDaq}
(\mathbb E\|M^i_{\infty}\|^p)^{\frac 1p}\leq \beta_{p,X}(\mathbb E\|M_{\infty}\|^p)^{\frac 1p}.
\end{equation}
\end{theorem}
\begin{proof}
{\em Step 1: finite dimensional case.} First assume that $X$ is finite dimensional. Then $M^a$ and $M^q$ exist and unique due to coordinate-wise applying of Theorem~\ref{thm:Yoeurpdec}.
Let $M = M^a + M^q$, $N= M^a$. Then for any $x^* \in X^*$, $t\geq 0$ by Theorem~\ref{thm:Yoeurpdec} and Lemma~\ref{lem:weakdefforaccjumpsandqlc}~a.s.
\begin{align*}
[\langle M, x^*\rangle]_t = [\langle M, x^*\rangle]^a_t + [\langle M, x^*\rangle]^q_t = [\langle M^a, x^*\rangle]_t + [\langle M^q, x^*\rangle]_t,
\end{align*}
and
\[
[\langle N, x^*\rangle]_t = [\langle N, x^*\rangle]^a_t + [\langle N, x^*\rangle]^q_t = [\langle M^a, x^*\rangle]_t .
\]
Therefore a.s.
\[
[\langle N, x^*\rangle]_t -[\langle N, x^*\rangle]_s \leq [\langle M, x^*\rangle]_t -[\langle M, x^*\rangle]_s,\;\;\; 0\leq s<t.
\]
Moreover $M_0 = N_0$.
Hence $N$ is weakly differentially subordinated to $M$ (see Section \ref{sec:weakdifsubandgenmart}), and \eqref{eq:thm:YoeurpdecUMDaq} for $i=a$ follows from \cite{Y17FourUMD}. By the same reason and since $M^q_0 = 0$, \eqref{eq:thm:YoeurpdecUMDaq} holds true for $i=q$.
{\em Step 2: general case.}
Now let $X$ be general. Let $\xi = M_{\infty}$. Without loss of generality we set $\mathcal F_{\infty} = \mathcal F_t$. Let $(\xi_n)_{n\geq 1}$ be a sequence of simple $\mathcal F_t$-measurable functions in $L^p(\Omega; X)$ such that $\xi_n \to \xi$ as $n\to \infty$ in $L^p(\Omega; X)$. For each $n\geq 1$ define $\mathcal F_t$-measurable $\xi_n^d$ and $\xi_n^c$ such that $M^{d,n} = (\mathbb E(\xi_n^d|\mathcal F_{s}))_{s\geq 0}$ and $M^{c,n}=(\mathbb E(\xi_n^c|\mathcal F_{s}))_{s\geq 0}$ are respectively purely discontinuous and continuous parts of a martingale $(\mathbb E(\xi_n|\mathcal F_{s}))_{s\geq 0}$ as in Remark \ref{rem:YoeMey}. Then thanks to Theorem \ref{thm:Meyer-Yoeurp}, $\xi^d_n \to \xi$ and $\xi_n^c \to 0$ in $L^p(\Omega; X)$ as $n\to \infty$ since $M$ is purely discontinuous.
Since for each $n\geq 1$ the random variable $\xi^d_n$ takes its values in a finite dimensional space, by Theorem \ref{thm:Yoeurpdec} there exist $\mathcal F_t$-measurable $\xi^a, \xi^q\in L^p(\Omega; X)$ such that purely discontinuous martingales $M^{a,n}=(\mathbb E(\xi_n^a|\mathcal F_{s}))_{s\geq 0}$ and $M^{q,n}=(\mathbb E(\xi_n^q|\mathcal F_{s}))_{s\geq 0}$ are respectively with accessible jumps and quasi-left continuous, $\mathbb E(\xi_n^q|\mathcal F_{0})=0$, and the decomposition
$M^{d,n} = M^{a, n} + M^{q,n}$
is as in Theorem~\ref{thm:Yoeurpdec}. Since $(\xi^d_n)_{n\geq 1}$ is a Cauchy sequence in $L^p(\Omega; X)$, by Step 1 both $(\xi^a_n)_{n\geq 1}$ and $(\xi^q_n)_{n\geq 1}$ are Cauchy in $L^p(\Omega; X)$ as well. Let $\xi^a$ and $\xi^q$ be their limits. Define martingales $M^a, M^q:\mathbb R_+ \times \Omega \to X$ in the following way:
\begin{align*}
M^a_s:= \mathbb E(\xi^a|\mathcal F_s),\;\; M^q_s:=\mathbb E(\xi^q|\mathcal F_s),\;\;\; s\geq 0.
\end{align*}
By Proposition \ref{prop:M^pqandM^paareclosed} $M^a$ is a martingale with accessible jumps, $M^q$ is quasi-left continuous, $M^q_0 = 0$ a.s., and therefore $M = M^a + M^q$ is the desired decomposition. Moreover, by Step 1 for each $n\geq 1$ and $i\in \{a,q\}$,
$(\mathbb E \|\xi^i_n\|^p)^{\frac 1p} \leq \beta_{p, X} (\mathbb E \|\xi^d_n\|^p)^{\frac 1p}$,
and hence the estimate \eqref{eq:thm:YoeurpdecUMDaq} follows by letting $n$ to infinity.
The uniqueness of the decomposition follows from Lemma \ref{lem:accjumps+qlc=cons}.
\end{proof}
The following theorem, as Theorem \ref{thm:exampleforlowboundofA}, illustrates that the decomposition in Theorem \ref{thm:YoeurpdecUMD} takes place only in the UMD space case.
\begin{theorem}\label{thm:exampleforlowboundofA^qA^a}
Let $X$ be a finite dimensional Banach space, $p\in(1,\infty)$, $\delta\in\bigl(0, \frac{\beta_{p, X}-1}{2}\bigr)$. Then there exist purely discontinuous martingales $M^a, M^q:\mathbb R_+ \times \Omega \to X$ such that $M^a$ has accessible jumps, $M^q$ is quasi-left continuous, $\mathbb E \|M^a_{\infty}\|^p$, $\mathbb E \|M^q_{\infty}\|^p<\infty$, $M^a_0 = M^q_0=0$, and for $M = M^a + M^q$ and $i\in \{a,q\}$ the following holds
\begin{equation}\label{eq:exampleforlowboundofA^aA^q}
(\mathbb E \|M^i_{\infty}\|^p)^{\frac 1p} \geq \Bigl(\frac{\beta_{p, X}-1}{2} - \delta \Bigr) (\mathbb E \|M_{\infty}\|^p)^{\frac 1p}.
\end{equation}
\end{theorem}
For the proof we will need the following lemma.
\begin{lemma}\label{lem:approxRadbyqlc}
Let $\varepsilon\in \bigl(0,\frac 12\bigr)$, $p\in (1,\infty)$. Then there exist martingales $M, M^a, M^q:[0,1] \times \Omega \to [-1-\varepsilon, 1+\varepsilon]$ with symmetric distributions such that $M^a$ is a martingale with accessible jumps, $\|M^a_1\|_{L^p(\Omega)} <\varepsilon$, $M^q$ is a quasi-left continuous martingale, $M^q_0=0$ a.s., $M = M^a+M^q$, $\mathrm{sign}\text{ } M_1$ is a Rademacher random variable and
\begin{equation}\label{eq:approxRadbyqlc}
\|M_1 - \mathrm{sign}\text{ } M_1\|_{L^p(\Omega)}< \varepsilon.
\end{equation}
\end{lemma}
\begin{proof}
Let $N^+, N^-:[0,1]\times \Omega \to \mathbb R$ be independent Poisson processes with the same intensity $\lambda_{\varepsilon}$ such that $\mathbb P(N^+_1=0)=\mathbb P(N^-_1=0) <\frac {\varepsilon^p} {2^p}$ (such $\lambda_{\varepsilon}$ exists since $N^+_1$ and $N^-_1$ have Poisson distributions, see \cite{KingPois}). Define a stopping time $\tau$ in the following way:
\[
\tau = \inf\{t: N^+_t \geq 1\} \wedge\inf\{t: N^-_t \geq 1\} \wedge 1.
\]
Let $M^q_t := N^+_{t\wedge \tau} - N^-_{t\wedge \tau}$, $t\in [0,1]$. Then $M^q$ is quasi-left continuous with a symmetric distribution. Let $r$ be an independent Rademacher variable, $M^a_t =\frac {\varepsilon}{2} r$ for each $t\in [0,1]$. Then $M^a$ is a martingale with accessible jumps and symmetric distribution, and $\|M^a_1\|_{L^p(\Omega)} = \frac {\varepsilon}{2}< \varepsilon$. Let $M = M^a + M^q$. Then a.s.
\begin{equation}\label{eq:lem:approxRadbyqlc}
M_1 \in \Bigl\{-1-\frac{\varepsilon}{2}, -1+\frac{\varepsilon}{2}, -\frac{\varepsilon}{2},\frac{\varepsilon}{2}, 1-\frac{\varepsilon}{2}, 1+\frac{\varepsilon}{2}\Bigr\},
\end{equation}
so $\mathbb P(M_1=0)=0$, and therefore $\mathrm{sign}\text{ } M_1$ is a Rademacher random variable. Let us prove \eqref{eq:approxRadbyqlc}. Notice that due to \eqref{eq:lem:approxRadbyqlc} if $|M^q_1| = 1$, then $ |M_1 - \mathrm{sign}\text{ } M_1| < \frac {\varepsilon}{2}$, and if $|M^q_1| = 0$, then $ |M_1 - \mathrm{sign}\text{ } M_1| <1$. Therefore
\begin{align*}
\mathbb E |M_1 - \mathrm{sign}\text{ } M_1|^p &= \mathbb E |M_1 - \mathrm{sign}\text{ } M_1|^p \mathbf 1_{|M^q_1| = 1} + \mathbb E |M_1 - \mathrm{sign}\text{ } M_1|^p \mathbf 1_{|M^q_1| = 0}\\
&< \frac {\varepsilon^p} {2^p} + \frac {\varepsilon^p} {2^p} < \varepsilon^p,
\end{align*}
so \eqref{eq:approxRadbyqlc} holds.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:exampleforlowboundofA^qA^a}]
The proof is analogous to the proof of Theorem \ref{thm:exampleforlowboundofA}, while one has to use Lemma \ref{lem:approxRadbyqlc} instead of Lemma \ref{lem:contapproxradem}.
\end{proof}
Theorem \ref{thm:exampleforlowboundofA^qA^a} yields the following characterization of the UMD property.
\begin{theorem}\label{thm:M^aM^qM^c}
Let $X$ be a Banach space. Then $X$ is a UMD Banach space if and only if for some (equivalently, for all) $p\in(1,\infty)$ there exists $c_{p,X} >0$ such that for any $L^p$-martingale $M:=\mathbb R_+ \times \Omega \to X$ there exist unique martingales $M^c, M^q, M^a:\mathbb R_+ \times \Omega \to X$ such that $M^c_0=M^q_0=0$, $M^c$ is continuous, $M^q$ is purely discontinuous quasi-left continuous, $M^a$ is purely discontinuous with accessible jumps, $M=M^c + M^q + M^a$, and
\begin{equation}\label{eq:coriffXUMDcqlcaj}
(\mathbb E \|M^c_{\infty}\|^p)^{\frac 1p}+ (\mathbb E \|M^q_{\infty}\|^p)^{\frac 1p} + (\mathbb E \|M^a_{\infty}\|^p)^{\frac 1p} \leq c_{p,X}(\mathbb E \|M_{\infty}\|^p)^{\frac 1p}.
\end{equation}
If this is the case, then the least admissible $c_{p,X}$ is in the interval $\bigl[\frac{3\beta_{p, X}\!-\!3}{2}\vee 1,3\beta_{p, X}\bigr]$.
\end{theorem}
The decomposition $M =M^c + M^q + M^a$ is called the {\em canonical decomposition} of the martingale $M$ (see \cite{Kal,Yoe76,DY17}).
\begin{proof}
The ``if and only if'' part follows from Theorem \ref{thm:charofUMDbyMeyYoe}, Theorem \ref{thm:YoeurpdecUMD} and Theorem \ref{thm:exampleforlowboundofA^qA^a}. The estimate $c_{p,X} \leq 3\beta_{p, X}$ follows from \eqref{eq:Meyer-Yoeurp} and \eqref{eq:thm:YoeurpdecUMDaq}. The estimate $c_{p,X} \geq \frac{3\beta_{p, X}-3}{2}\!\vee\!1$ follows from \eqref{eq:exampleforlowboundofA} and \eqref{eq:exampleforlowboundofA^aA^q}.
\end{proof}
\begin{corollary}
Let $X$ be a Banach space. Then $X$ is a UMD Banach space if and only if $\mathcal M_X^{p, d} = \mathcal M_X^{p,a} \oplus\mathcal M_X^{p,q}$ and $\mathcal M_X^p = \mathcal M_X^{p,c}\oplus \mathcal M_X^{p,q}\oplus \mathcal M_X^{p,a}$ for any filtration that satisfies the usual conditions.
\end{corollary}
\begin{proof}
The corollary follows from Theorem \ref{thm:YoeurpdecUMD}, Theorem \ref{thm:exampleforlowboundofA^qA^a} and Theorem \ref{thm:M^aM^qM^c}.
\end{proof}
\subsection{Stochastic integration}
The current subsection is devoted to application of Theorem \ref{thm:M^aM^qM^c} to stochastic integration with respect to a general martingale.
\begin{proposition*}\label{prop:intpreserves}
Let $H$ be a Hilbert space, $X$ be a Banach space, $M:\mathbb R_+ \times \Omega \to H$ be a martingale, $\Phi:\mathbb R_+\times \Omega \to \mathcal L(H, X)$ be elementary progressive. Then
\begin{itemize}
\item [(i)] if $M$ is continuous, then $\Phi \cdot M$ is continuous;
\item[(ii)] if $M$ is purely discontinuous, then $\Phi\cdot M$ is purely discontinuous;
\item[(iii)] if $M$ has accessible jumps, then $\Phi\cdot M$ has accessible jumps;
\item[(iv)] if $M$ is quasi-left continuous, then $\Phi\cdot M$ is quasi-left continuous.
\end{itemize}
\end{proposition*}
\begin{proposition}\label{prop:decHilSpacase}
Let $H$ be a Hilbert space, $M:\mathbb R_+ \times \Omega \to H$ be a local martingale. Then there exist unique martingales $M^c, M^q, M^a:\mathbb R_+ \times \Omega \to H$ such that $M^c$~is continuous, $M^q$ and $M^a$ are purely discontinuous, $M^q$ is quasi-left continuous, $M^a$~has accessible jumps, $M^c_0 = M^q_0=0$ a.s., and $M = M^c + M^q + M^a$.
\end{proposition}
\begin{proof}
Analogously to Theorem 26.14 and Corollary 26.16 in \cite{Kal}.
\end{proof}
\begin{theorem}\label{thm:stochintgenUMDcase}
Let $H$ be a Hilbert space, $X$ be a UMD Banach space, $p\in (1,\infty)$, $M:\mathbb R_+ \times \Omega \to H$ be a local martingale, $\Phi:\mathbb R_+ \times \Omega \to \mathcal L(H, X)$ be elementary progressive. Let $M=M^c + M^q + M^a$ be the canonical decomposition from Proposition \ref{prop:decHilSpacase}. Then
\begin{equation}\label{eq:stochintgenUMDcase}
\mathbb E\|(\Phi \cdot M)_{\infty}\|^p \eqsim_{p, X}\mathbb E\|(\Phi \cdot M^c)_{\infty}\|^p+ \mathbb E\|(\Phi \cdot M^q)_{\infty}\|^p + \mathbb E\|(\Phi \cdot M^a)_{\infty}\|^p .
\end{equation}
and if $(\Phi \cdot M)_{\infty} \in L^p(\Omega; X)$, then $\Phi \cdot M= \Phi \cdot M^c + \Phi \cdot M^q + \Phi \cdot M^a$ is the canonical decomposition from Theorem \ref{thm:M^aM^qM^c}.
\end{theorem}
\begin{proof}
The statement that $\Phi \cdot M= \Phi \cdot M^c + \Phi \cdot M^q + \Phi \cdot M^a$ is the canonical decomposition follows from Proposition \ref{prop:intpreserves}, Theorem \ref{thm:M^aM^qM^c} and the fact that a.s.\ $(\Phi \cdot M)_0 = (\Phi \cdot M^c)_0 = (\Phi \cdot M^q)_0=0$. \eqref{eq:stochintgenUMDcase} follows then from \eqref{eq:coriffXUMDcqlcaj} and the triangle inequality.
\end{proof}
\begin{remark}
Notice that the It\^o isomorphism for the term $\Phi \cdot M^c$ from \eqref{eq:stochintgenUMDcase} was explored in \cite{VY2016}. It remains open what to do with the other two terms, but positive results in this direction were obtained in the case of $X = L^q(S)$ in \cite{DY17}.
\end{remark}
\section{Weak differential subordination and general martingales}\label{sec:weakdifsubandgenmart}
This subsection is devoted to the generalization of the main theorem in work \cite{Y17FourUMD}. Namely, here we show the $L^p$-estimates for general $X$-valued weakly differentially subordinated martingales.
\begin{definition}
Let $X$ be a Banach space, $M, N:\mathbb R_+ \times \Omega \to X$ be local martingales. Then $N$ is {\em weakly differentially subordinated} to $M$ if $[\langle M, x^* \rangle] - [\langle N, x^* \rangle]$ is an increasing process a.s.\ for each $x^* \in X^*$.
\end{definition}
The following theorem have been proven in \cite{Y17FourUMD}.
\begin{theorem}\label{thm:ewakdiffsubUMDpurdisc}
Let $X$ be a Banach space. Then $X$ has the UMD property if and only if for some (equivalently, for all) $p\in (1, \infty)$ there exists $\beta>0$ such that for each pair of purely discontinuous martingales $M, N:\mathbb R_+ \times \Omega\to X$ such that $N$ is weakly differentially subordinated to $M$ one has that
\[
(\mathbb E \|N_{\infty}\|^p)^{\frac 1p} \leq \beta (\mathbb E \|M_{\infty}\|^p)^{\frac 1p}.
\]
If this is the case, then the least admissible $\beta$ is the UMD constant $\beta_{p, X}$.
\end{theorem}
The main goal of the current section is to prove the following generalization of Theorem~\ref{thm:ewakdiffsubUMDpurdisc} to the case of arbitrary martingales.
\begin{theorem}\label{thm:weakdiffsubprocmaingen}
Let $X$ be a UMD Banach space, $M, N:\mathbb R_+ \times \Omega \to X$ be two martingales such that $N$ is weakly differentially subordinated to $M$. Then for each $p\in (1,\infty)$, $t\geq 0$,
\begin{equation}\label{eq:weakdiffsubprocmaingen}
(\mathbb E \|N_t\|^p)^{\frac 1p} \leq \beta_{p,X}^2 (\beta_{p,X}+1)(\mathbb E \|M_t\|^p)^{\frac 1p}.
\end{equation}
\end{theorem}
The proof will be done in several steps. First we show an analogue of Theorem~\ref{thm:ewakdiffsubUMDpurdisc} for continuous martingales.
\begin{theorem*}\label{thm:weakdiffsubproccont}
Let $X$ be a Banach space. Then $X$ is a UMD Banach space if and only if for some (equivalently, for all) $p\in (1,\infty)$ there exists $c>0$ such that for any continuous martingales $M, N:\mathbb R_+ \times \Omega \to X$ such that $N$ is weakly differentially subordinated to $M$, $M_0 = N_0 = 0$, one has that
\begin{equation}\label{eq:weakdiffsubprocmaingencont}
(\mathbb E \|N_{\infty}\|^p)^{\frac 1p} \leq c_{p,X}(\mathbb E \|M_{\infty}\|^p)^{\frac 1p}.
\end{equation}
If this is the case, then the least admissible $c_{p,X}$ is in the segment $[\beta_{p, X}, \beta_{p, X}^2]$.
\end{theorem*}
For the proof we will need the following proposition, which demonstrates that one needs a slightly weaker assumption rather then in Theorem \ref{thm:weakdiffsubproccont} so that the estimate \eqref{eq:weakdiffsubprocmaingencont} holds in a UMD Banach space.
\begin{proposition}\label{prop:weakversofweak}
Let $X$ be a UMD Banach space, $1<p<\infty$, $M,N:\mathbb R_+ \times \Omega \to X$ be continuous $L^p$-martingales s.t.\ $M_0=N_0=0$ and for each $x^* \in X^*$ a.s.\ for each $t\geq 0$
\begin{equation}\label{eq:weakversofweak1}
[\langle N, x^*\rangle]_t \leq [\langle M, x^*\rangle]_t.
\end{equation}
Then for each $t\geq 0$
\begin{equation}\label{eq:weakversofweak2}
(\mathbb E \|N_t\|^p)^{\frac 1p} \leq \beta_{p, X}^2(\mathbb E \|M_t\|^p)^{\frac 1p}.
\end{equation}
\end{proposition}
\begin{proof}
Without loss of generality by a stopping time argument we assume that $M$ and $N$ are bounded and that $M_{\infty} = M_t$ and $N_{\infty} = N_t$.
One can also restrict to a finite dimensional case. Indeed, since $X$ is a
separable reflexive space, $X^*$ is separable as well. Let $(Y_m)_{m\geq 1}$ be
an increasing sequence of finite-dimensional subspaces of $X^*$ such that
$\overline{\bigcup_mY_m}=X^*$ and $\|\cdot\|_{Y_m} = \|\cdot\|_{X^*|_{Y_m}}$ for each
$m\geq 1$. Then for each fixed $m\geq 1$ there exists a linear operator
$P_m:X\to Y_m^*$ of norm $1$ defined as follows: $\langle P_mx, y\rangle =
\langle x,y\rangle$ for each $x\in X, y\in Y_m$. Therefore $P_mM$ and $P_m N$ are $Y_m^*$-valued martingales. Moreover, \eqref{eq:weakversofweak1} holds for $P_m M$ and $P_m N$ since there exists $P_m^*:Y_m \to X^*$, and for each $y\in Y_m$ we have that $\langle P_m M, y\rangle = \langle M, P_m y\rangle $ and $\langle P_m N, y\rangle = \langle N, P_m y\rangle$. Since $Y_m$ is a closed
subspace of $X^*$, \cite[Proposition 4.2.17]{HNVW1} yields $\beta_{p',Y_m}\leq
\beta_{p',X^*}$, consequently again by \cite[Proposition 4.2.17]{HNVW1}
$\beta_{p,Y_m^*}\leq \beta_{p,X^{**}} = \beta_{p,X}$. So if we prove the finite dimensional version, then
\[
(\mathbb E \|P_m N_t\|^p)^{\frac 1p} \leq \beta_{p, Y_m^*}^2(\mathbb E \|P_m M_t\|^p)^{\frac 1p}\leq \beta_{p, X}^2(\mathbb E \|P_m M_t\|^p)^{\frac 1p},
\]
and \eqref{eq:weakversofweak2} with $c_{p,X} = \beta_{p, X}^2$ will follow by letting $m\to \infty$.
Let $d$ be the dimension of $X$, $\vertiii{\cdot}$ be a Euclidean norm on $X\times X$. Let $L= (M,N):\mathbb R_+ \times \Omega \to X\times X$ be a continuous martingale. Since $(X\times X, \vertiii{\cdot})$ is a Hilbert space, $L$ has a continuous quadratic variation $[L]:\mathbb R_+ \times \Omega \to \mathbb R_+$ (see Remark \ref{rem:qvcont}). Let $A:\mathbb R_+ \times \Omega \to \mathbb R_+$ be such that $A_s = [L]_s+s$ for each $s\geq 0$.
Then $A$ is continuous strictly increasing predictable. Define a random time-change $(\tau_s)_{s\geq 0}$ as in Theorem~\ref{thm:apptimechange}.
Let $\mathbb G = (\mathcal G_s)_{s \geq 0} = (\mathcal F_{\tau_s})_{s\geq 0}$ be the induced filtration. Then thanks to the Kazamaki theorem \cite[Theorem 17.24]{Kal} $\widetilde L = L \circ \tau$ is a $G$-martingale, and $[\widetilde L] = [L]\circ \tau$. Notice that $\widetilde L = (\widetilde M, \widetilde N)$ with $\widetilde M = M\circ \tau$, $\widetilde N = N \circ \tau$, and since by Kazamaki theorem \cite[Theorem 17.24]{Kal} $[M\circ \tau] = [M]\circ \tau$, $[N\circ \tau] = [N]\circ \tau$, and $(M\circ \tau)_0 = (N\circ \tau)_0=0$, we have that by \eqref{eq:weakversofweak1} for each $x^*\in X^*$ a.s.\ for each $s\geq 0$
\begin{equation}\label{eq:weakversofweak3}
[\langle \widetilde N, x^*\rangle]_s = [\langle N, x^*\rangle]_{\tau_s} \leq [\langle M, x^*\rangle]_{\tau_s} = [\langle \widetilde M, x^*\rangle]_s
\end{equation}
Moreover, for all $0\leq u<s$ we have that a.s.
\begin{align*}
[\widetilde L]_s - [\widetilde L]_u = ([L]\circ \tau)_s - ([L]\circ \tau)_u&\leq ([L]\circ \tau)_s + \tau_s - ([L]\circ \tau)_u - \tau_u\\
&= ([L]_{\tau_s} + \tau_s) - ([L]_{\tau_u} + \tau_u) = s-u.
\end{align*}
Therefore $[\widetilde L]$ is a.s.\ absolutely continuous with respect to the Lebesgue measure on~$\mathbb R_+$. Consequently, due to Theorem \ref{thm:Brrepres}, there exists an enlarged probability space $(\widetilde {\Omega}, \widetilde {\mathcal F}, \widetilde {\mathbb P})$ with an enlarged filtration $\widetilde {\mathbb G} = (\widetilde{\mathcal G}_s)_{s \geq 0}$, a $2d$-dimensional standard Wiener process $W$, which is defined on~$\widetilde{\mathbb G}$, and a stochastically integrable progressively measurable function $f:\mathbb R_+ \times\widetilde{\Omega} \to \mathcal L(\mathbb R^{2d}, X\times X)$ such that $\widetilde L = f\cdot W$. Let $f^M, f^N:\mathbb R_+ \times \Omega \to \mathcal L(\mathbb R^{2d},X)$ be such that $f = (f^M, f^N)$. Then $\widetilde M = f^M\cdot W$ and $\widetilde N = f^N\cdot W$. Let $(\overline{\Omega}, \overline{\mathcal F}, \overline{\mathbb P})$ be an independent probability space with a filtration~$\overline{\mathbb G}$ and a $2d$-dimensional Wiener process $\overline W$ on it. Denote by $\overline {\mathbb E}$ the expectation on $(\overline{\Omega}, \overline{\mathcal F}, \overline{\mathbb P})$. Then because of the decoupling theorem \cite[Theorem 4.4.1]{HNVW1}, for each~$s\geq 0$
\begin{equation}\label{eq:weakdiffsubproccontdecopl}
\begin{split}
(\mathbb E \|\widetilde N_s\|^p)^{\frac 1p} = (\mathbb E \|(f^N\cdot W)_s\|^p)^{\frac 1p}\leq \beta_{p,X}(\mathbb E\, \overline{\mathbb E} \|(f^N\cdot \overline W)_s\|^p)^{\frac 1p},\\
\frac{1}{\beta_{p,X}}(\mathbb E\, \overline{\mathbb E} \|(f^M\cdot \overline W)_s\|^p)^{\frac 1p} \leq (\mathbb E \|(f^M\cdot W)_s\|^p)^{\frac 1p} = (\mathbb E \|\widetilde M_s\|^p)^{\frac 1p}.
\end{split}
\end{equation}
Due to the multidimensional version of \cite[Theorem 17.11]{Kal} and \eqref{eq:weakversofweak3} for each $x^* \in X^*$ we have that
\begin{equation}\label{eq:weakdiffsubproccontomega_0}
s\mapsto [\langle \widetilde M, x^*\rangle]_s - [\langle \widetilde N, x^*\rangle]_s = \int_0^s(|\langle x^*, f^M(r)\rangle|^2 - |\langle x^*, f^N(r)\rangle|^2)\ud r
\end{equation}
is nonnegative and absolutely continuous a.s. Since $X$ is separable, we can fix a set $\widetilde {\Omega}_0\subset \widetilde{\Omega}$ of full measure on which the function \eqref{eq:weakdiffsubproccontomega_0} is nonnegative for each $s\geq 0$.
Now fix $\omega \in \widetilde {\Omega}_0$ and $s\geq 0$. Let us prove that
$$
\overline{\mathbb E} \|(f^N(\omega)\cdot \overline W)_s\|^p\leq \overline{\mathbb E} \|(f^M(\omega)\cdot \overline W)_s\|^p.
$$
Since $f^M(\omega)$ and $f^N(\omega)$ are deterministic on $\overline {\Omega}$, and since due to \eqref{eq:weakdiffsubproccontomega_0} for each $x^* \in X^*$
\begin{align*}
\overline{\mathbb E}|\langle (f^N(\omega)\cdot \overline W)_s, x^*\rangle|^2 &= \int_0^s|\langle x^*, f^N(r,\omega)\rangle|^2\ud r\\
&\leq \int_0^s|\langle x^*, f^M(r,\omega)\rangle|^2\ud r = \overline{\mathbb E}|\langle (f^M(\omega)\cdot \overline W)_s, x^*\rangle|^2,
\end{align*}
by \cite[Corollary 4.4]{NW1} we have that $\overline{\mathbb E} \|(f^N(\omega)\cdot \overline W)_s\|^p\leq \overline{\mathbb E} \|(f^M(\omega)\cdot \overline W)_s\|^p$. Consequently, due to~\eqref{eq:weakdiffsubproccontdecopl} and the fact that $\widetilde{\mathbb P}(\Omega_0)=1$
\begin{align*}
(\mathbb E \|\widetilde N_s\|^p)^{\frac 1p} \leq \beta_{p,X}(\mathbb E\, \overline{\mathbb E} \|(f^N\cdot \overline W)_s\|^p)^{\frac 1p}&\leq \beta_{p,X}(\mathbb E\, \overline{\mathbb E} \|(f^M\cdot \overline W)_s\|^p)^{\frac 1p}\leq \beta_{p,X}^2(\mathbb E \|\widetilde M_s\|^p)^{\frac 1p}.
\end{align*}
Recall that $\widetilde M$ and $\widetilde N$ are bounded, so thanks to the dominated convergence theorem one gets \eqref{eq:weakversofweak2} with $c_{p,X} = \beta_{p, X}^2$ by letting $s$ to infinity.
\end{proof}
\begin{proof}[Proof of Theorem \ref{thm:weakdiffsubproccont}]
{\em The ``only if'' part \& the upper bound of $c_{p,X}$:} The ``only if'' part and the estimate $c_{p,X}\leq \beta_{p, X}^2$ follows from Proposition \ref{prop:weakversofweak} since \eqref{eq:weakversofweak1} holds for $M$ and $N$ because $N$ is weakly differentially subordinated to $M$.
{\em The ``if'' part \& the lower bound of $c_{p,X}$:} {\em See the supplement \cite{Y17MDSupp}.}
\end{proof}
\begin{remark}\label{rem:Hiltranfweakdiffsubcont}
Let $X$ be a Banach space. Then according to \cite{Bour83,Burk83,Gar85} the Hilbert transform $\mathcal H_X$ can be extended to $L^p(\mathbb R; X)$ for each $1<p<\infty$ if and only if $X$ is a UMD Banach space. Moreover, if this is the case, then
\[
\sqrt{\beta_{p, X}}\leq\|\mathcal H_X\|_{\mathcal L(L^p(\mathbb R; X))}\leq \beta_{p, X}^2.
\]
As it was shown in \cite{Y17FourUMD}, the upper bound $\beta_{p, X}^2$ can be also directly derived from the upper bound for $c_{p,X}$ in Theorem \ref{thm:weakdiffsubproccont}. The sharp upper bound for $\|\mathcal H_X\|_{\mathcal L(L^p(\mathbb R; X))}$ remains an open question (see \cite[pp. 496-497]{HNVW1}), so the sharp upper bound for $c_{p,X}$ is of interest.
\end{remark}
\begin{lemma*}\label{lemma:weakdiffsubdecomp}
Let $X$ be a Banach space, $M^c,N^c:\mathbb R_+ \times \Omega \to X$ be continuous martingales, $M^d, N^d:\mathbb R_+ \times \Omega \to X$ be purely discontinuous martingales, $M^c_0 = N^c_0=0$. Let $M := M^c +M^d$, $N:= N^c + N^d$. Suppose that $N$ is weakly differentially subordinated to $M$. Then $N^c$ is weakly differentially subordinated to $M^c$, and $N^d$ is weakly differentially subordinated to $M^d$.
\end{lemma*}
\begin{proof}[Proof of Theorem \ref{thm:weakdiffsubprocmaingen}]
By Theorem \ref{thm:Meyer-Yoeurp} there exist martingales $M^d, M^c, N^d, N^c:\mathbb R_+ \times \Omega \to X$ such that $M^d$ and $N^d$ are purely discontinuous, $M^c$ and $N^c$ are continuous, $M^c_0 = N^c_0=0$, and $M = M^d+ M^c$ and $N = N^d+ N^c$. By Lemma~\ref{lemma:weakdiffsubdecomp}, $N^d$ is weakly differentially subordinated to $M^d$ and $N^c$ is weakly differentially subordinated to $M^c$. Therefore for each $t\geq 0$
\begin{align*}
(\mathbb E \|N_t\|^p)^{\frac 1p} \stackrel{(i)}\leq (\mathbb E \|N^d_t\|^p)^{\frac 1p} + (\mathbb E \|N^c_t\|^p)^{\frac 1p}
&\stackrel{(ii)} \leq \beta_{p, X}^2(\mathbb E \|M^d_t\|^p)^{\frac 1p} + \beta_{p, X}(\mathbb E \|M^c_t\|^p)^{\frac 1p}\\
& \stackrel{(iii)}\leq\beta_{p, X}^2(\beta_{p,X}+1)(\mathbb E \|M_t\|^p)^{\frac 1p},
\end{align*}
where $(i)$ holds thanks to the triangle inequality, $(ii)$ follows from Theorem \ref{thm:ewakdiffsubUMDpurdisc} and Theorem \ref{thm:weakdiffsubproccont}, and $(iii)$ follows from \eqref{eq:Meyer-Yoeurp}.
\end{proof}
\begin{remark}
It is worth noticing that in a view of recent results the sharp constant in \eqref{eq:Meyer-Yoeurp} and \eqref{eq:thm:YoeurpdecUMDaq} can be derived and equals the {\em UMD$^{\{0,1\}}_p$ constant} $\beta_{p, X}^{\{0,1\}}$. In order to show that this is the right upper bound one needs to use a {\em $\{0,1\}$-Burkholder function} instead of the Burkholder function, while the sharpness follows analogously Theorem \ref{thm:exampleforlowboundofA} and \ref{thm:exampleforlowboundofA^qA^a}. See \cite{Y17UMD^A} for details.
\end{remark}
\begin{remark}
In the recent paper \cite{Y17GMY} the existence of the canonical decomposition of a general local martingale together with the corresponding weak $L^1$-estimates were shown. Again existence of the canonical decomposition of any $X$-valued martingale is equivalent to $X$ having the UMD property.
\end{remark}
\section*{Acknowledgements} The author would like to thank Mark Veraar for helpful comments; in particular for showing him \cite[Corollary 4.4]{NW1}. The author thanks Jan van Neerven for careful reading of parts of this article and useful suggestions. The author thanks the anonymous referee for his/her valuable comments.
\bibliographystyle{plain}
\def$'${$'$} \def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth
\lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}}
\def\polhk#1{\setbox0=\hbox{#1}{\ooalign{\hidewidth
\lower1.5ex\hbox{`}\hidewidth\crcr\unhbox0}}} \def$'${$'$}
| c735f846dc98ab635c4224f08f6e0115416fc5aa | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction
\label{sec:introduction}}
It is known that in the minimal supersymmetric standard model~(MSSM), the baryon asymmetry in our universe can be generated by Affleck-Dine mechanism~\cite{ad,add}, which produces a scalar field condensate with baryon number. In many models such as the gauge mediated SUSY breaking models, the spatial inhomogeneities of this condensate due to quantum fluctuations grow and fragment into non-topological solitons called Q-balls~\cite{ks,kkins,kkins2}, which are defined as spherical solutions in a global $U(1)$ theory which minimize the energy of the system with a fixed $U(1)$ charge~\cite{c}. In this case, the baryon number generated in Affleck-Dine mechanism is confined inside Q-balls, so that the baryon asymmetry in the universe is generated by baryons emitted by the decay of the Q-balls.
The Q-ball decay into other particles was first studied by Cohen {\it et al.}~\cite{coh}, who considered the Yukawa theory and calculated the neutrino pair production rate by leading semi-classical approximation treating the Q-ball as the classical background scalar field, where the Q-ball configuration was approximated as a step function. More realistic configurations are considered in Refs.~\cite{nur,ym}. In particular, the production rates of quarks and gravitinos from the Q-balls in supersymmetric theories were derived in Ref.~\cite{ym}, which can be used to estimate baryon-to-dark matter ratio in gauge mediated SUSY breaking models, where the gravitino is dark matter.
While the Q-ball is a non-topological soliton resulting from global $U(1)$ symmetry, whose generalization to local $U(1)$ symmetry was also proposed. The lowest-energy configuration with a fixed local $U(1)$ charge is called gauged Q-ball~\cite{gaugedqball}, which consists not only of the scalar field, but also of the $U(1)$ gauge field. The properties of gauged Q-ball solutions have been studied analytically and numerically in the literature~\cite{gaugedqball,5,7,8,9,10,11,12}, but their decay into other particles has not been considered. In this paper, we derive the decay rate of the gauged Q-ball into fermions, applying the leading semi-classical approximation used in Ref.~\cite{coh} to the gauged Q-ball. We assume that the scalar field in the Q-ball couples to fermions by Yukawa interaction. Since the gauged Q-ball can be interpreted as the electrically charged Q-ball, the decay rate into particles with the charge of the same sign is expected to be enhanced, compared to the case of a global Q-ball, and we show that this is indeed the case. We also show that, on the other hand, the decay rate of the gauged Q-ball is upperly bounded, due to the Pauli blocking at the surface of the Q-ball, just as pointed out in the case of the global Q-ball~\cite{coh}, and even more suppressed due to the Coulomb potential outside the Q-ball, since it plays the role of a potential barrier for the fermions coming from the inside, as we will see later.
The paper is organized as follows. In Sec.~\ref{sec:bp}, we review some basic properties of gauged Q-ball. In Sec.~\ref{sec:dr}, we present the theoretical setup for gauged Q-balls coupled to the massless fermions and calculate the gauged Q-ball decay rate into massless fermions by using the leading semi-classical approximation. Sec.~\ref{sec:conc} is devoted to the conclusions.
\section{Gauged Q-ball}
\label{sec:bp}
We consider a theory of a complex scalar field $\phi$ coupled to a $U(1)$ gauge field $A_\mu$. The Lagrangian density is written as follows.
\begin{align}
\mathcal{L}&=(D_\mu\phi)^\ast D^\mu\phi-V(\phi)-\frac14F_{\mu\nu}F^{\mu\nu},\\
D_\mu&\equiv\partial_\mu+ieA_\mu,
\end{align}
where $V(\phi)$ is a scalar potential and $F_{\mu\nu}=\del_\mu A_\nu-\del_\nu A_\mu$. We introduce the following ansatz on $\phi$,
\begin{align}
&\phi(x,t)\equiv \phi(r)e^{-i\omega t},
\label{eq:gqpara}
\end{align}
which is the same parametrization as that of a global Q-ball.
For the gauge field, we find spatially symmetric solution with no magnetic field, or no electric current:
\begin{align}
&A_0=A_0(r),\\
&A_i=0.
\end{align}
The equations of motion are then given by
\begin{align}
\label{eq:eomf}
&\frac{d^2\phi}{dr^2}+\frac2r\frac{d\phi}{dr}+\phi h^2-\frac{dV}{d\phi}=0,\\
&\frac{d^2h}{dr^2}+\frac2r\frac{dh}{dr}-e^2\phi^2h=0,
\label{eq:eomgaugedqball}
\end{align}
where we redefined the gauge field to absorb $\omega$ as $h\equiv-\omega+eA_0$.
We set boundary conditions as
\begin{align}
\label{eq:bdcga}
&\phi(\infty)=0,~~ \frac{d\phi}{dr}(0)=0,\\
&A_0(\infty)=0,~~\frac{dA_0}{dr}(0)=0,
\end{align}
especially to avoid singularities at $r=0$.
As a scalar potential, we choose a logarithmic potential $V(\phi)=m_\phi^4\ln(1+|\phi|^2/m_\phi^2)$, which is motivated by gauge mediation models. For $e=0$, the solution becomes a global Q-ball, which is called gauge mediation type Q-ball~\cite{dv,ls}.
It is known that this type of Q-balls with sufficiently large charge has the following approximate analytic solution.
\begin{align}
\phi(r)=\left\{
\begin{array}{ll}
\phi_0\sin\omega r/\omega r,
& (r\leq R\equiv\pi/\omega)
\\
\\
0, & (r>R)
\end{array}
\right..
\label{eq:sin}
\end{align}
The angular velocity $\omega$ is equal to $dE/dQ$, which is true for general Q-ball solutions, and has the following charge dependence,
\begin{align}
\omega=\frac{dE}{dQ}&\propto Q^{-1/4},\label{eq:q14}
\end{align}
which will be useful later.
The second derivative of $\phi$ becomes singular at $r=R$, which for actual Q-balls, becomes a peak of $\phi''(r)$. We define the size of a gauged Q-ball, which is the case $e\neq0$, as the point where $\phi'''(r)=0$ as well, even if the profile is somewhat pushed outward by the electric repulsion, as shown in Fig.~\ref{fig:gauex}. Indeed we can see that $\phi''(r=R)$ becomes singular for a large gauged Q-ball, just as for the global Q-ball, even when the Coulomb potential has a non-negligible effect on the profile. Later we consider the case of large gauged Q-balls when we discuss the saturation of fermion production, where specifying the size of a Q-ball becomes important.
\begin{figure}[t]
\begin{minipage}{.5\linewidth}
\includegraphics[width=\linewidth]{prfs.pdf}
\end{minipage}
\hspace{0cm}
\begin{minipage}{.5\linewidth}
\includegraphics[width=\linewidth]{prfl.pdf}
\end{minipage}
\caption{Examples of profile of the gauged Q-ball of gauge mediation type. The dimensionful parameters are in units of $m_\phi$. The dashed line denotes the global Q-ball with the same charge. We see that the profile of $\phi$ is pushed outward due to the electric repulsion. We can also see that $\phi''(r=R)$ becomes singular for a large gauged Q-ball, just as for a global Q-ball, even when the Coulomb potential has a non-negligible effect on the profile.}
\label{fig:gauex}
\end{figure}
The energy and charge of the gauged Q-ball are given by
\begin{align}
E&=\int d^3x\left[\frac12(\nabla\phi)^2+\frac1{2}(\nabla A_0)^2+\frac12\phi^2(\omega-eA_0)^2+V(\phi)\right],\\
Q&=\int d^3 x(\omega-eA_0)\phi^2,
\end{align}
and the relation $\omega=dE/dQ$ holds, just as in the case of a global Q-ball, whose proof is given in Ref.~\cite{7}.
The gauged Q-ball becomes unstable as the charge grows due to electric repulsion, which can be seen by the behavior of $\omega=dE/dQ$. We present the plot of $\omega$ as a function of $Q$ in Fig.~\ref{fig:omegg}~(Left). We can see that in contrast to the behavior of $\omega$ for global Q-balls, which is denoted by a dashed line, $\omega$ increases as the charge grows, which means that the Q-ball becomes unstable\footnote
If the charge grows further, $\omega$ becomes larger than $m_\phi$, which means that the Q-ball becomes unstable against decay into itself, and soon the solutions cease to exist. Here we simply focus on the case $\omega<m_\phi$.}. We also plot Coulomb energy at the surface of Q-ball, $e^2Q/4\pi R$ by a dotted line, whose contribution also becomes large as charge grows. However, we see that the Coulomb energy stays smaller than $\omega$, due to the growth of $\omega$, and also of $R$ by electric repulsion, which we illustrate in Fig.~\ref{fig:omegg}~(Right).
\begin{figure}[t]
\begin{minipage}{.47\linewidth}
\includegraphics[width=\linewidth]{oyf.pdf}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}{.47\linewidth}
\includegraphics[width=\linewidth]{ryf.pdf}
\end{minipage}
\caption{The plots of $\omega=dE/dQ$ and $R$ as functions of $Q$. We plot those for global Q-balls by dashed lines, for comparison. We see that $\omega$ becomes large as the charge grows, which means the Q-ball becomes unstable due to the electric charge, and $R$ becomes large due to the electric repulsion. We also present the Coulomb energy at the surface, $e^2Q/4\pi R$, which is denoted by a dotted line in the left figure.}
\label{fig:omegg}
\end{figure}
\section{Gauged Q-ball decay rates into fermions}
\label{sec:dr}
In this section, we derive the decay rate of the gauged Q-ball into fermions, by using the leading semi-classical approximation, where we treat the gauged Q-ball as the classical background. We calculate the production rate of the fermions in the presence of it.
We consider the following Lagrangian,
\begin{align}
\mathcal{L}_{\text{fermion}}=\chi^\dagger i\bar{\sigma}^\mu(\del_\mu+iq_\chi eA_\mu)\chi+\eta^\dagger i\bar{\sigma}^\mu(\del_\mu+iq_\eta eA_\mu)\eta-(g\phi^\ast\chi\eta+\text{h.c.}),
\end{align}
where $\chi,\eta$ are Weyl fermions, which couple to $\phi$ by Yukawa interaction, and $\bar{\sigma}^\mu=({\bf1},-\sigma^i)$, where $\sigma^i$ are the Pauli matrices. We note that $q_\chi+q_\eta=1$ must be satisfied due to the charge conservation.
Here we simply set $(q_\chi,q_\eta)=(1,0)$, which assigns the same sign of charge to~$\chi$.
The equations of motion are written as
\begin{align}
i\bar{\sigma}^\mu(\del_\mu+iq_\chi eA_\mu)\chi-g\phi\eta^\dagger&=0,\\
i\sigma^\mu(\del_\mu-iq_\eta eA_\mu)\eta^\dagger-g\phi^\ast\chi&=0.
\end{align}
Since $\phi$ is time dependent, the following modes mix with each other.
\begin{align}
\chi&\propto e^{-ik_+t},\\
\eta^\dagger&\propto e^{i(\omega-k_+)t}\equiv e^{ik_-t},
\end{align}
whose equations of motion become
\begin{align}
(k_++iq_\chi eA_0-i\boldsymbol{\sigma}\cdot\nabla)\chi-g\phi(r)\eta^\dagger&=0,\label{eq:em1}\\
(-k_--iq_\eta eA_0+i\boldsymbol{\sigma}\cdot\nabla)\eta^\dagger-g\phi(r)\chi&=0.\label{eq:em2}
\end{align}
First, we consider the case when $\chi,\eta$ are free fields, whose equations of motion are
\begin{align}
(k_+-i\boldsymbol{\sigma}\cdot\nabla)\chi&=0,\\
(-k_-+i\boldsymbol{\sigma}\cdot\nabla)\eta^\dagger&=0.
\end{align}
Then, we can write the following expansion of $\chi,\eta^\dagger$,
\begin{align}
\chi&=\sum_{j,m}\int_0^\infty dk_+\left[a_{\text{in}}(k_+,j,m)e^{-ik_+t}u^{(1)}(-k_+,j,m;{\bf r})\right.\nonumber\\
&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \left.+a_{\text{out}}(k_+,j,m)e^{-ik_+t}u^{(2)}(-k_+,j,m;{\bf r})+\text{terms for antiparticle} \right],\\
\eta^\dagger&=\sum_{j,m}\int_0^\infty dk_-\left[(-1)^{m_-}c_{\text{in}}^\dagger(k_-,j,-m)e^{ik_-t}u^{(1)}(-k_-,j,m;{\bf r})\right.\nonumber\\
&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \left.+(-1)^{m_-}c_{\text{out}}^\dagger(k_-,j,-m)e^{ik_-t}u^{(2)}(-k_-,j,m;{\bf r})+\text{terms for antiparticle}\right],
\end{align}
using the solution of $(k+i\boldsymbol{\sigma}\cdot\nabla)u^{(i)}=0$, which is defined as
\begin{align}
u^{(i)}(k,j,m;{\bf r})\equiv\frac{k}{\sqrt{\pi}}\left[h_{l'}^{(i)}(kr)\Phi(j,m,l')+ih_l^{(i)}(kr)\Phi(j,m,l)\right],\ (i=1,2)
\end{align}
where $(l,l')\equiv(j+1/2,j-1/2)$ and $h_l^{(i)}$ denote Spherical Hankel functions.
$\Phi(j,m,l)$ and $\Phi(j,m,l')$ are the Pauli spinors, which are defined as follows.
\begin{align}
\Phi(j,m,l\equiv j+1/2)&\equiv\left(\begin{array}{ccc}\frac{\sqrt{j-m+1}}{\sqrt{2(j+1)}}Y_l^{m-1/2}\\-\frac{\sqrt{j+m+1}}{\sqrt{2(j+1)}}Y_l^{m+1/2}\\\end{array}\right),\\
\Phi(j,m,l'\equiv j-1/2)&\equiv\left(\begin{array}{ccc}\frac{\sqrt{j+m}}{\sqrt{2j}}Y_{l'}^{m-1/2}\\\frac{\sqrt{j-m}}{\sqrt{2j}}Y_{l'}^{m+1/2}\\\end{array}\right).
\end{align}
We also used
$\eta^\dagger=i\sigma_2(\eta_\alpha)^\ast$, $i\sigma_2u^{(1,2)}(k,j,m;{\bf r})^\ast=(-1)^{m_+}u^{(2,1)}(k,j,-m;{\bf r})$, where $m_{\pm}\equiv m\pm1/2$.
One may expect that the fermions outside the gauged Q-ball are described by the solutions above, but the Coulomb field $A_0$, which behaves as $\sim 1/r$ outside the Q-ball, cannot be neglected compared to the fermions, which become spherical waves $\sim e^{ikr}/r$, which are asymptotic forms of the spherical Hankel functions. However, as we derive in the Appendix, $A_0$ only gives an additional phase factor~$e^{iq_{\chi,\eta}e^2Q\log(2kr)}$ to the spherical waves at infinity. Thus, we can still identify incoming and outgoing wave solutions as in the previous paragraph, only corrected by the phase factors.
The coefficients $a_{\text{out}},c_{\text{out}}^\dagger$ can be written as superpositions of reflecting, and transmuting solutions as follows,
\begin{align}
a_{\text{out}}(k_+,j,m)&=R_\chi(k_+,j)a_{\text{in}}(k_+,j,m)+T_\chi(k_+,j)(-1)^{m_-}c_{\text{in}}^\dagger(k_-,j,-m),\label{eq:nuc}\\
(-1)^{m_-}c_{\text{out}}^\dagger(k_-,j,-m)&=T_\eta(k_-,j)a_{\text{in}}(k_+,j,m)+R_\eta(k_-,j)(-1)^{m_-}c_{\text{in}}^\dagger(k_-,j,-m),
\end{align}
whose coefficients must satisfy the following conditions,
\begin{align}
&|T_\chi(k_+,j)|^2=|T_\eta(k_-,j)|^2,\label{eq:eqce}\\
&|R_\chi(k_+,j)|^2+|T_\chi(k_+,j)|^2=1,\\
&|R_\eta(k_-,j)|^2+|T_\eta(k_-,j)|^2=1,
\end{align}
due to the anticommutation relations of the creation and annihilation operators.
If we define the vacuum $|0_{\text{in}}\rangle$ by $a_{\text{in}}|0_{\text{in}}\rangle=c_{\text{in}}|0_{\text{in}}\rangle=0$ at infinity, we see that the number of outgoing $\chi$ becomes
\begin{align}
\langle0_{\text{in}}|a_{\text{out}}^{\dagger}(k_+,j,m)a_{\text{out}}(k_+',j',m')|0_{\text{in}}\rangle=|T_\chi(k_+,j)|^2\delta(k_+-k_+')\delta_{j,j'}\delta_{m,m'},
\end{align}
using Eq.~(\ref{eq:nuc}), and by summing over the states, the production rate $dQ_i/dt$ is calculated as follows,
\begin{align}
\frac{dQ_i}{dt}=\sum_{j=1/2}\int_0^\omega\frac{dk}{2\pi}(2j+1)\left|T_i(k,j)\right|^2,\ \ \ \ (i=\chi,\eta)\label{eq:dni}
\end{align}
where we averaged the particle number over time using $\delta(0)=T/2\pi$.
This is the decay rate of the gauged Q-ball into the particle species $i$. Note that $\eta$ with momentum $k_\eta$ must be produced by the same amount as $\chi$ with momentum $\omega-k_\eta$, using Eq.~(\ref{eq:eqce}), which is due to the relation $dE/dQ=\omega$ of the gauged Q-balls.
The coefficients $R_i,T_i$ are determined by matching with the interior solutions, where $\phi,A_0\neq0$. The solutions are written as
\begin{align}
\chi&=f_\chi(r)\Phi(j,m,l')+ig_\chi(r)\Phi(j,m,l),\\
\eta^\dagger&=f_\eta(r)\Phi(j,m,l')+ig_\eta(r)\Phi(j,m,l),
\end{align}
where, again, we expanded the solutions by the Pauli spinors. We numerically solve for $f_i,g_i$, using Eq.~(\ref{eq:em1}) and (\ref{eq:em2}), under the following boundary conditions.
\begin{align}
f_i'(0)=g_i'(0)=0,
\end{align}
which regularize the solutions at $r=0$.
In Fig.~\ref{fig:dn}, we present the results for the production rates.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{qeh.pdf}
\caption{The production rates of fermions from gauged Q-balls. We can see the enhancement due to the electric repulsion. The dashed line indicates the saturated rates for global Q-balls. We can see that for global Q-balls, the production rates saturate as the charge grows, while for gauged Q-balls the saturation is unclear from the figure.}
\label{fig:dn}
\end{figure}
Since the gauged Q-ball has electric charge, the decay rate into particles with the charge of the same sign is expected to be enhanced by the electric repulsion, compared to that of the global Q-ball with the same charge. In the figure, we can see that this is indeed the case, where more fermions are
produced for larger gauge coupling $e^2$.
On the other hand, since $\chi$ is fermion, the flux coming out of the surface of the Q-ball must have an upper bound due to the Pauli blocking. For the global Q-ball, it is obtained by integrating the fully occupied phase space~$(0<k_+<\omega)$, at the surface of the Q-ball~\cite{coh}:
\begin{align}
\left(\frac{dQ}{dt}\right)_{\text{sat}}\equiv\frac{\omega^3R^2}{24\pi},
\label{eq:sato}
\end{align}
which is called saturated rate. The gauge mediation type global Q-balls have the following properties:
\begin{align}
\omega&\propto Q^{-1/4},\\
R&\simeq\pi/\omega,
\end{align}
Thus, we see that the saturated rate has charge dependence of $Q^{-1/4}$, which is illustrated by a dashed line in the figure. The production rate saturates when the Yukawa interaction becomes strong enough, or when $g\phi_0/\omega\gg1$. Here $\phi_0$ denotes the maximal value of $\phi$. If the Q-ball becomes large, $g\phi_0/\omega$ becomes large so that the interaction effectively becomes strong, which is the reason why the production rate saturates as the charge grows.
For gauged Q-balls, however, $\omega$ becomes large as the charge grows, as pointed out in the previous section, hence $g\phi_0/\omega$ does not necessarily become large for a large charge.
But if we consider a gauged Q-ball with a certain charge and a large Yukawa coupling so that $g\phi_0/\omega\gg1$, we find that the production rate indeed saturates, as shown in Fig.~\ref{fig:satto}.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{hpf.pdf}
\caption{The production rate from a gauged Q-ball as a function of $g\phi_0/\omega$. we see that the production rate saturates for $g\phi_0/\omega\gg1$. We took a gauged Q-ball with large charge, in order to identify the size of the Q-ball as clearly as possible, so that we can compare the production rate to the classically defined saturated rate, which is illustrated by a dotted line. We note that the actual saturated rate is larger than the classical formula, since the fermions with classically forbidden momenta are produced at infinity by quantum tunneling, where Coulomb potential outside effectively becomes potential barrier for fermions coming from inside. We can see that the production rate is suppressed compared to the saturated rate when the Coulomb barrier outside does not exist.}
\label{fig:satto}
\end{figure}
We took a gauged Q-ball with large charge, in order to identify the size of the Q-ball as clearly as possible~(see Fig.~\ref{fig:gauex}), so that we can compare the production rate to the saturated rate defined by
\begin{align}
\left(\frac{dQ}{dt}\right)_{\text{sat}}^{\text{(gauged)}}\equiv\frac{\tilde{\omega}^3R^2}{24\pi},\label{eq:sagau}
\end{align}
where we replaced $\omega$ in Eq.~(\ref{eq:sato}) by $\tilde{\omega}\equiv\omega-e^2Q/4\pi R$, which is the maximal momentum of fermions at the surface of the gauged Q-ball. However, we note that the actual saturated rate is somewhat larger than the one predicted by the above classical formula, whose reason is as follows. Since the classically emitted fermions are ``accelerated", or the momentum is increased by the Coulomb potential outside~($\sim1/r$), the observed fermions must have momentum of $e^2Q/4\pi R<k<\omega$ at infinity. However, as we see in Fig.~\ref{fig:tc}, fermions with momentum of $0<k<e^2Q/4\pi R$ are also observed, which leads to the disagreement in Fig.~\ref{fig:satto}.
The production of fermions with momentum of $0<k<e^2Q/4\pi R$ can be understood as a quantum tunneling effect. If the Yukawa interaction becomes strong, the fermion fields mainly feel $\phi$ inside the Q-ball, which we confirmed numerically as well, and feel the Coulomb potential suddenly at $r=R$. This situation is approximately the same as the case where the fermions produced by $\phi$ come out as a saturated flux with momentum of $0<k<e^2Q/4\pi R$, and bump into the barrier of Coulomb potential at $r=R+\Delta R$~($\Delta R\ll R$), where again, the validity of the approximation is confirmed numerically. This means the production of fermions with momentum of $0<k<e^2Q/4\pi R$ at infinity is due to the tunneling effect, and in particular must be suppressed compared to the saturated rate,
\begin{align}
\left(\frac{dQ}{dt}\right)_{\text{sat}}^{(0)}\equiv\frac{\omega^3R^2}{24\pi},
\label{eq:satos}
\end{align}
when the Coulomb barrier outside does not exist. This can also be confirmed in Fig.~\ref{fig:satto}.
Thus, we conclude that the decay rate of the gauged Q-ball is bounded from above, due to the Pauli blocking at the surface of the Q-ball, and further suppressed due to the Coulomb potential outside the Q-ball, which effectively becomes a potential barrier for the fermions coming from the inside.
\begin{figure}[t]
\centering
\includegraphics[width=0.65\linewidth]{hppf.pdf}
\caption{ The production rate as a function of momentum. We can see that the fermions with classically forbidden momenta are produced by quantum effect at infinity.}
\label{fig:tc}
\end{figure}
Finally, we present the behavior of production rates when $g\phi_0/\omega\ll1$ in Fig.~\ref{fig:dns}.
\begin{figure}[t]
\begin{minipage}{.47\linewidth}
\includegraphics[width=\linewidth]{dnsf.pdf}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}{.47\linewidth}
\includegraphics[width=\linewidth]{dns.pdf}
\end{minipage}
\caption{The production rates as a function of $g\phi_0/\omega$, normalized by the classical saturated rates. We find the enhancement of the production rates for $g\phi_0/\omega\ll1$ when $e^2$ or $Q$ becomes large, which can be explained by the analogy to the case of the global Q-ball, where the production rate enhances for a step function-like profile, since the electric repulsion deforms the profile of the gauged Q-ball into a step function-like profile.}
\label{fig:dns}
\end{figure}
Here we consider the gauged Q-balls of weak Coulomb potential with $e^2Q/4\pi R\ll\omega$, and normalize the production rates by the classical saturated rates $\left(dQ/dt\right)_{\text{sat}}^{(\text{gauged})}$, defined by Eq.~(\ref{eq:sagau}). Since the saturated rates have the value between $\left(dQ/dt\right)_{\text{sat}}^{(\text{gauged})}$ and $\left(dQ/dt\right)_{\text{sat}}^{(0)}$, we see that the normalized rate must saturate close to unity for $e^2Q/4\pi R\ll\omega$, which is indeed the case in the figure. Even for such weak Coulomb potential, we note that there are some differences in the production rates when $g\phi_0/\omega\ll1$, depending on the gauge coupling $e^2$, and the charge $Q$. The production rates~(normalized) are enhanced as $e^2$ or $Q$ grows, as shown in the figure. In Ref.~\cite{ym}, it was pointed out that for the global Q-ball, the production rate~(normalized) is enhanced for a step function-like profile. The similar explanation can be valid for the gauged Q-ball as well, since the electric repulsion pushes the charge toward the surface, which makes the profile like a step function.
\section{Conclusions and discussion}
\label{sec:conc}
In this paper, we derived the decay rate of the gauged Q-ball into fermions, using the semi-classical method in Ref.~\cite{coh}. We assume that the scalar field that forms the gauged Q-ball couples to fermions by Yukawa interaction.
Since the gauged Q-ball is electrically charged, the decay rate into particles with the charge of the same sign is expected to be enhanced, compared to the case of the global Q-ball. We found that indeed more particles come out from the surface of the gauged Q-ball, compared to the case of the global Q-ball, due to the electric repulsion.
For global Q-balls, it is known that there is an upper bound on the flux of fermions coming out of the surface of the Q-ball, due to the Pauli blocking, which is called saturated rate. We found that the production rates from each gauged Q-ball also saturate when the Yukawa interaction becomes strong, just as in the case of the global Q-ball. However, the saturated rate is somewhat larger than the one predicted by the classical formula, which is obtained by integrating the fully occupied phase space~$(0<k_+<\omega-e^2Q/4\pi R)$, at the surface of the Q-ball. The disagreement arises since the fermions with classically forbidden momenta are produced by the quantum effect. We found that the production can be interpreted as a tunneling effect, where the fermions, which are mainly produced by $\phi$ due to the strong Yukawa interaction, come out as a saturated flux, and immediately bump into the Coulomb barrier and tunnel through it. The production must be suppressed compared to the saturated rate when the Coulomb barrier outside does not exist, which is also confirmed.
We also found the enhancement of the production rates~(normalized) for $g\phi_0/\omega\ll1$ when $e^2$ or $Q$ becomes large, which can be explained by the analogy to the case of the global Q-ball, where the production rate enhances for a step function-like profile, since the electric repulsion deforms the profile of the gauged Q-ball into a step function-like profile.
In our previous works, we considered the elecrically charged Q-ball dark matter scenarios~\cite{j1,j2,j3}. In this scenario, the Q-balls formed after the Affleck-Dine mechanism can become electrically charged if the flat direction consists of baryonic and leptonic components and only the leptonic component decays off while the baryonic component is stable. This is possible if the energy of outgoing particle is smaller than the baryon mass, but larger than the lepton mass. Then, we also implicitly assumed that the decay of the leptonic component is sufficiently fast, so that the electrically charged Q-balls are formed in the early universe. From the result of this paper, we can gain some insight on that matter. While we discussed small Q-balls with charge of $10^3$ to $10^6$, for the convenience in the numerical calculations, the Q-balls in the cosmological context, are usually very large, with charge of $10^{20}$ to $10^{30}$. Thus, it is likely that $g\phi_0/\omega\gg 1$ unless the Yukawa coupling $g$ is extremely small, which means that the production rates of leptons are saturated. It was also pointed out in Ref.~\cite{j1} that the electric charge of the Q-ball can grow only until $Q\sim O(100)$ due to the Schwinger effect, etc., and especially the size and the maximal momentum of outgoing particle at the surface are nearly the same as the case of the global Q-ball. Thus, the saturated rate almost does not change from that for the global Q-ball. In all, the decay rate of the leptonic component is approximately written by the saturated rate for the global Q-ball, which is typically known to be of the order of GeV, thus we can conclude that the decay is sufficiently fast, and our previous assumption was reasonable.
\vspace{1cm}
\section*{Acknowledgments}
J.H. would like to thank Masaki Yamada for helpful comments. This work is supported by MEXT KAKENHI Grant Number 15H05889 (M.K.) and JSPS KAKENHI Grant Number 17K05434 (M.K.). The work is also supported by World Premier International Research Center Initiative (WPI Initiative), MEXT, Japan.
\vspace{1cm}
\begin{appendix}
\section{Fermions outside the gauged Q-ball}\label{apdix}
It is known that the equations of motion for fermions in the presence of the Coulomb potential of the form~$\sim1/r$ have analytic solutions. Thus, we can give analytic forms for the solution outside the gauged Q-ball. In this appendix, we present the solutions and their asymptotic behaviors, where we especially show that the solutions become spherical waves with some additional phase factors.
The radial parts of Eq.~(\ref{eq:em1}), which are the equations of motion for $\chi$, are given by
\begin{align}
\left(k_+-q_\chi \frac{e^2Q}{r}\right)f_\chi+\left(\frac{\del}{\del r}+\frac{3/2+j}{r}\right)g_\chi=0,\label{eq:emap1}\\
\left(k_+-q_\chi \frac{e^2Q}{r}\right)g_\chi-\left(\frac{\del}{\del r}+\frac{1/2-j}{r}\right)f_\chi=0,\label{eq:emap2}
\end{align}
using $\boldsymbol{\sigma}\cdot\nabla\Phi(j,m,j\pm1/2)=\Phi(j,m,j\mp1/2)\left(\del/\del r+(1\pm(j+1/2))/r\right)$, and the solutions are written as the following forms.
\begin{align}
f_\chi(r)&=\frac{e^{ik_+r}}{r}\biggl[r^{s_0}C_+\biggl[{}_1F_1(s_0+iq_\chi e^2Q,2s_0+1;-2ik_+r)\biggr.\biggr.\nonumber\\
&+\left.\frac{s_0+iq_\chi e^2Q}{j+1/2} {}_1F_1(s_0+1+iq_\chi e^2Q,2s_0+1;-2ik_+r)\right]\nonumber\\
&+r^{-s_0}C_-\biggl[{}_1F_1(-s_0+iq_\chi e^2Q,-2s_0+1;-2ik_+r)\biggr.\nonumber\\
&+\left.\left.\frac{-s_0+iq_\chi e^2Q}{j+1/2} {}_1F_1(-s_0+1+iq_\chi e^2Q,-2s_0+1;-2ik_+r)\right]\right],
\\
g_\chi(r)&=i\frac{e^{ik_+r}}{r}\biggl[r^{s_0}C_+\biggl[{}_1F_1(s_0+iq_\chi e^2Q,2s_0+1;-2ik_+r)\biggr.\biggr.\nonumber\\
&+\left.\frac{s_0+iq_\chi e^2Q}{j+1/2} {}_1F_1(s_0+1+iq_\chi e^2Q,2s_0+1;-2ik_+r)\right]\nonumber\\
&-r^{-s_0}C_-\biggl[{}_1F_1(-s_0+iq_\chi e^2Q,-2s_0+1;-2ik_+r)\biggr.\nonumber\\
&+\left.\left.\frac{-s_0+iq_\chi e^2Q}{j+1/2} {}_1F_1(-s_0+1+iq_\chi e^2Q,-2s_0+1;-2ik_+r)\right]\right],
\end{align}
where $s_0=\sqrt{(j+1/2)^2-(q_\chi e^2Q)^2}$, and we used the confluent hypergeometric function, which is defined as follows.
\begin{align}
{}_1F_1(a,b;z)\equiv\sum_{k=0}^\infty\frac{a(a+1)\cdots(a+k-1)}{b(b+1)\cdots(b+k-1)}\frac{z^k}{k!}
\end{align}
Using the following asymptotic form of the confluent hypergeometric function,
\begin{align}
{}_1F_1(a,b;z)\sim\Gamma(b)\left(\frac{e^zz^{a-b}}{\Gamma(a)}+\frac{(-1)^{-a}z^{-a}}{\Gamma(b-a)}\right),\,\,\,\,\, |z|\gg1
\end{align}
we find that the solutions behave as
\begin{align}
f_\chi&\sim \left[C_+\frac{s_0+iq_\chi e^2Q}{j+1/2}\left((-1)^{s_0}\frac{(-i)^{-s_0+iq_\chi e^2Q}\Gamma(2s_0+1)}{\Gamma(s_0+1+iq_\chi e^2Q)}\right)\right.\nonumber\\
&\left.\,\,\,\,\,+C_-\frac{-s_0+iq_\chi e^2Q}{j+1/2}\left((-1)^{-s_0}\frac{(-i)^{s_0+iq_\chi e^2Q}\Gamma(-2s_0+1)}{\Gamma(-s_0+1+iq_\chi e^2Q)}\right)\right]\times\frac{e^{-ik_+r+iq_{\chi}e^2Q\log(2kr)}}{r}\nonumber\\
&\,\,\,\,\,+\left[C_+\left((-1)^{s_0}\frac{i^{-s_0-iq_\chi e^2Q}\Gamma(2s_0+1)}{\Gamma(s_0+1-iq_\chi e^2Q)}\right)\right.\nonumber\\
&\left.\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,+C_-\left((-1)^{-s_0}\frac{i^{s_0-iq_\chi e^2Q}\Gamma(-2s_0+1)}{\Gamma(-s_0+1-iq_\chi e^2Q)}\right)\right]\times\frac{e^{ik_+r-iq_{\chi}e^2Q\log(2kr)}}{r},\\
g_\chi&\sim \left[C_+\frac{s_0+iq_\chi e^2Q}{j+1/2}\left((-1)^{s_0}\frac{(-i)^{-s_0+iq_\chi e^2Q}\Gamma(2s_0+1)}{\Gamma(s_0+1+iq_\chi e^2Q)}\right)\right.\nonumber\\
&\left.\,\,\,\,\,+C_-\frac{-s_0+iq_\chi e^2Q}{j+1/2}\left((-1)^{-s_0}\frac{(-i)^{s_0+iq_\chi e^2Q}\Gamma(-2s_0+1)}{\Gamma(-s_0+1+iq_\chi e^2Q)}\right)\right]\times(+i)\frac{e^{-ik_+r+iq_{\chi}e^2Q\log(2kr)}}{r}\nonumber\\
&\,\,\,\,\,+\left[C_+\left((-1)^{s_0}\frac{i^{-s_0-iq_\chi e^2Q}\Gamma(2s_0+1)}{\Gamma(s_0+1-iq_\chi e^2Q)}\right)\right.\nonumber\\
&\left.\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,+C_-\left((-1)^{-s_0}\frac{i^{s_0-iq_\chi e^2Q}\Gamma(-2s_0+1)}{\Gamma(-s_0+1-iq_\chi e^2Q)}\right)\right]\times(-i)\frac{e^{ik_+r-iq_{\chi}e^2Q\log(2kr)}}{r},
\end{align}
at infinity, consisting of incoming and outgoing waves with additional phase factors,~$e^{\pm iq_{\chi}e^2Q\log(2kr)}$.
\end{appendix}
\vspace{1cm}
\section{Introduction
\label{sec:introduction}}
It is known that in the minimal supersymmetric standard model~(MSSM), the baryon asymmetry in our universe can be generated by Affleck-Dine mechanism~\cite{ad,add}, which produces a scalar field condensate with baryon number. In many models such as the gauge mediated SUSY breaking models, the spatial inhomogeneities of this condensate due to quantum fluctuations grow and fragment into non-topological solitons called Q-balls~\cite{ks,kkins,kkins2}, which are defined as spherical solutions in a global $U(1)$ theory which minimize the energy of the system with a fixed $U(1)$ charge~\cite{c}. In this case, the baryon number generated in Affleck-Dine mechanism is confined inside Q-balls, so that the baryon asymmetry in the universe is generated by baryons emitted by the decay of the Q-balls.
The Q-ball decay into other particles was first studied by Cohen {\it et al.}~\cite{coh}, who considered the Yukawa theory and calculated the neutrino pair production rate by leading semi-classical approximation treating the Q-ball as the classical background scalar field, where the Q-ball configuration was approximated as a step function. More realistic configurations are considered in Refs.~\cite{nur,ym}. In particular, the production rates of quarks and gravitinos from the Q-balls in supersymmetric theories were derived in Ref.~\cite{ym}, which can be used to estimate baryon-to-dark matter ratio in gauge mediated SUSY breaking models, where the gravitino is dark matter.
While the Q-ball is a non-topological soliton resulting from global $U(1)$ symmetry, whose generalization to local $U(1)$ symmetry was also proposed. The lowest-energy configuration with a fixed local $U(1)$ charge is called gauged Q-ball~\cite{gaugedqball}, which consists not only of the scalar field, but also of the $U(1)$ gauge field. The properties of gauged Q-ball solutions have been studied analytically and numerically in the literature~\cite{gaugedqball,5,7,8,9,10,11,12}, but their decay into other particles has not been considered. In this paper, we derive the decay rate of the gauged Q-ball into fermions, applying the leading semi-classical approximation used in Ref.~\cite{coh} to the gauged Q-ball. We assume that the scalar field in the Q-ball couples to fermions by Yukawa interaction. Since the gauged Q-ball can be interpreted as the electrically charged Q-ball, the decay rate into particles with the charge of the same sign is expected to be enhanced, compared to the case of a global Q-ball, and we show that this is indeed the case. We also show that, on the other hand, the decay rate of the gauged Q-ball is upperly bounded, due to the Pauli blocking at the surface of the Q-ball, just as pointed out in the case of the global Q-ball~\cite{coh}, and even more suppressed due to the Coulomb potential outside the Q-ball, since it plays the role of a potential barrier for the fermions coming from the inside, as we will see later.
The paper is organized as follows. In Sec.~\ref{sec:bp}, we review some basic properties of gauged Q-ball. In Sec.~\ref{sec:dr}, we present the theoretical setup for gauged Q-balls coupled to the massless fermions and calculate the gauged Q-ball decay rate into massless fermions by using the leading semi-classical approximation. Sec.~\ref{sec:conc} is devoted to the conclusions.
\section{Gauged Q-ball}
\label{sec:bp}
We consider a theory of a complex scalar field $\phi$ coupled to a $U(1)$ gauge field $A_\mu$. The Lagrangian density is written as follows.
\begin{align}
\mathcal{L}&=(D_\mu\phi)^\ast D^\mu\phi-V(\phi)-\frac14F_{\mu\nu}F^{\mu\nu},\\
D_\mu&\equiv\partial_\mu+ieA_\mu,
\end{align}
where $V(\phi)$ is a scalar potential and $F_{\mu\nu}=\del_\mu A_\nu-\del_\nu A_\mu$. We introduce the following ansatz on $\phi$,
\begin{align}
&\phi(x,t)\equiv \phi(r)e^{-i\omega t},
\label{eq:gqpara}
\end{align}
which is the same parametrization as that of a global Q-ball.
For the gauge field, we find spatially symmetric solution with no magnetic field, or no electric current:
\begin{align}
&A_0=A_0(r),\\
&A_i=0.
\end{align}
The equations of motion are then given by
\begin{align}
\label{eq:eomf}
&\frac{d^2\phi}{dr^2}+\frac2r\frac{d\phi}{dr}+\phi h^2-\frac{dV}{d\phi}=0,\\
&\frac{d^2h}{dr^2}+\frac2r\frac{dh}{dr}-e^2\phi^2h=0,
\label{eq:eomgaugedqball}
\end{align}
where we redefined the gauge field to absorb $\omega$ as $h\equiv-\omega+eA_0$.
We set boundary conditions as
\begin{align}
\label{eq:bdcga}
&\phi(\infty)=0,~~ \frac{d\phi}{dr}(0)=0,\\
&A_0(\infty)=0,~~\frac{dA_0}{dr}(0)=0,
\end{align}
especially to avoid singularities at $r=0$.
As a scalar potential, we choose a logarithmic potential $V(\phi)=m_\phi^4\ln(1+|\phi|^2/m_\phi^2)$, which is motivated by gauge mediation models. For $e=0$, the solution becomes a global Q-ball, which is called gauge mediation type Q-ball~\cite{dv,ls}.
It is known that this type of Q-balls with sufficiently large charge has the following approximate analytic solution.
\begin{align}
\phi(r)=\left\{
\begin{array}{ll}
\phi_0\sin\omega r/\omega r,
& (r\leq R\equiv\pi/\omega)
\\
\\
0, & (r>R)
\end{array}
\right..
\label{eq:sin}
\end{align}
The angular velocity $\omega$ is equal to $dE/dQ$, which is true for general Q-ball solutions, and has the following charge dependence,
\begin{align}
\omega=\frac{dE}{dQ}&\propto Q^{-1/4},\label{eq:q14}
\end{align}
which will be useful later.
The second derivative of $\phi$ becomes singular at $r=R$, which for actual Q-balls, becomes a peak of $\phi''(r)$. We define the size of a gauged Q-ball, which is the case $e\neq0$, as the point where $\phi'''(r)=0$ as well, even if the profile is somewhat pushed outward by the electric repulsion, as shown in Fig.~\ref{fig:gauex}. Indeed we can see that $\phi''(r=R)$ becomes singular for a large gauged Q-ball, just as for the global Q-ball, even when the Coulomb potential has a non-negligible effect on the profile. Later we consider the case of large gauged Q-balls when we discuss the saturation of fermion production, where specifying the size of a Q-ball becomes important.
\begin{figure}[t]
\begin{minipage}{.5\linewidth}
\includegraphics[width=\linewidth]{prfs.pdf}
\end{minipage}
\hspace{0cm}
\begin{minipage}{.5\linewidth}
\includegraphics[width=\linewidth]{prfl.pdf}
\end{minipage}
\caption{Examples of profile of the gauged Q-ball of gauge mediation type. The dimensionful parameters are in units of $m_\phi$. The dashed line denotes the global Q-ball with the same charge. We see that the profile of $\phi$ is pushed outward due to the electric repulsion. We can also see that $\phi''(r=R)$ becomes singular for a large gauged Q-ball, just as for a global Q-ball, even when the Coulomb potential has a non-negligible effect on the profile.}
\label{fig:gauex}
\end{figure}
The energy and charge of the gauged Q-ball are given by
\begin{align}
E&=\int d^3x\left[\frac12(\nabla\phi)^2+\frac1{2}(\nabla A_0)^2+\frac12\phi^2(\omega-eA_0)^2+V(\phi)\right],\\
Q&=\int d^3 x(\omega-eA_0)\phi^2,
\end{align}
and the relation $\omega=dE/dQ$ holds, just as in the case of a global Q-ball, whose proof is given in Ref.~\cite{7}.
The gauged Q-ball becomes unstable as the charge grows due to electric repulsion, which can be seen by the behavior of $\omega=dE/dQ$. We present the plot of $\omega$ as a function of $Q$ in Fig.~\ref{fig:omegg}~(Left). We can see that in contrast to the behavior of $\omega$ for global Q-balls, which is denoted by a dashed line, $\omega$ increases as the charge grows, which means that the Q-ball becomes unstable\footnote
If the charge grows further, $\omega$ becomes larger than $m_\phi$, which means that the Q-ball becomes unstable against decay into itself, and soon the solutions cease to exist. Here we simply focus on the case $\omega<m_\phi$.}. We also plot Coulomb energy at the surface of Q-ball, $e^2Q/4\pi R$ by a dotted line, whose contribution also becomes large as charge grows. However, we see that the Coulomb energy stays smaller than $\omega$, due to the growth of $\omega$, and also of $R$ by electric repulsion, which we illustrate in Fig.~\ref{fig:omegg}~(Right).
\begin{figure}[t]
\begin{minipage}{.47\linewidth}
\includegraphics[width=\linewidth]{oyf.pdf}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}{.47\linewidth}
\includegraphics[width=\linewidth]{ryf.pdf}
\end{minipage}
\caption{The plots of $\omega=dE/dQ$ and $R$ as functions of $Q$. We plot those for global Q-balls by dashed lines, for comparison. We see that $\omega$ becomes large as the charge grows, which means the Q-ball becomes unstable due to the electric charge, and $R$ becomes large due to the electric repulsion. We also present the Coulomb energy at the surface, $e^2Q/4\pi R$, which is denoted by a dotted line in the left figure.}
\label{fig:omegg}
\end{figure}
\section{Gauged Q-ball decay rates into fermions}
\label{sec:dr}
In this section, we derive the decay rate of the gauged Q-ball into fermions, by using the leading semi-classical approximation, where we treat the gauged Q-ball as the classical background. We calculate the production rate of the fermions in the presence of it.
We consider the following Lagrangian,
\begin{align}
\mathcal{L}_{\text{fermion}}=\chi^\dagger i\bar{\sigma}^\mu(\del_\mu+iq_\chi eA_\mu)\chi+\eta^\dagger i\bar{\sigma}^\mu(\del_\mu+iq_\eta eA_\mu)\eta-(g\phi^\ast\chi\eta+\text{h.c.}),
\end{align}
where $\chi,\eta$ are Weyl fermions, which couple to $\phi$ by Yukawa interaction, and $\bar{\sigma}^\mu=({\bf1},-\sigma^i)$, where $\sigma^i$ are the Pauli matrices. We note that $q_\chi+q_\eta=1$ must be satisfied due to the charge conservation.
Here we simply set $(q_\chi,q_\eta)=(1,0)$, which assigns the same sign of charge to~$\chi$.
The equations of motion are written as
\begin{align}
i\bar{\sigma}^\mu(\del_\mu+iq_\chi eA_\mu)\chi-g\phi\eta^\dagger&=0,\\
i\sigma^\mu(\del_\mu-iq_\eta eA_\mu)\eta^\dagger-g\phi^\ast\chi&=0.
\end{align}
Since $\phi$ is time dependent, the following modes mix with each other.
\begin{align}
\chi&\propto e^{-ik_+t},\\
\eta^\dagger&\propto e^{i(\omega-k_+)t}\equiv e^{ik_-t},
\end{align}
whose equations of motion become
\begin{align}
(k_++iq_\chi eA_0-i\boldsymbol{\sigma}\cdot\nabla)\chi-g\phi(r)\eta^\dagger&=0,\label{eq:em1}\\
(-k_--iq_\eta eA_0+i\boldsymbol{\sigma}\cdot\nabla)\eta^\dagger-g\phi(r)\chi&=0.\label{eq:em2}
\end{align}
First, we consider the case when $\chi,\eta$ are free fields, whose equations of motion are
\begin{align}
(k_+-i\boldsymbol{\sigma}\cdot\nabla)\chi&=0,\\
(-k_-+i\boldsymbol{\sigma}\cdot\nabla)\eta^\dagger&=0.
\end{align}
Then, we can write the following expansion of $\chi,\eta^\dagger$,
\begin{align}
\chi&=\sum_{j,m}\int_0^\infty dk_+\left[a_{\text{in}}(k_+,j,m)e^{-ik_+t}u^{(1)}(-k_+,j,m;{\bf r})\right.\nonumber\\
&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \left.+a_{\text{out}}(k_+,j,m)e^{-ik_+t}u^{(2)}(-k_+,j,m;{\bf r})+\text{terms for antiparticle} \right],\\
\eta^\dagger&=\sum_{j,m}\int_0^\infty dk_-\left[(-1)^{m_-}c_{\text{in}}^\dagger(k_-,j,-m)e^{ik_-t}u^{(1)}(-k_-,j,m;{\bf r})\right.\nonumber\\
&\ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \left.+(-1)^{m_-}c_{\text{out}}^\dagger(k_-,j,-m)e^{ik_-t}u^{(2)}(-k_-,j,m;{\bf r})+\text{terms for antiparticle}\right],
\end{align}
using the solution of $(k+i\boldsymbol{\sigma}\cdot\nabla)u^{(i)}=0$, which is defined as
\begin{align}
u^{(i)}(k,j,m;{\bf r})\equiv\frac{k}{\sqrt{\pi}}\left[h_{l'}^{(i)}(kr)\Phi(j,m,l')+ih_l^{(i)}(kr)\Phi(j,m,l)\right],\ (i=1,2)
\end{align}
where $(l,l')\equiv(j+1/2,j-1/2)$ and $h_l^{(i)}$ denote Spherical Hankel functions.
$\Phi(j,m,l)$ and $\Phi(j,m,l')$ are the Pauli spinors, which are defined as follows.
\begin{align}
\Phi(j,m,l\equiv j+1/2)&\equiv\left(\begin{array}{ccc}\frac{\sqrt{j-m+1}}{\sqrt{2(j+1)}}Y_l^{m-1/2}\\-\frac{\sqrt{j+m+1}}{\sqrt{2(j+1)}}Y_l^{m+1/2}\\\end{array}\right),\\
\Phi(j,m,l'\equiv j-1/2)&\equiv\left(\begin{array}{ccc}\frac{\sqrt{j+m}}{\sqrt{2j}}Y_{l'}^{m-1/2}\\\frac{\sqrt{j-m}}{\sqrt{2j}}Y_{l'}^{m+1/2}\\\end{array}\right).
\end{align}
We also used
$\eta^\dagger=i\sigma_2(\eta_\alpha)^\ast$, $i\sigma_2u^{(1,2)}(k,j,m;{\bf r})^\ast=(-1)^{m_+}u^{(2,1)}(k,j,-m;{\bf r})$, where $m_{\pm}\equiv m\pm1/2$.
One may expect that the fermions outside the gauged Q-ball are described by the solutions above, but the Coulomb field $A_0$, which behaves as $\sim 1/r$ outside the Q-ball, cannot be neglected compared to the fermions, which become spherical waves $\sim e^{ikr}/r$, which are asymptotic forms of the spherical Hankel functions. However, as we derive in the Appendix, $A_0$ only gives an additional phase factor~$e^{iq_{\chi,\eta}e^2Q\log(2kr)}$ to the spherical waves at infinity. Thus, we can still identify incoming and outgoing wave solutions as in the previous paragraph, only corrected by the phase factors.
The coefficients $a_{\text{out}},c_{\text{out}}^\dagger$ can be written as superpositions of reflecting, and transmuting solutions as follows,
\begin{align}
a_{\text{out}}(k_+,j,m)&=R_\chi(k_+,j)a_{\text{in}}(k_+,j,m)+T_\chi(k_+,j)(-1)^{m_-}c_{\text{in}}^\dagger(k_-,j,-m),\label{eq:nuc}\\
(-1)^{m_-}c_{\text{out}}^\dagger(k_-,j,-m)&=T_\eta(k_-,j)a_{\text{in}}(k_+,j,m)+R_\eta(k_-,j)(-1)^{m_-}c_{\text{in}}^\dagger(k_-,j,-m),
\end{align}
whose coefficients must satisfy the following conditions,
\begin{align}
&|T_\chi(k_+,j)|^2=|T_\eta(k_-,j)|^2,\label{eq:eqce}\\
&|R_\chi(k_+,j)|^2+|T_\chi(k_+,j)|^2=1,\\
&|R_\eta(k_-,j)|^2+|T_\eta(k_-,j)|^2=1,
\end{align}
due to the anticommutation relations of the creation and annihilation operators.
If we define the vacuum $|0_{\text{in}}\rangle$ by $a_{\text{in}}|0_{\text{in}}\rangle=c_{\text{in}}|0_{\text{in}}\rangle=0$ at infinity, we see that the number of outgoing $\chi$ becomes
\begin{align}
\langle0_{\text{in}}|a_{\text{out}}^{\dagger}(k_+,j,m)a_{\text{out}}(k_+',j',m')|0_{\text{in}}\rangle=|T_\chi(k_+,j)|^2\delta(k_+-k_+')\delta_{j,j'}\delta_{m,m'},
\end{align}
using Eq.~(\ref{eq:nuc}), and by summing over the states, the production rate $dQ_i/dt$ is calculated as follows,
\begin{align}
\frac{dQ_i}{dt}=\sum_{j=1/2}\int_0^\omega\frac{dk}{2\pi}(2j+1)\left|T_i(k,j)\right|^2,\ \ \ \ (i=\chi,\eta)\label{eq:dni}
\end{align}
where we averaged the particle number over time using $\delta(0)=T/2\pi$.
This is the decay rate of the gauged Q-ball into the particle species $i$. Note that $\eta$ with momentum $k_\eta$ must be produced by the same amount as $\chi$ with momentum $\omega-k_\eta$, using Eq.~(\ref{eq:eqce}), which is due to the relation $dE/dQ=\omega$ of the gauged Q-balls.
The coefficients $R_i,T_i$ are determined by matching with the interior solutions, where $\phi,A_0\neq0$. The solutions are written as
\begin{align}
\chi&=f_\chi(r)\Phi(j,m,l')+ig_\chi(r)\Phi(j,m,l),\\
\eta^\dagger&=f_\eta(r)\Phi(j,m,l')+ig_\eta(r)\Phi(j,m,l),
\end{align}
where, again, we expanded the solutions by the Pauli spinors. We numerically solve for $f_i,g_i$, using Eq.~(\ref{eq:em1}) and (\ref{eq:em2}), under the following boundary conditions.
\begin{align}
f_i'(0)=g_i'(0)=0,
\end{align}
which regularize the solutions at $r=0$.
In Fig.~\ref{fig:dn}, we present the results for the production rates.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{qeh.pdf}
\caption{The production rates of fermions from gauged Q-balls. We can see the enhancement due to the electric repulsion. The dashed line indicates the saturated rates for global Q-balls. We can see that for global Q-balls, the production rates saturate as the charge grows, while for gauged Q-balls the saturation is unclear from the figure.}
\label{fig:dn}
\end{figure}
Since the gauged Q-ball has electric charge, the decay rate into particles with the charge of the same sign is expected to be enhanced by the electric repulsion, compared to that of the global Q-ball with the same charge. In the figure, we can see that this is indeed the case, where more fermions are
produced for larger gauge coupling $e^2$.
On the other hand, since $\chi$ is fermion, the flux coming out of the surface of the Q-ball must have an upper bound due to the Pauli blocking. For the global Q-ball, it is obtained by integrating the fully occupied phase space~$(0<k_+<\omega)$, at the surface of the Q-ball~\cite{coh}:
\begin{align}
\left(\frac{dQ}{dt}\right)_{\text{sat}}\equiv\frac{\omega^3R^2}{24\pi},
\label{eq:sato}
\end{align}
which is called saturated rate. The gauge mediation type global Q-balls have the following properties:
\begin{align}
\omega&\propto Q^{-1/4},\\
R&\simeq\pi/\omega,
\end{align}
Thus, we see that the saturated rate has charge dependence of $Q^{-1/4}$, which is illustrated by a dashed line in the figure. The production rate saturates when the Yukawa interaction becomes strong enough, or when $g\phi_0/\omega\gg1$. Here $\phi_0$ denotes the maximal value of $\phi$. If the Q-ball becomes large, $g\phi_0/\omega$ becomes large so that the interaction effectively becomes strong, which is the reason why the production rate saturates as the charge grows.
For gauged Q-balls, however, $\omega$ becomes large as the charge grows, as pointed out in the previous section, hence $g\phi_0/\omega$ does not necessarily become large for a large charge.
But if we consider a gauged Q-ball with a certain charge and a large Yukawa coupling so that $g\phi_0/\omega\gg1$, we find that the production rate indeed saturates, as shown in Fig.~\ref{fig:satto}.
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{hpf.pdf}
\caption{The production rate from a gauged Q-ball as a function of $g\phi_0/\omega$. we see that the production rate saturates for $g\phi_0/\omega\gg1$. We took a gauged Q-ball with large charge, in order to identify the size of the Q-ball as clearly as possible, so that we can compare the production rate to the classically defined saturated rate, which is illustrated by a dotted line. We note that the actual saturated rate is larger than the classical formula, since the fermions with classically forbidden momenta are produced at infinity by quantum tunneling, where Coulomb potential outside effectively becomes potential barrier for fermions coming from inside. We can see that the production rate is suppressed compared to the saturated rate when the Coulomb barrier outside does not exist.}
\label{fig:satto}
\end{figure}
We took a gauged Q-ball with large charge, in order to identify the size of the Q-ball as clearly as possible~(see Fig.~\ref{fig:gauex}), so that we can compare the production rate to the saturated rate defined by
\begin{align}
\left(\frac{dQ}{dt}\right)_{\text{sat}}^{\text{(gauged)}}\equiv\frac{\tilde{\omega}^3R^2}{24\pi},\label{eq:sagau}
\end{align}
where we replaced $\omega$ in Eq.~(\ref{eq:sato}) by $\tilde{\omega}\equiv\omega-e^2Q/4\pi R$, which is the maximal momentum of fermions at the surface of the gauged Q-ball. However, we note that the actual saturated rate is somewhat larger than the one predicted by the above classical formula, whose reason is as follows. Since the classically emitted fermions are ``accelerated", or the momentum is increased by the Coulomb potential outside~($\sim1/r$), the observed fermions must have momentum of $e^2Q/4\pi R<k<\omega$ at infinity. However, as we see in Fig.~\ref{fig:tc}, fermions with momentum of $0<k<e^2Q/4\pi R$ are also observed, which leads to the disagreement in Fig.~\ref{fig:satto}.
The production of fermions with momentum of $0<k<e^2Q/4\pi R$ can be understood as a quantum tunneling effect. If the Yukawa interaction becomes strong, the fermion fields mainly feel $\phi$ inside the Q-ball, which we confirmed numerically as well, and feel the Coulomb potential suddenly at $r=R$. This situation is approximately the same as the case where the fermions produced by $\phi$ come out as a saturated flux with momentum of $0<k<e^2Q/4\pi R$, and bump into the barrier of Coulomb potential at $r=R+\Delta R$~($\Delta R\ll R$), where again, the validity of the approximation is confirmed numerically. This means the production of fermions with momentum of $0<k<e^2Q/4\pi R$ at infinity is due to the tunneling effect, and in particular must be suppressed compared to the saturated rate,
\begin{align}
\left(\frac{dQ}{dt}\right)_{\text{sat}}^{(0)}\equiv\frac{\omega^3R^2}{24\pi},
\label{eq:satos}
\end{align}
when the Coulomb barrier outside does not exist. This can also be confirmed in Fig.~\ref{fig:satto}.
Thus, we conclude that the decay rate of the gauged Q-ball is bounded from above, due to the Pauli blocking at the surface of the Q-ball, and further suppressed due to the Coulomb potential outside the Q-ball, which effectively becomes a potential barrier for the fermions coming from the inside.
\begin{figure}[t]
\centering
\includegraphics[width=0.65\linewidth]{hppf.pdf}
\caption{ The production rate as a function of momentum. We can see that the fermions with classically forbidden momenta are produced by quantum effect at infinity.}
\label{fig:tc}
\end{figure}
Finally, we present the behavior of production rates when $g\phi_0/\omega\ll1$ in Fig.~\ref{fig:dns}.
\begin{figure}[t]
\begin{minipage}{.47\linewidth}
\includegraphics[width=\linewidth]{dnsf.pdf}
\end{minipage}
\hspace{0.5cm}
\begin{minipage}{.47\linewidth}
\includegraphics[width=\linewidth]{dns.pdf}
\end{minipage}
\caption{The production rates as a function of $g\phi_0/\omega$, normalized by the classical saturated rates. We find the enhancement of the production rates for $g\phi_0/\omega\ll1$ when $e^2$ or $Q$ becomes large, which can be explained by the analogy to the case of the global Q-ball, where the production rate enhances for a step function-like profile, since the electric repulsion deforms the profile of the gauged Q-ball into a step function-like profile.}
\label{fig:dns}
\end{figure}
Here we consider the gauged Q-balls of weak Coulomb potential with $e^2Q/4\pi R\ll\omega$, and normalize the production rates by the classical saturated rates $\left(dQ/dt\right)_{\text{sat}}^{(\text{gauged})}$, defined by Eq.~(\ref{eq:sagau}). Since the saturated rates have the value between $\left(dQ/dt\right)_{\text{sat}}^{(\text{gauged})}$ and $\left(dQ/dt\right)_{\text{sat}}^{(0)}$, we see that the normalized rate must saturate close to unity for $e^2Q/4\pi R\ll\omega$, which is indeed the case in the figure. Even for such weak Coulomb potential, we note that there are some differences in the production rates when $g\phi_0/\omega\ll1$, depending on the gauge coupling $e^2$, and the charge $Q$. The production rates~(normalized) are enhanced as $e^2$ or $Q$ grows, as shown in the figure. In Ref.~\cite{ym}, it was pointed out that for the global Q-ball, the production rate~(normalized) is enhanced for a step function-like profile. The similar explanation can be valid for the gauged Q-ball as well, since the electric repulsion pushes the charge toward the surface, which makes the profile like a step function.
\section{Conclusions and discussion}
\label{sec:conc}
In this paper, we derived the decay rate of the gauged Q-ball into fermions, using the semi-classical method in Ref.~\cite{coh}. We assume that the scalar field that forms the gauged Q-ball couples to fermions by Yukawa interaction.
Since the gauged Q-ball is electrically charged, the decay rate into particles with the charge of the same sign is expected to be enhanced, compared to the case of the global Q-ball. We found that indeed more particles come out from the surface of the gauged Q-ball, compared to the case of the global Q-ball, due to the electric repulsion.
For global Q-balls, it is known that there is an upper bound on the flux of fermions coming out of the surface of the Q-ball, due to the Pauli blocking, which is called saturated rate. We found that the production rates from each gauged Q-ball also saturate when the Yukawa interaction becomes strong, just as in the case of the global Q-ball. However, the saturated rate is somewhat larger than the one predicted by the classical formula, which is obtained by integrating the fully occupied phase space~$(0<k_+<\omega-e^2Q/4\pi R)$, at the surface of the Q-ball. The disagreement arises since the fermions with classically forbidden momenta are produced by the quantum effect. We found that the production can be interpreted as a tunneling effect, where the fermions, which are mainly produced by $\phi$ due to the strong Yukawa interaction, come out as a saturated flux, and immediately bump into the Coulomb barrier and tunnel through it. The production must be suppressed compared to the saturated rate when the Coulomb barrier outside does not exist, which is also confirmed.
We also found the enhancement of the production rates~(normalized) for $g\phi_0/\omega\ll1$ when $e^2$ or $Q$ becomes large, which can be explained by the analogy to the case of the global Q-ball, where the production rate enhances for a step function-like profile, since the electric repulsion deforms the profile of the gauged Q-ball into a step function-like profile.
In our previous works, we considered the elecrically charged Q-ball dark matter scenarios~\cite{j1,j2,j3}. In this scenario, the Q-balls formed after the Affleck-Dine mechanism can become electrically charged if the flat direction consists of baryonic and leptonic components and only the leptonic component decays off while the baryonic component is stable. This is possible if the energy of outgoing particle is smaller than the baryon mass, but larger than the lepton mass. Then, we also implicitly assumed that the decay of the leptonic component is sufficiently fast, so that the electrically charged Q-balls are formed in the early universe. From the result of this paper, we can gain some insight on that matter. While we discussed small Q-balls with charge of $10^3$ to $10^6$, for the convenience in the numerical calculations, the Q-balls in the cosmological context, are usually very large, with charge of $10^{20}$ to $10^{30}$. Thus, it is likely that $g\phi_0/\omega\gg 1$ unless the Yukawa coupling $g$ is extremely small, which means that the production rates of leptons are saturated. It was also pointed out in Ref.~\cite{j1} that the electric charge of the Q-ball can grow only until $Q\sim O(100)$ due to the Schwinger effect, etc., and especially the size and the maximal momentum of outgoing particle at the surface are nearly the same as the case of the global Q-ball. Thus, the saturated rate almost does not change from that for the global Q-ball. In all, the decay rate of the leptonic component is approximately written by the saturated rate for the global Q-ball, which is typically known to be of the order of GeV, thus we can conclude that the decay is sufficiently fast, and our previous assumption was reasonable.
\vspace{1cm}
\section*{Acknowledgments}
J.H. would like to thank Masaki Yamada for helpful comments. This work is supported by MEXT KAKENHI Grant Number 15H05889 (M.K.) and JSPS KAKENHI Grant Number 17K05434 (M.K.). The work is also supported by World Premier International Research Center Initiative (WPI Initiative), MEXT, Japan.
\vspace{1cm}
\begin{appendix}
\section{Fermions outside the gauged Q-ball}\label{apdix}
It is known that the equations of motion for fermions in the presence of the Coulomb potential of the form~$\sim1/r$ have analytic solutions. Thus, we can give analytic forms for the solution outside the gauged Q-ball. In this appendix, we present the solutions and their asymptotic behaviors, where we especially show that the solutions become spherical waves with some additional phase factors.
The radial parts of Eq.~(\ref{eq:em1}), which are the equations of motion for $\chi$, are given by
\begin{align}
\left(k_+-q_\chi \frac{e^2Q}{r}\right)f_\chi+\left(\frac{\del}{\del r}+\frac{3/2+j}{r}\right)g_\chi=0,\label{eq:emap1}\\
\left(k_+-q_\chi \frac{e^2Q}{r}\right)g_\chi-\left(\frac{\del}{\del r}+\frac{1/2-j}{r}\right)f_\chi=0,\label{eq:emap2}
\end{align}
using $\boldsymbol{\sigma}\cdot\nabla\Phi(j,m,j\pm1/2)=\Phi(j,m,j\mp1/2)\left(\del/\del r+(1\pm(j+1/2))/r\right)$, and the solutions are written as the following forms.
\begin{align}
f_\chi(r)&=\frac{e^{ik_+r}}{r}\biggl[r^{s_0}C_+\biggl[{}_1F_1(s_0+iq_\chi e^2Q,2s_0+1;-2ik_+r)\biggr.\biggr.\nonumber\\
&+\left.\frac{s_0+iq_\chi e^2Q}{j+1/2} {}_1F_1(s_0+1+iq_\chi e^2Q,2s_0+1;-2ik_+r)\right]\nonumber\\
&+r^{-s_0}C_-\biggl[{}_1F_1(-s_0+iq_\chi e^2Q,-2s_0+1;-2ik_+r)\biggr.\nonumber\\
&+\left.\left.\frac{-s_0+iq_\chi e^2Q}{j+1/2} {}_1F_1(-s_0+1+iq_\chi e^2Q,-2s_0+1;-2ik_+r)\right]\right],
\\
g_\chi(r)&=i\frac{e^{ik_+r}}{r}\biggl[r^{s_0}C_+\biggl[{}_1F_1(s_0+iq_\chi e^2Q,2s_0+1;-2ik_+r)\biggr.\biggr.\nonumber\\
&+\left.\frac{s_0+iq_\chi e^2Q}{j+1/2} {}_1F_1(s_0+1+iq_\chi e^2Q,2s_0+1;-2ik_+r)\right]\nonumber\\
&-r^{-s_0}C_-\biggl[{}_1F_1(-s_0+iq_\chi e^2Q,-2s_0+1;-2ik_+r)\biggr.\nonumber\\
&+\left.\left.\frac{-s_0+iq_\chi e^2Q}{j+1/2} {}_1F_1(-s_0+1+iq_\chi e^2Q,-2s_0+1;-2ik_+r)\right]\right],
\end{align}
where $s_0=\sqrt{(j+1/2)^2-(q_\chi e^2Q)^2}$, and we used the confluent hypergeometric function, which is defined as follows.
\begin{align}
{}_1F_1(a,b;z)\equiv\sum_{k=0}^\infty\frac{a(a+1)\cdots(a+k-1)}{b(b+1)\cdots(b+k-1)}\frac{z^k}{k!}
\end{align}
Using the following asymptotic form of the confluent hypergeometric function,
\begin{align}
{}_1F_1(a,b;z)\sim\Gamma(b)\left(\frac{e^zz^{a-b}}{\Gamma(a)}+\frac{(-1)^{-a}z^{-a}}{\Gamma(b-a)}\right),\,\,\,\,\, |z|\gg1
\end{align}
we find that the solutions behave as
\begin{align}
f_\chi&\sim \left[C_+\frac{s_0+iq_\chi e^2Q}{j+1/2}\left((-1)^{s_0}\frac{(-i)^{-s_0+iq_\chi e^2Q}\Gamma(2s_0+1)}{\Gamma(s_0+1+iq_\chi e^2Q)}\right)\right.\nonumber\\
&\left.\,\,\,\,\,+C_-\frac{-s_0+iq_\chi e^2Q}{j+1/2}\left((-1)^{-s_0}\frac{(-i)^{s_0+iq_\chi e^2Q}\Gamma(-2s_0+1)}{\Gamma(-s_0+1+iq_\chi e^2Q)}\right)\right]\times\frac{e^{-ik_+r+iq_{\chi}e^2Q\log(2kr)}}{r}\nonumber\\
&\,\,\,\,\,+\left[C_+\left((-1)^{s_0}\frac{i^{-s_0-iq_\chi e^2Q}\Gamma(2s_0+1)}{\Gamma(s_0+1-iq_\chi e^2Q)}\right)\right.\nonumber\\
&\left.\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,+C_-\left((-1)^{-s_0}\frac{i^{s_0-iq_\chi e^2Q}\Gamma(-2s_0+1)}{\Gamma(-s_0+1-iq_\chi e^2Q)}\right)\right]\times\frac{e^{ik_+r-iq_{\chi}e^2Q\log(2kr)}}{r},\\
g_\chi&\sim \left[C_+\frac{s_0+iq_\chi e^2Q}{j+1/2}\left((-1)^{s_0}\frac{(-i)^{-s_0+iq_\chi e^2Q}\Gamma(2s_0+1)}{\Gamma(s_0+1+iq_\chi e^2Q)}\right)\right.\nonumber\\
&\left.\,\,\,\,\,+C_-\frac{-s_0+iq_\chi e^2Q}{j+1/2}\left((-1)^{-s_0}\frac{(-i)^{s_0+iq_\chi e^2Q}\Gamma(-2s_0+1)}{\Gamma(-s_0+1+iq_\chi e^2Q)}\right)\right]\times(+i)\frac{e^{-ik_+r+iq_{\chi}e^2Q\log(2kr)}}{r}\nonumber\\
&\,\,\,\,\,+\left[C_+\left((-1)^{s_0}\frac{i^{-s_0-iq_\chi e^2Q}\Gamma(2s_0+1)}{\Gamma(s_0+1-iq_\chi e^2Q)}\right)\right.\nonumber\\
&\left.\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,\,+C_-\left((-1)^{-s_0}\frac{i^{s_0-iq_\chi e^2Q}\Gamma(-2s_0+1)}{\Gamma(-s_0+1-iq_\chi e^2Q)}\right)\right]\times(-i)\frac{e^{ik_+r-iq_{\chi}e^2Q\log(2kr)}}{r},
\end{align}
at infinity, consisting of incoming and outgoing waves with additional phase factors,~$e^{\pm iq_{\chi}e^2Q\log(2kr)}$.
\end{appendix}
\vspace{1cm}
| 218f48a35d2ab4cd47e558016273be67a14ee177 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Introduction}
Presently, density functional theory (DFT) dominates the field of atomistic and molecular quantum chemistry calculations \cite{J01,B01,K01}. This is mainly due to its relatively low computational cost compared to many other atomistic approaches. The mainstream DFT (i.e. Kohn-Sham DFT (KS-DFT) \cite{K02}) is a slight alteration of the original work of Hohenberg and Kohn (HK-DFT) \cite{H01} where it was proved that the ground state of any many-electron system is completely characterized by its density and the energy functional which permits the system to attain its minimum at the density corresponding to the ground state. However, representing the contribution of the kinetic energy as a density functional (KEDF) \textcolor{black}{($T\left[ \rho\left( \textbf{x} \right) \right] = \int t \left( \textbf{x} \right) d\textbf{x}$) where $t \left( \textbf{x} \right)$ is the kinetic energy density (KED)} has proven to be challenging as the accuracy and applicability of the presently proposed KEDF are generally not sufficient for reliable calculations \cite{K01,P01,C01,A01,FH3,FH4}. So, Kohn and Sham (KS-DFT) suggested an approximate approach where the ``orbitals'' are reintroduced such that the sum of the orbitals densities equals to the exact density of the real system and the kinetic energy is defined as the kinetic energy of the introduced ``fictitious'' system. Computation-wise, this results in a conversion of the problem from 3-dimensional (3D) to $3N$-dimensional where $N$ orbitals are determined by solving the governing $N$-3D equations self-consistently and $N$ is the number of electrons in the calculations.
Recently, the search for an \textit{orbital-free} version of DFT (OF-DFT) has rapidly gained attention \cite{C02,E01,K03,G02,C04}. To apply ``orbital free'' approaches, it is essential to find highly accurate kinetic energy density functionals (KEDF). The subject is not new and its origins date back to the early years of quantum mechanics. Most of the early proposed KEDF are based on particular exactly solvable models, e.g. constant potential with plainwaves solution for Thomas-Fermi model (TF) \cite{T01,F01} or modified planewaves for von Weizsacker KEDF (vW) \cite{V01}. This period was followed by many extensions and further developments on ``specific physical models''. For a comprehensive collection of suggested functionals, we refer the reader to a review by Tran and Wesolowski \cite{T02}. Alternatively, some of the recent developments adopted statistical techniques. For example, Burke and coworkers recently used machine learning to approximate density functional \cite{S01,S02} based on statistical expressions.
In this paper, we revisit this from a different perspective. Instead of starting from a specific physical model or using mathematical tools to improve a KEDF, we begin by asking which forms of KEDF are physically acceptable. To answer this query, an axiomatic approach is used to determine the physically acceptable forms in a general $D$-dimension space. In axiomatic approaches, a set of axioms are used in conjunction to derive equations, expressions, or theorems.
As expected, the resulted expansion captures most of the known forms of one-point KEDFs. By statistically training the forms for a model problem of the non-interacting kinetic energy in 1D (6 terms only), we find that the mean relative accuracy for 1000 randomly generated potentials is orders of magnitudes better than that delivered by standard KEDFs. The accuracy improves with the number of occupied states, and it is better than $10^{-4}$ at four occupied states. Furthermore, it is shown that the free fitting of the coefficients associated with known KEDFs approaches the exactly known values.
\textcolor{black}{
\section*{Kinetic Energy Density Expansion for $D$-Dimensional Space}
}
Herein, an axiomatic approach is used to derive an expansion of physically acceptable forms of KEDF that satisfy the following essential physical requirements; dimensionality, finiteness, compatibility with the virial theorem, and non-negativity of $t \left( \textbf{x} \right)$ . Also, the functional derivative of the final expansion will be derived as it is needed for efficient energy minimization.
\subsection*{The expression of $t \left( \textbf{x} \right)$ as a function of density derivatives}
Generally, $t \left( \textbf{x} \right)$ can be defined as a scalar function of the density and its derivative, i.e.
\begin{equation}
\label{GEF}
t\left( \textbf{x} \right) \equiv f \left( \rho , \nabla \rho , \nabla^2 \rho , \dotsm , \nabla^{(n)} \rho \right) \, .
\end{equation}
For the 1-dimensional case (1D), Shao and Baltin \cite{S03} proved that the highest derivative order is 1 and all the derivatives for $n>1$ must be ruled out based on the compatibility with the differential virial theorem \cite{R01,H02,H03} and $t \left( \textbf{x} \right)$ non-negativity . The same concept can be further extended to $D$-dimensional cases.
\textcolor{black}{Here, we outline Shao and Baltin \cite{S03} proof for 1D case and how it is extended to $D$-dimension. In 1D, the differential virial theorem of fermionic system with Coulombic interaction is
\begin{equation}
\label{DVT1D}
\rho(x) v'(x) = \dfrac{1}{4} \rho''' - 2 \, t'(x) \, .
\end{equation}
At the beginning, we need to represent the above equation in terms of the density and its derivatives only. $t'(x)$, which is directly obtained from Eq. \ref{GEF} (for 1D), is
\begin{equation}
\label{R2}
t'(x) = \sum_{\nu=0}^n \rho^{(\nu+1)} \frac{\partial f}{\partial \rho^{(\nu)}}
\end{equation}
As for $v'(x)$, it is obtained by taking the derivative of Euler equation when applying Eq. \ref{GEF} within DFT framework. So,
\begin{equation}
\frac{d}{dx}\left( \frac{\delta E[\rho]}{\delta \rho} \right)=0
\end{equation}
and hence
\begin{equation}
\label{R3}
v'(x) = - \frac{d}{dx} \left( \frac{\delta E[\rho]}{\delta \rho} \right) = \sum_{\nu=0}^n (-1)^{\nu+1} \frac{d^{\nu+1}}{dx^{\nu+1}} \left( \frac{\partial f}{\partial \rho^{(\nu)}} \right)
\end{equation}
Here, we used the fact that the chemical potential $\mu$ is constant and hence its derivative is zero. By inserting Eq. \ref{R2} and Eq. \ref{R3} in Eq. \ref{DVT1D}, it becomes
\begin{equation}
\sum_{\nu=0}^n \left[ (-1)^{\nu+1} \rho \frac{d^{\nu+1}}{dx^{\nu+1}} \left( \frac{\partial f}{\partial \rho^{(\nu)}} \right) + 2 \rho^{(\nu+1)} \frac{\partial f}{\partial \rho^{(\nu)}} \right] = \dfrac{1}{4} \rho'''
\end{equation}
Finally, the left hand side of the above equation are rearranged \cite{S03} and due to the fact that only $\rho'''$ appears on the right hand side, it is found that for any $n \geq 2$,
\begin{equation}
\frac{\partial^2 f}{\partial \left[\rho^{(n)}\right]^2} = 0 \, .
\end{equation}
This implies that $f$ must be linear with respect of $\rho^{(n)}$. This clearly violates the non-negativity of KED and hence the dependence of $f$ on $\rho^{(n)}$ must be ruled out.}
\textcolor{black}{The extension of the above proof to higher dimensions is conceptually straightforward; but more tedious mathematically. First, the differential virial theorem is extended to $D$-dimension \cite{H03}, and it becomes:
\begin{equation}
\label{DVTDD}
\rho(x) \frac{dv}{dx_\alpha} = \dfrac{1}{4} \frac{d}{dx_\alpha} \left( \nabla^2 \rho \right) - z_\alpha
\end{equation}
where
\begin{equation}
z_\alpha = \sum_{\beta=1}^D \left( \frac{\partial}{\partial x_\beta'} + \frac{\partial}{\partial x_\beta''} \right) \frac{\partial^2 \gamma\left( \textbf{x} + \textbf{x}' ; \textbf{x} + \textbf{x}'' \right)}{\partial x_\alpha' \partial x_\beta''} \vert_{\textbf{x}' = \textbf{x}'' = \textbf{0}}
\end{equation}
and $\gamma\left( \textbf{x} + \textbf{x}' ; \textbf{x} + \textbf{x}'' \right)$ is the one-particle density matrix. $z_\alpha$ can be decomposed into two parts as follow:
\begin{equation}
z_\alpha = 2 \frac{dt}{dx_\alpha} + \tilde{z}_\alpha
\end{equation}
Then, we take the derivative of Euler equation in $D$-dimesion. The resulted equation is
\begin{equation}
\frac{dv}{dx_\alpha} = \sum_{\nu=0}^n (-1)^{\nu+1} \frac{d}{dx_\alpha} \left[ \nabla^\nu \left( \frac{\partial f}{\partial \left[ \nabla^\nu \rho \right]} \right) \right]
\end{equation}
Finally, by inserting the above equations in the differential virial equation (Eq. \ref{DVTDD}). It can be rewritten as
\begin{widetext}
\begin{equation}
\sum_{\nu=0}^n \left[ (-1)^{\nu+1} \rho \frac{d}{dx_\alpha} \left[ \nabla^\nu \left( \frac{\partial f}{\partial \left[ \nabla^\nu \rho \right]} \right) \right] + 2 \frac{d \left[ \nabla^\nu \rho \right]}{dx_\alpha} \frac{\partial f}{\partial \left[ \nabla^\nu \rho \right]} \right] + \tilde{z}_\alpha = \dfrac{1}{4} \frac{d}{dx_\alpha} \left( \nabla^2 \rho \right)
\end{equation}
\end{widetext}
With further arrangement and because only $\frac{d}{dx_\alpha} \left( \nabla^2 \rho \right)$ term appears on the right hand side, it can be shown that
\begin{equation}
\frac{\partial^2 f}{\partial \left[\nabla^n \rho\right]^2} = 0 \, .
\end{equation}
for $n \geq 2$. Again, this implies that $f$ must be linear with respect of $\nabla^n \rho$ and this clearly violates the non-negativity of KED. Thus, it must be ruled out.
}
On the other hand, the Laplacian (beside the gradient) arises naturally from the definition of kinetic energy operator \cite{A01}. However, there exists some ambiguity concerning it \cite{A01,S04,E01}. As for $t \left( \textbf{x} \right)$, unphysical pathological negativity arises if the dependence on $\nabla^2 \rho$ is of odd-power and $\nabla^2 \rho$ is negative itself. This agrees with the conclusion of Shao and Baltin eliminating dependence on $\nabla^2 \rho$. Thus, the scalar function \textcolor{black}{compatible with the differential virial theorem} becomes
\begin{equation}
\label{GEFr}
t\left( \textbf{x} \right) \equiv f \left( \rho , \nabla \rho \right) \, .
\end{equation}
In addition, it is known that the spatial extension of the density plays an important role in $t \left( \textbf{x} \right)$. Previously, this was accounted for by using either the total number of electrons $N$ or the physical coordinates, $\textbf{x}$ (please see \cite{T01} and the references within). Using $N$ --as a number-- ignores details associated with $\rho \left( \textbf{x} \right)$ while the explicit inclusion of $\textbf{x}$ violates the invariance under coordinate transformations. In this work, we incorporate the effect of the spatial extension of the density by a single measure, $r_d$, which is related to the trace of the covariance matrix $\boldsymbol{\sigma}=[\sigma_{ij}]$. It is defined conventionally as:
\begin{equation}
\label{rd1}
r_d^2 = \dfrac{1}{D} \sum_i^D \sigma_{ii}^2 \quad,
\end{equation}
where:
\begin{equation}
\label{cmg}
\sigma_{ij}^2 = \dfrac{\int \rho(\textbf{x})\left( x_i - \mu_i \right)\left( x_j - \mu_j \right) \textbf{dx}}{\int \rho(\textbf{x}) \, \textbf{dx}}
\end{equation}
and
\begin{equation}
\label{sav}
\mu_i = \dfrac{\int \rho(\textbf{x}) \, x_i \, \textbf{dx}}{\int \rho(\textbf{x}) \textbf{dx}}.
\end{equation}
By inserting Eqs. \ref{cmg}-\ref{sav} in Eq. \ref{rd1}, it becomes
\begin{equation}
\label{rd2}
r_d^2 = \dfrac{1}{D} \dfrac{\iint \rho(\textbf{x}) \rho(\textbf{x}') \, \textbf{x} \cdot \left( \textbf{x} -\textbf{x}' \right) \, \textbf{dx} \, \textbf{dx}'}{\left[ \int \rho(\textbf{x}) \textbf{dx} \right]^2} \,.
\end{equation}
Eq. \ref{rd2} may simply be reduced to the normalized standard deviation in the 1-dimensional case. Thus, the spatial extension is derived solely from $\rho(\textbf{x})$. More interestingly, the consideration of the spatial extension leads to an additional origin of non-locality in KEDF as anticipated by many \cite{H04,K01,C03,W01,W02,P02,P03}. This shall be investigated further in a future work.
However, it is important to highlight that $r_d$ could result in some inconsistencies. The covariance matrix, $\boldsymbol{\sigma}$, (which is used to estimate $r_d$) accounts for many spatial aspects concurrently. As an example, it accounts for the spatial dispersion of a single clustered density. On the contrary, in the case of a pair of ``non-overlapping densities'', $\boldsymbol{\sigma}$ is dominated by the distance between the centers of the two densities rather than their individual spatial extensions. Thus, there is still a need to have a more consistent ``measure'' for the spatial extension of the density. This shall not alter the form of KEDF; however, $\dfrac{\delta T}{\delta \rho}$ must be modified according to the new conventionally defined ``measure''.
With all these considerations plus the scalar nature of $t \left( \textbf{x} \right)$, a possible general expansion of KEDF is given by:
\begin{equation}
\label{DExp01}
t\left( \textbf{x} \right) = \sum_{sln} a_{sln} \frac{1}{r_d^s} \rho^{\frac{l}{D}} \left( \nabla \rho \cdot \nabla \rho \right)^{n/2},
\end{equation}
where $a_{sln}$ are the expansion coefficients and shall be determined through statistical training.
\subsection*{The limits of the sum}
Obviously, the general expansion (Eq. \ref{DExp01}) may have infinite terms as $s$, $l$, and $n$ can take any integer value. However, this is governed by physics and hence there are limits and interconnections between the three indices. By seeking the proper dimensionality of $t\left( \textbf{x} \right)$ ($L^{-D-2}$ in atomic unit, where $L$ is the dimension of the length), it is found that:
\begin{equation}
\label{lexp}
l = \left( 2-s-n \right) + \left( 1-n \right)D \,,
\end{equation}
This permits a description of $l$ entirely in terms of $s$ and $n$ hence the expansion is reduced to:
\begin{equation}
\label{DExp02}
t\left( \textbf{x} \right) = \sum_{sn} a_{sn} \frac{1}{r_d^s} \rho^{\frac{1}{D}\left[ \left( 2-s-n \right) + \left( 1-n \right)D \right]} \left( \nabla \rho \cdot \nabla \rho \right)^{n/2} \,.
\end{equation}
The limits of the sum are determined by other essential physical considerations. Fundamentally, since $\rho(\textbf{x})$ is finite, $n$ must be non-negative. Concerning $s$, for uniform density --which is a proper quantum mechanical case-- $r_d \rightarrow \infty$; thus, $s$ must be non-negative as well. The upper limits of $s$ and $n$ are determined by considering finite systems where $\rho\left( \textbf{x} \right)$ vanishes exponentially. For such systems, \textcolor{black}{$\rho(\textbf{x})=f(\textbf{x}) \exp\left( - \kappa \sqrt{x^2} \right)$} where \textcolor{black}{$\kappa$ is} positive and non-zero. By applying this restriction, it can be shown that for such systems, Eq. \ref{DExp01} is further reduced to:
\begin{equation}
\begin{split}
t\left( \textbf{x} \right) = & \sum_{sn} \frac{a_{sn}}{r_d^s} \left[ f^{\frac{l}{D}} + \left( \nabla f \cdot \nabla f + \kappa^2 f^2 \textcolor{black}{- 2 \, \kappa \, f \frac{\nabla f \cdot \textbf{x}}{x^2} } \right)^{n/2} \right] \\
& \times \exp \left\lbrace -\left( \frac{l}{D} + n \right) \textcolor{black}{ \kappa \sqrt{x^2} } \right\rbrace.
\end{split}
\end{equation}
By forcing $t\left( \textbf{x} \right)$ to vanish at infinity,
\begin{equation}
\frac{l}{D} + n > 0 \, ,
\end{equation}
and by using Eq. \ref{lexp}, we are lead to
\begin{equation}
s+n<D+2 \,,
\end{equation}
by using Eq. \ref{lexp}. This defines the upper limits of each expansion iterator. So, the general form Eq.\ref{DExp01} is reduced to the following expansion in $D$-dimensional space:
\begin{equation}
\label{DExp}
t\left( \textbf{x} \right) = \sum_{s=0}^{D+1} \sum_{n=0}^{D+1-s} a_{sn} \frac{1}{r_d^s} \rho^{\frac{1}{D}\left[ \left( 2-s-n \right) + \left( 1-n \right)D \right]} \left( \nabla \rho \cdot \nabla \rho \right)^{n/2} \,.
\end{equation}
This expansion captures most of the known forms of one-point KEDFs. For example, by setting, $s=n=0$, the resulted form $\rho^{(D+2)/D}$, which is simply TF KEDF in $D$-dimension. If $s=0$ and $n=2$, the resulted KEDF is vW, i.e. $\vert \nabla \rho \vert^2 / \rho$.
The axiomatic approach determines the possible KEDF forms that satisfy the essential physical requirements of dimensionality and finiteness. But, the expansion coefficients, $a_{ns}$, must be determined by other means. In this work, we will determine these $a_{ns}$ statistically in an analogous method to the machine learning approach. However, the derived functional forms are in principal universal and their coefficients must be accordingly universal. Later, we will show that this statistical approach leads to the properly known coefficients in the cases of the TF and vW limits.
\subsection*{The functional derivative}
As the application of DFT requires the minimization of the total energy as a functional of the density and subjected to some constraints, we must find the functional derivative of the kinetic energy functional:
\begin{equation}
\label{KEF}
T[\rho\left( \textbf{x} \right)] = \int t\left( \textbf{x} \right) \, \textbf{dx} \, .
\end{equation}
Using the general derived form of $t\left( \textbf{x} \right)$ as in Eq. \ref{DExp}, we found that:
\footnotesize
\begin{multline}
\label{DEexD}
\dfrac{\delta T}{\delta \rho} = \sum_{s=0}^{D+1} \sum_{n=0}^{D+1-s} a_{sn} \frac{1}{r_d^s} \left\lbrace \rho^\frac{l}{D} \left( \nabla \rho \cdot \nabla \rho \right)^{n/2-2} \left[ \frac{l}{D} \left( 1 - n \right) \frac{1}{\rho} \left( \nabla \rho \cdot \nabla \rho \right)^2 - n \left( \nabla \rho \cdot \nabla \rho \right) \nabla^2 \rho - \frac{n}{2} \left( n - 2 \right) \left[ \nabla \rho \cdot \nabla \left( \nabla \rho \cdot \nabla \rho \right) \right] \right] \right. \\
+ \left. s \int \rho(\textbf{x}')^{\frac{l}{D}} \left( \nabla \rho(\textbf{x}') \cdot \nabla \rho(\textbf{x}') \right)^{n/2} \textbf{dx}' \left[ \dfrac{1}{\int \rho(\textbf{x}') \textbf{dx}'} - \frac{1}{2} \frac{\int \rho(\textbf{x}') \left( \textbf{x}-\textbf{x}' \right) \cdot \left( \textbf{x}-\textbf{x}' \right) \textbf{dx}'}{\iint \rho(\textbf{x}') \rho(\textbf{x}'') \, \textbf{x}' \cdot \left( \textbf{x}' -\textbf{x}'' \right) \, \textbf{dx}' \, \textbf{dx}''} \right] \right\rbrace \, .
\end{multline}
\normalsize
\section*{Results and discussion for 1D cases}
In this work, we consider only non-interacting fermions in one-dimension where the occupied states are assumed to be doubly and fully occupied closed-shell. In such case, Eqs. \ref{DExp} and \ref{DEexD} are reduced to:
\begin{equation}
\label{DExp1D}
t\left( x \right) = \sum_{s=0}^{2} \sum_{n=0}^{2-s} a_{sn} \frac{1}{r_d^s} \rho^{\left( 3-s-2n \right)} \left( \rho'^2 \right)^{n/2} \,
\end{equation}
and
\begin{multline}
\label{DEexD1D}
\dfrac{\delta T}{\delta \rho} = \sum_{s=0}^{2} \sum_{n=0}^{2-s} a_{sn} \frac{1}{r_d^s} \left\lbrace \left(3-s-2n\right) \left( 1 - n \right) \rho^{\left(2-s-2n\right)} \left( \rho'^2 \right)^{n/2} - n \left( n - 1 \right) \, \rho^{\left( 3-s-2n \right)} \left( \rho'^2 \right)^{\left(\frac{n}{2}- 1 \right)} \rho'' \right. \\
\left. + s \int \rho^{\left(3-s-2n\right)} \left( \rho'^2 \right)^{n/2} dx' \left[ \dfrac{1}{\int \rho(x') dx'} - \frac{1}{2} \frac{\int \rho(x') \left( x-x' \right)^2 dx'}{\iint \rho(x') \rho(x'') \, x' \left( x' -x'' \right) \, dx' \, dx''} \right] \right\rbrace
\end{multline}
The resulted expansion terms of 1D KEDF are shown in Table-1. However, we still need a mechanism to determine the expansion coefficients. In principle, these coefficients must be universal and should be determined entirely by physical consideration. For example, in the TF limit, the known coefficient for $\rho^3$ (TF model) is $a_{00}=\pi^2/24$ while $a_{02}=1/8$ is the vW coefficient. However, and aforementioned in this work, these coefficients are determined statistically by using training sets of known kinetic energies and densities for given potentials. Then, these coefficients are used to calculate the kinetic energy for a new set of "test" potentials using $t\left( x \right)$ with densities both numerically obtained by solving Schr\"odinger equation and resulting from DFT minimization. This process is represented schematically in Figure-\ref{Prcs}. It was found that the expansion coefficients converged rapidly with the size of the training set for a given number of occupied states. To guarantee consistency, training sets of 1000 potentials are used throughout this paper.
\begin{table}[H]
\centering
\begin{tabular}{cccccc}
\hline
& $s$ & $n$ & & KEDF & \\
\hline
& 0 & 0 & & $\rho^3$ & \\
& 0 & 1 & & $\rho |\rho'|$ & \\
& 0 & 2 & & $\dfrac{\left( \rho' \right)^2}{\rho}$ & \\
& 1 & 0 & & $\dfrac{1}{r_d} \rho^2$ & \\
& 1 & 1 & & $\dfrac{1}{r_d} |\rho'|$ & \\
& 2 & 0 & & $\dfrac{1}{r_d^2} \rho$ & \\
\hline
\end{tabular}
\label{T1}
\caption{The expansion terms of KEDF in 1D.}
\end{table}
\begin{figure}[ht]
\centering
\includegraphics[width=3.5in]{ScrPrcs.eps}
\caption{The statistical method used to determine the expansion coefficients that are then applied in $t\left( x \right)$ to calculate the kinetic energy and to implement DFT minimization.}
\label{Prcs}
\end{figure}
The considered class of potentials is the one used by Burke and coworkers which consists of three different Gaussian dips (GD) confined to a 1D box of length $L=1$ and between two infinite walls \cite{S01,L01}. This class of potentials possesses the functional form:
\begin{equation}
\label{v1}
v(x) = - \sum_{i=1}^3 a_i \exp \left[ \dfrac{- \left( x-b_i \right)^2 }{2 c_i^2} \right] \quad,
\end{equation}
where $a_i$, $b_i$, and $c_i$ are generated randomly, and obey the following constraints: $1<a<10$, $0.4<b<0.6$, and $0.03<c<0.1$. An efficient spectral method is used to solve for the states and hence the densities with an accuracy greater than $10^{-12}$ for the exact non-interacting kinetic energy ($T_s$) \cite{FH1,FH2} .
\subsection*{$T[\rho]$ calculated using the density numerically obtained through solving the Schr\"odinger equation}
\begin{figure}[ht]
\centering
\includegraphics[width=3.2in]{Fig2.eps}
\caption{The relative error in the non-interacting kinetic energy $T_s$. The top three piles are for full fitting where three upper limits of $s$ are assumed; namely 0, 1, and 2. The middle set of three piles are for the case where we use $a_{00}=\pi^2/24$ and $a_{02}=1/8$ to enforce the TF and vW (TFvW) limits, respectively. The last three piles are for the standard TF, vW, and TFvW models.}
\label{F02}
\end{figure}
A thousand potentials are generated and solved; the exact $T_s$ is then calculated for various numbers of occupied states ($N=2,\,3,\,4,\,5$). The Exact $T_s$ are used to find the best least square fitting for the expansion coefficients. Figure-\ref{F02} gives the performance of the expansion in terms of mean relative error ($|T_s - \int_0^L t(\rho(x)) dx|/T_s$). The top three piles are for full fitting where three upper limits of $s$ are assumed; namely 0, 1, and 2. The middle set of three piles are for the case where we use $a_{00}=\pi^2/24$ and $a_{02}=1/8$ to enforce the TF and vW (TFvW) limits, respectively. The last three piles are for the standard TF, vW, and TFvW models.
Clearly, the new expansion (Eq.\ref{DExp}) results in better performance when compared to the standard models by at least two orders of magnitude. The accuracy is improved further as the number of the occupied states increases. Also, it is clear that the consideration of the spatial extension of the density -- by using $r_d$ -- improves the estimation even further. Note that only 4 parameters (6 under the case of full fitting) are needed to very accurately estimate the non-interacting kinetic energy for 1000 potentials with four different occupied states.
The second analysis was designed to test the validity of the expansion. At the beginning, 1000 training potentials are used to find the expansion coefficients. These are then used to estimate $T$ directly for another 1000 test potentials for $N=2,\,3,\,4,\,5$. The results are shown in Figure-\ref{F03} where the left histograms (blue) are for the training set while the right histograms (green) are for the test set. The left column is for full fitting while the right column is for the case where $a_{00}$ and $a_{02}$ are set to $\pi^2/24$ and $1/8$ and only 4 parameters are statically determined. Here, the error is given in kcal/mol. Distinctly, the statistically trained expansion coefficients were able to estimate $T_s$ very accurately. The errors mean absolute, standard deviation, and max absolute are shown in Table-2 and are given in kcal/mol. However, they are still much beyond the chemical accuracy limit (1 kcal/mol). Furthermore and as aforementioned, the family of potentials used is the same as was used by Burke and coworkers to find KEDF with machine learning \cite{S01,L01}. In that work, they used a process containing around 100,000 empirical parameters. They achieved accuracies below 1 kcal/mol and two order of magnitudes better than what is obtained in this work. However to iterate, in this work we used only 6 parameters.
\begin{figure}[ht]
\centering
\includegraphics[width=4.0in]{Fig3.eps}
\caption{The percentage histograms of error (in kcal/mol) of 1000 training potentials and 1000 test potentials and for different number of occupied states ($N=2,\,3,\,4,\,5$). The left column is for full fitting while the right column is for the case where $a_{00}$ and $a_{02}$ are set based on the TF and vW models. The left histograms (blue) are for the training set while the right histograms (green) are for the test set.}
\label{F03}
\end{figure}
\begin{table} [t]
\centering
\begin{tabular}{ccccccc}
\hline
$N$ & \multicolumn{3}{c}{Full Fitting} & \multicolumn{3}{c}{Fitting with TFvW}\\
& $\overline{|\Delta T|}$ & $|\Delta T|_{std}$ & $|\Delta T|_{max}$ & $\overline{|\Delta T|}$ & $|\Delta T|_{std}$ & $|\Delta T|_{max}$ \\
\hline
2 & 40.4 & 38.0 & 283.4 & 38.9 & 37.4 & 344.3 \\
3 & 22.3 & 21.7 & 164.0 & 37.9 & 35.8 & 330.1 \\
4 & 11.9 & 10.4 & 89.0 & 53.6 & 47.2 & 268.1 \\
5 & 12.0 & 10.1 & 55.4 & 19.1 & 18.9 & 123.2 \\
\hline
\end{tabular}
\label{Table2}
\caption{The errors mean absolute ($\overline{|\Delta T|}$), standard deviation ($|\Delta T|_{std}$), and max absolute ($|\Delta T|_{max}$) in kcal/mol.}
\end{table}
\subsection*{$T[\rho]$ calculated using the density resulted from DFT minimization}
\begin{figure*} [ht]
\centering
\includegraphics[width=6.0in]{DenMin.eps}
\caption{The calculated densities from energy minimization using both full fitting (blue lines) and the fitting with forced TFvW (red lines) beside the exact density (black lines) for two randomly generated potentials and for various number of occupied states ($N=$2, 3, 4, as 5 as indicated in the top right side of each of the panels). The top set is for the first potential (shown in the top left panel) while the bottom set is for the second potential (shown in the bottom left panel).}
\label{F04}
\end{figure*}
\begin{figure}[ht]
\centering
\includegraphics[width=4.0in]{Fig5.eps}
\caption{The percentage histograms of error in relative values (left column) and in kcal/mol (right column) of 1000 potentials and for different number of occupied states ($N=2,\,3,\,4,\,5$). In each panel, the left side (blue) is for the full fitting while the right one (green) is for the fitting with TFvW imposed.}
\label{Fig5}
\end{figure}
In the previous subsection, the exact density (as calculated using high order spectral methods) are used to estimate the kinetic energy. However in DFT, it is essential to find the density that minimize the energy while maintaining the number of involved particles. There are many possible techniques \cite{S05,F02}. In this work, we use a gradient-based trust-region self-consistent method \cite{T03,F03,S06,B02}. Figure-\ref{F04} shows the calculated densities resulted from energy minimization using both full fitting and the fitting with forced TFvW beside the exact density for two randomly generated potentials and for various number of occupied states ($N=$2, 3, 4, as 5). It is clear that the suggested $t\left( x \right)$ (as tested in 1D cases) accurately produces the minimum densities. Also, the results of full fitting are better that those of the fitting with imposed TFvW parameters. However, the main outcome is its capacity to maintain the shell structure, which is challenging for the standard approximation based on TFvW models.
Energy minimization is then applied to find the densities and the kinetic energies of the 1000 systems used in the previous subsection for $N=2,\,3,\,4,\,5$. The results are shown in Figure-\ref{Fig5} where the left column shows the relative error while the right column shows the absolute error in kcal/mol. In each panel, the left side (blue) is for the full fitting while the right one (green) is for the fitting with TFvW imposed. As can be seen, the error is less than 1\% and in most cases, it is less than 0.1\%. However, this is large as absolute value. Also as observed in the previous subsection, it is clear that full fitting results in more accurate calculations as expected from the calculated densities by DFT minimization. There could be many causes for the increased accuracy beside the additional degrees of freedom allowed by full fitting. However, we believe that this could be a numerical error due to the shape optimization. This becomes more apparent for higher occupied states.
\subsection*{On the expansion coefficients}
As stated earlier, the expansion coefficients are determined statistically in this work. However, they must be universal. The only two known parameters are $a_{00}$ which equal to $\pi^2/24$ (TF KEDF) and $a_{02}$ which equal to $1/8$ (vW KEDF). vW KEDF must be equal to $T_s$ for single occupied states. Thus, a fitting for various potentials with single occupied state must get reduced to vW KEDF. This was obtained for various number of training potentials while the other expansion coefficients in this case are fitted trivially to zero. However, they must not all vanish. Rather, the collective contributions of all the other KEDF terms -- in this limit -- must vanish.
In the other extreme, by increasing the number of the occupied states, $a_{00}$ approaches the TF limit as shown in Figure-\ref{F03}. In this analysis, the employed potentials are: a particle in a box (PiB) of a width of 1; a simple harmonic oscillator (SHO) with $\omega=1$; and three randomly generated DG potentials (Eq.\ref{v1}) but with a scaling of $L$, $b_i$, and $c_i$ by factors of 1, 2, and 3. The exactly solvable models (i.e. PiB and SHO) expeditiously reached the exact limit and the resulted $a_{00}$'s for both are hardly distinguishable.
\begin{figure}[ht]
\centering
\includegraphics[width=3.5in]{a00.eps}
\caption{The evolution of $a_{00}$ with the number of the occupied states for various potentials. The red line is the exact limit of $\pi^2/24$. (PiB: a particle in a box of a width of 1, SHO: a simple harmonic oscillator with, $\omega=1$, and GD$_m$: the randomly generated Gaussian dips potential (Eq.\ref{v1}) with a scaling of $L$, $b_i$, and $c_i$ by factors of 1, 2, and 3).}
\label{F05}
\end{figure}
The other expansion coefficients shall be determined by physical considerations rather than statistical fitting. This is one of many imperative goals and should permit implementing the "OF-DFT" forms to calculate the non-interacting kinetic energy in KS-DFT without solving the computationally expensive KS equations. Also, the expansion shall be implemented within a real 3D environment by application to the Coulombic potentials, where there are 15 expansion terms.
\section*{Conclusion}
The general expansion -- which was derived axiomatically through considering physical necessities -- can become a systematic approach to develop density functionals, one of the main challenges existing in the field. This shall reattain the 3D nature of DFT calculations (i.e. OF-DFT), which shall permit conduction of large-scale calculations which were inconceivable by KS-DFT. For example, Gavini and others \cite{G01} studied metallic systems with multi-million atoms by using OF-DFT. And recently, EA Carter and others analyzed a system of more than 1 million lithium atoms \cite{C04}. However, using OF-DFT with the standard KEDF for non-metallic system is inaccurate as they are known to be sufficient only for nearly-uniform density systems like metals. So, the proposed KEDF expansion should allow a systematic improvement of OF-DFT accuracy and applicability.
\nocite{*}
| c5031db9e115b573a18674b366a2b1d62c6e1b45 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The physics of excitons and associated optical phenomena was greatly influenced by recent discoveries in the domain of transition metal dichalcogenide (TMD) materials [\onlinecite{Novoselov2016}]. They can both exist in the bulk and monolayer configurations and possess a set of peculiar properties which differ them from other semiconductor materials. TMD monolayers are of particular interest in this context --- they are atomically thin, are characterized by the direct bandgap favorable for optical interband transitions and exciton formation and demonstrate peculiar interplay of spin and valley effects. All this give them certain advantages for the use in optoelectronics as compared to semimetallic graphene.
Extensive studies of excitonic properties of TMD monolayers started immediately after their discovery [\onlinecite{Mak2010}]. In striking contrast to bulk and quasi-2D structures, the different screening in 2D monolayer governs the deviation of the interparticle Coulomb interaction from the standard form, and ultimately leads to unusual properties of the excitons in TMD structures [\onlinecite{Keldysh,Cudazzo2011}].
The exciton binding energies and absorption spectra in various TMD monolayers were measured experimentally [\onlinecite{Zhao,Shan,Zhu,Ceballos,Shang,Schmidt}] and calculated from the first principles [\onlinecite{Komsa1,Komsa2,Yakobson,Qiu1,Qiu2,Rama,Berkelbach,Lambrecht}]. The results have shown huge increase of the exciton binding energy (up to 1~eV) [\onlinecite{Rasmussen2015}], as compared to conventional semiconductors, and the non-hydrogenic behavior of the excitonic series [\onlinecite{ChernikovPRL}].
Further investigations cover measurements of exciton lifetimes and linewidths in monolayers [\onlinecite{Schaibley,Robert,Selig}], as well as electric field control of the excitonic properties [\onlinecite{Heinz,Scharf}]. Moreover, the rich many-body physics in TMD materials was confirmed by observation of more complex particles, such as trions and biexcitons [\onlinecite{You,Zhang,Sie,Plechinger}] as well as interlayer excitons in bilayer structures [\onlinecite{Rivera,Bellus,Butov}]. Additionally, the hybrid exciton-electron systems in TMDs were considered [\onlinecite{Sidler2017}].
\begin{figure}
\includegraphics[width=0.8\linewidth]{sketch.eps}
\caption{Sketch of the system. A transition metal dichalcogenide monolayer hosts excitonic quasiparticles formed by electrons (blue circles) and holes (red circles). The scattering between two excitons corresponds to the Coulomb interaction between carriers, consisting of the direct and exchange contributions. The latter is dependent on the exciton wavefunction overlap, shown in green.}
\label{fig:sketch}
\end{figure}
Excellent optical properties of TMD monolayers put them as a prominent platform for optoelectronical applications. For instance, the large binding energy of excitons allowed to study excitonic physics at elevated temperatures, and observe the excitons with high principal quantum numbers [\onlinecite{ChernikovPRL}]. The large oscillator strength allows to couple excitons strongly to an optical microcavity mode, and study strong light-matter coupling at room temperature [\onlinecite{Dufferwiel,Lundt1,Lundt2}]. The particular spin-orbit interaction for the bands leads to non-trivial valley dynamics and spin properties, also suggested to be potentially interesting for the quantum information processing [\onlinecite{Wang2016}]. Finally, $\chi_2$ nonlinear response of TMD monolayers was predicted, making it suitable for the observation of the nonlinear quantum optical effects [\onlinecite{Glazov2017}]. %
There are several experimental investigations of TMD monolayer properties in the strong excitation regime, manifesting itself in various intriguing phenomena, including spectral peak broadening [\onlinecite{Sim}], exciton-exciton annihilation [\onlinecite{Kumar}], and giant bandgap renormalization (up to 500 meV) in the vicinity of Mott transition [\onlinecite{ChernikovNat,Aivazian}]. However, to the best of our knowledge, theoretical investigations of the interexciton interactions in TMD monolayers are lacking so far.
Motivated by the aforementioned advances, we consider the nonlinear properties of excitons in a TMD monolayer. The system is well-suitable for the observation of highly excited states of excitons, and similarly to bulk semiconductors [\onlinecite{Kazimierczuk2014}], can allow for studying nonlinear interaction between Rydberg excitons. In the paper, we calculate the exciton-exciton interaction in TMD structures, considering both ground and excited states of excitons. We find that the interaction of excited states exhibits non-monotonic dependence on the exchanged momentum and is attractive. We provide the analytical formula to quantitatively estimate the maximal exciton-exciton interaction strength, which differs from those for the III-V group semiconductors. Finally, we calculate the exciton-electron matrix elements of scattering for both direct and exchange terms.
\section{Excitonic spectrum in TMD monolayer}
To study the interparticle interactions in TMD monolayers, one should take into account structural peculiarities of such materials. Namely, the atomic thickness of the layer and discontinuity of the dielectric screening on the monolayer interface modifies the Coulomb interaction to the following form [\onlinecite{Keldysh}]:
\begin{equation}
V(r)=\frac{e_1 e_2}{4 \pi \varepsilon_0} \frac{\pi }{2 r_0}\left[H_0\left(\frac{r}{r_0}\right)-Y_0\left(\frac{r}{r_0}\right)\right],
\label{pot_0}
\end{equation}
where $e_1$, $e_2$ denote the charge of particles, $r$ is the interparticle distance, and $r_0$ is a quantity describing the polarizability of the monolayer. $H_0$ and $Y_0$ are zeros order Struve and Bessel functions of the first kind, respectively. The modification of Coulomb interaction results in the qualitative change of the excitonic spectrum [\onlinecite{Berkelbach,ChernikovPRL}], which in this case cannot be considered as common 2D hydrogenic spectrum of the form $E_n=\mu e^4 /[2(4 \pi \varepsilon_0\varepsilon)^2\hbar^2(n-1/2)^2]$, where $n$ is a principal quantum number of the exciton, $\mu$ is reduced mass of an electron- hole pair, $\varepsilon$ corresponds to the static dielectric screening constant, and $\varepsilon_0$ is the vacuum permittivity.
The excitonic states should be found as eigenstates of the Hamiltonian
\begin{equation}
\hat{H}_{\mathrm{exc}}=-\frac{\hbar^2}{2\mu}\Delta+V(r),
\label{ExcHam}
\end{equation}
with $V(r)$ is taken in the form of Eq. (\ref{pot_0}). As first approximation one can use variational method, where the trial functions are similar to the conventional 2D excitonic functions and excitonic Bohr radius plays a role of variational parameter [\onlinecite{Portnoi}]:
\begin{align}
\label{psi_n}
&\psi_{n,m}(r)=\frac{1}{\sqrt{2}\lambda_{n}}\sqrt{\frac{(n-|m|-1)!}{(n+|m|-1)!(n-1/2)^3}} \\ \notag
&\hspace{10mm} \left(\frac{r}{(n-1/2)\lambda_{n}}\right)^m \exp\left[-\frac{r}{(2n-1)\lambda_{n}}\right] \\ \notag
&\hspace{10mm} L_{n-|m|-1}^{2|m|}\left[\frac{r}{(n-1/2)\lambda_{n}}\right] \frac{1}{\sqrt{2\pi}} e^{im\varphi}.
\end{align}
Here $L_n^m[x]$ denotes associated Laguerre polynomial, $\lambda_{n}$ is a variational parameter, and $m$ is an angular momentum quantum number. Contrary to the conventional quantum well exciton, where all states have the same radial characteristic --- two-dimensional Bohr radius, in the case of a monolayer the spatial parameter $\lambda_n$ changes from state to state.
To be specific, we consider WS$_2$ monolayer, noting however that all results are of general character and are applicable for the whole family of TMD monolayers. Accurate calculation of exciton series confirmed by experimental data was done in Ref. [\onlinecite{ChernikovPRL}], where the value of polarizibility parameter $r_0$ was found to be equal to $7.5$~nm. Here, we reproduce these results by the binding energy minimization using $\lambda_n$ as a variational parameter. The corresponding values of the exciton energies and spatial characteristics $\lambda_n$ are presented in the Table 1. Note that while the energies of the lower states are essentially non-hydrogenic, for the states starting from $n=3$ the conventional $n^{-2}$ energy dependence can be observed. Correspondingly, the saturation of the $\lambda_n$ values can be seen for higher states.
\begin{table}
\centering
\begin{tabular}{|c|c|c|}
\hline
n & $\lambda_{n}~(\mathrm{nm})$ & $E_{n}$ (meV) \\ \hline
1 & 1.7 & 320 \\ \hline
2 & 0.65 & 160 \\ \hline
3 & 0.45 & 90 \\ \hline
4 & 0.35 & 60 \\ \hline
5 & 0.3 & 50 \\ \hline
\end{tabular}
\caption{Spatial characteristics ($\lambda_n$) and energies of excitons ($E_n$) of different states $n$ calculated for the WS$_2$ monolayer.}
\end{table}
\section{Exciton-exciton interaction}
Analyzing the asymptotic behavior of the potential given by Eq. (\ref{pot_0}) one can find its accurate approximate expression [\onlinecite{Cudazzo2011}]
\begin{equation}
V(r)= - \frac{e_1 e_2}{4 \pi \varepsilon_0} \frac{1}{r_0}\left[\mathrm{ln}\left(\frac{r}{r+r_0}\right)-(\gamma-\mathrm{ln}2)e^{-\frac{r}{r_0}} \right],
\label{pot}
\end{equation}
which is used in further calculations. $\gamma$ denotes Euler gamma constant. To calculate interactions between TMD monolayer excitons in the ground and excited states, we employ the method similar to those used by us before for the case of III-V semiconductor quantum well structures [\onlinecite{Shahnazaryan2016}]. It represents the generalization of the Coulomb scattering formalism for the ground state excitons in quantum wells developed in the Ref. [\onlinecite{Tassone1999,Ciuti1998}]. The wavefunction of an exciton with a wave vector $\textbf{Q}$ can be written in the form
\begin{equation}
\Psi_{\mathbf{Q},n,m}(\mathbf{r}_e,\mathbf{r}_h)=\frac{1}{\sqrt{A}}\exp[i\mathbf{Q}(\beta_e\mathbf{r}_e+\beta_h\mathbf{r}_h)]\psi_{n,m}(|\mathbf{r}_e-\mathbf{r}_h|),
\label{Psi}
\end{equation}
where $\mathbf{r}_e, \mathbf{r}_h$ are the radius vectors of an electron and a hole, respectively, $A$ denotes the normalization area. The coefficients $\beta_e, \beta_h$ are defined as $\beta_{e(h)}=m_{e(h)}/(m_e+m_h)$, where $m_{e(h)}$ is the mass of an electron (hole). The wavefunction of relative motion of an electron and a hole motion is described by Eq. (\ref{psi_n}).
\begin{figure}[t]
\includegraphics[width=1.0\linewidth]{scheme.eps}
\caption{The schematic representation of the exciton-exciton scattering \textbf{(a)-(d)}. Panels correspond to: (a) direct, (b) exciton exchange, (c) electron exchange, and (d) hole exchange interactions. Blue and red solid lines denote an electron ($e_1$) and a hole ($h_1$) of the first exciton exciton, and the dashed lines correspond to an electron and a hole of the second exciton marked with $e_2$ and $h_2$. The exciton-electron scattering diagrams are shown in panels \textbf{(e)-(f)}, describing the direct (e) and exchange (f) interaction. Green solid line ($e_c$) denotes a free electron.}
\label{fig:scheme}
\end{figure}
We consider the interaction of the excitons in the same states with parallel spin projections. In this case the process of Coulomb scattering in reciprocal space with transfer of wave vector $\mathbf{q}$ can be presented in the form
\begin{equation}
(n,m,\mathbf{Q}) + (n,m,\mathbf{Q}') \rightarrow (n,m,\mathbf{Q}+\mathbf{q})+(n,m,\mathbf{Q}'-\mathbf{q}),
\label{scattering}
\end{equation}
which can be represented graphically by scattering diagrams in Fig. \ref{fig:scheme}.
Using the wave function symmetrization procedure the total interaction may be presented as linear combination of the interaction channels, including direct interaction, and electron, hole, exciton exchange terms, as schematically depicted in Fig. \ref{fig:scheme} (a)-(d). It was shown previously [\onlinecite{Ciuti1998,Shahnazaryan2016,Tassone1999}] that in the wide region of exchanged wave vectors $q\leq 1/\lambda_1$ the interaction of excitons is determined by the exchange terms, while the direct interaction in negligibly small. The latter becomes dominant for large values of $\textbf{q}$, governing the long range behavior of the interaction. Assuming the initial wave vectors being equal and setting $\textbf{Q}=\textbf{Q}'=0$, for the total interaction we have (see Appendix A for the details and definitions):
\begin{align}
\label{Htot}
&V_{\mathrm{tot}}(n,m,\mathbf{q})=\frac{e^2}{4 \pi \varepsilon_0}\frac{\lambda_{1}}{A} I_{\mathrm{tot}}(n,m,q\lambda_1),\\
&I_{\mathrm{tot}}(n,m,q\lambda_1) =I_{\mathrm{dir}}(n,m,q\lambda_1)+I^X_{\mathrm{exch}}(n,m,q\lambda_1) \notag \\
&+I^e_{\mathrm{exch}}(n,m,q\lambda_1)+I^h_{\mathrm{exch}}(n,m,q\lambda_1) \approx 2 I^e_{\mathrm{exch}}(n,m,q\lambda_1),
\label{Itot}
\end{align}
where indices $e$, $h$, $X$ stand for the electron, hole, and exciton exchange integrals, respectively.
In principle, the interaction processes between excitons with different spin projections can be accounted for. However, they involve spin-flip processes, and typically contain only direct interaction channel [\onlinecite{Kyriienko2012}].
\begin{figure}[t]
\includegraphics[width=1.0\linewidth]{exc_exc.eps}
\caption{Dimensionless integrals corresponding to the direct interaction matrix element (a), and total interaction energy (b) of excitons in TMD monolayer, plotted as function of the transferred wave vector. Solid lines correspond to $s$ states ($m=0$) and dashed lines denote $p$ states ($m=1$). The interaction has similar form for both type excitons, being attractive for the excited states. The difference appears for $n=2$ state, where for $s$ state there is attraction with the absolute maxima at intermediate momenta, and for $p$ state there are attraction and repulsion regions.}
\label{exc}
\end{figure}
We calculate direct and total interaction as a function of the scattered momentum exploiting the multidimensional Monte-Carlo integration [\onlinecite{Hahn_VEGAS}]. The results of the calculation are shown in Fig. \ref{exc}.
The direct interaction as function of the exchanged momentum is repulsive, and its peak-shaped dependence becomes narrower with increase of the principal quantum number of the scattered excitons. The total interaction is fully governed by the exchange term, which is non-zero at $q \lambda_1 \rightarrow 0$. It is repulsive for the ground state and attractive for the excited states. This behavior is qualitatively similar to quantum well exciton interaction [\onlinecite{Shahnazaryan2016}].
However, the screened nature of Coulomb interaction imposes peculiarities in the TMD exciton-exciton interaction behavior. Namely, the crucial difference of the monolayer exciton interaction appears in the dependence of $2s$ state interactions, which demonstrate potential minima for the nonzero exchange momenta $q$. It should be noted that the interaction of $2p$ excitons demonstrates similar properties, being repulsive for zero exchange momenta and having attraction peak at intermediate momenta. This non-monotonic behavior can be expected to lead to different condensation processes for TMD polaritons [\onlinecite{Matuszewski2012}].
\begin{figure}[t]
\includegraphics[width=1.0\linewidth]{scaling.eps}
\caption{ (a) The dependence of exciton radius (blue curve) and energy (red curve) on the principal quantum number. (b) Exchange interaction energy as a function of the principal quantum number of a TMD exciton. The green curve corresponds to the numerical calculation, and the red line shows the qualitative estimate using Eq. (\ref{Uexch_simp}).}
\label{fig:scaling}
\end{figure}
Next, we search for the compact analytical formula to describe the exciton-exciton interaction in TMD monolayers, considering both ground and excite states scattering. Previously it was shown that the exchange interaction of GaAs quantum well ground state excitons can be described by the formula $V_{\mathrm{exch}}^{\mathrm{QW}}= {6 E_b a_B^2}/A$, where $a_B$ and $E_b$ denote Bohr radius and binding energy of quantum well exciton, respectively [\onlinecite{Tassone1999}]. The numerical prefactor 6 comes from the calculation of exchange integrals.
Following the analogy, we search for similar dependence for the exciton series in TMD monolayer. Fig.~\ref{fig:scaling}(a) presents the dependence of the radius and energy of exciton states on their principal quantum number. While the radius increases quadratically (as in the case of the conventional Rydberg series), the energy dependence for the first few states drops superpolynomially with $n$. The latter allows us to approximate the exchange interaction dependence by the formula
\begin{equation}
\label{Uexch_simp}
V_{\mathrm{exch}}(n)={\alpha E_n R_n^2}/A,
\end{equation}
where $R_n$ and $E_n$ denote the radius and energy of $n$-th exciton state, respectively, and $\alpha$ is a fitting constant. The green line in Fig. \ref{fig:scaling}(b) denotes the dependence of the exchange interaction strength on the principal quantum number, unveiling close-to-linear dependence starting from the $n=2$ state. The red curve shows the estimate by Eq. (\ref{Uexch_simp}), where we chose the parameter $\alpha=2.07$, which gives the exact fitting for the ground state. It is worth mentioning that despite the smaller pre-factor, for the case of the quantum well with the similar material parameters the interaction would be weaker. Namely, taking the reduced effective mass characteristic to WS$_2$ monolayer, $\mu = 0.16 m_0$ [\onlinecite{ChernikovPRL}], the interaction strength between ground state excitons in a quantum well can be estimated as $V_{\mathrm{exch}}^{\mathrm{QW}}= 6 \frac{\hbar^2}{2\mu A}$. Comparing it to the TMD estimate $V_{\mathrm{exch}}^{\mathrm{TMD}}= 2.07 E_{1s} \lambda_{1s}^2/A$, where $E_{1s}$ and $\lambda_{1s}$ values are taken from Table 1, we get the ratio
\begin{equation}
V_{\mathrm{exch}}^{\mathrm{TMD}} / V_{\mathrm{exch}}^{\mathrm{QW}} = 1.34.
\label{ratio}
\end{equation}
The reason beyond this is peculiar interaction screening in TMD monolayers, leading to the exciton effective radius value larger than for the conventional Coulomb potential.
It should be noted that the close agreement between the exact calculation of interaction and its qualitative estimate is possible only because of the rapid decrease of the exciton energy for the lower excitonic states in TMD monolayer. On the contrary, in semiconductor heterostructures the energy drops quadratically, $E_n \sim n^{-2}$, obeying Rydberg rule. The corresponding estimate thus predicts quadratic growth of the interaction strength. However the exact calculation shows the linear dependence of the exchange term on quantum number for quantum well [\onlinecite{Shahnazaryan2016}], meaning that the estimate is not reasonable for that case.
\begin{figure}[t]
\includegraphics[width=1.0\linewidth]{screening.eps}
\caption{(a) Maximum of exchange interaction energy plotted as a function of the screening length $r_0$. Interaction between different excitonic states is considered. (b) Maximum of exchange interaction energy of ground excitonic states plotted as a function of the screening length $r_0$. Here, $r_0$ axis is extended to show small screening length limit, and we use the logarithmic scale. }
\label{fig:screening}
\end{figure}
Finally, we proceed with the discussion of the influence of screening on the interexciton interaction. While previously we focused on a structure of particular configuration, discussed in Ref. [\onlinecite{ChernikovPRL}], the obtained results are expected to be qualitatively valid for other configurations as well. For instance, an additional factor is the presence of a substrate, which can substantially modify the optical properties of a sample. Particularly, the strong modulation of monolayer bandgap by the dielectric environment was studied in Ref. [\onlinecite{Ryou}]. In our model the screening length $r_0$ is influenced by the dielectric permittivity of a monolayer ($\varepsilon$) and substrate ($\varepsilon_{1,2}$) as $r_0=d \varepsilon / (\varepsilon_1+\varepsilon_2)$, where $d$ denotes the thickness of a monolayer, and $\varepsilon_{1,2}$ stand for dielectric permittivity of substrate and cover layer, respectively [\onlinecite{Keldysh, Berkelbach}]. Here we vary the screening length in a wide range and study it influence on the exciton-exciton interaction strength. In Fig. \ref{fig:screening}(a) we plot the exchange interaction energy as a function of screening length for ground ($11$) and excited ($22,33,44$) excitonic $s$ states. Here, we choose the realistically achievable values of $r_0$, which can be tuned by the substrate choice. We observe that for the excited states the growth of $r_0$ leads to the decrease of interaction energy, despite the actual increase of exciton radius $\lambda_n$. This effect can be explained by the evidence that the interaction potential (\ref{pot}) itself decreases rapidly, thus overcoming the impact of interexciton interaction enhancement coming from the exciton wavefunction spread. Hence, one may further increase the XX interaction strength by the reduction of $r_0$, which can be reached by the choice of substrate with large dielectric permittivity.
The situation is different for the exchange interaction of ground state excitons. Fig. \ref{fig:screening} (a) indicates that in the plotted range of screening length the interaction strength varies weakly. This can be seen as a consequence of the non-hydrogenic nature of $1$s excitons in TMDs, and explained as a mutual compensation of interaction enhancement from exciton radius growth and intracarrier interaction decrease. For better understanding we explore the limit $r_0 \rightarrow 0$, where the interaction potential (\ref{pot}) reduces to conventional 2D Coulomb form [\onlinecite{Cudazzo2011}]. In Fig. \ref{fig:screening} (b) we plot the interaction of ground state excitons as a function of screening length in logarithmic scale. We observe that in the above mentioned limit the interaction rapidly decreases, and becomes about $1.3$ times smaller than in screened interaction limit (large $r_0$). Notably, this value is in a close agreement with the previously presented estimate of the ratio (\ref{ratio}) between interactions of excitons in QW and TMD.
We would like to remark also, that in the limit $r\rightarrow 0$ the interaction of excited excitons do not undergo rapid changes, and continues smooth increase (not shown). Such a striking difference of screening length dependence of interaction for ground state and excited states is a direct consequence of the fact, that ground excitonic state in the TMD materials is essentially non-Rydbergian, while excited states demonstrate Rydberg-like behavior [\onlinecite{ChernikovPRL}].
\section{Exciton-electron scattering}
In this section we consider n-doped TMD monolayer with excess of the free electrons which can interact with optically created excitons. This nonlinear process is especially relevant for up to date TMD experiments [\onlinecite{Heinz}], can contribute to the exciton line broadening [\onlinecite{Selig}], and determines the physics of TMD exciton-polarons [\onlinecite{Sidler2017}].
\begin{figure}
\includegraphics[width=1.0\linewidth]{e_exc.eps}
\caption{Exciton electron interaction energy as a function of transferred momentum. Ground state direct and exchange interactions (a), direct (b), and exchange(c) interaction of the excited states.}
\label{e_exc}
\end{figure}
We proceed with the calculation of the exciton scattering with conduction band electrons. We restrict our consideration to $s$ states, noting that for $p$-type excitons the results are expected to be similar. The conduction band electron wave function is given by a plane wave $f_{\mathbf{K}}(\mathbf{\rho})=(1/\sqrt{A}) e^{i \mathbf{K} \mathbf{\rho}}$, where $\mathbf{K}$ denotes an electron momentum. We consider the process of Coulomb scattering of an exciton with an electron, corresponding to the momentum transfer process
\begin{equation}
(n,\mathbf{Q})+(\mathbf{K}) \rightarrow (n,\mathbf{Q}+\mathbf{q})+(\mathbf{K}-\mathbf{q}).
\end{equation}
Possible interaction channels include direct interaction and the electron exchange term, as shown in Fig. \ref{fig:scheme}(e,f). Correspondingly, one can present the total interaction as sum of the direct and electron exchange contributions:
\begin{align}
&U(n,\mathbf{Q},\mathbf{K},\mathbf{q}) = U_{\mathrm{dir}}(n,q)+U_{\mathrm{exch}}(n,\textbf{q}, \mathbf{K}-\beta_e\mathbf{Q}) \notag \\
&= \frac{e^2}{4\pi \varepsilon_0} \frac{\lambda_1}{A} \left[u_{\mathrm{dir}}(n,q\lambda_1)+u_{\mathrm{exch}}(n,q\lambda_1, (K-\beta_e Q)\lambda_1)\right],
\end{align}
where the explicit form of the corresponding terms is given in Appendix B. It should be noted that the described approach is in agreement with the method previously used to characterize the exciton-electron scattering in quantum well heterostructures [\onlinecite{Ramon}]. We calculated the scattering of free electron with ground and excited state TMD excitons. Without the loss of generality, it is convenient to put the condition $\mathbf{K}-\beta_e\mathbf{Q}=0$. Fig. \ref{e_exc}(a) illustrates the direct and exchange terms of $1s$ exciton scattering with an electron. Similarly to QW heterostructure, the interaction is governed by exchange contribution. Fig. \ref{e_exc}(b,c) shows direct and exchange interaction of excited excitons with electron, respectively. One can see that similarly to the exciton-exciton interaction both components of scattering amplitudes increase with principal quantum number, conserving the domination of the exchange component. An additional feature is that unlike for the ground state, for the excited states the interaction is attractive and has maxima appearing at intermediate exchange momenta. Moreover, with the increase of quantum number both interaction components become more peak shaped.
\section{Conclusion}
In conclusion, we considered theoretically the exciton-exciton and exciton-electron scattering processes in transition metal dichalcogenide monolayers. We found that unusual screening of the Coulomb interaction characteristic to TMD monolayers leads to the non-monotonic dependence of the exchange interaction on the transferred momentum. We have shown that contrary to the conventional quantum well excitons the interaction can be accurately estimated by a simple analytical formula. It is proportional to the product of the exciton binding energy and the square of exciton radius, and exhibits linear growth with the principal quantum number of exciton. We have studied the dependence of interaction on the dielectric permittivity of a substrate, and have shown that while for excited exciton states interaction increases for the samples with high dielectric permittivity substrates, the ground state interaction cannot be enhanced.
Additionally, we calculated the exciton-electron interaction in TMD monolayers, relevant for systems with excess of free electrons. This interaction is characterized by dominant attractive contribution of the exchange component increasing with the principal quantum number of exciton. The results provide the basis for quantitative description for nonlinear effects in TMD systems, and are important for the design of corresponding nonlinear optoelectronic devices.
\section*{Acknowledgments}
The authors are grateful to D. Gulevich and M. Glazov for fruitful discussions. This work was supported by Icelandic Research Fund, Grant No. 163082-051 and mega-grant No. 14.Y26.31.0015 of the Ministry of Education and Science of Russian Federation. O.K. thanks University of Iceland for the hospitality during the work on the project, and acknowledges the funding by FP7 ERC Grant QIOS (Grant No. 306576). V.S and I.A.S acknowledge support from Ministry of Education and Science of Russian Federation, goszadanie no 3.2614.2017/4.6 and Horizon2020 RISE project CoExAN.
| b96f5cda26e8bc58d83b88ba8b3b28043293d469 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:intro}
In targeted online advertising, the main goal is to figure out the best opportunities by showing an advertisement to an online user, who is most likely to take a desired action, such as ordering a product or signing up for an account. The complexity of realizing this goal is so high that advertisers need specialized technology solutions called demand-side platforms (DSP).
In a DSP, each individual advertiser usually sets up a list of campaigns that can be thought of as plans for delivering advertisements. For each campaign, the advertiser specifies the characteristics of the audience segments that she would like to target (e.g., males, ages 18-35, who view news articles on espn.com) along with the particular media that she would like to display to the target audience (e.g., a video ad for beer). In addition, the advertiser specifies a budget amount, time schedule, pacing details, and performance goals for each campaign. The performance goals typically can be specified by minimizing cost-per-click (CPC) or cost-per-action (CPA).
The DSP manages its active campaigns for many different advertisers simultaneously across multiple ad exchanges where ad impressions can be acquired through a real-time bidding (RTB) process. In the RTB process, the DSP interacts with several ad exchanges where bids are placed for potential impressions on behalf of those advertisers. This interaction happens in real time when an ad request is submitted to an ad exchange (which may happen, for example, when a user views a news story on a webpage). In this scenario, the DSP needs to offer a solution to decide, among the list of all campaigns associated with its advertiser clients, which campaign to bid on behalf of and how much to offer for the corresponding bid. The fundamental problem we consider here is how to make these decisions in real time to maximize the profit for the DSP while ensuring that all of its advertiser clients are satisfied.
Generally speaking, today most DSPs offer different pricing models and enhancement schemes to help advertisers manage their campaigns. Those pricing models include cost per thousand impressions (CPM), cost per click (CPC) and cost per action (CPA). Advertisers often like to choose the CPC/CPA pricing model because the return on investment (ROI) is always positive. However, this pricing model might introduce revenue loss for the DSP since the DSP only earns revenue when a click or action occurs. Therefore, in CPC/CPA pricing model, the DSP needs to convert the CPC/CPA bid to an expected cost per thousand impressions (eCPM) bid in order to sensibly participate in auctions in the RTB exchanges. In this paper, we focus on CPC/CPA pricing model as it is a very challenging problem for the DSP.
It is challenging for a DSP to perform such profit optimization with CPC/CPA pricing model in a RTB environment for several reasons. First, top DSPs typically receive as many as a million ad requests per second. The short latency and high throughput requirements introduce extreme time sensitivity on the process. Second, a large amount of information is missing in the real time evaluation of the individual ad requests, e.g., the feedback on previous decisions normally has a long delay in practice. Therefore, most of the DSPs today only apply a greedy approach by selecting the ad with the highest bid among all the qualified ads for each incoming request.
In this paper, we propose a novel approach based on a precise mathematical formulation to optimize the overall DSP profit. We appropriately model the uncertainty in impression arrival, auction, and click/action processes and develop an optimization formulation to maximize profit for the DSP while ensuring that each campaign remains under budget. Our formulation is aimed at optimizing with respect to both impression allocation and bid price decisions, and due to the additional complexity of accounting for both of these decisions the formulation is a large-scale, nonconvex model. However, due to the properties of second-price auctions we are able to effectively use the technique of Lagrangian relaxation. We construct a dual problem and establish that subgradients of the dual function may be efficiently computed. Our overall approach is based on a two-phase procedure, wherein we solve the dual problem in the first phase and use the dual solution to naturally recover a primal solution in the second phase. We conduct several computational experiments on synthetic datasets and demonstrate that our Lagrangian relaxation based approach is able to significantly increase DSP profits relative to a baseline greedy approach.
Revenue optimization in online advertising has been extensively studied in recent literature from different perspectives, such as optimization \cite{mehta2007adwords, balseiro2014yield, chen2011real, zhang2014optimal}, game theory and mechanism design \cite{balseiro2015repeated, maehara2015budget}, and contract design \cite{mirrokni2017deals}. Due to space limitations, we only review several directly relevant papers. \cite{balseiro2014yield, chen2014dynamic, chen2011real} focus on the publisher's revenue management problem. Specifically, \cite{balseiro2014yield, chen2014dynamic} study how publishers should optimally trade off guaranteed contracts with RTB. \cite{chen2011real} studies how a publisher should optimally allocate impressions and set up bid prices for campaigns, under the implicit assumption that the publisher is a ``central planner''. On the other hand, \cite{ciocan2012model} studies ad networks' revenue management problem based on model predictive control. Finally, \cite{zhang2014optimal} studies advertisers' optimal bidding problem in RTB. Unlike our paper, \cite{zhang2014optimal} focuses on optimal bidding and its framework does not consider impression allocation.
The rest of the paper is organized as follows.
In Section \ref{sec:model}, we describe the notation and problem statement and we set up the model. The profit optimization formulation is presented in Section \ref{sec:optalg}. Our proposed Lagrangian relaxation based algorithm to approximately solve the profit optimization problem is specified in Section \ref{sec:LAD}. Experimental results are presented in Section \ref{sec:expts}, and we conclude with some remarks about possible future work.
\section{Model Foundations}\label{sec:model}
We assume that the planning takes place over a fixed time horizon (e.g., 24 hours). To simplify the presentation, we also assume that the DSP interacts with a single ad exchange. For our purposes, the ad exchange simply represents a pool of potential impressions that the DSP may bid on. Thus, in the likely scenario that the DSP interacts with more than one ad exchange, we may simply group all of the different ad exchanges together into a representative ad exchange.
Let $\mathcal{I}$ denote the set of \emph{impression types} and let $\mathcal{K}$ denote the set of all campaigns associated with advertisers managed by the DSP. Before discussing the details of our assumptions, let us describe the basic flow of events in the model. When an impression of type $i \in \mathcal{I}$ is submitted to the ad exchange, a real-time second-price auction is held for which the DSP has an opportunity to bid. Thus the DSP has an opportunity to make two strategic decisions related to each real-time auction: \emph{(i)} how to select a campaign $k \in \mathcal{K}$ to bid on behalf of in the auction, and \emph{(ii)} how to set the corresponding bid price $b_{ik}$. If the DSP wins the auction on behalf of campaign $k$, then the DSP must pay the ad exchange an amount equal to the second largest price and an ad from campaign $k$ is displayed. The advertiser corresponding to campaign $k$ is charged only if a user clicks on the ad.
\paragraph{Impression Types}
It is important that the set of impression types $\mathcal{I}$ represents a partition of all possible impressions that are submitted to the ad exchange. Thus, every impression submitted to the ad exchange is associated with a particular impression type $i \in \mathcal{I}$. It is most natural to define $\mathcal{I}$ in terms of features associated with impressions. For example, if the DSP determines that there are only two relevant attributes associated with each impression -- say gender and whether or not the viewer is 18 years or older -- then the DSP would choose $\mathcal{I} = \{(M, 18-), (M, 18+), (F, 18-), (F, 18+)\}$. Then, all impressions corresponding to male viewers who are under 18 years of age would be assigned to impression type $i = (M, 18-)$, etc. Note that the construction of the set $\mathcal{I}$ is part of the modeling process and consideration should be given to the trade-off between computational limitations and the potential for higher profits due to a more fine-grained construction of $\mathcal{I}$. Nevertheless, the algorithmic schemes we propose in Section \ref{sec:optalg} are scalable to problem instances where the size of $\mathcal{I}$ is extremely large. We use the following notation and make the following assumptions about the impression types:
\begin{itemize}
\item Let $S_i$ denote the number of impressions of type $i$ that arrive during the planning horizon, and assume that $S_i$ is a random variable with mean $s_i$.
\end{itemize}
\paragraph{Campaigns}
Recall that $\mathcal{K}$ denotes the set of all campaigns that are managed by the DSP. That is, $\mathcal{K}$ is the union over all advertisers (who are managed by the DSP) of the sets of campaigns run by each advertiser. We use the following notation and make the following assumptions about the campaigns:
\begin{itemize}
\item $m_k$ denotes the (advertiser selected) budget for campaign $k$ during the planning horizon.
\item $\mathcal{I}_k$ denotes the set of impression types that campaign $k$ targets. For example, if an advertiser wishes to create a particular campaign to target female users, then in this case $\mathcal{I}_k$ would denote the set of all impression types corresponding to female users (e.g., $\mathcal{I}_k = \{(F, 18-), (F, 18+)\}$ in the example described above).
Note that each advertiser can create multiple campaigns to achieve different targeting goals.
\item In this model, it is assumed that advertisers are charged on a \emph{CPC (cost per click)} basis. That is, campaign $k$ is charged an amount $q_k > 0$, called the CPC price, each time a user clicks on an advertisement from campaign $k$. (Note that a ``click'' may also be thought of more generally as an ``action'' whereby our model easily extends to campaigns that are charged on a CPA basis. Moreover, our model may be easily extended to allow for multiple actions, each with their own rewards.)
\end{itemize}
\paragraph{Auctions}
When an impression is submitted to the ad exchange, an instantaneous real-time auction occurs to determine who gets to display an advertisement. We assume that these are \emph{second-price} auctions, which are very common in practice. In a second-price auction, the bidder who submits the highest bid is the winner, but the amount that the winner pays is the amount of the \emph{second highest} bid. It is well known that, in a second-price auction, a dominant strategy for each participant is to bid truthfully \cite{vickrey1961counterspeculation}.
Herein we assume that the DSP takes a \emph{probabilistic} approach to modeling the behavior of the other bidders in the auction. Namely, we make the following assumptions:
\begin{itemize}
\item For each impression type $i \in I$, let $B^{\max}_i$ be a random variable representing the maximum, among all other bidders excluding the DSP, of the bid prices entered in an auction for an impression of type $i$. It is assumed that $B^{\max}_i > 0$ with probability one. Let $\rho_i(\cdot) : \mathbb{R} \to [0,1]$ denote the cumulative distribution function (CDF) of $B^{\max}_i$, so that $\rho_i(b) := \mathbb{P}(B^{\max}_i \leq b)$ is the probability that the DSP wins an auction for an impression of type $i$ when the DSP enters the bid $b$. Note that the functions $\rho_i(\cdot)$ are typically estimated using a bid landscape model (see, e.g., \cite{cui2011bid}).
(In the case of a tie, here we assume that the DSP automatically wins the auction. Our framework may be easily adapted to the case of a fair tie-breaking mechanism.)
\item Furthermore, let $\beta^{\max}_i(b) := \mathbb{E}[B^{\max}_i ~|~ B^{\max}_i \leq b]$ denote the expected value of the second highest bid price (i.e., the value of the payment to the ad exchange) given that the DSP enters a bid price of $b$ and $b$ is the largest bid price entered.
\end{itemize}
\paragraph{Click Events}
After the DSP has won an auction on behalf of campaign $k$, an ad for campaign $k$ is displayed to the user corresponding to the impression for which the auction was held. For a given impression type $i \in \mathcal{I}$ and a given campaign $k \in \mathcal{K}$, let $\theta_{ik} \in [0,1]$ denote the \emph{click-through-rate} for users corresponding to impression type $i$ and when the ad corresponds to campaign $k$. That is, $\theta_{ik}$ represents the fraction of users corresponding to impression type $i$ that click on an ad associated with campaign $k$, i.e., the probability that the user clicks on the ad that is shown.
Although the true click-through-rates are not available, the DSP is typically able to leverage a vast amount of historical data and use predictive models to produce accurate predictions of these values, even when $\mathcal{K}$ and $\mathcal{I}$ are extremely large (see, e.g., \cite{mcmahan2013ad}).
Finally, given an impression type $i \in \mathcal{I}$ and a campaign $k \in \mathcal{K}$, let $r_{ik}$ denote the \emph{expected cost per impression (eCPI)} value, namely $r_{ik} := q_k\theta_{ik}$ where $q_k$ is the CPC price defined earlier. Note that $r_{ik}$ is the expected amount of revenue that the DSP earns each time an ad for campaign $k$ is shown to an impression of type $i$, and $r_{ik}$ also corresponds to the optimal bid price when campaign $k$ has unlimited budget.
\paragraph{Decision Variables and Additional Notation}
As mentioned previously, when an auction for impression type $i \in \mathcal{I}$ arrives to the ad exchange, the DSP decides which campaign $k \in \mathcal{K}$ to bid of behalf of and also selects the value of the corresponding bid price. Let $\mathcal{E} \subseteq \mathcal{I} \times \mathcal{K}$ denote the edges of an undirected bipartite graph between $\mathcal{I}$ and $\mathcal{K}$, whereby there is an edge $e = (i, k) \in \mathcal{E}$ whenever campaign $k$ targets impression type $i$, i.e, $\mathcal{E} := \{(i, k) : i \in \mathcal{I}_k\}$. Let $\mathcal{K}_i := \{k \in \mathcal{K} : (i, k) \in \mathcal{E}\}$ denote the set of campaigns that target impression type $i$.
When a new auction for impression type $i$ arrives to the ad exchange, we say that the DSP selects campaign $k$ for the auction if the DSP chooses to bid on behalf campaign $k$ in the auction. For each edge $(i,k) \in \mathcal{E}$, we define two decision variables as follows: {\em (i)} $x_{ik}$ is the probability that the DSP selects campaign $k$, and {\em (ii)} $b_{ik}$ is the corresponding bid price that the DSP submits in the auction. Interpreted differently, $x_{ik}$ represents a proportional allocation, i.e., the fraction of auctions for impression type $i$ that are allocated to campaign $k$ on average. Note that $b_{ik}$ represents the bid price that the DSP submits to an auction for impression type $i$ \emph{conditional} on the fact that the DSP has selected campaign $k$ for the auction. Related approaches (e.g., as in \cite{chen2011real}) also use bid prices to rank advertisers -- in our approach, the selection of which campaign to bid on behalf of is completely captured by the $x_{ik}$ decision variables and thus the $b_{ik}$ decision variables only determine the actual bid price decisions. Let $\bf{x}, \bf{b} \in \mathbb{R}^{|\mathcal{E}|}$ denote vectors of these quantities, which will represent decision variables in our model.
Let us also define some additional notation used herein. For a given set $S$ and a function $f(\cdot) : S \to \mathbb{R}$, let $\arg\max_{x \in S} f(x)$ denote the (possibly empty) set of maximizers of the function $f(\cdot)$ over the set $S$. If $f(\cdot) : \mathbb{R}^n \to \mathbb{R}$ is a convex function then, for a given $x \in \mathbb{R}^n$, $\partial f(x)$ denotes the set of subgradients of $f(\cdot)$ at $x$, i.e., the set of vectors $g$ such that $f(y) \geq f(x) + g^T(y - x)$ for all $y \in \mathbb{R}^n$. Finally, let $\mathbf{1}(\cdot)$ denote an indicator function that is equal to $1$ whenever the argument of $\mathbf{1}(\cdot)$ is true and equal to $0$ otherwise.
\section{Optimization Formulation}
\label{sec:optalg}
Problem \eqref{poi_deterministic} presents our formulation of the allocation and real-time bidding planning problem faced by the DSP.
\begin{equation}\label{poi_deterministic}
\begin{array}{lcl}
\underset{\mathbf{x}, \mathbf{b}}{\mathrm{maximize}} & & \displaystyle \sum_{(i,k) \in \mathcal{E}} [r_{ik} - \beta^{\max}_i(b_{ik})]s_ix_{ik}\rho_i(b_{ik}) \\
& & \\
\vspace{0.1cm}
\mathrm{subject \ to} & & \textstyle \sum_{i \in \mathcal{I}_k}r_{ik}s_ix_{ik}\rho_i(b_{ik}) ~\leq~ m_k \ \ \forall k \in \mathcal{K} \\
\vspace{0.1cm}
& & \textstyle \sum_{k \in \mathcal{K}_i} x_{ik} ~\leq~ 1 \ \ \forall i \in \mathcal{I} \\
& & \mathbf{x}, \mathbf{b} ~\geq~ 0 \ .
\end{array}
\end{equation}
Herein, let $\pi(\mathbf{x}, \mathbf{b}) := \sum_{(i,k) \in \mathcal{E}} [r_{ik} - \beta^{\max}_i(b_{ik})]s_ix_{ik}\rho_i(b_{ik})$ denote the objective function of \eqref{poi_deterministic}. Let us now briefly describe the interpretation of each part of the formulation \eqref{poi_deterministic} above. First, note that the formulation is based on the idea of ``deterministic approximation,'' whereby we assume that all random quantities deterministically take on their expected values. In this formulation, the DSP seeks to maximize its total profit over the planning horizon, while ensuring that each campaign does not spend more than its budget. Indeed, the expected number of times during the planning horizon that the DSP selects campaign $k$ for an auction of impression type $i$ is $s_ix_{ik}$ and the expected number of such auctions that the DSP wins is $s_ix_{ik}\rho_i(b_{ik})$. Furthermore, for each instance that the DSP selects campaign $k$ and wins the corresponding auction for impression type $i$, the expected profit for the DSP is $r_{ik} - \beta^{\max}_i(b_{ik})$. Therefore, the objective function of \eqref{poi_deterministic} represents the expected total profit earned by the DSP throughout the planning horizon. The first set of constraints in \eqref{poi_deterministic} represent the \emph{budget constraints} for the campaigns, which ensure that, in expectation, each campaign does not spend more than its pre-specified budget level. Finally, the second set of constraints in \eqref{poi_deterministic} are referred to as the \emph{supply constraints} for the impression types, which (along with the nonnegativity constraints on $\mathbf{x}$) ensure that the variables $\mathbf{x}$ represent valid probabilities. Note also that these probabilities may sum to a value strictly less than 1, in which case $1 - \sum_{k \in \mathcal{K}_i} x_{ik}$ represents the probability of electing not to bid when an impression of type $i$ arrives to the ad exchange. For ease of notation, let us denote the supply and nonnegativity constraints on $\mathbf{x}$ using $\mathcal{S} := \left\{\mathbf{x} : \sum_{k \in \mathcal{K}_i} x_{ik} ~\leq~ 1 \ \text{for all } i \in \mathcal{I}, \ \mathbf{x} \geq 0\right\}$.
Note that, due to the joint optimization over both $\bf{x}$ and $\bf{b}$, problem \eqref{poi_deterministic} is generally a nonconvex optimization problem (this is clearly evident, for example, when $B^{\max}_i$ is taken to be uniformly distributed on $[0,1]$ and hence $\rho_i(b_{ik}) = b_{ik}$ for $b_{ik} \in [0,1]$). Despite its nonconvexity, problem \eqref{poi_deterministic} does have some important structural properties that we now highlight. First, if we consider the bid prices $\mathbf{b}$ to be fixed, then the resulting problem in $\mathbf{x}$ is a \emph{linear optimization} problem -- in other words the objective function and constraints can all be expressed as linear functions of $\mathbf{x}$ -- and may be solved very efficiently using off-the-shelf solvers or perhaps a specialized algorithm. Conversely, if we consider $\mathbf{x}$ to be fixed, then the resulting problem in $\mathbf{b}$ is generally still nonconvex but the main ``difficulty'' arises from the budget constraints. Indeed, due to the presence of budget constraints, it may be optimal for the DSP to underbid on a relatively less valuable impression due to the possibility of a more valuable impression arriving in the future. Therefore, whenever a campaign has unlimited budget, it is optimal for the DSP to set $b_{ik} = r_{ik}$, i.e., to bid truthfully. The following Proposition, which will be useful in the development of the Lagrangian relaxation algorithm in Section \ref{sec:LAD}, formalizes this intuition.
\begin{proposition}\label{prop:truth}
Consider the following modification of \eqref{poi_deterministic} without budget constraints:
\begin{equation}\label{poi_prop}
\begin{array}{lcl}
\underset{\mathbf{x}, \mathbf{b}}{\mathrm{maximize}} & & \displaystyle \sum_{(i,k) \in \mathcal{E}} [r_{ik} - \beta^{\max}_i(b_{ik})]s_ix_{ik}\rho_i(b_{ik}) \\
& & \\
\mathrm{subject \ to} & & \mathbf{x} \in \mathcal{S} \qquad
\textrm{ and } \qquad \mathbf{b} ~\geq~ 0 \ ,
\end{array}
\end{equation}
where $\mathcal{S} := \left\{\mathbf{x} : \sum_{k \in \mathcal{K}_i} x_{ik} ~\leq~ 1 \ \text{for all } i \in \mathcal{I}, \ \mathbf{x} \geq 0\right\}$, as defined earlier.
Define $(\mathbf{x}^\ast, \mathbf{b}^\ast)$ by $b_{ik}^\ast := r_{ik}$ for all $(i,k) \in \mathcal{E}$ and by letting $\mathbf{x}^\ast$ be an arbitrary optimal solution of the resulting linear optimization problem, i.e.,
$\mathbf{x}^\ast \in \arg\max\limits_{\mathbf{x} \in \mathcal{S}}\left\{\sum_{(i,k) \in \mathcal{E}}\pi_{ik}x_{ik}\right\} $,
where $\pi_{ik} := [r_{ik} - \beta^{\max}_i(r_{ik})]s_i\rho_i(r_{ik})$ for all $(i,k) \in \mathcal{E}$.
Then, \eqref{poi_prop} is finite and $(\mathbf{x}^\ast, \mathbf{b}^\ast)$ is an optimal solution of \eqref{poi_prop}.
\end{proposition}
\begin{proof}
Fix an arbitrary $\mathbf{x} \in \mathcal{S}$ and consider the resulting problem in the $\mathbf{b}$ variables:
\begin{equation}\label{poi_prop2}
\begin{array}{lcl}
\underset{\mathbf{b}}{\mathrm{maximize}} & & \displaystyle \sum_{(i,k) \in \mathcal{E}} [r_{ik} - \beta^{\max}_i(b_{ik})]s_ix_{ik}\rho_i(b_{ik}) \\
\mathrm{subject \ to} & & \mathbf{b} ~\geq~ 0 \ .
\end{array}
\end{equation}
We now demonstrate that, regardless of the value of $\mathbf{x}$, an optimal solution of \eqref{poi_prop2} is given by $\mathbf{b}^\ast$, i.e., by setting $b_{ik}^\ast = r_{ik}$ for all $(i,k) \in \mathcal{E}$. Indeed, note that \eqref{poi_prop2} is completely separable across the $b_{ik}$ variables. Thus, since $x_{ik} \geq 0$, for each $(i,k) \in \mathcal{E}$ we simply need to independently solve
\begin{equation}\label{poi_prop3}
\begin{array}{lcl}
\underset{b_{ik}}{\mathrm{maximize}} & & [r_{ik} - \beta^{\max}_i(b_{ik})]\rho_i(b_{ik})\mathbf{1}(x_{ik} > 0) \\
\mathrm{subject \ to} & & b_{ik} ~\geq~ 0 \ ,
\end{array}
\end{equation}
where $\mathbf{1}(x_{ik} > 0)$ is an indicator function that is equal to 1 whenever $x_{ik} > 0$ and 0 otherwise. If $x_{ik} = 0$, then the objective function in \eqref{poi_prop3} is always 0 and hence any value of $b_{ik} \geq 0$, in particular $b_{ik}^\ast := r_{ik}$, is an optimal solution. Otherwise, if $x_{ik} > 0$, then the objective function in \eqref{poi_prop3} is just the expected utility when entering a bid of $b_{ik}$ into a second price auction when the valuation is equal to $r_{ik} \geq 0$, for which the dominant strategy is to bid truthfully. Thus, in either case, it is clear that $b_{ik}^\ast := r_{ik}$ is an optimal solution of \eqref{poi_prop3} and it follows that the vector $\mathbf{b}^\ast$ is an optimal solution of \eqref{poi_prop2}. Recall that $\pi(\mathbf{x}, \mathbf{b})$ denotes the objective function of \eqref{poi_deterministic} and hence also the objective function of \eqref{poi_prop}. Then we have shown that $\pi(\mathbf{x}, \mathbf{b}^\ast) \geq \pi(\mathbf{x}, \mathbf{b})$ for all $\mathbf{b} \geq 0$.
Thus, since $\mathbf{x} \in \mathcal{S}$ was selected arbitrarily, we have:
\begin{equation*}
\pi(\mathbf{x}^\ast, \mathbf{b}^\ast) \geq \pi(\mathbf{x}, \mathbf{b}^\ast) \geq \pi(\mathbf{x}, \mathbf{b}) \ \ \text{for all } \mathbf{x} \in \mathcal{S}, \mathbf{b} \geq 0 \ ,
\end{equation*}
which shows that $(\mathbf{x}^\ast, \mathbf{b}^\ast)$ is an optimal solution of \eqref{poi_prop}.
\end{proof}
\begin{remark}\label{calScompute}
Given coefficients $\pi_{ik} \in \mathbb{R}$ for each $(i,k) \in \mathcal{E}$, an optimal solution $\mathbf{x}^\ast$ of the linear optimization problem
$\max\limits_{\mathbf{x} \in \mathcal{S}}\left\{\sum_{(i,k) \in \mathcal{E}}\pi_{ik}x_{ik}\right\}$
may be computed efficiently in $O(|\mathcal{E}|)$ time using a simple greedy algorithm. Namely, for each $i \in \mathcal{I}$, we compute $k^\ast(i) \in \arg\max_{k \in \mathcal{K}_i}\left\{\pi_{ik}\right\}$, set $x_{ik^\ast(i)} = \mathbf{1}(\pi_{ik^\ast(i)} > 0)$, and set $x_{ik} = 0$ for all other $k \in \mathcal{K}_i$.
\end{remark}
\section{Lagrangian Dual and Algorithmic Scheme}\label{sec:LAD}
We begin this section with a high-level description of our approach for solving \eqref{poi_deterministic}. Our algorithmic approach is based on a two phase procedure. In the first phase, we construct a suitable dual of \eqref{poi_deterministic}, which turns out to be a convex optimization problem that can be efficiently solved with most subgradient based algorithms. A solution of the dual problem naturally suggests a way to set the bid prices $\mathbf{b}$. In the second phase, we set the bid prices using the previously computed dual solution then we solve the linear optimization problem that results when $\mathbf{b}$ is fixed in order to recover allocation probabilities $\mathbf{x}$.
Let us now construct a Lagrangian dual of the deterministic approximation problem \eqref{poi_deterministic} by relaxing the ``difficult'' budget constraints. We show that the resulting dual problem is a convex optimization problem with only very simple box constraints and that subgradients of the objective function may be efficiently computed. Since our formulation is quite general (for example, there are no strong assumptions made about the distribution of $B^{\max}_i$), we are unable to exploit any special structure of the dual function $L^\ast(\cdot)$ and must resort to simple subgradient based algorithms to solve the dual problem. Nevertheless, subgradient methods offer the advantage of being highly scalable and parallelizable, and moreover our overall two phase procedure does not necessitate a high accuracy solution of the dual problem.
To start, we introduce multipliers $\lambda \in \mathbb{R}^{|\mathcal{K}|}_{+}$ for the budget constraints in \eqref{poi_deterministic} and form the Lagrangian function:
\begin{equation}\label{original_lag}
\begin{split}
L(\mathbf{x}, \mathbf{b}, \lambda) ~:=~ & \textstyle \sum_{(i,k) \in \mathcal{E}}[r_{ik} - \beta^{\max}_i(b_{ik})]s_ix_{ik}\rho_i(b_{ik}) \\
+ & \textstyle \sum_{k \in \mathcal{K}} \lambda_k \left[ m_k ~-~ \sum_{i \in \mathcal{I}_k}r_{ik}s_ix_{ik}\rho_i(b_{ik}) \right] \ .
\end{split}
\end{equation}
After rearranging, we may re-express the Lagrangian function:
\begin{equation}\label{rearrange}
\begin{split}
L(\mathbf{x}, \mathbf{b}, \lambda) = & \textstyle \sum_{(i,k) \in \mathcal{E} }[(1 - \lambda_k)r_{ik} - \beta^{\max}_i(b_{ik})]s_ix_{ik}\rho_i(b_{ik})\\
& \textstyle + ~\sum_{k \in \mathcal{K}} \lambda_k m_k
\end{split}
\end{equation}
Note that the above implies a natural interpretation of the dual variables as related to scaling factors, namely $1 - \lambda_k$ for each $k \in \mathcal{K}$, to reduce the bid prices based on the fact that each campaign has limited budget. The dual function is defined in the standard way:
\begin{equation}\label{dual_fcn_def}
L^\ast(\lambda) := \max_{\mathbf{x} \in \mathcal{S}, \mathbf{b} \geq 0} \ L(\mathbf{x}, \mathbf{b}, \lambda) \ ,
\end{equation}
and we define the dual problem as:
\begin{equation}\label{dual_prob}
\begin{array}{lcl}
\underset{\lambda}{\mathrm{minimize}} & & \displaystyle L^\ast(\lambda) \\
\mathrm{subject \ to} & & 0 \leq \lambda_k \leq 1 \ \ \text{for all } k \in \mathcal{K} \ .
\end{array}
\end{equation}
Note that the nonnegativity constraints above are standard in Lagrangian duality to ensure that \eqref{dual_prob} provides a valid upper bound on \eqref{poi_deterministic}. The upper bound constraints, i.e., $\lambda_k \leq 1$ for all $k \in \mathcal{K}$, are without loss of generality since whenever $\lambda_k > 1$ the dual function $L^\ast(\lambda)$ is only improved by instead setting $\lambda_k = 1$. It is always the case that $L^\ast(\cdot)$ is a convex function, but in general it may not be differentiable. Nevertheless, Procedure \ref{algo-subgrad} precisely describes how to compute a \emph{subgradient} of $L^\ast(\cdot)$ at $\lambda$ and is based on Proposition \ref{prop:truth}. Theorem \ref{prop_dual_fcn} demonstrates that Procedure \ref{algo-subgrad} computes valid subgradients and also summarizes the most important properties of the dual function $L^\ast(\cdot)$ and the dual problem \eqref{dual_prob}.
\floatname{algorithm}{Procedure}
\begin{algorithm}
\caption{Computing a Subgradient of $L^\ast(\cdot)$}\label{algo-subgrad}
\begin{algorithmic}
\STATE {\bf Input:} $\lambda \in \mathbb{R}^{|\mathcal{K}|}$ such that $0 \leq \lambda_k \leq 1$ for all $k$ . \\
\STATE 1. For each $(i,k) \in \mathcal{E}$, set:
\begin{equation*}
\begin{split}
&b^\ast(\lambda)_{ik} \gets (1 - \lambda_k)r_{ik} \ , \text{ and } \\
&\pi(\lambda)_{ik} \gets [b^\ast(\lambda)_{ik} - \beta^{\max}_i(b^\ast(\lambda)_{ik})]s_i\rho_i(b^\ast(\lambda)_{ik})
\end{split}
\end{equation*}
\vspace{-3mm}
\STATE 2. Compute $\mathbf{x}^\ast(\lambda) \in \arg\max_{\mathbf{x} \in \mathcal{S}}\left\{ \sum_{(i,k) \in \mathcal{E}}\pi(\lambda)_{ik}x_{ik} \right\}$ using the greedy algorithm described in Remark \ref{calScompute}.
\STATE 3. For each $k \in \mathcal{K}$, set:
\begin{equation*}
\textstyle g(\lambda)_k \gets m_k ~-~ \sum_{i \in \mathcal{I}_k}r_{ik}s_ix^\ast(\lambda)_{ik}\rho_i(b^\ast(\lambda)_{ik}) \ .
\end{equation*}
\vspace{-3mm}
\STATE {\bf Output:} $g(\lambda) \in \partial L^\ast(\lambda)$ .
\end{algorithmic}
\end{algorithm}
\medskip
\begin{theorem}\label{prop_dual_fcn}
We have the following properties:
\begin{itemize}
\item[\emph{(i)}] $L^\ast(\cdot)$ is finite and convex everywhere on $\mathbb{R}^{|\mathcal{K}|}$.
\item[\emph{(ii)}] For any $\lambda \in \mathbb{R}^{|\mathcal{K}|}$ such that $0 \leq \lambda_k \leq 1$ for all $k$, Procedure \ref{algo-subgrad} computes a valid subgradient $g(\lambda) \in \partial L^\ast(\lambda)$.
\item[\emph{(iii)}] Let $P^\ast$ denote the optimal objective function value of the primal problem \eqref{poi_deterministic}, and let $D^\ast$ denote the optimal objective function value of the dual problem \eqref{dual_prob}. Then, for any $(\mathbf{x}, \mathbf{b})$ that is feasible for \eqref{poi_deterministic} and any $\lambda$ that is feasible for \eqref{dual_prob}, it holds that:
\begin{equation*}
\pi(\mathbf{x}, \mathbf{b}) \leq P^\ast \leq D^\ast \leq L^\ast(\lambda) \ .
\end{equation*}
\end{itemize}
\end{theorem}
\begin{proof}
Fix $\lambda \in \mathbb{R}^{|\mathcal{K}|}$ such that $\lambda_k \leq 1$ for all $k$ and consider Proposition \ref{prop:truth} where we replace $r_{ik}$ with the modified value $(1 - \lambda_k)r_{ik}$. It is clear that Proposition \ref{prop:truth} also holds with these modified values. Then, in this case, by \eqref{rearrange} it follows that the subproblem appearing in the definition of $L^\ast(\cdot)$ in \eqref{dual_fcn_def} is exactly the same as \eqref{poi_prop} (except for the term $\sum_{k \in \mathcal{K}} \lambda_k m_k$, which is just a constant since $\lambda$ is fixed).
Therefore, by Proposition \ref{prop:truth}, it holds that $L^\ast(\lambda)$ is finite. The case when $\lambda_k$ is possibly greater than 1 for some $k$ is requires a simple extension of Proposition \ref{prop:truth} that allows for possibly negative values of $r_{ik}$. Convexity of $L^\ast(\cdot)$ follows since $L^\ast(\cdot)$ is a pointwise maximum of linear functions.
To prove {\em (ii)}, again by Proposition \ref{prop:truth} it follows that Steps (1.) and (2.) of Procedure \ref{algo-subgrad} are computing a solution of the subproblem in the definition of $L^\ast(\cdot)$ given in \eqref{dual_fcn_def}, i.e., it holds that $(\mathbf{x}^\ast(\lambda), \mathbf{b}^\ast(\lambda)) \in \arg\max_{\mathbf{x} \in \mathcal{S}, \mathbf{b} \geq 0}L(\mathbf{x}, \mathbf{b}, \lambda)$. By \eqref{original_lag}, Step (3.) of Procedure \ref{algo-subgrad} is computing the partial gradient of $L(\mathbf{x}^\ast(\lambda), \mathbf{b}^\ast(\lambda), \lambda)$ (holding $(\mathbf{x}^\ast(\lambda), \mathbf{b}^\ast(\lambda))$ fixed), i.e., it holds that $g(\lambda) = \nabla_{\lambda}L(\mathbf{x}^\ast(\lambda), \mathbf{b}^\ast(\lambda), \lambda)$. Therefore, for any $\lambda^\prime \in \mathbb{R}^{|\mathcal{K}|}$ it holds that:
\begin{align*}
L^\ast(\lambda^\prime) &\geq L(\mathbf{x}^\ast(\lambda), \mathbf{b}^\ast(\lambda), \lambda^\prime) \\
&\geq L(\mathbf{x}^\ast(\lambda), \mathbf{b}^\ast(\lambda), \lambda) + g(\lambda)^T(\lambda^\prime - \lambda) \\
&= L^\ast(\lambda) + g(\lambda)^T(\lambda^\prime - \lambda) \ ,
\end{align*}
which by definition implies that $g(\lambda) \in \partial L^\ast(\lambda)$. The first inequality above follows from the definition of $L^\ast(\cdot)$, the second inequality holds since $L(\mathbf{x}, \mathbf{b}, \lambda)$ is convex in $\lambda$, and the equality holds by $(\mathbf{x}^\ast(\lambda), \mathbf{b}^\ast(\lambda)) \in \arg\max_{\mathbf{x} \in \mathcal{S}, \mathbf{b} \geq 0}L(\mathbf{x}, \mathbf{b}, \lambda)$.
Finally, {\em (iii)} is standard in Lagrangian duality and we omit its proof.
\end{proof}
Algorithm \ref{algo-twophase} presents our two phase procedure for obtaining an approximate solution $(\hat\mathbf{x}, \hat\mathbf{b})$ of problem \eqref{poi_deterministic}. In Phase 1, we solve the dual problem \eqref{dual_prob}. As mentioned previously, we suggest the use of simple subgradient methods (see, for example, \cite{nesterov2013introductory} and the references therein), with the use of Procedure \ref{algo-subgrad} to compute subgradients, in order to solve this problem. In our experiments in Section \ref{sec:expts}, we use the basic method of projected subgradient descent with step-sizes proportional to $1/\sqrt{T}$ where $T$ is the iteration counter. In this case, as evident from Procedure \ref{algo-subgrad}, the subgradients will remain bounded and therefore we may apply classical convergence results for this method, which state that the objective function value optimality gap converges to zero at the rate of $O(1/\sqrt{T})$ \cite{nesterov2013introductory}. Moreover, the per iteration cost of this method is dominated by the cost of computing a subgradient, which, as is clear from Procedure \ref{algo-subgrad} and Remark \ref{calScompute}, is $O(|\mathcal{E}|)$.
\floatname{algorithm}{Algorithm}
\begin{algorithm}
\caption{Two Phase Lagrangian Relaxation-based Scheme for Problem \eqref{poi_deterministic}}\label{algo-twophase}
\begin{algorithmic}
\STATE {\bf Phase 1: Solve Lagrangian Relaxation}
\INDSTATE Solve the dual problem \eqref{dual_prob} to near global optimality
\INDSTATE using a subgradient method, and return dual variables $\hat\lambda$
\INDSTATE and dual objective value $\hat D \gets L^\ast(\hat \lambda)$.
\STATE {\bf Phase 2: Primal Recovery}
\INDSTATE 1. Set bid prices $\hat{\mathbf{b}}$: $\hat{b}_{ik} \gets (1 - \hat{\lambda}_k)r_{ik}$ for all $(i,k) \in \mathcal{E}$.
\INDSTATE 2. Consider the primal problem \eqref{poi_deterministic} with the $\mathbf{b}$ variables
\INDSTATE fixed at the values $\hat{\mathbf{b}}$, and solve the resulting linear
\INDSTATE optimization problem to obtain allocation probabilities $\hat{\mathbf{x}}$.
\STATE {\bf Output:} Approximate primal solution $(\hat{\mathbf{x}}, \hat{\mathbf{b}})$, primal objective value $\pi(\hat\mathbf{x}, \hat\mathbf{b})$, and dual upper bound $\hat D$.
\end{algorithmic}
\end{algorithm}
In Phase 2 of Algorithm \ref{algo-twophase}, we suggest a heuristic to construct an approximate primal solution $(\hat{\mathbf{x}}, \hat{\mathbf{b}})$ based on the previously computed dual solution $\hat \lambda$. First, we use the natural correspondence suggested by \eqref{rearrange} to set the bid prices $\hat{\mathbf{b}}$. Then, we fix these bid prices and solve the resulting linear optimization problem from \eqref{poi_deterministic} to obtain allocation probabilities $\hat{\mathbf{x}}$. In our experiments, we use the off-the-shelf solver Gurobi although more sophisticated and scalable approaches may also be employed. Finally, Algorithm \ref{algo-twophase} outputs the approximate primal solution $(\hat{\mathbf{x}}, \hat{\mathbf{b}})$ along with the dual objective value $\hat D$. Using item {\em (iii)} of Theorem \ref{prop_dual_fcn}, we may use $\hat D$ to obtain a useful bound on the suboptimality of $(\hat{\mathbf{x}}, \hat{\mathbf{b}})$, namely $\pi(\hat{\mathbf{x}}, \hat{\mathbf{b}}) \leq P^\ast \leq \hat D$.
\section{Computational Experiments}\label{sec:expts}
In this section, we present the results of several computational experiments wherein we applied our two-phase solution procedure to synthetic data examples and compared its performance to a baseline policy. Before discussing the experimental results, it is important to clarify how our model, and in particular the output of Algorithm \ref{algo-twophase}, should be applied in a practical, online environment. Policy \ref{algo-online} precisely describes the sequence of events that occur when a new impression of type $i$ arrives to ad exchange and also describes how the decision variables $(\hat{\mathbf{x}}, \hat{\mathbf{b}})$ resulting from Algorithm \ref{algo-twophase} would be used to make decisions in real-time. (Recall that the decision variables $\mathbf{x}$ represent probabilities of selecting campaigns, and that $1 - \sum_{k \in \mathcal{K}_i} x_{ik}$ represents the probability of not participating in an auction for a new impression of type $i$. In Policy \ref{algo-online} below, the symbol ``0'' is used to encode a ``null campaign'' that represents this option of refraining from participating in the auction.)
\floatname{algorithm}{Policy}
\begin{algorithm}
\caption{Online Policy Implied by Algorithm \ref{algo-twophase}
}\label{algo-online}
\begin{algorithmic}
\STATE {\bf Input:} Approximate primal solution $(\hat{\mathbf{x}}, \hat{\mathbf{b}})$ from Algorithm \ref{algo-twophase} and new impression arrival $i \in \mathcal{I}$.
\\ $ \ $ \\
\STATE 1. Sample a campaign $\tilde{k} \in \mathcal{K}_i \cup \{0\}$ according to the distribution implied by the values $x_{ik}$ for $k \in \mathcal{K}_i$ and $1 - \sum_{k \in \mathcal{K}_i} x_{ik}$. If $\tilde k = 0$ (the sampled campaign is null) or the sampled campaign $\tilde k$ has depleted its budget, then break.
\STATE 2. Enter bid price $\hat{b}_{i\tilde{k}}$. If the auction is won, then pay the ad exchange an amount to equal to the second price. If the auction is not won, then break.
\STATE 3. Show an ad for campaign $\tilde k$. If a click happens, then deduct $q_{\tilde k}$ from the budget of campaign $\tilde k$ and earn revenue $q_{\tilde k}$.
\end{algorithmic}
\end{algorithm}
We also refer to Policy \ref{algo-online} as the ``Lagrangian policy.'' Notice that Policy \ref{algo-online} is particularly conservative in dealing with what happens when a campaign depletes its budget. Indeed, if a campaign with depleted budget is sampled then the DSP does not participate in the auction. Policy \ref{algo-online} may be improved by incorporating the idea of model predictive control whereby Algorithm \ref{algo-twophase} is rerun every time a campaign depletes its budget, or possibly at periodic time intervals. We compare Algorithm \ref{algo-twophase} and correspondingly Policy \ref{algo-online} against a simple ``greedy policy'' that is often employed in practice. Indeed, the greedy policy has the same basic flow of events as Policy \ref{algo-online} with two major differences in how decisions are made: {\em (i)} at Step (1.) the greedy policy selects, among those campaigns in $\mathcal{K}_i$ with budgets that are not yet depleted, the campaign $\tilde k$ with the largest eCPI value of $r_{ik}$, and {\em (ii)} at Step (2.) the greedy policy uses $r_{i\tilde{k}}$, the eCPI value of the selected campaign, as the bid price.
\paragraph{Synthetic Data Examples} Let us now describe how the synthetic data examples were generated and how the corresponding simulations were conducted. Throughout this discussion, all relevant random variables are generated independently unless otherwise mentioned. Furthermore, throughout our experiments, it is assumed that the optimization model developed herein is \emph{correctly specified} in that all distributional information used by our model (e.g., in \eqref{poi_deterministic} and Algorithm \ref{algo-twophase}) accurately reflect the corresponding random variables in our simulations.
Now, to generate our synthetic instances, we first fix the sizes of the sets $\mathcal{K}$ and $\mathcal{I}$, and for each campaign $k \in \mathcal{K}$ (and also for each impression type $i \in \mathcal{I}$) we generate a ``quality score'' $Q_k$ (resp., $Q_i$) that is uniformly distributed on $[0,1]$. The quality scores are intended to reflect the ``desirability'' of each impression type and each campaign and are used to generate the main parameters of our model. Indeed, for each impression type $i \in \mathcal{I}$, the set $\mathcal{K}_i$ is constructed by sampling edges independently with probability $Q_i$. Hence $|\mathcal{K}_i| \sim \text{Bin}(|\mathcal{K}|, Q_i)$, where $\text{Bin}(n, p)$ denotes a binomial random variable with $n$ trials and success probability parameter $p$. Moreover, for each impression type $i \in \mathcal{I}$, $B^{\max}_i$ is taken to be the maximum of $\text{Bin}(M, Q_i)$ independent random variables that are uniformly distributed on $[0,1]$, where $M$ is an integer parameter dictating the ``size of the market.'' The click-through-rate values $\theta_{ik}$ are defined by $\theta_{ik} := Q_i \cdot Q_k$.
In all of our experiments, we set $|\mathcal{K}| = 100$, $M = 10$, $s_i = 5000$ for all $i \in \mathcal{I}$, and the CPC value $q_k = 1$ for all $k \in \mathcal{K}$. In our first experiment, we generated a single problem instance, referred to as Example A, that additionally had $|\mathcal{I}| = 100$ and the budget parameters set to $m_k = 50$ for all $k \in \mathcal{K}$. In this case, as verified by the dual upper bound $\hat D$, Algorithm \ref{algo-twophase} was able to solve problem \eqref{poi_deterministic} to within 13\% of optimality. We compared the Policy \ref{algo-online} implied by Algorithm \ref{algo-twophase} to the ``greedy policy'' described earlier by simulating the impression arrival, real-time bidding, and click processes. In our simulations, we assume that impressions arrive to the ad exchange according to $|\mathcal{I}|$ independent Poisson processes that are ``merged'' together. This assumption implies that $S_i$ is a Poisson random variable, and the time horizon $T$ of the overall arrival process is set so that $s_i = 5000$ for each $i \in \mathcal{I}$.
\paragraph{Results} The top left table in Figure \ref{thefig} presents the main results of our first experiment on Example A. We ran 500 simulation runs comparing our Lagrangian relaxation approach, i.e., Policy \ref{algo-online} to the greedy baseline policy. Each policy saw the same exact sequence of impression arrivals and the same sequence of realized $B^{\max}_i$ values during each individual simulation run. For each simulation run, we computed the relative profit, relative cost, and relative revenue of the two policies, where each relative statistic is computed as the Lagrangian statistic relative to the greedy baseline, e.g., $\text{Relative Profit} := \frac{\text{Lagrangian Profit}}{\text{Greedy Profit}}$. The results in the tables are averaged over 500 such simulation runs. As the top left table demonstrates, the Lagrangian policy is able to achieve significantly higher profit levels and lower costs than the greedy baseline. Interestingly, the Lagrangian policy also achieves lower revenue levels than the greedy policy. This makes good intuitive sense since the Lagrangian policy uses bid prices that are shaded down by a factor of $1 - \lambda_k$ as compared to the greedy policy, and moreover the Lagrangian policy should make smarter allocation decisions whereby, under the Lagrangian policy, a particularly valuable campaign would wait for better opportunities before depleting its budget as compared to the greedy policy. The bottom left table in Figure \ref{thefig} reports the budget utilization (defined as total revenue divided by $\sum_{k \in \mathcal{K}}m_k$) and profit margin (defined as total profit divided by total revenue) statistics for each policy, which confirms our intuition and the results presented in top left table. Example B constitutes our second experiment, whereby we took the same exact problem instance and made one modification, namely instead of using constant budgets across the different campaigns we allowed the budget to be correlated with the quality score of each campaign so that $m_k := 50Q_k$. As Figure \ref{thefig} demonstrates, the profit improvement of the Lagrangian policy over the greedy baseline is even more dramatic in this case.
\begin{figure}[h!]
\vspace{-2mm}
\begin{multicols}{2}
\scalebox{0.68}{
\begin{tabular}{c}
\begin{tabular}{ccc}
\hline
{\bf (Lag./Gr.)} & {\bf Example A} & {\bf Example B} \\ \hline
{\bf Relative Profit} & {\Large 1.257} & {\Large 1.576} \\
{\bf Relative Cost} & {\Large 0.286} & {\Large 0.431} \\
{\bf Relative Revenue} & {\Large 0.759} & {\Large 0.677} \\ \hline
\end{tabular}
\\ \\ \\
\begin{tabular}{ccccc}
& \multicolumn{2}{c}{{\bf Example A}} & \multicolumn{2}{c}{{\bf Example B}} \\ \hline
& {\bf Lag.} & {\bf Gr.} & {\bf Lag.} & {\bf Gr.} \\ \hline
{\bf Budget Util.} & {\Large 0.483} & {\Large 0.636} & {\Large 0.542} & {\Large 0.801} \\
{\bf Profit/Revenue} & {\Large 0.807} & {\Large 0.487} & {\Large 0.500} & {\Large 0.215} \\ \hline
\end{tabular}
\end{tabular}
}
\vfill\null
\columnbreak
\includegraphics[scale = 0.2]{plot.pdf}
\end{multicols}
\vspace{-8mm}
\caption{Lagrangian two-phase policy vs. greedy policy}
\label{thefig}
\end{figure}
In our third experiment, we generated a fresh problem instance using all of the same parameters as before except for two differences: {\em (i)} we set $|\mathcal{I}| = 10$, and {\em (ii)} we varied the budget parameters $m_k \in \{5,10,15,20,25,30,35,40,45,50\}$ (also we took $m_k$ to be constant across the different campaigns) and reran Algorithm \ref{algo-twophase} for each of the 10 budget values. The right side of Figure \ref{thefig} plots the relative profit of the two policies versus these budget values, and for each of the 10 budget values the relative profit statistic was averaged over 500 simulation runs. Clearly, the Lagrangian policy has significantly larger profits over the greedy policy for small budget values, but as the budget increases this improvement is diminished. This makes sense since the Lagrangian policy is based on accounting for the budget constraints in \eqref{poi_deterministic} via dual variables. Indeed, as the budget values become larger, the budget constraints are less active and Proposition \ref{prop:truth} implies that Policy \ref{algo-online} and the greedy policy are exactly the same when $m_k$ is large enough for each $k \in \mathcal{K}$.
Let us conclude this section by mentioning a few directions for future research. First, it would be very valuable to also perform some computational experiments comparing the Lagrangian policy to the greedy policy using a real advertising dataset. Second, it would be very interesting to extend our methodology, in particular problem \eqref{poi_deterministic} and Algorithm \ref{algo-twophase} to different pricing models, such as the CPM pricing model, with performance constraints. Finally, it would be interesting to examine the benefits of more sophisticated stochastic or robust optimization approaches that more carefully account for the uncertainty in the impression arrivals and the real-time bidding environment. The authors plan to pursue all of these directions in future research.
| 052f2ebf43bde1bf3dbae107008b6097d89832d5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{ Nomenclature}
\input{Introduction}
\input{SecondOrderDSMCTheory}
\input{CaseStudies2}
\input{Conclusion}
\small
\section*{ACKNOWLEDGMENT}
This material is based upon the work supported by the National Science Foundation under Grant No. 1434273. Dr. Ken Butts from Toyota Motor Engineering $\&$ Manufacturing North America is gratefully acknowledged for his technical comments during the course of this study. \vspace{-0.3cm}
\bibliographystyle{unsrt}
\section{Case Study: Automotive Engine Control}
\label{sec:CS}\vspace{-0.15cm}
Here, application of the proposed method in Section~\ref{sec:UncertaintyPrediction} is demonstrated for a physics-based spark ignition (SI) combustion engine model~\cite{Shaw} during cold start. Our proposed algorithm fits the requirements of this automotive control problem well, as it contains complicated plant model dynamics prone to uncertainty in slowly-fluctuating environments, yet require uncertainty mitigation to achieve tracking of desired trajectory behavior
The engine model~\cite{Shaw} is parameterized for a 2.4-liter, 4-cylinder, DOHC 16-valve Toyota 2AZ-FE engine. The engine rated power is 117kW $@$ 5600~RPM, and it has a rated torque of 220 Nm $@$ 4000~RPM. The experimental validation of different components of the engine model is available in~\cite{Sanketi}. The nonlinear model has four states including the exhaust gas temperature~($T_{exh}$), fuel mass flow rate into the cylinders (${\dot{m}_f}$), the engine speed~(${\omega_e}$), and the mass of air inside the intake manifold ($m_{a}$). The control problem is defined to steer $T_{exh}$, ${\omega_e}$, and air-fuel ratio ($AFR$) to their pre-defined desired values. A set of four SISO DSMCs is designed to achieve this objective. Four states of the model and corresponding dynamics and controllers will be discussed in the following sections. Details of the functions and constants in the engine model are found in the Appendix and~\cite{Sanketi}.
$\bullet {\textbf{~~Exhaust~Gas~Temperature~Controller:}}$~Discretized model for exhaust gas temperature ($T_{exh}$) is:
\vspace{-0.2cm}
\begin{gather}
\label{eq:Engine_discretized_1}
T_{exh}(k+1)=(1-\frac{T}{\tau_e})T_{exh}(k)\\
+\frac{T}{\tau_e}(7.5\Delta(k)+600)AFI(k) \nonumber
\end{gather}
where $\Delta(k)$ is the control input. The sliding surface for $T_{exh}$ controller is defined to be the error in tracking the desired exhaust gas temperature ($s_{1}=T_{exh}-{T_{exh,d}}$). The dynamics of the exhaust gas temperature ($f_{{T_{exh}}}$) with multiplicative unknown term ($\alpha_{T_{exh}}$) is: \vspace{-0.20cm}
\begin{gather}
\label{eq:Engine_discretized_Texh}
f_{{T_{exh}}}=\alpha_{T_{exh}}\Big(\frac{1}{\tau_e}[600AFI-T_{exh}]\Big)
\end{gather}
The exhaust gas time constant ($\tau_{e}$) has a significant role in the exhaust gas temperature dynamics (Eq.~(\ref{eq:Engine_discretized_Texh})). This means that any error in estimating the time constant ($\tau_{e}$) directly affects the dynamics and causes deviation from the nominal model. Multiplicative uncertainty term ($\alpha_{T_{exh}}$) is assumed to represent any error in estimating $\tau_{e}$. The error in the modeled $T_{exh}$ dynamics is removed by using the following adaptation law with respect to Eq.~(\ref{eq:D2SMC_14}): \vspace{-0.15cm}
\begin{gather}
\label{eq:adaptive_Texh}
\hat{\alpha}_{T_{exh}}(k+1)=\hat{\alpha}_{T_{exh}}(k)+\frac{T(s_1(k))}{\tau_e \rho_{\alpha_1}}(600AFI-T_{exh}(k))
\end{gather}
By incorporating Eq.~(\ref{eq:Engine_discretized_1}) and $\hat{\alpha}_{T_{exh}}$ from Eq.~(\ref{eq:adaptive_Texh}) into Eq.~(\ref{eq:D2SMC_6}), the second order adaptive DSMC for exhaust gas temperature becomes: \vspace{-0.25cm}
\begin{gather}\label{eq:Engine_DSMC_Final_1}
\Delta(k)=\frac{\tau_e}{7.5\,.\,AFI\,.\,T}[-\hat{\alpha}_{T_{exh}}(k)\frac{T}{\tau_e}(600\,.\,AFI\\
-T_{exh}(k))-(\beta_1+1)s_1(k)+T_{exh,d}(k+1)-T_{exh,d}(k)] \nonumber
\end{gather} \vspace{-0.75cm}
$\bullet { \textbf{~~Fuel~Flow~Rate~Controller:}}$~The discretized difference equation for the fuel flow rate is: \vspace{-0.2cm}
\begin{gather}
\label{eq:Engine_discretized_2}
\dot{m}_f(k+1)=\dot{m}_f(k)+\frac{T}{\tau_f}[\dot{m}_{fc}(k)-\dot{m}_f(k)]
\end{gather}
The fuel flow dynamic ($f_{\dot{m}_f}$) with multiplicative uncertainty term ($\alpha_{\dot{m}_f}$) is as follows: \vspace{-0.25cm}
\begin{gather}
\label{eq:Engine_discretized_mdotf}
f_{\dot{m}_f}=-\alpha_{\dot{m}_f}\Big(\frac{1}{\tau_f}\dot{m}_f(k)\Big)
\end{gather}
The sliding variable for the fuel flow controller is defined to be the error in tracking the desired fuel mass flow ($s_{2}={\dot{m}_{f}}-{\dot{m}_{f,d}}$). In a similar manner to $T_{exh}$ dynamics, the fuel evaporation time constant $\tau_{f}$ dictates the dynamics of the fuel flow into the cylinder. Consequently, any error in estimating $\tau_{f}$ leads to a considerable deviation from the nominal model. $\alpha_{\dot{m}_f}$ is introduced to the fuel flow dynamics to represent the uncertainty in estimating $\tau_{f}$. The adaptation law for $\alpha_{\dot{m}_f}$ is:
\vspace{-0.4cm}
\begin{gather}
\label{eq:adaptive_mdotf}
\hat{\alpha}_{\dot{m}_f}(k+1)=\hat{\alpha}_{\dot{m}_f}(k)-\frac{T(s_2(k))}{\tau_f \rho_{\alpha_2}}\dot{m}_f(k)
\end{gather}
where, $\dot{m}_{f,d}$ in $s_2$ is calculated according to desired AFR. The adaptive control law for $\dot{m}_{fc}$ is: \vspace{-0.15cm}
\begin{gather}\label{eq:Engine_DSMC_Final_2}
\dot{m}_{fc}(k)=\frac{\tau_f}{T}[\hat{\alpha}_{\dot{m}_f}(k)\frac{T}{\tau_f}\dot{m}_f(k) \\-(\beta_2+1)s_2(k)+\dot{m}_{f,d}(k+1)-\dot{m}_{f,d}(k)] \nonumber
\end{gather} \vspace{-0.5cm}
$\bullet { \textbf{~~Engine~Speed~Controller:}}$~The rotational dynamics of the engine is described by using the following equation:\vspace{-0.25cm}
\begin{gather}
\label{eq:Engine_discretized_3}
\omega_e(k+1)=\omega_e(k)+\frac{T}{J}T_E(k)
\end{gather}
where $T_E(k)$ is the engine torque and is found by $30000~m_a(k)-(0.4~\omega_e(k)+100)$. There is no direct control input for modulating the engine speed, therefore $m_a$ is considered as the synthetic control input. The calculated $m_a$ from engine speed controller will be used as the desired trajectory in intake air mass flow rate controller. $f_{\omega_e}$ of the engine with multiplicative uncertainty ($\alpha_{\omega_e}$) is as follows: \vspace{-0.25cm}
\begin{gather}
\label{eq:Engine_discretized_we}
f_{\omega_e}=-\alpha_{\omega_e}\Big(\frac{1}{J}T_{loss}\Big)
\end{gather}
where $T_{loss}=0.4\omega_e+100$. $T_{loss}$ represents the torque losses on the crankshaft. Thus, the multiplicative uncertainty $\alpha_{\omega_e}$ compensates for any error in estimated torque loss. The sliding variable for the engine speed controller is defined to be $s_3=\omega_e-{\omega_{e,d}}$. $\alpha_{\omega_e}$ is driven to ``1'' using the following adaptation law: \vspace{-0.3cm}
\begin{gather}
\label{eq:adaptive_we}
\hat{\alpha}_{\omega_e}(k+1)=\hat{\alpha}_{\omega_e}(k)-\frac{T(s_3(k))}{J.\rho_{\alpha_3}}(0.4\omega_e(k)+100)
\end{gather}
Finally, the desired synthetic control input ($m_{a,d}$) is:\vspace{-0.3cm}
\begin{gather}\label{eq:Engine_DSMC_Final_3}
m_{a,d}(k)=\frac{J}{30,000\,T}[\hat{\alpha}_{\omega_e}(k)\frac{T}{J}(100+0.4\omega_e(k))\\\nonumber-(\beta_3+1)s_3(k)
+\omega_{e,d}(k+1)-\omega_{e,d}(k)]
\end{gather}
$\bullet { \textbf{~~Air~Mass~Flow~Controller:}}$
The following state difference equation describes the air mass flow behaviour: \vspace{-0.2cm}
\begin{gather}
\label{eq:Engine_discretized_3}
m_a(k+1)=m_a(k)+T[\dot{m}_{ai}(k)-\dot{m}_{ao}(k)]
\end{gather}
The calculated $m_{a,d}$ from Eq.~(\ref{eq:Engine_DSMC_Final_3}) is used as the desired trajectory to obtain $\dot{m}_{ai}$ as the control input of $m_a$ controller. The last sliding surface for the air mass flow controller is defined to be $s_4=m_a-{m_{a,d}}$. The intake air manifold mass dynamic with the unknown term ($\alpha_{m_a}$) is: \vspace{-0.25cm}
\begin{gather}
\label{eq:Engine_discretized_ma}
f_{m_a}=-\alpha_{m_a}\dot{m}_{ao}(k)
\end{gather} \vspace{-0.35cm}
where, air mass flow into the cylinder is determined by~\cite{Shaw}: \vspace{-0.2cm}
\begin{gather}
\label{eq:Engine_discretized_ma2}
\dot{m}_{ao}=k_1\eta_{vol}m_a\omega_e
\end{gather}
$\eta_{vol}$ is the volumetric efficiency. As can be seen from Eq.~(\ref{eq:Engine_discretized_ma}) and~(\ref{eq:Engine_discretized_ma2}), the multiplicative uncertainty term in the intake air manifold dynamics ($\alpha_{m_a}$) represents the uncertainty in $\dot{m}_{ao}$ that is extracted from $\eta_{vol}$ map. $\alpha_{m_a}$ is updated using the following adaptation law: \vspace{-0.2cm}
\begin{gather}
\label{eq:adaptive_ma}
\hat{\alpha}_{m_a}(k+1)=\hat{\alpha}_{m_a}(k)-\frac{T(s_4(k))}{\rho_{\beta_4}}\dot{m}_{ao}
\end{gather}
Finally, the controller input is: \vspace{-0.2cm}
\begin{gather}\label{eq:Engine_DSMC_Final_4}
\dot{m}_{ai}(k)=\frac{1}{T}[\hat{\alpha}_{m_a}(k)\dot{m}_{ao}(k)T-(\beta_4+1)s_4(k)\\
+m_{a,d}(k+1)-m_{a,d}(k)] \nonumber
\end{gather}
In the absence of model uncertainties ($\alpha_{T_{exh}}=\alpha_{\dot{m}_f}=\alpha_{\omega_e}=\alpha_{{m}_a}=1$), Figures~\ref{fig:Engine_2DSMC_SamplingComparison_10ms} and~\ref{fig:Engine_2DSMC_SamplingComparison_40ms} show the results of tracking the desired $AFR$, $T_{exh}$, and engine speed trajectories, using the first and second order DSMCs for sampling times of $10~ms$ and $40~ms$, respectively. The mean tracking errors for both controllers are listed in Table~\ref{table:tracking_Results}. It can be observed from Fig.~~\ref{fig:Engine_2DSMC_SamplingComparison_10ms} and Table~\ref{table:tracking_Results} when the signals at the controller I/O are sampled every $10~ms$, both first and second order DSMCs illustrate acceptable tracking performances, while the second order controller is 50\% more robust on average in terms of the tracking errors. As long as the Shannon's sampling theorem criteria, which states that the sampling frequency must be at least twice the maximum frequency of the measured analog signal, is satisfied, increasing the sampling time helps to reduce the computation cost. Upon increasing the sampling rate from $10~ms$ to $40~ms$, the first order DSMC performance degrades significantly. On the other side, despite the increase in the sampling time, the second order DSMC still presents smooth and accurate tracking results. By comparing the first and second order DSMC results at $T=40~ms$, it can be concluded that the proposed second order DSMC outperforms the first order controller by up to 85\% in terms of the mean tracking errors. \vspace{-0.5cm}
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,width= \columnwidth]{Engine_2DSMC_SamplingComparison_10ms.eps} \vspace{-0.75cm}
\caption{\label{fig:Engine_2DSMC_SamplingComparison_10ms}Engine tracking results by the first and second order DSMCs with $T=10~ms$.} \vspace{-0.8cm}
\end{center}
\end{figure}
\begin{table} [htbp!]
\small
\begin{center}
\caption{Mean ($\bar{e}$) of Tracking Errors. Values Inside the Parentheses Show the Resulting Improvement from the Second Order DSMC Compared to the First Order DSMC.} \linespread{1.15}
\label{table:tracking_Results}\vspace{-0.15cm}
\begin{tabular}{lccccc}
\hline\hline
\multicolumn{1}{c}{} & \multicolumn{2}{c}{$\bar{e}~(T=10~ms)$} & & \multicolumn{2}{c}{$\bar{e}~(T=40~ms)$} \\
\cline{2-3} \cline{5-6}
\textbf{}& {$1^{st}$-Order} & {$2^{nd}$-Order} & & {$1^{st}$-Order} & {$2^{nd}$-Order} \\
\textbf{}& {DSMC} & {DSMC} & & {DSMC} & {DSMC} \\
\textbf{} & {\textcolor{blue}{Reference}}& \textbf{} & & {\textcolor{blue}{Reference}} & \textbf{} \\ \hline
AFR& 0.028 & 0.010 & & 0.126 & 0.019 \\ \vspace{0.05cm}
[-] & & \textcolor{blue}{(-64\%)} & & & \textcolor{blue}{(-84.9\%)} \\ \hline \vspace{0.05cm}
$T_{exh}$ & 0.2 & 0.1 & & 1.8 & 0.2 \\ \vspace{0.05cm}
[$^o$C] & & \textcolor{blue}{(-50\%)} & & & \textcolor{blue}{(-88.9\%)}\\ \hline \vspace{0.05cm}
$N$ & 0.1 & 0.06 & & 1.9 & 0.3\\ \vspace{0.05cm}
[RPM] & & \textcolor{blue}{(-40\%)} & & & \textcolor{blue}{(-84.2\%)} \\
\hline\hline
\end{tabular} \vspace{-0.5cm}
\end{center}
\end{table}
\linespread{1}
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,width= \columnwidth]{Engine_2DSMC_SamplingComparison_40ms.eps} \vspace{-0.75cm}
\caption{\label{fig:Engine_2DSMC_SamplingComparison_40ms}Engine tracking results by the first and second order DSMCs with $T=40~ms$.} \vspace{-0.8cm}
\end{center}
\end{figure}
The effect of the unknown multiplicative terms (up to 25\%) on the engine plant's dynamics ($f$) is shown in Fig.~\ref{fig:Engine_2DSMC_DynamicImpact_40ms}. The uncertainty terms in the model introduce a permanent error in the estimated dynamics compared to the nominal model. If these errors are not removed in the early seconds of the controller operation, the tracking performance will be affected adversely. Upon activation of the adaptation mechanism, as it can be observed from Fig.~\ref{fig:Engine_2DSMC_DynamicImpact_40ms}, the model with error is steered towards the nominal model in less than 2 $sec$. Consequently, the errors in the model are removed. Fig.~\ref{fig:Engine_2DSMC_ParamConv_40ms} shows the results of unknown multiplicative uncertainty term ($\hat{\alpha}$) estimation against the actual (nominal) values ($\alpha$). \vspace{-0.35cm}
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,width= \columnwidth]{Engine_2DSMC_DynamicImpact_40ms.eps} \vspace{-0.75cm}
\caption{\label{fig:Engine_2DSMC_DynamicImpact_40ms}The effect of the model uncertainty terms on the engine dynamics when using the second order DSMC and how the adaptation mechanism drives the model with error to its nominal value: (a) $T_{exh}$, (b) $\dot{m}_f$, (c) $\omega_e$, and (d) $m_a$ ($T=40~ms$).} \vspace{-0.75cm}
\end{center}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,width= \columnwidth]{Engine_2DSMC_ParamConv_40ms.eps} \vspace{-0.75cm}
\caption{\label{fig:Engine_2DSMC_ParamConv_40ms}Estimation of unknown multiplicative parameters in adaptive DSMC ($T=40~ms$).} \vspace{-0.7cm}
\end{center}
\end{figure}
Fig.~\ref{fig:Engine_2DSMC_AdaptiveTracking_40ms} shows the comparison between the tracking performances of the non-adaptive and adaptive second order DSMCs. As expected, the non-adaptive DMSC fails to track the desired trajectories, which explains the importance of handling the model uncertainties in the body of the DSMC. On the other hand, once the adaptation algorithm is enabled and the convergence period of the unknown parameters is over, the adaptive DSMC tracks all the desired trajectories smoothly with the minimum error under $T=40~ms$
\vspace{-0.4cm}
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,width= \columnwidth]{Engine_2DSMC_AdaptiveTracking_40ms.eps} \vspace{-0.75cm}
\caption{\label{fig:Engine_2DSMC_AdaptiveTracking_40ms}Comparison between adaptive and non-adaptive second order DSMCs for the engine model with uncertainties: (a) $AFR$, (b) $T_{exh}$, and (c) $N$~($T=40~ms$).} \vspace{-0.9cm}
\end{center}
\end{figure}
\section{Case Studies} \label{sec:CS}\vspace{-0.15cm}
Here, application of the proposed method in Section~\ref{sec:UncertaintyPrediction} is demonstrated for one linear (DC motor) and one nonlinear (SI engine) case studies. Our proposed algorithm fits the requirements of these automotive control problems well, as they contain complicated plant model dynamics prone to uncertainty in slowly-fluctuating environments, yet require uncertainty mitigation to achieve tracking of desired trajectory behavior.\vspace{-0.15cm}
\subsection{Linear Case Study: DC Motor Speed Control} \label{CS_DCMotor} \vspace{-0.10cm}
DC motors are common automotive actuators for control applications that require rotary and transitional motions. For speed regulation of a DC motor, the control input is voltage ($V$) to the motor's armature and the output is rotation speed (${\theta}$) of the shaft. Assuming a constant magnetic field and linear relationship between motor torque and armature current ($\mathcal{I}$), by choosing the rotor speed and current as the state variables, the following discretized linear time-invariant state-space representation can be used to describe the dynamics of the DC motor~\cite{DCMotor_ref}, in which $\alpha_{pq}$ represents the unknown multiplicative term on the plant's parameters: \vspace{-0.25cm}
\begin{subequations}
\begin{gather}\label{eq:DC_motor_linear_DSMC_Uncertain}
\theta(k+1)=T\left([-{\alpha}_{11}\frac{k_f}{J}]\theta(k)\right) \\
+T\left(\frac{k_m}{J}{\mathcal{I}}(k)+\frac{1}{J} \Gamma\right) +\theta(k) \nonumber \\
{\mathcal{I}}(k+1)=T\left([-\alpha_{21}\frac{k_b}{L}]\theta(k)+[-\alpha_{22}\frac{R}{L}]{\mathcal{I}}(k)\right)\\
+\frac{T}{L}V(k)+{\mathcal{I}}(k) \nonumber
\end{gather}
\end{subequations}
where $J$ is the rotor's moment of inertia, $R$ is the electrical resistance, $L$ is the electric inductance, $\Gamma$ is the generated torque on the rotor, $k_f$ is the mechanical damping, $k_m$ is the motor torque constant, and $k_b$ is the electromotive force constant. The DC motor model constants are listed in
the Appendix. Three multiplicative ($\alpha_{pq}$) unknown parameters are derived to their nominal values by solving the adaptation law in Eq.~(\ref{eq:D2SMC_14}). A second order DSMC is designed to regulate the DC motor rotational speed with respect to the desired speed profile ($\theta_d$). The first sliding surface is defined as the error in tracking the desired speed profile ($s_{1}=\theta-\theta_d$). Since there is no direct control input on DC motor rotational speed, $\mathcal{I}_d$ is defined as the synthetic control input for controlling the shaft speed. $\mathcal{I}_d$ is used to define the second sliding surface ($s_{2}=\mathcal{I}-\mathcal{I}_d$) in which the control input is voltage. The final synthetic ($\mathcal{I}_d$) and physical ($V$) control inputs are calculated as follows: \vspace{-0.25cm}
\begin{subequations} \label{eq:DC_motor_adaptiveDSMC_modified}
\begin{gather}
{\mathcal{I}}_d(k)=\frac{J}{k_m}\left(\frac{1}{T}\left[-\beta_1(\theta(k)-\theta_d(k))+\theta_d(k+1)\right.\right.\\
\left.\left.-\theta(k)\right]+[\hat{\alpha}_{11}\frac{k_f}{J}\theta(k)]-\frac{1}{J}\Gamma\right) \nonumber \\
V(k)=L\left(\frac{1}{T_s}\left[-\beta_2({\mathcal{I}}(k)-{\mathcal{I}}_d(k))+{\mathcal{I}}_d(k+1)
\right.\right.\\
\left.\left.-{\mathcal{I}}(k)\right]
+[\hat{\alpha}_{21}\frac{k_b}{L}]\theta(k)+\hat{\alpha}_{22}\frac{R}{L}{\mathcal{I}}(k)\right) \nonumber
\end{gather}
\end{subequations}
Simulations are done in MATLAB Simulink$^{\textregistered}$~which allows for testing the controller in a model-in-the-loop (MIL) platform for different sampling rates. In the absence of model uncertainties ($\alpha_{pq}=1$), Fig.~\ref{fig:DC_2DSMC_SamplingComparison} shows the comparison between the first~\cite{Pan_Discrete} and second order DSMCs for different sampling rates. Although at lower sampling rates (e.g., $200~ms$) both controllers show similar performances, by increasing the sampling rate to $800~ms$, the higher robustness characteristics of the second order DSMC in comparison with the first order controller are revealed. The comparison results show that the second order DSMC is improving the tracking errors by 69\% on average for different sampling rates, compared to the first order DSMC.\vspace{-0.35cm}
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,width= \columnwidth]{DC_2DSMC_SamplingComparison.eps} \vspace{-0.85cm}
\caption{\label{fig:DC_2DSMC_SamplingComparison}Comparison between the first and second order DSMCs in tracking the desired speed trajectory of the DC motor for different sampling rates. (a) First order DSMC, (b) Second order DSMC. Values inside the parentheses show the resulting improvement ($\downarrow$) from the second order DSMC compared to the first order DSMC.} \vspace{-0.55cm}
\end{center}
\end{figure}
The convergence results of the unknown parameters are shown in Fig.~\ref{fig:DC_2DSMC_ParamConv}. This is done by simultaneously solving Eq.~(\ref{eq:D2SMC_14}) and Eq.~(\ref{eq:DC_motor_adaptiveDSMC_modified}). The performance of the first~\cite{Amini_DSCC2016}, and second order adaptive DSMCs are shown in Fig.~\ref{fig:DC_2DSMC_SpeedTracking} for tracking the desired speed profile under 200 $ms$ of sampling time, and up 50\% uncertainty on the plant's parameters. As can be observed, the second order DSMC is able to significantly improve the tracking performance by 60\% compared to the first order adaptive DSMC.\vspace{-0.45cm}
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,width= \columnwidth]{DC_2DSMC_ParamConv.eps} \vspace{-0.85cm}
\caption{\label{fig:DC_2DSMC_ParamConv}Convergence results of the unknown multiplicative ($\alpha$) terms in the DC motor model (sampling time=200 ms).} \vspace{-0.75cm}
\end{center}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,width= \columnwidth]{DC_2DSMC_SpeedTracking.eps} \vspace{-0.85cm}
\caption{\label{fig:DC_2DSMC_SpeedTracking}DC motor speed control using the first and second order \textit{adaptive} DSMCs (sampling time=200 ms).} \vspace{-1cm}
\end{center}
\end{figure}
\subsection{Nonlinear Case Study: Automotive Engine Control}\label{subsec:Engine_Control}\vspace{-0.10cm}
In this section, the performance of the proposed adaptive second order DSMC is studied for a physics-based spark ignition (SI) combustion engine model~\cite{Shaw} during cold start. The engine model~\cite{Shaw} is parameterized for a 2.4-liter, 4-cylinder, DOHC 16-valve Toyota 2AZ-FE engine. The engine rated power is 117kW $@$ 5600~RPM, and it has a rated torque of 220 Nm $@$ 4000~RPM. The experimental validation of different components of the engine model is available in~\cite{Sanketi}. The nonlinear model has four states including the exhaust gas temperature~($T_{exh}$), fuel mass flow rate into the cylinders (${\dot{m}_f}$), the engine speed~(${\omega_e}$), and the mass of air inside the intake manifold ($m_{a}$). The control problem is defined to steer $T_{exh}$, ${\omega_e}$, and air-fuel ratio ($AFR$) to their pre-defined desired values. A set of four SISO DSMCs is designed to achieve this objective. Four states of the model and corresponding dynamics and controllers will be discussed in the following sections. Details of the functions and constants in the engine model are found in the Appendix and~\cite{Sanketi}.
$\bullet {\textbf{~~Exhaust~Gas~Temperature~Controller:}}$~Discretized model for exhaust gas temperature ($T_{exh}$) is:
\vspace{-0.2cm}
\begin{gather}
\label{eq:Engine_discretized_1}
T_{exh}(k+1)=(1-\frac{T}{\tau_e})T_{exh}(k)\\
+\frac{T}{\tau_e}(7.5\Delta(k)+600)AFI(k) \nonumber
\end{gather}
where $\Delta(k)$ is the control input. The sliding surface for $T_{exh}$ controller is defined to be the error in tracking the desired exhaust gas temperature ($s_{1}=T_{exh}-{T_{exh,d}}$). The dynamics of the exhaust gas temperature ($f_{{T_{exh}}}$) with multiplicative unknown term ($\alpha_{T_{exh}}$) is: \vspace{-0.20cm}
\begin{gather}
\label{eq:Engine_discretized_Texh}
f_{{T_{exh}}}=\alpha_{T_{exh}}\Big(\frac{1}{\tau_e}[600AFI-T_{exh}]\Big)
\end{gather}
The exhaust gas time constant ($\tau_{e}$) has a significant role in the exhaust gas temperature dynamics (Eq.~(\ref{eq:Engine_discretized_Texh})). This means that any error in estimating the time constant ($\tau_{e}$) directly affects the dynamics and causes deviation from the nominal model. Multiplicative uncertainty term ($\alpha_{T_{exh}}$) is assumed to represent any error in estimating $\tau_{e}$. The error in the modeled $T_{exh}$ dynamics is removed by using the following adaptation law with respect to Eq.~(\ref{eq:D2SMC_14}): \vspace{-0.15cm}
\begin{gather}
\label{eq:adaptive_Texh}
\hat{\alpha}_{T_{exh}}(k+1)=\hat{\alpha}_{T_{exh}}(k)+\frac{T(s_1(k))}{\tau_e \rho_{\alpha_1}}(600AFI-T_{exh}(k))
\end{gather}
By incorporating Eq.~(\ref{eq:Engine_discretized_1}) and $\hat{\alpha}_{T_{exh}}$ from Eq.~(\ref{eq:adaptive_Texh}) into Eq.~(\ref{eq:D2SMC_6}), the second order adaptive DSMC for exhaust gas temperature becomes: \vspace{-0.25cm}
\begin{gather}\label{eq:Engine_DSMC_Final_1}
\Delta(k)=\frac{\tau_e}{7.5\,.\,AFI\,.\,T}[-\hat{\alpha}_{T_{exh}}(k)\frac{T}{\tau_e}(600\,.\,AFI\\
-T_{exh}(k))-(\beta_1+1)s_1(k)+T_{exh,d}(k+1)-T_{exh,d}(k)] \nonumber
\end{gather} \vspace{-0.75cm}
$\bullet { \textbf{~~Fuel~Flow~Rate~Controller:}}$~The discretized difference equation for the fuel flow rate is: \vspace{-0.2cm}
\begin{gather}
\label{eq:Engine_discretized_2}
\dot{m}_f(k+1)=\dot{m}_f(k)+\frac{T}{\tau_f}[\dot{m}_{fc}(k)-\dot{m}_f(k)]
\end{gather}
The fuel flow dynamic ($f_{\dot{m}_f}$) with multiplicative uncertainty term ($\alpha_{\dot{m}_f}$) is as follows: \vspace{-0.25cm}
\begin{gather}
\label{eq:Engine_discretized_mdotf}
f_{\dot{m}_f}=-\alpha_{\dot{m}_f}\Big(\frac{1}{\tau_f}\dot{m}_f(k)\Big)
\end{gather}
The sliding variable for the fuel flow controller is defined to be the error in tracking the desired fuel mass flow ($s_{2}={\dot{m}_{f}}-{\dot{m}_{f,d}}$). In a similar manner to $T_{exh}$ dynamics, the fuel evaporation time constant $\tau_{f}$ dictates the dynamics of the fuel flow into the cylinder. Consequently, any error in estimating $\tau_{f}$ leads to a considerable deviation from the nominal model. $\alpha_{\dot{m}_f}$ is introduced to the fuel flow dynamics to represent the uncertainty in estimating $\tau_{f}$. The adaptation law for $\alpha_{\dot{m}_f}$ becomes:
\vspace{-0.4cm}
\begin{gather}
\label{eq:adaptive_mdotf}
\hat{\alpha}_{\dot{m}_f}(k+1)=\hat{\alpha}_{\dot{m}_f}(k)-\frac{T(s_2(k))}{\tau_f \rho_{\alpha_2}}\dot{m}_f(k)
\end{gather}
where, $\dot{m}_{f,d}$ is calculated according to desired AFR. The adaptive control law for $\dot{m}_{fc}$ is: \vspace{-0.15cm}
\begin{gather}\label{eq:Engine_DSMC_Final_2}
\dot{m}_{fc}(k)=\frac{\tau_f}{T}[\hat{\alpha}_{\dot{m}_f}(k)\frac{T}{\tau_f}\dot{m}_f(k) \\-(\beta_2+1)s_2(k)+\dot{m}_{f,d}(k+1)-\dot{m}_{f,d}(k)] \nonumber
\end{gather} \vspace{-0.75cm}
$\bullet { \textbf{~~Engine~Speed~Controller:}}$~The rotational dynamics of the engine is described by using the following difference equation:\vspace{-0.4cm}
\begin{gather}
\label{eq:Engine_discretized_3}
\omega_e(k+1)=\omega_e(k)+\frac{T}{J}T_E(k)
\end{gather}
where $T_E(k)$ is the engine torque and is found by $30000~m_a(k)-(0.4~\omega_e(k)+100)$. There is no direct control input for modulating the engine speed, therefore $m_a$ is considered as the synthetic control input. The calculated $m_a$ from engine speed controller will be used as the desired trajectory in intake air mass flow rate controller. $f_{\omega_e}$ of the engine with multiplicative uncertainty ($\alpha_{\omega_e}$) is as follows: \vspace{-0.25cm}
\begin{gather}
\label{eq:Engine_discretized_we}
f_{\omega_e}=-\alpha_{\omega_e}\Big(\frac{1}{J}T_{loss}\Big)
\end{gather}
where $T_{loss}=0.4\omega_e+100$. $T_{loss}$ represents the torque losses on the crankshaft. Thus, the multiplicative uncertainty $\alpha_{\omega_e}$ compensates for any error in estimated torque loss. The sliding variable for the engine speed controller is defined to be $s_3=\omega_e-{\omega_{e,d}}$. $\alpha_{\omega_e}$ is driven to ``1'' using the following adaptation law: \vspace{-0.3cm}
\begin{gather}
\label{eq:adaptive_we}
\hat{\alpha}_{\omega_e}(k+1)=\hat{\alpha}_{\omega_e}(k)-\frac{T(s_3(k))}{J.\rho_{\alpha_3}}(0.4\omega_e(k)+100)
\end{gather}
Finally, the desired synthetic control input ($m_{a,d}$) is:\vspace{-0.3cm}
\begin{gather}\label{eq:Engine_DSMC_Final_3}
m_{a,d}(k)=\frac{J}{30,000\,T}[\hat{\alpha}_{\omega_e}(k)\frac{T}{J}(100+0.4\omega_e(k))\\\nonumber-(\beta_3+1)s_3(k)
+\omega_{e,d}(k+1)-\omega_{e,d}(k)]
\end{gather}
$\bullet { \textbf{~~Air~Mass~Flow~Controller:}}$
The following state difference equation describes the air mass flow behaviour in discrete time: \vspace{-0.2cm}
\begin{gather}
\label{eq:Engine_discretized_3}
m_a(k+1)=m_a(k)+T[\dot{m}_{ai}(k)-\dot{m}_{ao}(k)]
\end{gather}
The calculated $m_{a,d}$ from Eq.~(\ref{eq:Engine_DSMC_Final_3}) is used as the desired trajectory to obtain $\dot{m}_{ai}$ as the control input of $m_a$ controller. The last sliding surface for the air mass flow controller is defined to be $s_4=m_a-{m_{a,d}}$. The intake air manifold mass dynamic with the unknown term ($\alpha_{m_a}$) is: \vspace{-0.25cm}
\begin{gather}
\label{eq:Engine_discretized_ma}
f_{m_a}=-\alpha_{m_a}\dot{m}_{ao}(k)
\end{gather} \vspace{-0.35cm}
where, air mass flow into the cylinder is determined by~\cite{Shaw}: \vspace{-0.2cm}
\begin{gather}
\label{eq:Engine_discretized_ma2}
\dot{m}_{ao}=k_1\eta_{vol}m_a\omega_e
\end{gather}
$\eta_{vol}$ is the volumetric efficiency. As can be seen from Eq.~(\ref{eq:Engine_discretized_ma}) and~(\ref{eq:Engine_discretized_ma2}), the multiplicative uncertainty term in the intake air manifold dynamics ($\alpha_{m_a}$) represents the uncertainty in $\dot{m}_{ao}$ that is extracted from $\eta_{vol}$ map. $\alpha_{m_a}$ is updated using the following adaptation law: \vspace{-0.2cm}
\begin{gather}
\label{eq:adaptive_ma}
\hat{\alpha}_{m_a}(k+1)=\hat{\alpha}_{m_a}(k)-\frac{T(s_4(k))}{\rho_{\beta_4}}\dot{m}_{ao}
\end{gather}
Finally, the controller input is: \vspace{-0.2cm}
\begin{gather}\label{eq:Engine_DSMC_Final_4}
\dot{m}_{ai}(k)=\frac{1}{T}[\hat{\alpha}_{m_a}(k)\dot{m}_{ao}(k)T-(\beta_4+1)s_4(k)\\
+m_{a,d}(k+1)-m_{a,d}(k)] \nonumber
\end{gather}
In the absence of model uncertainties ($\alpha_{T_{exh}}=\alpha_{\dot{m}_f}=\alpha_{\omega_e}=\alpha_{{m}_a}=1$), Fig.~\ref{fig:Engine_2DSMC_SamplingComparison_10ms} and~\ref{fig:Engine_2DSMC_SamplingComparison_40ms} show the results of tracking the desired $AFR$, $T_{exh}$, and engine speed trajectories, using the first and second order DSMCs for sampling times of $10~ms$ and $40~ms$, respectively. The mean tracking errors for both controllers are listed in Table~\ref{table:tracking_Results}. It can be observed from Fig.~~\ref{fig:Engine_2DSMC_SamplingComparison_10ms} and Table~\ref{table:tracking_Results} when the signals at the controller I/O are sampled every $10~ms$, both first and second order DSMCs illustrate acceptable tracking performances, while the second order controller is 50\% more robust on average in terms of the tracking errors. Upon increasing the sampling rate from $10~ms$ to $40~ms$, the first order DSMC performance degrades significantly. On the other side, despite the increase in the sampling time, the second order DSMC still presents smooth and accurate tracking results. By comparing the first and second order DSMC results at $T=40~ms$, it can be concluded that the proposed second order DSMC outperforms the first order controller by up to 85\% in terms of the mean tracking errors.
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,width= \columnwidth]{Engine_2DSMC_SamplingComparison_10ms.eps} \vspace{-0.75cm}
\caption{\label{fig:Engine_2DSMC_SamplingComparison_10ms}Engine tracking results by the first and second order DSMCs with $T=10~ms$.} \vspace{-0.55cm}
\end{center}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,width= \columnwidth]{Engine_2DSMC_SamplingComparison_40ms.eps} \vspace{-0.75cm}
\caption{\label{fig:Engine_2DSMC_SamplingComparison_40ms}Engine tracking results by the first and second order DSMCs with $T=40~ms$.} \vspace{-0.40cm}
\end{center}
\end{figure}
\begin{table} [htbp!]
\small
\begin{center}
\caption{Mean ($\bar{e}$) of Tracking Errors. Values Inside the Parentheses Show the Resulting Improvement from the Second Order DSMC Compared to the First Order DSMC.} \linespread{1.15}
\label{table:tracking_Results}\vspace{-0.15cm}
\begin{tabular}{lccccc}
\hline\hline
\multicolumn{1}{c}{} & \multicolumn{2}{c}{$\bar{e}~(T=10~ms)$} & & \multicolumn{2}{c}{$\bar{e}~(T=40~ms)$} \\
\cline{2-3} \cline{5-6}
\textbf{}& {$1^{st}$-Order} & {$2^{nd}$-Order} & & {$1^{st}$-Order} & {$2^{nd}$-Order} \\
\textbf{}& {DSMC} & {DSMC} & & {DSMC} & {DSMC} \\
\textbf{} & {\textcolor{blue}{Reference}}& \textbf{} & & {\textcolor{blue}{Reference}} & \textbf{} \\ \hline
AFR& 0.028 & 0.010 & & 0.126 & 0.019 \\ \vspace{0.05cm}
[-] & & \textcolor{blue}{(-64\%)} & & & \textcolor{blue}{(-84.9\%)} \\ \hline \vspace{0.05cm}
$T_{exh}$ & 0.2 & 0.1 & & 1.8 & 0.2 \\ \vspace{0.05cm}
[$^o$C] & & \textcolor{blue}{(-50\%)} & & & \textcolor{blue}{(-88.9\%)}\\ \hline \vspace{0.05cm}
$N$ & 0.1 & 0.06 & & 1.9 & 0.3\\ \vspace{0.05cm}
[RPM] & & \textcolor{blue}{(-40\%)} & & & \textcolor{blue}{(-84.2\%)} \\
\hline\hline
\end{tabular} \vspace{-0.5cm}
\end{center}
\end{table}
\linespread{1}
The effect of the unknown multiplicative terms (up to 25\%) on the engine plant's dynamics ($f$) is shown in Fig.~\ref{fig:Engine_2DSMC_DynamicImpact_40ms}. The uncertainty terms in the model introduce a permanent error in the estimated dynamics compared to the nominal model. If these errors are not removed in the early seconds of the controller operation, the tracking performance will be affected adversely. Upon activation of the adaptation mechanism, as it can be observed from Fig.~\ref{fig:Engine_2DSMC_DynamicImpact_40ms}, the model with error is steered towards the nominal model in less than 2 $sec$. Consequently, the errors in the model are removed. Fig.~\ref{fig:Engine_2DSMC_ParamConv_40ms} shows the results of unknown multiplicative uncertainty term ($\hat{\alpha}$) estimation against the actual (nominal) values ($\alpha$).
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,width= \columnwidth]{Engine_2DSMC_DynamicImpact_40ms.eps} \vspace{-0.75cm}
\caption{\label{fig:Engine_2DSMC_DynamicImpact_40ms}The effect of the model uncertainty terms on the engine dynamics when using the second order DSMC and how the adaptation mechanism drives the model with error to its nominal value: (a) $T_{exh}$, (b) $\dot{m}_f$, (c) $\omega_e$, and (d) $m_a$ ($T=40~ms$).} \vspace{-0.55cm}
\end{center}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,width= \columnwidth]{Engine_2DSMC_ParamConv_40ms.eps} \vspace{-0.75cm}
\caption{\label{fig:Engine_2DSMC_ParamConv_40ms}Estimation of unknown multiplicative parameters in adaptive DSMC ($T=40~ms$).} \vspace{-0.55cm}
\end{center}
\end{figure}
Fig.~\ref{fig:Engine_2DSMC_AdaptiveTracking_40ms} shows the comparison between the tracking performances of the non-adaptive and adaptive second order DSMCs. As expected, the non-adaptive DMSC fails to track the desired trajectories, which explains the importance of handling the model uncertainties in the body of the DSMC. On the other hand, once the adaptation algorithm is enabled and the convergence period of the unknown parameters is over, the adaptive DSMC tracks all the desired trajectories smoothly with the minimum error under $40~ms$ sampling time
\vspace{-0.4cm}
\begin{figure}[h!]
\begin{center}
\includegraphics[angle=0,width= \columnwidth]{Engine_2DSMC_AdaptiveTracking_40ms.eps} \vspace{-0.75cm}
\caption{\label{fig:Engine_2DSMC_AdaptiveTracking_40ms}Comparison between adaptive and non-adaptive second order DSMCs for the engine model with uncertainties: (a) $AFR$, (b) $T_{exh}$, and (c) $N$~($T=40~ms$).} \vspace{-0.75cm}
\end{center}
\end{figure}
\section{Summary and Conclusions} \label{sec:Conclusion} \vspace{-0.1cm}
A new adaptive second order discrete sliding mode controller (DSMC) formulation for nonlinear uncertain systems was introduced in this paper. Based on the discrete Lyapunov stability theorem, an adaptation law was determined for removing generic unknown multiplicative uncertainty terms within the nonlinear difference equation of the plant's model. The proposed controller was examined for a spark ignition combustion engine control problem to track desired air-fuel ratio, engine speed, and exhaust gas temperature trajectories. Comparing to the first order DSMC, the second order DSMC shows significantly better robustness against data sampling imprecisions, and can provide up to 80\% improvement in terms of the tracking errors. The better performance of the second order DSMC can be traced in driving the higher order derivatives (difference functions) of the sliding variable to zero. In the presence of the model uncertainties, it was shown that the adaptation mechanism is able to remove the errors in the modeled dynamics quickly, and steer the dynamics towards their nominal values. Increasing the sampling time raises the required time for the adaptation law to compensate for the uncertainties in the models. This required time was increased by two times, when the sampling time was increased from $10~ms$ to $40~ms$ in the engine tracking control problem, though the adaptation mechanism still could remove the model uncertainties in less than two seconds.\vspace{-0.25cm}
\section{INTRODUCTION} \label{Sec:Intro} \vspace{-0.2cm}
The key feature of sliding mode control (SMC) is converting a high dimensional tracking control problem into a lower dimensional stabilization control problem~\cite{Slotine}. SMC is well known for its robust characteristics against model uncertainty/mismatch and external disturbances, while it requires low computational efforts. However, there are challenging issues that arise during digital implementation of SMC, among which chattering phenomenon has been widely reported in the literature~\cite{utkin2013sliding}. One effective approach for reducing the oscillation due to chattering is the use of higher order SMC for continuous-time systems. This approach was first introduced in the 1980s~\cite{nollet2008observer}. The basic idea of the higher order SMC is to drive all the higher order derivatives of the sliding variable to the sliding manifold, in addition to the zero convergence condition of the sliding variable. In this approach, the chattering caused by the discontinuity is transferred to the higher derivatives of the sliding variable. The final control input is calculated by integrating the $r-1$ derivatives of the input for $r-1$ times, and the result would be a continuous chattering-free signal of a $r^{th}$-order SMC~\cite{nollet2008observer}. Higher order SMC leads to less oscillations; however, it adds complexity to the calculations.
In addition to the high frequency oscillations issue, it is shown in the literature~\cite{Amini_DSC,AminiSAE2016,Hansen_DSCC} that upon digital implementation of the baseline SMC software, controller performance degrades from its expected behaviour significantly. The gap between the designed and implemented SMCs is mostly created due to data sampling errors that are introduced by the analog-to-digital (ADC) converter unit at the controller input/output (I/O).
Discrete sliding mode control (DSMC) was shown to be an effective approach to mitigate the ADC implementation imprecisions and enhancing the controller robustness against ADC uncertainties~\cite{Pan_Discrete,Amini_ACC2016}. However, the chattering phenomenon due to the discontinuous nature of the discrete controller is more problematic for the DSMC and can even lead to instability since the
sampling rate is far from infinite~\cite{sira1991non}.
Similar to continuous-time SMC, it was shown in~\cite{mihoub2009real} that a second order DSMC shows less oscillations compared to a first order DSMC. The second order DSMC in~\cite{mihoub2009real} is formulated for linear systems without consideration of the uncertainties in the model. Moreover, the study in~\cite{mihoub2009real} lacks the stability analysis of the closed-loop system. In this paper, a new second order DSMC formulation is developed for a general class of single-input single-output (SISO) \textit{uncertain nonlinear systems}. Moreover, the asymptotic stability of the new controller is guaranteed via a Lyapunov stability argument.
Similar to implementation imprecisions, any uncertainty in the plant model, which is used for designing the model-based controller, results in a significant gap between the designed and implemented controllers.
The previous works in the literature that aimed to handle uncertainties in the model via an adaptive SMC structure are limited to continuous-time domain~\cite{Slotine}, and linear systems~\cite{Chan_Automatica}. The adaptive DSMC formulation from our previous works in~\cite{Pan_DSC,Amini_DSCC2016,Amini_CEP} presents a generic solution for removing the model uncertainties for nonlinear systems based on a first order DSMC formulation. The proposed second order DSMC formulation from this paper allows us to derive the adaptation laws via a Lyapunov stability argument to remove the uncertainty in the plant's model quickly.
The contribution of this paper is threefold. First, a new second order DSMC is formulated for a general class of nonlinear affine systems. Second, the proposed controller is extended to handle the multiplicative type of model uncertainty using a discrete Lyapunov stability argument that also guarantees the asymptotic stability of the closed-loop system. Third, this paper presents the first application of the second order DSMC for an automotive combustion engine control problem. The proposed second order DSMC not only demonstrates robust behavior against data sampling imprecisions compared to a first order DSMC, but it also removes the uncertainties in the model quickly and steers the dynamics to their nominal values.
\vspace{-0.15cm}
\section{Second Order Sliding Mode Control} \vspace{-0.15cm} \label{sec:UncertaintyPrediction}
\subsection{Continuous Second Order Sliding Mode Control}\vspace{-0.10cm} \label{sec:ContinousTimeSecondSMC}
A general class of continuous-time SISO nonlinear systems can be expressed as follows:
\vspace{-0.25cm}
\begin{gather}\label{eq:C2SMC_1}
\dot{x}=f(t,x,u)
\end{gather}
where $x{\in{\mathbb{R}^{n}}}$ and $u{\in{\mathbb{R}}}$ are the state and the input variables, respectively. The sliding mode order is the number of continuous successive derivatives of the differentiable sliding variable $s$, and it is a measure of the degree of smoothness of the sliding variable in the vicinity of the sliding manifold. For the continuous-time systems, the $r^{th}$ order sliding mode is determined by the following equalities~\cite{salgado2004robust}:
\vspace{-0.22cm}
\begin{gather}\label{eq:C2SMC_2}
s(t,x)=\dot{s}(t,x)=\ddot{s}(t,x)=...=s^{r-1}(t,x)=0
\end{gather}
The sliding variable ($s$) is defined as the difference between desired ($x_d$) and measured signal ($x$):
\vspace{-0.22cm}
\begin{gather}\label{eq:C2SMC_3}
s(t,x)=x(t)-x_d(t)
\end{gather}
For the second order SMC design,
a new sliding variable ($\xi$) is defined according to $s$ and $\dot{s}$
\vspace{-0.22cm}
\begin{gather}\label{eq:C2SMC_4}
\xi(t,x)=\dot{s}(t,x)+\lambda s(t,x),~\lambda>0
\end{gather}
Eq.~(\ref{eq:C2SMC_4}) describes the sliding surface of a system with a relative order equal to one, in which the input is $\dot{u}$ and output is $\xi(t,x)$~\cite{sira1990structure}. The control input is obtained according to the following law:
\vspace{-0.22cm}
\begin{gather}\label{eq:C2SMC_5}
\dot{\xi}(t,x)=0 \Rightarrow \ddot{s}(t,x)+\lambda \dot{s}(t,x)=0
\end{gather}
which according to the sliding variable definition needs the second derivative of the state variable ($\ddot{x}(t)$).~
By substituting Eq.~(\ref{eq:C2SMC_3}) and $\ddot{x}(t)$ into Eq.~(\ref{eq:C2SMC_5}), $\dot{u}$ is calculated as follows:
\vspace{-0.25cm}
\begin{gather}\label{eq:C2SMC_9}
\dot{u}(t)=\frac{1}{\frac{\partial}{\partial u}f(t,x,u)}\Big(-\frac{\partial}{\partial t}f(t,x,u) \\-\big(\frac{\partial}{\partial x}f(t,x,u)\big)f(t,x,u)+\ddot{x}_d(t)-\lambda\dot{s}(t,x)\Big) \nonumber
\end{gather}
and finally the control input is:
\vspace{-0.25cm}
\begin{gather}\label{eq:C2SMC_10}
u(t)=\int\dot{u}(t)dt
\end{gather}
This approach guarantees asymptotic convergence of the sliding variable and its derivative to zero~\cite{mihoub2009real}. \vspace{-0.15cm}
\subsection{Discrete Adaptive Second Order Sliding Mode Control} \label{sec:DiscreteTimeSecondDSMC}
The affine form of the nonlinear system in Eq.~(\ref{eq:C2SMC_1}) with an unknown multiplicative term ($\alpha$) can be presented using the following state space equation:
\vspace{-0.15cm}
\begin{gather}\label{eq:D2SMC_1}
\dot{x}(t)=\alpha f(x(t))+g(x(t))u(t)
\end{gather}
where $g(x(t))$ is a non-zero input coefficient and $f(x(t))$ represents the dynamics of the plant and does not depend on the inputs. $\alpha$ is an unknown constant, and represents the errors in the modeled plant's dynamic. By applying the first order Euler approximation the continuous model in Eq.~(\ref{eq:D2SMC_1}) is descritized as follows:
\vspace{-0.15cm}
\begin{gather}\label{eq:D2SMC_2}
x(k+1)=T\alpha f(x(k))+Tg(x(k))u(k)+x(k)
\end{gather}
where $T$ is the sampling time. Similar to Eq.~(\ref{eq:C2SMC_4}), a new discrete sliding variable is defined:
\vspace{-0.25cm}
\begin{gather}\label{eq:D2SMC_3}
\xi(k)={s}(k+1)+\beta s(k),~\beta>0
\end{gather}
where $s(k)=x(k)-x_d(k)$, and $\beta$ is the new sliding variable gain. Substituting Eq.~(\ref{eq:D2SMC_2}) into Eq.~(\ref{eq:D2SMC_3}) yields:
\vspace{-0.2cm}
\begin{gather}\label{eq:D2SMC_4}
\xi(k)=T\alpha f(x(k))+Tgu(k)+x(k)-x_d(k+1)+\beta s(k)
\end{gather}
The second order discrete sliding law is defined as~\cite{mihoub2009real}:
\vspace{-0.25cm}
\begin{gather}\label{eq:D2SMC_5}
\xi(k+2)=\xi(k+1)=\xi(k)=0
\end{gather}
Applying Eq.~(\ref{eq:D2SMC_5}) to the nonlinear system in Eq.~(\ref{eq:D2SMC_2}) results in the following control input:
\vspace{-0.15cm}
\begin{gather}\label{eq:D2SMC_6}
u(k)=\frac{1}{gT}\Big(-T\hat{\alpha}(k)f(x(k))-x(k)+x_d(k+1)-\beta s(k)\Big)
\end{gather}
where $\hat{\alpha}$ is the estimation of the unknown multiplicative uncertainty term in the plant's model. By incorporating the control law ($u$) into the second order sliding variable ($\xi$), we have:
\vspace{-0.3cm}
\begin{gather}\label{eq:D2SMC_7}
\xi(k)=Tf(\alpha-\hat{\alpha}(k))=Tf\tilde{\alpha}(k)
\end{gather}
where ($\tilde{\alpha}$) is the difference between the unknown and estimated multiplicative uncertainty terms ($\tilde{\alpha}(k)=\alpha-\hat{\alpha}(k)$). In order to determine the stability of the closed-loop system, and derive the adaptation law to remove the uncertainty in the model, a Lyapunov stability analysis is employed here. The following Lyapunov candidate function is proposed:
\begin{gather}\label{eq:D2SMC_8}
V(k)=\frac{1}{2}\Big({s}^2(k+1)+\beta{s}^2(k)\Big)\\
+\frac{1}{2}\rho_{\alpha}\Big(\tilde{\alpha}^2(k+1)+\beta \tilde{\alpha}^2(k)\Big) \nonumber
\end{gather}
where $\rho_{\alpha}>0$ is a tunable parameter (adaptation gain) chosen for the numerical sensitivity of the unknown parameter estimation. The proposed Lyapunov function in Eq.~(\ref{eq:D2SMC_8}) is positive definite and quadratic with respect to the sliding variable ($s(k)$) and the unknown parameter estimation error ($\tilde{\alpha}(k)$). In the discrete time domain, the negative semi-definite condition is required for the difference function of $V$ to guarantee the asymptotic stability of the closed-loop system~\cite{Pan_DSC,Amini_CEP}. The Lyapunov difference function is calculated using a Taylor series expansion: \vspace{-0.20cm}
\begin{gather}\label{eq:D2SMC_9}
V(k+1)=V(k)+\frac{\partial V(k)}{\partial s(k)}\Delta s(k)\\
+\frac{\partial V(k)}{\partial s(k+1)}\Delta s(k+1)+\frac{\partial V(k)}{\partial \tilde{\alpha}(k)}\Delta \tilde{\alpha}(k) \nonumber \\
+\frac{\partial V(k)}{\partial \tilde{\alpha}(k+1)}\Delta \tilde{\alpha}(k+1)
+\frac{1}{2} \frac{\partial^2 V(k)}{\partial {s}^2(k)}\Delta {s}^2(k) \nonumber \\
+\frac{1}{2} \frac{\partial^2 V(k)}{\partial {s}^2(k+1)}\Delta {s^2(k+1)}+
\frac{1}{2} \frac{\partial^2 V(k)}{\partial {\tilde{\alpha}}^2(k)}\Delta {\tilde{\alpha}}^2(k) \nonumber \\
+\frac{1}{2} \frac{\partial^2 V(k)}{\partial {\tilde{\alpha}}^2(k+1)}\Delta {\tilde{\alpha}}^2(k+1)+... \nonumber
\end{gather
where $\Delta s(k)\equiv s(k+1)-s(k)$ and $\Delta \tilde{\alpha}(k) \equiv \tilde{\alpha}(k+1)-\tilde{\alpha}(k)$.
Next, the Lyapunov difference function ($\Delta V(k)=V(k+1)-V(k)$) is calculated by substituting the values of the partial derivatives into Eq.~(\ref{eq:D2SMC_9}): \vspace{-0.15cm}
\begin{gather}\label{eq:D2SMC_11}
\Delta V(k)=\beta s(k)\Delta s(k)+s(k+1)\Delta s(k+1) \\
+\rho_{\alpha}\beta\tilde{\alpha}(k)\Delta\tilde{\alpha}(k)+\rho_{\alpha}\tilde{\alpha}(k+1)\Delta\tilde{\alpha}(k+1) \nonumber \\
+\frac{1}{2} \beta\Delta {s}^2(k)+\frac{1}{2}\Delta{s}^2(k+1)\nonumber \\
+\frac{1}{2}\rho_{\alpha}\beta\Delta{\tilde{\alpha}}^2(k)+\frac{1}{2}\rho_{\alpha}\Delta{\tilde{\alpha}}^2(k+1)+... \nonumber
\end{gather}
where cross term second order derivatives are zero. Eq.~(\ref{eq:D2SMC_11}) can be simplified after substituting Eq.~(\ref{eq:D2SMC_7}) at $k$ and $k+1$ time steps: \vspace{-0.2cm}
\begin{gather}\label{eq:D2SMC_12}
\Delta V(k)=-\beta(\beta+1)s^2(k)-(\beta+1)s^2(k+1)\\
+\beta s(k)Tf\tilde{\alpha}(k)+\rho_{\alpha}\beta \tilde{\alpha}(k)\Delta \tilde{\alpha}(k)\nonumber \\
+s(k+1)Tf\tilde{\alpha}(k+1)+\rho_{\alpha}\tilde{\alpha}(k+1)\Delta\tilde{\alpha}(k+1) \nonumber \\
+O\left(\Delta {s}^2(k),\Delta{s}^2(k+1),\Delta\tilde{\alpha}^2(k),\Delta\tilde{\alpha}^2(k+1)\right)+... \nonumber
\end{gather}
which yields: \vspace{-0.25cm}
\begin{gather}\label{eq:D2SMC_13}
\Delta V(k)=-(\beta+1)\big(s^2(k+1)+\beta s^2(k)\big)\\
+\rho_{\alpha}\beta\tilde{\alpha}(k)\Big(\frac{s(k)Tf}{\rho_{\alpha}}+\Delta\tilde{\alpha}(k)\Big) \nonumber \\
+\rho_{\alpha}\tilde{\alpha}(k+1)\Big(\frac{s(k+1)Tf}{\rho_{\alpha}}+\Delta\tilde{\alpha}(k+1)\Big) \nonumber \\
+O\left(\Delta {s}^2(k),\Delta{s}^2(k+1),\Delta\tilde{\alpha}^2(k),\Delta\tilde{\alpha}^2(k+1)\right)+... \nonumber
\end{gather}
in which the higher order ($>2$) terms are zero. As can be seen from Eq.~(\ref{eq:D2SMC_13}), the first term is negative definite when $\beta>0$. To guarantee the asymptotic stability of the closed-loop system, and minimize the tracking errors, the Lyapunov difference function should be at least negative semi-definite~\cite{Amini_DSC}. To this end, the second and third terms in Eq.~(\ref{eq:D2SMC_13}) should become zero, which leads to the following adaptation law:\vspace{-0.35cm}
\begin{gather}\label{eq:D2SMC_14}
\tilde{\alpha}(k+1)=\tilde{\alpha}(k)-\frac{s(k)Tf}{\rho_{\alpha}}
\end{gather}
By using Eq.~(\ref{eq:D2SMC_14}) to estimate the unknown uncertainty term, the Lyapunov difference function becomes: \vspace{-0.20cm}
\begin{gather}\label{eq:D2SMC_15}
\Delta V(k)=-(\beta+1)\big(s^2(k+1)+\beta s^2(k)\big)\\
+O\left(\Delta {s}^2(k),\Delta{s}^2(k+1),\Delta\tilde{\alpha}^2(k),\Delta\tilde{\alpha}^2(k+1)\right) \nonumber
\end{gather}
Let us assume that by using Eq.~(\ref{eq:D2SMC_14}), the uncertainty in the model will be removed. This means that the error in estimating the unknown parameter converges to zero ($\tilde{\alpha}(k+1)=\tilde{\alpha}(k)=0$). Thus, by expanding the second order terms ($O(.)$), and assuming a small enough sampling time ($T$), which means all terms that contain $T^2$ can be neglected, Eq.~(\ref{eq:D2SMC_11}) can be re-arranged as follows: \vspace{-0.20cm}
\begin{gather}\label{eq:D2SMC_16}
\Delta V(k)=\beta s(k)(s(k+1)-s(k))\\
+s(k+1)(s(k+2)-s(k+1))+\frac{1}{2} \beta(s(k+1)-s(k))^2 \nonumber \\
+\frac{1}{2}(s(k+2)-s(k+1))^2+... \nonumber
\end{gather}
Since it was assumed that the uncertainty in the model is compensated by Eq.~(\ref{eq:D2SMC_14}), $s(k+1)$ and $s(k+2)$ can be replaced by $-\beta s(k)$ and $\beta^2 s(k)$, respectively, according to Eq.~(\ref{eq:D2SMC_5}). Thus, Eq.~(\ref{eq:D2SMC_16}) can be simplified as: \vspace{-0.25cm}
\begin{gather}\label{eq:D2SMC_18}
\Delta V(k)=-\frac{1}{2}\beta\big(-{\beta}^3-{\beta}^2+\beta+1\big)s^2(k)
\end{gather}
$-{\beta}^3-{\beta}^2+\beta+1$ is positive if $1>\beta>0$. In other words, if $1>\beta>0$, then $\Delta V(k)\leq 0$, which guarantees the asymptotic stability of the system: \vspace{-0.15cm}
\begin{gather}\label{eq:D2SMC_19}
V\rightarrow 0 \Rightarrow s(k+1)~\&~s(k) \rightarrow 0 \xRightarrow[]{1>\beta>0} \xi(k) \rightarrow 0
\end{gather}
It was shown that the second order sliding mode (Eq.~(\ref{eq:D2SMC_5})) and the adaptation law from Eq.~(\ref{eq:D2SMC_14}) guarantee the negative semi-definite condition of the Lyapunov difference function. This means that the sliding variable ($s$) and the error in estimating the unknown parameter ($\tilde{\alpha}$) converge to zero in finite time. Moreover, since the second order DSMC steers both first and second derivatives (difference functions) of the sliding variable to the origin, it provides better tracking performance, lower chattering, and higher robustness against data sampling imprecisions, compared to the first order DSMC. Fig.~\ref{fig:AdaptiveDSMC_Schematic} shows the overall schematic of the proposed second order adaptive DSMC along with the adaptation mechanism. \vspace{-0.4cm}
\begin{figure}[h!]
\begin{center}
\includegraphics[width=\columnwidth]{AdaptiveDSMC_Schematic.eps} \vspace{-0.75cm}
\caption{\label{fig:AdaptiveDSMC_Schematic} Schematic of the proposed second order adaptive DSMC.} \vspace{-0.7cm}
\end{center}
\end{figure}
| 27c1b6a0bef3229b49f673e63b47385746906fd3 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} \label{intro}
Advances in a wide range of medical imaging technologies have revolutionized how we view functional and pathological events in the body and define anatomical structures in which these events take place. X-ray, CAT, MRI, Ultrasound, nuclear medicine, among other medical imaging technologies, enable 2D or tomographic 3D images to capture in-vivo structural and functional information inside the body for diagnosis, prognosis, treatment planning and other purposes.
One fundamental problem in medical image analysis is image segmentation, which identifies the boundaries of objects such as organs or abnormal regions (e.g. tumors) in images. Since manually annotation can be very time-consuming and subjective, an accurate and reliable automatic segmentation method is valuable for both clinical and research purpose. Having the segmentation result makes it possible for shape analysis, detecting volume change, and making a precise radiation therapy treatment plan.
In
the literature of image processing and computer vision, various theoretical frameworks have been proposed for automatic segmentation. Traditional unsupervised methods such as thresholding~\cite{otsu1979threshold}, region growing~\cite{adams1994seeded}, edge detection and grouping~\cite{canny1986computational}, Markov Random Fields (MRFs)~\cite{manjunath1991unsupervised}, active contour models~\cite{kass1988snakes}, Mumford-Shah functional based frame partition~\cite{mumford1989optimal}, level sets~\cite{malladi1995shape}, graph cut~\cite{shi2000normalized}, mean shift~\cite{comaniciu2002mean}, and their extensions and integrations~\cite{gooya2011deformable, lee2008segmenting, lefohn2003interactive} usually utilize constraints about image intensity or object appearance. Supervised methods~\cite{menze2015multimodal,cobzas20073d, geremia2011spatial, wels2008discriminative,ronneberger2015u, havaei2017brain}, on the other hand, directly learn from labeled training samples, extract features and context information in order to perform a dense pixel (or voxel)-wise classification.
Convolutional Neural Networks (CNNs) have been widely applied to visual recognition problems in recent years, and they are shown effective in learning a hierarchy of features at multiple scales from data. For pixel-wise semantic segmentation, CNNs have also achieved remarkable success. In~\cite{long2015fully}, Long \emph{et al.} first proposed a fully convolutional networks (FCNs) for semantic segmentation. The authors replaced conventional fully connected layers in CNNs with convolutional layers to obtain a coarse label map, and then upsampled the label map with deconvolutional layers to get per pixel classification results. Noh \emph{et al.}~\cite{noh2015learning} used an encoder-decoder structure to get more fine details about segmented objects.
With multiple unpooling and deconvolutional layers in their architecture, they avoided the coarse-to-fine stage in~\cite{long2015fully}. However, they still needed to ensemble with FCNs in their method to capture local dependencies between labels. Lin \emph{et al.}~\cite{lin2016efficient} combined Conditional Random Fields (CRFs) and CNNs to better explore spatial correlations between pixels, but they also needed to implement a dense CRF to refine their CNN output.
In the field of medical image segmentation, deep CNNs have also been applied with promising results. Ronneberger \emph{et al.}~\cite{ronneberger2015u} presented a FCN, namely U-net, for segmenting neuronal structures in electron microscopic stacks. With the idea of skip-connection from~\cite{long2015fully}, the U-net achieved very good performance and has since been applied to many different tasks such as image translation~\cite{isola2016image}.
In addition, Havaei \emph{et al.}~\cite{havaei2017brain} obtained good performance for medical image segmentation with their InputCascadeCNN. The InputCascadeCNN has image patches as inputs and uses a cascade of CNNs in which the output probabilities of a first-stage CNN are taken as additional inputs to a second-stage CNN.
Pereira \emph{et al.}~\cite{pereira2016brain} applied deep CNNs with small kernels for brain tumor segmentation.
They proposed different architectures for segmenting high grade and low grade tumors, respectively.
Kamnitsas \emph{et al.}~\cite{kamnitsas2017efficient} proposed a 3D CNN using two pathways with inputs of different resolutions. 3D CRFs were also needed to refine their results.
Although these previous approaches using CNNs for segmentation have achieved promising results, they still have limitations. All above methods utilize a pixel-wise loss, such as softmax, in the last layer of their networks, which is insufficient to learn both local and global contextual relations between pixels. Hence they always need models such as CRFs ~\cite{chen14semantic} as a refinement to enforce spatial contiguity in the output label maps. Many previous methods \cite{havaei2017brain,kamnitsas2017efficient,pereira2016brain} address this issue by training CNNs on image patches and using multi-scale, multi-path CNNs with different input resolutions or different CNN architectures. Using patches and multi-scale inputs could capture spatial context information to some extent.
Nevertheless, as described in U-net~\cite{ronneberger2015u}, the computational cost for patch training is very high and there is a trade-off between localization accuracy and the patch size.
Instead of training on small image patches, current state-of-the-art CNN architectures such as U-net are trained on whole images or large image patches and use skip connections to combine hierarchical features for generating the label map. They have shown potential to implicitly learn some local dependencies between pixels. However, these methods are still limited by their pixel-wise loss function, which lacks the ability to enforce the learning of multi-scale spatial constraints directly in the end-to-end training process.
Compared with patch training, an issue for CNNs trained on entire images is label or class imbalance. While patch training methods can sample a balanced number of patches from each class, the numbers of pixels belonging to different classes in full-image training methods are usually imbalanced. To mitigate this problem, U-net uses a weighted cross-entropy loss to balance the class frequencies. However, the choice of weights in their loss function is task-specific and is hard to optimize. In contract to the weighted loss in U-net, a general loss that could avoid class imbalance as well as extra hyper-parameters would be more desirable.
In this paper, we propose a novel end-to-end Adversarial Network architecture, called SegAN, with a multi-scale $L_1$ loss function, for semantic segmentation. Inspired by the original GAN~\cite{goodfellow2014generative}, the training procedure for SegAN is similar to a two-player min-max game in which a segmentor network ($S$) and a critic network ($C$) are trained in an alternating fashion to respectively minimize and maximize an objective function. However, there are several major differences between our SegAN and the original GAN that make SegAN significantly better for the task of image segmentation.
\begin{itemize}
\vspace{-6pt}
\item
In contrast to classic GAN with separate losses for generator and discriminator, we propose a novel multi-scale loss function for both segmentor and critic.
Our critic is trained to maximize a novel multi-scale $L_{1}$ objective function that takes into account CNN feature differences between the predicted segmentation and the ground truth segmentation at multiple scales (i.e. at multiple layers).
\item We use a fully convolutional neural network (FCN) as the segmentor $S$, which is trained with only gradients flowing through the critic, and with the objective of minimizing the same loss function as for the critic.
\item Our SegAN is an end-to-end architecture trained on whole images, with no requirements for patches, or inputs of multiple resolutions, or further smoothing of the predicted label maps using CRFs.
\end{itemize}
By training the entire system end-to-end with back propagation and alternating the optimization of $S$ and $C$, SegAN can directly learn spatial pixel dependencies at multiple scales. Compared with previous methods that learn hierarchical features with multi-scale multi-path CNNs~\cite{farabet2013learning}, our SegAN network applies a novel multi-scale loss to enforce the learning of hierarchical features in a more straightforward and efficient manner.
Extensive experimental results demonstrate that the proposed SegAN achieves comparable or better results than the state-of-the-art CNN-based architectures including U-net.
The rest of this paper is organized as follows. Section \ref{methods} introduces our SegAN architecture and methodology. Experimental results are presented in Section 3. Finally, we conclude this paper in Section 4.
\section{Methodology} \label{methods}
\begin{figure*}[tb]
\centerline{
\includegraphics[width=0.99\textwidth,height = 350pt]{framework.pdf}}
\caption{The architecture of the proposed SegAN with segmentor and critic networks. $4 \times 4$ convolutional layers with stride $2$ (S2) and the corresponding number of feature maps (e.g., N64) are used for encoding, while image resize layers with a factor of $2$ (R2) and $3 \times 3$ convolutional layers with stride $1$ are used for decoding. Masked images are calculated by pixel-wise multiplication of a label map and (the multiple channels of) an input image. Note that, although only one label map (for whole tumor segmentation) is illustrated here, multiple label maps (e.g. also for tumor core and Gd-enhanced tumor core) can be generated by the segmentor in one path. }
\label{fig:architecture}
\end{figure*}
As illustrated in Figure~\ref{fig:architecture}, the proposed SegAN consists of two parts: the segmentor network $S$ and the critic network $C$.
The segmentor is a fully convolutional encoder-decoder network that generates a probability label map from input images.
The critic network is fed with two inputs: original images masked by ground truth label maps, and original images masked by predicted label maps from $S$.
The $S$ and $C$ networks are alternately trained in an adversarial fashion: the training of $S$ aims to minimize our proposed multi-scale $L_1$ loss, while the training of $C$ aims to maximize the same loss function.
\vspace{-8pt}
\subsection{The multi-scale $L_1$ loss}
The conventional GANs \cite{goodfellow2014generative} have an objective loss function defined as:
\begin{equation}
\begin{split}
\min_{\theta_G} &\max_{\theta_D} \mathcal{L}(\theta_G,\theta_D)\\
&= \mathbb{E}_{x\sim P_\mathrm{data}}[\log D(x)] + \mathbb{E}_{z\sim P_z}\log (1 - D(G(z)))] \enspace .
\end{split}
\end{equation}
In this objective function, $x$ is the real image from an unknown distribution $P_\mathrm{data}$, and $z$ is a random input for the generator, drawn from a probability distribution (such as Gaussion) $P_z$. $\theta_G$ and $\theta_D$ represent the parameters for the generator and discriminator in GAN, respectively.
In our proposed SegAN, given a dataset with $N$ training images $x_n$ and corresponding ground truth label maps $y_n$, the multi-scale objective loss function $\mathcal{L}$ is defined as:
\begin{equation}
\begin{split}
\min\limits_{\theta_S} &\max\limits_{\theta_C}\mathcal{L}(\theta_S,\theta_C) \\
&= \frac{1}{N}\sum_{n=1}^N
\ell_\mathrm{mae}(f_C(x_n \circ S(x_n)), f_C(x_n \circ y_n)) \enspace ,
\label{eq:SegGANloss}
\end{split}
\end{equation}
where $\ell_\mathrm{mae}$ is the Mean Absolute Error (MAE) or $L_1$ distance;
$x_n \circ S(x_n)$ is the input image masked by segmentor-predicted label map (i.e., pixel-wise multiplication of predicted\_label\_map and original\_image); $x_n \circ y_n$ is the input image masked by its ground truth label map (i.e., pixel-wise multiplication of ground\_truth\_label\_map and original\_image); and $f_C(x)$ represents the hierarchical features extracted from image $x$ by the critic network. More specifically, the $\ell_\mathrm{mae}$ function is defined as:
\begin{equation}
\ell_\mathrm{mae}(f_C({x}), f_C({x}^{\prime})) = \frac{1}{L}\sum_{i=1}^L ||f_C^i({x}) - f_C^i({x}^{\prime})||_{1} \enspace ,\label{eq:mae}
\end{equation}
where $L$ is the total number of layers (i.e. scales) in the critic network, and $f_C^i({x})$ is the extracted feature map of image ${x}$ at the $i$th layer of $C$.
\subsection{SegAN Architecture}
\textbf{Segmentor.} We use a fully convolutional encoder-decoder structure for the segmentor $S$ network.
We use the convolutional layer with kernel size $4 \times 4$ and stride $2$ for downsampling, and perform upsampling by image resize layer with a factor of $2$ and convolutional layer with kernel size $3 \times 3$ stride $1$.
We also follow the U-net and add skip connections between corresponding layers in the encoder and the decoder.
\noindent\textbf{Critic.}
The critic $C$ has the similar structure as the decoder in $S$.
Hierarchical features are extracted from multiple layers of $C$ and used to compute the multi-scale $L_1$ loss.
This loss can capture long- and short-range spatial relations between pixels by using these hierarchical features, i.e., pixel-level features, low-level (e.g. superpixels) features, and middle-level (e.g. patches) features.
More details including activation layers (e.g., leaky ReLU), batch normalization layer and the number of feature maps used in each convolutional layers can be found in Figure~\ref{fig:architecture}.
\subsection{Training SegAN}
The segmentor $S$ and critic $C$ in SegAN are trained by back-propagation from the proposed multi-scale $L_1$ loss. In an alternating fashion, we first fix $S$ and train $C$ for one step using gradients computed from the loss function, and then fix $C$ and train $S$ for one step using gradients computed from the same loss function passed to $S$ from $C$.
As shown in (\ref{eq:SegGANloss}), the training of $S$ and $C$ is like playing a min-max game: while $G$ aims to minimize the multi-scale feature loss, $C$ tries to maximize it. As training progresses, both the $S$ and $C$ networks become more and more powerful. And eventually, the segmentor will be able to produce predicted label maps that are very close to the ground truth as labeled by human experts. We also find that the $S$-predicted label maps are smoother and contain less noise than manually-obtained ground truth label maps.
We trained all networks using RMSProp solver with batch size $64$ and learning rate $0.00002$. We used a grid search method to select the best values for the number of up-sampling blocks and the number of down-sampling blocks for the segmentor (four, in both cases), and for the number of down-sampling blocks for the critic (three).
\vspace{-25pt}
\subsection{Proof of training stability and convergence}
\vspace{-10pt}
Having introduced the multi-scale $L_1$ loss, we next prove that our training is stable and finally reaches an equilibrium. First, we introduce some notations.
Let $f:\mathcal{X} \rightarrow \mathcal{X}^{\prime}$ be the mapping between an input medical image and its corresponding ground truth segmentation, where $\mathcal{X}$ represents the compact space of medical images\footnote[1]{Although the pixel value ranges of medical images can vary, one can always normalize them to a certain value range such as [0,1], so it is compact.} and $\mathcal{X}^{\prime}$ represents the compact space of ground truth segmentations. We approximate this ground truth mapping $f$ with a segmentor neural network $g_\theta : \mathcal{X} \rightarrow \mathcal{X}^{\prime}$ parameterized by vector $\theta$ which takes an input image, and generates a segmentation result. Assume the best approximation to the ground truth mapping by a neural network is the network $g_{\hat{\theta}}$ with optimal parameter vector $\hat{\theta}$.
Second, we introduce a lemma about the Lipschitz continuity of either the segmentor or the critic neural network in our framework.
\begin{lemma}\label{lem1}
Let $g_\theta$ be a neural network parameterized by $\theta$, and $x$ be some input in space $\mathcal{X}$, then $g_\theta$ is Lipschitz continuous with a bounded Lipschitz constants $K(\theta)$ such that
\begin{equation}||g_{\theta}(x_1) - g_{\theta}(x_2)||_1 \leqslant K(\theta)(||x_1 - x_2||_1) \enspace ,
\label{eq:lemma1_1}
\end{equation}
and for different parameters with same input we have
\begin{equation}
||g_{\theta_1}(x) - g_{\theta_2}(x)||_1 \leqslant K(x)||\theta_1 - \theta_2||_1
\enspace ,
\label{eq:lemma1_2}
\end{equation}
\end{lemma}
Now we prove Lemma~\ref{lem1}.
\begin{proof}
Note that the neural network consists of several affine transformations and pointwise nonlinear activation functions such as leaky ReLU (see Figure~\ref{fig:architecture}). All these functions are Lipschitz continuous because all their gradient magnitudes are within certain ranges. To prove Lemma~\ref{lem1}, it's equivalent to prove the gradient magnitudes of $g_\theta$ with respect to $x$ and $\theta$ are bounded. We start with a neural network with only one layer: $g_\theta(x) = A_1(W_1x)$ where $A_1$ and $W_1$ represent the activation and weight matrix in the first layer. We have $\nabla_{x}g_{\theta}(x) = W_1 D_1$ where $D_1$ is the diagonal Jacobian of the activation, and we have
$\nabla_{\theta} g_{\theta}(x) = D_1 x$ where $\theta$ represents the parameters in the first layer.
Then we consider the neural network with $L$ layers.
We apply the chain rule of the gradient and we have $\nabla_{x}g_{\theta}(x) = \prod_{k=1}^{L}W_k D_k$ where $k$ represent the $k$-th layer of the network. Then we have
\begin{equation}
\begin{split}
||\nabla_{x} g_{\theta}(x)||_1 = ||\prod_{k=1}^{L}W_k D_k||_1\enspace .
\end{split}
\end{equation}
Due to the fact that all parameters and inputs are bounded, we have proved (\ref{eq:lemma1_1}).
Let's denote the first $i$ layers of the neural network by $g^i$ (which is another neural network with less layers), we can compute the gradient with respect to the parameters in $i$-th layer as $\nabla_{\theta_i} g_{\theta}(x) = (\prod_{k=i+1}^{L}W_k D_k)D_ig^{i-1}(x)$. Then we sum parameters in all layers and get
\begin{equation}
\begin{split}
||\nabla_{\theta} g_{\theta}(x)||_1 &= ||\sum_{i=1}^{L}(\prod_{k=i+1}^{L}W_k D_k)D_ig^{i-1}(x)||_1\\
&\leqslant \sum_{i=1}^{L}||((\prod_{k=i+1}^{L}W_k D_k)D_i)g^{i-1}(x)||_1\enspace .
\end{split}
\end{equation}
Since we have proved that $g(x)$ is bounded, we finish the proof of (\ref{eq:lemma1_2}).
\end{proof}
Based on Lemma~\ref{lem1}, we then prove that our multi-scale loss is bounded and won't become arbitrarily large during the training, and it will finally converge.
\begin{theorem}
Let $\mathcal{L}_t(x)$ denote the multi-scale loss of our SegAN at training time $t$ for input image $x$, then there exists a small constant $C$ so that
\begin{equation}
\lim_{t\rightarrow + \infty} \mathbb{E}_{x \in \mathcal{X}} \mathcal{L}_t(x) \leqslant C\enspace .
\end{equation}
\label{thm:thm1}
\end{theorem}
\begin{proof}
Let $g$ and $d$ represent the segmentor and critic neural network, $\theta$ and $w$ be the parameter vector for the segmentor and critic, respectively. Without loss of generality, we omit the masked input for the critic and rephrase (\ref{eq:SegGANloss}) and (\ref{eq:mae}) as
\begin{equation}
\min_{\theta} \max_{w} \mathcal{L}_t = \mathbb{E}_{x \in \mathcal{X}} \frac{1}{L}\sum_{i=1}^L ||d^i(g_{\theta}({x})) - d^i(g_{\hat{\theta}}({x}))||_{1}\enspace ,
\label{eq9}
\end{equation}
recall that $g_{\hat{\theta}}$ is the ground truth segmentor network and $d^i$ is the critic network with only first $i$ layers. Let's firstly focus on the critic. To make sure our multi-scale loss won't become arbitrarily large, inspired by~\cite{arjovsky2017wasserstein}, we clamp the weights of our critic network to some certain range (e.g., $[-0.01,0.01]$ for all dimensions of parameter) every time we update the weights through gradient descent. That is to say, we have a compact parameter space $\mathcal{W}$ such that all functions in the critic network are in a parameterized family of functions $\{d_w\}_{w \in \mathcal{W}}$. From Lemma~\ref{lem1}, we know that $||d_{w}(x_1) - d_{w}(x_2)||_1 \leqslant K(w)(||x_1 - x_2||_1)$. Due to the fact that $\mathcal{W}$ is compact, we can find a maximum value for $K(w)$, $K$, and we have
\begin{equation}
||d(x_1) - d(x_2)||_1 \leqslant K||x_1 - x_2||_1\enspace .
\end{equation}
Note that this constant $K$ only depends on the space $\mathcal{W}$ and is irrelevant to individual weights, so it is true for any parameter vector $w$ after we fix the vector space $\mathcal{W}$. Since Lemma~\ref{lem1} applies for the critic network with any number of layers, we have
\begin{equation}
\frac{1}{L}\sum_{i=1}^L||d^i (g_{\theta}(x)) - d^i (g_{\hat{\theta}}(x))||_1 \leqslant K||g_{\theta}(x) - g_{\hat{\theta}}(x)||_1\enspace .
\label{eq11}
\end{equation}
Now let's move to the segmentor. According to Lemma~\ref{lem1}, we have $||g_{\theta}(x) - g_{\hat{\theta}}(x)||_1 \leqslant K(x)||\theta - \hat{\theta}||_1$, then combined with (\ref{eq11}) we have
\begin{equation}
\frac{1}{L}\sum_{i=1}^L||d^i (g_{\theta}(x)) - d^i (g_{\hat{\theta}}(x))||_1 \leqslant K(x)K||\theta - \hat{\theta}||_1\enspace .
\label{eq12}
\end{equation}
We know $\mathcal{X}$ is compact, so there's a maximal value for $K(x)$ and it only depends on the difference between the ground truth parameter vector $\hat{\theta}$ and the parameter vector of the segmentor $\theta$. Since we don't update weights in the segmentor when we update weights in the critic, there's an upper bound for $\mathcal{L}_t$ when we update the critic network and it won't be arbitrarily large during the min-max game.
When we update the parameters in the segmentor, we want to decrease the loss. This makes sense because smaller loss indicates smaller difference between $\hat{\theta}$ and $\theta$. When $\theta \rightarrow \hat{\theta}$, $\mathcal{L}_t$ converges to zero because the upper bound of $\mathcal{L}$ becomes zero. However, we may not be able to find the global optimum for $\theta$. Now let us denote a reachable local optimum for $\theta$ in the segmentor by $\theta_0$, we will keep updating parameters in the segmentor through gradient descent and gradually approaches $\theta_0$. Based on (\ref{eq9}) and (\ref{eq12}), we denote the maximum of $K(x)$ by $K^\prime$ and have
\begin{equation}
\lim_{t\rightarrow + \infty} \mathcal{L}_t(x) \leqslant K K^{\prime}||\hat{\theta} - \theta_0||_1 = C\enspace .
\end{equation}
Since the constant $C$ does not depend on input $x$, we have proved Theorem~\ref{thm:thm1}.
\end{proof}
\section{Experiments}
In this section, we evaluated our system on the fully-annotated MICCAI BRATS datasets~\cite{menze2015multimodal}. Specifically,
we trained and validated our models using the BRATS 2015 training dataset, which consists of $220$ high grade subjects and $54$ low grade subjects with four modalities: T1, T1c, T2 and Flair. We randomly split the BRATS 2015 training data with the ratio $9 : 1$ into a training set and a validation set. We did such split for the high grade and low grade subjects separately, and then re-combined the resulting sets for training and validation. Each subject in BRATS 2015 dataset is a 3D brain MRI volume with size $240 \times 240 \times 155$. We center cropped each subject into a subvolume of $180 \times 180 \times 128$, to remove the border black regions while still keep the entire brain regions. We did our final evaluation and comparison on the BRATS 2015 test set using the BRATS online evaluation system, which has \emph{Dice}, \emph{Precision} and \emph{Sensitivity} as the evaluation metrics. The Dice score is is identical to the F-score which normalizes the number of true positives to the average size of the two segmented regions:
\begin{equation}
\mathrm{Dice} = \frac{2|P \cap T|}{|P|+|T|}
\end{equation}
where $P$ and $T$ represent the predicted region and the ground truth region, respectively.
Since the BRATS 2013 dataset is a subset of BRATS 2015, we also present our results on BRATS 2013 leaderboard set.
Although some work with 3D patching CNNs have been done for medical image segmentation, due to the limitation of our hardware memory and for the reason that brain images in BRATS dataset are inconsistent in third dimension, we built a 2D SegAN network to generate the label map for each axial slice of a 3D volume and then restack these 2D label maps to produce the 3D label map for brain tumor. Since each subject was center cropped to be a $180 \times 180 \times 128$ volume, it yields 128 axial slices each with the size $180 \times 180$. These axial slices were further randomly cropped to size $160 \times 160$ during training for the purpose of data augmentation. They were centered cropped to size $160 \times 160$ during validation and testing.
We used three modalities of these MRI images: T1c, T2, FLAIR. Corresponding slices of T1c, T2, FLAIR modalities are concatenated along the channel dimension and used as the multi-channel input to our SegAN model, as shown in Figure~\ref{fig:architecture}.
The segmentor of SegAN outputs label maps with the same dimensions as the input images.
As required by the BRATS challenge~\cite{menze2015multimodal}, we did experiments with the objective to generate label maps for three types of tumor regions: \emph{whole tumor}, \emph{tumor core} and \emph{Gd-enhanced tumor core}.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.48\textwidth]{component_analysis.pdf
\caption{Average dice scores of different architectures on BRATS validation set}
\label{fig:component}
\end{figure}
\subsection{Choice of components in SegAN architecture}
In this section, we compare different implementations of the proposed SegAN architecture and also evaluate the effectiveness of the proposed multi-scale $L_1$ loss on the BRATS validation set for the brain tumor segmentation task.
Specifically, we compare the following implementations:
\begin{itemize}
\vspace{-6pt}
\item \textbf{S1-1C.} A separate SegAN is built for every label class, i.e. one segmentor (S1) and one critic (1C) per label.
\item \textbf{S3-1C:} A SegAN is built with one segmentor and one critic, but the segmentor generates a three-channel label map, one channel for each label. Therefore, each 3-channel label map produces three masked images (one for each class), which are then concatenated in the channel dimension and fed into the critic.
\item \textbf{S3-3C.} A SegAN is built with one segmentor that generates a three-channel (i.e. three-class) label map, but three separate critics, one for each label class. The networks, one $S$ and three $C$s, are then trained end-to-end using the average loss computed from all three $C$s.
\item \textbf{S3-3C single-scale loss models.} For comparison, we also built two single-scale loss models: S3-3C-s0 and S3-3C-s3. S3-3C-s0 computes the loss using features from only the input layers (i.e., layer $0$) of the critics, and S3-3C-s3 calculates the loss using features from only the output layers (i.e., layer $3$) of the critics.
\end{itemize}
As shown in Figure~\ref{fig:component}, models S1-1C and S3-3C give similar performance which is the best among all models. Since the computational cost for S1-1C is higher than S3-3C, S3-3C is more favorable and we use it to compare our SegAN model with other methods in Section \ref{sec:comparison}.
In contrast, while model S3-1C is the simplest requiring the least computational cost, it sacrifices some performance; but by using the multi-scale loss, it still performs better than any of the two single-scale loss models especially for segmenting tumor core and Gd-enhanced tumor core regions.
\begin{figure}
\includegraphics[width=0.48\textwidth]{examples}
\caption{Example results of our SegGAN (right) with corresponding T2 slices (left) and ground truth (middle) on BRATS validation set.}
\label{fig:examples}
\label{sec:comparison}
\end{figure}
\begin{table*}[!htbp]
\caption{Comparison to previous methods and a baseline implementation of U-net with softmax loss for segmenting three classes of brain tumor regions: whole, core and Gd-enhanced (Enha.)}
\label{tb:result}
\centering
\begin{tabular}{c|c|c|c|c|c|c|c|c|c|c}
\cline{1-11}
\multirow{2}{2cm}{}&\multirow{2}{*}{Methods} & \multicolumn{3}{|c}{Dice} &
\multicolumn{3}{|c}{Precision} & \multicolumn{3}{|c}{Sensitivity} \\
\cline{3-11}
& & Whole& Core & Enha. & Whole& Core & Enha. & Whole& Core & Enha.\\
\hline
\multirow{3}{1.8cm}{BRATS 2013 Leaderboard}
&Havaei~\cite{havaei2017brain} &
\textbf{0.84} &0.71 &0.57 &\textbf{0.88} &0.79 &0.54 &0.84 &0.72 &0.68 \\
\cline{2-11}
&Pereira~\cite{pereira2016brain} &
\textbf{0.84} &\textbf{0.72} &0.62 &0.85 &\textbf{0.82} &0.60 &\textbf{0.86} &\textbf{0.76} &0.68\\
\cline{2-11}
&\textbf{SegAN} &
\textbf{0.84} & 0.70 & \textbf{0.65} &0.87 &0.80 &\textbf{0.68} &0.83 & 0.74 &\textbf{0.72}\\
\hline
\hline
\multirow{3}{1.8cm}{BRATS 2015 Test} &Kamnitsas~\cite{kamnitsas2017efficient} &
\textbf{0.85} &0.67 &{0.63} &0.85 &\textbf{0.86} &0.63 &\textbf{0.88} &0.60 &\textbf{0.67}\\
\cline{2-11}
&U-net &
0.80 &0.63 &0.64 &0.83 &0.81 &\textbf{0.78} &0.80 &0.58 &0.60\\
\cline{2-11}
&\textbf{SegAN} &
\textbf{0.85} &\textbf{0.70} &\textbf{0.66} &\textbf{0.92} &0.80 &0.69 &0.80 &\textbf{0.65} &0.62\\
\hline
\end{tabular}
\end{table*}
\subsection{Comparison to state-of-the-art}
In this subsection, we compare the proposed method, our S3-3C SegAN model, with other state-of-the-art methods on the BRATS 2013 Leaderboard~\cite{havaei2017brain,pereira2016brain} Test and the BRATS 2015 Test~\cite{kamnitsas2017efficient}.
We also implemented a U-net model \cite{ronneberger2015u} for comparison. This U-net model has the exact same architecture as our SegAN segmentor except that the multi-scale SegAN loss is replaced with the softmax loss in the U-net. Table~\ref{tb:result} gives all comparison results. From the table, one can see that our SegAN compares favorably to the existing state-of-the-art on BRATS 2013 while achieves better performance on BRATS 2015.
Moreover, the dice scores of our SegAN outperform the U-net baseline for segmenting all three types of tumor regions. Another observation is that our SegAN-produced label maps are smooth with little noise. Figure~\ref{fig:examples} illustrates some example results of our SegAN; in the figure, the segmented regions of the three classes (whole tumor, tumor core, and Gd-enhanced tumor core) are shown in yellow, blue, and red, respectively. One possible reason behind this phenomenon is that the proposed multi-scale $L_1$ loss from our adversarial critic network encourages the segmentor to learn both global and local features that capture long- and short-range spatial relations between pixels, resulting fewer noises and smoother results.
\section{Discussion}
To the best of our knowledge, our proposed SegAN is the first GAN-inspired framework adapted specifically for the segmentation task that produces superior segmentation accuracy.
While conventional GANs have been successfully applied to many unsupervised learning tasks (e.g., image synthesis~\cite{zhang2016stackgan}) and semi-supervised classification~\cite{salimans2016improved}, there are very few works that apply adversarial learning to semantic segmentation. One such work that we found by Luc \emph{et al.}~\cite{luc2016semantic} used both the conventional adversarial loss of GAN and pixel-wise softmax loss against ground truth.
They showed small but consistent gains on both the Stanford Background dataset and the PASCAL VOC 2012 dataset; the authors observed that pre-training only the adversarial network was unstable and suggested an alternating scheme for updating the segmenting networkβs and the adversarial networkβs weights.
We believe that the main reason contributing to the unstable training of their framework is: the conventional adversarial loss is based on a single scalar output by the discriminator that classifies a whole input image into real or fake category.
When inputs to the discriminator are generated \emph{vs.} ground truth dense pixel-wise label maps as in the segmentation task, the real/fake classification task is too easy for the discriminator and a trivial solution is found quickly. As a result, no sufficient gradients can flow through the discriminator to improve the training of generator.
In comparison, our SegAN uses a multi-scale feature loss that measures the difference between generated segmentation and ground truth segmentation at multiple layers in the critic, forcing both the segmentor and critic to learn hierarchical features that capture long- and short-range spatial relationships between pixels. Using the same loss function for both $S$ and $C$, the training of SegAN is end-to-end and stable.
\section{Conclusions}
In this paper, we propose a novel end-to-end Adversarial Network architecture, namely SegAN, with a new multi-scale loss for semantic segmentation. Experimental evaluation on the BRATS brain tumor segmentation dataset shows that the proposed multi-scale loss in an adversarial training framework is very effective and leads to more superior performance when compared with single-scale loss or the conventional pixel-wise softmax loss.
As a general framework, our SegAN is not limited to medical image segmentation applications. In our future work, we plan to investigate the potential of SegAN for general semantic segmentation tasks.
\section{Acknowledgements}
This research was supported in part by the Intramural Research Program of the National Institutes of Health (NIH), National Library of Medicine (NLM), and Lister Hill National Center for Biomedical Communications (LHNCBC), under Contract HHSN276201500692P.
\bibliographystyle{spbasic} | ee3a14370bed8e3d8150f22e61e4d3ec832e01f8 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
\input{text-16p/intro}
\section{Problem Formulation}
\label{sec:pf}
\input{text-16p/pf2}
\section{Concurrent Estimation and Sensor Planning}
\label{sec:sol}
\input{text-16p/sol2}
\section{Main Results}
\label{sec:assume}
\input{text-16p/assume2}
\section{Convergence of Algorithm~\ref{alg:filter+opt}}
\label{sec:converge}
\input{text-16p/converge2}
\section{Numerical Experiments}
\label{sec:sims}
\input{text-16p/sims2}
\section{Conclusion}
\label{sec:conc}
\input{text-16p/conc}
\appendices
\section{Proof of Lemmas~\ref{lem:summable_constraints}~and~\ref{lem:summable_grads}}
\label{app:summability}
\input{text-16p/proofs1-2}
\section{Proof of Lemmas~\ref{lem:basic}-\ref{lem:disagree}}
\label{app:lemmaproofs2}
\input{text-16p/proofs2}
\bibliographystyle{IEEEtran}
\subsection{Bounds}
The norms of $\nu_{s,k}$ and $\boldsymbol{\delta}_{s,k}$ are bounded, i.e., for some scalars $D, N > 0$, there holds for all $k \in \mbN$ and $s \in \ccalS$ \textit{a.s.}:
$
|\nu_{s,k}|\le N$ and $\|\boldsymbol{\delta}_{s,k}\| \le D.
$
We calculate the exact values of $D$ and $N$ in Corollaries~\ref{cor:N} and \ref{cor:D}, respectively.
The set $\ccalX_0 = \Gamma \times \ccalU$ is convex and compact.
Therefore, there must exist a constant $C_{\mathbf{z}} > 0$ such that for any $\mathbf{z}_1, \mathbf{z}_2 \in \mathcal{X}_0$
\begin{equation}\label{eqn:C_z}
\|\mathbf{z}_1 -\mathbf{z}_2\| \le C_{\mathbf{z}}.
\end{equation}
Also, note that the functions $f(\cdot)$ and $h(\mathbf{z};\ccalD,i)$ for $i \in \ccalI$ are convex (not necessarily differentiable) over some open set that contains $\mathcal{X}_0$.
A direct consequence of this and the compactness of $\mathcal{X}_0$ is that the subdifferentials $\partial f (\mathbf{z})$
and $\partial h_+ (\mathbf{z};\ccalD,i)$ are nonempty over $\mathcal{X}_0$,
where note that here we use the symbol $\partial$ to refer to the subdifferential set and not a partial derivative.
It also implies that
the subgradients $f'(\mathbf{z}) \in \partial f (\mathbf{z})$ and $h_+'(\mathbf{z};\ccalD,i) \in \partial h_+ (\mathbf{z};\ccalD,i)$ are uniformly bounded over the set $\mathcal{X}_0$. That is, there is a scalar $L_f$ such that for all $f'(\mathbf{z}) \in \partial f (\mathbf{z})$ and $\mathbf{z}\in \mathcal{X}_0$,
\begin{subequations}
\begin{equation}\label{eqn:L_f1}
\|f'(\mathbf{z})\| \le L_f,
\end{equation}
and for any $\mathbf{z}_1, \mathbf{z}_2 \in \mathcal{X}_0$,
\begin{equation}\label{eqn:L_f2}
|f(\mathbf{z}_1)-f(\mathbf{z}_2)| \le L_f \|\mathbf{z}_1-\mathbf{z}_2\|.
\end{equation}
\end{subequations}
Note that for the objective function $f$ in \eqref{eq:distributed}, we have $L_f = 1.$
Also, there is a scalar $L_h$ such that for all $h_+'(\mathbf{z};\ccalD,i)\in \partial h_+ (\mathbf{z};\ccalD,i)$, $\mathbf{z} \in \mathcal{X}_0$, $i \in \ccalI$,
\begin{equation}\label{e:Lh}
\|h_+'(\mathbf{z};\ccalD,i)\| \le L_h.
\end{equation}
We calculate the exact value of $L_h$ in Corollary~\ref{cor:D}.
\begin{comment}
Similarly, we know the function $h(\mathbf{z};\ccalD,i)$, for each $i \in \ccalI$, is defined and convex in $\mathbf{z}$ (not necessarily differentiable) over some open set that contains $\mathcal{X}_0$.
A direct consequence is that the subdifferentials $\partial h_+ (\mathbf{z};\ccalD,i)$ are nonempty over $\mathcal{X}_0$. It also implies that
the subgradients $h_+'(\mathbf{z};\ccalD,i) \in \partial h_+ (\mathbf{z};\ccalD,i)$ are uniformly bounded over the set $\mathcal{X}_0$. That is, there is a scalar $L_h$ such that for all $h_+'(\mathbf{z};\ccalD,i)\in \partial h_+ (\mathbf{z};\ccalD,i)$, $\mathbf{z} \in \mathcal{X}_0$, $i \in \ccalI$,
\[
\|h_+'(\mathbf{z};\ccalD,i)\| \le L_h.
\]
We calculate the exact value of $L_h$ in Corollary~\ref{cor:D}.
\end{comment}
From relation \eqref{eq:differror} we have for any $i\in \ccalI$, $s \in \ccalS$, $k \in \mbN$, and $\mathbf{z}\in \mathcal{X}_0$
\begin{subequations}
\begin{equation}\label{eqn:L_h1}
\|h_+'(\mathbf{z};\ccalD_{s,k},i)\| = \|h_+'(\mathbf{z};\ccalD,i) + \bbdelta_{s,k}\|\le L_h +D,
\end{equation}
and for any $\mathbf{z}_1, \mathbf{z}_2 \in \mathcal{X}_0$
\begin{equation}\label{eqn:L_h2}
|h_+(\mathbf{z}_1;\ccalD_{s,k},i) \! - \! h_+(\mathbf{z}_2;\ccalD_{s,k},i)| \le (L_h+D) \|\mathbf{z}_1 \! - \! \mathbf{z}_2\|.
\end{equation}
\end{subequations}
The compactness of $\mathcal{X}_0$, the boundedness of the data sequences $\ccalD_{s,k}$, and the continuity of the functions $h_+(\mathbf{z};\ccalD,i)$ for $i \in \ccalI$ imply that there exist constants $C_h >0$ such that for any $s \in \ccalS$, $k \in \mbN$, $i \in \ccalI$ and $\mathbf{z} \in \mathcal{X}_0$
\begin{subequations}
\begin{equation}\label{eqn:C_h1}
|h_+(\mathbf{z};\ccalD,i)| \le C_h,
\end{equation}
\begin{equation}\label{eqn:C_h2}
|h_+(\mathbf{z};\ccalD_{s,k},i)| = |h_+(\mathbf{z};\ccalD,i)+\nu_{s,k}| \le C_h+N,
\end{equation}
\end{subequations}
where the second relation follows from \eqref{eq:constrainterror}.
Furthermore, when $h_+(\mathbf{z};\ccalD_{s,k},i) \neq 0$, we have $h_+'(\mathbf{z};\ccalD_{s,k},i) \neq \bb0$. Therefore, there must exist a constant $c_h >0$ such that
\begin{align}\label{eqn:dhbnd}
\|h_+'(\mathbf{z};\ccalD_{s,k},i)\| \ge c_h
\end{align}
for all $\mathbf{z} \in \mathcal{X}_0$, $s \in \ccalS$, $k \in \mbN$, and $i \in \ccalI$.
This and relation (\ref{eqn:C_h2}) imply that
\begin{align}\label{eqn:betabnd}
\beta_{s,k} = \frac{h_+(\mathbf{z};\ccalD_{s,k},\omega_{s,k})}{\|h'_+(\mathbf{z};\ccalD_{s,k},\omega_{s,k})\|^2} \le \frac{C_h+N}{c_h^2}.
\end{align}
Note that when $h_+(\mathbf{z};\ccalD_{s,k},\omega_{s,k}) = 0$, the above bound holds trivially.
\subsection{Assumptions}
In the preceding sections, we have made extensive use of the function $Q\colon \reals^q \times \reals^p \to {\rm Sym}_{+}(p,\reals).$
For our algorithm to converge, we require the following to hold true regarding this information model.
\begin{assumption}
\label{assume:Q}
We assume that the information function $Q$
\begin{enumerate}[(a)]
\item is bounded,\label{a:bd}
\item is twice subdifferentiable\label{a:sub}
\item has bounded subdifferentials up to the second order, and\label{a:bd_sub}
\item has relatively few critical points, i.e., the sets of critical points of $Q$ and its partial derivatives up to the second order are measure zero. \label{a:crit}
\end{enumerate}
Denote the bound on the magnitude of $Q$ by $\eta_0$, the bound on its first derivatives by $\eta_1$, and the bound on its second derivatives by $\eta_2.$
\end{assumption}
The technical restrictions on the information model $Q$ are not overly restrictive in practice.
In particular, items~\ref{a:bd}, \ref{a:sub}, and \ref{a:bd_sub} imply that one cannot obtain infinite information and that, by changing sensor configuration, the information rate (and the rate of change of the rate) cannot change infinitely quickly.
Item~\ref{a:crit} essentially allows us to distinguish between sensor configurations, i.e., the set of configurations that offer ``optimal'' information rates are relatively sparse with respect to the space of signal source positions and sensor configurations $\reals^q \times \reals^p.$
The next two assumptions are related to the random variables $\omega_{s,k}$.
At each iteration $k$ of the inner-loop, recall that each sensor $s$ randomly generates $\omega_{s,k} \in \ccalI$. We assume that they are \textit{i.i.d.} samples from some probability distribution on $\ccalI$.
\begin{assumption}\label{assume:Pr}
In the $k$-th iteration of the inner-loop,
each element $i$ of $\ccalI$ is generated with nonzero probability,
i.e., for any $s \in \ccalS$ and $k \in \mbN$ it holds that
$
\pi_i \triangleq \textrm{Pr}\{ \omega_{s,k} = i\} > 0.
$
\end{assumption}
Assumption~\ref{assume:Pr} is easy to satisfy because $\ccalI$ is a finite set.
\begin{assumption}\label{assume:c}
For all $s \in \ccalS$ and $k \in \mbN$,
there exists a constant $\kappa> 0$ such that for all $\mathbf{z} \in \mathcal{X}_0$
$
\mathrm{dist}^2 (\mathbf{z},\mathcal{X}) \le \kappa \mathsf{E}\left[h_+^2(\mathbf{z};\ccalD_{s,k},\omega_{s,k})\right].
$
\end{assumption}
\noindent The upper bound in Assumption \ref{assume:c} is known as \textit{global error bound} and is crucial for the convergence analysis of Algorithm \ref{alg:filter+opt}.
Sufficient conditions for this bound have been shown in \cite{pang-book,Lewis96errorbounds,Bauschke:1996,Gubin19671}, which includes the case when each function $h(\cdot;\cdot,i)$ is linear in $\bbz$,
or when the feasible set $\mathcal{X}$ has a nonempty interior.
\begin{assumption}\label{assume:network}
For all $k \in \mbN$, the weighted graphs $\mathcal{G}_k = (\ccalS,\mathcal{E}_k,W_k)$ satisfy:
\begin{enumerate}[(a)]
\item There exists a scalar $\eta \in (0,1)$ such that $[W_k]_{sj} \ge \eta$ if $j \in \mathcal{N}_{s,k}$. Otherwise, $[W_k]_{sj} = 0$.
\item The weight matrix $W_k$ is doubly stochastic, i.e., $\sum_{s\in\ccalS}[W_k]_{sj}=1$ for all $j \in \ccalS$ and $\sum_{j\in\ccalS}[W_k]_{sj}=1$ for all $s \in \ccalS$.
\item There exists a scalar $Q > 0$ such that the graph $\left(\ccalS,\cup_{\ell=0,\ldots,Q-1} \mathcal{E}_{t+\ell}\right)$ is strongly connected for any $t \ge 1$.
\end{enumerate}
\end{assumption}
\noindent This assumption ensures a balanced communication between sensors. It also ensures that there exists a path from one sensor to every other sensor infinitely often even if the underlying graph topology is time-varying.
\subsection{Main result}
Our main result shows the almost sure convergence of Algorithm~\ref{alg:filter+opt}.
Specifically, the result states that the sensors asymptotically reach an agreement to a random point $\bbz^*$ which is in the optimal set $\ccalX^*$ almost surely (a.s.), as given in the following theorem.
\begin{theorem}[Asymptotic Convergence Under Noise]
\label{prop:main}
Let Assumptions
\ref{assume:Q} - \ref{assume:network}
hold.
Let also the stepsizes be such that $\sum_{k=1}^{\infty} \alpha_k = \infty$ and $\sum_{k=1}^{\infty} \alpha_k^2 < \infty$.
Then, the iterates $\{\bbz_{s,k}\}$ generated by each agent $s \in \ccalS$ via Algorithm~\ref{alg:filter+opt}
converge almost surely to the same point in the optimal set $\ccalX^*$, i.e., for a random point $\bbz^*\in \ccalX^*$
$
\lim_{k\to\infty} \bbz_{s,k} = \bbz^*, \quad \forall s \in \ccalS \quad a.s.
$
\end{theorem}
\subsection{Preliminary Results}
\begin{lemma}[Non-expansiveness \cite{BNO}]
\label{lem:proj}
Let $\mathcal{X} \subseteq \mathbb{R}^d$ be a nonempty closed convex set.
The function $\mathsf{\Pi}_{\mathcal{X}}: \mathbb{R}^d \rightarrow \mathcal{X}$ is nonexpansive, i.e.,
$
\|\mathsf{\Pi}_{\mathcal{X}}[\bba]-\mathsf{\Pi}_{\mathcal{X}}[\bbb]\| \leq \|\bba - \bbb\|
$
for all
$
\bba , \bbb \in \mathbb{R}^d.
$
\end{lemma}
\begin{lemma}[\!\!{\cite[Lemma 10-11, p. 49-50]{polyak}}]
\label{thm:super}
Let
$
\mathcal{F}_k \triangleq \set{v_\ell, u_\ell, a_\ell, b_\ell}_{\ell=0}^k
$
denote a collection of nonnegative real random variables for $k \in \mbN \cup \set{\infty}$ such that
$
\mathsf{E}[v_{k+1}|\mathcal{F}_k] \leq (1+a_k)v_k - u_k + b_k
$
for all
$
k \in \set{0}\cup \mbN \quad a.s.
$
Assume further that
$
\set{a_k}
$
and
$
\set{b_k}
$
are \textit{a.s.} summable.
Then, we have \textit{a.s.} that (i)
$
\set{u_k}
$
is summable and (ii) there exists a nonnegative random variable
$
v
$
such that
$
\set{v_k} \to v.
$
\end{lemma}
In the next lemma, we relate the two iterates $\mathbf{p}_{s,k}$ and $\mathbf{z}_{s,k}$ in Line \ref{line:p} and \ref{line:z} of Algorithm \ref{alg:filter+opt}.
In particular, we show a relation of $\mathbf{p}_{s,k}$ and $\mathbf{z}_{s,k-1}$
associated with any convex function $g$
which will be often used in the analysis. For example, $g(\mathbf{z}) = \|\mathbf{z}-\mathbf{a}\|^2$ for some $\mathbf{a}\in\mathbb{R}^n$ or
$g(\mathbf{z}) = \mathrm{dist}^2 (\mathbf{z},\mathcal{X})$.
\begin{lemma}[Convexity and Double Stochasticity]
\label{lem:ds}
For any convex function $g : \mathbb{R}^n \to \mathbb{R}$, we have
\[
\sum\nolimits_{s\in \ccalS} g(\mathbf{p}_{s,k}) \le \sum\nolimits_{s\in \ccalS} g(\mathbf{z}_{s,k-1})
\]
\end{lemma}
\begin{proof}
The double stochasticity of the weights plays a crucial role in this lemma. From the definition of $\mathbf{p}_{s,k}$ in Line \ref{line:p} of Algorithm \ref{alg:filter+opt}, we obtain
\begin{align*}
\sum_{s\in \ccalS} g(\mathbf{p}_{s,k}) \le~& \sum_{s\in \ccalS} \sum_{j\in \ccalS} [W_k]_{sj} g(\mathbf{z}_{j,k-1})\\
=~& \sum_{j\in \ccalS} \left(\sum_{s\in \ccalS} [W_k]_{sj}\right)g(\mathbf{z}_{j,k-1})= \sum_{j\in \ccalS} g(\mathbf{z}_{j,k-1}).
\end{align*}
\end{proof}
\begin{comment}
\begin{proof}
The doubly stochasticity of the weights plays a crucial role in this lemma. From the definition of $p_{i,k}$ in (\ref{eqn:algo1}),
\begin{align*}
\sum_{i\in \mathcal{V}} h(p_{i,k}) \le~& \sum_{i\in \mathcal{V}} \sum_{j\in \mathcal{V}} [W_k]_{ij} h(x_{j,k-1})\\
=~& \sum_{j\in \mathcal{V}} \left(\sum_{i\in \mathcal{V}} [W_k]_{ij}\right)h(x_{j,k-1})\\
=~& \sum_{j\in \mathcal{V}} h(x_{j,k-1}). \quad \square
\end{align*}
\end{proof}
\end{comment}
Lastly, for the convergence proof of our algorithm, we use a result from \cite{Ram2012}, which ensures successful consensus in the presence of a well behaved disturbance sequence.
\begin{lemma}[Perturbed Consensus]
\label{lemma:ram}
Let Assumption \ref{assume:network} hold.
Consider the iterates generated by
\begin{equation}\label{eqn:rule}
\boldsymbol{\theta}_{s,k} \hspace{-0.5mm}= \hspace{-0.5mm}\sum_{v\in\ccalS} [W_k]_{sv} \boldsymbol{\theta}_{v,k-1} + \mathbf{e}_{s,k},~ \forall s \in \ccalS,
\end{equation}
Suppose there exists a nonnegative nonincreasing scalar sequence $\{\alpha_k\}$ such that
$
\sum_{k=1}^{\infty} \alpha_k \|\mathbf{e}_{s,k}\|< \infty \text{ for all } s \in \ccalS.
$
Then, for all $s,v \in \ccalS$,
$
\sum_{k=1}^{\infty} \alpha_k \|\boldsymbol{\theta}_{s,k} -\boldsymbol{\theta}_{v,k}\| < \infty.
$
\end{lemma}
In addition to the well-known results of Lemmas~\ref{lem:proj}-\ref{lemma:ram}, we need the following intermediate results before the main result in Theorem~\ref{prop:main} can be proven.
The proofs of all of these Lemmas are deferred to Appendices~\ref{app:summability} and \ref{app:lemmaproofs2}.
In the first two of them, we posit that the two error sequences $\{\nu_{s,k}\}$
and $\{\bbdelta_{s,k}\}$ are summable.
\begin{lemma}[Summable Constraint Errors]
\label{lem:summable_constraints}
For almost every bounded sequence $\set{\bbz_{s,k}},$
the error in the constraint $\{|\nu_{s,k}|\}_{k \in \mbN}$ is summable a.s., i.e.,
$
\sum_{k=0}^\infty \abs{\nu_{s,k} }<\infty.
$
\end{lemma}
\begin{lemma}[Summable Constraint Violation Errors]
\label{lem:summable_grads}
For almost every bounded sequence $\set{\bbz_{s,k}},$
the error in the gradient of the constraint $\{\|\bbdelta_{s,k}\|\}_{k \in \mbN}$ is summable a.s., i.e.,
$
\sum_{k=0}^\infty \norm{\bbdelta_{s,k} }<\infty.
$
\end{lemma}
In the next lemma, we relate the two iterates $\mathbf{p}_{s,k}$ and $\mathbf{z}_{s,k}$ in Line \ref{line:p} and \ref{line:z} of Algorithm \ref{alg:filter+opt}.
\begin{lemma}[Basic Iterate Relation]
\label{lem:basic}
Consider the sequences $\{\mathbf{z}_{s,k}\}_{k =0}^\infty$ and $\{\mathbf{p}_{s,k}\}_{k =0}^\infty$ for $s \in \ccalS$ generated by Algorithm \ref{alg:filter+opt}. Then, for any $\mathbf{z},\check{\mathbf{z}} \in \mathcal{X}$, $s \in \ccalS$ and $k \in \mbN$, we have \textit{a.s.}:
\begin{align*}
&\|\mathbf{z}_{s,k}-\mathbf{z}\|^2
\le \|\mathbf{p}_{s,k}-\mathbf{z}\|^2 - 2\alpha_k(f(\check{\mathbf{z}})-f(\mathbf{z}))\\
&~+ 2\frac{C_{\mathbf{z}}(C_h+N)}{c_h^2}\|\boldsymbol{\delta}_{s,k}\| +2\frac{C_h+N}{c_h^2}|\nu_{s,k}|\\
&~ + \frac{1}{4\eta}\|\mathbf{p}_{s,k} - \check{\mathbf{z}}\|^2 -\frac{\tau-1}{\tau(L_h+D)^2}h_+^2(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\\
&~ + A_{\eta,\tau}\alpha_k^2L_f^2,
\end{align*}
where $A_{\eta,\tau} = 1+4\eta+\tau$ and $\eta,\tau>0$ are arbitrary.
\end{lemma}
Since we use a random approximate projection, we cannot guarantee the feasibility of the sequences $\mathbf{p}_{s,k}$ and $\mathbf{z}_{s,k}$ for every $k \in \mbN$ and $s \in \ccalS$. In the next lemma, we prove that $\mathbf{p}_{s,k}$ for all $s \in \ccalS$ asymptotically achieve feasibility even under the effect of the disturbances $\{\nu_{s,k}\}$
and $\{\bbdelta_{s,k}\}$.
\begin{lemma}[Asymptotic Feasibility Under Noise]
\label{lem:feas}
Let Assumption
\ref{assume:Pr} and
\ref{assume:c} hold.
Let $\alpha_k$ be square summable Consider the sequences $\{\mathbf{p}_{s,k}\}$ for $s \in \ccalS$ generated by Algorithm \ref{alg:filter+opt}.
Then, for any $s \in \ccalS$, the sequence
$
\set{ \mathrm{dist}(\mathbf{p}_{s,k},\mathcal{X}) }_{k \in \mbN}
$
is \textit{a.s.} square summable.
\end{lemma}
In the final Lemma, we show that $\|\mathbf{z}_{s,k}-\mathbf{p}_{s,k}\|$ eventually converges to zero for all $s \in \ccalS$.
This result combined with Lemma \ref{lem:feas} implies that the sequences $\set{ \mathbf{z}_{s,k} }_{k \in \mbN} $ also achieve asymptotic feasibility.
\begin{lemma}[Network Disagreement Under Noise]
\label{lem:disagree}
Let Assumptions
\ref{assume:Pr}-\ref{assume:network} hold. Let the sequence
$\{\alpha_k\}$ is such that $\sum_{k=1}^\infty \alpha_k^2<\infty$.
Define
$\mathbf{e}_{s,k}=\mathbf{z}_{s,k}-\mathbf{p}_{s,k}$ for all $s\in \ccalS$ and $k\ge 1$.
Then, we have \textit{a.s.}:
\begin{enumerate}[(a)]
\item
$\sum_{k=1}^\infty \|\mathbf{e}_{s,k}\|^2 <\infty \quad \text{for all } s \in \ccalS$,
\item
$\sum_{k=1}^\infty\alpha_k\|\check{\mathbf{z}}_{s,k}-\bar {\check{\mathbf{z}}}_k\|<\infty \quad \hbox{for all $s \in \ccalS$}$,
\end{enumerate}
where $\check{\mathbf{z}}_{s,k} = \mathsf{\Pi}_{\mathcal{X}}[\mathbf{p}_{s,k}]$ and $\bar {\check{\mathbf{z}}}_k = \frac{1}{n}\sum_{s \in \ccalS}\check{\mathbf{z}}_{s,k}$.
\end{lemma}
Note that the sequences $\{\mathbf{z}_{s,k}\}_{k \in \mbN}$ for $s \in \ccalS$ generated by Algorithm \ref{alg:filter+opt} can be
represented as relation \eqref{eqn:rule}. That is,
$
\mathbf{z}_{s,k}= \sum_{s\in\ccalS} [W_k]_{sj} \mathbf{z}_{j,k-1} + \mathbf{e}_{s,k}
$
for all $s \in \ccalS$
and $\mathbf{e}_{s,k} = \mathbf{z}_{s,k}-\mathbf{p}_{s,k}$.
Therefore, from Lemma \ref{lem:disagree}(a), we have
$
\sum_{k=1}^\infty \alpha_k\|\mathbf{e}_{s,k}\| = \frac{1}{2}\sum_{k=1}^\infty \alpha_k^2 + \frac{1}{2}\sum_{k=1}^\infty \|\mathbf{e}_{s,k}\|^2<\infty.
$
Invoking Lemma \ref{lemma:ram}, we see that there is consensus among $\mathbf{z}_{s,k}$ for $s \in \ccalS$.
\subsection{Proof of Theorem~\ref{prop:main}}
We invoke Lemma \ref{lem:basic} with $\check{\mathbf{z}} = \check{\mathbf{z}}_{s,k}$ (Note that $\check{\mathbf{z}}_{s,k}$ is defined as $\mathsf{\Pi}_{\mathcal{X}}[\mathbf{p}_{s,k}]$ in Lemma \ref{lem:disagree}), $\tau = 4$ and $\eta = \kappa(L_h+D)^2$.
We also let
$\mathbf{z} = \mathbf{z}^*$ for an arbitrary $\mathbf{z}^* \in \mathcal{X}^*$. Therefore, for any $\mathbf{z}^* \in \mathcal{X}^*$, $s \in \ccalS$ and $k \in \mbN$, we have \textit{a.s.}:
\begin{align}\label{eqn:pr1}
&\|\mathbf{z}_{s,k}-\mathbf{z}^*\|^2
\le \|\mathbf{p}_{s,k}-\mathbf{z}^*\|^2 - 2\alpha_k(f(\check{\mathbf{z}}_{s,k})-f(\mathbf{z}^*))\nonumber\\
&~+ 2\frac{C_{\mathbf{z}}(C_h+N)}{c_h^2}\|\boldsymbol{\delta}_{s,k}\| +2\frac{C_h+N}{c_h^2}|\nu_{s,k}|\nonumber\\
&~ + \frac{1}{4\kappa(L_h+D)^2}\mathrm{dist}^2(\mathbf{p}_{s,k},\mathcal{X})+A\alpha_k^2L_f^2\nonumber\\
&~-\frac{3}{4(L_h+D)^2}h^2_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k}),
\end{align}
where $A = 5+4\kappa(L_h+D)^2$. From Assumption \ref{assume:c}, we know that
$
\mathrm{dist}^2(\mathbf{p}_{s,k},\mathcal{X}) \le \kappa \mathsf{E}\left[h^2_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\mid \mathcal{F}_{k-1}\right].
$
Denote by $\mathcal{F}_k$ the sigma-field induced by the history of the algorithm up to time $k$, i.e.,
$
\mathcal{F}_k = \{\mathbf{z}_{s,0},~(\omega_{s,\ell},~ 1 \le \ell \le k),~ s\in \ccalS\} \quad \text{for } k \in \mbN,
$
and $\mathcal{F}_0 = \{\mathbf{z}_{s,0},~ s\in \ccalS\}$.
Taking the expectation conditioned on $\mathcal{F}_{k-1}$ in relation (\ref{eqn:pr1}), summing this over $s \in \ccalS$, and using the above relation, we obtain
\begin{align}\label{eqn:pr2}
&\sum_{s\in\ccalS}\mathsf{E}\left[\|\mathbf{z}_{s,k}-\mathbf{z}^*\|^2\mid \mathcal{F}_{k-1}\right]\\
&~\le \sum_{s\in\ccalS}\|\mathbf{z}_{s,k-1}-\mathbf{z}^*\|^2 - 2\alpha_k\sum_{s\in \ccalS}(f(\check{\mathbf{z}}_{s,k})-f(\mathbf{z}^*))\nonumber\\
&~~+ 2\frac{C_{\mathbf{z}}(C_h+N)}{c_h^2}\sum_{s\in\ccalS}\mathsf{E}[\|\boldsymbol{\delta}_{s,k}\|\mid \mathcal{F}_{k-1}] \nonumber\\ &~~+2\frac{C_h+N}{c_h^2}\sum_{s\in\ccalS}\mathsf{E}[|\nu_{s,k}|\mid \mathcal{F}_{k-1}]\nonumber\\
&~ -\frac{1}{2\kappa_2(L_h+D)^2}\sum_{s\in\ccalS}\mathrm{dist}^2(\mathbf{p}_{s,k},\mathcal{X}) + A\alpha_k^2nL_f^2,\nonumber
\end{align}
where we used Lemma \ref{lem:ds} for the first term on the right-hand side.
Recall that $f(\mathbf{z})=\sum_{s\in\ccalS} f(\mathbf{z})$ and $\bar{\check{\mathbf{z}}}_k \triangleq \frac{1}{n}\sum_{\ell \in \ccalS} \check{\mathbf{z}}_{\ell,k}$.
Using $\bar{\check{\mathbf{z}}}_k$ and $f$,
we can rewrite the term $\sum_{s\in\ccalS}(f(\check{\mathbf{z}}_{s,k})-f(\mathbf{z}^*))$ as follows:
\begin{subequations}
\begin{align}
\sum_{s\in\ccalS} &(f(\check{\mathbf{z}}_{s,k})-f(\mathbf{z}^*)) \nonumber\\
&~= \sum_{s\in\ccalS} (f(\check{\mathbf{z}}_{s,k})-f(\bar{\check{\mathbf{z}}}_k)) + (f(\bar{\check{\mathbf{z}}}_k)-f^*)\label{eqn:rewrite1}\\
&~\le \sum_{s\in\ccalS} \langle f'(\bar{\check{\mathbf{z}}}_k),\bar{\check{\mathbf{z}}}_k-\check{\mathbf{z}}_{s,k}\rangle+ (f(\bar{\check{\mathbf{z}}}_k)-f^*)\label{eqn:rewrite2}\\
&~ \le \sum_{s\in\ccalS} \|f'(\bar{\check{\mathbf{z}}}_k)\|\,\|\check{\mathbf{z}}_{s,k}-\bar{\check{\mathbf{z}}}_k\|+ (f(\bar{\check{\mathbf{z}}}_k)-f^*)\label{eqn:rewrite3}\\
&~ \le L_f\sum_{s\in\ccalS} \|\check{\mathbf{z}}_{s,k}-\bar{\check{\mathbf{z}}}_k\|+ (f(\bar{\check{\mathbf{z}}}_k)-f^*),\label{eqn:rewrite4}
\end{align}
\end{subequations}
where (\ref{eqn:rewrite1}) follows from adding and subtracting $f(\bar{\check{\mathbf{z}}}_k)$;
(\ref{eqn:rewrite2}) follows from the convexity of the function $f$;
(\ref{eqn:rewrite3}) follows from the Schwarz inequality;
and (\ref{eqn:rewrite4}) follows from relation \eqref{eqn:L_f1} and $\bar{\check{\mathbf{z}}}_k \in \mathcal{X} \subseteq \mathcal{X}_0$.
Combining (\ref{eqn:rewrite4}) with (\ref{eqn:pr2}), we obtain
\begin{align*}
&\sum_{s\in\ccalS}\mathsf{E}\left[\|\mathbf{z}_{s,k}-\mathbf{z}^*\|^2\mid \mathcal{F}_{k-1}\right]\\
&~\le \sum_{s\in\ccalS}\|\mathbf{z}_{s,k-1}-\mathbf{z}^*\|^2 - 2\alpha_k(f(\bar{\check{\mathbf{z}}}_k)-f^*)\\
&~~+ 2\frac{C_{\mathbf{z}}(C_h+N)}{c_h^2}\sum_{s\in\ccalS}\mathsf{E}[\|\boldsymbol{\delta}_{s,k}\|\mid \mathcal{F}_{k-1}] \\ &~~+2\frac{C_h+N}{c_h^2}\sum_{s\in\ccalS}\mathsf{E}[|\nu_{s,k}|\mid \mathcal{F}_{k-1}]\\
&~~ +L_f\sum_{s\in\ccalS} \alpha_k \|\check{\mathbf{z}}_{s,k}-\bar{\check{\mathbf{z}}}_k\| + A\alpha_k^2nL_f^2,
\end{align*}
where we omitted the negative term.
Since $\bar{\check{\mathbf{z}}}_k \in \mathcal{X}$, we have $f(\bar{\check{\mathbf{z}}}_k) - f^* \ge 0$.
Thus, under the assumption $\sum_{k=0}^{\infty} \alpha_k^2 < \infty$ and Lemma \ref{lem:disagree}(b), the above relation satisfies all the conditions of Lemma \ref{thm:super}. Using this, we have the following results.
\textit{Result 1:} For some $\bbz^* \in \ccalX^*$ and any $s \in \ccalS$, the sequence $\{\|\mathbf{z}_{s,k}-\mathbf{z}^*\|\}$ is
convergent \textit{a.s.}
\textit{Result 2:}
$
\sum_{k=1}^{\infty} \alpha_k(f(\bar{\check{\mathbf{z}}}_k)-f^*) < \infty \quad a.s.
$
As a direct consequence of \textit{Result 1},
we know that the sequences $\{\|\mathbf{p}_{s,k}-\mathbf{z}^*\|\}$ and $\{\|\check{\mathbf{z}}_{s,k}-\mathbf{z}^*\|\}$ are also
convergent \textit{a.s.}
(This is straightforward from Line \ref{line:p} of Algorithm \ref{alg:filter+opt} and Lemma \ref{lem:feas} with the knowledge that
$\mathrm{dist}^2(\mathbf{p}_{s,k},\mathcal{X}) = \|\mathbf{p}_{s,k}-\check{\mathbf{z}}_{s,k}\|^2$.)
Since $\|\bar{\check{\mathbf{z}}}_k-\mathbf{z}^*\| \le \frac{1}{N} \sum_{i\in \ccalS}\|\check{\mathbf{z}}_{s,k}-\mathbf{z}^*\|$,
we further know that the sequence $\{\|\bar{\check{\mathbf{z}}}_k-\mathbf{z}^*\|\}$ is also
convergent \textit{a.s.}
As a direct consequence of \textit{Result 2}, since $\alpha_k$ is not summable,
$
\liminf_{k\rightarrow \infty} (f(\bar{\check{\mathbf{z}}}_k)-f^*)=0 \; a.s.
$
From this relation and the continuity of $f$, it follows that
the sequence $\{\bar{\check{\mathbf{z}}}_k\}$ must have one accumulation point in the set $\mathcal{X}^*$ \textit{a.s.}
This and the fact that $\{\|\bar{\check{\mathbf{z}}}_k-\mathbf{z}^*\|\}$ is convergent \textit{a.s.} for every $\mathbf{z}^* \in \mathcal{X}^*$ imply that
\begin{equation}\label{eqn:z_final1}
\lim_{k \rightarrow \infty} \bar{\check{\mathbf{z}}}_k = \mathbf{z}^* \quad a.s.
\end{equation}
Also, by Lemma \ref{lem:disagree}(b), we have
\begin{align}\label{eqn:liminfv1}
\liminf_{k \rightarrow \infty} \|\check{\mathbf{z}}_{s,k}-\bar{\check{\mathbf{z}}}_k\| = 0 \quad \text{for all } s \in \ccalS \quad a.s.
\end{align}
The fact that $\{\|\check{\mathbf{z}}_{s,k}-\mathbf{z}^*\|\}$ is convergent \textit{a.s.} for all $s\in\ccalS$, together with~\eqref{eqn:z_final1}
and~\eqref{eqn:liminfv1} implies that
\begin{equation}\label{eqn:consensus1}
\lim_{k \rightarrow \infty} \check{\mathbf{z}}_{s,k} = \mathbf{z}^* \quad \text{for all } s \in \ccalS \quad a.s.
\end{equation}
Since $\|\mathbf{z}_{s,k}-\check{\mathbf{z}}_{s,k}\| \le \|\mathbf{z}_{s,k}-\mathbf{p}_{s,k}\| + \|\mathbf{p}_{s,k}-\check{\mathbf{z}}_{s,k}\|$, by invoking Lemma \ref{lem:feas} and \ref{lem:disagree}(a), we have
$
\lim_{k\to \infty} \|\mathbf{z}_{s,k}-\check{\mathbf{z}}_{s,k}\| = 0$ for all $ s \in \ccalS \quad a.s.
$
This and relation (\ref{eqn:consensus1}) give us the desired result, which is
$
\lim_{k \rightarrow \infty} \mathbf{z}_{s,k} = \mathbf{z}^\star$ for all $ s \in \ccalS \quad a.s.
$
\subsection{Sensor Model}\label{sec:mod}
We assume that the landmarks live in $\reals^2$ and denote the configuration of landmark $i$ by $\bbx_i.$
We also assume that the configuration space of the robot is $\reals^2$.
Generic models for the measurement covariance matrix for sparse landmark localization have been proposed \cite{chung06, simonetto11, jalalkamali12}.
These models have two important characteristics that apply to a wide variety of localization sensors:
\begin{enumerate*}[label=(\itshape\roman*\upshape)]
\item Measurement quality is inversely proportional to viewing distance, and
\item The direction of maximum uncertainty is the viewing direction, i.e., the sensors are more proficient at sensing bearing than range.
\end{enumerate*}
The general idea is to use the vector $\hbbx_{i}(t)-\bbr_s(t)$ to construct a diagonal matrix in a coordinate frame relative to the sensor, then rotate the matrix to a global coordinate frame so that it can be compared to other observations.
Following the accepted generic sensor models, we express $Q(\bbr_s, \hbbx_{i})$ by its eigenvalue decomposition
$
Q= R_{\phi_{i} } \Lambda_{i}R_{\phi_{i}}^\top,
$
where the argument $(\bbr_s, \hbbx_{i})$ of $Q$, $\Lambda_i$, and $R_{\phi_i}$ is implicit.
The angle $\phi_{i} \in \left(-\frac{\pi}{2},\frac{\pi}{2} \right]$ is defined so that the first column of $R_{\phi_{i}}$, which by convention is $[\cos \phi_{i} \; \sin \phi_{i}]^\top \! \! , \;$ is parallel to the subspace spanned by the vector $\hbbx_{i}-\bbr_{s}$.
This angle is given by
$
\phi_{i} = \tan^{-1} \left(
( [\hbbx_{i}]_2-[\bbr_{s}]_2)/([\hbbx_{i}]_1 - [\bbr_{s}]_1)
\right),
$
where the subscripts outside the brackets refer to the first and second coordinate of the vector representing the location of the landmark and sensor located at $\hbbx_{i}$ and $\bbr_s$.
The eigenvalue matrix is given by $\Lambda_{i} = \textrm{diag}(\lambda_{1,i}, \lambda_{2,i}),$ where
$
\lambda_{1,i} = c_0(/1 + c_1\norm{\hbbx_{i} -\bbr_{s}}_2 ) \and \lambda_{2,i} =\rho \lambda_{1,i} .
$
In the body frame of sensor $s$, the eigenvalues $\lambda_{1,i}$ and $\lambda_{2,i} $ control the shape of the confidence ellipses for individual measurements of the $i$-th target.
The parameter $c_0 >0$ represents the overall sensor quality and scales the whole region equally, $c_1 > 0$ controls the sensitivity to depth, and $\rho >1$ controls eccentricity of the confidence regions associated with the measurements.
Figure~\ref{fig:params} illustrates the measurement model for one sensor and two targets.
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{figures/meas.pdf}
\vspace{-1.0cm}
\caption{A sensor at $\bbr_s$ observes targets it believes to be at $\hbbx_i$ and $\hbbx_j$.
Possible confidence regions are shown in grey.
Dotted lines are parallel to the global $\hat{\imath}$ axis.
Red lines guide the eye from the sensor to targets
}
\vspace{-0.7cm}
\label{fig:params}
\end{center}
\end{figure}
Since $\tan^{-1}$ is not continuous at zero, we need to impose the following restriction so that Assumption~\ref{assume:Q} holds.
Essentially, we assume that, for the selected $\delta>0$, which recall has an affect on the robot's ability to translate in $\reals^2,$ we choose a $\tau_i$ such that
$
\tau_i < c_0 / (1 + c_1\delta).
$
Thus, if the robot is within $\delta$ of a target $i$, the constraint $h(\bbz, \ccalD, i)$ is trivially satisfied, and therefore we will never evaluate $Q$ or any of its derivatives in this region.
\subsection{Results}
\begin{figure}[!t]
\centering
\includegraphics[width=\columnwidth]{figures/traj_stills.pdf}
\vspace{-0.7cm}
\caption{Four sensors localize 100 landmarks.
Gray $\times$'s indicate landmarks with uncertainty above the threshold, and black $\circ$'s indicate localized landmarks.
Gray lines connecting the agents indicate the communication links.
}
\vspace{-0.7cm}
\label{fig:traj}
\end{figure}
Figure~\ref{fig:traj} shows an example of the resulting active sensing trajectories for a network of four sensors that cooperatively localize a uniformly random distribution of 100 landmarks in a square $100 \,{{\rm m}}^2$ workspace to a desired accuracy of $1\, {{\rm m}}$.
This corresponds to an eigenvalue tolerance of $\tau_1=\cdots=\tau_m=9\,{{\rm m}}^{-2}.$
The prior distribution on the landmarks was uninformative, i.e., the initial mean estimate and information matrix are given by the initial observations $\set{\bby_{i,s} (0)\mid i \in \ccalI, s \in \ccalS}$ and the corresponding information matrices were set to $\set{Q(\bbr_s (0), \bby_{i,s} (0))}_{ i \in \ccalI, s \in \ccalS}.$
The parameters of the sensing model $Q$ were set to $c_0 = \frac{1}{2}, c_1 = 10, \and \rho = 30$.
The feasible control set $\ccalU$ was set to
$
\ccalU = \set{\bbu_s \in \reals^2 \mid \norm{\bbu_s}\leq 1\,{{\rm m}}}.
$
In the figure, note that between $t=194$ and $t=242$, the agent in the bottom right moves north in order to maintain connectivity while the other agents move to finish the localization task.
It can be seen
from the four snapshots of the trajectories in Fig.~\ref{fig:traj}, that the robots effectively ``divide and conquer'' the task based on proximity to unlocalized targets.
\begin{figure}[!t]
\centerin
\includegraphics[width=\columnwidth]{code/paper_plots.pdf}
\vspace{-0.7cm}
\caption{
Top panel:
Evolution of the minimum eigenvalues.
The center curve is the mean among all 100 targets, with the grey block extending to the lowest and largest eigenvalues at each time instant.
The dotted line is at the tolerance $\tau=9\,{{\rm m}}^{-2}$.
Bottom panel:
Evolution of localization error.
The center curve is the mean among all 100 targets, with the grey block extending to the minimum and maximum errors at each time instant.
The dotted line is at the desired accuracy of $1\,{{\rm m}}$.
Note that the scale on vertical axis is large because the initial uncertainty is infinite.
}
\vspace{-0.3cm}
\label{fig:eigs_and_errs}
\end{figure}
Fig.~\ref{fig:eigs_and_errs} shows the evolution of the minimum eigenvalues of the information matrices $S_i(t)$ from the simulation shown in Fig.~\ref{fig:traj}.
Note that, even after the minimum eigenvalue associated with a landmark has surpassed $\tau$, more information is still collected, but it does not affect the control signal.
Note also that the maximal uncertainty in the bottom panel does not go below the desired threshold, even though the minimal eigenvalue does go above the threshold in the top panel.
This is expected because we use the 95\% confidence interval to generate the threshold $\tau$, thus we expect 5\% of landmarks to have errors larger than the desired threshold.
In this case study, we selected the termination criterion for Algorithm~\ref{alg:filter+opt} to be at most 30 iterations.
When testing the algorithm with $\alpha_k = 1/k$, we observed that the relative change in the iterates $\bbz_{s,k}$ and the data $\ccalD_{s,k}$ would often reach levels at or below 10\% inside of thirty iterations, so we allowed the message passing to terminate early if this criterion was met.
\begin{figure}[!t]
\centering
\includegraphics[width=0.9\columnwidth]{code/constraint-viols.eps}
\caption{
Plotting the amount of constraint violation of the original nonlinear constraint that will occur because we solve a linearized version.
Twenty lines are drawn, one for each landmark in a simulation with $m=20.$
The horizontal axis represents the outer loop iterations $t$.
Lines are drawn to guide the eye
}
\vspace{-0.7cm}
\label{f:viols}
\end{figure}
Figure~\ref{f:viols} measures the amount of constraint violation of the original nonlinear constraints for the solution obtained from the linearized version that we observed in a simulation with twenty landmarks.
For the $\ccalU$ and the parameters for the function $Q$ used in these simulations, the constraint violation was very small, and driven to zero typically for $t > 10.$
We tested larger feasible sets $\ccalU$ and (predictably) observed a general increase in the constraint violation, although we still observed reasonable robot behaviors.
\subsection{Distributed Information Filtering}
We solve problem~\eqref{eq:distributed} using an iterative approach.
Let $\ccalD_{s,k}$ denote the copy of the data set $\ccalD$ that is local to sensor $s$ at iteration $k \in \mbN$, where $\mbN = \set{1,2,3,\dots}$ denotes the natural numbers.
Additionally, let $\bbz_{s,k}$ denote the decision variables of the sensor $s$ at iteration $k$.
The first step of the algorithm is for sensor $s$ to broadcast $\bbz_{s,k-1} $ and a summary of $\ccalD_{s,k-1}.$
In Algorithm~\ref{alg:filter+opt}, the \emph{data summary for state} $i$ is captured by the intermediate information matrix $\bbXi_{i, s, k}$ and information vector $\bbxi_{i, s, k},$ defined in lines \ref{line:Xi} and \ref{line:xi}.
We refer the reader to \cite{kamal12} for a detailed discussion of Information Consensus Filtering, but we note briefly here that if the sensors each compute $\bbXi_{i,s,0}$ and $\bbxi_{i,s,0}$, and carry out the weighted averaging that is standard in all consensus algorithms, given in lines \ref{line:Xi_new} and \ref{line:xi_new}, then $n \bbXi_{i, s, k}$ and $\bbxi_{i, s, k}$ converge geometrically to the consensus posterior information.
In other words, under the ICF, we have that, for all $s \in \ccalS,$
$
\set{ \left( n\Xi_{1, s, k}, \dots, n\Xi_{m,s,k} \right)}_{k \in \mbN}\to S
$
for all
$i \in \ccalI$
and
$
\set{\big((n\Xi_{1,s,k})^{-1} \bbxi_{1, s, k}, \dots, (n\Xi_{m,s,k})^{-1} \bbxi_{m,s,k}\big)}_{k \in \mbN} \to \hbbx.
$
In particular, let $\ccalG_k = (\ccalS, \ccalE_k)$ represent a communication graph such that $(s,v) \in \ccalE_k$ if and only if sensors $s$ and $v$ can communicate at iteration $k$.
This defines the \emph{neighbor set},
$
\ccalN_{s,k} \triangleq \set{s} \cup \set{
v \in \ccalS \mid (v,s) \in \ccalE_k
}.
$
Sensor $s$ receives $\set{\bbz_{v,k-1}, \set{\bbxi_{i,v,k},\bbXi_{i,v,k}}_{i \in \ccalI} \mid v \in \ccalN_{s,k}}.$
With this new information, the sensor updates its own optimization variable and data summary via using weighted averaging.
In particular, let $W_k$ be an $n \times n$ row stochastic matrix such that
$
[W_k]_{sv} > 0 \iff v \in \ccalN_s.
$
The exact assumptions for $W_k$ will be introduced later, when they are necessary for the proof of our main result.
The consensus steps are depicted in lines~\ref{line:p}, \ref{line:Xi_new}, and \ref{line:xi_new} of Algorithm~\ref{alg:filter+opt}.
Then, the agent can compute $\ccalD_{s,k}$ as shown in line~\ref{line:D}, concluding the IF portion of iteration $k.$
Note that, given initial conditions $\set{ \ccalD_{s, 0}}_{s \in \ccalS},$ the data for any sensor $s$ at time $k$ is bounded by the initial conditions as
\begin{equation}\label{eq:bound_data}
\norm{ \ccalD_{s,k} } \leq
\max_{v \in \ccalS}
\norm{
\left( \ccalD_{v,0}
\right)
}
,
\end{equation}
where $\norm{\cdot}$ is any norm in the space of data $\reals^{pm} \times \left( {\rm Sym}_{++} (p, \reals) \right)^m.$
This implies that the set of possible data is bounded with respect to any norm in the vector space $\reals^{pm} \times\left( {\rm Sym} (p, \reals) \right)^m.$
Note that, by the same arguments as can be found in \cite{kamal12}, under any norm, all sensors' data converge to the same consensus value, i.e.,
$
\norm{ \ccalD_{s,k} - \ccalD_{v,k } }\to 0
$
and
$
\norm{ \ccalD_{s,k} - \ccalD }\to 0
$
for all $s , v \in \ccalS.$
\subsection{Random Approximate Projections}
The optimization part of the algorithm begins in line \ref{line:v} by computing a negative subgradient of the $f$ at the local iterate, which we denote by $- f'(\bbz_{s,k})$, taking a gradient descent step, and projecting the iterate back to the simple constraint set $\ccalX_0$.
The positive step size $\alpha_k$ is diminishing, non-summable and square-summable, i.e, $\alpha_k \to 0$, $\sum_{k \in \mbN}\alpha_k = \infty$ and $\sum_{k \in \mbN}\alpha_k^2 < \infty$.
The last step of our method is not a projection to
the feasible set, but instead a subgradient step in the direction
that minimizes the violation of a \textit{single} LMI constraint.
Essentially, each node selects one of
the LMIβs randomly, measures the violation of that constraint,
and takes an additional gradient descent step with step size
$\beta_{s,k}$ in order to minimize this violation.
We call this process \textit{random approximate projection}.
Specifically, line~\ref{line:o} randomly assigns $\omega_{s,k} \in \ccalI$, effectively choosing to activate $h\left(\cdot; \cdot, \omega_{s,k} \right)$.
The idea behind line~\ref{line:z} is to minimize a scalar metric that measures the amount of violation of $h\left(\cdot; \cdot, \omega_{s,k} \right)$, which we denote by $h_+\left(\cdot; \cdot, \omega_{s,k} \right),$ while remaining feasible.
Let $\mathsf{\Pi}_+ \colon {\rm Sym}(p,\reals) \to {\rm Sym}_+(p,\reals)$ denote the projection operator onto the set of positive semidefinite matrices, given by
\begin{equation}\label{eq:pi}
\mathsf{\Pi}_+X = E \textrm{diag}\left[ (\lambda_1)_+, \dots, (\lambda_p)_+\right] E^\top,
\end{equation}
where
$(\cdot)_+ \triangleq \max \set{\cdot, 0}$
and $\lambda_1, \dots, \lambda_p$ are the eigenvalues and $E$ is a matrix with columns $\bbe_1, \dots, \bbe_p$ that are the corresponding eigenvectors of $X$.
A measure of the ``positive definiteness'' of $X$ is thus
$
\norm{\mathsf{\Pi}_+ X }_F
\equiv \sqrt{ (\lambda_1)_+^2 + \cdots + (\lambda_p)_+^2}.
$
This defines our measure of constraint violation,
\begin{align}
h_+ \left(\bbz; \ccalD, i\right) &\triangleq \norm{ \mathsf{\Pi}_+ h \left(\bbz; \ccalD, i\right) }_F.
\end{align}
We use special notation to denote the vector of derivatives of $h_+$ with respect to the decision variables $\bbz$ evaluated at $(\bbz; \ccalD, i)$.
In particular, define the gradient
$
h'_+ \left(\bbz; \ccalD, i\right)= \left(\nabla_1 h_+ (\bbz; \ccalD,i), \dots , \nabla_{ m+qn} h_+ (\bbz ; \ccalD,i) \right)^\top
= \nabla_\bbz h_+(\bbz; \ccalD, i).
$
Defining the function
$
\psi_{ij} \colon {\rm Sym} (p, \reals) \times \reals^{m +qn} \times \left(\reals^{pm} \times \left( {\rm Sym}_{+} (p, \reals) \right)^m \right) \to \reals
$
for notational convenience, the $j$-th entry of $h'_+ \! \left(\bbz; \ccalD, i\right)$ is given by
\begin{equation}\label{eq:dh+_dzj}
\nabla_{j} h_+ (\bbz; \ccalD,i)
= \psi_{ij} \Big( \mathsf{\Pi}_+ \big( h\left(\bbz;\ccalD, i \right) \big), \bbz, \ccalD\Big),
\end{equation}
where
\begin{equation}\label{eq:psi}
\psi_{ij} (X,\bbz,\ccalD)\triangleq\begin{cases}
\frac{
{{\rm Tr}}
\left(
\nabla_j h (\bbz, \ccalD,i)
X
\right) }{\norm{X}_F}
&{{\rm if}} \; \norm{X}_F\! >0\\
d&{{\rm else}}
\end{cases}
\end{equation}
and $d >0$ is some arbitrary constant.
The function $\psi_{ij}$ is not continuous at $X=\bb0,$ and we set $d \neq 0$ for technical reasons discussed below.
The step size
\begin{equation}\label{eqn:beta}
\beta_{s,k} = \frac{h_+(\bbv_{s,k}; \ccalD_{s,k}, \omega_{s,k})}{\norm{h'_+(\bbv_{s,k}; \ccalD_{s,k}, \omega_{s,k})}^2}
\end{equation}
given in line \ref{line:b} is a variant of the Polyak step size \cite{polyak-sg}.
Note that this is a well defined step size as we let
$
h_+' (\bbv_{s,k}; \ccalD_{s,k}, \omega_{s,k}) = d {\bf 1}_{m+qn}
$
(cf. Eq \eqref{eq:psi})
whenever
$
h_+(\bbv_{s,k}; \ccalD_{s,k}, \omega_{s,k})=0
$
and
it is nonzero elsewhere by construction.
\subsection{Proof of the Lemmas in \ref{subsec:lem}}
\begin{proof}[Proof of Lemma \ref{lem:basic}]
From Line \ref{line:z} of Algorithm \ref{alg:filter+opt}, the following chain of relations hold for any feasible point $\mathbf{z} \in \mathcal{X} \subseteq \mathcal{X}_0$ \textit{a.s.}:
\begin{subequations}
\begin{align}
&\|\mathbf{z}_{s,k}-\mathbf{z}\|^2 \le \|\mathbf{v}_{s,k}-\mathbf{z} \! - \! \beta_{s,k}h'_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})\|^2\label{eqn:basica}\\
&~\le \|\mathbf{v}_{s,k}-\mathbf{z}\|^2 \! - \! 2\beta_{s,k} \langle h'_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k}), \mathbf{v}_{s,k}-\mathbf{z} \rangle \nonumber\\
&~+ \beta_{s,k}^2 \|h'_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})\|^2\label{eqn:basicb}\\
&~\le \|\mathbf{v}_{s,k}-\mathbf{z}\|^2 \! - \! 2\beta_{s,k} \langle h'_+(\mathbf{v}_{s,k};\ccalD,\omega_{s,k}), \mathbf{v}_{s,k} \! - \! \mathbf{z} \rangle \label{eqn:basicc} \\
&~~+\! 2\beta_{s,k} \|\boldsymbol{\delta}_{s,k}\|\|\mathbf{v}_{s,k} \! - \! \mathbf{z} \|
\! + \! \beta_{s,k}^2 \|h'_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})\|^2 \nonumber \\
&~\le \|\mathbf{v}_{s,k}-\mathbf{z}\|^2 \! - \! 2\beta_{s,k} h_+(\mathbf{v}_{s,k};\ccalD,\omega_{s,k})\label{eqn:basicd}\\
&~+ 2\beta_{s,k} \|\boldsymbol{\delta}_{s,k}\|\|\mathbf{v}_{s,k} \! - \! \mathbf{z} \|
+ \beta_{s,k}^2 \|h'_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})\|^2,\nonumber
\end{align}
\end{subequations}
where (\ref{eqn:basica}) follows from the nonexpansiveness of the projection operator; (\ref{eqn:basicb}) follows from the expansion of $\|\cdot\|^2$; (\ref{eqn:basicc}) follows from relation (\ref{eq:differror}) and the Schwarz inequality;
and (\ref{eqn:basicd}) follows from the convexity of the function $h_+(\cdot;\cdot,\omega_{s,k})$, and $h_+(\mathbf{z};\ccalD,\omega_{s,k})=0$ due to the feasibility of the point $\mathbf{z}$, i.e.,
$
h_+(\mathbf{v}_{s,k};\ccalD,\omega_{s,k})
\le \langle h'_+(\mathbf{v}_{s,k};\ccalD,\omega_{s,k}), \mathbf{v}_{s,k}-\mathbf{z} \rangle.
$
From relation (\ref{eq:constrainterror}) and the definition of $\beta_{s,k}$ in (\ref{eqn:beta}), we further have
\begin{align*}
&- 2\beta_{s,k} h_+(\mathbf{v}_{s,k};\ccalD,\omega_{s,k})+ \beta_{s,k}^2 \|h'_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})\|^2\\
&~= - 2\beta_{s,k} h_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k}) + 2\beta_{s,k}|\nu_{s,k}| \\
&~~+ \beta_{s,k}^2 \|h'_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})\|^2\\
&~= 2\beta_{s,k}|\nu_{s,k}| - \frac{h^2_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})}{\|h'_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})\|^2}\\
&~\le 2\beta_{s,k}|\nu_{s,k}| - \frac{h_+^2(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})}{(L_h+D)^2},
\end{align*}
where the last inequality follows from relation \eqref{eqn:L_h1}.
Combining this with inequality (\ref{eqn:basicd}), we obtain for any $\mathbf{z} \in \mathcal{X}$ \textit{a.s.}:
\begin{align}\label{eqn:basic1}
\|\mathbf{z}_{s,k}-\mathbf{z}\|^2
\le &~ \|\mathbf{v}_{s,k}-\mathbf{z}\|^2
+ 2\beta_{s,k} \|\boldsymbol{\delta}_{s,k}\|\|\mathbf{v}_{s,k}-\mathbf{z} \|\\
&~+2\beta_{s,k}|\nu_{s,k}|
- \frac{h_+^2(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})}{(L_h+D)^2}.\nonumber
\end{align}
For the last term on the right-hand side, we can rewrite
$
h_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k}) = h_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k}) -h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})
+h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k}).
$
Therefore,
\begin{align}\label{eqn:gfnpre}
&h_+^2(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})\ge 2h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\nonumber\\
&\hspace{2cm}\times\left(h_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k}) -h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\right)\nonumber\\
&\hspace{3cm} + h_+^2(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k}).
\end{align}
Regarding the first term on the right-hand side of (\ref{eqn:gfnpre}), the following chain of relations holds:
\begin{subequations}\label{eqn:basice}
\begin{align}
&2h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\nonumber\\
&~~\times\left(h_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k}) -h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\right)\nonumber\\
&~\ge-2h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\nonumber\\
&~~\times\left|h_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k}) -h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\right|\label{eqn:basicf}\\
&~\ge -2(L_h+D)\|\mathbf{v}_{s,k}-\mathbf{p}_{s,k}\|h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\label{eqn:basicg}\\
&~\ge -2\alpha_k (L_h+D)L_fh_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\label{eqn:basich}\\
&~\ge -\tau \alpha_k^2(L_h+D)^2L_f^2 - \frac{1}{\tau}h_+^2(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k}),\label{eqn:basici}
\end{align}
\end{subequations}
where (\ref{eqn:basicf}) follows from the Schwarz inequality; (\ref{eqn:basicg}) follows from relation (\ref{eqn:L_h2}); (\ref{eqn:basich}) follows from Line \ref{line:v} of Algorithm \ref{alg:filter+opt}; and (\ref{eqn:basici}) follows from using
$
2|a||b| \le \tau a^2 + \frac{1}{\tau} b^2
$
with $a=\alpha_k (L_h+D)L_f$ and $b=h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})$, and $\tau >0$ is arbitrary.
Using relations (\ref{eqn:gfnpre})-(\ref{eqn:basice}) in (\ref{eqn:basic1}), we obtain for all $\mathbf{z} \in \mathcal{X}$ \textit{a.s.}:
\begin{align}\label{eqn:basic2}
&\|\mathbf{z}_{s,k}-\mathbf{z}\|^2 \le \|\mathbf{v}_{s,k}-\mathbf{z}\|^2
+ 2\beta_{s,k} \|\boldsymbol{\delta}_{s,k}\|\|\mathbf{v}_{s,k}-\mathbf{z} \|\\
&~+2\beta_{s,k}|\nu_{s,k}|-\frac{\tau-1}{\tau (L_h+D)^2}h_+^2(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})
+\tau \alpha_k^2 L_f^2.\nonumber
\end{align}
Similarly to (\ref{eqn:basica})-(\ref{eqn:basicd}), from Line \ref{line:v} of Algorithm \ref{alg:filter+opt}, the following chain of relations hold for any $\mathbf{z}\in \mathcal{X} \subseteq \mathcal{X}_0$ \textit{a.s.}:
\begin{subequations}
\begin{align}
&\|\mathbf{v}_{s,k}-\mathbf{z} \|^2\le \|\mathbf{p}_{s,k}-\mathbf{z}-\alpha_kf'(\mathbf{p}_{s,k})\|^2\label{eqn:basicj}\\
&~\le \|\mathbf{p}_{s,k}-\mathbf{z}\|^2 - 2\alpha_k\langle f'(\mathbf{p}_{s,k}), \mathbf{p}_{s,k}-\mathbf{z}\rangle + \alpha_k^2\|f'(\mathbf{p}_{s,k})\|^2\label{eqn:basick}\\
&~\le \|\mathbf{p}_{s,k}-\mathbf{z}\|^2 - 2\alpha_k(f(\mathbf{p}_{s,k})-f(\mathbf{z})) + \alpha_k^2L_f^2,\label{eqn:basicm}
\end{align}
\end{subequations}
where (\ref{eqn:basicj}) follows from the nonexpansive projection;
(\ref{eqn:basick}) follows from the expansion of $\|\cdot\|^2$;
(\ref{eqn:basicm}) follows from the convexity of the function $f$ and relation \eqref{eqn:L_f1}.
For the second term on the right-hand side, we further have for any $\check{\mathbf{z}} \in \mathcal{X}$:
\begin{subequations}
\begin{align}
&- 2\alpha_k(f(\mathbf{p}_{s,k})-f(\mathbf{z})) \nonumber\\
&~= -2\alpha_k(f(\mathbf{p}_{s,k}) - f(\check{\mathbf{z}}) + f(\check{\mathbf{z}})-f(\mathbf{z}))\label{eqn:basicn}\\
&~\le - 2\alpha_k \langle f'(\check{\mathbf{z}}), \mathbf{p}_{s,k} - \check{\mathbf{z}}\rangle - 2\alpha_k(f(\check{\mathbf{z}})-f(\mathbf{z}))\label{eqn:basico}\\
&~\le 2\alpha_k L_f \|\mathbf{p}_{s,k} - \check{\mathbf{z}}\| - 2\alpha_k(f(\check{\mathbf{z}})-f(\mathbf{z}))\label{eqn:basicp}\\
&~\le 4\eta\alpha_k^2 L_f^2 + \frac{1}{4\eta}\|\mathbf{p}_{s,k} - \check{\mathbf{z}}\|^2 - 2\alpha_k(f(\check{\mathbf{z}})-f(\mathbf{z})),\label{eqn:basicq}
\end{align}
\end{subequations}
where (\ref{eqn:basicn}) follows from adding and subtracting $f(\check{\mathbf{z}})$;
(\ref{eqn:basico}) follows from the convexity of the function $f$;
(\ref{eqn:basicp}) follows from the Schwarz inequality;
and (\ref{eqn:basicq}) follows from using
$
2|a||b| \le 4\eta a^2 + \frac{1}{4\eta} b^2
$
with $a = \alpha_kL_f$ and $b = \|\mathbf{p}_{s,k} - \check{\mathbf{z}}\|$.
Combining (\ref{eqn:basicq}) in (\ref{eqn:basicm}), we have for any $\mathbf{z}, \check{\mathbf{z}} \in \mathcal{X} \subseteq \mathcal{X}_0$ \textit{a.s.}:
\begin{align*}
\|\mathbf{v}_{s,k}-\mathbf{z} \|^2
\le&~ \|\mathbf{p}_{s,k}-\mathbf{z}\|^2 - 2\alpha_k(f(\check{\mathbf{z}})-f(\mathbf{z})) \\
&~+ \frac{1}{4\eta}\|\mathbf{p}_{s,k} - \check{\mathbf{z}}\|^2 + (1+4\eta)\alpha_k^2L_f^2.
\end{align*}
Substituting this inequality in relation (\ref{eqn:basic2}), we obtain
\begin{align*}
&\|\mathbf{z}_{s,k}-\mathbf{z}\|^2
\le \|\mathbf{p}_{s,k}-\mathbf{z}\|^2 - 2\alpha_k(f(\check{\mathbf{z}})-f(\mathbf{z}))\\
&~+ 2\beta_{s,k} \|\boldsymbol{\delta}_{s,k}\|\|\mathbf{v}_{s,k}-\mathbf{z} \|+2\beta_{s,k}|\nu_{s,k}|\\
&~ + \frac{1}{4\eta}\|\mathbf{p}_{s,k} - \check{\mathbf{z}}\|^2 + A_{\eta,\tau}\alpha_k^2L_f^2\\
&~-\frac{\tau-1}{\tau(L_h+D)^2}h^2_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k}),
\end{align*}
where $A_{\eta,\tau} = 1+4\eta+\tau$.
From relation \eqref{eqn:C_z}, we have
$
\|\mathbf{v}_{s,k}-\mathbf{z} \|\le C_{\mathbf{z}}.
$
Using this and the upper estimate for $\beta_{s,k}$ in (\ref{eqn:betabnd}) for bounding the three error terms completes the proof.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:feas}]
We use Lemma~\ref{lem:basic} with $\check{\mathbf{z}}=\mathbf{z} = \mathsf{\Pi}_{\mathcal{X}}[\mathbf{p}_{s,k}]$.
Therefore, for any $s\in \ccalS$, and $k \ge 1$, we obtain \textit{a.s.}:
\begin{align}\label{eqn:lem2a}
&\|\mathbf{z}_{s,k}-\mathsf{\Pi}_{\mathcal{X}}[\mathbf{p}_{s,k}]\|^2
\le \|\mathbf{p}_{s,k}-\mathsf{\Pi}_{\mathcal{X}}[\mathbf{p}_{s,k}]\|^2 \\
&~+ 2\frac{C_{\mathbf{z}}(C_h+N)}{c_h^2}\|\boldsymbol{\delta}_{s,k}\| +2\frac{C_h+N}{c_h^2}|\nu_{s,k}|\nonumber\\
&~ + \frac{1}{4\eta}\|\mathbf{p}_{s,k} - \mathsf{\Pi}_{\mathcal{X}}[\mathbf{p}_{s,k}]\|^2 + A_{\eta,\tau}\alpha_k^2L_f^2\nonumber\\
&~ -\frac{\tau-1}{\tau(L_h+D)^2}h_+^2(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k}),\nonumber
\end{align}
where $A_{\eta,\tau} = 1+4\eta+\tau$ and $\eta,\tau > 0$ are arbitrary.
By the definition of the projection, we have
$
\mathrm{dist}(\mathbf{p}_{s,k},\mathcal{X}) = \|\mathbf{p}_{s,k} - \mathsf{\Pi}_{\mathcal{X}}[\mathbf{p}_{s,k}]\|$
and
$
\mathrm{dist}(\mathbf{z}_{s,k},\mathcal{X})
= \|\mathbf{z}_{s,k}-\mathsf{\Pi}_{\mathcal{X}}[\mathbf{z}_{s,k}]\| \le \|\mathbf{z}_{s,k}-\mathsf{\Pi}_{\mathcal{X}}[\mathbf{p}_{s,k}]\|.
$
Upon substituting these estimates in relation (\ref{eqn:lem2a}), we obtain
\begin{align}\label{eqn:lem2b}
&\mathrm{dist}^2(\mathbf{z}_{s,k},\mathcal{X})
\le \mathrm{dist}^2(\mathbf{p}_{s,k},\mathcal{X}) \\
&\hspace{2cm}+ 2\frac{C_{\mathbf{z}}(C_h+N)}{c_h^2}\|\boldsymbol{\delta}_{s,k}\| +2\frac{C_h+N}{c_h^2}|\nu_{s,k}|\nonumber\\
&\hspace{2cm} + \frac{1}{4\eta}\mathrm{dist}^2(\mathbf{p}_{s,k},\mathcal{X}) + A_{\eta,\tau}\alpha_k^2L_f^2\nonumber\\
&\hspace{2cm} -\frac{\tau-1}{\tau(L_h+D)^2}h_+^2(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k}).\nonumber
\end{align}
Taking the expectation conditioned on $\mathcal{F}_{k-1}$ and noting that $\mathbf{p}_{s,k}$ is fully determined by $\mathcal{F}_{k-1}$,
we have for any $s\in \ccalS$ and $k \ge 1$ \textit{a.s.}:
\begin{align}\label{eqn:lem2c}
&\mathsf{E}\left[\mathrm{dist}^2(\mathbf{z}_{s,k},\mathcal{X}) \mid \mathcal{F}_{k-1}\right]
\le \mathrm{dist}^2(\mathbf{p}_{s,k},\mathcal{X}) \\
&+ \hspace{-0.5mm}2\frac{C_{\mathbf{z}}(C_h+N)}{c_h^2}\mathsf{E}[\|\boldsymbol{\delta}_{s,k}\|\mid \mathcal{F}_{k-1}] \hspace{-0.5mm}+\hspace{-0.5mm}2\frac{C_h+N}{c_h^2}\mathsf{E}[|\nu_{s,k}|\mid \mathcal{F}_{k-1}]\nonumber\\
& + \frac{1}{4\eta}\mathrm{dist}^2(\mathbf{p}_{s,k},\mathcal{X})
+ A_{\eta,\tau}\alpha_k^2L_f^2\nonumber\\
&~ -\frac{\tau-1}{\tau(L_h+D)^2}\mathsf{E}\left[h^2_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\mid\mathcal{F}_{k-1}\right].\nonumber
\end{align}
We now choose $\tau =4$, $\eta = \kappa(L_h+D)^2$ and use Assumption \ref{assume:c} to yield
\begin{align*}
&\mathsf{E}\left[\mathrm{dist}^2(\mathbf{z}_{s,k},\mathcal{X}) \mid \mathcal{F}_{k-1}\right]
\le \mathrm{dist}^2(\mathbf{p}_{s,k},\mathcal{X}) \\
&+ \hspace{-0.5mm}2\frac{C_{\mathbf{z}}(C_h+N)}{c_h^2}\mathsf{E}[\|\boldsymbol{\delta}_{s,k}\| \mid \mathcal{F}_{k-1}] \hspace{-0.5mm}+2\hspace{-0.5mm}\frac{C_h+N}{c_h^2}\mathsf{E}[|\nu_{s,k}| \mid \mathcal{F}_{k-1}]\nonumber\\
& -\frac{1}{2\kappa(L_h+D)^2}\mathrm{dist}^2(\mathbf{p}_{s,k},\mathcal{X}) + A\alpha_k^2L_f^2,\nonumber
\end{align*}
where $A = 5+4\kappa(L_h+D)^2$.
Finally, by summing over all $s \in \ccalS$ and using Lemma \ref{lem:ds} with $h(x) = \mathrm{dist}^2(x,\mathcal{X})$,
we arrive at the following relation:
\begin{align} \label{eqn:lem2e}
&\sum_{s\in\ccalS}\mathsf{E}\left[\mathrm{dist}^2(\mathbf{z}_{s,k},\mathcal{X}) \mid \mathcal{F}_{k-1}\right]
\le \sum_{s\in\ccalS}\mathrm{dist}^2(\mathbf{z}_{s,k-1},\mathcal{X}) \\
&~+ 2\frac{C_{\mathbf{z}}(C_h+N)}{c_h^2}\sum_{s\in\ccalS}\mathsf{E}[\|\boldsymbol{\delta}_{s,k}\| \mid \mathcal{F}_{k-1}] \nonumber\\ &~+2\frac{C_h+N}{c_h^2}\sum_{s\in\ccalS}\mathsf{E}[|\nu_{s,k}| \mid \mathcal{F}_{k-1}]\nonumber\\
&~ -\frac{1}{2\kappa(L_h+D)^2}\sum_{s\in\ccalS}\mathrm{dist}^2(\mathbf{p}_{s,k},\mathcal{X}) + A\alpha_k^2nL_f^2.\nonumber
\end{align}
Therefore, from $\sum_{k\ge 0} \alpha_k^2 < \infty$ and Lemmas~\ref{lem:summable_constraints} and \ref{lem:summable_grads}, all the conditions in the convergence theorem (Theorem~\ref{thm:super}) are satisfied and the desired result follows.
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:disagree}(a)]
From Line \ref{line:p}-\ref{line:v} and \ref{line:z} of Algorithm \ref{alg:filter+opt}, we define $\mathbf{e}_{s,k} \triangleq \mathbf{z}_{s,k} - \mathbf{p}_{s,k}$ for $s\in \ccalS$ and $k \ge 0$. Hence, $\mathbf{e}_{s,k}$ can be viewed as the perturbation that we make on $\mathbf{p}_{s,k}$ after the network consensus step in Line \ref{line:p} of Algorithm \ref{alg:filter+opt}.
Consider $\|\mathbf{e}_{s,k}\|$, for which we can write
\begin{align} \label{eqn:lem3a}
\|\mathbf{e}_{s,k}\|
\le{}& \|\mathbf{z}_{s,k} - \mathbf{v}_{s,k}\| +\|\mathbf{v}_{s,k} - \mathbf{p}_{s,k}\|\nonumber\\
= {}&\left\|\mathsf{\Pi}_{\mathcal{X}_0}\left[\mathbf{v}_{s,k} - \beta_{s,k}h_+'(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})\right] -\mathbf{v}_{s,k}\right\| \nonumber\\
{}&+\left\|\mathsf{\Pi}_{\mathcal{X}_0}[\mathbf{p}_{s,k}- \alpha_k f'(\mathbf{p}_{s,k})]- \mathbf{p}_{s,k}\right\|\nonumber\\
\le{}&\frac{h_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})}{c_h} +\alpha_kL_f,
\end{align}
where the second inequality follows from fact that $\mathbf{p}_{s,k}, \mathbf{v}_{s,k} \in \mathcal{X}_0$, the nonexpansiveness of the projection operator, relations \eqref{eqn:L_f1} and \eqref{eqn:dhbnd}.
Applying $(a+b)^2 \le 2a^2 + 2b^2$ in inequality (\ref{eqn:lem3a}), we have for all $s\in \ccalS$ and $k\ge 1$,
\begin{equation}\label{eqn:lem3b}
\|\mathbf{e}_{s,k}\|^2 \leq 2\frac{h_+^2(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})}{c_h^2} +2\alpha_k^2L_f^2
\end{equation}
For the term $h_+^2(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k})$ in (\ref{eqn:lem3b}), we have the following chain of relations:
\begin{align}
&h_+^2(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k}) \nonumber\\
=&~ \left(h_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k}) -h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\right)^2 \label{eqn:dis1}\\
&~+\hspace{-0.7mm} h_+^2(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\nonumber\\
&~ + 2h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\nonumber\\
&~ \times\left(h_+(\mathbf{v}_{s,k};\ccalD_{s,k},\omega_{s,k}) -h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\right)\nonumber\\
\le &~ (L_h+D)^2\|\mathbf{v}_{s,k}-\mathbf{p}_{s,k}\|^2 + h^2_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\label{eqn:dis2}\\
&~+ 2(L_h+D)\|\mathbf{v}_{s,k}-\mathbf{p}_{s,k}\|h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\nonumber\\
\le &~ \alpha_k^2(L_h+D)^2L_f^2 + h^2_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\label{eqn:dis3}\\
&~+ 2\alpha_k(L_h+D)L_fh_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})\nonumber\\
\le &~ 2\alpha_k^2(L_h+D)^2L_f^2 + h_+^2(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k}),\label{eqn:dis4}
\end{align}
where (\ref{eqn:dis1}) follows from adding and subtracting $h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})$;
(\ref{eqn:dis2}) follows from relation (\ref{eqn:L_h2});
(\ref{eqn:dis3}) follows from Line \ref{line:v} of Algorithm \ref{alg:filter+opt};
(\ref{eqn:dis4}) follows from using the relation $2|a||b| \le a^2 + b^2$ with $a = \alpha_k(L_h+D)L_f$ and $b = h_+(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})$.
By combining (\ref{eqn:lem3b}) and (\ref{eqn:dis4}), we obtain
\begin{align*}
\|\mathbf{e}_{s,k}\|^2 \leq &~ 2\alpha_k^2L_f^2\left(1+\frac{2(L_h+D)^2}{c_h^2}\right) + \frac{2h_+^2(\mathbf{p}_{s,k};\ccalD_{s,k},\omega_{s,k})}{c_h^2}\\
\leq &~ 2\alpha_k^2L_f^2\left(1+\frac{2(L_h+D)^2}{c_h^2}\right) + \frac{2L_h^2 \mathrm{dist}^2(\mathbf{p}_{s,k},\mathcal{X})}{c_h^2},
\end{align*}
where the last inequality is from \eqref{e:Lh}.
Therefore, from Lemma \ref{lem:feas} and $\sum_{k=1}^\infty\alpha_k^2<\infty$, we conclude that
$
\sum_{k=1}^\infty \|\mathbf{e}_{s,k}\|^2 <\infty$ for all $ s \in \ccalS \quad a.s.$
\end{proof}
\begin{proof}[Proof of Lemma \ref{lem:disagree}(b)]
The proof of this part of Lemma will coincide with \cite[Lemma 6(b)]{lee15approx}.
\end{proof}
| 0168cd123032e1cd53e5460752f109d0036927ff | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The study of deep-inelastic scattering (DIS) of leptons off nuclei reveals
an appearance of a significant nuclear effect (for a review see, e.g.,~\cite{Arneodo:1992wf,Rith:2014tma}).
It was first observed by the European Muon Collaboration~\cite{Aubert:1983xm} in the valence quark dominance region;
hence the name. This observation rules out the naive picture of a nucleus as being a system of quasi-free nucleons.
There in general are two mainstream approaches to studying the EMC effect. In the first one, which is at present
more popular, nuclear parton distribution functions (nPDFs) are extracted from the global fits to nuclear
data by using empirical parametrizations of their normalizations (see~\cite{Eskola:2009uj,Eskola:2016oht,Khanpour:2016pph}).
This is completely analogous to respective studies of usual (nucleon) PDFs (see recent analyses in~\cite{fits}).
Both PDFs and nPDFs are obtained from the numerical solution to Dokshitzer-Gribov-Lipatov-Altarelli-Parisi (DGLAP)
equations~\cite{DGLAP}
\footnote{Sometimes, in the analyses of DIS experimental data it is convenient to use an exact
solution to DGLAP equations in the Mellin moment space and reconstruct SF $F_2$ from the moments
(see recent paper~\cite{Kotikov:2016ljf} and references and discussions therein). The studies of
nuclear effects in such a type of analyses can be found in~\cite{Krivokhizhin:2005pt}, though its
consideration is beyond the scope of the present study.}.
The second strategy is based upon some models of nuclear PDFs (see different models in,
for example,~\cite{Kulagin:2004ie}--\cite{Close:1984zn}
and a recent review~\cite{Kulagin:2016fzf}).
Here we will follow the rescaling model~\cite{Jaffe:1983zw,Close:1984zn}, which was very popular some time ago.
The model is based on a suggestion~\cite{Close:1983tn} that the effective confinement size of gluons and quarks in the
nucleus is greater than in a free nucleon. In the framework of perturbative QCD it was found \cite{Jaffe:1983zw,Close:1984zn,Close:1983tn}
that such a change in the confinement scale predicts that nPDFs and PDFs can be related by simply rescaling their arguments
(see Eq.~(\ref{va.1a}) below). Thus, in a sense, the rescaling model lies in-between two above approaches: in its framework there
are certain relations between usual and nuclear PDFs that result from shifting the values of kinematical variable $\mu^2$;
however, both densities obey DGLAP equations.
At that time, the model was established for the valence quark dominance region $0.2 \leq x \leq 0.8$.
The aim of our paper is to extend its applicability to the region of small $x$ values,
where the rescaling values can be different for gluons and quarks. To see it clearly
we use the generalized double-scaling approach (DAS)~\cite{Munich,Q2evo}. The latter is
based upon the analytical solution to DGLAP equations in the small $x$ region and generalizes
earlier studies~\cite{BF1}.
A few years ago most analyses of nPDFs have been done in the leading order
(LO) of perturbation theory, but now the situation is drastically changed and the standard level of accuracy in current
analyses is at the next-to-leading order (NLO) one (see~\cite{Eskola:2009uj,Eskola:2016oht}).
Even more, there have already appeared a global analysis~\cite{Khanpour:2016pph}
performed at the next-to-next-to-leading order.
Nevertheless the present analysis will be carried out in LO. We note that the analysis to this
level of accuracy is just for the start and can be considered as a first step in our investigations
in this direction. We are going to improve the accuracy at least to the NLO level in the future works.
\section{SF $F_2$ at low $x$}
A reasonable agreement between HERA data~\cite{H1ZEUS} and predictions made
by perturbative Quantum Chromodynamics (QCD) was observed for $Q^2 \geq 2$ GeV$^2$
\cite{CoDeRo}, thereby promising that perturbative QCD is capable of describing the
evolution of parton densities down to very low $Q^2$ values.
Some time ago ZEUS and H1 Collaborations have presented new precise combined data~\cite{Aaron:2009aa}
on the structure function (SF) $F_2$.
An application of the generalized DAS approach~\cite{Q2evo} at NLO
shows that theoretical predictions are well compatible with experimental data at $Q^2 \geq 3\div 4$ GeV$^2$
(see recent results in~\cite{Kotikov:2012sm}).
In the present paper we perform a LO analysis of the combined data~\cite{Aaron:2009aa}
where the SF $F_2$ has the following form
\begin{eqnarray}
F_2(x,\mu^2) &=& e \,
f_q(x,\mu^2),
\label{8a}
\end{eqnarray}
where
$e=(\sum_1^f e_i^2)/f$ is an average of the squared quark charges.
Notice that the approach used in these analyses will be analogous to that exploited in NLO ones carried out
in~\cite{Kotikov:2012sm}--\cite{Cvetic1}.
The small-$x$ asymptotic expressions for parton densities $f_a$ can be written as follows
\begin{eqnarray}
f_a(x,\mu^2) &=&
f_a^{+}(x,\mu^2) + f_a^{-}(x,\mu^2),~~(\mbox{hereafter}~~~a=q,g) \nonumber \\
f^{+}_g(x,\mu^2) &=& \biggl(A_g + \frac{4}{9} A_q \biggl)
\tilde{I}_0(\sigma) \; e^{-\overline d_{+} s} + O(\rho), \nonumber \\
f^{+}_q(x,\mu^2) &=&
\frac{f}{9} \biggl(A_g + \frac{4}{9} A_q \biggl) \rho \tilde{I}_1(\sigma) \; e^{-\overline d_{+} s}
+ O(\rho),
\label{8.01} \\
f^{-}_g(x,\mu^2) &=& -\frac{4}{9} A_q e^{- d_{-} s} \, + \, O(x),~~
f^{-}_q(x,\mu^2) ~=~ A_q e^{-d_{-}(1) s} \, + \, O(x),
\label{8.02}
\end{eqnarray}
where $I_{\nu}$ ($\nu=0,1$)
are the modified Bessel functions
with
\begin{eqnarray}
&&s=\ln \left( \frac{a_s(\mu^2_0)}{a_s(\mu^2)} \right),~~
\sigma = 2\sqrt{\left|\hat{d}_+\right| s
\ln \left( \frac{1}{x} \right)} \ ,~~~ \rho=\frac{\sigma}{2\ln(1/x)},~~ \nonumber \\
&&a_s(\mu^2) \equiv \frac{\alpha_s(\mu^2)}{4\pi} = \frac{1}{\beta_0\ln(\mu^2/\Lambda^2_{\rm LO})}
\label{intro:1a}
\end{eqnarray}
and
\begin{equation}
\hat{d}_+ = - \frac{12}{\beta_0},~~~
\overline d_{+} = 1 + \frac{20f}{27\beta_0},~~~
d_{-} = \frac{16f}{27\beta_0}
\label{intro:1b}
\end{equation}
denote singular $\hat{d}_+$ and regular $\overline d_{+}$ parts of the ``anomalous dimensions''
$d_{+}(n)$ and $d_{-}(n)$
\footnote{Note that the variables $d_{\pm}(n)$ are ratios $\gamma_{\pm}^{(\rm LO)}(n)/(2\beta_0)$
of LO anomalous dimensions $\gamma_{\pm}^{(\rm LO)}(n)$ and LO coefficient $\beta_0$ of QCD
$\beta$-function.},
respectively, in the limit $n\to1$.
By using the expressions given above we have analyzed H1 and ZEUS data for $F_2$ \cite{Aaron:2009aa}.
In order to keep the analysis as simple as possible, here we take $\mu^2=Q^2$ and $\alpha_s(M^2_Z)=0.1168$
in agreement with ZEUS results presented in~\cite{H1ZEUS}.
Moreover, we use the fixed flavor scheme with two different values $f=3$ and $f=4$ of active quarks.
As can be seen from Table 1, the twist-two approximation looks reasonable for $Q^2 \geq 3.5$ GeV$^2$.
It is almost completely compatible with NLO analyses done in~\cite{Kotikov:2012sm}--\cite{Cvetic1}.
Moreover, these results are rather close to original analyses
(see~\cite{Cooper-Sarkar:2016foi} and references therein) performed by the HERAPDF group.
As in the case of~\cite{Cooper-Sarkar:2016foi} our $\chi^2/DOF \sim 1$ unless combined H1 and ZEUS experimental
data analyzed are kept according to $Q^2 \geq 3.5$ GeV$^2$.
At lower $Q^2$ there is certain disagreement, which is we believe to be explained by the higher-twist (HT)
corrections playing their important role. These HT corrections have rather cumbersome form
at low $x$~\cite{HT}. As it was shown~\cite{Cvetic1}, it is very promising to use
infrared modifications of the strong coupling constant in our analysis.
Such types of coupling constants modify the low $\mu^2$ behavior of parton densities
and structure functions. What is important, they do not generate additional free parameters.
Moreover, the present results will be applied in the analyses of NMC data (see Sect.~5 and~6) accumulated
at very low $Q^2$ values, where the HT expansion ($\sim 1/Q^{2n}$) is thought to be not applicable.
So, following~\cite{Cvetic1}, we are going to use the so-called ``frozen'' $a_{\rm fr}(\mu^2)$ \cite{Badelek:1996ap}
and analytic $a_{\rm an}(\mu^2)$ \cite{Shirkov:1997wi} versions
\begin{equation}
a_{\rm fr}(\mu^2) = a_s(\mu^2 + M^2_{g}),~~~
a_{\rm an}(\mu^2) = a_s(\mu^2) - \frac{1}{\beta_0} \, \frac{\Lambda_{\rm LO}^2}{\mu^2-\Lambda_{\rm LO}^2} \, ,
\label{Ana}
\end{equation}
where $M_{g}$ is a gluon mass with $M_{g}$=1 GeV$^2$ (see~\cite{Shirkov:2012ux} and references therein
\footnote{There are a number of various approaches to define the value of this gluon mass and even
the form of its momentum dependence (see, e.g., a recent review~\cite{Deur:2016tte}).}).
It is seen that the results of the fits carried out when $a_{\rm fr}(\mu^2)$ and $a_{\rm an}(\mu^2)$ are used,
are very similar to the corresponding ones obtained in~\cite{Kotikov:2012sm}.
Moreover, note that the fits in the cases with ``frozen'' and analytic strong coupling constants look
very much alike (see also~\cite{Cvetic1,KoLiZo}) and describe fairly well the data in the low $Q^2$ region,
as opposed to the fits with a standard coupling constant, which largely fails here.
The results are presented in Table~1. With the number of active quarks $f=4$, they are shown also in Fig.~1.
Just like the previous analyses~\cite{Kotikov:2012sm,Cvetic1,KoLiZo} we observe strong improvement
in the agreement between theoretical predictions and experimental data once ``frozen'' and analytic
modifiations to the coupling constant are applied. When the data are cut by $Q^2 \geq $ 1 GeV$^2$,
$\chi^2$ value drops by more than two times. Ditto for the analyses of data with $Q^2 \geq $ 3.5 GeV$^2$ imposed.\\
{\bf Table 1.}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
$f=3$ & $a_s(Q^2)$ & $a_s(Q^2)$ & $a_{an}(Q^2)$ & $a_{an}(Q^2)$ & $a_{fr}(Q^2)$ & $a_{fr}(Q^2)$ \\
$Q^2 \geq $ & 1 GeV$^2$ & $3.5$ GeV$^2$ & $1$ GeV$^2$ & $3.5$ GeV$^2$ & $1$ GeV$^2$ & $3.5$ GeV$^2$ \\
\hline
$A_g$ & 0.46 $\pm$ 0.02 & 0.74 $\pm$ 0.04 & 1.16 $\pm$ 0.03 & 1.30 $\pm$ 0.04 & 0.96 $\pm$ 0.03 &
1.06 $\pm$ 0.04 \\
$A_q$ & 1.58 $\pm$ 0.04 & 1.48 $\pm$ 0.06 & 1.16 $\pm$ 0.04 & 1.21 $\pm$ 0.07 & 1.23 $\pm$ 0.08 &
1.32 $\pm$ 0.07 \\
$Q_0^2$ & 0.40 $\pm$ 0.01 & 0.46 $\pm$ 0.01 & 0.20 $\pm$ 0.01 & 0.16 $\pm$ 0.01 & 0.49 $\pm$ 0.01 &
0.53 $\pm$ 0.01 \\
$\chi^{2}$ & 365.7 & 69.7 & 149.7 & 42.9 & 140.4 & 47.6 \\
\hline \hline
$f=4$ & $a_s(Q^2)$ & $a_s(Q^2)$ & $a_{an}(Q^2)$ & $a_{an}(Q^2)$ & $a_{fr}(Q^2)$ & $a_{fr}(Q^2)$ \\
$Q^2 \geq $ & 1 GeV$^2$ & $3.5$ GeV$^2$ & $1$ GeV$^2$ & $3.5$ GeV$^2$ & $1$ GeV$^2$ & $3.5$ GeV$^2$ \\
\hline
$A_g$ & 0.47 $\pm$ 0.02 & 0.54 $\pm$ 0.03 & 0.65 $\pm$ 0.02 & 0.76 $\pm$ 0.03 & 0.96 $\pm$ 0.03 &
0.77 $\pm$ 0.03 \\
$A_q$ & 1.58 $\pm$ 0.04 & 1.09 $\pm$ 0.06 & 0.95 $\pm$ 0.03 & 0.96 $\pm$ 0.04 & 1.23 $\pm$ 0.05 &
0.95 $\pm$ 0.06 \\
$Q_0^2$ & 0.40 $\pm$ 0.01 & 0.37 $\pm$ 0.01 & 0.16 $\pm$ 0.01 & 0.19 $\pm$ 0.01 & 0.49 $\pm$ 0.01 &
0.43 $\pm$ 0.01 \\
$\chi^{2}$ & 366.0 & 57.0 & 166.3 & 43.6 & 140.0 & 40.6 \\
\hline
\end{tabular}
\end{center}
\begin{figure}[t]
\centering
\vskip 0.5cm
\includegraphics[height=0.45\textheight,width=0.8\hsize]{hera1.eps}
\vskip -0.3cm
\caption{$x$ dependence of $F_2(x,Q^2)$ in bins of $Q^2$.
The combined experimental data from H1 and ZEUS Collaborations
\cite{Aaron:2009aa} are
compared with the LO fits for $Q^2\geq1$~GeV$^2$ implemented with a standard strong coupling constant (solid lines),
and its frozen (dash-dotted lines) and analytic (dashed lines) modifications.}
\label{fig:F1}
\end{figure}
Recent NLO analyses (see the third paper in~\cite{Kotikov:2012sm})
have been carried out within the framework of the fixed flavor scheme with $f=3$ active
light flavors and with a purely perturbative charm quark generated in a photon-gluon fusion (PGF) process.
Such type of analyses for the complete SF $F_2(x,Q^2)$ cannot be done at LO.
\footnote{Notice that the SF $F_{2c}(x,Q^2)$, the charm part of $F_2(x,Q^2)$, appears with $a_s(Q^2)$ and can be confronted
already at LO with the data produced in a PGF process (see Sect. 7 below).}
Therefore, we should use some fixed values of active quarks.
Nevertheless, we would like to note that the results obtained here and those in~\cite{Kotikov:2012sm}---\cite{Cvetic1},
where various schemes were used, are very stable and close to each other.
\section{Rescaling model}
In the rescaling model~\cite{Close:1984zn} SF $F_2$ and, therefore, valence part of quark densities,
gets modified in the case of a nucleus $A$ at intermediate and large $x$ values $(0.2 \leq x \leq 0.9)$ as follows
\begin{equation}
F_2^A(x,\mu^2) =
F_2(x,\mu^2_{A,v}),~~~
f_{NS}^A(x,\mu^2) =
f_{NS}(x,\mu^2_{A,v}),
\label{va.1}
\end{equation}
where a new scale $\mu^2_{A,v}$ is related with $\mu^2$ as
\begin{equation}
\mu^2_{A,v} = \xi^A_v(\mu^2)\mu^2,~~~
\xi^A_v(\mu^2) = {\left(\frac{\lambda_A^2}{\lambda_N^2}\right)}^{a_s(\tilde{\mu}^2)/a_s(\mu^2)} \,
\label{va.1a}
\end{equation}
where some additional scale $\tilde{\mu}^2=0.66$ GeV$^2$, which was in its turn an initial point
in a $\mu^2$-evolution performed in~\cite{Close:1984zn}; it is then estimated in Appendix~A of that paper.
The quantity $\lambda_A/\lambda_N$ stands for the ratio of quark confinement radii in a nucleus $A$ and nucleon.
The values of $\lambda_A/\lambda_N$ and $\xi^A_v(\mu^2)$ at $\mu^2=20$ GeV$^2$
were evaluated for different nuclei and presented in Tables I and II in~\cite{Close:1984zn}.
Since the factor $ \xi^A_v(\mu^2)$ is $\mu^2$ dependent, it is convenient to transform it
to some $\mu^2$ independent one. To this end, we consider the variable
$\ln(\mu^2_{A,v}/\Lambda^2)$, which has the following form (from Eq.~(\ref{va.1a}))
\begin{equation}
\ln\left(\frac{\mu^2_{A,v}}{\Lambda^2}\right) = \ln\left(\frac{\mu^2}{\Lambda^2}\right) \cdot \Bigl(1+ \delta^A_v\Bigr)
\label{va.1b}
\end{equation}
where the nuclear correction factor $\delta^A_v$ becomes $\mu^2$ independent:
\begin{equation}
\delta^A_v = \frac{1}{\ln\left(\tilde{\mu}^2/\Lambda^2\right)} \,
\ln\left(\frac{\lambda_A^2}{\lambda_N^2}\right)\,,
\label{delta}
\end{equation}
where it is seen that two parameters, namely, the scale $\tilde{\mu}$ and ratio $\lambda_A/\lambda_N$,
are combined to form a $Q^2$-independent quantity.
Using Eqs.~(\ref{va.1b}) and/or~(\ref{delta}), we can recover results for $\delta^A_v$, which are presented in Table~2.
{\bf Table 2.}
\begin{center}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
$A$ & ${}^2$D & ${}^4$He & ${}^7$Li & ${}^{12}$C & ${}^{40}$Ca \\
$N$ & & 11 & 16 & 16 & 11 \\
\hline
$\delta^A_v$ & 0.01 & 0.06 & 0.05 & 0.08 & 0.11 \\
$\delta^{AD}_v$ & 0 & 0.05 & 0.04 & 0.07 & 0.10 \\
-$\delta^{AD}_{+,an}$ & 0 & 0.06 $\pm$ 0.01 & 0.06 $\pm$ 0.01 & 0.11 $\pm$ 0.01 & 0.19 $\pm$ 0.01 \\
-$\delta^{AD}_{-,an}$ & 0 & 0.24 $\pm$ 0.08 & 0.22 $\pm$ 0.07 & 0.41 $\pm$ 0.04 & 0.51 $\pm$ 0.04 \\
$\chi^{2}_{an}$ & 0 & 4.68 & 17 & 9.68 & 12 \\
-$\delta^{AD}_{+,fr}$ & 0 & 0.06 $\pm$ 0.01 & 0.06 $\pm$ 0.01 & 0.12 $\pm$ 0.01 & 0.21 $\pm$ 0.02 \\
-$\delta^{AD}_{-,fr}$ & 0 & 0.32 $\pm$ 0.08 & 0.28 $\pm$ 0.07 & 0.54 $\pm$ 0.04 & 0.71 $\pm$ 0.04 \\
$\chi^{2}_{fr}$ & 0 & 5 & 35 & 26 & 37 \\
\hline
\end{tabular}
\end{center}
Since our parton densities contain the variable $s$ defined in Eq.~(\ref{intro:1a}),
it is convenient to consider its $A$ modification. It has the following simple form:
\begin{equation}
s^A_v \equiv \ln \left(\frac{\ln\left(\mu^2_{A,v}/\Lambda^2\right)}{\ln\left(\mu^2_{0}/\Lambda^2\right)}\right)
= s +\ln\Bigl(1+\delta^A_v\Bigr) \approx s +\delta^A_v,~~~
\label{sA}
\end{equation}
i.e. the nuclear modification of the basic variable $s$ depends on the
$\mu^2$ independent parameter $\delta^A_v$, which possesses very small values.
\section{Rescaling model al low $x$}
Standard evidence coming from earlier studies contains conclusion about inapplicability
of the rescaling model at small $x$ values (see, for example,~\cite{Efremov:1986mt}).
It looks like it can be related with some simplifications of low $x$ analyses (see, for example,~\cite{Kotikov:1988aa},
where the rise in EMC ratio was wrongly predicted at small $x$ values).
Using an accurate study of DGLAP equations at low $x$ within the framework of the generalized DAS approach,
it is possible to achieve nice agreement with
the experimental data for the DIS structure functon $F_2$ (see previous section)\footnote{Moreover,
using an analogous approach, good agreement was also found with the corresponding data for jet multiplicites~\cite{Bolzoni:2012ii}.}.
Therefore, we believe that all these indicate toward success in describing the EMC ratio by using the same approach.
We note that the main difference between global fits and DAS approach is in the restriction of
applicability of the latter by
low $x$ region only, while the advantage of the DAS approach lies in the analytic solution to DGLAP equations.
Thus, we are trying to apply the DAS approach to low $x$ region of EMC effect using a simple fact that the rise of parton
densities increases with increasing $Q^2$ values. This way, with scales of PDF evolutions
less than $Q^2$ (i.e. $\mu^2 \leq Q^2$) in nuclear cases, we can directly reproduce the shadowing effect which is observed
in the global fits. Since there are two components~(\ref{8.01}) for each parton density, we have two free parameters
$\mu_{\pm}$ to be fit in the analyses of experimental data for EMC effect at low $x$ values.
An application of the rescaling model at low $x$ can be incorporated at LO as follows:
\begin{eqnarray}
F^A_2(x,\mu^2) &=& e \, f^A_q(x,\mu^2),~~ F^N_2(x,\mu^2) = e \, f_q(x,\mu^2),
\nonumber \\
f^A_a(x,\mu^2) &=&
f_a^{A,+}(x,\mu^2) + f_a^{A,-}(x,\mu^2),~~(a=q,g),~~
f^{A,\pm}_a(x,\mu^2) =
f^{\pm}_a(x,\mu^2_{A,\pm}) \, ,
\label{8.02A}
\end{eqnarray}
with a similar definition of $\mu^2_{A,\pm}$ as in the previous section (up to replacement
$v \to \pm$). The expressions for $f^{\pm}_a(x,\mu^2)$ are given in Eqs.~(\ref{8.01}) and~(\ref{8.02}).
Then, the corresponding values of $s^A_{\pm}$ are found to be
\begin{equation}
s^A_{\pm} \equiv \ln \left(\frac{\ln\left(\mu^2_{A,\pm}/\Lambda^2\right)}{\ln\left(\mu^2_{0}/\Lambda^2\right)}\right)
= s +\ln\Bigl(1+\delta^A_{\pm}\Bigr)\, ,
\label{sApm}
\end{equation}
because of the saturation at low $x$ values for all considered $Q^2$ values, which in our case should be related
with decreasing the arguments of ``$\pm$'' component. Therefore, the values of $\delta^A_{\pm}$ should be negative.
\section{Analysis of the low $x$ data for nucleus}
Note that it is usually convenient to study the following ratio
(see Fig.~1 in Ref.~\cite{Kulagin:2016fzf})
\begin{equation}
R^{AD}_{F2}(x,\mu^2) = \frac{F^A_2(x,\mu^2)}{F^D_2(x,\mu^2)}\,.
\label{AD}
\end{equation}
Using the fact that the nuclear effect in a deutron is very small (see Table~1 for the values
of $\delta^A_{v}$ and discussions in~\cite{Kulagin:2016fzf})
\footnote{The study of nuclear effects in a deutron can be found in~\cite{AKP},
which also contains short reviews of preliminary investigations.},
we can suggest that
\begin{eqnarray}
F^D_2(x,\mu^2) &=& e \, f_q(x,\mu^2),~~
F^A_2(x,\mu^2) =
e \, \overline{f}^{A}_q(x,\mu^2), \nonumber \\
\overline{f}^{A}_a(x,\mu^2) &=&
\overline{f}_a^{A,+}(x,\mu^2) + \overline{f}_a^{A,-}(x,\mu^2),~~(a=q,g),~~
\overline{f}^{A,\pm}_a(x,\mu^2) =
f^{\pm}_a(x,\mu^2_{AD,\pm}) \, ,
\label{AD1}
\end{eqnarray}
i.e.
\begin{eqnarray}
\overline{f}^{A,+}_g(x,\mu^2) &=& \biggl(A_g + \frac{4}{9} A_q \biggl)
I_0(\sigma^{AD}_{+}) \; e^{-\overline d_{+} s^{AD}_{+}}
+ O(\rho^{AD}_{+}), \nonumber \\
\overline{f}^{A,+}_q(x,\mu^2) &=&
\frac{f}{9} \biggl(A_g + \frac{4}{9} A_q \biggl) \rho^{AD}_{+}
I_1(\sigma^{AD}_{+}) \; e^{-\overline d_{+} s^{AD}_{+}} + O(\rho^{AD}_{+}),
\label{8.01AD} \\
\overline{f}^{A,-}_g(x,\mu^2) &=& -\frac{4}{9} A_q e^{- d_{-} s^{AD}_{-}} \,
+ \, O(x),~~
\overline{f}^{A,-}_q(x,\mu^2) = A_q e^{-d_{-}(1) s^{AD}_{-}} \, + \, O(x),
\label{8.02AD}
\end{eqnarray}
where
\begin{eqnarray}
&&\sigma^{AD}_{+}=\sigma(s\to s^{AD}_{+}),~~\rho^{AD}_{+}=\rho(s\to s^{AD}_{+}),~~ \nonumber \\
&&s^{AD}_{\pm} \equiv \ln \left(\frac{\ln\left(\mu^2_{AD,\pm}/\Lambda^2\right)}{\ln\left(\mu^2_{0}/\Lambda^2\right)}\right)
= s +\ln\Bigl(1+\delta^{AD}_{\pm}\Bigr)
\, .
\end{eqnarray}
We obtain the values of $\delta^{AD}_{+}$ and $\delta^{AD}_{-}$
by fitting NMC experimenal data~\cite{Arneodo:1995cs}
for the EMC ratio at low $x$ in the case of different nuclei.
Since the experimental data for lithium and carbon are most precise
and contain the maximal number of points (16 points for each nucleus),
we preform combined fits of these data.
Obtained results (with $\chi^{2}_{an}$=27 and $\chi^{2}_{fr}$=43 for 32 points)
are presented in Table~3 and shown in Fig.~2.
\newpage
{\bf Table 3.}
\begin{center}
\begin{tabular}{|c|c|c|c|c|}
\hline
& -$\delta^{AD}_{+,an}$ & -$\delta^{AD}_{-,an}$ & -$\delta^{AD}_{+,fr}$ & -$\delta^{AD}_{-,fr}$ \\
\hline
${}^7$Li & 0.061 $\pm$ 0.006 & 0.216 $\pm$ 0.065 & 0.073 $\pm$ 0.012 & 0.348 $\pm$ 0.067 \\
${}^{12}$C & 0.105 $\pm$ 0.007 & 0.411 $\pm$ 0.042 & 0.139 $\pm$ 0.013 & 0.590 $\pm$ 0.041\\
\hline
\end{tabular}
\end{center}
As can be seen in Fig.~2 there is large difference between the fits with ``frozen''
and analytic versions of the strong couling constant. This is in contrast with the analysis done in Section~1
and results done in the earlier papers~\cite{KoLiZo}.
It seems that this difference comes about because we include in the analysis the region of very low $Q^2$ values,
where ``frozen'' and analytic strong couling constants are observed to be rather different (see also~\cite{Shirkov:2012ux}).
\begin{figure}[!hbt]
\centering
\vskip 0.5cm
\includegraphics[height=0.45\textheight,width=0.8\hsize]{emcLiCnew.eps}
\vskip -0.3cm
\caption{small $x$ dependence of $R^{AD}_{a}(x,\mu^2)$ for lithium and carbon.
The combined experimental data from NMC~\cite{Arneodo:1995cs}
are fitted by LO expressions
implemented with the frozen (solid lines) and analytic (dashed lines)
modifications of the strong coupling constant.}
\end{figure}
\vskip 0.5cm
\section{ $A$ dependence at low $x$ }
Taking NMC experimental data~\cite{Arneodo:1995cs} along with E665 and HERMES Collaborations~\cite{Adams:1995is}
for the EMC ratio at low $x$ in the case of different nuclei, we can find the $A$ dependence of $\delta^{AD}_{\pm}$,
which can be parameterized as follows
\begin{equation}
- \delta^{AD}_{\pm} = c^{(1)}_{\pm} + c^{(2)}_{\pm} A^{1/3}.
\label{AD2}
\end{equation}
As it was already mentioned in the previous section, usage of the analytic coupling constant leads
to the fits with smaller $\chi^2$ values. For example, the values of $c^{(1)}_{\pm}$ and $c^{(2)}_{\pm}$ found
in the combined fit of the data (76 points) when the analytic coupling constant is used (with $\chi^2=89$)
look like
\begin{eqnarray}
&& c^{(1)}_{+,an} = -0.055 \pm 0.015,~~ c^{(2)}_{+,an} = 0.068 \pm 0.006,~~ \nonumber \\
&& c^{(1)}_{-,an} = 0.071 \pm 0.101,~~ c^{(2)}_{-,an} = 0.120\pm 0.039 \, .
\end{eqnarray} \label{AD2.an}
Now, using the $A$ dependence (\ref{AD2}), $R^{AD}_{F2}(x,\mu^2)$ values for any nucleus $A$ can be predicted.
What is more, we can consider also the ratios $R^{AD}_{a}(x,\mu^2)$ of parton densities in a nucleus and deutron themselves,
\begin{equation}
R^{AD}_{a}(x,\mu^2) = \frac{\overline{f}^A_a(x,\mu^2)}{f_a(x,\mu^2)},~~ (a=q,g) \, ,
\label{ADa}
\end{equation}
with $\overline{f}^A_a(x,\mu^2)$ and ${f_a(x,\mu^2)}$ defined in Eqs.~(\ref{AD1})---(\ref{AD2}) and~(\ref{8.01})---(\ref{intro:1b}), respecively.
Indeed, at LO $R^{AD}_{q}(x,\mu^2)=R^{AD}_{F2}(x,\mu^2)$; therefore, results for $R^{AD}_{q}(x,\mu^2)$ are already known.
Since all the parameters of PDFs found within the framework of the generalized DAS approach are now fixed
we can predict the ratio $R^{AD}_{g}(x,\mu^2)$ of the gluon densities in a nucleus and nucleon given
in Eqs.~(\ref{8.01}), (\ref{8.02}), (\ref{8.01AD}) and (\ref{8.02AD}), which is currently under intensive studies
(see a recent paper~\cite{Frankfurt:2016qca} and review~\cite{Armesto:2006ph} along with references and discussion therein).
\begin{figure}[t]
\centering
\vskip 0.5cm
\includegraphics[height=0.45\textheight,width=0.8\hsize]{RF2Rg_comp_new.eps}
\vskip -0.3cm
\caption{$x$ dependence of $R^{AD}_{F2}(x,\mu^2)$ and $R^{AD}_{g}(x,\mu^2)$ at $\mu^2$=10 GeV$^2$ for lead data.
A green line with pink band (shows 90$\%$ uncertainties) is taken from the second paper of~\cite{Armesto:2006ph},
while a black one with light green band is obtained in the present paper.}
\end{figure}
The results for $R^{AD}_{F2}(x,\mu^2)$ and $R^{AD}_{g}(x,\mu^2)$, depicted in Fig.~3, show some
difference between these ratios. It is also seen that the difference is similar to that
obtained in a recent EPPS16 analysis (see the first paper in~\cite{Eskola:2016oht})
\footnote{
Note that the result for $R^{AD}_{g}(x,\mu^2)$ along with its uncertainty is completely determined
by both the rescaling model and the analytic form for parton densities at low $x$ values we've used.
Therefore, it is clear that the light green band for $R^{AD}_{g}(x,\mu^2)$ should become broader
due to a freedom in using various models.
Also note that a comparison between two uncertainty bands shown in Fig.~3 is in some sense misleading.
The pink band is much broader since the EPPS16 global analysis included a fit to all available data
across quite a wide range in $x$ as opposed to small $x$ consideration adopted in the present paper.
Nonetheless, we decided to quote it here just to give the reader an idea about the subject, at least qualitatively.}.
However, what for $R^{AD}_{F2}(x,\mu^2)$ and $R^{AD}_{g}(x,\mu^2)$ themselves (irrespective of other results),
we obtain a bit stronger effect at lowest $x$ values, which does in fact not contradict
the experimental data collected by the LHCb experiment (see recent review in~\cite{Winn:2017kwv}). Such a strong
effect is also well compatible with the leading order EPPS09 analysis (which can also be found in~\cite{Winn:2017kwv}).
It will be interesting to delve into more in-depth studies of the
ratio $R^{AD}_{g}(x,\mu^2)$, which is one of our aims in the future.
\section{SF $F_{2c}$ at low $x$}
Several years ago H1~\cite{Aaron:2009jy} and ZEUS~\cite{Chekanov:2009kj} Collaborations at HERA
have separately presented their new data on the charm structure function $F_{2c}$\footnote{Open charm production was
also observed in the COMPASS fixed target experiment~\cite{Adolph:2012ca}.} and more recently
they have combined these data on $F_{2c}(x,\mu^2)$~\cite{Abramowicz:1900rp}.
The SF $F_{2c}$ was found to be around 25\% of $F_{2}$, which is considerably larger than what was observed
by the European Muon Collaboration (EMC) at CERN \cite{Aubert:1982tt} at larger $x$ values, where it was only aroung
1\% of $F_{2}$.
Ensuing and very extensive theoretical analyses were carried out to establish that the $F_{2c}$ data can be described
through the perturbative generation of charm in QCD \cite{Frixione:1994dv}.
In view of this, a PGF process in experiments with nucleon and nucleus targetsis one of the most
effective and promising studies of gluon density (see a recent review~\cite{Chudakov:2016ytj}).
Following~\cite{Illarionov:2008be} the SF $F_{2c}$ at low $x$ can be represented
in the framework of the generalized DAS approach as follows
\begin{equation}
F_{2c}(x,\mu^2) = e^2_c \, a_s(\mu_c) \, C_{2,g}(1, z_c(\mu^2)) f_g(x,\mu^2),~~~ z_c(\mu^2)=\frac{m^2_c(\mu^2)}{\mu^2},~~~ e_c=\frac{2}{3} \, ,
\label{c.1}
\end{equation}
where $C_{2,g}(1, z_c(\mu^2)) $ is a first Mellin moment of the LO PGF coefficient function
$\tilde{C}_{2,g}(x, z_c(\mu^2))$. It can be obtained from the QED case~\cite{Baier:1966bf} by adjusting
the coupling constants
(see also the direct calculations in~\cite{Witten:1975bh,Kotikov:2001ct}).
The Mellin moment $C_{2,g}(1, z_c(\mu^2)) $ has a very compact form~\cite{Illarionov:2008be}:
\begin{equation}
C_{2,g}(1, z) = \frac{2}{3} \,
\left[ 1- \frac{2(1-z)}{\sqrt{1+4z}} \, \ln
\frac{\sqrt{1+4z}-1}{\sqrt{1+4z}+1} \right] \, .
\label{c.1a}
\end{equation}
The gluon density $f_g(x,\mu^2)$ is determined in~(\ref{8.01}) and (\ref{8.02}).
The scale $\mu_c$ in~(\ref{c.1}) is actually not fixed because the results for $F_{2c}$ are at LO. There are two
widespread scales, $\mu_c^2=4m_c^2$ \cite{Chudakov:2016ytj,Gluck:1993dpa} and $\mu_c^2=4m_c^2+\mu^2$
\cite{Aaron:2009jy,Chekanov:2009kj,Abramowicz:1900rp,Illarionov:2008be}. We will use below both of them
(see Subsect.~7.1).
In the framework of the rescaling model the SF $F^A_{2c}(x,\mu^2)$ for nucleus $A$ can be represented as follows
\begin{equation}
F^A_{2c}(x,\mu^2) = e_c^2 \, \sum_{i=\pm} a_s(\mu_c(\mu^2_{A,i})) \, C_{2,g}(1,z_c(\mu^2_{A,i})) f^{i}_g(x,\mu^2_{A,i})\,,
\label{c.2}
\end{equation}
where the scale $\mu^2_{A,i}$
looks like
\begin{equation}
\mu^2_{A,\pm}= \Lambda^2 {\left(\frac{\mu^2}{\Lambda^2}\right)}^{1+\delta^A_{\pm}} =
\mu^2 {\left(\frac{\mu^2}{\Lambda^2}\right)}^{\delta^A_{\pm}}
\, ,
\label{c.3}
\end{equation}
as it follows from~(\ref{va.1}) with the replacement $v \to \pm$.
The results for the ratios $R^{A}_{F2}(x,\mu^2)$, $R^{A}_{g}(x,\mu^2)$ and
\begin{equation}
R^{A}_{c}(x,\mu^2)= \frac{F^A_{2c}(x,\mu^2)}{F_{2c}(x,\mu^2)}
\end{equation}
should be rather similar. Moreover, they have similar $x$-dependences, as it will be shown in the following subsection.
\subsection{Analysis of the low $x$ data}
To have as close a relation with analyses in Sect.~5 as possible, let us consider the ratio
\begin{equation}
R^{AD}_{c}(x,\mu^2)= \frac{F^A_{2c}(x,\mu^2)}{F^D_{2c}(x,\mu^2)}\,.
\label{c.4}
\end{equation}
As in Sect.~5, we will use the following expressions for the SFs
\begin{eqnarray}
&& F^D_{2c}(x,\mu^2) = e^2_c \, a_s(\mu_c) \,C_{2,g}(1, a_c(\mu^2)) \, f_g(x,\mu^2), \nonumber \\
&& F^A_{2c}(x,\mu^2) =e_c^2 \, \sum_{i=\pm} a_s(\mu_c(\mu^2_{AD,i})) \, C_{2,g}(1, z_c(\mu^2_{AD,i})) \,
\overline{f}^{A,\pm}_g(x,\mu^2) \,,
\label{c.6}
\end{eqnarray}
where the gluon density $\overline{f}^{A,\pm}_a(x,\mu^2)= f^{\pm}_a(x,\mu^2_{AD,\pm})$ is defined
in~(\ref{8.01AD}) and (\ref{8.02AD}). The scale $\mu^2_{AD,\pm}$
can be obtained from~(\ref {c.3}) with the replacement $\delta^A_{\pm} \to \delta^{AD}_{\pm}$,
by analogy with analyses in Sect.~5.
\begin{figure}[t]
\centering
\vskip 0.5cm
\includegraphics[height=0.45\textheight,width=0.8\hsize]{RQ2eq10.eps}
\vskip -0.3cm
\caption{$x$ dependence of $R^{AD}_{c}(x,\mu^2)$, $R^{AD}_{cg}(x,\mu^2)$ and $R^{AD}_{c2}(x,\mu^2)$ at
$\mu^2$=10 GeV$^2$ for lead data and two choices of $\mu_c$ scale: $\mu_c^2=4m_c^2$ and $\mu_c^2=4m_c^2+\mu^2$
are shown by black, blue and pink lines, respectively.
A band represents 90$\%$ level uncertainties in determining $R^{AD}_{c}(x,\mu^2)$ values.}
\end{figure}
The results for the ratios $R^{AD}_{c}(x,\mu^2)$,
\begin{equation}
R^{AD}_{cg}(x,\mu^2)= \frac{R^{AD}_{c}(x,\mu^2)}{R^{AD}_{g}(x,\mu^2)}~~\mbox{ and }
R^{AD}_{c2}(x,\mu^2)= \frac{R^{AD}_{c}(x,\mu^2)}{R^{AD}_{F2}(x,\mu^2)}
\label{c.7}
\end{equation}
are presented in Fig.~4 for $\mu^2$ = 10 GeV$^2$.
Since the $\mu^2$-dependence of $m_c$ is not strong, we use fixed $m_c=1.27$ GeV~\cite{Breview}
in our analysis.
As can be seen in Fig.~4, results look very much the same for both scales of $\mu_c$.
What is more, a behavior of the ratio $R^{AD}_{c}(x,\mu^2)$ is a little bit weaker than
that of $R^{AD}_{F2}(x,\mu^2)$ and a bit stronger than that observed for $R^{AD}_{g}(x,\mu^2)$.
We hope that the $x$-dependence of the ratio $R^{AD}_{c}(x,\mu^2)$, along with that of $R^{AD}_{g}(x,\mu^2)$,
can be measured at a future Electron--Ion Collider (see~\cite{Chudakov:2016ytj} and discussion therein).
\section{Conclusion}
Using a recent progress in the application of double-logarithmic approximations~(see \cite{Q2evo,Kotikov:2012sm}
and~\cite{Bolzoni:2012ii}) to the studies of small $x$ behavior of the structure and fragmentation functions,
respectively, we applied the DAS approach~\cite{Munich,Q2evo} to examine an EMC $F_2$ structure function ratio
between various nuclei and a deutron. Within a framework of the rescaling model~\cite{Close:1984zn,Close:1983tn}
good agreement between theoretical predictions and respective experimental data is achieved.
The theoretical formul\ae ~contain certain parameters, whose values were fit in
the analyses of experimental data. Once the fits are carried out we have predictions for
the corresponding ratios of parton densities without free parameters. These results were used to
predict small $x$ behavior of the gluon density in nuclei, which is at present poorly known.
The ratios $R^{AD}_{a}(x,\mu^2)$ $(a=q,g)$ predicted in the present paper are compatible with those
given by various groups working in the area. From our point of view, it is quite valuable
that the application of the rescaling model~\cite{Close:1984zn,Close:1983tn} provided us with
very simple forms for these ratios.
It should also be mentioned that without any free parameters we can predict the ratio $R^{AD}_{c}(x,\mu^2)$
of charm parts, $F^{A}_{2c}(x,\mu^2)$ and $F^{D}_{2c}(x,\mu^2)$, of the respective structure functions.
This latter ratio
has a simple form and it is very similar to the corresponding ratio of the complete structure functions
$F^{A}_{2}(x,\mu^2)$ and $F^{D}_{2}(x,\mu^2)$.
Following~\cite{Q2evo,Kotikov:2012sm} we plan to extend our analysis
to the NLO level of approximation, the accuracy that is currently a standard in nPDF studies.
Also, we are going to consider a rather broad range of the Bjorken variable $x$ by
using parametrizations of parton densities, which will be constructed by analogy
with the one obtained earlier in the valence quark case (see~\cite{Illarionov:2010gy}).
The usage of such type of parametrizations will make it possible to carry out the present analysis of the data
accumulated within the range of intermediate $x$ values, which is presently under active considerations.\\
Support by the National Natural Science Foundation of China (Grant
No. 11575254) is acknowledged. A.V.K. and B.G.S. thank Institute of Modern Physics
for invitation. A.V.K. is also grateful to the CAS President's International Fellowship Initiative
(Grant No.~2017VMA0040) for support.
The work of A.V.K. and B.G.S. was in part supported by the RFBR Foundation through the Grant No.~16-02-00790-a.
| de913c10aa62b4029f02008598e5e9466139f59b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
One of the most important applications in the calculus of variation is to maximize (minimize) the functional
\begin{equation}\label{integralfunctional}
Q[y]=\int_{a}^{b} \left( \sqrt{p(x)} (y^\prime(x))^2 + \frac{ h(y(x))}{ \sqrt{p(x)}}\right) dx,
\end{equation}
where $p(x)$ is a positive and differentiable function on some open interval $(a,b)\subset \R,$ and $h(x)$ is a differentiable function. In fact, the functional
\begin{equation}\label{integralfunctional2}
Q[y]=\int_{a}^{b} \left( \sqrt{p(x)} (y^\prime(x))^2 + \frac{ h(y(x))}{ \sqrt{p(x)}}\right) dx
\end{equation}
attains its extreme values at a function $y(x)\in C^2(a,b)$ that $y(x)$ satisfies the Euler's-Lagrange differential equation \cite{Arfken, Hand, Makarets},
\begin{equation}\label{leequation}
\frac{\partial F}{\partial y}(x,y,y^\prime) - \frac{d}{dx}\left(\frac{\partial F}{\partial y^\prime}(x,y,y^\prime)\right)= 0,
\end{equation}
where
\begin{equation}\label{functionform}
F(x,y,y^\prime):=\sqrt{p(x)} (y^\prime(x))^2 + \frac{h(y(x))}{ \sqrt{p(x)}}.
\end{equation}
Therefore, the problem of maximizing (minimizing) $Q[y]$ is reduced to solve the differential equation \eqref{leequation}. i.e., to solve
\begin{equation}\label{primerydiffequation}
p(x)y^{\prime\prime}(x)+\frac{1}{2}p^\prime(x)y^\prime(x)=\frac{1}{2}h^\prime(y(x)).
\end{equation}
Hence, it is a matter to solve such differential equations. In the first part of this paper, we solve the following class of second order differential equation:
\begin{equation}\label{main}
p(x)y^{\prime\prime}(x)+\frac{1}{2}p^\prime(x)y^\prime(x)+f(\sqrt{p(x)}\;y^\prime(x),y(x))=0
\end{equation}
which generalizes the differential equation \eqref{primerydiffequation}. Here, we assume that $p(x)$ is a positive and differentiable function on some open interval $ (a,b)\subset \R$, and $f(\sqrt{p(x)}\;y^\prime(x),y(x))$ is continuous function on some domain $D\subset \R^2$. In fact, equation \eqref{main} not only generalizes \eqref{primerydiffequation}, but also it generalizes many of well known differential equation. For example
\begin{enumerate}
\item the Chebyshev's Differential Equation ~\cite{Arfken, Zw},
\[
(1 -x^2)y^{\prime\prime}(x)-x\;y^\prime(x)+n^2 y(x)=0,\;\; |x|<1,
\]
\item the Cauchy-Euler's Differential Equation ~\cite{Boyce, Kr, MURPHY},
\[
ax^2y^{\prime\prime}(x)+a\;x\;y^\prime(x)
+b\;y(x)=0,\;\; x>0,
\]
\item the Nonlinear Chebyshev's Equation,
\[
(1 -x^2)y^{\prime\prime}(x)+\left(\alpha\sqrt{1-x^2}-x\right)y^\prime(x) +f(y(x)) = 0,\;\; \;\;|x|<1,
\]
\item the Hypergeometric Differential Equation ~\cite{Anderws,Zw},
\[
x(1-x)y^{\prime\prime}(x) + \left[c-(a+b+1)x \right] y^\prime(x) - a\,b\;y(x) = 0, \;0<x<1, \;\text{ with } c=1/2, a=-b,
\]
and
\item the Nonlinear Hypergeometric Differential Equation,
\begin{equation*
x(1-x)y^{\prime\prime}(x)+ \left(\frac{1}{2}-x+\alpha\sqrt{x(1-x)}\right)y^\prime(x)+f(y(x))=0,\; 0<x<1.
\end{equation*}
\end{enumerate}
Throughout this paper, we call the class of differential equation in \eqref{main} by Chebyshev's-type of differential equations.
In the second part of this paper, we introduce an approach to solve
the differential equation
\begin{equation}\label{part2eq}
a_2\left(f^\prime(y)y^{\prime\prime}+(y^\prime)^2f^{\prime\prime}(y)\right)+a_1f^\prime(y)y^\prime+a_0f(y)=g(x),
\end{equation}
where $a_0,\;a_1$ and $a_2$ are constants, and $f\in C^2(a,b)$, for some open interval $(a,b) \subset \R$. We also give an approach to solve the differential equation
\begin{equation}\label{part2eeq}
p(x)\left(f^\prime(y)y^{\prime\prime}+(y^\prime)^2f^{\prime\prime}(y)\right)+\frac{1}{2}p^\prime(x)f^\prime(y)y^\prime+a_0f(y)=0,
\end{equation}
where $p(x)$ is a positive and differentiable function on some open interval $(a,b) \subset \R$, and $f\in C^2(c,d)$, for some open interval $(c,d) \subset \R$.
Throughout this paper, we call these classes of second order differential equations by $f-$type second order differential equations.
In the third part of this paper, we introduce an approach to solve
the second order nonlinear differential equation
\begin{equation}
a_2(x,y,y^\prime)\left(f^\prime(y)y^{\prime\prime}+f^{\prime\prime}(y)(y^\prime)^2\right)+a_1(x,y,y^\prime)(f^\prime(y)y^\prime)+a_0(x,y,y^\prime)=0,
\end{equation}
where $f(y)$ is an invertible function ($y=f^{-1}(z)$), and $f\in C^2(a,b)$ where $(a,b)$ is the open interval in $\R.$ To solve this class of differential equations, we assume that
\begin{equation}\label{exact}
a_2\left(x,f^{-1}(z),\frac{z^\prime}{f^\prime\left(f^{-1}(z)\right)}\right)z^{\prime\prime}+a_1\left(x,f^{-1}(z),\frac{z^\prime}{f^\prime\left(f^{-1}(z)\right)}\right)z^\prime+a_0\left(x,f^{-1}(z),\frac{z^\prime}{f^\prime\left(f^{-1}(z)\right)}\right)=0
\end{equation}
is exact differential equation. The differential equation \eqref{exact} is called exact if the conditions
\begin{equation}\label{exactcond2}
\frac{\partial a_2}{\partial z}=\frac{\partial a_1}{\partial z^\prime}, \;\; \frac{\partial a_2}{\partial x}=\frac{\partial a_0}{\partial z^\prime},\; \text{and}\; \;\frac{\partial a_1}{\partial x}=\frac{\partial a_0}{\partial z}.
\end{equation}
hold \cite{AlAhmad,Aljararha}.
In this case, the first integral of \eqref{exact} exists and it is given by
\[
\int_{x_0}^{x}a_0(\alpha,z,z^\prime)d\alpha+\int_{z_0}^{z}a_1(x_0,\beta,z^\prime)d\beta+\int_{z^\prime_0}^{z^\prime}a_2(x_0,z_0,\gamma)d\gamma=c.
\]
Throughout this paper, we call this class of differential equations by $f-$type second order differential equations that can be transformed into exact second order differential equations.
The layout of the paper: In the first section, we solve Chebyshev's-type of Second Order Differential Equations. In the second section, we solve the $f-$type of second order differential equations. In the third section, we solve $f-$type second order differential equation that can be transformed into exact second order differential equations. The fourth section is devoted for the concluding remarks.
\section{Solving Chebyshev's-type of Second Order Differential Equations}
In this section, we present an approach to solve Chebyshev's-type of second order differential equations
\begin{equation}\label{mainsec1}
p(x)y^{\prime\prime}(x)+\frac{1}{2}p^\prime(x)y^\prime(x)+f(\sqrt{p(x)}\;y^\prime(x),y(x))=0,
\end{equation}
where $p(x)$ is a positive and differentiable function on some open interval $\mathbb \mathbb (a,b) \in \R$, and $f(\sqrt{p(x)}\;\\y^\prime(x),y(x))$ is continuous function on some domain $D\subset \R^2$. The approach is described in the following theorem:
\begin{theorem}\label{maintheorem}
Assume that $p(x)$ be a positive and differentiable function on the open interval $\mathbb (a,b)\subset \R$. Let $x_0$ be any point in the interval $(a,b)$. Then
\[
t=\int_{x_0}^x\frac{d\xi}{\sqrt{p(\xi)}}
\]
transforms the differential equation ~\eqref{mainsec1} into the second order differential equation
\begin{equation}\label{main22}
y^{\prime\prime}(t)+f(y(t),y^\prime(t))=0.
\end{equation}
\end{theorem}
\begin{proof} Let $$t=\int_{x_0}^x\frac{d\xi}{\sqrt{p(\xi)}}.$$ Since,
\[
\frac{dy}{dx}=\frac{dy}{dt}\frac{dt}{dx}.
\]
Hence,
\begin{equation}\label{main4}
\frac{dy}{dt}= \sqrt{p(x)}\;\frac{dy}{dx}.
\end{equation}
Therefore,
\begin{equation}\label{main5}
\frac{d^2y}{dt^2}
=\frac{d}{dt}\left(\sqrt{p(x)}\;\frac{dy}{dx}\right)=\frac{d}{dx}\left(\sqrt{p(x)}\;\frac{dy}{dx}\right)\frac{dx}{dt}=p(x)y^{\prime\prime}+\frac{1}{2} p^\prime(x)y^\prime.
\end{equation}
By substituting ~\eqref{main4} and ~\eqref{main5} in Equation ~\eqref{main}, we get the result.\quad
\end{proof}
\begin{remark}\label{R2}
The differential equation ~\eqref{main22} is independent of the variable $t$, and so, it is easy to solve by setting $\eta(t)=y^\prime(t)$. Hence, it reduces into the following first order differential equation:
\begin{equation}\label{main3}
\eta\frac{d\eta}{dy}+f(y,\eta)=0.\
\end{equation}
In case that $f(\sqrt {p(x)}\;y^\prime,y)=f(y)$, we get
\[
\eta^2(t)=-2\int^y f(\xi)d\xi+c.
\]
Hence,
\[
y^\prime(t)=\left(c-2\int ^y f(\xi)d\xi\right)^{\frac{1}{2}}
\]
where $c$ is the integration constant.
\end{remark}
Next, we present some examples to explain this approach
\begin{example}
Consider the nonlinear Chebyshev's differential equation
\begin{equation}\label{hh}
\left\{
\begin{array}{ll}
(1 -x^2)y^{\prime\prime}(x)-x y^\prime(x) +4\sqrt{1-x^2}\;y^{\prime}y(x) = 0,\\
\\
y(0)=\frac{1}{2},\;y^\prime(0)=-\frac{1}{2}\,.
\end{array}
\right.
\end{equation}
Then
\[
t=\int^x\frac{d\xi}{\sqrt{1-\xi^2}}=\arcsin(x)
\]
transforms \eqref{hh} into
\[
y^{\prime\prime}(t)+4y^\prime(t) y(t)=0.
\]
Set $\eta(t)=y^\prime(t)$. The above equation becomes
\[
\eta \frac{d \eta}{dy}+4\eta y=0.
\]
The solution of this equation is $y(t)=\displaystyle\frac{1}{2(t+1)}.$ Therefore, $y(x)=\displaystyle\frac{1}{2(\arcsin(x)+1)}.$
\end{example}
\begin{example} Consider the initial value problem
\begin{equation}\label{ex2eq}
\left\{
\begin{array}{ll}
x^2y^{\prime\prime}+xy^\prime- 3y^2=0, \;x>0, \\
\;y(1) = 2, \;y^\prime(1) = 4.
\end{array}
\right.
\end{equation}
Then
$$t=\int_{1}^x \frac{d\xi}{\xi}d\xi=\ln(x)$$ transforms the \eqref{ex2eq} into
\begin{equation}\label{transone}
\left\{
\begin{array}{ll}
y^{\prime\prime}-3y^2=0,\\
y(0) = 2,\; y^\prime(0) = 4.
\end{array}
\right.
\end{equation}
The solution of the above differential equation is
\[
y (t)= \frac{2}{(1 - t)^2}.
\]
Hence,
\[
y(x) = \frac{2}{(1 - \ln(x))^2}\,.
\]
\end{example}
\begin{example}\label{ex3}
Consider the linear form of ~\eqref{main}
\begin{equation}\label{ex1eq}
\phi(x)y^{\prime\prime}+\frac{1}{2}\phi^\prime(x)y^\prime+\lambda^2 y=0,
\end{equation}
where $\lambda\in \R,$ and $\phi(x)$ is a positive and differentiable function on some open interval $\mathbb (a,b) \subset \R$. By applying the transformation
\[
t=\int_{x_0}^x \frac{d\xi}{\sqrt{\phi(\xi)}}d\xi,
\]
equation \eqref{ex1eq} can be transformed into the following second order differential equation:
$$\frac{d^2y}{dt^2}+\lambda^2 y=0.$$
The solution of this different5ial equation is
$$y(t)=C_1\sin(\lambda t)+C_2\cos(\lambda t).$$
Hence, the general solution of equation ~\eqref{ex1eq} is
$$y(x)=C_1\sin\left(\lambda\int_{x_0}^x \frac{d\xi}{\sqrt{\phi(\xi)}}d\xi \right)+C_2\cos\left(\lambda\int_{x_0}^x \frac{d\xi}{\sqrt{\phi(\xi)}}d\xi \right).$$
\end{example}
\begin{example}
Consider the following second order linear differential equation (see Eq. 239, p. 335 in \cite{MURPHY}):
\begin{equation}\label{part1ex4}
4xy^{\prime\prime}+2y^\prime+y=0.
\end{equation}
From the previous example, the general solution of this equation is given by
$$y(x)=C_1\sin\left(\int_{x_0}^x \frac{d\xi}{2\,\sqrt{\xi}}\,d\xi \right)+C_2\cos\left(\int_{x_0}^x \frac{d\xi}{2\,\sqrt{\xi}}\,d\xi \right)$$
and so,
\[
y(x)=C_1\sin\left(\sqrt{x}\right)+C_2\cos\left(\sqrt{x}\right).
\]
\end{example}
\begin{remark} \label{R3}
Consider the second order linear differential equation
\begin{equation}
(\phi(x))^2y^{\prime\prime}(x)+\phi(x)\phi^\prime(x)y^\prime(x)+\lambda y(x)=0, \;\;\; x \in (a,b),
\end{equation}
and assume that $\phi(x)$ is a positive and differentiable function on an open interval $(a,b)\subset \R.$ Moreover, assume that $\phi(a)=\phi(b)=0$. Define the linear differential operator
\[
L[y]:=-\left( (\phi(x))^2y^{\prime\prime}(x)+\phi(x)\phi^\prime(x)y^\prime(x))\right)=\lambda y(x).
\]
Then the boundary value problem
\begin{equation}\label{bvpp}
\left\{ \begin{array}{ll}
L[y]=-\left((\phi(x))^2y^{\prime\prime}(x)+\phi(x)\phi^\prime(x)y^\prime(x)\right)=\lambda y(x),& a<x<b,\\
\phi(a)=\phi(b)=0,
\end{array}
\right.
\end{equation}
satisfies the Lagrange Identity $\int_a^b \phi L[\psi]dx=\int_a^b \psi L[\phi]dx$, where $\phi$ and $\psi$ satisfy the above boundary value problem. Therefore, the operator $L[y]$ is self-adjoint. Hence, the boundary value problem \eqref{bvpp} has an orthogonal set of eigenfunctions $\{\phi_n(x)\}_{n=1}^\infty$ with corresponding eigenvalues $\{\lambda_n\}_{n=1}^\infty$. Since the above boundary value problem is a special case of \eqref{main}. Then, by using the approach described in Theorem \ref{maintheorem}, it is easy to find its orthogonal set of eigenfunctions.
\end{remark}
By using the same approach described in Theorem \ref{maintheorem}. We can solve the following class of second order linear differential equations:
\begin{equation}\label{mainlin1}
\left[P(x)\right]^2y^{\prime\prime}(x)+P(x)\left[\alpha+ P^\prime(x) \right]y^\prime(x)+\beta y(x)=0,
\end{equation}
where $P(x)>0$, $P(x)\in C^1(a,b)$, and $\alpha$ and $\beta$ are constants.
In fact, the transformation
\begin{equation} \label{transformation}
t=\displaystyle\int_{x_0}^x\frac{d\xi}{P(\xi)},
\end{equation}
where $x_0,\; x \in (a,b),$ transforms Eq. \eqref{mainlin1} into the following second order differential equation:
\[
y^{\prime\prime}(t)+\alpha y^\prime(t)+\beta y(t)=0.
\]
This differential equation is with constant coefficients which can be solved by using the elementary techniques of solving second order differential equations. For illustration, we present the following examples:
\begin{example}
Consider the well-known Cauchy-Euler's Equation
\[
x^2y^{\prime\prime}(x)+(\alpha +1) xy^\prime(x)+\beta y=0,\;x>0.
\]
Then $P(x)=x$, and the $t-$transformation is $t=\ln(x)$, which transforms the equation into
\[
y^{\prime\prime}(x)+\alpha y^\prime(x)+\beta y=0
\]
\end{example}
\begin{example}
Consider the Chebyshev's Equation
\[
\left[1-x^2\right]y^{\prime\prime}(x)-2xy^\prime(x)+n^2 y=0,\;|x|<1.
\]
Then $P(x)=\sqrt{1-x^2}$, and the $t-$transformation is $t=\sin^{-1}(x)$, which transforms the equation into
\[
y^{\prime\prime}(x)+n^2 y=0.
\]
Using this transformation, the solution of Chebyshev's Equation is given by
\[
y(x)=A\cos(n\sin^{-1}(x))+B\sin(n\sin^{-1}(x)).
\]
\end{example}
\begin{example}
Consider the Hypergeometric Equation
\begin{equation}\label{hypereq}
x(1-x)y^{\prime\prime}(x)+\frac{1}{2}(1-2x)y^\prime(x)+a^2 y=0,\;x\in(0,1).
\end{equation}
Then $P(x)=\sqrt{x(1-x)}$, and the $t-$transformation is $t=\sin^{-1}(2x-1).$ This transforms the equation into
\[
y^{\prime\prime}(t)+a^2 y(t)=0.
\]
Hence, the solution of \eqref{hypereq} is given by
\[
y(x)=A\cos(a\sin^{-1}(2x-1))+B\sin(a\sin^{-1}(2x-1)).
\]
\end{example}
For certain functions, $h(x)\in C(a,b)$, for some open interval $(a,b)\in \R$ , we can solve the nonhomogeneous second order differential equation
\begin{equation}\label{main33}
\left[P(x)\right]^2y^{\prime\prime}(x)+P(x)\left[\alpha+ P^\prime(x) \right]y^\prime(x)+\beta y=h(x).
\end{equation}
Particularly, when $h(x)$ can be written in the form $H(t),$ where $t=\int_{x_0}^x\frac{d\xi}{P(\xi)}.$ The following example shows this idea:
\begin{example}
Consider the nonhomogeneous differential equation
\begin{equation}\label{linexample}
x(1-x)y^{\prime\prime}(x)+\frac{1}{2}(1-2x)y^\prime(x)+a^2 y=2x,\;x\in(0,1).
\end{equation}
Then $P(x)=\sqrt{x(1-x)}$. The $t-$transformation is $t=\sin^{-1}(2x-1)$. This transforms the equation into
\[
y^{\prime\prime}(t)+a^2 y(t)=1+\sin(t).
\]
Hence, the solution of equation \eqref{linexample} is given by
\begin{equation*}
y(x)=\left\{
\begin{array}{ll}
A\cos(a\sin^{-1}(2x-1))+B\sin(a\sin^{-1}(2x-1))+\displaystyle \frac{2x-1}{a^2-1}+\displaystyle\frac{1}{a^2},\;\text{if}\;\;a\neq \pm 1,\\
A\cos(\sin^{-1}(2x-1))+B(2x-1)+\displaystyle \frac{1}{2}(1-2x)\sin^{-1}(2x-1)+1,\;\text{if}\;\;a= \pm 1.
\end{array}
\right.
\end{equation*}
\end{example}
\section{Solving $f-$type Second Order Differential Equations}
In this section, we solve the following class of second order nonlinear differential equation
\begin{equation}\label{part2eq1}
a_2\left(f^\prime(y)y^{\prime\prime}+(y^\prime)^2f^{\prime\prime}(y)\right)+a_1f^\prime(y)y^\prime+a_0f(y)=g(x),
\end{equation}
where $a_2,a_1$ and $a_0$ are constants, and $f\in C^2(a,b)$, for some open interval $(a,b) \subset \R$. In this section, we also solve the following class of second order nonlinear differential equation:
\begin{equation}\label{part2eq2}
p(x)\left(f^\prime(y)y^{\prime\prime}+(y^\prime)^2f^{\prime\prime}(y)\right)+\frac{1}{2}p^\prime(x)f^\prime(y)y^\prime+a_0f(y)=0,
\end{equation}
where $p(x)$ is a positive and differentiable function on an open interval $(a,b) \subset \R$, and $f\in C^2(c,d)$, for some open interval $(c,d)\subset \R$. To solve \eqref{part2eq1}, let $z=f(y)$. Hence, $z^\prime=f^\prime(y)y^\prime$, and $z^{\prime\prime}=f^\prime(y)y^{\prime\prime}+(y^\prime)^2f^{\prime\prime}(y).$ Substitute $z$, $z^\prime$ and $z^{\prime\prime}$ in equation \eqref{part2eq1}, we ge
\begin{equation}\label{part2eqcc}
a_2z^{\prime\prime}+a_1z^\prime+a_0z=g(x),
\end{equation}
Similarly, equation \eqref{part2eq2} becomes
\begin{equation}\label{part2eq2l}
p(x)z^{\prime\prime}+\frac{1}{2}p^\prime(x)z^\prime+a_0z=0,
\end{equation}
which is the linear form of \eqref{main}. Therefore, it can be solved by using the technique described in Example \ref{ex3}. To illustrate the procedure of solving \eqref{part2eq1} and \eqref{part2eq2}, we present the following examples
\begin{example}
Consider Langumir Equation, with a slightly modification,
\begin{equation}\label{part2ex1}
3yy^{\prime\prime}+3(y^\prime)^2+4yy^\prime+y^2=1.
\end{equation}
The original Langumir Equation is given by
\begin{equation*}
3yy^{\prime\prime}+(y^\prime)^2+4yy^\prime+y^2=1
\end{equation*}
which originally appears in connection with the theory of current flow from hot cathode to an anode in a hight vacuum \cite{Ames,Langmuir}. To solve \eqref{part2ex1}, we let $z=\displaystyle\frac{y^2}{2}$. Then $z^\prime=yy^\prime$ and $z^{\prime\prime}=yy^{\prime\prime}+(y^\prime)^2.$ Hence, equation \eqref{part2ex1} becomes
\[
3z^{\prime\prime}+4z^\prime+2z=1.
\]
The solution of this equation is $$z(x)=e^{-\frac{2}{3}x}\left(A\cos\left(\frac{\sqrt 2}{3}x\right)+B\sin\left(\frac{\sqrt 2}{3}x\right)\right)+\frac{1}{2}.$$ Hence, the solution of \eqref{part2ex1} is given by
\[
y^2=2e^{-\frac{2}{3}x}\left(A\cos\left(\frac{\sqrt 2}{3}x\right)+B\sin\left(\frac{\sqrt 2}{3}x\right)\right)+1.
\]
\end{example}
\begin{example}
Consider the initial value problem
\[
\left\{
\begin{array}{ll}
y^{\prime\prime}+(y^\prime)^2+1=(\cos \omega x)e^{-y},\; \;\omega\neq \pm 1,\\
y(0)=y^\prime(0)=0.
\end{array}
\right.
\]
This problem is equivalent to
\[
\left\{
\begin{array}{ll}
\left(y^{\prime\prime}+(y^\prime)^2\right)e^y+e^y=(\cos \omega x),\; \;\omega\neq \pm 1,\\
y(0)=y^\prime(0)=0.
\end{array}
\right.
\]
Let $z=e^y.$ Then $z^\prime=y^\prime e^y$ and $z^{\prime\prime}=y^{\prime\prime}e^y+(y^\prime)^2e^y$. By substituting $z$, $z^\prime$ and $z^{\prime\prime}$ in the above initial value problem, we ge
\[
\left\{
\begin{array}{ll}
z^{\prime\prime}+z=\cos \omega x,\;\; \omega \neq \pm 1,\\
z(0)=1,\;z^\prime(0)=0.
\end{array}
\right.
\]
The solution of this problem is $z(x)=\frac{1}{1-\omega^2}\left( \cos \omega x-\omega^2\cos x\right),\;\omega\neq \pm1.$ Therefore,
$y(x)=\ln \left(\frac{1}{1-\omega^2}\left(\cos \omega x-\omega^2\cos x \right)\right),\; \omega \neq \pm 1.$
\end{example}
\begin{example} Let $\phi(x)$ be a positive and differentiable function on an open interval $(a,b) \subset \R$, and consider the differential equation
\[
\phi(x)\left(y^{\prime\prime}+(y^\prime)^2\right)+\frac{1}{2}\phi^\prime(x) y^\prime +\lambda =0.
\]
By multiplying this equation by $e^y$, we get
\[
\phi(x)\left(y^{\prime\prime}+(y^\prime)^2\right)e^y+\frac{1}{2}\phi^\prime(x)y^\prime e^y+\lambda e^y=0.
\]
Let $z=e^y.$ Then $z^\prime=y^\prime e^y$ and $z^{\prime\prime}=y^{\prime\prime}e^y+(y^\prime)^2e^y$. By substituting $z$, $z^\prime$ and $z^{\prime\prime}$ in the above differential equation, we get
\[
\phi(x)z^{\prime\prime}+\frac{1}{2}\phi^\prime(x)z^\prime+\lambda z=0.
\]
The solution of this equation is (see Example \ref{ex3})
\[
z(x)=C_1\sin\left(\lambda\int_{x_0}^x \frac{d\xi}{\sqrt{\phi(\xi)}}d\xi \right)+C_2\cos\left(\lambda\int_{x_0}^x \frac{d\xi}{\sqrt{\phi(\xi)}}d\xi \right).
\]
Therefore,
\[
y(x)=\ln\left[C_1\sin\left(\lambda\int_{x_0}^x \frac{d\xi}{\sqrt{\phi(\xi)}}d\xi \right)+C_2\cos\left(\lambda\int_{x_0}^x \frac{d\xi}{\sqrt{\phi(\xi)}}d\xi \right)\right].
\]
\end{example}
\section{Solving $f$-type Second Order Differential Equations that can be Transformed into Exact Second Order Differential Equations}
In this section, we solve the following class of second order nonlinear differential equations:
\begin{equation}\label{part3eq1}
a_2(x,y,y^\prime)\left(f^\prime(y)y^{\prime\prime}+f^{\prime\prime}(y)(y^\prime)^2\right)+a_1(x,y,y^\prime)(f^\prime(y)y^\prime)+a_0(x,y,y^\prime)=0,
\end{equation}
where $f(y)$ is an invertible function and $f\in C^2(a,b)$. To solve this class of differential equations, we let $z=f(y)$. Then $z^\prime=f^\prime(y)y^\prime$ and $z^{\prime\prime}=f^{\prime\prime}(y)(y^\prime)^2+f^\prime(y)y^{\prime\prime}.$ Moreover, we let $y=f^{-1}(z).$ Then $y^\prime=\frac{z^\prime}{f^{\prime}\left(f^{-1}(z)\right)}.$ Hence, equation \eqref{part3eq1} can be transformed into the following differential equation:
\begin{equation}\label{part3eq2}
a_2\left(x,f^{-1}(z),\frac{z^\prime}{f^\prime\left(f^{-1}(z)\right)}\right)z^{\prime\prime}+a_1\left(x,f^{-1}(z),\frac{z^\prime}{f^\prime\left(f^{-1}(z)\right)}\right)z^\prime+a_0\left(x,f^{-1}(z),\frac{z^\prime}{f^\prime\left(f^{-1}(z)\right)}\right)=0
\end{equation}
Assume that \eqref{part3eq2} is exact, then it can be solved. To explain the procedure of solving such differential equations, we consider the following example:
\begin{example}
Consider the second order nonlinear differential equation
\begin{equation}\label{example6}
\left\{
\begin{array}{ll}
e^y\left[y^{\prime\prime}+(y^\prime)^2\right]+12xe^{4y}y^\prime+\left(3e^{4y}-1\right)=0,\\
y(0)=\ln2,\;y^\prime(0)=0.
\end{array}
\right.
\end{equation}
Let $z=e^y$. Then $z^\prime=e^{y}y^\prime$ and $z^{\prime\prime}=e^{y}y^{\prime\prime}+e^y(y^\prime)^2.$ Hence, Eq. \eqref{example6} becomes
\begin{equation}\label{ex6exact}
\left\{
\begin{array}{ll}
z^{\prime\prime}+12xz^3z^\prime+\left(3z^4-1\right)=0,\\
z(0)=2,\;z^\prime(0)=0.
\end{array}
\right.
\end{equation}
Therefore, $a_2(x,z,z^\prime)=1,$ $a_1(x,z,z^\prime)=12xz^3,$ and $a_0(x,z,z^\prime)=\left(3z^4-1\right).$ In addition, we have
\begin{equation}\label{exactcond22}
\frac{\partial a_2}{\partial z}=\frac{\partial a_1}{\partial z^\prime}=0, \;\; \frac{\partial a_2}{\partial x}=\frac{\partial a_0}{\partial z^\prime}=0,\; \text{and}\; \;\frac{\partial a_1}{\partial x}=\frac{\partial a_0}{\partial z}=12z^3,
\end{equation}
Therefore, equation \eqref{ex6exact} is exact differential equation. Hence, its first integral exists and it is given by
\[
z^\prime+3xz^4-x=0.
\]
For which an implicit solution of this equation can be obtained by separating the variables, and so, $y(x)=\ln(z(x)).$
\end{example}
\begin{remark}
Assume that \eqref{part3eq2} is not exact. Then an integrating factor of \eqref{part3eq2} could be exist. Hence, it can be transformed into an exact differential equation (see \cite{AlAhmad}).
To explain the procedure of solving \eqref{part3eq2} in case it is not exact, we present the following example:
\end{remark}
\begin{example} Consider the second order nonlinear differential equation
\begin{equation}\label{part3ex21}
xe^y\left(2x+e^y\right)\left(y^{\prime\prime}+(y^\prime)^2\right)+x\left(x+e^y\right)y^\prime+\left(3x+e^y\right)=0.
\end{equation}
By multiplying this equation by $e^y$, we get
\begin{equation}\label{part3ex22}
xe^{2y}\left(2x+e^y\right)\left(y^{\prime\prime}+(y^\prime)^2\right)+x\left(x+e^y\right)e^yy^\prime+e^y\left(3x+e^y\right)=0.
\end{equation}
Let $z=e^y$. Then $z^\prime=e^{y}y^\prime$ and $z^{\prime\prime}=e^{y}y^{\prime\prime}+e^y(y^\prime)^2.$ Hence, by substituting $z$, $z^\prime$ and $z^{\prime\prime}$ in \eqref{part3ex22}, we get
\begin{equation}\label{part3ex23}
xz(2x+z)z^{\prime\prime}+x(x+z)z^\prime+z(3x+z)=0.
\end{equation}
This equation is not exact since $\frac{\partial a_2}{\partial z}=2(x+z)\neq0=\frac{\partial a_1}{\partial z^\prime}$. An integrating factor of this second order nonlinear differential equation exists, and it is given by $\mu(x,z)=\displaystyle \frac{1}{xz(2x+z)}.$ Multiplying \eqref{part3ex23} by $\mu(x,z)$, we get
\begin{equation}\label{part3ex24}
z^{\prime\prime}+\frac{(x+z)}{z(2x+z)}z^\prime+\frac{(3x+z)}{x(2x+z)}=0.
\end{equation}
Clearly,
\begin{equation}
\frac{\partial a_2}{\partial z}=\frac{\partial a_1}{\partial z^\prime}=0, \;\; \frac{\partial a_2}{\partial x}=\frac{\partial a_0}{\partial z^\prime}=0,\; \text{and}\; \;\frac{\partial a_1}{\partial x}=\frac{\partial a_0}{\partial z}=\frac{-1}{(2x+z)^2}\,.
\end{equation}
Therefore, the differential equation \eqref{part3ex24} is exact, and its first integral is given by
\begin{equation}
c= z^\prime+\ln\left(xz\sqrt{2x+z}\right).
\end{equation}
This first order differential equation can be solved by using the elementary techniques of solving first order differential equations. Hence, $y(x)=\ln(z(x))$.
\end{example}
Finally, we consider the nonhomogeneous second order linear differential equation
\[
a_2(x)y^{\prime\prime}+a_1(x)y^\prime+a_0(x)y=h(x),
\]
where $a_2(x)\neq 0$, $a_1(x)$, and $a_0(x)$ are differentiable functions on an open interval $(a,b) \subset \R$. This equation admits an integrating factor $\mu(x)=\displaystyle\frac{1}{a_2(x)}$ provided that $W(a_2,a_1)(x)=a_0(x)a_2(x),$ where $W(a_2,a_1)(x)=a_2(x)a^\prime_1(x)-a_1(x)a^\prime_2(x)$. For this case, we present the following example:
\begin{example}
consider the second order linear differential equation
\[
e^x y^{\prime\prime}+\cos x y^\prime -(\cos x+\sin x)y=h(x).
\]
By multiplying this equation by the integrating factor $e^{-x}$, we get
\[
y^{\prime\prime}+e^{-x}\cos x y^\prime -e^{-x}(\cos x+\sin x)y=h(x)e^{-x}.
\]
This equation can be written as
\[
\frac{d}{dx}\left[y^\prime+(e^{-x}\cos x) y\right]=h(x)e^{-x}
\]
Hence, its first integral is given by
\[
y^\prime+(e^{-x}\cos x) y=\displaystyle\int^x h(\xi)e^{-\xi}d\xi+c_1
\]
which can be solved by using the elementary techniques of solving first order differential equations.
\end{example}
\section{Concluding Remarks}
In this paper, we solved some classes of second order differential equation. In fact, we solved the following classes of second order differential equations:
\begin{enumerate}
\item The Chebyshev's type of second order differential equation \begin{equation}\label{CRmain}
p(x)y^{\prime\prime}(x)+\frac{1}{2}p^\prime(x)y^\prime(x)+f(\sqrt{p(x)}\;y^\prime(x),y(x))=0,\;\;\; x\in (a,b),
\end{equation}
where $p(x)$ is a positive and differentiable function on an open interval $ (a,b)\subset \R$, and $f(\sqrt{p(x)}\;y^\prime(x),y(x))$ is a continuous function on some domain $D\subset \R^2$.
\item The $f-$type of second order differential equations
\begin{itemize}
\item [a)] \begin{equation}\label{CRpart2eq}
a_2\left(f^\prime(y)y^{\prime\prime}+(y^\prime)^2f^{\prime\prime}(y)\right)+a_1f^\prime(y)y^\prime+a_0f(y)=g(x),
\end{equation}
where $a_2,a_1$ and $a_0$ are constants, and the function $f(y)$ is of $C^2-$class on some open interval $ (a,b) \subset \R$, and
\item [b)]
\begin{equation}\label{CRpart2eeq}
p(x)\left(f^\prime(y)y^{\prime\prime}+(y^\prime)^2f^{\prime\prime}(y)\right)+\frac{1}{2}p^\prime(x)f^\prime(y)y^\prime+a_0f(y)=0,
\end{equation}
where $p(x)$ is a positive and differentiable function on some open interval $(a,b) \subset \R$, and $f\in C^2(c,d)$, for some open interval $(c,d) \subset \R$.
\end{itemize}
\item $f$-type second order differential equations that can be transformed into exact second order Differential Equations
\begin{equation}
a_2(x,y,y^\prime)\left(f^\prime(y)y^{\prime\prime}+f^{\prime\prime}(y)(y^\prime)^2\right)+a_1(x,y,y^\prime)(f^\prime(y)y^\prime)+a_0(x,y,y^\prime)=0,
\end{equation}
where the function $f(y)$ is an invertible function and $f\in C^2(a,b)$, for some open interval $(a,b) \subset \R$.
\end{enumerate}
Moreover, we presented some examples to explain our approach of solving the above classes of second order differential equation.
| c1b232d542083d5e05b1c163b06371871b0ba911 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:Introduction}
Over the past decades, biometric authentication systems have gained popularity due to the reliability and adaptability. Existing biometric authentication systems generally include physiological and behavioral ones. The former is based on individuals' unique intrinsic features (e.g., face \cite{givens2013biometric}, iris \cite{latman2013field}, retina \cite{sadikoglu2016biometric}, voice \cite{goldstein2016methods}, and fingerprint \cite{unar2014review}) and the latter is based on individuals' behavior patterns such as gait analysis \cite{boulgouris2013gait}.
Recently, biometrics (e.g., fingerprint, face) based authentication systems face an increasing threat of being deceived as a result of the rapid development of manufacturing industry and technologies. For example, individuals can easily trick a fingerprint-based authentication system by using a fake fingerprint film\footnote{\url{http://www.instructables.com/id/How-To-Fool-a-Fingerprint-Security\-System-As-Easy-/}}
or an expensive face recognition-based authentication systems by simply wearing a two-hundred-dollar
anti-surveillance mask\footnote{\url{http://www.urmesurveillance.com/urme-prosthetic/}}.
Thus, fake-resistance characteristics are becoming a more significant requirement for any authentication system.
To address the aforementioned issues, EEG (Electroencephalography) signal-based cognitive biometrics and gait-based systems have been attracting increasing attention.
\begin{table}[t]
\centering
\caption{Comparison of various biometrics. EEG and Gait have considerable fake-resistance which is the most significant characteristic of authentication systems. $\uparrow$ denotes the higher the better while $\downarrow$ denotes the lower the better.}
\label{tab:comparison_1}
\resizebox{\textwidth}{!}{\begin{tabular}{lllllllll}
\hline
& \textbf{Biometrics} & \begin{tabular}[c]{@{}l@{}}\textbf{Fake-}\\ \textbf{-resistance}\end{tabular} $\uparrow$ & \textbf{Universality} $\uparrow$ & \textbf{Uniqueness} $\uparrow$ & \textbf{Stability} $\uparrow$ & \textbf{Accessibility} $\uparrow$ & \textbf{Performance} $\uparrow$ & \begin{tabular}[c]{@{}l@{}}\textbf{Computational}\\ \textbf{Cost}\end{tabular} $\downarrow$ \\ \hline
\multirow{8}{*}{\begin{tabular}[c]{@{}l@{}}\textbf{Uni-}\\ \textbf{-modal}\end{tabular}} & Face/Vedio & Medium & Medium & Low & Low & High & Low & High \\
& Fingerprint/Palmprint & Low & High & High & High & Medium & High & Medium \\
& Iris & Medium & High & High & High & Medium & High & High \\
& Retina & High & Medium & High & Medium & Low & High & High \\
& Signature & Low & High & Low & Low & High & Low & Medium \\
& Voice & Low & Medium & Low & Low & Medium & Low & Low \\
& Gait & High & Medium & High & Medium & Medium & High & Low \\
& EEG & High & Low & High & Low & Low & High & Low \\ \hline
\multirow{6}{*}{\begin{tabular}[c]{@{}l@{}}\textbf{Mulit-}\\ \textbf{-modal}\end{tabular}} & Fingerprint+face & Low & High & Low & Low & High & Medium & High \\
& Iris+pupil & Medium & Medium & High & High & Medium & High & High \\
& Iris+face & Medium & High & Medium & Medium & Medium & Medium & High \\
& ECG+fingerprint & High & Medium & Medium & High & Low & High & Medium \\
\hline
& \textbf{ EEG+gait}& \textbf{ High} & \textbf{ Low} & \textbf{ High} & \textbf{ High} & \textbf{ Low} & \textbf{ High} & \textbf{ Low} \\ \hline
\end{tabular}
}
\end{table}
EEG signal-based authentication systems are an emerging approach in physiological biometrics.
EEG signals measure the brain's response and record the electromagnetic, invisible, and untouchable electrical neural oscillations. Many research efforts have been made on EEG-based biometric authentication for the uniqueness and reliability .
EEG data are unique for each person and almost impossible to be cloned and duplicated. Therefore, an EEG-based authentication system has the potential to uniquely identify humans and ingenious enough to protect against faked identities~\cite{chuang2013think}.
For instance, Chuang et al. \cite{chuang2013think}propose a single-channel EEG-based authentication system, which achieves an accuracy of 0.99. Sarineh Keshishzadeh et al. \cite{keshishzadeh2016improved} employ a statistical model for analyzing EEG signals and achieves an accuracy of 0.974. Generally, EEG signals have the following inherent advantages:
\begin{itemize}
\item \textit{Fake-resistibility.} EEG data are unique for each person and almost impossible to be cloned and duplicated.
EEG signals are individual-dependent. Therefore, an EEG-based authentication system has the potential to verify human identity and ingenious enough to protect against faked identities\cite{chuang2013think}.
\item \textit{Reliability.} An EEG-based authentication system can prevent the subjects under abnormal situations (e.g., dramatically spiritual fluctuating, hysterical, drunk, or under threaten) since EEG signals are sensitive to human stress and mood.
\item \textit{Feasibility.} We have seen an important trend to build authentication systems based on EEG because the equipment for collecting EEG data is cheap and easy to acquire, and it is expected to be more precise, accessible, and economical in the future.
\end{itemize}
In comparison, gait-based authentication systems have been an active direction for years \cite{yao2018compressive,qian2018enabling}. Gait data are more generic and can be gathered easily from popular inertial sensors. Gait data are also unique because they are determined by intrinsic factors (e.g., gender, height, and limb length),
temporal factors \cite{callisaya2009population} (e.g., step length, walking speed, and cycle time) and kinematic factors (e.g., joint rotation of the hip, knee, and ankle, mean joint angles of the hip/knee/ankle, and thigh/trunk/foot angles). In addition, a person's gait behavior is established inherently in the long-term and therefore difficult to be faked.
Hoang et al. \cite{hoang2014secure} propose a gait-based authentication biometric system to analyze gait data gathered by mobile devices, adopt error correcting codes to process the variation in gait measurement, and finally achieve a false acceptance rate (FAR) of 3.92\% and a false rejection rate (FRR) of 11.76\%.
Cola et al. \cite{cola2016gait} collect wrist signals and train gait patterns to detect invalid subjects (unauthenticated people). The proposed method achieves an equal error rate (EER) of 2.9\%.
Despite the tremendous efforts,
various other challenges still remain in single EEG/gait based authentication systems:
(i) the solo EEG/gait authentication system generally obtains a False Acceptance Rate (FAR, which is extremely crucial in high-confidential authentication scenarios)
higher than $3\%$. It is not precise enough for highly confidential places such as military bases, the treasuries of banks and political offices where tiny misjudges could provoke great economic or political catastrophes; (ii) the single authentication system may break down under attack but no backup plan is provided; (iii) the solo EEG-based authentication system is easy to be corrupted by environmental factors (e.g., noise) and subjective factors (e.g., mental state) although it has high rake-resistance; (iv) the solo gait-based authentication system has relatively low performance although it is more stable over different scenarios.
In this paper, we propose Deepkey, a novel biometric authentication system that enables dual-authentication leveraging on the advantages of both gait-based and EEG-based systems. Compared with either gait-based or EEG-based authentication system, a dual-authentication system offers more reliable and precise identification.
Table~\ref{tab:comparison_1} summarizes the overall comparison of Deepkey with some representative works on seven key aspects.
Deepkey consists of three main components: the Invalid ID Filter Model to eliminate invalid subjects, the EEG Identification Model to identify EEG IDs, and the Gait Identification Model to identify gait IDs.
An individual is granted access only after she/he passes all the authentication components. Our main contributions
are highlighted as follows:
\begin{itemize}
\item We present Deepkey, a dual-authentication system that exploits both EEG and gait biological traits. To the best of our knowledge, Deepkey is the first two-factor authentication system for person authentication
using EEG and gaits.
Deepkey is empowered high-level fake-resistance and reliability because both EEG and gait signals are invisible and
hard to be reproduced.
\item We design a robust framework that includes an attention-based RNN to detect and classify of multimodal sensor data, and to decode the large diversity in how people perform gaits and brain activities simultaneously. The delta band of EEG data is decomposed for the rich discriminative information.
\item We validate and evaluate Deepkey on several locally collected datasets. The results show that Deepkey significantly outperforms a series of baseline models and the state-of-the-art methods, achieving FAR of 0 and FRR of $1\%$. Further, we design extensive experiments to investigate the impact of key elements.
\end{itemize}
The remainder of this paper is organized as follows. Section~\ref{sec:Related Work} introduces the EEG-based, gait-based, and
multimodal biometric systems briefly. Section~\ref{sec:The Proposed approach} present the methodology framework and three key models (Invalid ID Filter Model, Gait Identification Model, and EEG Identification Model) of the DeepKey authentication system in detail. Section~\ref{sec:experiment_and_results} evaluates the proposed approach on the public Gait and EEG dataset and provides analysis of the experimental results. Finally, Section~\ref{sec:discussions_and_future_work} discusses the opening challenges of this work and highlight the future scope while Section~\ref{sec:conclusion} summarizes the key points this paper.
\section{Related Work}
\label{sec:Related Work}
In this section, we introduce the related studies on several topics: biometric authentication technologies, EEG-based authentication, gait-based authentication, and multimodal biometric authentication.
\subsection{Biometric authentication technologies}
\label{sec:biometric_authentication_technologies}
Since biometric features cannot be stolen or duplicated easily, biometric authentication is becoming increasingly a commonplace. Currently, the most mature biometric authentication technology is fingerprint-based authentication which has been demonstrated to high matching accuracy and been used for decades \cite{maio2002fvc2002}. Iris recognition is another popular approach for biometric authentication owing to its unique and stable pattern \cite{pillai2014cross}.
In 1993, Daugman \cite{daugman1993high} proposes to use Gabor phase information and Hamming distance for iris code matching, which still is the most classic iris recognition method. Based on \cite{daugman1993high}, a flurry of research \cite{pillai2014cross} has emerged offering solutions to ameliorate iris authentication problems.
For example, Pillai et al. \cite{pillai2014cross} introduce kernel functions to represent transformations of iris biometrics. This method restrains both the intra-class and inter-class variability to solve the sensor mismatch problem. Face recognition techniques \cite{guzman2013thermal, drira20133d, zhou2014recent} is the most commonly used and accepted by the public for its unique features and non-intrusiveness. Since face recognition systems require tackling different challenges including expression, image quality, illumination, and disguise to achieve high accuracy, infrared (IR) \cite{guzman2013thermal} and 3D \cite{drira20133d} systems have attracted much attention. According to \cite{zhou2014recent}, multimodal recognition combining traditional visual textual features and IR or 3D systems can achieve higher accuracy than single modal systems. \par
\subsection{EEG-based Authentication}
\label{sec:eeg_based_authentication}
Since EEG can be gathered in a safe and non-intrusive way, researchers have paid great attention to exploring this kind of brain signals. For person authentication, EEG is, on the one hand, promising for being confidential and fake-resistant but, on the other hand, complex and hard to be analyzed. Marcel and Mill{\'a}n \cite{marcel2007person} use Gaussian Mixture Models and train client models with maximum A Posteriori (MAP). Ashby et al. \cite{ashby2011low} extract five sets of features from EEG electrodes and inter-hemispheric data, combine them together, and process the final features with support vector machine (SVM). The study shows that EEG authentication is also feasible with less-expensive devices. Altahat et al. \cite{altahat2015analysing} select Power Spectral Density (PSD) as the feature instead of the widely used autoregressive (AR) models to achieve higher accuracy. They also conduct channel selection to determine the contributing channels among all 64 channels. Thomas and Vinod \cite{thomas2016utilizing} take advantage of individual alpha frequency (IAF) and delta band signals to compose specific feature vector. They also prefer PSD features but only perform the extraction merely on gamma band.
\subsection{Gait-based Authentication}
\label{sec:gait_based_authentication}
As the most basic activity in our daily lives, walking is an advanced research hotspot for activity recognition \cite{yao2016learning, huang2016exploiting, mennicken2016s}. Differing from previous work, our work focuses on human gait, a spatio-temporal biometric that measures a person's manner
on walking. On the one hand, gait can be collected remotely without human interaction
compared
to
other aforementioned biometric features \cite{wang2012human}.
On the other hand, it is challenging to eliminate the influence of exterior factors including clothing, walking surface, shoes, carrying stuff, and other environmental factors. Existing gait recognition approaches sit in two categories. One is {\em model-based approach} \cite{nixon2009model}, which models gait information with mathematical structures, and the other is
{\em appearance-based approach}, which extracts features in a straightforward way irrespective of the mathematical structure.
Due to its high efficiency and remarkable performance, Gait Energy Image (GEI) \cite{man2006individual} has become one of the most popular appearance-based methods in recent years. Based on GEIs,
conducts enhancing process and autocorrelation on the contour profile. Besides, the cross-view variance is also a concern of gait identification \cite{kusakunniran2013new, kusakunniran2014recognizing}. For example, Wu et al. \cite{wu2017comprehensive} consider not only the cross-view variance but also deep convolutional neural networks (CNNs) for robust gait identification.
\subsection{Multimodal Biometric Authentication}
\label{sec:multimodal_biometric_authentication}
Since traditional unimodal authentication suffers from the negative influence of loud noise, low universality, and intra-class variation, it cannot achieve higher accuracy in a wide range of applications. To address this issue, multimodal biometric authentication
which combines and uses biometric traits in different ways, is becoming popular. Taking the commonness into consideration, most works choose two biometrics from face, iris, and fingerprints and make the fusion \cite{telgad2014combination, khiari2016quality, kumar2013feature}. In \cite{derawi2014fusion}, an innovative combination between gait and electrocardiogram (ECG) is shown to be effective. Manjunathswamy et al. \cite{manjunathswamy2015multimodal} combine ECG and fingerprint at the score level.
To the best of our knowledge, the approach proposed in this paper is the first
to combine EEG and gaits for person authentication. Taking advantages of both EEG and gait signals, the combination is expected to improve the reliability of authentication systems.
This dual-authentication system is partially based on our previous work MindID \cite{zhang2018mindid} which is an EEG-based identification system. We emphasize several difference compared to \cite{zhang2018mindid}: 1) this work is an authentication system with an invalid ID filter while \cite{zhang2018mindid} only focuses on identification; 2) this work adopt two biometrics including EEG and gait while \cite{zhang2018mindid} only exploit EEG signals; 3) this work conduct extensive real world experiments to collect gait signals.
\section{Deepkey Authentication System}
\label{sec:The Proposed approach}
In this section, we first report the workflow of Deepkey
to give
an overview of the authentication algorithm, and then present the technical details for each component.
\subsection{Deepkey System Overview}
\label{sec:system_overview}
\begin{figure}
\includegraphics[width=\linewidth]{figure/scenario.pdf}
\caption{Workflow of Deepkey authentication system. The data collection of EEG and gait are cascade.}
\label{fig:scenario}
\end{figure}
The Deepkey system is supposed to be deployed in access to confidential locations (e.g., bank vouchers, military bases, and government confidential residences). As shown in Figure~\ref{fig:scenario}, the overall workflow of the Deepkey authentication system consists of the following four steps:
\begin{enumerate}
\item Step 1: EEG data collection. The subject, who
requests for authentication, is required to wear the EEG headset and stays in a relaxed state.
The EEG data ($\mathcal{E}$) will typically take 2 seconds.
\item Step 2: Gait data collection. The subject takes off the EEG headset and puts three IMUs (Inertial Measurement Unit) and walks through an aisle to collect gait data $\mathcal{G}$ by IMUs.
\item Step 3: Authentication. The gathered EEG and gait data are flatten and associated with input data $\mathcal{I} =[\mathcal{E}:\mathcal{G}]$ to be fed into the Deepkey authentication algorithm.
\item Step 4: Decision. An \textit{Approve} or \textit{Deny} decision will be made according to the Deepkey authentication results.
\end{enumerate}
The most crucial component among above the steps is the third step, where the Deepkey authentication system receives the associated input data $\mathcal{I}$ and accomplishes two goals: authentication and identification.
For the former goal, we employ EEG signals to justify the impostor for its high fake-resistance. EEG signals are invisible and unique, making them difficult to be duplicated and hacked.
For the latter goal, we adopt a deep learning model to extract the distinctive features and feed them into a non-parametric neighbor-based classifier for ID identification. In summary, the Deepkey authentication algorithm contains several key stages, namely {\em Invalid ID Filter}, {\em Gait-based Identification}, {\em EEG-based Identification} and {\em Decision Making}.
The overall authentication contains the following several stages (Figure~\ref{fig:attention}):
\begin{enumerate}
\item Based on the EEG data, the Invalid ID Filter decides the subject is an impostor or a genuine. If the subject is an impostor, the request will be denied.
\item If the individual is determined as genuine, the EEG/Gait Identification Model will identify the individual's authorized EEG/Gait ID. This model is pre-trained off-line with the attention-based Long Short-Term Member (LSTM) model (Section~\ref{sec:eeg_id_identification_model}). The output is the ID number associated with the person's detailed personal information.
\item The final stage is to check the consistency of the EEG ID and the Gait ID. If they are identical,
the system will grant an approval, otherwise, deny the subject and take corresponding security measures.
\end{enumerate}
\begin{table}%
\caption{Notation}
\label{tab:notation}
\begin{minipage}{\columnwidth}
\begin{center}
\begin{tabular}{ll}
\toprule
\textbf Parameters& \textbf Explanation\\
\hline
$L^{\circ}$ & the number of authorized subjects (genuine)\\
$E$ & the set of EEG signals \\
$G$ & the set of gait signals \\
$n_s$ & the number of features in input data sample\\
$G_i$ & the $i$-th gait data\\
$n_g$ & the number of features in per gait sample\\
$E_i$ & the $i$-th EEG data\\
$n_e$ & the number of features in EEG data sample\\
$X^i$ & the data in the $i$-th layer of attention RNN\\
$N^i$ &the number of dimensions in $X^i$\\
$K$ & the number of participants (genuine and impostor) \\
$c^{i'}_{j} $ & the hidden state in the $j$-th LSTM cell \\
$\mathcal{T}(\cdot)$ & the linear calculation of dense neural layers \\
$\mathcal{L}(\cdot)$ & The LSTM calculation process \\
$\odot$ & the element\-wise multiplication \\
$C_{att}$ & attention based code \\
\bottomrule
\end{tabular}
\end{center}
\end{minipage}
\end{table}%
\begin{figure}[t]
\includegraphics[width=0.8\linewidth]{figure/EEG_topo.pdf}
\caption{EEG topography of different subjects under different frequency bands. The inter-subject EEG signal cosine-similarity is calculated under each band and the results are reported as $0.1313$ (full bands), $0.0722$ (Delta band), $0.1672$ (Theta band), $0.2819$ (Alpha band), $0.0888$ (Beta band), and $0.082$ (Gamma band). This illustrates that the delta band with the lowest similarity contains the most distinguishable features for person identification.}
\label{fig:topo}
\end{figure}
\subsection{Invalid ID Filter Model}
\label{sec:invalid_detection_model}
Since the high fake-resistance of EEG data and rich distinguishable features in EEG signal, which lead to competitive invalid filter performance, in the Invalid ID Filter Model, only the EEG data $E_i$ is utilized to detect the genuine\footnote{Through our preliminary experiments, the gait signal is not effectiveness enough for invalid ID filtering.}.
The subjects in an authentication system are categorized into two classes: {\em authorized} and {\em unauthorized}. Since the unauthorized data are not available in training stage, therefore, an unsupervised learning algorithm is required to identify the invalid ID. In this work, we apply one-class SVM to sort out the unauthorized subjects.
Given a set of authorized subjects $\mathcal{S}=\left \{S_{i},i=1,2,\cdots,L^{\circ}\right \}$,
$S_{i}\in R^{n_s}$, where $L^{\circ}$ denotes the number of authorized subjects and $n_s$ denotes the number of dimensions
of the input data.
The input data consist of
EEG data $\mathcal{E}=\left \{E_{i},i=1,2,\cdots,L^{\circ}\right \}$, $E_{i}\in R^{n_e}$ and gait data $\mathcal{G}=\left \{ G_{i},i=1,2,\cdots,L^{\circ}\right \}, G_{i}\in R^{n_g}$.
$n_g$ and $n_e$ denote the number of dimensions
of the
gait data and EEG data, respectively, and
$n_s=n_g+n_e$. The notation can be found in Table~\ref{tab:notation}.
For each authentication, the collected EEG data $E_i$ includes a number of samples. Each sample is a vector with shape $[1,14]$ where $14$ denotes the number of electric-nodes in Emotiv headset. To trade-off the authentication efficiency (less collecting and waiting time) and
computational performance,
based on the experimental experience, we fed 200 samples ([200, 14]) into the Invalid ID Filter. 200 EEG samples are collected in 1.56 seconds which is acceptable.
The final filter result is the mean of the results on all the samples.
\begin{table*}[!tb]
\centering
\caption{Characteristics of EEG frequency bands. Awareness Degree denotes the degree of being aware of an external world.}
\label{tab:bands}
\resizebox{\textwidth}{!}{
\begin{tabular}{llllll}
\hline
\textbf{Name} & \textbf{Frequency ($Hz$)} & \textbf{Amplitude} & \textbf{Brain State} & \textbf{Awareness Degree} & \textbf{Produced Location} \\ \hline
\textbf{Delta} & 0.5-3.5 & Higher & Deep sleep pattern & Lower & Frontally and posteriorly \\
\textbf{Theta} & 4-8 & High & Light sleep pattern & Low & Entorhinal cortex, hippocampus \\
\textbf{Alpha} & 8-12 & Medium & Closing the eyes, relax state & Medium & Posterior regions of head \\
\textbf{Beta} & 12-30 & Low & \begin{tabular}[c]{@{}l@{}}Active thinking, focus, \\high alert, anxious \end{tabular} & High & Most evident frontally \\
\textbf{Gamma} & 30-100 & Lower & \begin{tabular}[c]{@{}l@{}}During cross-modal \\sensory processing \end{tabular} & Higher & Somatosensory cortex\\ \hline
\end{tabular}
}
\end{table*}
\begin{figure}[t]
\includegraphics[width=0.8\linewidth]{figure/workflow.pdf}
\caption{Authentication workflow. If the input data can not pass the invalid ID filter, it would directly regarded as an impostor and deny access. If pass, the Delta pattern and gait signals are parallelly fed into an attention-based RNN structure to study the distinctive features $C_{att}$. The learned features are classified by the EEG and Gait classifier in order to identify the subject's EEG and Gait ID. The subject is approved only if the EEG ID is match with Gait ID.}
\label{fig:attention}
\end{figure}
\subsection{EEG Identification Model}
\label{sec:eeg_id_identification_model}
Compared to
gait data,
EEG data contain more noise which is more challenging to handle. Given the complexity of EEG signals, the data pre-processing is necessary. In practical EEG data analysis, the assembled EEG signals can be divided into several different frequency patterns (delta, theta, alpha, beta, and gamma) based on the strong intra-band correlation with a distinct behavioral state.
The EEG frequency patterns and the corresponding characters are listed in Table~\ref{tab:bands} \cite{zhang2018mindid}. Figure~\ref{fig:topo} reports the topography of EEG signals of different subjects under different frequency bands and demonstrates that the Delta wave, compared to other bands, enriches distinctive features. In detail, we calculate the inter-subject EEG signal cosine-similarity which measure the average similarity among different subjects under all the EEG bands. The results are reported as 0.1313 (full bands), 0.0722(Delta band), 0.1672 (Theta band), 0.2819 (Alpha band), 0.0888 (Beta band), and 0.082 (Gamma band). This
illustrates that the delta band with the lowest similarity contains the most distinguishable features for person
identification. Our previous work \cite{zhang2018mindid} has demonstrated that Delta pattern, compare to other EEG patterns contains the most distinctive information and is the most stable pattern in different environments by qualitative analysis and empirical experiment results. Thus, in this paper, we adopt a bandpass (0.5Hz-3.5Hz) butter-worth filter to extract Delta wave signal for further authentication. For simplicity, we denote the filtered EEG data as $\mathcal{E}$.
Since different EEG channels record different aspects of the brain signals, some of which are more representative of the individual, an approach that assumes all dimensions to be equal may not be suitable. Thus, we attempt to develop a novel model which can pay more attention to the most informative signals. In particular, the proposed approach is supposed to automatically learn the importance of the different parts of the EEG signal and focus on the valuable part. The effectiveness of attention-based RNN has been demonstrated in various domains including natural language processing \cite{wang2016attention} and speech recognition \cite{chan2017speech}.
Inspired by the wide success of this approach, we introduce the attention mechanism to the Encoder-Decoder RNN model to assign varying weights to different dimensions of the EEG data.
After EEG filtering, the composed Delta pattern $\mathcal{E}$ is fed into an attention-based Encoder-Decoder RNN structure \cite{wang2016attention} aiming to learn more representative features for user identification. The general Encoder-Decoder RNN framework regards all the feature dimensions of input sequence has the same weights, no matter how important the dimension is for the output sequence. In this paper, the different feature dimensions of the EEG data
correspond
to the different nodes of the EEG equipment. For example, the first dimension (first channel) collects the EEG data from the $AF3$\footnote{Both $AF3$ and $O1$ are EEG measurement positions in the International 10-20 Systems.} node
located at the frontal lobe of the scalp while the 7-th dimension is gathered from $O1$ node at the occipital lobe.
To assign
various
weights to different dimensions of $\mathcal{E}$, we introduce the attention mechanism to the Encoder-Decoder RNN model. The proposed attention-based Encoder-Decoder RNN
consists of three components (Figure~\ref{fig:attention}): the encoder, the attention module, and the decoder. The encoder is designed to compress the input Delta $\delta$ wave into a single intermediate code $C$; the attention module helps the encoder
calculate a better intermediate code $C_{att}$ by generating a sequence of the weights $W_{att}$ of different dimensions; the decoder accepts the attention-based code $C_{att}$ and decodes it to the output layer $Y'$.
Suppose the data in $i$-th layer can be denoted by $X^i=(X^i_j;i\in[1,2,\cdots,I], j\in[1,2,\cdots, N^i])$
where $j$ denotes the $j$-th dimension of $X^i$. $I$ represents the number of neural network layers while $N^i$ denotes the number of dimensions in $X^i$. Take the first layer as an example, we have $X^1=\mathcal{E}$ which indicates that the input sequence is the Delta pattern. Let the output sequence be $Y=(Y_k; k\in[1,2,\cdots,K])$ where K denotes the number of users.
In this paper, the user ID is represented by the one-hot label with length $K$.
For simplicity, we define the operation $\mathcal{T}(\cdot)$ as:$\mathcal{T}(X^i)=X^iW+b$.
Further more, we have
$$\mathcal{T}(X^{i-1}_j,X^i_{j-1})=X^{i-1}_j*W'+X^i_{j-1}*W''+b'$$
where $W$, $b$, $W'$, $W''$, $b'$ denote the corresponding weights and biases parameters.
The encoder component contains several non-recurrent fully-connected layers and one recurrent Long Short-Term Memory (LSTM) layer. The non-recurrent layers are employed to construct and fit into a non-linear function to purify the input Delta pattern, the necessity is demonstrated by the preliminary experiments\footnote{Some optimal designs like the neural network layers are validated by the preliminary experiments but the validation procedure will not be reported in this paper for space limitation.}.
The data flow in these non-recurrent layers can be calculated by
$$X^{i+1}=tanh(\mathcal{T}(X^i))$$
where $tanh$ is the activation function. We engage the $tanh$ as an activation function instead of $sigmoid$ for the stronger gradient \cite{lecun1998efficient}. The LSTM layer is adopted to compress the output of non-recurrent layers to a length-fixed sequence which is regarded as the intermediate code $C$. Suppose LSTM is the $i'$-th layer, the code equals to the output of LSTM, which is $C=X^{i'}_j$. The $X^{i'}_j$ can be measured by
\begin{equation}
\label{equ:1}
X^{i'}_j=\mathcal{L}(c^{i'}_{j-1},X^{i-1}_j,X^{i'}_{j-1})
\end{equation}
where $c^{i'}_{j-1}$ denotes the hidden state of the $(j-1)$-th LSTM cell. The operation $\mathcal{L}(\cdot)$ denotes the calculation process of the LSTM structure, which can be inferred from the following equations.
\[X^{i'}_{j}=f_o\odot tanh(c^{i'}_{j})\]
\[c^{i'}_{j}=f_f\odot c^{i'}_{j-1}+f_i\odot f_m\]
\[f_o=sigmoid(\mathcal{T}(X^{i'-1}_{j},X^{i'}_{j-1}))\]
\[f_f=sigmoid(\mathcal{T}(X^{i'-1}_{j},X^{i'}_{j-1}))\]
\[f_i=sigmoid(\mathcal{T}(X^{i'-1}_{j},X^{i'}_{j-1}))\]
\[f_m=tanh(\mathcal{T}(X^{i'-1}_{j},X^{i'}_{j-1}))\]
where $f_o,f_f, f_i$ and $f_m$ represent the output gate, forget gate, input gate and input modulation gate
,
respectively, and $\odot$ denotes the element-wise multiplication.
The attention module accepts the final hidden states as the unnormalized attention weights $W'_{att}$ which can be measured by the mapping operation $\mathcal{L}'(\cdot)$ (similar with Equation~\ref{equ:1})
$$W'_{att}=\mathcal{L}'(c^{i'}_{j-1},X^{i-1}_j,X^{i'}_{j-1})$$
and calculate the normalized attention weights $W_{att}$
$$W_{att}=softmax(W'_{att})$$
The softmax function is employed to normalize the attention weights into the range of $[0,1]$. Therefore, the weights can be explained as the probability that how the code $C$ is relevant to the output results.
Under the attention mechanism, the code $C$ is weighted to $C_{att}$
$$C_{att}=C\odot W_{att}$$
Note, $C$ and $W_{att}$ are trained instantaneously.
The decoder receives the attention-based code $C_{att}$ and decodes it to the output $Y'$.
Since $Y'$ is predicted at the output layer of the attention based RNN model ($Y'=X^I$), we have
$$Y'=\mathcal{T}(C_{att})$$
At last, we employ the cross-entropy cost function and $\ell_2$-norm (with parameter $\lambda$) is selected to prevent overfitting. The cost is optimized by the AdamOptimizer algorithm \cite{kingma2014adam}.
The iterations threshold of attention-based RNN is set as $n^E_{iter}$.
The weighted code $C_{att}$ has a
direct linear relationship with the output layer and the predicted results. If the model is well trained with low cost, we could regard the weighted code as a high-quality representation of the user ID. We set the learned deep feature $X_D$ to $C_{att}$, $X_D=C_{att}$, and feed it into a lightweight nearest neighbor classifier. The EEG ID, which is denoted by $E_{ID}$, can be directly predicted by the classifier.
\begin{algorithm}[!t]
\caption{Deepkey System}
\label{alg:Deepkey}
\begin{algorithmic}[1]
\renewcommand{\algorithmicrequire}{\textbf{Input:}}
\renewcommand{\algorithmicensure}{\textbf{Output:}}
\REQUIRE EEG data $\mathcal{E}$ and Gait data $\mathcal{G}$
\ENSURE Authentication Decision: Approve/Deny
\STATE \#Invalid ID Filter:
\FOR{$\mathcal{E}, \mathcal{G}$}
\STATE Genuine/Impostor $\gets \mathcal{E}$
\IF{Impostor}
\RETURN Deny
\ELSIF{Genuine}
\STATE \#EEG Identification Model:
\WHILE{$iteration<n^E_{iter}$}
\STATE $X^{i+1}=tanh(\mathcal{T}(X^i))$ \COMMENT{$X^1 = \mathcal{E}$}
\STATE $C=X^{i'}_j=\mathcal{L}(c^{i'}_{j-1},X^{i-1}_j,X^{i'}_{j-1})$
\STATE $W_{att}=softmax(\mathcal{L}'(c^{i'}_{j-1},X^{i-1}_j,X^{i'}_{j-1}))$
\STATE $C_{att}=C\odot W_{att}$
\STATE $E_{ID}\gets C_{att}$
\ENDWHILE
\STATE \#Gait Identification Model:
\WHILE{$iteration<n^G_{iter}$}
\STATE $G_{ID}\gets \mathcal{G}$
\ENDWHILE
\IF{$E_{ID}=G_{ID}$}
\RETURN Approve
\ELSE
\RETURN Deny
\ENDIF
\ENDIF
\ENDFOR
\end{algorithmic}
\end{algorithm}
The Gait Identification Model
works similar as the EEG Identification Model except the frequency band filtering. The iterations threshold of attention-based RNN is set as $n^G_{iter}$.
The Gait Identification Model receives subjects' gait data $\mathcal{G}$ from the input data $\mathcal{I}$ and
maps to the user's Gait ID $G_{ID}$. All the model structures, hyper-parameters, optimization, and other settings in the EEG and Gait Identification Models remain the same to keep the lower model complexity of the Deepkey system.
\begin{figure}[!t]
\includegraphics[width=\linewidth]{figure/collection.pdf}
\caption{Data collection. Two collection steps are cascade to eliminate the impact on EEG data of walking. The first step collects the solo EEG signals while the second step collects the gait signals.}
\label{fig:collection}
\end{figure}
\section{Experiments and results}
\label{sec:experiment_and_results}
In this section, we first outline the experimental setting including dataset, hyper-parameters settings, and evaluation metrics. Then we systematically investigate:
1) the comparison with the state-of-the-art authentication systems in both system-level and component-level;
2) the impact of key paramters like single/multiple sessions\footnote{Single session refers to the dataset collected in one session (the period from one subject putting the EEG headset on until all the experiments are finished then putting off). Multi-session represents the EEG data is collected from different sessions, which considered the effect on EEG data quality caused by the headset position errors.}, EEG band, and datasize;
3) the authentication latency?
\subsection{Experimental Settings}
\label{sub:setting}
\begin{table}[!bt]
\centering
\caption{Datasets description. \#-D denotes the number of dimensions.}
\label{tab:dataset}
\resizebox{0.6\textwidth}{!}{
\begin{tabular}{llllll}
\rowcolor[HTML]{FFFFFF}
\hline
\textbf{Dataset} & \textbf{Biometric} & \textbf{\#-D} & \textbf{Session} & \textbf{Frequency} & \textbf{Samples} \\ \hline
\textbf{EID-S} & EEG & 14 & Single & 128Hz & 49,000 \\
\textbf{EID-M} & EEG & 14 & Multiple & 128Hz & 147,000 \\
\textbf{GID-S} & Gait & 27 & Single & 80Hz & 140,000 \\
\textbf{GID-M} & Gait & 27 & Multiple & 80Hz & 420,000 \\ \hline
\end{tabular}
}
\vspace{-3mm}
\end{table}
\begin{table*}[!bt]
\centering
\caption{System-level comparison between Deepkey and other biometrics authentication systems. The performance of our methods are evaluated on multi-session datasets (EID-M, GID-M).}
\label{tab:system_comparison}
\resizebox{\textwidth}{!}{\begin{tabular}{lllllllll}
\hline \hline
& \textbf{Reference} & \textbf{Biometric} & \textbf{Method} & \#-Subject & Dataset & \textbf{Accuracy} & \textbf{FAR} & \textbf{FRR} \\ \hline
\multirow{11}{*}{\begin{tabular}[c]{@{}l@{}}\textbf{Uni-}\\ \textbf{-modal}\end{tabular}} & \cite{cola2016gait} & Gait & semi-supervised anomaly detection+NN & 15 & Local & 97.4 & & \\
& \cite{muramatsu2015gait} & Gait & AVTM-PdVs &100 & Public & 77.72 & & \\
& \cite{al2017unobtrusive} & Gait & MLP & 60 &Local & 99.01 & & \\
& \cite{sun2014gait} & Gait & Artificial features + voting classifier & 10 &Local & 98.75 & & \\
& \cite{konno2015gait} & Gait & Two SVMs & 50 &Local & & 1.0 & 1.0 \\
& \cite{thomas2017eeg} & EEG & PSD + cross-correlation values &109 &Public & & 1.96 &1.96 \\
& \cite{chuang2013think} & EEG & Customized Threshold &15 &Local & & 0& 2.2 \\
& \cite{gui2014exploring} & EEG &Low-pass filter+wavelets+ ANN &32 &Local & 90.03 & & \\
& \cite{bashar2016human} & EEG & Bandpass FIR filter +ECOC + SVM & 9 & Local & 94.44 & & \\
& \cite{thomas2016utilizing} & EEG & \begin{tabular}[c]{@{}l@{}}IAF + delta band EEG \\ + Cross-correlation values\end{tabular} &109 &Public & 90.21 & & \\
& \cite{jayarathne2016brainid} & EEG & CSP +LDA & 12 & Local & 96.97 & & \\
& \cite{zhang2018mindid} & EEG & Attention-based RNN + XGB & 8 &Local & 98.82 & & \\
& \cite{keshishzadeh2016improved} & EEG & AR + SVM & 104 & Public& 97.43 & & \\ \hline \hline
\multirow{18}{*}{\begin{tabular}[c]{@{}l@{}}\textbf{Multi-}\\ \textbf{-modal}\end{tabular}} & \multirow{3}{*}{\cite{long2012multimodal}} & Fingerprint & \multirow{3}{*}{ZM + RBF Neural Network} & & & 92.89 & 7.108 & 7.151 \\
& & Face & & & & & 11.52 & 13.47 \\
& & Fusion & & 40 &Public & & 4.95 & 1.12 \\ \cline{2-9}
& \multirow{3}{*}{\cite{yano2012multimodal}} & Iris & \multirow{3}{*}{\begin{tabular}[c]{@{}l@{}}Gabor 2D wavelets + Gabor 2D wavelets\\ + Hamming distance \end{tabular}} & & & & 13.88 & 13.88 \\
& & Pupil & & & & &5.47 & 5.47 \\
& & Fusion & & 59 &Public & &2.44 & 2.44 \\ \cline{2-9}
& \multirow{3}{*}{\cite{manjunathswamy2015multimodal}} & ECG & wavelet decomposition & & & & 2.37 & 9.52 \\
& & Fingerprint & Histogram manipulation Image Enhancement & & & & 7.77 & 5.55 \\
& & Fusion & Score Fusion &50 &Public & & 2.5 & 0 \\ \cline{2-9}
& \multirow{3}{*}{\cite{derawi2014fusion}} & ECG & \multirow{2}{*}{linear time interpolation + cross correlation} & & & & 4.2 & 4.2 \\
& & Gait & & & & & 7.5 & 7.5 \\
& & Fusion & Score Fusion & 30 &Local & & 1.26 &1.26 \\ \cline{2-9}
& \multirow{3}{*}{Ours} & EEG & Delta wave + Attention-based RNN +KNN & & & 99.96 & & \\
& & Gait & Attention-based RNN +KNN & & & 99.61 & & \\
& & Fusion & &7 &Local & \textbf{99.57} & \textbf{0} & \textbf{1.0} \\ \hline \hline
\end{tabular}}
\end{table*}
\begin{figure*}[t]
\hspace{5mm}
\centering
\begin{subfigure}[t]{0.25\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/eeg_1trial_cm.png}
\caption{EID-S CM}
\end{subfigure}%
\begin{subfigure}[t]{0.25\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/eeg_3trial_cm.png}
\caption{EID-M CM}
\end{subfigure}%
\begin{subfigure}[t]{0.25\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/gait_1trial_cm.png}
\caption{GID-S CM}
\end{subfigure}%
\begin{subfigure}[t]{0.25\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/gait_3trial_cm.png}
\caption{GID-M CM}
\end{subfigure}%
\centering
\begin{subfigure}[t]{0.25\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/deepkey_eeg_1trial_roc.png}
\caption{EID-S ROC}
\end{subfigure}%
\begin{subfigure}[t]{0.25\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/deepkey_eeg_3trial_roc.png}
\caption{EID-M ROC}
\end{subfigure}%
\begin{subfigure}[t]{0.25\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/deepkey_gait_1trial_roc.png}
\caption{GID-S ROC}
\end{subfigure}%
\begin{subfigure}[t]{0.25\textwidth}
\centering
\includegraphics[width=\textwidth]{figure/deepkey_gait_3trial_roc.png}
\caption{GID-M ROC}
\end{subfigure}%
\caption{Confusion matrix and ROC curves of the datasets. CM denotes confusion matrix. The AUC are provided on the figures. }
\label{fig:cm_roc}
\end{figure*}
\begin{table*}[tb]
\centering
\caption{Classification report of the datasets including precision, recall and F-1 score. The proposed approach gains impressive results (higher than 99\%) on all the metrics over all the 7 subjects.
}
\label{tab:classification_report}
\resizebox{\textwidth}{!}{
\begin{tabular}{l|lll|lll|lll|lll}
\hline
\textbf{Datasets} & \multicolumn{3}{c|}{\textbf{EID-S}} & \multicolumn{3}{c|}{\textbf{EID-M}} & \multicolumn{3}{c|}{\textbf{GID-S}} & \multicolumn{3}{c}{\textbf{GID-M}} \\ \hline
\textbf{Metrics} & \textbf{Precision} & \textbf{Recall} & \textbf{F1-score} & \textbf{Precision} & \textbf{Recall} & \textbf{F1-score} & \textbf{Precision} & \textbf{Recall} & \textbf{F1-score} & \textbf{Precision} & \textbf{Recall} & \textbf{F1-score} \\ \hline
\textbf{0} & 1.0 & 1.0 & 1.0 & 0.998 & 1.0 & 0.999 & 1.0 & 1.0 & 1.0 & 1.0 & 1.0 & 1.0 \\
\textbf{1} & 1.0 & 1.0 & 1.0 & 1.0 & 0.9996 & 0.9998 & 0.996 & 0.996 & 0.996 & 0.9915 & 0.9904 & 0.9909 \\
\textbf{2} & 1.0 & 1.0 & 1.0 & 1.0 & 1.0 & 1.0 & 1.0 & 1.0 & 1.0 & 0.9999 & 1.0 & 0.9999 \\
\textbf{3} & 1.0 & 1.0 & 1.0 & 1.0 & 0.9992 & 0.9996 & 0.9992 & 0.9988 & 0.999 & 0.9937 & 0.9948 & 0.9942 \\
\textbf{4} & 1.0 & 1.0 & 1.0 & 1.0 & 1.0 & 1.0 & 0.9976 & 0.998 & 0.9978 & 0.9949 & 0.996 & 0.9955 \\
\textbf{5} & 1.0 & 1.0 & 1.0 & 0.9996 & 0.9992 & 0.9994 & 0.9969 & 0.9984 & 0.9977 & 0.9963 & 0.996 & 0.9961 \\
\textbf{6} & 1.0 & 1.0 & 1.0 & 0.9996 & 0.9993 & 0.9994 & 0.9988 & 0.9972 & 0.998 & 0.9962 & 0.9953 & 0.9958 \\
\textbf{Average} & \textbf{1.0} & \textbf{1.0} & \textbf{1.0} & \textbf{0.9996} & \textbf{0.9996} & \textbf{0.9996} & \textbf{0.9983} & \textbf{0.9983} & \textbf{0.9983} & \textbf{0.9961} & \textbf{0.9961} & \textbf{0.9961} \\ \hline
\end{tabular}
}
\end{table*}
\subsubsection{Datasets}
\label{sub:datasets}
We design real-world experiments to collect EEG data and gait data in cascade. The experiments (Figure~\ref{fig:collection}) are conducted by 7 healthy participants aged $26\pm2$ including 4 males and 3 females.
In Step 1, each participant remains standing and
relaxed with eye closed.
The EEG data are collected by EPOC+ Emotiv headset\footnote{\url{https://www.emotiv.com/product/emotiv-epoc-14-channel-mobile-eeg/}} which integrates 14 electrodes (corresponding to 14 EEG channels) with 128Hz sampling rate.
In Step 2, each participant walks in an aisle to generate the gait data. In the gait collection procedure, three IMUs are attached to the participants' left wrist, the middle of the back, and the left ankle, respectively. Each IMU (PhidgetSpatial 3/3/3\footnote{\url{https://www.phidgets.com/?&prodid=48}}) with 80Hz sampling rate gathering 9 dimensional motor features contains a 3-axis accelerometers, 3-axis gyroscopes, and 3-axis magnetometers.
To investigate the impact of dataset sessions, both the EEG and gait data are collected in three sessions. In every cycle of single session, the subject puts on the equipment (headset/IMUs), gathers data, and then takes the equipment off. Therefore, in different sessions, the positions of the equipment may have slight deviation.
Table~\ref{tab:dataset} reports the details of the datasets used in this paper. Each EEG or gait sample contains 10 continuous instances without overlapping.
The single session datasets (EID-S and GID-S) are collected in single experiment session whilst the multi-session datasets (EID-M and GID-M) are gathered in three sessions. All the sessions are conducted in the same place but three different days (each session in one day). The EEG data are easily influenced if the emotional or physical states has changed, thus, we believe the collected data are diverse because of the varying environmental factors (e.g., noise and temperature) and subjective factors (e.g., participants' mental state and fatigue state). Similarly, the gait signals could be affected by lots of variables like different shoes (comfortable/uncomfortable). In this paper, as an exploratory work, we focus to develop a robust discriminative deep learning model which is strong enough to prevent the corruption of the aforementioned influencing factors. The investigation of the detail effect brought by each specific factor will be left as a future research direction.
\subsubsection{Parameter Settings}
\label{sub:parameter_settings}
The Invalid ID Filter attempts to recognize the unauthorized subject based on the unique EEG data. The filter chooses the RBF kernel with $nu = 0.15$. In the EEG Identification Model, the Delta band ($[0.5Hz,3.5Hz]$) is filtered by 3 order butter-worth bandpass filter. In the attention-based RNN, for both EEG and gait, each hidden layer contains 64 nodes, while the learning rate and $\lambda$ are both set to 0.001. The weighted code is produced after 1,000 iterations.
87.5\% of the datasets are used for training while the remaining datasets are used for testing. The user ID is ranged from 0 to 6 and represented in the one-hot label.
\begin{table*}[t]
\centering
\caption{Component-level comparison. DL denotes Deep Learning. Att-RNN denotes attention-based RNN.
}
\label{tab:component_level_comparison}
\resizebox{\textwidth}{!}{
\begin{tabular}{l|l|llll|llll}
\hline \hline
\multirow{2}{*}{\textbf{Baseline}} & \multirow{2}{*}{\textbf{Methods}} & \multicolumn{4}{c|}{\textbf{EID-S}} & \multicolumn{4}{c}{\textbf{EID-M}} \\ \cline{3-10}
& & \textbf{Accuracy} & \textbf{Precision} & \textbf{Recall} & \textbf{F1-score} & \textbf{Accuracy} & \textbf{Precision} & \textbf{Recall} & \textbf{F1-score} \\ \hline
\multirow{5}{*}{\textbf{Non-DL Baseline}} & \textbf{SVM} & 0.4588 & 0.5848 & 0.4588 & 0.4681 & 0.7796 & 0.7815 & 0.7796 & 0.7796 \\
& \textbf{RF} & 0.9875 & 0.9879 & 0.9876 & 0.9876 & 0.8124 & 0.8139 & 0.8124 & 0.812 \\
& \textbf{KNN} & 0.9897 & 0.9899 & 0.9898 & 0.9898 & 0.8211 & 0.8232 & 0.8211 & 0.8197 \\
& \textbf{AB} & 0.2872 & 0.3522 & 0.2871 & 0.2337 & 0.3228 & 0.3224 & 0.3228 & 0.2815 \\
& \textbf{LDA} & 0.1567 & 0.1347 & 0.1567 & 0.1386 & 0.3082 & 0.285 & 0.3082 & 0.2877 \\ \hline
\multirow{3}{*}{\textbf{DL Baseline}} & \textbf{LSTM} & 0.9596 & 0.9601 & 0.9596 & 0.9597 & 0.8482 & 0.8509 & 0.8483 & 0.8489 \\
& \textbf{GRU} & 0.9633 & 0.9636 & 0.99631 & 0.9631 & 0.862 & 0.8638 & 0.8626 & 0.8629 \\
& \textbf{CNN} & 0.8822 & 0.8912 & 0.8813 & 0.8912 & 0.7647 & 0.7731 & 0.7854 & 0.7625 \\ \hline
\multirow{3}{*}{\textbf{State-of-the-art}} & \textbf{\cite{jayarathne2016brainid}} & 0.5843 & 0.5726 & 0.5531 & 0.5627 & 0.5735 & 0.5721 & 0.5443 & 0.5579 \\
& \textbf{\cite{keshishzadeh2016improved}} & 0.8254 & 0.8435 & 0.8617 & 0.8525 & 0.8029 & 0.7986 & 0.8125 & 0.8055 \\
& \textbf{\cite{gui2014exploring}} & 0.8711 & 0.8217 & 0.7998 & 0.8106 & 0.8567 & 0.8533 & 0.8651 & 0.8592 \\ \hline
\textbf{} & \textbf{Att-RNN} & 0.9384 & 0.9405 & 0.9388 & 0.9391 & 0.9324 & 0.9343 & 0.9322 & 0.9326 \\
\textbf{} & \textbf{Ours} & \textbf{1.0} & \textbf{1.0} & \textbf{1.0} & \textbf{1.0} & \textbf{0.9996} & \textbf{0.9996} & \textbf{0.9996} & \textbf{0.9996} \\ \hline \hline
\multirow{2}{*}{\textbf{Baselines}} & \multirow{2}{*}{\textbf{Methods}} & \multicolumn{4}{c|}{\textbf{GID-S}} & \multicolumn{4}{c}{\textbf{GID-M}} \\ \cline{3-10}
& & \textbf{Accuracy} & \textbf{Precision} & \textbf{Recall} & \textbf{F1-score} & \textbf{Accuracy} & \textbf{Precision} & \textbf{Recall} & \textbf{F1-score} \\ \hline
\multirow{5}{*}{\textbf{Non-DL Baseline}} & \textbf{SVM} & 0.9981 & 0.9981 & 0.9981 & 0.9981 & 0.993 & 0.993 & 0.993 & 0.993 \\
\textbf{} & \textbf{RF} & 0.9878 & 0.9878 & 0.9878 & 0.9878 & 0.9954 & 0.9954 & 0.9954 & 0.9954 \\
\textbf{} & \textbf{KNN} & 0.9979 & 0.9979 & 0.9979 & 0.9979 & 0.9953 & 0.9953 & 0.9953 & 0.9953 \\
\textbf{} & \textbf{AB} & 0.5408 & 0.5689 & 0.5409 & 0.4849 & 0.5401 & 0.5135 & 0.542 & 0.4985 \\
\textbf{} & \textbf{LDA} & 0.688 & 0.6893 & 0.688 & 0.6855 & 0.6933 & 0.693 & 0.6933 & 0.6915 \\ \hline
\multirow{3}{*}{\textbf{DL Baseline}} & \textbf{LSTM} & 0.9951 & 0.9951 & 0.9951 & 0.9951 & 0.9935 & 0.9936 & 0.9936 & 0.9936 \\
\textbf{} & \textbf{GRU} & 0.9949 & 0.9949 & 0.9949 & 0.9949 & 0.9938 & 0.9938 & 0.9938 & 0.9938 \\
\textbf{} & \textbf{CNN} & 0.9932 & 0.9932 & 0.9932 & 0.9932 & 0.9845 & 0.9845 & 0.9845 & 0.9845 \\ \hline
\textbf{State-of-the-art} & \textbf{\cite{cola2016gait}} & 0.9721 & 0.9789 & 0.9745 & 0.9767 & 0.9653 & 0.9627 & 0.9669 & 0.9648 \\
\textbf{} & \textbf{\cite{al2017unobtrusive}} & 0.9931 & 0.9934 & 0.9957 & 0.9945 & 0.9901 & 0.9931 & 0.9942 & 0.9936 \\
\textbf{} & \textbf{\cite{sun2014gait}} & 0.9917 & 0.9899 & 0.9917 & 0.9908 & 0.9875 & 0.9826 & 0.9844 & 0.9835 \\ \hline
\textbf{} & \textbf{Att-RNN} & 0.99 & 0.99 & 0.99 & 0.99 & 0.9894 & 0.9895 & 0.9895 & 0.9895 \\
\textbf{} & \textbf{Ours} & \textbf{0.9983} & \textbf{0.9983} & \textbf{0.9983} & \textbf{0.9983} & \textbf{0.9961} & \textbf{0.9961} & \textbf{0.9961} & \textbf{0.9961} \\ \hline \hline
\end{tabular}
}
\end{table*}
\subsubsection{Metrics}
\label{sub:evaluation}
The adopted evaluation metrics are accuracy, ROC, AUC, along with {\em False Acceptance Rate} (FAR) and {\em False Rejection Rate } (FRR).
Deepkey is very sensitive to the invalid subject and can acquire very high accuracy in Invalid ID Filter Model for the reason that even tiny misjudgment may lead to catastrophic consequences. Therefore, FAR is more important than other metrics. Therefore, in Deepkey, FAR has higher priority compared to other metrics such as FRR.
\subsection{Overall Comparison}
\label{sub:overall_comparison}
\subsubsection{System-level Comparison}
\label{sub:system_level_comparison}
To evaluate the performance of Deepkey, by comparing it with a set of the state-of-the-art authentication systems. Deepkey is empowered to solve both the authentication and identification problems. As shown in Table~\ref{tab:system_comparison}, Deepkey achieves a FAR of 0 and a FRR of 1\%, outperforming other uni-modal and multimodal authentication systems. Specifically, our approach, compared to the listed uni-modal systems, achieves the highest EEG identification accuracy (99.96\%) and Gait identification accuracy (99.61\%).
\subsubsection{Component-level Comparison}
\label{sub:component_level_comparison}
To have a closer observation, we provide the detailed performance study of each component.
In the Invalid ID Filter, to enhance the accuracy and robustness of the classifier, EEG samples are separated into different segments, with each segment (without overlapping) having $200$ continuous samples.
Six in seven subjects are labeled as genuine while the other subject is labeled as
an impostor. In the training stage, all the EEG segments are fed into the one-class SVM with rbf kernel for pattern learning. In the test stage, 1,000 genuine segments and 1,000 impostor segments are randomly selected to asses the performance. We use the leave-one-out-cross-validation training strategy and achieve \textbf{ a FAR of 0 and a FRR of 0.006}.
In the EEG/Gait Identification Model, the proposed approach achieves an accuracy of 99.96\% and 99.61\% over the multi-session datasets, respectively. The detailed confusion matrix, ROC curves with AUC scores, and the classification reports (precision, recall, and F1-score) over all the datasets are presented in Figure~\ref{fig:cm_roc} and Table~\ref{tab:classification_report}. The above evaluation metrics demonstrate that the proposed approach achieves a performance of over 99\% on all the metrics over each subject and each dataset.
Furthermore, the overall comparison between our model and other state-of-the-art baselines
are listed in Table~\ref{tab:component_level_comparison}. RF denotes Random Forest, AdaB denotes Adaptive Boosting, LDA denotes Linear Discriminant Analysis. In addition, the key parameters of the baselines are listed here: Linear SVM ($C=1$), RF ($n=200$), KNN ($k=3$). The settings of LSTM are the same as the attention-based RNN classifier, along with the GRU (Gated Recurrent Unit).
The CNN contains 2 stacked convolutional layers (both with stride $[1,1]$, patch $[2,2]$, zero-padding, and the depth are 4 and 8, separately.) and followed by one pooling layer (stride $[1,2]$, zero-padding) and one fully connected layer (164 nodes). Relu activation function is employed in the CNN.
The methods used for comparison (3 for EEG-based authentication and 3 for gait-based authentication) are introduced as follows:
\begin{itemize}
\item Jayarathne et al. \cite{jayarathne2016brainid} feed EEG data to a bandpass filter ($8Hz-30Hz$), extract CSP (Common Spatial Pattern) and recognize the user ID by LDA.
\item Keshishzadeh et al. \cite{keshishzadeh2016improved} extract autoregressive coefficients as the features and identify the subject by SVM.
\item Gui et al. \cite{gui2014exploring} employ low-pass filter (60Hz) and wavelet packet decomposition to generate features and distinguish unauthorized person through deep neural network.
\item Cola et al. \cite{cola2016gait}
hire neural networks to analyze user gait pattern by artificial features such as kurtosis, peak-to-peak amplitude and skewness.
\item Al-Naffakh et al. \cite{al2017unobtrusive} propose to utilize time-domain statistical features and Multilayer Perception (MLP) for person identify.
\item Sun et al. \cite{sun2014gait} adopt a weighted voting classifier to process the extracted features like gait frequency, symmetry coefficient, and dynamic range.
\end{itemize}
\begin{table*}[t]
\centering
\caption{EEG bands comparison. The full band denotes the raw EEG data with full frequency bands.}
\label{tab:bands_comparison}
\resizebox{\textwidth}{!}{
\begin{tabular}{lllllllllll}
\hline \hline
\multirow{2}{*}{\textbf{Dataset}} & \multirow{2}{*}{\textbf{Baseline}} & \multirow{2}{*}{\textbf{Methods}} & \multicolumn{6}{c}{\textbf{EEG Bands}} & \multirow{2}{*}{\textbf{Best Level}} & \multirow{2}{*}{\textbf{Best Band}} \\ \cline{4-9}
& & & \textbf{Delta} & \textbf{Theta} & \textbf{Alpha} & \textbf{Beta} & \textbf{Gamma} & \textbf{Full} & & \\ \hline
\multirow{13}{*}{\textbf{EID-S}} & \multirow{5}{*}{\textbf{\begin{tabular}[c]{@{}l@{}}Non-DL\\ Baseline\end{tabular}}} & \textbf{SVM} & 0.4588 & 0.4682 & 0.7484 & 0.5955 & 0.5239 & 0.8788 & 0.8788 & Full \\
& & \textbf{RF} & 0.9875 & 0.8006 & 0.7729 & 0.6376 & 0.5469 & 0.8931 & 0.9875 & Delta \\
& & \textbf{KNN} & 0.9897 & 0.8084 & 0.7465 & 0.5553 & 0.4606 & 0.8792 & 0.9897 & Delta \\
& & \textbf{AB} & 0.2872 & 0.2879 & 0.2318 & 0.3069 & 0.2922 & 0.3289 & 0.3289 & Full \\
& & \textbf{LDA} & 0.1567 & 0.1802 & 0.1957 & 0.1502 & 0.1306 & 0.4547 & 0.4547 & Full \\ \cline{2-11}
& \multirow{3}{*}{\textbf{\begin{tabular}[c]{@{}l@{}}DL\\ Baseline\end{tabular}}} & \textbf{LSTM} & 0.9596 & 0.8126 & 0.8277 & 0.6906 & 0.6027 & 0.9273 & 0.9596 & Delta \\
& & \textbf{GRU} & 0.9633 & 0.7996 & 0.8082 & 0.6902 & \textbf{0.6985} & 0.9251 & 0.9633 & Delta \\
& & \textbf{CNN} & 0.8822 & 0.7416 & 0.8079 & \textbf{0.6918} & 0.6059 & 0.8985 & 0.8985 & Full \\ \cline{2-11}
& & \textbf{\cite{jayarathne2016brainid}} & 0.5843 & 0.4487 & 0.2918 & 0.3017 & 0.4189 & 0.5112 & 0.5843 & Delta \\
& & \textbf{\cite{keshishzadeh2016improved}} & 0.8254 & 0.7935 & 0.7019 & 0.6368 & 0.6621 & 0.8018 & 0.8254 & Delta \\
& \multirow{-3}{*}{\textbf{\begin{tabular}[c]{@{}l@{}}State-of\\ -the-art\end{tabular}}} & \textbf{\cite{gui2014exploring}} & 0.8711 & 0.8531 & 0.7556 & 0.6882 & 0.5101 & 0.7819 & 0.8711 & Delta \\ \cline{2-11}
& \textbf{} & \textbf{Att-RNN} & 0.9384 & 0.7928 & 0.8318 & 0.6854 & 0.6046 & 0.9238 & 0.9384 & Delta \\
& \textbf{} & \textbf{Ours} & \textbf{1.0} & \textbf{0.9285} & \textbf{0.8366} & 0.5529 & 0.4558 & 0.9417 & 1.0 & Delta \\ \hline \hline
\multirow{13}{*}{\textbf{EID-M}} & \multirow{5}{*}{\textbf{\begin{tabular}[c]{@{}l@{}}Non-DL \\ Baseline\end{tabular}}} & \textbf{SVM} & 0.7796 & 0.5424 & 0.5664 & 0.6522 & 0.4915 & 0.7477 & 0.7796 & Delta \\
& & \textbf{RF} & 0.8124 & 0.7194 & 0.7351 & \textbf{0.6842} & 0.4765 & 0.8121 & 0.7194 & Delta \\
& & \textbf{KNN} & 0.8211 & 0.7501 & 0.7649 & 0.6611 & 0.3821 & 0.8162 & 0.7501 & Delta \\
& & \textbf{AB} & 0.3228 & 0.3095 & 0.2478 & 0.2548 & 0.2529 & 0.3189 & 0.3228 & Delta \\
& & \textbf{LDA} & 0.3082 & 0.1681 & 0.1621 & 0.1824 & 0.1311 & 0.2995 & 0.3082 & Delta \\ \cline{2-11}
& \multirow{3}{*}{\textbf{\begin{tabular}[c]{@{}l@{}}DL \\ Baseline\end{tabular}}} & \textbf{LSTM} & 0.8482 & 0.6926 & 0.7438 & 0.5726 & 0.5008 & 0.8185 & 0.8482 & Delta \\
& & \textbf{GRU} & 0.862 & 0.6935 & 0.7531 & 0.5672 & \textbf{0.5072} & 0.8221 & 0.862 & Delta \\
& & \textbf{CNN} & 0.7647 & 0.6712 & 0.7191 & 0.5588 & 0.4949 & 0.7749 & 0.7749 & Full \\ \cline{2-11}
& \multirow{3}{*}{\textbf{\begin{tabular}[c]{@{}l@{}}State-of\\ -the-art\end{tabular}}} & \textbf{\cite{jayarathne2016brainid}} & 0.9721 & 0.7019 & 0.7091 & 0.4189 & 0.4089 & 0.8195 & 0.9721 & Delta \\
& & \textbf{\cite{keshishzadeh2016improved}} & 0.9931 & 0.6891 & 0.6988 & 0.5124 & 0.3397 & 0.7963 & 0.9931 & Delta \\
& &\textbf{\cite{gui2014exploring}} & 0.9917 & 0.7199 & 0.6572 & 0.4911 & 0.3977 & 0.8011 & 0.9917 & Delta \\ \cline{2-11}
& \textbf{} & \textbf{Att-RNN} & 0.9324 & 0.6847 & 0.6846 & 0.5732 & 0.4941 & 0.7976 & 0.9324 & Delta \\
& & \textbf{Ours} & \textbf{0.9996} & \textbf{0.9013} & \textbf{0.8989} & 0.4428 & 0.3661 & \textbf{0.8858} & 0.9996 & Delta\\ \hline \hline
\end{tabular}
}
\end{table*}
\begin{figure*}[t]
\centering
\begin{minipage}[b]{0.48\linewidth}
\centering
\includegraphics[width=\textwidth]{figure/session-eps-converted-to.pdf}
\caption{Impact of Session}
\label{fig:sessions}
\end{minipage}
\begin{minipage}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{figure/deepkey_latency-eps-converted-to.pdf}
\caption{Latency}
\label{fig:latency}
\end{minipage}
\end{figure*}
The primary conclusions from Table~\ref{tab:component_level_comparison} are summarized as follows:
\begin{itemize}
\item Our approach achieves the highest performance over both EEG and gait datasets under single session and multi-session settings;
\item For most of the baselines, we observe lower accuracy in EEG dataset compared to that in gait dataset, implying the EEG-based authentication is still challenging and our approach has room to improve; nevertheless, our model outperforms others and show superiority to both EEG-based or Gait-based methods;
\item The results on single-session datasets are better than that on multi-session datasets. This is reasonable, and it demonstrates that the number of sessions does affect the authentication accuracy. This problem will be further analyzed in Section~\ref{sub:impact_of_sessions};
\item Our model achieves better performance than Att-RNN. Since the diversity between our model and Att-RNN is that we employ an external classifier, this observation demonstrates that the external classifier is effective and efficient.
\end{itemize}
In Deepkey, the subjects passing the Invalid ID Filter are regarded as genuine only if their recognized IDs are consistent, i.e., $E_{ID}=G_{ID}$. It can be inferred easily that the FAR of Deepkey is
0 as well. However, the FRR depends on one or more of these
three scenarios: the false rejection of Invalid ID Filter; the incorrect Gait identification; the incorrect EEG identification. In summary, the overall FAR is \textbf{0} and the overall FRR is calculated as $\mathbf{1\%}=(0.006+0.994*((1-0.9961)+0.9961*(1-0.9996)))$.
\subsection{Impact of Key Parameters}
\label{sub:impact_of_key_parameters}
\subsubsection{Impact of sessions}
\label{sub:impact_of_sessions}
In practical applications, sessions in different scenarios may result in a minor difference in equipment position, signal quality, and other factors. To investigate the impact of sessions, we conduct external experiments by comparing the performance between single-session datasets and multi-session datasets. The comprehensive evaluation metrics and the comparison over various baselines are listed in Table~\ref{tab:classification_report} and Table~\ref{tab:component_level_comparison} while the comparison is summarized in Figure~\ref{fig:sessions}. The experiment results show that on the multi-session datasets, compared the single-session datasets, we achieve a slightly lower but still highly competitive performance.
\subsubsection{Impact of EEG Band}
\label{sub:impact_of_eeg_band}
A series of comparison experiments are designed to explore the optimal EEG frequency band which contains the most discriminative features. The results presented in Table~\ref{tab:bands_comparison} illustrate that:
\begin{itemize}
\item The Delta band consistently provides higher identification accuracy compared to other frequency bands for both single and multiple sessions. This observation shows that Delta pattern contains the most discriminative information for person identification.
\item Our method gains the best outcome on both datasets with different sessions. This validates the robustness and adaptability of the proposed approach.
\end{itemize}
Why Delta pattern could outperform other patterns since Delta wave mainly appear in deep sleep state? Here we give one possible reason. We know the fact that the EEG patterns are associated with individuals' mental and physical states (organics and systems). For example, while the subject is under deep sleep and producing Delta pattern, the most parts of physical functions of the body (such as sensing, thinking, even dreaming) are completely detached. Only the very essential life-support organs and systems (such as breathing, heart beating, and digesting) keep working, which indicates the brain areas corresponding to life-support functions are active. While the subject is awake (e.g., relax state) and producing Alpha pattern, the subject has more activated functions such as imaging, visualizing and concentrating. Also, more brain functions like hearing, touching, and thinking are attached, which means that more physical brain areas (such as frontal lobe, temporal lobe, and parietal lobe) are activated. At this time, the life-support organs are still working. In short, only the life-support organs related brain areas are active in the first scenario (Delta pattern) whilst the brain areas controlling life-support and high-level functions (e.g., concentrating) are active in the second scenario (Alpha pattern). Thus, we infer that the delta pattern is correspond to the life-support organs and systems, which is the most stable function in different scenarios and the most discriminative signal in inter-subject classification.
\subsubsection{Impact of Datasize}
\label{sub:impact_of_datasize}
Datasize is one crucial influence element in deep learning based algorithms. In this section, we conduct experiments to train the proposed method over various data sizes. As shown in the radar chart (Figure~\ref{fig:datasize}), four datasets are evaluated under different proportion of training datasize. It has five equi-angular spokes which represent the proportion of datasize (20\%, 40\%, 60\%, 80\%, 100\%), respectively. The four concentric circles indicate the accuracy which are 85\%, 90\%, 95\%, and 100\%, respectively. Each closed line represent a dataset and have five intersections with the five spokes. Each intersection node represents the classification accuracy of a specific dataset with a specific proportion datasize. For example, the intersection of EID-S line and 20\% spoke is about 0.92, which denote that our approach achieves the accuracy of 92\% over EID-S dataset with 20\% datasize.
The radar chart infers that gait datasets (GID-S and GID-M) obtain competitive accuracy even with 20 percent datasize; nevertheless, EID-S and EID-M highly rely on the datasize. This phenomenon is reasonable because EEG data has lower signal-to-noise ratio and requires more samples to learn the latent distribution pattern.
\subsection{Latency Analysis}
\label{sub:latency_analysis}
We also study the latency of Deepkey since low delay is highly desirable in real-world deployment.
The latency of Deepkey is compared to the response time of several state-of-the-art authentication approaches. The comparison results are shown in Figure~\ref{fig:latency}.
The latency of our method is 0.39 sec, which is competitive compared
to the state-of-the-art baselines.
Specifically, the reaction time of Deepkey is composed of three components, with the Invalid ID Filter taking 0.06 sec and the ID Identification taking 0.33 sec.
\begin{figure}
\centering
\includegraphics[width=0.5\textwidth]{figure/radar.png}
\caption{Impact of Datasize. The spokes of this radar chart represent the datasize proportion (20\%, 40\%, 60\%, 80\%, 100\%) while the concentric circles indicate the classification accuracy (85\%, 90\%, 95\%, and 100\%).}
\label{fig:datasize}
\end{figure}
\section{Discussions and Future Work}
\label{sec:discussions_and_future_work}
In this paper, we propose a biometric identification system based both EEG and gait information. In this section, we discuss the open challenges and potential future research directions.
First, the datasets used in this paper only has limited participants. The extensive evaluations over more subjects are necessary. However, compare to some existing works (\cite{zhang2018mindid}, \cite{bashar2016human}, \cite{sun2014gait} have 8, 9, 10 subjects in their experiments, respectively), we believe our participants amount is acceptable.
Our work has already demonstrate that Deepkey can be used in settings such as small offices which are accessed by a small group of people.
In addition, evaluation can be improved by extending observations on how the system performs in different conditions. For example, considering changes in EEG signals during more trails, and longer times (hours, days or even months) to understand if these are consistent and reliable for detection.
Second, wearable sensors like EEG headset and wearable IMUs are required in the data collection stage of Deepkey system. Extensive experiments are meaningful in the future to investigate how the placement position of the wearable sensor matters. The EEG headset position on the head and the IMU position on the arms may affect the authentication performance. With the development of hardware related techniques, the EEG headset is becoming more portable and affordable. For example, the developed cEEGrids\footnote{http://ceegrid.com/home/}, an flex-printed, multi-channel sensor arrays that are placed around the ear using an adhesive, are easy and comfortable to wear and dispatch. This a good tendency of EEG acquisition equipment although the cEEGrids are currently expensive.
In the future, the widely deployment of DeepKey authentication system in real world environment is possible.
Moreover, another future scope is to develop device-free authentication techniques to overstep the inconvenient brought by wearable sensors. We have analyzed the computational latency in Section~\ref{sub:latency_analysis}. However, the overall system latency not only include the computation latency but also include the data collection latency such as the time cost when wear the EEG headset and the IMUs. The data collection latency of our system is about ten seconds while EEG collection requires about four seconds and gait collection requires above six seconds. Compare to other authentication system such as fingerprint, our data collection latency is much higher. This is one of the major future directions of our work. A potential solution is develop device-free authentication system and measure gait signals by environmental sensors such as RFID (Radio Frequency IDentification) tags.
The Received Signal Strength Indication (RSSI) of RFID tag measures the present received signal power which reflects the target subject's walking information.
Additionally, a promising future work is the real-world online deployment of the proposed DeepKey system. Since we have demonstrated that the effectiveness and efficiency of Deepkey in off-line situation, the future step is to build a real-time authentication environment to evaluate the online performance.
Furthermore, the proposed DeepKey still facing the challenge of `in the wild' scenario since the gathered EEG data are easily corrupted by physical actions like walking. In this work, the EEG and gait data are gathered in two separate steps. However, in the outdoor environments, the user is hardly stand still and wait for the authentication. Fortunately, our system has competitive performance in a fixed indoor environment (such as bank vouchers) which can provide a stable data collection environment and mainly concerned about the high fake-resistance.
\section{Conclusion}
\label{sec:conclusion}
Taking the advantages of both EEG- and gait-based systems for fake resistance, we propose Deepkey, a multimodal biometric authentication system, to overcome the limitations of traditional unimodal biometric authentication systems.
This authentication system Deepkey contains three independent models: an Invalid ID Filter Model, a Delta band based EEG Identification Model, and a Gait Identification Model, to detect invalid EEG data, recognize the EEG ID and Gait ID, respectively.
The Deepkey system outperforms the state-of-the-art baselines by achieving a FAR of $0$ and a FRR of $1\%$. In addition, the key parameters (such as sessions and EEG frequency band) and the system latency are also investigated by extensive experiments.
This work sheds the light on further research on multimodal biometric authentication systems based on EEG and gait data. Our future work will focus on deploying the Deepkey system in an online real-world environment. In addition, the gait signals are currently gathered by three wearable IMUs, which may obstruct the large-scale deployment in
practice. Therefore, another direction in the future is to collect gait data from non-wearable gait solutions (e.g., sensors deployed in environments).
\bibliographystyle{ACM-Reference-Format}
| 41cae1db66a99ad671d2afdd3a4f16b4caa88bb3 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Learning distributions is a fundamental problem in statistics and computer science, and has numerous applications in machine learning and signal processing.
The problem can be stated as:
{\begin{quote}{Given an i.i.d.\ sample generated from an unknown probability distribution $g$, find a distribution $\hat{g}$ that is close to $g$ in total variation distance.\footnote{Total variation distance is a prominent distance measure between distributions. For a discussion on this and other choices see~\cite[Chapter~5]{devroye_book}.
}
}\end{quote}}
This strong notion of learning is not possible in general using a finite number of samples. However, if we assume that the target distribution belongs to or can be approximated by a family of distributions, then there is hope to acquire algorithms with finite-sample guarantees. In this paper, we study the important family of mixture models within this framework.
Notice that we consider PAC learning of distributions (a.k.a.\ density estimation), which is different from parameter estimation.
In the parameter estimation problem, it is assumed that the target distribution belongs to some parametric class, and the goal is to learn/identify the parameters (see, e.g.,~\cite{dasgupta1999learning,Belkin,moitravaliant}).
As an example of our setting, assume that the target distribution is a Gaussian mixture with $k$ components in $\mathbb{R}^d$. Then, how many examples do we need to find a distribution that is $\epsilon$-close to the target? This \emph{sample complexity} question, as well as the corresponding \emph{computational complexity} question, has received a lot of attention recently (see, e.g.~\cite{axis_aligned,onedimensional,spherical,robustestimation,gaussian_mixture,nearlylinear}).
In this paper, we consider a scenario in which we are given a method for learning a class of distributions (e.g., Gaussians). Then, we ask whether we can use it, as a black box, to come up with an algorithm for learning a mixture of such distributions (e.g., mixture of Gaussians). We will show that the answer to this question is affirmative.
We propose a generic method for learning mixture models. Roughly speaking, we show that by going from learning a single distribution from a class
to learning a mixture of $k$ distributions from the same class, the sample complexity is multiplied by a factor of at most ${(k\log^2 k)}/{\epsilon^2}$. This result is general, and yet it is surprisingly tight in many important cases.
In this paper, we assume that the algorithm knows the number of components $k$.
As a demonstration, we show that our method provides a better sample complexity for learning mixtures of Gaussians than the state of the art. In particular, for learning mixtures of $k$ Gaussians in $\mathbb{R}^d$, our method requires $\widetilde{O}({d^2k}/{\epsilon^4})$ samples, improving by a factor of $k^2$ over the $\widetilde{O}({d^2k^3}/{\epsilon^4})$ bound of~\cite{gaussian_mixture}.
Furthermore, for the special case of mixtures of axis-aligned Gaussians, we provide an upper bound of $\widetilde{O}({dk}/{\epsilon^{4}})$, which is the first optimal bound with respect to $k$ and $d$ up to logarithmic factors, and improves upon the
$\widetilde{O}({dk^9}/{\epsilon^{4}})$ bound of~\cite{spherical}, which is only shown for the subclass of spherical Gaussians.
One merit of our approach is that it can be applied in the agnostic (a.k.a.\ robust) setting, where the target distribution does not necessarily belong to the mixture model of choice.
To guarantee such a result, we need the black box to work in the agnostic setting.
For example, an agnostic learning method for learning Gaussians can be lifted to work for Gaussian mixtures in the agnostic setting.
We would like to emphasize that our focus is on the information-theoretic aspects of learning rather than the computational ones; in particular, although our framework is algorithmic, its running time is exponential in the number of required samples. Proving sample complexity bounds is important in understanding the statistical nature of various classes of distributions (see, e.g., the recent work of \cite{logconcave}), and may pave the way for developing efficient algorithms.
\subsection{Our Results}
Let $\mathcal F$ be a class of probability distributions, and let $\mathcal{F}^k$ denote the class of $k$-mixtures of elements of $\mathcal F$.
In our main result, Theorem~\ref{thm:main}, assuming the existence of a method for learning $\mathcal F$ with sample complexity $m_{\mathcal{F}}(\epsilon)$,
we provide a method for learning $\mathcal F^k$ with sample complexity
$O({k\log^2 k \cdot m_{\mathcal F}(\varepsilon) }/{\epsilon^{2}})$.
Our mixture learning algorithm has the property that, if the $\mathcal F$-learner is proper, then the $\mathcal F^k$-learner would be proper as well (i.e., the learner will always output a member of $\mathcal{F}^k$).
Furthermore, the algorithm works in the more general agnostic setting provided that the base learners are agnostic learners.
We provide several applications of our main result.
In Theorem~\ref{axisalignedupperbound}, we show that the class of mixtures of $k$ axis-aligned Gaussians in $\mathbb{R}^d$
is PAC-learnable in the agnostic setting with
sample complexity $O({kd\log^2 k}/{\epsilon ^ 4})$ (see Theorem~\ref{thm:lowerbound}).
This bound is tight in terms of $k$ and $d$ up to logarithmic factors.
In Theorem~\ref{gaussianupperbound}, we show that the class of mixtures of $k$ Gaussians in $\mathbb{R}^d$
is PAC-learnable in the agnostic setting with
sample complexity
$O({kd^2\log^2 k}/{\epsilon ^ 4})$. Finally, in Theorem~\ref{logconcave}, we prove that the class of mixtures of $k$ log-concave distributions over $\mathbb{R}^d$ is PAC-learnable using $\widetilde{O}(d^{(d+5)/2}\varepsilon^{-(d+9)/2}k)$ samples. To the best of our knowledge, this is the first upper bound on the sample complexity of learning this class.
\subsection{Related Work}
PAC learning of distributions was introduced by~\cite{Kearns}, we refer the reader to~\cite{Diakonikolas2016} for a recent survey. A closely related line of research in statistics (in which more emphasis is on sample complexity) is density estimation, for which the book by \cite{devroye_book} is an excellent resource.
One approach for studying the sample complexity of learning a class of distributions is bounding the VC-dimension of its associated Yatrocas class (see Definition~\ref{def_yatrocas}), and applying results such as Theorem~\ref{thm:distributionVC}\footnote{These VC-dimensions have mainly been studied for the purpose of proving generalization bounds for neural networks with sigmoid activation functions.}. In particular, the VC-dimension bound of \cite[Theorem~8.14]{AB99} -- which is based on the work of~\cite{Karpinski} -- implies a sample complexity upper bound of
$O((k^4d^2+k^3d^3)/\varepsilon^2)$
for PAC learning mixtures of
axis-aligned Gaussians,
and an upper bound of $O(k^4d^4/\varepsilon^2)$ for PAC learning mixtures of general Gaussians (both results hold in the more general agnostic setting).
A sample complexity upper bound of $O(d^2 k^3 \log^2k / \varepsilon^4)$ for learning mixtures of Gaussians
in the realizable setting was proved in~\cite[Theorem~A.1]{gaussian_mixture} (the running time of their algorithm is not polynomial).
Our algorithm is motivated by theirs, but we have introduced several new ideas in the algorithm and in the analysis,
which has resulted in improving the sample complexity bound by a factor of $k^2$ and an algorithm that works in the more general agnostic setting.
For mixtures of spherical Gaussians,
a polynomial time algorithm for the realizable setting with sample complexity
$O(dk^9\log^2(d)/\varepsilon^4)$
was proposed in \cite[Theorem~11]{spherical}.
We improve their sample complexity by a factor of $\widetilde{O}(k^8)$, and moreover our algorithm works in the agnostic setting, too.
In the special case of $d=1$, a non-proper agnostic polynomial time algorithm with the optimal sample complexity of $\widetilde{O}(k/\varepsilon^2)$
was given in~\cite{onedimensional}, and a proper agnostic algorithm with the same sample complexity and better running time was given in~\cite{gmm_1d_proper}.
An important question, which we do not address in this paper, is finding polynomial time algorithms for learning distributions.
See~\cite{gaussian_mixture} for the state-of-the-art results on computational complexity of learning mixtures of Gaussians.
Another important setting is computational complexity in the agnostic learning, see, e.g.,~\cite{robustestimation} for some positive results.
A related line of research is parameter estimation for mixtures of Gaussians, see, e.g.,~\cite{dasgupta1999learning,Belkin,moitravaliant}, who gave polynomial time algorithms for this problem assuming certain separability conditions (these algorithms are polynomial in the dimension and the error tolerance but exponential in the number of components).
Recall that parameter estimation is a more difficult problem and any algorithm for parameter estimation requires some separability assumptions for the target Gaussians\footnote{E.g.,\ consider the case that $k=2$ and the two components are identical; then there is no way to learn their mixing weights.}, whereas for density estimation no such assumption is needed.
We finally remark that characterizing the sample complexity of learning a class of distributions in general is an open problem, even for the realizable (i.e., non-agnostic) case (see~\cite[Open Problem~15.1]{Diakonikolas2016}).
\section{The Formal Framework}
Generally speaking, a \emph{distribution learning method} is an algorithm that takes a
sample of i.i.d.\ points from distribution $g$
as input, and outputs (a description) of a distribution $\hat{g}$ as an estimation for $g$.
Furthermore, we assume that $g$ belongs to or can be approximated by class $\mathcal{F}$ of distributions,
and we may require that $\hat{g}$ also belongs to this class (i.e., proper learning).
Let $f_1$ and $f_2$ be two probability distributions defined over the Borel $\sigma$-algebra $\mathcal{B}$. The total variation distance between $f_1$ and $f_2$ is defined as
\[\|f_1- f_2\|_{TV} = \sup_{B\in \mathcal{B}}|f_1(B) - f_2(B)| = \frac{1}{2}\|f_1 - f_2\|_1 \:,
\]
where
\[\|f\|_1\coloneqq \int_{-\infty}^{+\infty} |f(x)|\mathrm{d} x\]
is the $L_1$ norm of $f$.
In the following definitions, $\mathcal{F}$ is a
class of probability distributions, and $g$ is a distribution not necessarily in $\mathcal{F}$.
Denote the set $\{1,2,...,m\}$ by $[m]$.
All logarithms are in the natural base.
For a function $g$ and a class of distributions $\mathcal{F}$, we define
\[
\opt(\mathcal{F}, g) \coloneqq \inf_{f\in \mathcal{F}}\|f-g\|_1
\]
\begin{definition}[$\varepsilon$-approximation, $(\varepsilon, C)$-approximation]
A distribution $\hat{g}$ is an \emph{$\varepsilon$-approximation} for $g$ if $ \|\hat{g}- g\|_1 \leq \varepsilon$.
A distribution $\hat{g}$ is an \emph{$(\epsilon, C)$-approximation} for $g$ with respect to $\mathcal{F}$ if
\[ \|\hat{g}- g\|_1 \leq C\times \opt(\mathcal{F}, g) + \varepsilon\]
\end{definition}
\begin{definition}[PAC-Learning Distributions, Realizable Setting]
A distribution learning method is called a \emph{(realizable) PAC-learner} for $\mathcal{F}$ with sample complexity $m_{\mathcal{F}}(\epsilon, \delta)$, if for all distribution $g\in\mathcal F$ and all $\epsilon, \delta >0$, given $\epsilon$, $\delta$, and a sample of size $m_{\mathcal{F}}(\epsilon, \delta)$, with probability at least $1-\delta$ outputs an $\epsilon$-approximation of $g$.
\end{definition}
\begin{definition}[PAC-Learning Distributions, Agnostic Setting]
For $C>0$, a distribution learning method is called a \emph{$C$-agnostic PAC-learner for $\mathcal{F}$} with sample complexity $m_{\mathcal{F}}^C(\epsilon, \delta)$, if for all distributions $g$ and all $\epsilon, \delta >0$, given $\epsilon$, $\delta$, and a sample of size $m_{\mathcal{F}}^C(\epsilon, \delta)$, with probability at least $1-\delta$ outputs an $(\epsilon, C)$-approximation of $g$.\footnote{Note that in some papers, only the case $C\leq1$ is called agnostic learning, and the case $C>1$ is called semi-agnostic learning.}
\end{definition}
Clearly, a $C$-agnostic PAC-learner (for any constant $C$)
is also a realizable PAC-learner, with the same error parameter $\varepsilon$.
Conversely a realizable PAC-learner can be thought of an $\infty$-agnostic PAC-learner.
\section{Learning Mixture Models}
Let $\Delta_n$ denote the $n$-dimensional simplex:
\[
\Delta_n \:= \{ (w_1,\dots,w_n) : w_i\geq 0, \sum_{i=1}^k w_i=1\}
\]
\begin{definition}
Let $\mathcal{F}$ be a class of probability distributions. Then the class of \emph{$k$-mixtures of $\mathcal{F}$}, written $\mathcal{F}^k$, is defined as
$$\mathcal{F}^k \coloneqq \left\{\sum_{i=1}^{k}w_{i}f_{i}: (w_1,\dots,w_k)\in \Delta_k ,
f_1,\dots,f_k\in\mathcal F
\right\}.$$
\end{definition}
Assume that we have a method to PAC-learn $\mathcal{F}$. Does this mean that we can PAC-learn $\mathcal{F}^k$? And if so, what is the sample complexity of this task?
Our main theorem gives an affirmative answer to the first question, and provides a bound for sample complexity of learning $\mathcal{F}^k$.
\begin{theorem}
\label{thm:main}
Assume that $\mathcal{F}$ has a $C$-agnostic PAC-learner
with sample complexity $m_{\mathcal{F}}^C(\epsilon, \delta) = {\lambda(\mathcal F,\delta)}/{\epsilon^\alpha}$ for some $C>0$, $\alpha \geq 1$ and some function
$\lambda(\mathcal F,\delta) = \Omega(\log(1/\delta))$.
Then there exists a $3C$-agnostic PAC-learner for the class
$\mathcal{F}^k$ requiring $m_{\mathcal{F}^k}^{3C}(\epsilon, \delta) =$
\[ O\left(\frac{\lambda (\mathcal F, \frac{\delta}{3k}) k\log k}{\epsilon^{\alpha+2}}\right)
=
O\left(\frac{k\log k \cdot m_{\mathcal F}(\varepsilon, \frac{\delta}{3k}) }{\epsilon^{2}}\right)
\]
samples.
\end{theorem}
Since a realizable PAC-learner is an $\infty$-agnostic PAC-learner, we immediately obtain the following corollary.
\begin{corollary}
Assume that $\mathcal{F}$ has a realizable PAC-learner
with sample complexity $m_{\mathcal{F}}(\epsilon, \delta) = {\lambda(\mathcal F,\delta)}/{\epsilon^\alpha}$ for some $\alpha \geq 1$ and some function
$\lambda(\mathcal F,\delta) = \Omega(\log(1/\delta))$.
Then there exists a realizable PAC-learner for the class
$\mathcal{F}^k$ requiring $m_{\mathcal{F}^k}(\epsilon, \delta) =$
\[ O\left(\frac{\lambda (\mathcal F, \frac{\delta}{3k}) k\log k}{\epsilon^{\alpha+2}}\right)
=
O\left(\frac{k\log k \cdot m_{\mathcal F}(\varepsilon, \frac{\delta}{3k}) }{\epsilon^{2}}\right)
\]
samples.
\end{corollary}
Some remarks:
\begin{enumerate}
\item
Our mixture learning algorithm has the property that, if the $\mathcal F$-learner is proper, then the $\mathcal F^k$-learner is proper as well.
\item
The computational complexity of the resulting algorithm is exponential in the number of required samples.
\item
The condition
$\lambda(\mathcal F,\delta) = \Omega(\log(1/\delta))$
is a technical condition that holds for all interesting classes $\mathcal F$.
\item
One may wonder about tightness of this theorem.
In Theorem~2 in \cite{spherical}, it is shown that
if $\mathcal F$ is the class of spherical Gaussians, we have
$m_{\mathcal{F}^k}^{O(1)}(\epsilon, \delta) = \Omega(k m_{\mathcal F}(\varepsilon, \delta/k) )$, therefore, the factor of $k$ is necessary in general. However, it is not clear whether the additional factor of
$\log k /\varepsilon^2$ in the theorem is tight.
\item
The constant 3 (in the $3C$-agnostic result) comes from
\cite[Theorem~6.3]{devroye_book} (see Theorem~\ref{thm:candidates}), and it is not clear whether it is necessary. If we allow for randomized algorithms (which produce a random distribution whose expected distance to the target is bounded by $\varepsilon$), then the constant can be improved to 2, see \cite[Theorem~22]{density_estimation_lineartime}.
\end{enumerate}
In the rest of this section we prove
Theorem~\ref{thm:main}.
Let $g$ be the true data generating distribution,
and let
\begin{equation}\label{def_rho}
g^* = \argmin_{f \in\mathcal{F}^k} \|g-f\|_1
\textnormal{ and }
\rho = \|g^*-g\|_1=\opt(\mathcal F^k,g)\:.
\end{equation}
Note that although $g^*\in\mathcal{F}^k$,
$g$ itself is not necessarily in the form of a mixture.
Since our algorithm works for mixtures, we would first like to write $g$ in the form of a mixture of $k$ distributions, such that they are on average close to being in $\mathcal{F}$.
This is done via the following lemma.
\begin{lemma}\label{lem:projection}
Suppose that $g$ is a probability density function with $\opt(\mathcal{F}^k,g)=\rho$.
Then we may write
\(g = \sum_{i\in [k]} w_i {G_i},\)
such that $w\in\Delta_k$,
each $G_i$ is a density, and
we have
$\sum_{i\in[k]} w_i \opt(\mathcal F,G_i) = \rho$.
\end{lemma}
\begin{proof}
Let $f\in \mathcal{F}^k$ be such that $\|g-f \|_1\leq \rho$, and let $X=\{x: g(x) < f(x)\}$.
Suppose $f = \sum_{i\in[k]} w_i f_i$, where $f_i\in \mathcal{F}$.
Define
\[
G_i(x) = \left\{\begin{array}{lr}
{f_i(x)g(x)}/{f(x)}, & \text{for } x\in X \\
f_i(x)+\Delta_i(x), & \text{for } x\notin X\\
\end{array}\right.
\]
where
\[
\Delta_i(x) = {(g(x)-f(x))\left(\int_X f_i(x)(f(x)-g(x)) \mathrm{d} x /f(x) \right)} \bigg/ \left({\int_X (f(x)-g(x)) \mathrm{d} x}\right).
\]
Observe that each $G_i$ is a density and that $g=\sum_{i\in[k]}w_i G_i$.
Finally, note that $f_i(x)>G_i(x)$ precisely on $X$. Thus,
\begin{align*} \rho=\|g-f \|_1 =
2\int_X (f(x) - \sum_i w_i G_i(&x))
=2\int_X \sum_i w_i (f_i(x)-G_i(x)) \\&=2\int_X \sum_i w_i |f_i(x)-G_i(x)| = \sum_i w_i \|f_i-G_i\|_1. \qedhere
\end{align*}
\end{proof}
By Lemma~\ref{lem:projection}, we have
\(g = \sum_{i\in [k]} w_i {G_i},\)
where each $G_i$ is a probability distribution.
Let \(\rho_i \coloneqq \opt (\mathcal{F}, G_i)\),
and by the lemma we have
\begin{equation}
\sum_{i \in[k]} w_i \rho_i = \rho.\label{sumrhoiisrho}.
\end{equation}
The idea now is to learn each of the $G_i$'s separately using the agnostic learner for $\mathcal{F}$.
We will view $g$ as a mixture of $k$ distributions
$G_1,G_2,\dots,G_k$.
For proving Theorem~\ref{thm:main}, we will use the following theorem on learning finite classes of distributions,
which immediately follows from \cite[Theorem~6.3]{devroye_book} and a standard Chernoff bound.
\begin{theorem}\label{thm:candidates}
Suppose we are given $M$ candidate distributions $f_1,\dots,f_M$ and we have access to i.i.d.\ samples from an unknown distribution $g$.
Then there exists an algorithm that given the $f_i$'s and $\varepsilon>0$, takes
$\log (3M^2/\delta)/2\varepsilon^2$ samples from $g$, and with probability $\geq 1-\delta/3$ outputs an index $j\in[M]$ such that
\[
\|f_j-g\|_1 \leq 3 \min_{i\in[M]} \|f_i-g\|_1 + 4\varepsilon \:.
\]
\end{theorem}
We now describe an algorithm that with probability $\geq 1-\delta$ outputs a distribution with $L_1$ distance $13\varepsilon + 3C\rho$ to $g$ (the error parameter is $13\varepsilon$ instead of $\varepsilon$ just for convenience of the proof; it is clear that this does not change the order of magnitude of sample complexity).
The algorithm, whose pseudocode is shown in Figure~\ref{fig:alg}, has two main steps.
In the first step we generate a set of candidate distributions, such that at least one of them is $(3\varepsilon+\rho)$-close to $g$ in $L_1$ distance.
These candidates are of the form $\sum_{i=1}^{k} \widehat{w}_i \widehat{G}_i$, where the $\widehat{G}_i$'s are extracted from samples and are estimates for the real components $G_i$,
and the $\widehat{w}_i$'s come from a fixed discretization of $\Delta_k$, and are estimates for the real mixing weights $w_i$.
In the second step, we use Theorem~\ref{thm:candidates} to obtain a distribution that is $(13\varepsilon+3C\rho)$-close to $g$.
\begin{figure}
\begin{tcolorbox}
Input: $k, \epsilon, \delta$ and an iid sample $S$\\
0. Let $\widehat{W}$ be an $(\epsilon/k)$-cover for $\Delta_k$ in $\ell_{\infty}$ distance.\\
1. $\mathcal C = \emptyset$ (set of candidate distributions)\\
2. For each $(\widehat{w}_1,\dots,\widehat{w}_k)\in \widehat{W}$ do:\\
\phantom{aa}3. For each possible partition of $S$ into \\
\phantom{aa aaaaaa}$A_1, A_2, ...,A_{k}$:\\
\phantom{aaaa}4. Provide $A_i$ to the $\mathcal F$-learner, and let $\widehat{G}_i$ \\\phantom{aaaa aaaaaa}be its output.\\
\phantom{aaaa}5. Add the candidate distribution \\ \phantom{aaaa aaaaaa}$ \sum_{i\in [k]} \widehat{w}_i \widehat{G}_i $ to $\mathcal C$.\\
6. Apply the algorithm for finite classes (Theorem~\ref{thm:candidates}) to $\mathcal C$ and output its result.
\end{tcolorbox}
\caption{Algorithm for learning the mixture class $\mathcal F^k$}
\label{fig:alg}
\end{figure}
We start with describing the first step.
We take
\begin{equation}
\label{s_def}
s=
\max \left \{
\frac
{2k\lambda(\mathcal F, \delta/3k)}
{ \varepsilon^{\alpha} },
\frac{16 k \log(3k/\delta)}{\varepsilon}
\right\}
\end{equation}
i.i.d.\ samples from $g$.
Let $S$ denote the set of generated points.
Note that $\lambda(\mathcal F,\delta)=\Omega(\log(1/\delta))$ implies
\[s=O({k\lambda (\mathcal F, \delta/3k) }\times {\epsilon^{-\alpha}}).\]
Let $\widehat{W}$ be an $\varepsilon/k$-cover for $\Delta_k$ in $\ell_{\infty}$ distance of cardinality
$(k/\varepsilon+1)^k$.
That is, for any $x\in \Delta_k$ there exists $w\in\widehat{W}$ such that $\|w-x\|_{\infty}\leq\varepsilon/k$.
This can be obtained from a grid in $[0,1]^k$ of side length $\varepsilon/k$, which is an $\varepsilon/k$-cover for $[0,1]^k$, and projecting each of its points onto $\Delta_k$.
By an \emph{assignment}, we mean a function
$A:S\to [k]$.
The role of an assignment is to ``guess'' each sample point is coming from which component, by mapping them to a component index.
For each pair $(A,(\widehat{w}_1,\dots,\widehat{w}_k))$, where
$A$ is an assignment and $(\widehat{w}_1,\dots,\widehat{w}_k)\in \widehat{W}$,
we generate a candidate distribution as follows:
let $A^{-1}(i)\subseteq S$ be those sample points that are assigned to component $i$.
For each $i\in[k]$, we provide the set $A^{-1}(i)$ of samples to our $\mathcal F$-learner, and the learner provides us with a distribution $\widehat{G}_i$.
We add the distribution $\sum_{i\in [k]} \widehat{w}_i \widehat{G}_i $
to the set of candidate distributions.
\begin{lemma}
\label{somecandidateisclose}
With probability $\geq 1-2\delta/3$, at least one of the generated candidate distributions is $(3\varepsilon+C\rho)$-close to $g$.
\end{lemma}
Before proving the lemma, we show that it implies our main result, Theorem~\ref{thm:main}.
By the lemma, we obtain a set of candidates such that at least one of them is $(3\varepsilon+C\rho)$-close to $g$ (with failure probability $ \leq 2\delta/3$).
This step takes
$s=
O({k\lambda (\mathcal F, \delta/3k) }\times {\epsilon^{-\alpha}})$ many samples.
Then, we apply Theorem~\ref{thm:candidates} to output one of those candidates that is $(13\varepsilon+3C\rho)$-close to $g$ (with failure probability $\leq \delta/3$), therefore using $\log(3M^2/\delta)/2\varepsilon^2$ additional
samples.
Note that the number of generated candidate distributions is
$M = k^s \times (1+k/\varepsilon)^k$.
Hence, in the second step of our algorithm,
we take
\begin{align*}
\log(3M^2/\delta)/2\varepsilon^2
=
O\left(\frac{\lambda (\mathcal F, \delta/3k) k\log k}{\epsilon^{\alpha+2}}\right)
=
O\left(\frac{m_{\mathcal F}(\varepsilon, \delta/3k) k\log k}{\epsilon^{2}}\right)
\end{align*}
additional samples.
The proof is completed noting the total failure probability is at most $\delta$ by the union bound.
We now prove Lemma~\ref{somecandidateisclose}.
We will use the following concentration inequality, which holds for any binomial random variable $X$ (see \cite[Theorem~4.5(2)]{Mitzenmacher}):
\begin{equation}
\label{chernoffhalf}
\Pr\{X < \mathbf{E} X/2\} \leq \exp(-\mathbf{E} X/8)\:.
\end{equation}
Say a component $i$ is \emph{negligible}
if
\[ w_i \leq \frac{8 \log(3k/\delta)}{s}
\]
Let $L\subseteq[k]$ denote the set of negligible components.
Let $i$ be a non-negligible component.
Note that, the number of points coming from component $i$
is binomial with parameters
$s$ and $w_i$ and thus has mean
$ s w_i$,
so (\ref{chernoffhalf})
implies that, with probability at least $1-\delta/3k$, $S$ contains at least
$w_is/2 $ points from $i$.
Since we have $k$ components in total, the union bound implies that,
with probability at least $1-\delta/3$,
uniformly for all $i\notin L$,
$S$ contains at least
$w_is/2 $ points from component $i$.
Now consider the pair $(A,(\widehat{w}_1,\dots,\widehat{w}_k))$ such that $A$ assigns samples to their correct indices,
and has the property that $|\widehat{w}_i-w_i|\leq\varepsilon/k$ for all $i\in[k]$.
We claim that the resulting candidate distribution is $(3\varepsilon+C\rho)$-close to $g$.
Let $\widehat{G}_1,\dots,\widehat{G}_k$ be the distributions provided by the learner.
For each $i\in[k]$ define
\[
\varepsilon_i \coloneqq \left ( \frac
{2\lambda(\mathcal F, \delta/3k)}
{ w_is }
\right)^{1/\alpha}
\]
For any $i\notin L$,
since there exists at least
$ w_is/2 $ samples for component $i$,
and since
\[
w_is/2 = \lambda(\mathcal F, \delta/3k) \varepsilon_i^{-\alpha} = m_{\mathcal F} (\varepsilon_i, \delta/3k)\:,
\]
we are guaranteed that
$\|\widehat{G}_i-G_i\|_1\leq C\rho_i + \varepsilon_i$ with probability $1-\delta/3k$
(recall that each $G_i$ is $\rho_i$-close to the class $\mathcal{F}$).
Therefore,
$\|\widehat{G}_i-G_i\|_1\leq C\rho_i+\varepsilon_i$ holds uniformly over all $i\notin L$, with probability $\geq 1-\delta/3$.
Note that since $\alpha\geq1$, the function $w_i^{1-1/\alpha}$ is concave in $w_i$, so by Jensen's inequality we have
\[
\sum_{i\in [k]} w_i^{1-1/\alpha}
\leq k \left ( (\sum_{i\in [k]} w_i / k)^{1-1/\alpha}\right)
= k^{1/\alpha}\:,
\]
hence
\begin{align*}
\sum_{i\notin L}
w_i \varepsilon_i =
\left ( \frac
{2\lambda(\mathcal F, \delta/3k)}
{s}
\right)^{1/\alpha}
\sum_{i\notin L} w_i^{1-1/\alpha}
\leq
\left ( \frac
{2k\lambda(\mathcal F, \delta/3k)}
{s }
\right)^{1/\alpha}.
\end{align*}
Also recall from (\ref{sumrhoiisrho}) that $\sum_{i \in[k]}w_i \rho_i \leq \rho$.
Proving the lemma is now a matter of careful applications of the triangle inequality:
\begin{align*}
\left\| \sum_{i\in [k]} \widehat{w}_i \widehat{G}_i
- g\right\|_1
&=
\left\| \sum_{i\in [k]} \widehat{w}_i \widehat{G}_i
-
\sum_{i\in [k]} w_i G_i
\right\|_1\\
&
\leq
\left\|\sum_{i\in [k]} {w}_i (\widehat{G}_i
- G_i)
\right\|_1
+
\left\|\sum_{i\in [k]} (\widehat{w}_i-w_i) \widehat{G}_i
\right\|_1
\\
& \leq
\left\|\sum_{i\in L}{w_i} (\widehat{G}_i
- G_i)
\right\|_1
+
\left\|\sum_{i\notin L}{w_i} (\widehat{G}_i
- G_i)
\right\|_1
+
\sum_{i\in [k]} |\widehat{w}_i-w_i| \left\|\widehat{G}_i
\right\|_1
\\
& \leq
2\sum_{i\in L}{w_i}
+
\sum_{i\notin L}{w_i} (\varepsilon_i + C\rho_i)
+
\sum_{i\in [k]} \varepsilon/k \times 1
\\
&\leq
2k \times \frac{8 \log(3k/\delta)}{s}
+
\left ( \frac
{2k\lambda(\mathcal F, \delta/3k)}
{ s }
\right)^{1/\alpha}
+ C \rho + \varepsilon
\\
&\leq \varepsilon+\varepsilon+\varepsilon + C \rho\:,
\end{align*}
where for the last inequality we used the definition of $s$ in (\ref{s_def}).
This completes the proof of Lemma~\ref{somecandidateisclose}.
\section{Learning Mixtures of Gaussians}
Gaussian Mixture Models (GMMs) are probably the most widely studied mixture classes with numerous applications; yet, the sample complexity of learning this class is not fully understood, especially when the number of dimensions is large. In this section, we will show that our method for learning mixtures can improve the state of the art for learning GMMs in terms of sample complexity. In the following, $\mathcal{N}_d(\mu, \Sigma)$ denotes a Gaussian density function defined over $\mathbb{R}^d$, with mean $\mu$ and covariance matrix $\Sigma$.
\subsection{Mixture of Axis-Aligned Gaussians}
\label{sec:axisaligned}
A Gaussian is called \emph{axis-aligned} if its covariance matrix $\Sigma$ is diagonal.
The class of axis-aligned Gaussian Mixtures is an important special case of GMMs that is thoroughly studied in the literature (e.g.~\cite{axis_aligned}).
\begin{theorem}{\label{thm:AAGaussianUpperBound}}
Let $\mathcal F$ denote the class of $d$-dimensional axis-aligned Gaussians.
Then $\mathcal F$ is 3-agnostic PAC-learnable with
$m_{\mathcal F}^3(\varepsilon,\delta)=O((d+\log(1/\delta))/{\epsilon ^ 2})$.
\end{theorem}
We defer the proof of this result to the appendix.
Combining this theorem with Theorem~\ref{thm:main} we obtain the following result:
\begin{theorem}
\label{axisalignedupperbound}
The class $\mathcal{F}^k$ of mixtures of $k$ axis-aligned Gaussians in $\mathbb{R}^d$
is 9-agnostic PAC-learnable with
sample complexity
$m_{\mathcal F^k}^9(\epsilon,\delta) = O({kd\log k \log(k/\delta)}/{\epsilon ^ 4})$. Accordingly, it is also PAC-learnable in the realizable case with the same number of samples.
\end{theorem}
This theorem improves the upper bound of
$O(dk^9\log^2(d/\delta)/\varepsilon^4)$
proved in \cite[Theorem~11]{spherical} for spherical Gaussians in the realizable setting.
Spherical Gaussians are special cases of axis-aligned Gaussians in which all eigenvalues of the covariance matrix are equal, i.e., $\Sigma$ is a multiple of the identity matrix.
The following minimax lower bound (i.e., worst-case on all instances) on the sample complexity of learning mixtures of spherical Gaussians is proved in the same paper.
\begin{theorem}[Theorem~2 in \cite{spherical}]\label{thm:lowerbound}
The class $\mathcal{F}^k$ of mixtures of $k$ axis-aligned Gaussians in $\mathbb{R}^d$
in the realizable setting has
$m_{\mathcal{F}^k}(\epsilon,1/2)=\Omega(dk/\epsilon^2)$.
\end{theorem}
Therefore, our upper bound of Theorem~\ref{axisalignedupperbound} is optimal in terms of dependence on $d$ and $k$ (up to logarithmic factors) for axis-aligned Gaussians.
\subsection{Mixture of General Gaussians}
For general Gaussians, we have the following result.
\begin{theorem}{\label{thm:GaussianUpperBound}}
Let $\mathcal F$ denote the class of $d$-dimensional Gaussians.
Then, $\mathcal F$ is 3-agnostic PAC-learnable with
$m_{\mathcal F}^3(\varepsilon,\delta)=O((d^2+\log(1/\delta))/{\epsilon ^ 2})$.
\end{theorem}
We defer the proof of this result to the appendix.
Combining this theorem with Theorem~\ref{thm:main}, we obtain the following result:
\begin{theorem}
\label{gaussianupperbound}
The class $\mathcal{F}^k$ of mixtures of $k$ Gaussians in $\mathbb{R}^d$
is 9-agnostic PAC-learnable with
sample complexity
$m_{\mathcal F^k}^9(\epsilon,\delta) = O({kd^2\log k \log(k/\delta)}/{\epsilon ^ 4})$. Accordingly, it is also PAC-learnable in the realizable case with the same number of samples.
\end{theorem}
This improves by a factor of $k^2$ the upper bound
of $O(k^3d^2 \log k / \varepsilon^4)$
in the realizable setting,
proved in~\cite[Theorem~A.1]{gaussian_mixture}.
Note that Theorem \ref{thm:lowerbound}
gives a lower bound of
$\Omega(kd/{\epsilon ^ 2})$ for
$m_{\mathcal F^k}(\epsilon,\delta)$,
hence the dependence of Theorem~\ref{gaussianupperbound} on $k$ is optimal (up to logarithmic factors).
However, there is a factor of $d/\epsilon^2$ between the upper and lower bounds.
\section{Mixtures of Log-Concave Distributions}
A probability density function over $\mathbb{R}^d$ is log-concave if its logarithm is a concave function. The following result about the sample complexity of learning log-concave distributions is the direct consequence of the recent work of \cite{logconcave}.
\begin{theorem} Let $\mathcal{F}$ be the class of distributions corresponding to the set of all log-concave pdfs over $\mathbb{R}^d$. Then $\mathcal{F}$ is $3$-agnostic PAC learnable using $m^{3}(\epsilon, \delta)=O((d/\varepsilon)^{(d+5)/2}\log^2(1/\varepsilon))$ samples.
\end{theorem}
Using Theorem~\ref{thm:main}, we come up with the first result about the sample complexity of learning mixtures of log-concave distributions.
\begin{theorem}\label{logconcave}
The class of mixtures of $k$ log-concave distributions over $\mathbb{R}^d$ is $9$-agnostic PAC-learnable using $\widetilde{O}(d^{(d+5)/2}\varepsilon^{-(d+9)/2}k)$ samples.
\end{theorem}
\section{Conclusions}
We studied PAC learning of classes of distributions that are in the form of mixture models, and proposed a generic approach for learning such classes in the cases where we have access to a black box method for learning a single-component distribution. We showed that by going from one component to a mixture model with $k$ components, the sample complexity is multiplied by a factor of at most $(k\log^2 k)/\epsilon^2$.
Furthermore, as a corollary of this general result, we provided upper bounds for the sample complexity of learning GMMs and axis-aligned GMMs---$O({kd^2\log^2 k}/{\epsilon ^ 4})$ and $O({kd\log^2 k}/{\epsilon ^ 4})$ respectively. Both of these results improve upon the state of the art in terms of dependence on $k$ and $d$.
It is worthwhile to note that for the case of GMMs, the dependence of our bound is $1/ \epsilon^4$. Therefore, proving an upper bound of $kd^2/\epsilon^2$ remains open.
Also, note that our result can be readily applied to the general case of mixtures of the exponential family. Let $\mathcal{F}_d$ denote the $d$-parameter exponential family. Then the VC-dimension of the corresponding Yatrocas class (see Definition~\ref{def_yatrocas}) is $O(d)$ (see Theorem 8.1 in \cite{devroye_book}) and therefore by Theorem \ref{thm:distributionVC}, the sample complexity of PAC learning $\mathcal{F}_d$ is $O(d/\epsilon^2)$. Finally, applying Theorem \ref{thm:main} gives a sample complexity upper bound of $\widetilde{O}(k d/ \epsilon^4)$ for learning $\mathcal{F}_d^k$
\paragraph{Addendum.}
After an earlier version of this work was presented in AAAI 2018 (and appeared on arXiv), we obtained improved results for learning mixtures of Gaussians: the class of mixtures of $k$ axis-aligned
Gaussians in $\mathbb{R}^d$ is agnostic PAC-learnable with sample
complexity $\widetilde{O}(kd/\varepsilon^2)$,
and
the class of mixtures of $k$ general
Gaussians in $\mathbb{R}^d$ is agnostic PAC-learnable with sample
complexity $\widetilde{O}(kd^2/\varepsilon^2)$,
The proof uses novel techniques,
see~\cite{2017-abbas} for details.
| 760bd5ebebea86c195646b36b6032fcb2e81b70f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:intro}
In multi-user communication, the signal received by any user (UE) is a superposition of the desired information and the information intended for other UEs. Transmit beamforming at a base station (BS) plays a pivotal role in
focusing the energy of the desired signal and suppressing the multi-user interference to achieve better throughput at UEs \cite{ShangIT11,GSO}.
For effective beamforming, the number of UEs
usually does not exceed the number of transmit antennas. Using more transmit antennas and thus increasing the dimensionality of beamforming vectors can improve the UEs' throughput with reduced transmit power. However, allocating more transmit power under fixed number of transmit antennas still does not necessarily improve the UEs' throughput.
Non-orthogonal multiple access (NOMA) \cite{DSP16,DingMag17} has been introduced to improve
the UEs' throughput by allowing UEs with better channel conditions to access
and decode the signals, which are intended for the UEs with poorer channel conditions. In other words, the UEs with
poorer channel conditions can achieve higher throughput by compromising their information privacy in NOMA \cite{DingTWC16}.
This privacy compromise for the UEs with poor channel conditions is unavoidable in NOMA. Moreover, by restricting all beamforming vectors in the same space, beamforming in NOMA still needs that the dimension of this space, which is
equal to the number of transmit antennas, should not be less than the number of UEs to allow the suppression of multiuser interference \cite{Dietal17}.
It is noteworthy that the UEs with good channel conditions may need only a fraction of the time unit to achieve their needed throughput.
Therefore, by serving them only per fractional time, the BS still has the remaining time room to serve
the UEs with poor channel conditions. In this way, the information privacy for each UE is preserved because all UEs are allowed to
decode their own information only. More importantly, the number of UEs supported at the same fractional
time is effectively reduced. Thus, the BS will not need more transmit antennas to suppress the multi-user interference.
In this letter, the fractional time allocation to UEs with similar channel conditions and beamforming are enhanced
for improving the system's throughput and meeting the UEs' quality-of-service (QoS) in terms of the throughput requirement.
This problem is mathematically modelled by a highly nonconvex optimization problem, for which a path-following computational procedure
of low complexity is then developed for its computation. Finally, the numerical examples are provided to demonstrate the
advantage of the proposed optimization scheme.
\textit{Notation.} We use bold upper-case letters for matrices, bold lower-case letters for column vectors, lower-case letters for scalars. $\Re\{x\}$ denotes the real part of a complex number $x$. The notation $(\cdot)^{H}$ stands for the Hermitian transpose. $\mathbf{x}\sim\mathcal{CN}(\boldsymbol{\eta},\boldsymbol{Z})$ means that $\mathbf{x}$ is a random vector following a complex circular Gaussian distribution with mean $\boldsymbol{\eta}$ and covariance matrix $\boldsymbol{Z}$.
\section{System Model and Problem Formulation}\label{sec:sys_model}
Consider a multiuser downlink system over a given frequency band with
a BS equipped with $N_t > 1$ antennas in serving $2K$ single-antenna UEs as illustrated by Fig. \ref{fig:Layout}.
There are $K$ UEs $(1,k)$, $k=1,\dots K$, which are located in a zone nearer to the BS, called by zone-1,
and $K$ UEs $(2,k)$, $k=1,\dots, K$, which are located in a zone farer from the BS, called by zone-2.
Denote by $\mathcal{K} \triangleq\{1,2,\dots, K\}$ and $\mathcal{M}=\{1,2\}\times \mathcal{K}$.
Within one time unit, BS uses the fraction time (FT) $\tau_1:=\tau\;(0 < \tau < 1)$ to serve UEs $(1,k)$
and uses the remaining FT $\tau_2:=(1 - \tau)$ to serve UEs $(2,k)$.
The BS deploys a transmit beamformer $\mathbf{w}_{i,k}\in\mathbb{C}^{N_t\times 1}$ to deliver the information signal $x_{i,k}$ with $\mathbb{E}\{|x_{i,k}|^2 \} =1$ to UE $(i,k)$.
Let $\mathbf{h}_{i,k}\in\mathbb{C}^{N_t\times 1}$ be the channel vector from the BS to UE $(i,k)$, which is assumed to follow frequency flat fading with the effects of both large-scale pathloss and small-scale fading counted. The complex baseband signal received by UE $(i,k)$ can be expressed as
\[
y_{i,k} = \mathbf{h}_{i,k}^H\mathbf{w}_{i,k}x_{i,k} + \sum_{j\in\mathcal{K}\setminus \{k\}}\mathbf{h}_{i,k}^H\mathbf{w}_{i,j}x_{i,j} + n_{i,k}
\]
where the first term is the desired signal, the second term is the multi-user interference, and the third term
$n_{i,k}\sim\mathcal{CN}(0,\sigma^2_{i,k})$ is additive noise.
{\color{black}For $\mathbf{w}_i\triangleq (\mathbf{w}_{i,k})_{k\in{\cal K}}$,
the throughput at UE $(i,k)$ is
\[
{\cal R}_{i,k}\bigl(\mathbf{w}_i,\tau_i \bigr) = \tau_i\ln\Bigl(1+\frac{ |\mathbf{h}_{i,k}^H\mathbf{w}_{i,k}|^2}{\sum_{j\in\mathcal{K}\setminus \{k\}}|\mathbf{h}_{i,k}^H\mathbf{w}_{i,j}|^2 + \sigma^2_{i,k}}\Bigl)
\]
which by \cite{WLP} can be equivalently reformulated by
\begin{equation}\label{frate}
{\cal R}_{i,k}\bigl(\mathbf{w}_i,\tau_i \bigr) = \tau_i\ln\Bigl(1+\frac{ (\Re\{\mathbf{h}_{i,k}^H\mathbf{w}_{i,k}\})^2}{\sum_{j\in\mathcal{K}\setminus \{k\}}|\mathbf{h}_{i,k}^H\mathbf{w}_{i,j}|^2 + \sigma^2_{i,k}}\Bigl)
\end{equation}
under the additional condition
\begin{equation}\label{ad}
\Re\{\mathbf{h}_{i,k}^H\mathbf{w}_{i,k}\} \geq 0, (i,k)\in {\cal M}.
\end{equation}
}
The main advantage of this FT-based beamforming scheme is that there is no inter-zone interference in (\ref{frate})
that is in contrast with the
conventional scheme to concurrently serve all UEs, under which the throughput at UE $(i,k)$ is
\begin{equation}\label{crate}
{\cal R}_{i,k}^{'}(\mathbf{w}) = \ln\Bigl(1+ \frac{(\Re\{\mathbf{h}_{i,k}^H\mathbf{w}_{i,k}\})^2}{\sum_{(i,j)\in\mathcal{M}\setminus \{(i,k)\}}|\mathbf{h}_{i,k}^H\mathbf{w}_{i,j}|^2 + \sigma^2_{i,k}}\Bigl)
\end{equation}
with the full inter-zone interference. Here and in the sequence $\mathbf{w}\triangleq (\mathbf{w}_1, \mathbf{w}_2)$ and $\boldsymbol{\tau}\triangleq(\tau_1,\tau_2)$.
We are interested in the following problem of jointly designing FT $(\tau_1,\tau_2)$ and the beamformers $(\mathbf{w}_1, \mathbf{w}_2)$ to maximize the system sum throughput (ST):
\begin{subequations}\label{eq:SRoptimi}
\begin{eqnarray}
&\underset{\mathbf{w},\boldsymbol{\tau}}{\max}& \sum_{(i,k)\in{\cal M}} {\cal R}_{i,k}(\mathbf{w}_i,\tau_i) \quad {\mathrm{s.t.}} \quad (\ref{ad}), \label{eq:SRoptimi:a}\\
&&{\cal R}_{i,k}\bigl(\mathbf{w}_i,\tau_i \bigr) \geq {\mathsf{\bar R}}_{i,k},\ \forall(i,k)\in\mathcal{M}, \label{eq:SRoptimi:b}\\
&&\tau_1\|\mathbf{w}_{1}\|^2 + \tau_2\|\mathbf{w}_{2}\|^2 \leq P_{bs}^{\max},\label{eq:SRoptimi:d}\\
&& \tau_1\geq 0, \tau_2\geq 0, \tau_1+\tau_2\leq 1. \label{eq:SRoptimi:e}
\end{eqnarray}
\end{subequations}
Here ${\mathsf{\bar R}}_{i,k}$ sets a minimum throughput requirement for UE $(i,k)$ and $P_{bs}^{\max}$ is a given power budget.
Since ${\cal R}_{i,k}(\mathbf{w}_i,\tau_i)$ is a nonconcave function, the optimization problem (\ref{eq:SRoptimi}) is regarded
as a highly nonconvex optimization problem, for which finding a feasible point is already computationally difficult. The next
section is devoted to a computational path-following procedure for its solution.
\begin{figure}[t]
\centering
\includegraphics[width=0.28\textwidth,trim={0cm 0.0cm -0cm -0cm}]{Layout.eps}
\caption{Scenario with $K=4$.}
\label{fig:Layout}
\end{figure}
\section{Convex Quadratic-based Path-following method}\label{sec:CQBI}
The most important step is to provide an effective lower bounding approximation for the function
${\cal R}_{i,k}(\mathbf{w}_i,\tau_i)$ defined by (\ref{frate}) to facilitate
a path-following computational procedure of the problem (\ref{eq:SRoptimi}).
We use the variable changes $\pmb{\alpha}\triangleq (\alpha_1, \alpha_2)=(1/\tau_1,1/\tau_2)$,
which satisfy the following convex constraint:
\begin{equation}\label{eq:changevariables:b}
1/\alpha_1+1/\alpha_2 \leq 1, \alpha_1>0, \alpha_2>0.
\end{equation}
The problem \eqref{eq:SRoptimi} can be equivalently expressed as
\begin{subequations}\label{eq:SRMequi}
\begin{eqnarray}
&\underset{\mathbf{w}, \pmb{\alpha}}{\max}& \Phi(\mathbf{w},\pmb{\alpha})\triangleq \sum_{(i,k)\in{\cal M}}
{\cal R}_{i,k}(\mathbf{w}_i,1/\alpha_i)\quad {\mathrm{s.t.}}\quad (\ref{ad}), \eqref{eq:changevariables:b},\label{eq:SRMequi:a}\qquad\\
&&{\cal R}_{i,k}(\mathbf{w}_i,1/\alpha_i) \geq {\mathsf{\bar R}}_{i,k},\ \forall(i,k)\in{\cal M},\label{eq:SRMequi:b}\\
&&(1-1/\alpha_2)\|\mathbf{w}_{1}\|^2 + \|\mathbf{w}_{2}\|^2/\alpha_2 \leq P_{bs}^{\max}.\label{eq:SRMequi:d}
\end{eqnarray}
\end{subequations}
At a feasible point $(\mathbf{w}^{(\kappa)},\pmb{\alpha}^{(\kappa)})$, define
$x_{i,k}^{(\kappa)}\triangleq \Re\{\mathbf{h}_{i,k}^H\mathbf{w}_{i,k}^{(\kappa)}\}>0$,
$y_{i,k}^{(\kappa)}\triangleq\sum_{j\in\mathcal{K}\setminus \{k\}}|\mathbf{h}_{i,k}^H\mathbf{w}_{i,j}^{(\kappa)} |^2 + \sigma^2_{i,k} $,
$d_{i,k}^{(\kappa)} \triangleq (x_{i,k}^{(\kappa)})^2/y_{i,k}^{(\kappa)}>0$,
$a^{(\kappa)}_{i,k}\triangleq 2{\cal R}_{i,k}(\mathbf{w}_i^{(\kappa)},1/\alpha_i^{(\kappa)})+d_{i,k}^{(\kappa)}/\alpha_i^{(\kappa)}(d_{i,k}^{(\kappa)}+1)>0$,
$b^{(\kappa)}_{i,k}\triangleq (d_{i,k}^{(\kappa)})^2/\alpha_i^{(\kappa)}(d_{i,k}^{(\kappa)}+1)>0$, and
$c_{i,k}^{(\kappa)}\triangleq {\cal R}_{i,k}(\mathbf{w}_i^{(\kappa)},1/(\alpha_i^{(\kappa)})^2)>0$.
It follows from the inequality \eqref{inq1} in the appendix
that
\[
{\cal R}_{i,k}(\mathbf{w}_i,1/\alpha_i)\geq {\cal R}_{i,k}^{(\kappa)}(\mathbf{w}_i,\alpha_i)
\]
over the trust region
\begin{IEEEeqnarray}{rCl}
2\Re\{\mathbf{h}_{i,k}^H\mathbf{w}_{i,k}\} - \Re\{\mathbf{h}_{i,k}^H\mathbf{w}_{i,k}^{(\kappa)}\} > 0,\ \forall(i,k) \in\mathcal{M},
\label{eq:Rate1ktrust}
\end{IEEEeqnarray}
for the concave function
\[
{\cal R}_{i,k}^{(\kappa)}(\mathbf{w}_i,\alpha_i)\triangleq
a^{(\kappa)}_{i,k}-
b^{(\kappa)}_{i,k}\frac{\sum_{j\in\mathcal{K}\setminus \{k\}}|\mathbf{h}_{i,k}^H\mathbf{w}_{i,j}|^2 + \sigma^2_{i,k} }{x_{i,k}^{(\kappa)}(2\Re\{\mathbf{h}_{i,k}^H\mathbf{w}_{i,k}\}-x_{i,k}^{(\kappa)})}
- c_{i,k}^{(\kappa)}\alpha_i.
\]
Next, due to the convexity of function $\|\mathbf{w}_{1}\|^2/\alpha_2$, it is true that
$ \|\mathbf{w}_{1}\|^2/\alpha_2\geq 2\Re\{(\mathbf{w}_{1}^{(\kappa)})^H\mathbf{w}_{1}\}/\alpha_2^{(\kappa)}- (\|\mathbf{w}_{1}^{(\kappa)}\|^2/(\alpha_2^{(\kappa)})^2)\alpha_2$.
An inner convex approximation of nonconvex constraint \eqref{eq:SRMequi:d} is then given by
\begin{eqnarray}
\|\mathbf{w}_{1}\|^2 + \|\mathbf{w}_{2}\|^2/\alpha_2 - 2\Re\{(\mathbf{w}_{1}^{(\kappa)})^H\mathbf{w}_{1}\}/\alpha_2^{(\kappa)}&\nonumber\\
+ (\|\mathbf{w}_{1}^{(\kappa)}\|^2/(\alpha_2^{(\kappa)})^2)\alpha_2 &\leq P_{bs}^{\max}.\label{eq:SRMequi:d2}
\end{eqnarray}
Initialized by a feasible point $(\mathbf{w}^{(0)},\pmb{\alpha}^{(0)})$ for (\ref{eq:SRMequi}), the following convex quadratic program (QP)
is solved at the $\kappa$-th iteration to generate the next feasible point $(\mathbf{w}^{(\kappa+1)},\pmb{\alpha}^{(\kappa+1)})$:
\begin{eqnarray}
\displaystyle\max_{\mathbf{w},\pmb{\alpha}}\ \Phi^{(\kappa)}(\mathbf{w},\pmb{\alpha})\triangleq \sum_{(i,k)\in{\cal M}}{\cal R}_{i,k}^{(\kappa)}(\mathbf{w}_i,\alpha_i)\quad{\mathrm{s.t.}}\nonumber\\
{\cal R}_{i,k}^{(\kappa)}(\mathbf{w}_i,\alpha_i)\geq {\mathsf{\bar R}}_{i,k},\ \forall(i,k) \in\mathcal{M}, (\ref{ad}),
\eqref{eq:changevariables:b}, \eqref{eq:Rate1ktrust}, \eqref{eq:SRMequi:d2}. \label{QPproblem}
\end{eqnarray}
As problem \eqref{QPproblem} involves
$m=2(3K+1)$ quadratic and linear constraints, and $n=2(KN_t + 1)$ real decision variables, its
computational complexity is $\mathcal{O}(n^2m^{2.5}+m^{3.5})$.
Note that $\Phi(\mathbf{w},\pmb{\alpha})\geq \Phi^{(\kappa)}(\mathbf{w},\pmb{\alpha})$ $\forall (\mathbf{w},\pmb{\alpha})$, and
$\Phi(\mathbf{w}^{(\kappa)},\pmb{\alpha}^{(\kappa)})=\Phi^{(\kappa)}(\mathbf{w}^{(\kappa)},\pmb{\alpha}^{(\kappa)})$. Moreover,
$\Phi^{(\kappa)}(\mathbf{w}^{(\kappa+1)},\pmb{\alpha}^{(\kappa+1)})>\Phi^{(\kappa)}(\mathbf{w}^{(\kappa)},\pmb{\alpha}^{(\kappa)})$
whenever $(\mathbf{w}^{(\kappa+1)},\pmb{\alpha}^{(\kappa+1)})\neq (\mathbf{w}^{(\kappa)},\pmb{\alpha}^{(\kappa)})$ because the former and
the latter, respectively, are the optimal solution and feasible point for (\ref{QPproblem}). Therefore,
$\Phi(\mathbf{w}^{(\kappa+1)},\pmb{\alpha}^{(\kappa+1)})\geq \Phi^{(\kappa)}(\mathbf{w}^{(\kappa+1)},\pmb{\alpha}^{(\kappa+1)})>\Phi^{(\kappa)}(\mathbf{w}^{(\kappa)},\pmb{\alpha}^{(\kappa)})=
\Phi(\mathbf{w}^{(\kappa)},\pmb{\alpha}^{(\kappa)})$, showing that $(\mathbf{w}^{(\kappa+1)},\pmb{\alpha}^{(\kappa+1)})$ is a better feasible point
than $(\mathbf{w}^{(\kappa)},\pmb{\alpha}^{(\kappa)})$ for (\ref{eq:SRMequi}). The sequence $\{(\mathbf{w}^{(\kappa)},\pmb{\alpha}^{(\kappa)})\}$
of improved feasible points for (\ref{eq:SRMequi}) thus converges at least to a locally optimal solution satisfying the
Karush-Kuh-Tucker conditions \cite{MW78}. We summarize the proposed QP-based
path-following procedure in Algorithm~\ref{alg_SCALE_FW}.
\begin{algorithm}[t]
\begin{algorithmic}[1]
\protect\caption{QP-based path-following algorithm for ST maximization problem \eqref{eq:SRoptimi}}
\label{alg_SCALE_FW}
\global\long\def\textbf{Initialization:}{\textbf{Initialization:}}
\REQUIRE Iterate \eqref{qosconvex} for an initial feasible point $(\mathbf{w}^{(0)},\pmb{\alpha}^{(0)})$. Set $\kappa:=0$
\REPEAT
\STATE Solve convex quadratic program \eqref{QPproblem} to obtain the optimal solution: $(\mathbf{w}^{(\kappa+1)},\pmb{\alpha}^{(\kappa+1)})$.
\STATE Set $\kappa:=\kappa+1.$
\UNTIL Convergence
\end{algorithmic} \end{algorithm}
\textit{Generation of an initial point.}: Initialized from a feasible point $(\mathbf{w}^{(0)},\pmb{\alpha}^{(0)})$ for constraints
\eqref{eq:changevariables:b} and (\ref{eq:SRMequi:d2}), we iterate the convex program
\begin{equation}\label{qosconvex}
\max_{\mathbf{w}, \pmb{\alpha}} \min_{(i,k) \in\mathcal{M}} {\cal R}_{i,k}^{(\kappa)}(\mathbf{w}_i,\alpha_i)/\bar{\mathsf{R}}_{i,k}\quad
{\mathrm{s.t.}}\quad (\ref{ad}), \eqref{eq:changevariables:b}, \eqref{eq:Rate1ktrust},\eqref{eq:SRMequi:d2}
\end{equation}
till reaching
$\min_{(i,k)\in{\cal M}} {\cal R}_{i,k}^{(\kappa)}(\mathbf{w}_i^{(\kappa+1)},\alpha_i^{(\kappa+1)})/\bar{\mathsf{R}}_{i,k} \geq 1$
to make $(\mathbf{w}^{(\kappa+1)},\pmb{\alpha}^{(\kappa+1)})$ feasible for (\ref{eq:SRMequi}) and thus usable as an initial feasible
point for implementing Algorithm \ref{alg_SCALE_FW}.
\section{Numerical Results}\label{sec:simulation}
Monte Carlo simulations have been implemented to evaluate the performance of the proposed algorithm for $K=4$ ($8$ UEs)
and $N_t=5$ per the scenario in Fig.~\ref{fig:Layout}. The channel vector $\mathbf{h}_{i,k}$ between the BS and
UE $(i,k)$ at a distance $d_{i,k}$ (in kilometres) is generated as $\mathbf{h}_{i,k}=\sqrt{10^{-\sigma_{\mathsf{PL}}/10}}\tilde{\mathbf{h}}_{i,k}$, where $\sigma_{\mathsf{PL}}$ is the path loss (PL) in dB and $\tilde{\mathbf{h}}_{i,k}\sim\mathcal{CN}(0,\mathbf{I}_{N_t})$ represents small-scale effects.
The other parameters are given by Table \ref{parameter}. Without loss of generality, $\bar{\mathsf{R}}_{i,k}\equiv \bar{\mathsf{R}}$ is set. The numerical results are obtained using the parser YALMIP \cite{L04}.
We compare the performance of the proposed FT-based beamforming scheme with five other beamforming schemes:
$(i)$ ``Conventional DL,'' under which the problem of ST maximization is formulated similarly as:
$\max_{\mathbf{w}}\sum_{(i,k)\in {\cal M}} {\cal R}_{i,k}^{'}(\mathbf{w})\ {\mathrm{s.t.}}\ {\cal R}_{i,k}^{'}(\mathbf{w})\geq {\mathsf{\bar R}}_{i,k}, (i,k)\in{\cal M},
\|\mathbf{w}_1\|^2+\|\mathbf{w}_2\|^2\leq P_{bs}^{\max}$ under the definition (\ref{crate});
$(ii)$ ``NOMA'': {\color{black}each UE in zone-1 is paired with an UE in zone-2 according to the clustering algorithm
in \cite{Kietal13} to create a virtual cluster.} In each cluster, both UEs decode the signal intended for the UE in zone-2
and then the UE in zone-1 processes successive interference cancellation (SIC) to cancel the interference of the UE in zone-2 in decoding its own signal;
$(iii)$ ``FT + NOMA in both zones'': under FT, NOMA is zone-wide adopted;
$(iv)$ ``FT + NOMA in zone-1'': under FT, NOMA is adopted only in zone-1; and
$(v)$ ``FT + NOMA in zone-2'': under FT, NOMA is adopted only in zone-2. {\color{black}The reader is referred to \cite[Sec. V]{Dietal17}
for beamforming under NOMA, which is used in these five schemes. {\color{black}The
computational complexity of each iteration in NOMA is similar to that of (\ref{QPproblem}).}
Note that
the performance of NOMA-based beamforming can be improved by involving more UEs in virtual clusters \cite[Sec VI]{Dietal17} but the UEs' privacy is more compromised.}
\begin{table}[t]
\caption{Simulation Parameters}
\label{parameter}
\centering
\begin{tabular}{l|l}
\hline
Parameters & Value \\
\hline\hline
Noise power density & -174 [dBm/Hz] \\
Path loss from the BS to UE $(i,k)$, $\sigma_{\mathsf{PL}}$ & 128.1 + 37.6$\log_{10}(d_{i,k})$ [dB]\\
Radius of cell & 500 [m]\\
Coverage of zone-1 UEs & 200 [m]\\
Distance between the BS and nearest user & $>$ 10 [m]\\
\hline
\end{tabular}
\end{table}
On average, Algorithm~\ref{alg_SCALE_FW} requires about $10$ iterations for convergence.
\begin{figure}
\begin{center}
\begin{subfigure}[$\bar{\mathsf{R}} = 0 $ bps/Hz.]{
\includegraphics[width=0.32\textwidth]{SRvsPbs_R0.eps}}
\label{fig:SRMa}
\end{subfigure}
\begin{subfigure}[$\bar{\mathsf{R}} = 1 $ bps/Hz.]{
\includegraphics[width=0.32\textwidth]{SRvsPbs_R1.eps}}
\label{fig:SRMb}
\end{subfigure}
\caption{Average sum throughput versus $P_{bs}^{\max}$.}\label{fig:SRvsPmax}
\end{center}
\end{figure}
Fig.~\ref{fig:SRvsPmax} plots the average achievable ST versus the transmit power $P_{bs}^{\max}$ for $N_t = 5$. For $\bar{\mathsf{R}} = 0 $ bps/Hz shown in Fig.~\ref{fig:SRvsPmax}(a), one can see that the ST of
the proposed FT-based beamforming is higher than that achieved by the other schemes
in the high transmit power region. On the other
hand, the conventional DL outperforms NOMA and FT+NOMA schemes for high transmit power
by making the ST concentrated at those UEs with good channel conditions.
Apparently, NOMA does not look efficient when there is no UEs' QoS requirement.
In Fig.~\ref{fig:SRvsPmax}(b) with $\bar{\mathsf{R}} = 1 $ bps/Hz, the conventional DL performs worse than
NOMA at high $P_{bs}^{\max}$. The low STs of FT+NOMA-based schemes are probably attributed to the fact that NOMA is more efficient by exploiting their channel differentiation. Increasing $P_{bs}^{\max}$ leads to a remarkable gain in ST
by the proposed FT compared with the other schemes. In addition, the proposed FT is feasible in all range of $P_{bs}^{\max}$ while the other schemes cannot offer such high QoS to UEs at low $P_{bs}^{\max}$.
\begin{figure}
\centering
\includegraphics[width=0.32\textwidth]{SRvsRth.eps}
\caption{Average sum throughput versus $\bar{\mathsf{R}}$ for $P_{bs}^{\max} = 30$ dBm.}\label{fig:SRvsRth}
\end{figure}
The plot of the ST versus QoS requirement threshold $\bar{\mathsf{R}}\in[0.2,\;1.2]$ bps/Hz is illustrated in Fig.~\ref{fig:SRvsRth}. We can observe that the proposed FT-based beamforming performs quite well and only slightly degrades when $\bar{\mathsf{R}}$ increases. The performance gap between the proposed FT-based beamforming and other schemes substantially increases by increasing $\bar{\mathsf{R}}$. It is expected because the proposed
FT-based beamforming can tune the power allocation in meeting zone-2 UEs' QoS requirements without causing interference to the zone-1's UEs.
Next, we look for the max-min UE throughput optimization problem
\begin{equation}\label{eq:maxminoptimi}
\max_{\mathbf{w}, \boldsymbol{\tau}}\min_{(i,k) \in\mathcal{M}}{\cal R}_{i,k}(\mathbf{w}_i,\tau_i)\quad
{\mathrm{s.t.}}\quad \eqref{eq:SRoptimi:d}, \eqref{eq:SRoptimi:e}
\end{equation}
which can be addressed similarly by solving the convex program (\ref{qosconvex})
(with $\bar{\mathsf{R}}_{i,k}\equiv 1$) at each iteration.
\begin{figure}
\centering
\includegraphics[width=0.32\textwidth]{SRvsPbs_maxmin.eps}
\caption{Average max-min throughput versus $P_{bs}^{\max}$.}
\label{fig:MaxminPbs}
\end{figure}
Fig.~\ref{fig:MaxminPbs} plots the average UEs' worst throughput as a function of $P_{bs}^{\max}$.
It shows that the worst throughput achieved by
the conventional DL and FT + NOMA-based schemes is saturated once $P_{bs}^{\max}$ is beyond a threshold. It also
reveals that the proposed FT-based beamforming
and FT + NOMA in zone-1 achieve worst throughput that is higher than that achieved by others schemes.
\section{Conclusions}\label{sec:conclusion}
The paper has proposed a fractional time-based beamforming scheme at a base station to serve two groups of users, which
is able to improve the network throughput while preserving the information privacy for the users. Accordingly,
a path-following computational procedure for the joint design of fractional times and
beamforming vectors to maximize the network throughput has been developed. Extensive simulations have been
provided to demonstrate the superior performance of the proposed scheme over the exiting schemes.
\section*{Appendix} \label{Appendix:A}
For all $ x > 0$, $\bar{x} > 0$, $y>0$, $\bar{y}>0$, $t>0$, and $\bar{t}>0$, it is true that
\begin{eqnarray}
\displaystyle\frac{\ln(1+x^2/y)}{t}&\geq&\displaystyle a - b\frac{y}{x^2} - ct\label{ing1a}\\
&\geq&\displaystyle a - b\frac{y}{\bar{x}(2x-\bar{x})} - ct\label{inq1}
\end{eqnarray}
over the trust region
\begin{equation}\label{trust2}
2x-\bar{x}>0,
\end{equation} where
$a = 2 [\ln(1+d)]/\bar{t} +
d/\bar{t}(d+1) > 0, b=d^2/\bar{t}(d+1) > 0,\\
c = [\ln(1+d)]/\bar{t}^2> 0, d=\bar{x}^2/\bar{y}.$
Inequality (\ref{ing1a}) follows
from \cite{Ngetal17} while inequality (\ref{inq1}) is obtained by using $x^2\geq \bar{x}(2x-\bar{x})>0$ over the
trust region (\ref{trust2}).
\bibliographystyle{IEEEtran}
\balance
| 9c1af8182c4f922626e1ca2ed15f770d128f8b44 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{section:introduction}
The dominant approach to inference in regression analysis in the social sciences takes a sampling perspective on uncertainty. This perspective relies on the assumption that the observed units can be viewed as a sample drawn randomly from a large population of interest. In many cases this random sampling perspective is a natural and attractive one. For example, if one analyzes individual-level data from the U.S. Current Population Survey, the Panel Study of Income Dynamics, or the 1\% public use sample from the U.S. Census, it is natural to regard the sample as a small random subset of the population of interest. In many other settings, however, this sampling perspective is less attractive. For example, suppose that the data set to be analyzed contains information on all 50 states of the United States, all the countries in the world, or all visits to a website. If, for all units in this data set, we observe an outcome and some attributes at a single point in time, and we ask how the average outcome varies across two subpopulations defined by these attributes, the answer is a quantity that is known with certainty. Hence, the standard error should be zero. However, researchers analyzing this type of data typically report standard errors that are formally justified by the random sampling perspective. This widespread practice implicitly forces the object of interest to be a data generating process, or superpopulation, from which the actual population is drawn at random. In such a setting, uncertainty arises from lack of observability of the superpopulation. While this may be an appealing framework in some instances, it is clearly not so in cases where the interest resides in an actual finite population and, in any event, a researcher may want to first define the object of interest and then use an appropriate mode of inference, rather than allowing the mode of inference to implicitly define the object of interest of her/his investigation.
In this article, we provide an alternative framework for the interpretation of uncertainty in regression analysis regardless of whether a fraction of the population or the entire population is included in the sample. While our framework accommodates sampling-based uncertainty, it also takes into account design-based uncertainty, which arises when the parameter of interest is defined in terms of the unobserved outcomes that some units would attain under a certain intervention. Design-based uncertainty is often explicitly accounted for in the analysis of randomized experiments where it is the basis of randomization inference (\citeauthor{neyman1923}, \citeyear{neyman1923}; \citeauthor{rosenbaum_book}, \citeyear{rosenbaum_book}; \citeauthor{imbens2015causal}, \citeyear{imbens2015causal}), but it is rarely explicitly acknowledged in regression analyses or, more generally, in observational studies (exceptions in special cases include \citeauthor{samii2012equivalencies}, \citeyear{samii2012equivalencies}; \citeauthor{Freedman2008}, \citeyear{Freedman2008} and \citeyear{ freedman2008regression}; \citeauthor{lin}, \citeyear{lin}).
To illustrate the differences between sampling-based inference and design-based inference, we present two examples in Tables \ref{tabel_sampling} and \ref{tabel_assignment}. In the example of Table \ref{tabel_sampling}, there is a finite population consiting of $n$ units, each characterized by the values of a pair of variables, $Y_i$ and $Z_i$. Here, we can define an estimand as a function of the pairs $\{(Y_i,Z_i)\}_{i=1}^n$ for the entire population. For example, the estimand could be the difference in the population average value of the outcome, $Y_i$, by values of the attribute, $Z_i$. Uncertainty about the estimand exists when we observe the values $(Y_i,Z_i)$ only for a subset of the population, the sample. In Table \ref{tabel_sampling} inclusion of unit $i$ in a sample is coded by the binary variable $R_i$. In this setting, an estimator can be naturally defined as the difference in the average value of the outcome, $Y_i$, by values of the attribute, $Z_i$, in the sample. Sampling-based inference uses information about the process that determines $R_1,\ldots, R_n$, to assess the variability of estimators across different potential samples.
\vskip1cm
\begin{table}[ht]
\caption{\textsc{: Sampling-based Uncertainty ($\tick$ is observed, $\tock$ is missing)}}
\label{tabel_sampling}\vskip1cm
\begin{center}
\begin{tabular}{c|cccccccccccccc}
& \multicolumn{3}{c}{Actual} &&
\multicolumn{3}{c}{Alternative}&&
\multicolumn{3}{c}{Alternative}&& $\hdots$\\
Unit & \multicolumn{3}{c}{Sample} &&
\multicolumn{3}{c}{Sample I}&&
\multicolumn{3}{c}{Sample II}&& $\hdots$\\
& $Y_i$ & $Z_i$&$R_i$&\hskip0.5cm \ \ \ & $Y_i$ & $Z_i$ & $R_i$
&\hskip0.5cm \ \ \ & $Y_i$ & $Z_i$ & $R_i$&\hskip0.5cm& $\hdots$
\\
\hline \\
$1$ & $\tick$ & $\tick$&1& & $\tock$ & $\tock$&0& & $\tock$ & $\tock$&0&& $\hdots$\\
$2$ & $\tock$ & $\tock$&0& & $\tock$ & $\tock$&0& & $\tock$ & $\tock$&0&& $\hdots$\\
$3$ & $\tock$ & $\tock$&0& & $\tick$ & $\tick$&1& & $\tick$ & $\tick$&1&& $\hdots$\\
$4$ & $\tock$ & $\tock$&0&& $\tick$ & $\tick$&1& & $\tock$ & $\tock$&0&& $\hdots$\\
\vdots & \vdots &\vdots& \vdots& & \vdots & \vdots&\vdots& & \vdots & \vdots&\vdots&& $\hdots$\\
$n$ & $\tick$ & $\tick$&1&& $\tock$ & $\tock$&0& & $\tock$ & $\tock$&0&& $\hdots$
\end{tabular}
\end{center}
\end{table}
Table \ref{tabel_assignment} depicts a different scenario. We encounter again a finite population of size $n$. For each population unit we now observe the value of one of two variables, either $Y_i(1)$ or $Y_i(0)$, but not both. $Y_i(1)$ and $Y_i(0)$ represent the potential outcomes that unit $i$ would attain under exposure or lack of exposure to certain intervention (or treatment) of interest. In Table \ref{tabel_assignment} exposure to the intervention is coded by the binary treatment variable, $X_i$. We observe $Y_i(1)$ if $X_i=1$, and $Y_i(0)$ if $X_i=0$. The estimand is a function of the full set of pairs $\{(Y_i(1),Y_i(0))\}^n_{i=1}$, for example, the average causal effect $(1/n)\sum_{i=1}^n (Y_i(1)-Y_i(0))$. As in the first example, the estimator is a function of the observed data, e.g., the difference in the average of observed values of $Y_i(1)$ and $Y_i(0)$. Design-based inference uses information about the process that determines $X_1,\ldots, X_n$, to assess the variability of estimators across different potential samples. Notice that, under this mode of inference, uncertainty about the estimand remains even when we observe the entire population, as in Table \ref{tabel_assignment}.
\vskip1cm
\begin{table}[ht]
\caption{\textsc{: Design-based Uncertainty ($\tick$ is observed, $\tock$ is missing)}}
\label{tabel_assignment}\vskip1cm
\begin{center}
\begin{tabular}{c|cccccccccccccc}
& \multicolumn{3}{c}{Actual} &&
\multicolumn{3}{c}{Alternative}&&
\multicolumn{3}{c}{Alternative}&& $\hdots$\\
Unit & \multicolumn{3}{c}{Sample} &&
\multicolumn{3}{c}{Sample I}&&
\multicolumn{3}{c}{Sample II}&& $\hdots$\\
& $Y_i(1)$ & $Y_i(0)$& $X_i$ &\hskip0.5cm \ \ \ & $Y_i(1)$ & $Y_i(0)$& $X_i$ &\hskip0.5cm \ \ \ & $Y_i(1)$ & $Y_i(0)$& $X_i$ &\hskip0.5cm& $\hdots$\\
\hline \\
$1$ & $\tick$ & $\tock$&1& & $\tick$ & $\tock$&1& & \tock & \tick&0&& $\hdots$\\
$2$ & $\tock$ & $\tick$&0& & $\tock$ & $\tick$&0& & \tock & \tick&0&& $\hdots$\\
$3$ & $\tock$ & $\tick$&0& & $\tick$ & $\tock$&1& & \tick& \tock&1&& $\hdots$\\
$4$ & $\tock$ & $\tick$&0& & $\tock$ & $\tick$&0& & \tick & \tock&1&& $\hdots$\\
\vdots & \vdots & \vdots&\vdots && \vdots&\vdots & \vdots && \vdots&\vdots&\vdots&& $\hdots$\\
$n$ & $\tick$ & $\tock$&1& & $\tock$ & $\tick$&0& & \tock & \tick&0&& $\hdots$
\end{tabular}
\end{center}
\end{table}
More generally, of course, we can have complex missing data processes that combine features of these two examples, with some units not included in the sample at all, and with one of the two potential outcomes not observed for the sample units. The inferential procedures proposed in this article address both sources of variability. As the examples in Tables \ref{tabel_sampling} and \ref{tabel_assignment} illustrate, articulating the exact nature of the estimand of interest and the source of uncertainty that makes the estimator stochastic are crucial steps to valid inference. For this purpose, it will be useful to distinguish between descriptive estimands, where uncertainty stems solely from not observing all units in the population of interest, and causal estimands, where the uncertainty stems, at least partially, from unobservability of potential outcomes.
The main formal contribution of this article is to generalize the results for the approximate variance for multiple linear regression estimators associated with the work by \citet{eicker}, \citet{huber}, and \citet{white1980robust,white1980using,white1982maximum}, \ehww from hereon, in two directions. First, we allow sampling from a finite population and, second, we allow for design-based uncertainty in addition to, or instead of, the sampling-based uncertainty that the \ehww results based on. The first generalization decreases the variance, and the second increases the variance. Incorporating these generalizations requires developing a new framework for regression analysis with assumptions that differ from the standard ones. This framework nests as special cases the \citet{neyman1923}, \citet{samii2012equivalencies}, \citet{Freedman2008}, \citet{freedman2008regression}, and \citet{lin} regression analyses for data from randomized experiments. We show that in large samples the widely used \ehww robust standard errors are conservative. Moreover, we show that the presence of attributes -- that is, immutable characteristics of the units -- can be exploited to improve on the \ehww variance estimator, and we propose variance estimators that do so. Finally, we show that in some special cases, in particular the case where the regression function is correctly specified, the \ehww standard errors are asymptotically correct.
One important practical advantage of our framework is that it justifies non-zero standard errors in cases where we observe all units in the population but design-based uncertainty remains. A second advantage of the formal separation into sampling-based and design-based uncertainty is that it allows us to discuss the distinction between internal and external validity (\citeauthor{shadishcookcampbell}, \citeyear{shadishcookcampbell}; \citeauthor{manski2013public}, \citeyear{manski2013public}; \citeauthor{deaton2010}, \citeyear{deaton2010}) in terms of these two sources of uncertainty. For internal validity there are no assumptions required on the sampling process, and conversely, for external validity there are no assumptions required on the design.
\section{A Simple Example}
\label{section:cases}
In this section we set the stage for the problems discussed in the current article by discussing least squares estimation in a simple example with a single binary regressor. We make four points. First, we show how design-based uncertainty affects the variance of regression estimators. Second, we show that the standard Eicker-Huber-White (EHW) variance estimator remains conservative when we take into account design-based uncertainty. Third, we show that there is a simple finite-population correction to the \ehww variance estimator for descriptive estimands but not for causal estimands. Fourth, we discuss the relation between the two sources of uncertainty and the notions of internal and external validity of the estimand.
We focus on a setting with a finite population of size $n$. We sample $N$ units from this population, with $R_i\in\{0,1\}$ indicating whether a unit was sampled ($R_i=1$) or not $(R_i=0$), so that $N=\sum_{i=1}^n R_i$. There is a single binary regressor, $X_i\in\{0,1\}$, and $n_x$ (resp.\ $N_x$) are the number of units in the population (resp.\ the sample) with $X_i=x$. To make the discussion specific, suppose the binary regressor $X_i$ is an indicator for a state regulation, say the state having a minimum wage higher than the federal minimum wage. We view the regressor not as a fixed attribute or characteristic of each unit, but instead as a cause or policy variable whose value could have been different from the observed value. This generates missing data of the type shown in Table \ref{tabel_assignment}, where only some of the states of the world are observed, implying that there is design-based uncertainty. Formally, using the Rubin causal model or potential outcome framework (\citeauthor{neyman1923}, \citeyear{neyman1923}; \citeauthor{rubin1974estimating}, \citeyear{rubin1974estimating}; \citeauthor{holland1986statistics}, \citeyear{holland1986statistics}; \citeauthor{imbens2015causal}, \citeyear{imbens2015causal}), we postulate the existence of two potential outcomes for each unit, denoted by $Y_i(1)$ and $Y_i(0)$, for state average earnings without and with a state minimum wage, with $Y_i$, the realized outcome, given the actual or prevailing minimum wage, defined as:
\begin{equation*}
Y_i=Y_i(X_i)= \left\{ \begin{array}{ll} Y_i(1) & \mathrm{if}\ X_i=1,\\ Y_i(0) & \mathrm{if}\ X_i=0. \end{array} \right.
\end{equation*}
These potential outcomes are viewed as non-stochastic attributes for unit $i$, irrespective of the realized value of $X_i$. They, as well as the additional observed attributes, $Z_i$, remain fixed in repeated sampling thought experiments, whereas $R_i$ and $X_i$ are stochastic and, as a result, so are the realize outcomes, $Y_i$. In the current section we abstract from the presence of fixed observed attributes, $Z_i$, which will play an important role in Section \ref{main}. Let $\by$, $\by(1)$, $\by(0)$, $\br$, and $\bx$ be the $n$-vectors with $i$-th element equal to $Y_i$, $Y_i(1)$, $Y_i(0)$, $R_i$, and $X_i$ respectively. For sampled units (units with $R_i=1$) we observe $X_i$, and $Y_i$.
In general, estimands are functions of the full set of population values $(\by(1),\by(0),\bx,\br)$. We consider two types of estimands, descriptive and causal. If an estimand can be written as a function of $(\by,\bx)$, free of dependence on $\br$ and on the potential outcomes beyond the realized outcome, we label it a {\it descriptive} estimand. Intuitively a descriptive estimand is an estimand whose value would be known with certainty if we observe all the realized values of all variables for all units in the population. If an estimand cannot be written as a function of $(\by,\bx,\br)$ because it depends on the potential outcomes $\by(1)$ and $\by(0)$, then we label it a {\it causal} estimand.
We now consider in our binary regressor example three closely related estimands, one descriptive and two causal. The first estimand is the difference in population averages by the prevailing minimum wage,
\begin{equation*}
\theta^{\descr} =\theta^\descr(\by,\bx) =\frac{1}{n_{1}}\sum_{i=1}^nX_iY_i- \frac{1}{n_0}\sum_{i=1}^n (1-X_i)Y_i.
\end{equation*}
This estimand is a function of $(\by,\bx)$ and so it is a descriptive estimand. The second estimand is the sample average causal effect,
\begin{equation*}
\theta^{\causal,\sample} =\theta^{\causal,\sample} (\by(1),\by(0),\br) = \frac{1}{N}\sum_{i=1}^n R_i\, \Bigl( Y_i(1)-Y_i(0)\Bigr).
\end{equation*}
This estimand is a causal estimand: it cannot be written as a function of $(\by,\bx,\br)$ because it depends on the potential outcomes $\by(1)$ and $\by(0)$. The third estimand is the population version of $\theta^{\causal,\sample}$:
\begin{equation*}
\theta^\causal=\theta^\causal(\by(1),\by(0)) = \frac{1}{n}\sum_{i=1}^n \Bigl( Y_i(1)-Y_i(0)\Bigr).
\end{equation*}
This is again a causal estimand.
Now let us turn to estimators. In general an estimator is a function of the values of $Y_i$, $X_i$ and $Z_i$ for the units in the sample, that is for the units with $R_i=1$. We focus on the properties of a particular estimator:
\begin{equation*}
\widehat\theta
=\frac{1}{N_1}\sum_{i=1}^n R_iX_iY_i-
\frac{1}{N_0}\sum_{i=1}^n R_i(1-X_i)Y_i,
\end{equation*}
which can be interpreted as a least squares estimator of the coefficient on $X_i$ for the regression of $Y_i$ on $X_i$ and a constant. There are two sources of randomness in this estimator: a sampling component arising from the randomness of $\bs R$ and a design component arising from the randomness of $\bs X$. We refer to the uncertainty generated by the randomness in the sampling component as {\it sampling-based uncertainty}, and the uncertainty generated by the design component as {\it design-based uncertainty}.
We will next consider the the first two moments of $\widehat\theta$ under combinations of two assumptions. The first assumption is on the sampling mechanism.
\begin{assumption}
\label{assumption:randomsampling1}
{\sc (Random Sampling / External Validity)}
\begin{equation*}
\Pr\left(\bs R=\bs r\right) =
1\biggl/\left(\begin{array}{c}n\\N \end{array}\right),
\end{equation*}
for all $n$-vectors ${\bbr}$ with $\sum_{i=1}^n r_{i}=N$.
\end{assumption}
The second assumption is on the assignment mechanism.
\begin{assumption}
\label{assumption:randomass1}
{\sc (Random Assignment / Internal Validity)}
\begin{equation*}
\Pr\left({\bx}={\bbx}|\br\right) =
1\biggl/\left(\begin{array}{c}n\\n_1 \end{array}\right),
\end{equation*}
for all $n$-vectors ${\bbx}$ with $\sum_{i=1}^n X_i=n_1$.
\end{assumption}
We start by studying the first moment of the estimator, conditional on $(N_1,N_0)$, and only for the cases where
$N_1\geq 1$ and $N_0\geq 1$ (and thus $n_1\geq 1$ and $n_0\geq 1$). We leave this latter conditioning implicit in the notation throughout this section.
A supplementary appendix contains proofs of the results in this section. First, taking the expectation only over the random sampling, under Assumption \ref{assumption:randomsampling1}:
\begin{equation}
\label{equation:condexpx}
E\big[\widehat{\theta}\,|\,\bx,N_1,N_0\big]=\theta^{\descr}.
\end{equation}
Notice that this result does not require random assignment. Second, taking the expectation only over the random assignment, under Assumption \ref{assumption:randomass1}:
\begin{equation}
\label{equation:condexpr}
E\big[\widehat{\theta}\,|\,\br,N_1,N_0\big]=\theta^{\causal,\sample}.
\end{equation}
This equality does not require random sampling. Third, taking the expectation over both the sampling and the assignment, maintaining both Assumptions \ref{assumption:randomsampling1} and \ref{assumption:randomass1}:
\begin{equation*}
E\big[\widehat{\theta}\,|\,N_1,N_0\big]
=E\big[\theta^{\descr}\,|\,N_1,N_0\big]
=E\big[\theta^{\causal,\sample}\,|\,N_1,N_0\big]
=\theta^{\mathrm{causal}}.
\end{equation*}
Next we look at the variance of the estimator. Here we maintain both the random assignment and random sampling assumption. From Equations (\ref{equation:condexpx}) and (\ref{equation:condexpr}), it follows that $\mbox{var}(\widehat{\theta}\,|\, \bx,N_1,N_0)$ measures dispersion with respect to $\theta^\descr$, while $\mbox{var}(\widehat{\theta}\,|\, \br,N_1,N_0)$ measures dispersion with respect to $\theta^{\causal,\sample}$. By the law of total variance, we can decompose:
\begin{align}
\label{equation:vardecomp}
\mbox{var}\big(\widehat{\theta}\,|\, N_1,N_0\big)&=
E\left[\mbox{var}\big(\widehat{\theta}\,|\, \bx,N_1,N_0\big)\,\big|\, N_1,N_0\right]
+\mbox{var}\big(\theta^{\descr}\,|\, N_1, N_0\big)\nonumber\\
&=E\left[\mbox{var}\big(\widehat{\theta}\,|\, \br,N_1,N_0\big)\,|\, N_1,N_0\right]
+\mbox{var}\big(\theta^{\causal,\sample}\,|\, N_1,N_0\big).
\end{align}
Let
\[
S^2_1=\frac{1}{n-1}\sum_{i=1}^n \left(Y_i(1)-\frac{1}{n}\sum_{j=1}^n Y_j(1)\right)^2.
\]
$S^2_0$ and $S^2_{\theta}$ are analogously defined for $Y_1(0), \ldots, Y_n(0)$ and $\theta_1,\ldots, \theta_n$, respectively, where $\theta_i = Y_i(1)-Y_i(0)$.
The variance of $\widehat\theta$ can be expressed as
\begin{align}
\label{equation:totalvariance}
V^\total(N_1,N_0,n_1,n_0)&=\mbox{var}\big(\widehat{\theta}\,|\, N_1,N_0\big),\nonumber\\
&=\frac{S^2_1}{N_1}+\frac{S^2_0}{N_0}-\frac{S^2_{\theta}}{n},
\end{align}
which is a variant of the result in \cite{neyman1923}.
For the first decomposition in equation (\ref{equation:vardecomp}), the sampling-based component of the total variance is
\begin{align*}
V^\sampling(N_1,N_0,n_1,n_0)&=E\left[\mbox{var}\big(\widehat{\theta}\,|\, \bx,N_1,N_0\big)\,\big|\, N_1,N_0\right],\\
&=\frac{S^2_1}{N_1} \left(
1-\frac{N_1}{n_1}\right) +\frac{S^2_0}{N_0}\left(1-\frac{N_0}{n_0}\right),
\end{align*}
and the design-based component, beyond the sampling-based component, is
\begin{align*}
V^{\design|\sampling}(N_1,N_0,n_1,n_0)&=\mbox{var}\big(\theta^{\descr}\,|\, N_1, N_0\big)\\
&=\frac{S^2_1}{n_1}+\frac{S^2_0}{n_0}-\frac{S^2_{\theta}}{n}.
\end{align*}
For the second decomposition in equation (\ref{equation:vardecomp}), the design-based component of the variance is
\begin{align*}
V^{\design}(N_1,N_0,n_1,n_0)&=E\left[\mbox{var}\big(\widehat{\theta}\,|\, \br,N_1,N_0\big)\,|\, N_1,N_0\right]\\
&=\frac{S^2_1}{N_1}+\frac{S^2_0}{N_0}-\frac{S^2_{\theta}}{N},
\end{align*}
and the sample-based component, beyond the design-based component, is
\begin{align*}
V^{\sampling|\design}(N_1,N_0,n_1,n_0)&=\mbox{var}\big(\theta^{\causal,\sample}\,|\, N_1, N_0\big)\\
&=\frac{S^2_{\theta}}{N}\left(1-\frac{N}{n}\right).
\end{align*}
\begin{comment}
{\sc Causal versus Descriptive Estimands}\\
A key comparison is between the sampling variance for the estimator for the descriptive estimand and the design variance for the estimator for the sample average causal effect
\[
V^\sampling(N_1,N_0,n_1,n_0)=\frac{S^2_1}{N_1}\left(
1-\frac{N_1}{n_1}\right) +\frac{S^2_0}{N_0}\left(1-\frac{N_0}{n_{0}}\right),
\]
versus
\[V^{\design}(N_1,N_0,n_1,n_0)=\frac{S^2_1}{N_{1}}+
\frac{ S^2_0}{N_0}-\frac{S^2_{\theta}}{N}.\]
In general these variances cannot be ranked: the sampling variance can be very close to zero if the sampling rate $N/n$ is close to one, but it can also be larger than the design variance if the sampling rate is small and the variance of the treatment effect, $S^2_{\theta}$, is substantial.\hfill$\square$
\end{comment}
\begin{comment}{\sc Finite Population Correction}\\
If the estimand is $\theta^\causal$ or $\theta^\descr$,
ignoring the fact that the population is finite generally leads to an overstatement of the variance:
\[
V^\total(N_1,N_0,\infty,\infty)-V^\total(N_1,N_0,n_1,n_0)=\frac{S^2_{\theta}}{n}\geq 0,
\]
\[
V^\sampling(N_1,N_0,\infty,\infty)-V^\sampling(N_1,N_0,n_1,n_0)=
\frac{S^2_1}{n_1} +\frac{S^2_0}{n_0}
\geq 0.
\]
If the estimand is $\theta^{\causal,\sample}$, however, the population size is irrelevant:
\[
V^\design(N_1,N_0,\infty,\infty)=
V^\design(N_1,N_0,n_1,n_0).\hfill\square
\]
\end{comment}
\begin{comment}{\sc Large Population {\it vs} Sample is identical to Population}\\
If the population is large relative to the sample,
the incremental design-based component is zero, and the sampling-based variance component is equal to the total variance:
\[
V^{\design|\sampling}(N_1,N_0,\infty,\infty)=0,\quad
V^\sampling(N_1,N_0,\infty,\infty)=
V^\total(N_1,N_0,\infty,\infty).
\]
In other words, if the population is large relative to the sample, it is sufficient to consider the sampling-based variance.
This can be viewed as the implicit justification for the common practice of ignoring design-based uncertainty.
If, at the other extreme, the sample is equal to the population, the sampling-based variance component is zero and the design-based component is equal to the total variance:
\[
V^\sampling(N_1,N_0,N_1,N_0)=0,\quad
V^{\design|\sampling}(N_1,N_0,N_1,N_0)=
V^\total(N_1,N_0,N_1,N_0).
\]
\hfill$\square$
\end{comment}
\begin{comment}{\sc Internal versus External Validity}\\
Often researchers are concerned about both the internal and external validity of estimands and estimators
(\citeauthor{shadishcookcampbell}, \citeyear{shadishcookcampbell}; \citeauthor{manski2013public}, \citeyear{manski2013public}; \citeauthor{deaton2010}, \citeyear{deaton2010}). The distinction between sampling and design-based uncertainty allows us to clarify these concerns. Internal validity bears on the question of whether $\widehat\theta$ is a good estimator for $\theta^{\causal,\sample}$. This relies on random assignment of the treatment. Whether or not the sampling is random is irrelevant for this question because $\theta^{\causal,\sampling}$ conditions on which units were sampled. External validity bears on the question whether $\widehat\theta$ is a good estimator for $\theta^\descr$. This relies on the random sampling assumption and is not affected by the assumptions on the assignment process. However, for $\widehat\theta$ to be a good estimator for $\theta^\causal$, which is often the most interesting estimand, we need both internal and external validity, and thus both random assignment and random sampling.
\hfill$\square$
\end{comment}
For the binary regressor example the \ehww variance estimator can be written as
\[
\widehat V^\ehw=\frac{N_1-1}{N^2_1} \widehat S_1^2 + \frac{N_0-1}{N_0^2} \widehat S_0^2,\]
where $\widehat S_1^2$ is the sample counterpart of $S^2_1$,
\[
\widehat S_1^2=\frac{1}{N_1-1}\sum_{i=1}^n R_iX_i\left(Y_i-\frac{1}{N_1}\sum_{i=1}^n R_iX_iY_i\right)^2,
\]
and $\widehat S_0^2$ is defined analogously.
If we adjust the degrees of freedom, using the modification proposed in \citet{mackinnon1985some}, specialized to this binary regressor example, we get the modified \ehww variance estimator,
\[
\widetilde V^\ehw=\frac{\widehat S_1^2}{N_1}+\frac{\widehat S_0^2}{N_0},
\]
with expectation equal to the sampling variance in the infinite population case,
\[
V^\ehw=E\left[\widetilde V^\ehw\right]=\frac{S_1^2}{N_1}+\frac{S_0^2}{N_0}=V^\sampling(N_1,N_0,\infty,\infty).
\]
This variance is also the one proposed by \citet{neyman1923}.
Bootstrapping the estimator would approximately give the same variance.
\begin{comment}{\sc The \ehww Variance Estimator}\\
Compare the expected value of the modified \ehww variance estimator to $V^{\total}(N_1,N_0,n_1,n_0) $. There are two differences. The \ehww variance $V^\ehw=V^\sampling(N_1,N_0,\infty,\infty)$ ignores the fact that the population may be finite, and it ignores the design component of the variance. The combination of these two differences renders the \ehww variance estimator conservative:
\[
V^\ehw=V^\sampling(N_1,N_0,\infty,\infty)
\geq V^{\total}(N_1,N_0,n_1,n_0) .\hfill\square
\]
\end{comment}
\begin{comment}{\sc Can We Improve on the EHW variance Estimator?}\\
The difference between $V^{\ehw}$ and the total variance is
\[
V^{\ehw}-V^\total(N_1,N_0,n_1,n_0)=\frac{S^2_\theta}{n}.
\]
There is no good estimator for $S^2_\theta$ because we do not observe $Y_i(1)$ and $Y_i(0)$ together. Although we may be able to come up with a lower bound for $S^2_\theta$ that is strictly positive, there is no unbiased estimator and, typically, this term is ignored in analyses of randomized experiments (see \citet{imbens2015causal}). In Section \ref{main} we propose a new variance estimator that exploits the presence of fixed attributes.\hfill$\square$
\end{comment}
\section{The General Case}
\label{main}
This section contains the main formal results in the article. The setting we consider here allows for the presence of covariates of the causal type ({\it e.g.}, state institution or a regulation such as the state minimum wage), which can be discrete or continuous, as well as for the presence of covariates of the fixed attribute or characteristic type ({\it e.g.}, an indicator whether a state is landlocked or coastal), which again can be discrete or continuous. We allow potential causes and attributes to be systematically correlated, and we allow for general misspecification of the regression function. The conceptual difference between the causal variables and the attributes is that the value of the causal variables may depend on the design while the value of the attributes does not. This requires us to postulate potential outcomes corresponding to causes but not for the attributes.
\subsection{Set Up}
Consider a sequence of finite populations indexed by population size, $n$. Unit $i$ in population $n$ is characterized by a set of fixed attributes $Z_{n,i}$ (including an intercept) and by a potential outcome function, $Y_{n,i}()$, which maps causes, $U_{n,i}$, into outcomes, $Y_{n,i}=Y_{n,i}(U_{n,i})$. $Z_{n,i}$ and $U_{n,i}$ are real-valued column vectors, while $Y_{n,i}$ is scalar. We do not place restrictions on the types of the variables: they can be continuous, discrete, or mixed.
There is a sequence of samples associated with the population sequence. We will use $R_{n,i}=1$ to indicate that unit $i$ of population $n$ is sampled, and $R_{n,i}=0$ to indicate that it is not sampled. For each unit in sample $n$, we observe the triple, $(Y_{n,i},U_{n,i},Z_{n,i})$.
A key feature of the analysis in this section relative to Section \ref{section:cases} is that we now allow for more complicated assignment mechanisms. In particular, we relax the assumption that the causes have identical distributions.
\begin{assumption}
\label{assumption:assignment}\textsc{(Assignment Mechanism)} The
assignments $U_{n,1},\ldots ,U_{n,n}$ are jointly independent, and independent of $R_{n,1},\ldots ,R_{n,n}$, but not (necessarily)
identically distributed (i.n.i.d.).
\end{assumption}
For what follows, it is convenient to work with a transformation $X_{n,1},\ldots ,X_{n,n}$ of $U_{n,1},\ldots ,U_{n,n}$ such that
\begin{equation}
\label{equation:uncorrelated}
E\left[\sum_{i=1}^n X_{n,i}Z_{n,i}'\right]=\sum_{i=1}^n E[X_{n,i}]Z_{n,i}'=0.
\end{equation}
This can be accomplished in the following way. We assume that the population matrix $\sum_{i=1}^n Z_{n,i}Z_{n,i}'$ is full-rank.
Then, equation (\ref{equation:uncorrelated}) holds for
\begin{equation}
\label{equation:transformation}
X_{n,i}=U_{n,i}-\Lambda_n Z_{n,i}
\end{equation}
where
\begin{equation*}
\Lambda_n = \left(\sum_{i=1}^n E[U_{n,i}]Z_{n,i}'\right)\left(\sum_{i=1}^n Z_{n,i}Z_{n,i}'\right)^{-1}.
\end{equation*}
It is important to notice that, because $\Lambda_n Z_{n,i}$ is deterministic in our setting and $U_{n,1}, \ldots, U_{n,n}$ are i.n.i.d., the variables $X_{n,1}, \ldots, X_{n,n}$ are i.n.i.d. too.
For population $n$, let $\bs Y_n$, $\bs X_n$, $\bs Z_n$, $\bs R_n$, and $\bs Y_n()$ be matrices that collect outcomes, causes, attributes, sampling indicators, and potential outcome functions, where each population unit has the same row index in each of the matrices. In our setting, the sampling indicators $\bs R_n$ and the causes $\bs X_n$ are stochastic. The attributes $\bs Z_n$ and the potential outcome functions $\bs Y_n()$ are taken as fixed. Expectations are taken over the distribution of $(\bs R_n, \bs X_n)$.
We analyze the properties of the estimator $\widehat\theta_n$
obtained by minimizing least square errors in the sample:
\[
(\widehat\theta_n,\widehat\gamma_n)= \operatorname*{argmin}_{(\theta,\gamma)}\,
\sum_{i=1}^n R_{n,i}\big(Y_{n,i}-X_{n,i}'\theta-Z_{n,i}'\gamma\big)^2.
\]
The properties of the population regression residuals, $e_{n,i}=Y_{n,i}-X_{n,i}'\theta_n-Z_{n,i}'\gamma_n$, depend on the exact nature of the estimands, $(\theta_n,\gamma_n)$. In what follows, we will consider alternative target parameters, which in turn will imply different properties for $e_{n,i}$. Notice also that, although the transformation in (\ref{equation:transformation}) is typically unfeasible (because the values of $E[U_{n,i}]$ may not be known), $\widehat\theta_n$ is not affected by the transformation in the sense that the least squares estimators $(\widetilde\theta_n,\widetilde\gamma_n)$, defined as
\[
(\widetilde\theta_n,\widetilde\gamma_n)= \operatorname*{argmin}_{(\theta,\gamma)}\,
\sum_{i=1}^n R_{n,i}\big(Y_{n,i}-U_{n,i}'\theta-Z_{n,i}'\gamma\big)^2,
\]
satisfy $\widehat\theta_n=\widetilde\theta_n$ (although, in general, $\widehat\gamma_n\neq \widetilde\gamma_n$). As a result, we can analyze the properties of $\widehat\theta_n$ focusing on the properties of the regression on $X_{n,1},\ldots, X_{n,n}$ instead of on $U_{n,1},\ldots, U_{n,n}$.
We assume random sampling with some conditions on the sampling rate to ensure that the sample size increases with the population size.
\begin{assumption}
\label{assumption:randomsampling} \textsc{(Random Sampling)} (i) There is a sequence of sampling probabilities, $\rho_n$, such that
\begin{equation*}
\Pr\left(\bs R_n=\bs r\right) =\rho_n^{\,\sum_{i=1}^n r_i} \left(1-\rho_n\right)
^{n-\sum_{i=1}^n r_i},
\end{equation*}
for all $n$-vectors $\bs r$ with $i$-th element $r_i\in \{0,1\}$.
(ii) The sequence of sampling rates, $\rho_n$, satisfies $n\rho_n\rightarrow\infty$ and $\rho_n\rightarrow\rho\in [0,1]$.
\end{assumption}
Assumption \ref{assumption:randomsampling}$(i)$ states that each population unit is sample with probability $\rho_n$ independently of the others. The first part of Assumption \ref{assumption:randomsampling}$(ii)$ guarantees that as the population size increases, the (expected) sample size also increases. The second part of Assumption \ref{assumption:randomsampling}$(ii)$ allows for the possibility that asymptotically the sample size is a negligible fraction of the population size so that the \ehww results, corresponding to $\rho=0$, are included as a special case of our results.
Next assumption is a regularity condition bounding moments.
\begin{assumption}\textsc{(Moments)}
\label{assumption:moments}
There exists some $\delta>0$ such that the sequences
\[
\frac{1}{n}\sum_{i=1}^n E[|Y_{n,i}|^{4+\delta}],\qquad \frac{1}{n}\sum_{i=1}^n E[\|X_{n,i}\|^{4+\delta}],\qquad\frac{1}{n}\sum_{i=1}^n \|Z_{n,i}\|^{4+\delta}
\]
are uniformly bounded.
\end{assumption}
Let
\[
W_n = \frac{1}{n}\sum_{i=1}^n \left(\begin{array}{c}Y_{n,i}\\X_{n,i}\\Z_{n,i}\end{array}\right)
\left(\begin{array}{c}Y_{n,i}\\X_{n,i}\\Z_{n,i}\end{array}\right)',\quad
\Omega_n=\frac{1}{n}\sum_{i=1}^n E\left[\left(\begin{array}{c}Y_{n,i}\\X_{n,i}\\Z_{n,i}\end{array}\right)
\left(\begin{array}{c}Y_{n,i}\\X_{n,i}\\Z_{n,i}\end{array}\right)'\right].
\]
So $\Omega_n=E[W_n]$, where the expectation is taken over the distribution of $\bs X_n$. We will consider also sample counterparts of $W_n$ and $\Omega_n$:
\[
\widetilde W_n = \frac{1}{N}\sum_{i=1}^N R_{n,i}\left(\begin{array}{c}Y_{n,i}\\X_{n,i}\\Z_{n,i}\end{array}\right)
\left(\begin{array}{c}Y_{n,i}\\X_{n,i}\\Z_{n,i}\end{array}\right)',\quad
\widetilde\Omega_n = \frac{1}{N}\sum_{i=1}^n R_{n,i} E\left[\left(\begin{array}{c}Y_{n,i}\\X_{n,i}\\Z_{n,i}\end{array}\right)
\left(\begin{array}{c}Y_{n,i}\\X_{n,i}\\Z_{n,i}\end{array}\right)'\right],
\]
where $\widetilde\Omega_n =E[\widetilde W_n|\br_n]$. We will use superscripts to indicate submatrices. For example,
\[
W_n=\left(\begin{array}{ccc}W_n^{YY} & W_n^{YX} & W_n^{YZ}\\ W_n^{XY} & W_n^{XX} & W_n^{XZ}\\ W_n^{ZY} & W_n^{ZX} & W_n^{ZZ}\end{array}\right),
\]
with analogous partitions for $\Omega_n$, $\widetilde W_n$, and $\widetilde\Omega_n$. Notice that the transformation in (\ref{equation:transformation}) implies that $\Omega_n^{XZ}$ and $\Omega_n^{ZX}$ are matrices will all zero entries.
We first obtain converge results for the sample objects, $\widetilde W_n$ and $\widetilde \Omega_n$.
\begin{lemma}
\label{lemma:convergence2moments}
Suppose Assumptions \ref{assumption:assignment}-\ref{assumption:moments} hold. Then, $\widetilde W_n - \Omega_n \stackrel{p}{\rightarrow} 0$ , $\widetilde \Omega_n - \Omega_n \stackrel{p}{\rightarrow} 0$ and $\widetilde W_n-W_n\stackrel{p}{\rightarrow} 0$.
\end{lemma}
Next assumption imposes convergence of second moments in the population.
\begin{assumption}\textsc{(Convergence of moments)}
\label{assumption:convergence}
$\Omega_n\rightarrow \Omega$, which is full rank.
\end{assumption}
\subsection{Descriptive and Causal Estimands}
\label{subsection:descriptiveandcausal}
We now define the descriptive and causal estimands that generalize $\theta ^{\descr}$, $\theta ^
\causal,\sample}$, and $\theta^{\causal}$ from Section \re
{section:cases} to a regression context.
\begin{definition}{\sc Causal and Descriptive Estimands}\\
For a given population $n$, with potential outcome functions $\bs Y_n()$, causes $\bs X_n$, attributes $\bs Z_n$, and sampling indicators $\bs R_n$:\vspace*{-5pt}
\begin{itemize}\setlength\itemsep{-2pt}
\item[(i)] Estimands are functionals of $(\bs Y_n(),\bs X_n ,\bs Z_n,\bs R_n)$, exchangeable in the rows of the arguments.
\item[(ii)] Descriptive estimands are estimands that can be written in terms of $\bs Y_n$, $\bs X_n$, and $\bs Z_n$, free of dependence on $\bs R_n$, and free of dependence on $\bs Y_n()$ beyond dependence on $\bs Y_n$.
\item[(iii)] Causal estimands are estimands that
cannot be written in terms of $\bs Y_n$, $\bs X_n$, $\bs Z_n$, and $\bs R_n$, because they depend on the potential outcome functions $\bs Y_n()$ beyond the realized outcomes, $\bs Y_n$.
\end{itemize}
\end{definition}
Causal estimands depend on the values of potential outcomes beyond the values that can be inferred from the realized outcomes.
Given a sample, the only reason we may not be able to infer the value of a descriptive estimand is that we may not see all the units in the population. In contrast, even if we observe all units in a population, we may not be able to infer the value of a causal estimand because its value depends on potential outcomes.
We define three estimands of interest,
\begin{equation}
\label{equation:descriptive}
\left(\begin{array}{c}\theta^{\desc}_n\\ \gamma^{\desc}_n\end{array}\right)=\left(\begin{array}{cc} W^{XX}_n& W^{XZ}_n\\ W^{ZX}_n& W^{ZZ}_n\end{array}\right)^{-1}\left(\begin{array}{c} W^{XY}_n\\ W^{ZY}_n\end{array}\right),
\end{equation}
\begin{equation}
\label{equation:causalsample}
\left(\begin{array}{c}\theta^{\causal, \sample}_n\\ \gamma^{\causal, \sample}_n\end{array}\right)=\left(\begin{array}{cc} \widetilde \Omega^{XX}_n& \widetilde\Omega^{XZ}_n\\ \widetilde\Omega^{ZX}_n& \widetilde\Omega^{ZZ}_n\end{array}\right)^{-1}\left(\begin{array}{c} \widetilde\Omega^{XY}_n\\ \widetilde\Omega^{ZY}_n\end{array}\right),
\end{equation}
and
\begin{equation}
\label{equation:causal}
\left(\begin{array}{c} \theta^{\causal}_n\\ \gamma^{\causal}_n\end{array}\right)=\left(\begin{array}{cc} \Omega^{XX}_n& \Omega^{XZ}_n\\ \Omega^{ZX}_n& \Omega^{ZZ}_n\end{array}\right)^{-1}\left(\begin{array}{c} \Omega^{XY}_n\\ \Omega^{ZY}_n\end{array}\right).
\end{equation}
Alternatively, the estimands in (\ref{equation:descriptive}) to (\ref{equation:causal}) can be defined as the coefficients
that correspond to the orthogonality conditions in terms of the residuals
$e_{n,i}=Y_{n,i}-X_{n,i}'\theta_n-Z'_{n,i}\gamma_n$,
\[
\frac{1}{n}\sum_{i=1}^n \left(\begin{array}{c}X_{n,i}\\ Z_{n,i}\end{array}\right)e_{n,i}=0,\quad
\frac{1}{n}\sum_{i=1}^n R_{n,i}E\left[\left(\begin{array}{c}X_{n,i}\\ Z_{n,i}\end{array}\right)e_{n,i}\right]=0,\quad
\frac{1}{n}\sum_{i=1}^n E\left[\left(\begin{array}{c}X_{n,i}\\ Z_{n,i}\end{array}\right)e_{n,i}\right]=0,
\]
respectively. We will study the properties of the least squares estimator, $\widehat\theta_n$, defined by
\[
\left(\begin{array}{c}\widehat\theta_n\\ \widehat\gamma_n\end{array}\right)=\left(\begin{array}{cc}\widetilde W^{XX}_n& \widetilde W^{XZ}_n\\ \widetilde W^{ZX}_n&\widetilde W^{ZZ}_n\end{array}\right)^{-1}\left(\begin{array}{c}\widetilde W^{XY}_n\\ \widetilde W^{ZY}_n\end{array}\right),
\]
as an estimator of the parameters defined in equations (\ref{equation:descriptive}) to (\ref{equation:causal}).
Notice that, by the law of total expectation, and because the potential outcome functions are fixed in our framework, $\theta^{\causal, \sample}_n$ and $\theta^{\causal}_n$ are causal estimands, according to our definition, while, $\theta^{\desc}_n$ is not. The fact that an estimand is causal according to our definition does not imply it has an interpretation as an average causal effect. In Section \ref{section:causal} we present conditions under which the regression estimand does have such an interpretation.
\subsection{Causal Interpretations of the Estimands}
\label{section:causal}
By construction, the descriptive estimand can be interpreted as the set of coefficients of a population best linear predictor (least squares). A more challenging question concerns the interpretation of the two causal estimands, and in particular their relation to the potential outcome functions. In this section we investigate this question.
The first part of our proposed set of conditions for a causal interpretation of $\theta^{\causal,\sample}_n$ and $\theta^{\causal}_n$ generalizes random assignment.
\begin{assumption}
\label{assumption:unconf}\textsc{(Linearity of the Expected Assignment)}
There exists a sequence of real matrices $B_n$ such that
\begin{equation*}
E[U_{n,i}]=B_n Z_{n,i}.
\end{equation*}
for $n$ large enough.
\end{assumption}
Because of Lemma \ref{lemma:convergence2moments} and Assumption \ref{assumption:convergence}, $\Lambda_n=B_n$, which implies that $E[X_{n,i}]=0$. Assumption \ref{assumption:unconf} looks very different from conventional exogeneity or unconfoundedness conditions, where the residuals, $e_{n,i}$, are assumed to be (mean-) independent of the regressors, and so it merits some discussion. A special case of this assumption arises in the context of a randomized assignment, when $E[U_{n,i}]$ is constant across units. In that case Assumption \ref{assumption:unconf} holds as long as there is an intercept in the set of attributes. More generally, Assumption \ref{assumption:unconf} relaxes the completely randomized assignment setting, by allowing the distribution of $U_{n,i}$ to depend on the attributes. However, this dependence is restricted in that the mean of $U_{n,i}$ is linear in $Z_{n,i}$. For example, Assumption \ref{assumption:unconf} holds automatically when $U_{n,1},\ldots, U_{n,n}$ are identically distributed and $Z_{n,i}$ contains a saturated set on indicators for all possible values of the attributes. Later in this section, we will show that under a set of conditions that includes Assumption \ref{assumption:unconf}, $\theta_n^\causal$ and $\theta^{\causal,\sample}$ can be interpreted as weighted averages of unit-level causal effects. The connection between linearity in the ``propensity score'' (\citet{rosenbaum1983central}, in our analysis represented by $E[U_{n,i}]=B_n Z_{n,i}$) and the interpretation of population regression coefficients as weighted averages of heterogeneous causal effects has been previously noticed in related contexts (see \citeauthor{angrist1998}, \citeyear{angrist1998}; \citeauthor{angristpischke}, \citeyear{angristpischke}; \citeauthor{aronow2016}, \citeyear{aronow2016}; \citeauthor{tymon2017}, \citeyear{tymon2017}).
\begin{assumption}
\label{assumption:linear}\textsc{(Linearity of Potential Outcomes)}
\[
Y_{n,i} = U_{n,i}'\theta_{n,i}+\xi_{n,i}
\]
almost surely, where $\theta_{n,i}$ and $\xi_{n,i}$ are non-stochastic.
\end{assumption}
In this formulation, any dependence of $Y_{n,i}$ on $Z_{n,i}$ or on unobserved attributes is subsumed by $\theta_{n,i}$ and $\xi_{n,i}$, which are non-stochastic. Each element of the vector $\theta_{n,i}$ represents the causal effect on $Y_{n,i}$ of increasing the corresponding value of $U_{n,i}$ in one unit.
\begin{theorem}
\label{theorem:causal}
Suppose Assumptions
\ref{assumption:assignment}-\ref{assumption:linear}
hold. Then, for $n$ large enough,
\begin{align*}
\theta_n^{\causal}&=\left(\sum_{i=1}^n E\left[W_{n,i}^{XX}
\right]\right)^{-1}\sum_{i=1}^nE\left[W_{n,i}^{XX}\right]\theta_{n,i},\\ \shortintertext{and, with probability approaching one,}
\theta_n^{\causal, \sample}&=\left(\sum_{i=1}^n R_{n,i}E\left[W_{n,i}^{XX}
\right]\right)^{-1}\sum_{i=1}^n R_{n,i}E\left[W_{n,i}^{XX}\right]\theta_{n,i},
\end{align*}
where $W_{n,i}^{XX}=X_{n,i}X_{n,i}'$.
\end{theorem}
The linearity in Assumption \ref{assumption:linear} is a strong restriction in many settings. In some other settings, in particular, when the causal variable is binary or, more generally when the causal variable takes on only a finite number of values, it is immediate to enforce this assumption by including in $U_{n,i}$ indicator variables representing each but one of the possible values of the cause. Assumption \ref{assumption:linear} can be relaxed at the cost of introducing additional complication in the interpretation of the estimands.
\begin{theorem}
\label{theorem:nonlinear}
Suppose that Assumptions \ref{assumption:assignment}-\ref{assumption:unconf} hold. Moreover, assume that $X_{n,1}, \ldots, X_{n,n}$ are continuous random variables with convex and compact supports, and that the potential outcome functions, $Y_{n,i}()$ are continuously differentiable. Then, there exist random variables $v_{n,1},\ldots, v_{n,n}$ such that, for $n$ sufficiently large,
\begin{align*}
\theta_n^{\causal}&=\left(\sum_{i=1}^n E\left[W_{n,i}^{XX}
\right]\right)^{-1}\sum_{i=1}^nE\left[W_{n,i}^{XX}\varphi_{n,i}\right],\\ \shortintertext{and}
\theta_n^{\causal, \sample}&=\left(\sum_{i=1}^n R_{n,i}E\left[W_{n,i}^{XX}
\right]\right)^{-1}\sum_{i=1}^n R_{n,i}E\left[W_{n,i}^{XX}\varphi_{n,i}\right],
\end{align*}
where $\varphi_{n,i}$ is the derivative of $Y_{n,i}()$ evaluated at $v_{n,i}$.
\end{theorem}
\begin{comment} Here, we provide a simple example that shows how the result in Theorems \ref{theorem:causal} and \ref{theorem:nonlinear} may not hold in the absence of Assumption \ref{assumption:unconf}.
\begin{table}[ht]
\caption{}
\label{table:example}\vskip1cm
\begin{center}
\begin{tabular}{c|rrrc}
Unit & $Y_{i}(x)$ & $Z_i$ & $E[U_i]$ & $\mbox{var}(U_i)$\\
\hline\\[-2ex]
$1$ & $a$ & $-1$ & $b$ & $1$\\
$2$ & $0$ & $0$ & $-2b$ & $1$\\
$3$ & $2a$ & $1$ & $b$ & $1$
\end{tabular}
\end{center}
\end{table}
Consider the population with three units described in Table \ref{table:example} (where, for simplicity, we drop the subscript $n$). In this example, $E[U_{i}]=3b Z_i^2-2b$ is a non-linear function of $Z_{i}$. Notice that
\[
\sum_{i=1}^3 E[U_i]/3=\sum_{i=1}^3 E[U_i] Z_i/3=0,
\]
so that $X_i=U_i$. Therefore, $E[X_i^2]=E[U_i^2]$. Also, because potential outcomes do not depend on $X_i$, it follows that $E[X_iY_i]=E[X_i]Y_i=E[U_i]Y_i$. As a result,
\[
\theta^{\causal}=
\left(\sum_{i=1}^3 E[X_i^2]\right)^{-1}\sum_{i=1}^3 E[X_iY_i]=\frac{ab}{2b^2+1},
\]
which is different from zero as long as $ab\neq 0$. In this example all the potential outcome functions $Y_i(x)$ are flat as a function of $x$, so all unit-level causal effects of the type $Y_i(x)-Y_i(x')$ are zero, and yet the causal least squares estimand can be positive or negative depending on the values of $a$ and $b$.\hfill$\square$
\end{comment}
\subsection{The Asymptotic Distribution of The Least Squares Estimator}
\label{section:distribution}
In this section we present one of the main results of the article, describing
the properties of the least squares estimator viewed as an estimator of the
causal estimands and, separately, viewed as an estimator of the descriptive
estimand.
In contrast to Section \ref{section:cases}, we do not have exact results, relying instead on asymptotic results based on sequences of populations.
First, we define the population residuals, denoted by $\varepsilon _{n,i}$, relative to the population causal estimands,
\begin{equation}
\label{equation:causalres}
\varepsilon _{n,i}=Y_{n,i}-X_{n,i}'\theta _n^\causal-Z_{n,i}'\gamma _n^\causal.
\end{equation}
\begin{comment}
The definition of the residuals, $\varepsilon_{n,1}, \ldots, \varepsilon_{n,n}$, mirrors that in conventional regression analysis, but their properties are conceptually different. For instance, the residuals need not be stochastic. If they are stochastic, they are so because of their dependence on $\bs X_n$.\hfill$\square$
\end{comment}
Under the assumption that the $X_{n,i}$ are jointly independent (but not necessarily identically distributed), the $n$ products $X_{n,i} \varepsilon _{n,i}$ are jointly independent but not identically distributed. Most importantly, in general the expectations $E[X_{n,i} \varepsilon _{n,i}] $ may vary across $i$, and need not all be zero. However, as shown in Section \ref{subsection:descriptiveandcausal}, the {\it averages} of these expectations over the entire population are guaranteed to be zero by the definition of $(\theta^\causal_n,\gamma^\causal_n)$. Define the limits of the population variance,
\begin{equation*}
\Delta^{\cond}=\lim_{n\rightarrow \infty}\frac{1}{n}\sum_{i=1}^{n}\mbox{var}\left(X_{n,i} \varepsilon _{n,i}\right),
\end{equation*}
and the expected outer product
\begin{equation*}
\Delta^{\ehw}=\lim_{n\rightarrow \infty}\frac{1}{n}\sum_{i=1}^{n} E\left[X_{n,i}\varepsilon_{n,i}^2X_{n,i}'\right].
\end{equation*}
The difference between $\Delta^\ehw$ and $\Delta^\cond$ is the limit of the average outer product of the means,
\begin{equation*}
\Delta^\mu=\Delta^\ehw-\Delta^\cond=\lim_{n\rightarrow \infty}\frac{1}{n}\sum_{i=1}^{n}
E[X_{n,i}\varepsilon_{n,i}]E[X_{n,i}\varepsilon_{n,i}]',
\end{equation*}
which is positive semidefinite. We assume existence of these limits.
\begin{assumption} \textsc{(Existence of Limits)}
\label{assumption:limits}
$\Delta^\cond$ and $\Delta^\ehw$ exist and are positive definite.
\end{assumption}
\begin{theorem}
\label{theorem:asympdist}
Suppose Assumptions \ref{assumption:assignment}-\ref{assumption:limits} hold, and let $H = \Omega^{XX}=\lim_{n\rightarrow\infty} \Omega^{XX}_n$. Then,
\vspace*{-10pt}
\begin{itemize}\setlength\itemsep{-2pt}
\item[(i)]
\begin{equation*}
\sqrt{N}\left(
\widehat{\theta}_n-\theta_n^{\causal}\right) \overset{d}{\longrightarrow }\mathcal{N}\left(0,H^{-1}\left(\rho\Delta^{\cond}+(1-\rho) \Delta^\ehw
\right) H^{-1}\right) ,
\end{equation*}
\item[(ii)]\begin{equation*}
\sqrt{N}\left(\widehat{\theta}_n-\theta_n^{\causal,\sample}
\right) \overset{d}{\longrightarrow}\mathcal{N}\left(0,H^{-1} \Delta^{\cond}H^{-1}\right) ,
\end{equation*}
\item[(iii)]
\begin{equation*}
\sqrt{N}\left(
\widehat{\theta}_n-\theta_n^{\descr}
\right) \overset{d}{\longrightarrow }\mathcal{N}\left(0,(1-\rho )H^{-1}\Delta^\ehw H^{-1}\right) .
\end{equation*}
\end{itemize}
\end{theorem}
\begin{comment}
For both the population causal and the descriptive estimand the asymptotic variance in the case with $\rho =0$ reduces to the standard \ehww variance, $H^{-1}\Delta^{\ehw}H^{-1}$. If the sample size is non-negligible as a fraction of the population size, $\rho >0$, the difference between the \ehww variance and the finite population causal variance is positive semi-definite and equal to $\rho H^{-1} (\Delta^{\ehw}-\Delta^\cond)H^{-1}$.\hfill$\square$
\end{comment}
\subsection{The Variance Under Correct Specification}
Consider a constant treatment effect assumption, which is required for a correct specification of a linear regression function as a function that describes potential outcomes.
\begin{assumption}\textsc{(Constant Treatment Effects)}
\label{assumption:cte}
\[
Y_{n,i} = U_{n,i}'\theta_n+\xi_{n,i}
\]
almost surely, where $\theta_n$ and $\xi_{n,i}$ are non-stochastic.
\end{assumption}
This strengthens Assumption \ref{assumption:linear} by requiring that the $\theta_{n,i}$ do not vary by $i$.
Under Assumption \ref{assumption:cte}, Theorem \ref{theorem:causal} implies that $\theta_n^\causal=\theta_n$ (although it need not be the case that $\theta^\desc=\theta_n$). Then, for
\[
\lambda_n = \left(\sum_{i=1}^n Z_{n,i}Z_{n,i}'\right)^{-1}\sum_{i=1}^n Z_{n,i}\xi_{n,i}
\]
we obtain that equation (\ref{equation:causalres}) holds for $\gamma_n^\causal=\Lambda_n'\theta_n+\lambda_n$ and $\varepsilon_{n,i}=\xi_{n,i}-Z_{n,i}'\lambda_n$. In this case, the residuals, $\varepsilon_{n,i}$, are non-stochastic. As a result, $E[X_{n,i}\varepsilon_{n,i}]=E[X_{n,i}]\varepsilon_{n,i}=0$, with implies $\Delta^\mu=\Delta^\ehw-\Delta^\cond=0$. This leads to the following result.
\begin{theorem} \label{theorem:neyman}
Suppose that Assumptions \ref{assumption:assignment}-\ref{assumption:cte} hold. Then, \begin{equation*} \sqrt{N}\left( \widehat{\theta}-\theta^{\causal}_n \right) \overset{d}{\longrightarrow }\mathcal{N}\left(0,H^{-1}\Delta^\ehw H^{-1}\right) , \end{equation*} irrespective of the value of $\rho$.
\end{theorem}
Notice that the result of the theorem applies also with $\theta_n^{\causal,\sample}$ replacing $\theta_n^{\causal}$ because the two parameter vectors are identical (with probability approaching one) under Assumption \ref{assumption:cte}.
\begin{comment}
The key insight in this theorem is that the asymptotic variance of $\widehat\theta_n$ does not depend on the ratio of the sample to the population size when the regression function is correctly specified. Therefore, it follows that the usual \ehww variance matrix is correct for $\widehat\theta_n$ under these assumptions. For the case with $X_{n,i}$ binary and no attributes beyond the intercept, this result can be inferred directly from Neyman's results for randomized experiments (\citeauthor{neyman1923}, \citeyear{neyman1923}). In that case, the result of Theorem \ref{theorem:neyman} follows from the restriction of constant treatment effects, $Y_{n,i}(1)-Y_{n,i}(0)=\theta_n$, which is extended to the more general case of non-binary regressors in Assumption \ref{assumption:cte}. The asymptotic variance of $\widehat{\gamma}_n$, the least squares estimator of the coefficients on the attributes, still depends on the ratio of sample to population size, and it can be shown that the conventional robust \ehww estimator continues to over-estimate the variance of $\widehat{\gamma}_n$.\hfill$\square$
\end{comment}
\section{Estimating the Variance}
\label{estimatingvariance}
Now let us turn to the problem of estimating the variance for the descriptive and causal estimands. There are four components to the asymptotic variance, $\rho$, $H$, $\Delta^\ehw$ and $\Delta^\cond$. The first three are straightforward to estimate. The ratio $\rho$ can be estimated as $N/n$. To estimate $H$, first estimate $\Lambda_n$ as
\[
\widehat{\Lambda}_n=\left(\sum_{i=1}^n R_{n,i} U_{n,i} Z_{n,i}'\right)\left(\sum_{i=1}^n R_{n,i} Z_{n,i} Z_{n,i}'\right)^{-1}.
\]
Then one can estimate $H$ as the average of the matrix of outer products over the sample:
\begin{equation*}
\widehat H_n=\frac{1}{N}\sum_{i=1}^{n}R_{n,i} \left(
U_{n,i}-\widehat{\Lambda}_n Z_{n,i}
\right) \left(
U_{n,i}-\widehat{\Lambda}_n Z_{n,i}
\right)'.
\end{equation*}
It is also straightforward to estimate $\Delta^\ehw$. First we estimate the residuals for the units in the sample,
$\widehat{\varepsilon}_{n,i}=Y_{n,i}-(U_{n,i}-\widehat{\Lambda}_n Z_{n,i})'\widehat\theta_n
-Z_{n,i}'\widehat\gamma_n,$ and then we estimate $\Delta^\ehw$ as:
\begin{equation*}
\widehat{\Delta}^\ehw_n= \frac{1}{N}\sum_{i=1}^n R_{n,i}
(U_{n,i}-\widehat{\Lambda}_n Z_{n,i})\,\widehat{\varepsilon}^2_{n,i}\,
(U_{n,i}-\widehat{\Lambda}_n Z_{n,i})'.
\end{equation*}
The \ehww variance, $V^\ehw=H^{-1}\Delta^\ehw H^{-1}$, is then estimated as
\begin{equation*}
\widehat V^\ehw_n=\widehat H^{-1}_n\widehat\Delta^\ehw_n\widehat H^{-1}_n.
\end{equation*}
\begin{lemma}\label{lemma:convergvar}
Suppose Assumptions \ref{assumption:assignment}-\ref{assumption:unconf} and \ref{assumption:limits} hold with $\delta=4$. Then,
\[
\widehat V^\ehw_n\stackrel{p}{\longrightarrow} V^\ehw.
\]
\end{lemma}
Alternatively one can use resampling methods such as the bootstrap \citep[e.g.,][]{efron1987jackknife}.
It is more challenging to estimate $\Delta^\cond$. The reason is the same that makes it impossible to obtain unbiased estimates of the variance of the estimator for the average treatment effect in the example in Section \ref{section:cases}. In that case there are three terms in the expression for the variance in equation (\ref{equation:totalvariance}). The first two are straightforward to estimate, but the third one, $S^2_\theta/n$ cannot be estimated consistently because we do not observe both potential outcomes for the same units. Often, researchers use the conservative estimator based on ignoring $S^2_\theta/n$. If we proceed in the same fashion for the regression context of Section \ref{main}, we obtain the conservative estimator $\widehat V^\ehw$, based on ignoring $\Delta^\mu$. We show, however, that in the presence of attributes we can improve the variance estimator. We build on \citet{abadie2008estimation}, \citet{abadie2014inference}, and \citet{fogarty2016regression} who, in contexts different than the one studied in this article, have used the explanatory power of attributes to improve variance estimators. While \citet{abadie2008estimation}, \citet{abadie2014inference} use nearest-neighbor techniques, here we follow \citet{fogarty2016regression} and apply linear regression techniques. The proposed estimator replaces the expectations $E[X_{n,i}\varepsilon_{n,i}]$, which cannot be consistently estimated, with predictors from a linear least squares projection of estimates of $X_{n,i}\varepsilon_{n,i}$ on the attributes, $Z_{n,i}$. Let $\widehat X_{n,i}= U_{n,i}-\widehat\Lambda_n Z_{n,i}$, and
\[
\widehat G_n = \left(\frac{1}{N}\sum_{i=1}^n R_{n,i} \widehat X_{n,i}\widehat\varepsilon_{n,i}Z_{n,i}'\right)\left(\frac{1}{N}\sum_{i=1}^n R_{n,i}Z_{n,i}Z_{n,i}'\right)^{-1}.
\]
The matrix $\widehat G_n$ contains the coefficients of a least squares regression of $ \widehat X_{n,i}\widehat\varepsilon_{n,i}$ on $Z_{n,i}$.
The next assumption ensures convergence of $\widehat G_n$.
\begin{assumption}
\label{assumption:convergG}
\[
\frac{1}{n}\sum_{i=1}^n E[X_{n,i}\varepsilon_{n,i}]Z_{n,i}'
\]
has a limit.
\end{assumption}
Consider now the following estimator,
\[
\widehat \Delta^Z_n = \frac{1}{N}\sum_{i=1}^n R_{n,i}\left(\widehat X_{n,i}\widehat\varepsilon_{n,i}-\widehat G_nZ_{n,i}\right)\left(\widehat X_{n,i}\widehat\varepsilon_{n,i}-\widehat G_nZ_{n,i}\right)'.
\]
which uses $\widehat G_nZ_{n,i}$ in lieu of a consistent estimator of $E[X_{n,i}\varepsilon_{n,i}]$. Notice that we do not assume that $E[X_{n,i}\varepsilon_{n,i}]$ is linear in $Z_{n,i}$. However, we will show that, as long as the attributes can linearly explain some of the variance in $\widehat X_{n,i}\widehat\varepsilon_{n,i}$, the estimator $\widehat \Delta^Z_n$ is smaller (in a matrix sense) than $\widehat\Delta^\ehw_n$. Moreover, $\widehat\Delta^Z_n$ remains conservative in large samples. These results are provided in the following lemma.
\begin{lemma}\label{lemma:convergvar2}
Suppose Assumptions \ref{assumption:assignment}-\ref{assumption:unconf}, \ref{assumption:limits} and \ref{assumption:convergG} hold with $\delta=4$. Then, $0\leq \widehat\Delta^Z_n\leq \widehat\Delta^\ehw_n$, and $\widehat\Delta^Z_n\stackrel{p}{\rightarrow} \Delta^Z$, where $\Delta^\cond\leq \Delta^Z\leq \Delta^\ehw$ (all inequalities are to be understood in a matrix sense).
\end{lemma}
Variance estimators follow immediately from Lemma \ref{lemma:convergvar2} by replacing $\Delta^\cond$ with the estimate $\widehat\Delta^Z_n$ in the asymptotic variance formulas of Theorem \ref{theorem:asympdist}. These estimators are not larger (and typically smaller) than estimators based on $\widehat\Delta^\ehw_n$, and they remain conservative in large samples. For simplicity, Lemma \ref{lemma:convergvar2} is based on a linear predictor for $E[X_{n,i}\varepsilon_{n,i}]$. Modifications that accommodate nonlinear predictors are immediate, at the cost of additional assumptions.
\begin{comment}
A special case of the adjusted variance estimate is an estimate obtained
from stratifying the sample on the basis of attributes $Z_{n,i}$. In
particular, if $Z_{n,i}$ includes exhaustive, mutually exclusive dummy
variables -- or, if we reduce the information in $Z_{n,i}$ down to such
indicators -- then $\hat{\Delta}_{n}^{Z}$ reduces to the middle of the
sandwich in a commonly used estimator in the context of standard stratified
sampling. (See, for example, (\citet{wooldridge2010econometric}, Section 20.2.2).) Then, the
residuals from regressing $\hat{X}_{n,i}\hat{\varepsilon}_{n,i}$ on $Z_{n,i}$
are simply stratum-specific demeaned versions of $\hat{X}_{n,i}\hat
\varepsilon}_{n,i}$. Such a variance estimator is easy to obtain using
standard software packages that support regression with survey samples.
$\square$
\end{comment}
\section{Inference for Alternative Questions}
\label{section:OtherQuestions}
This article has focused on inference for descriptive and causal estimands in a single cross-section. For example, we might have a sample that includes outcomes from all countries in a particular year, say 2013. In words, we analyze inference for estimands of parameters that answer the following causal question: ``What is the difference between what the average outcome would have been in those countries in the year 2013 if all had been treated, and what the average outcome would have been if all had not been treated?'' We also analyze inference for estimands of parameters that can be used to answer descriptive questions, such as ``What was the difference in outcomes between Northern and Southern countries in the year 2013?''
These are not the only questions a researcher could focus on. An alternative question might be, ``what is the expected difference in average outcomes between Northern and Southern countries in a future year, say the year 2020,'' or ``what is the difference between what the average outcome would be in those countries in the year 2020 if all would be treated, and what the average outcome would be if none would be treated?'' Arguably in most empirical analyses that are intended to inform policy the object of interest depends on future, not simply on past, outcomes. This creates substantial problems for inference. Here we discuss some of the complications, but much of this is left for future work.
Our two main points are, first, that it is important to be explicit about the estimand, and second,
that the conventional robust standard errors were not designed to solve these problems and do not do so without strong, typically implausible, assumptions.
Formally questions that involve future values of outcomes for countries could be formulated in terms of a population of interest that includes as its units each country in a variety of different states of the world that might be realized in future years. This population is large if there are many possible realizations of states of the world (e.g., rainfall, local political conditions, natural resource discoveries, etc.), with a potentially complex dependence structure. Given such a population the researcher may wish to estimate, say the difference in average 2020 outcomes for two sets of countries, and calculate standard errors based on values for the outcomes for the same set of countries in an earlier year, say 2020. A natural estimator for the difference in average values for Northern and Southern countries in 2020 would be the corresponding difference in average values in 2013. However, even though such data would allow us to infer without uncertainty the difference in average outcomes for Northern and Southern countries in 2013, there would be uncertainty regarding the true value of that difference in the year 2020.
In order to construct confidence intervals for the difference in 2020, the researcher must make some assumptions about how country outcomes will vary from year to year. An extreme assumption is that outcomes in 2013 and 2020 for the same country are independent conditional on attributes, which would justify the conventional \ehww variance estimator. However, assuming that there is no correlation between outcomes for the same country in successive years appears highly implausible. In fact any assumption about the magnitude of this correlation in the absence of direct information about it in the form of panel data would appear to be controversial. Such assumptions would also depend heavily on the future year for which we would wish to estimate the difference in averages, again highlighting the importance of being precise about the estimand.
Although in this case there is uncertainty regarding the difference in average outcomes in 2020 despite the fact that the researchers observes (some) information on all countries in the population of interest, we emphasize that the assumptions required to validate the application of \ehww standard errors in this setting are strong and arguably implausible. Moreover, researchers rarely formally state the population of interest, let alone state and justify the assumptions that justify inference.
Generally, if future predictions are truly the primary question of interest, it seems prudent to explicitly state the assumptions that justify particular calculations for standard errors. Especially in the absence of panel data, the results are likely to be sensitive to such assumptions. With panel data the researcher may be able to estimate the dynamic process underlying the potential outcomes in order to obtain standard errors for the future predictions.
In practice it may be useful to report standard errors for various estimands. For example, if the primary estimand is an average causal effect in the future, it may still be useful to report estimates and standard errors for the same contemporaneous average causal effect, in combination with estimates and standard errors for the future average causal effect, in order to understand the additional uncertainty that comes with predictions for a future period.
We leave this direction for future work.
\section{Conclusion}
In this article we study the interpretation of standard errors in regression analysis when the assumption that the sample is drawn randomly from a large population of interest is not attractive. The conventional robust standard errors justified by the random sampling assumption do not necessarily apply in this case. We show that, by viewing covariates as potential causes in a Rubin Causal Model or potential outcome framework, we can provide a coherent interpretation for standard errors that allows for uncertainty coming from both random sampling and from conditional random assignment. The proposed standard errors may be different from the conventional ones.
In the current article we focus exclusively on regression models, and we provide a full analysis of inference for only a certain class of regression models with some of the covariates causal and some attributes. Thus, this article is only a first step in a broader research program. The concerns we have raised in this article arise in many other settings and for other kinds of hypotheses, and the implications would need to be worked out for those settings. Section \ref{section:OtherQuestions} suggests some directions we think are particularly natural to consider.
\newpage
| 3efc7009bee3c041cd0535337dadf639b22c7a02 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The social impact of online services built on information posted by their users
has also turned them into a lucrative medium for fraudulently influencing
public opinion~\cite{RRCC16,BLLTZ16,LZ16,SWEKVZZ13}. The need to aggressively
promote disinformation has created a black market for social network fraud,
that includes fake opinions and reviews, likes, followers and app
installs~\cite{MMLSV11,TSM,RR,RL,AV,AS,AR}. For instance, in
$\S$~\ref{sec:model:adversary}, we show that in fraud markets, a fake review
can cost between \$0.5 and \$3 and a fake social networking ``like'' can cost
\$2. The profitability of fraud suggests that current solutions that focus on
fraud detection, are unable to control organized fraud.
In this paper we introduce the concept of {\it fraud preemption systems},
solutions deployed to defend online systems such as social networks and app
markets. Instead of reacting to fraud posted in the past, fraud preemption
systems seek to discourage fraudsters from posting fraud in the first place. We
propose FraudSys, the first real-time fraud preemption system that reduces the
profitability of fraud from the perspective of both crowdsourced fraud workers
and the people who hire them. FraudSys imposes computational penalties: the
activity of a user (e.g., review, like) is posted online only after his device
solves a computational puzzle. Puzzles reduce the profitability of fraud by (i)
limiting the amount of fraud per time unit that can be posted for any subject
hosted on the online system, and (ii) by consuming the computational resources
of fraudsters. For instance, Figure~\ref{fig:timeline} shows the timelines of
daily penalties assigned by FraudSys to two fraudsters detected in Google
Play. Based only on the recorded activities, FraudSys frequently assigned
hundreds of hours of daily computational penalties to a single fraudster.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{graphs/gplay/penalty/worker_penalties.eps}
\caption{Timeline of daily penalties (in hours) assigned by FraudSys to the
Google Play activities of two fraudsters we identified in Freelancer.com.
FraudSys imposes daily penalties of up to 1,247 hours to the fraudster at the
top and 3,079 hours for the fraudster at the bottom. As a result, the
fraudsters need to consume significant computational resources, while their
fraud is significantly delayed. This in turn reduces the number of payments
they would receive, and impacts their profitability.}
\label{fig:timeline}
\vspace{-15pt}
\end{figure}
\vspace{0.05in}\noindent
{\bf Challenges}.
Implementing a fraud preemption system raises several challenges. First,
FraudSys needs to detect fraud in real-time, whenever a user performs an
online system activity. Once assigned, a puzzle cannot be rescinded.
This is in contrast to existing systems (e.g., Yelp) that detect fraud
retroactively and can update previous decisions when new information surfaces.
Second, FraudSys needs to impose difficult puzzles on fraudsters, but minimally
impact the experience of honest users. This is made even more complex by the
fact that fraudsters can attempt to bypass detection and even obscure their
true ability to solve puzzles. Third, a stateful FraudSys service that
maintains state for millions of issued and active puzzles is expensive and
vulnerable to DoS attacks.
\vspace{0.05in}\noindent
\noindent
{\bf Our Contributions}.
Through FraudSys, we introduce several innovative solutions. To address the
first challenge, we exploit observations of fraudulent behaviors gleaned from
crowdsourcing sites and online systems, to propose a real-time graph based
algorithm to infer an {\it activity fraud score}, the chance that a user
activity is fraudulent [$\S$~\ref{sec:fraudsys:fraud}]. More specifically, we
introduce features that group fraudulent activities according to their human
creator: FraudSys identifies densely connected components in the co-review
graph of the subject targeted by the user activity, each presumably controlled
by a different fraudster. It then quantifies the connectivity of the user
account performing the action, to each component, and uses the highest
connectivity as features that may indicate that the user account and the
corresponding component are controlled by the same fraudster. FraudSys then
leverages supervised learning algorithms trained on these features to infer the
activity fraud score.
To address the second challenge, we develop adaptive hashrate inference
techniques to detect the computational capabilities of even adversarial
controlled devices to solve puzzles [$\S$~\ref{sec:fraudsys:all}], and devise
mechanisms to convert fraud scores to appropriate temporal penalty and puzzle
difficulty values [$\S$~\ref{sec:fraudsys:all}]. The puzzles assigned by
FraudSys do not alter the online experience of users, as they are solved on
their devices, in the background. However, the puzzles (1) significantly delay
detected fraudulent activities, posted only when the device returns the correct
puzzle solutions and (2) consume the computational resources of the fraudsters
who control the devices.
To address the third challenge, we propose the notion of {\it stateless
computational puzzles}, computational tasks that impose no storage overhead on
the fraud preemption system provider, but enable it to efficiently verify their
authenticity and the correctness of their solutions
[$\S$~\ref{sec:fraudsys:puzzler}]. Thus, the fraud preemption system can assign
a puzzle to a device from which an activity was performed on the online system,
without storing any state about this task. The device can return the results of
the puzzle in 5 seconds or 1 day, and the provider can verify that the task is
authentic, and its results are correct. This makes our approach resistant to
DoS attacks that attempt to exhaust the provider's storage space for assigned
puzzles.
We show that the computational penalty imposed by FraudSys on a fraudulent
activity is a function of the capabilities of the device from which it is
performed, and the probability that the activity is fraudulent. We introduce
and prove upper bounds on the profitability of fraud and the amount of fraud
that can be created for a single subject, per time unit
[$\S$~\ref{sec:properties}] .
We evaluate FraudSys on 23,028 fraudulent reviews (posted by 23 fraudsters from
2,664 user accounts they control), and 1,061 honest reviews we collected from
Google Play, as well as 274,297 fake and 180,400 honest likes from Facebook.
Even with incomplete data, FraudSys imposes temporal penalties that can be as
high as 3,079 hours per day for a single fraudster.
We also show that fraud does not pay off. At today's fraud payout, a
fraudster equipped with an AntMiner S7 (Bitcoin mining hardware) will earn
through fraud less than half the payout of honest Bitcoin mining.
\section{Related Work}
\noindent
{\bf Computation Based Fraud Preemption}.
Dwork and Naor~\cite{DN92} were the first to propose the use of computation to
prevent fraud, in particular spam, where the sender of an e-mail needs to
include the solution to a ``moderately hard function'' computed over a function
of the e-mail. Juels and Brainard~\cite{JB99} proposed to use puzzles to
prevent denial of service attacks, while Borisov~\cite{B06} introduced puzzles
that deter Sybils in peer-to-peer networks. In Borisov~\cite{B06}, newly joined
peers need to solve a puzzle to which all the other peers have contributed.
FraudSys not only seeks to adapt computational puzzles to prevent online system
fraud, but also needs to solve the additional challenges of building puzzles
whose difficulty is a function of the probability that an activity is
fraudulent, while handling heterogeneous user devices (e.g., ranging from
smartphones to machines that specialize in such puzzles).
\vspace{0.05in}\noindent
{\bf Graph Based Fraud Detection}.
Graphs have been used extensively to model relationships and detect fraudulent
behaviors in online systems. Ye and Akoglu~\cite{YL15} quantified the chance of
a subject to be a spam campaign target, then clustered spammers on a 2-hop
subgraph induced by the subjects with the highest chance values. Lu et
al.~\cite{LZXL13} proposed a belief propagation approach implemented on a
review-to-reviewer graph, that simultaneously detects fake reviews and
spammers (fraudsters).
Mukherjee et al.~\cite{MLG12} proposed a suite of features to identify reviewer
groups, as users who review many subjects in common but not much else, post
their reviews within small time windows, and are among the first to review the
subject. Hooi et al.~\cite{HSBSSF16} have recently shown that fraudsters have
evolved to hide their traces, by adding spurious reviews to popular items. To
identify ``camouflaged'' fraud, Hooi et al.~\cite{HSBSSF16} introduced
``suspiciousness'' metrics that apply to bipartite user-to-item graphs, and
developed a greedy algorithm to find the subgraph with the highest
suspiciousness. Akoglu et al.~\cite{ATK15} survey graph based online
fraud detection. \cite{FH16} provide a survey of community detection methods,
evaluation scores and techniques for general networks.
Unlike previous work, FraudSys assigns fraud scores to individual user
activities in {\it real time}, thus uses only partial information. To achieve
this, FraudSys develops and leverages features that quantify the connectivity
of the user activity to other groups of activities {\it previously} performed
by other fraudsters on the same subject. Further, FraudSys also imposes
computation and temporal penalties to discourage fraud creation.
\section{System and Adversary Model}
\label{sec:model}
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{./figures/fraudsys/system.model.eps}
\caption{System model. The user performs actions on the online service, from a
device that can range from a smartphone to a Bitcoin miner. The online service
implements and posts the activity only if and after the FraudSys service
validates it. The FraudSys functionality can be implemented by the online
service or by a third party provider.}
\label{fig:system}
\vspace{-15pt}
\end{figure}
Figure~\ref{fig:system} illustrates the three main components of the system
model. First, the online service (the \textit{service}) hosts the system
functionality, and stores information about user accounts and featured
\textit{subjects}. Subjects can be apps in stores like Google Play, or pages
for businesses, accounts and stories in social networks like Facebook.
Second, the users: they register with the service, record profile information
(e.g., name) and receive initial service credentials, including a unique id.
Users can access the online service from a variety of devices. For this, they
need to install a client (e.g., app) on each device they use. The online
service stores and maintains information about each device that the user has
used, e.g., to provide compatibility information on Google Play apps.
Users are encouraged to {\it act} on existing subjects. The activities include
posting reviews, comments, or likes, installing mobile apps, etc. The online
service associates statistics over the activities performed for each supported
subject. The statistics have a significant impact on the popularity and search
rank of subjects~\cite{Ankit.Jain,RankGooglePlay}, thus are targets of manipulation by
fraudsters (see $\S$~\ref{sec:model:adversary}).
The third component of the system model is the FraudSys service, whose goal is
to validate user activities. For increased flexibility, Figure~\ref{fig:system}
shows FraudSys as an independent provider. However, FraudSys can also be a
component of the online service.
\subsection{Adversary Model}
\label{sec:model:adversary}
We consider two types of adversaries -- adversarial owners and crowdsourced
fraud workers.
\vspace{0.05in}\noindent
{\bf Adversarial owners}.
Adversarial behaviors start with the subject owners. Adversarial owners seek to
fraudulently promote their subjects (or demote competitor subjects) in order to
bias the popularity and public opinion of specific subjects. For instance,
fraudulent promotions seek to make subjects more profitable~\cite{LZ16,AM12},
increase the ``reachability'' of malware (through more app installs), and boost
the impact of fake news.
\vspace{0.05in}\noindent
{\bf Fraud workers ($=$ fraudsters)}.
We assume that adversarial owners crowdsource this promotion task (also known
as search rank fraud) to {\it fraud workers}, or fraudsters. In this paper we
focus on two types of fraudulent activities: writing fake reviews in Google
Play and posting fake ``Likes'' in Facebook. We have studied fraudster
recruitment jobs in crowdsourcing sites and fraud posted in Google Play and
Facebook. This has allowed us to collect fraud data (see $\S$~\ref{sec:data})
and to identify several fraud behaviors: (i) more than one fraudster can target
the same subject; (ii) user accounts controlled by a fraudster tend to have a
significant history of common activities, i.e., performed on the same subjects;
and (iii) accounts controlled by different fraudsters tend to have few common
past activities.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{./figures/fraudsites/price_per_review_2.eps}
\caption{Price per review (minimum, average and maximum), for crowdsourcing
sites that focus on app market fraud. The sites offer ``fraud packages'' and
even discounts for bulk fake review purchases. A fake review costs between
\$0.5-\$3.
\label{fig:fraudsites}}
\vspace{-15pt}
\end{figure}
\vspace{0.05in}\noindent
{\bf Fraud incentives}.
We assume that fraud workers are rational, motivated by financial incentives.
That is, given an original investment in expertise and equipment, a fraud
worker seeks to maximize his revenue achieved per time unit.
Figure~\ref{fig:fraudsites} shows the minimum, average and maximum cost per
fraudulent activity, as advertised by several crowdsourcing and
fraud-as-a-service (FAAS) sites: a fake review for an app is worth between
\$0.5-\$3, while a fake social networking ``like'' can cost \$2. In contrast,
an adversarial owner may have both financial incentives (e.g., increased
market share for his subject, thus revenue), and external incentives (e.g.,
malware or fake news distribution).
\subsection{Fraud Preemption System Definition}
\label{sec:model:problem}
We introduce the concept of {\it fraud preemption systems}, that seek to
restrict the profitability of fraud for both fraudsters and the people who hire
the fraudsters (i.e., adversarial owners). Specifically, let $Sys =
(\mathcal{U}, \mathcal{S}, \mathcal{F}, P)$ be a system that consists of finite
sets of users ($\mathcal{U}$), subjects ($\mathcal{S}$) and fraudsters
($\mathcal{F}$) that interact through a set of procedures $P$. In the adversary
model of $\S$~\ref{sec:model:adversary}, we say that $Sys$ is a (p,a)-{\it
fraud preemption system} if it satisfies the following two conditions:
\begin{enumerate}
\item
{\bf Fraudster deterrence}:
The average payout per time unit of any fraudster in $\mathcal{F}$ does not
exceed $p$.
\item
{\bf Adversarial owner deterrence}:
The average number of fraudulent activities allowed for any subject in
$\mathcal{S}$ per time unit does not exceed $a$.
\end{enumerate}
\noindent
In addition, a puzzle-based fraud preemption system needs to satisfy the
following requirements:
\begin{enumerate}
\item
{\bf Real-time fraud detection}.
Detect fraud at the time it is created, with access to only limited information
(i.e., no knowledge of the future).
\item
{\bf Penalty accuracy}.
Impose difficult puzzles on fraudsters, but minimally impact the online
experience of honest users.
\item
{\bf Device heterogeneity}.
Both honest and fraudulent users may register and use multiple devices to
access the online service. Malicious users may obfuscate the computational
capabilities of their devices.
\item
{\bf Minimize system resource consumption}.
The high number of issued, active puzzles will consume the resources of the
FraudSys provider, and open it to DoS attacks.
\end{enumerate}
\section{FraudSys}
\label{sec:fraudsys}
\begin{figure}
\centering
\includegraphics[width=0.49\textwidth]{./figures/fraudsys/fraudsys.eps}
\caption{FraudSys architecture. The Fraud Detector module uses supervised
learning to assign a fraud score to user activities. The Fraud2Penalty module
converts the fraud score to a {\it time penalty}. The Hashrate Inference module
estimates the computational capabilities of the user device. Finally, the
Puzzler module generates a puzzle that the device should take approximately the
time penalty to solve.
\label{fig:fraudsys}}
\vspace{-15pt}
\end{figure}
We introduce FraudSys, a real-time fraud preemption system that requires users
to verify commitment through an imposed resource consumption action for each
activity they perform on the online system. Specifically,
FraudSys requires the device from which the activity was issued, to solve a
{\it computational puzzle}. FraudSys consists of the modules illustrated in
Figure~\ref{fig:fraudsys}: The Fraud Detection module takes as input a user
activity and the current state of the subject, and outputs a {\it fraud score}.
The Fraud2Penalty module converts the fraud score to a {\it time penalty}: the
time that the user's device will need to spend working on a computational
puzzle. The Hashrate Inference module interacts with the user device in order
to learn its puzzle solving capabilities. Finally, the Puzzler module uses the
inferred device capabilities to generate a puzzle that the device will take
a time approximately equal to {\it time penalty} to solve.
To address requirement \#1, the Fraud Detection module exploits the fraudulent
behaviors described in $\S$~\ref{sec:model:adversary}. It builds {\it
co-activity graphs} and extracts features that model the relationships between
the user performing the activity and other users that have earlier performed
similar activities for the same subject.
We address requirement \#2 through a two-pronged approach. First, the Fraud
Detection and Fraud2Penalty modules ensure that the difficulty of a FraudSys
puzzle will be a function of the detected probability of fraud: activities
believed to be honest will be assigned trivial puzzles, while increasingly
fraudulent activities will be assigned increasingly difficult puzzles. Second,
FraudSys does not change the experience of the user on the online system: the
user writes the review or clicks on the like button, then continues browsing or
quits the app. The assigned puzzle is solved in the background by the device on
which the activity was performed. However, FraudSys delays the publication of
the activity, until the device produces the correct puzzle solution.
To address requirement \#3, the Hashrate Inference module estimates the
hashrate of the device performing the activity, and provides the tool to punish
devices that cheat about their puzzle solving capabilities. To solve
requirement \#4, the Puzzler module generates puzzles that outsource the
storage constraints from the FraudSys service to the user devices that solve
the puzzles. In the following we detail each FraudSys module, starting with the
central puzzle creation module.
\begin{table}
\setlength{\tabcolsep}{.16677em}
\centering
\textsf{
\small
\begin{tabular}{l r}
\toprule
\textbf{Notation} & \textbf{Definition}\\
\midrule
$U$, $D$, $S$, $A$ & user, device, subject, activity\\
T & time of puzzle issue\\
$r$ & activity fraud score\\
\midrule
$\Delta$ & puzzle difficulty\\
$\eta_D$ & hashrate of device $D$\\
$\Gamma$ & puzzle cookie\\
$\Pi$ & puzzle\\
$target$ & puzzle target value\\
$\tau$ & temporal penalty\\
$q$ & number of shares (puzzle solutions)\\
\midrule
$K$ & secret key of FraudSys\\
\bottomrule
\end{tabular}
}
\caption{FraudSys symbol table.}
\label{fig:symbols}
\vspace{-15pt}
\end{table}
\subsection{The Puzzler Module: Stateless Puzzles}
\label{sec:fraudsys:puzzler}
Let $U$ be a user that performs an activity $A$ from a device $D$, on a subject
$S$ hosted by the online service. Table~\ref{fig:symbols}
summarizes the notations we use. The FraudSys service stores minimal state for
each registered user, and {\it serializes} his activities, see
Figure~\ref{fig:serialization}: the devices from which a user performs a
sequence of activities on the online service, are assigned one puzzle per
activity, each with its own timeout. The device needs to return the puzzle
solutions before the associated timeout. To implement this, for
each user $U$, the FraudSys service stores the following entry:
\[
U, [\langle D_i, \eta_i \rangle]_{i=1..d}, timeout,
\]
where, for each of the $i = 1 .. d$ devices registered by $U$, $D_i$ is the
device identifier and $\eta_i$ is its hashrate (puzzle solving capabilities
measure, see following), and $timeout$ is the latest time by which one of these
devices needs to return puzzle solutions.
FraudSys builds on the computational puzzles of Bitcoin, see~\cite{N08}. Let
$H^2(M)$ denote the double SHA-256 hash of a message $M$. Then, the FraudSys
puzzle issued to device $D$ consists of a $target$ value and a fixed string
$F$. We detail $F$ shortly. To solve the puzzle, $D$ needs to randomly choose
32 byte long $nonce$ values until it finds at least one that satisfies:
\vspace{-5pt}
\begin{equation}
H^2(nonce || F) < target
\label{eq:puzzle}
\end{equation}
That is, the double hash of the $nonce$ concatenated with $F$, needs to be
smaller than the $target$ value, another 32 byte long value. A smaller $target$
implies a harder puzzle. The largest $target$ acceptable by the system
is called $target\_1$, or {\it target of difficulty 1}.
Bitcoin has two drawbacks. First, the current difficulty of Bitcoin puzzles
requires computational capabilities that greatly exceed those of devices used
to access online services. Second, Bitcoin requires the network to maintain
state about issued puzzles. State storage exposes FraudSys to attacks, while
not storing state can enable adversaries to lower the difficulty of their
assigned puzzles. To address these problems we (i) change the $target\_1$
difficulty to allow trivial puzzles, and (ii) introduce {\it puzzle cookies},
special values that authenticate puzzles with minimal FraudSys state, see
following.
\vspace{0.05in}\noindent
{\bf Device hashrate and puzzle difficulty}.
We set the $target\_1$ value to be a 32 byte long value with one zero at the
beginning, e.g., $2^{255}-1$. In addition, the {\it hashrate} $\eta_D$ of a
device $D$ is a measure that describes the ability of the device to solve
puzzles. Since the puzzles need to be solved in a brute force approach, the
hashrate is measured in hashes per second. A relevant concept is the notion of
\textit{difficulty}, denoted by $\Delta$, a measure of how difficult it is to
solve a puzzle whose input values hash below a given target. Its relationship
to the above $target$ value is given by:
\begin{figure}
\centering
\includegraphics[width=0.43\textwidth,height=0.9in]{./figures/fraudsys/serialization.eps}
\vspace{-5pt}
\caption{Puzzle serialization: a user can perform multiple activities, but each
receives a different puzzle with its own timeout, authenticated through the
cookie $\Gamma$.}
\label{fig:serialization}
\vspace{-10pt}
\end{figure}
\begin{equation}
\Delta = \frac{target\_1}{target} = \frac{2^{255}-1}{target}
\label{eq:delta:target}
\end{equation}
\noindent
Given $\eta_D$, we derive the time
$\tau$ taken by $D$ to solve a puzzle with difficulty $\Delta$, as follows.
First, the number of hashes smaller than a given target is equal to the target.
For instance, the number of hashes smaller than $target\_1$ is $2^{255}-1$ .
Then, the probability $p$ of finding an input that hashes to a value smaller
than the target is equal to the target divided by the total number of hashes
($2^{256}$). Furthermore, the expected number of hashes, $E$, before achieving
the target is given by $1/p$. Thus:
\[
E = \eta_D \times \tau = \frac{2^{256}}{target} =
\frac{ 2^{256} }{target\_1}\times \frac{target\_1}{target} \approx 2\times \Delta
\]
and conclude that
\begin{equation}
\tau = \frac {2 \times \Delta}{\eta_D}
\label{eq:time:to:hashrate:puzzler}
\end{equation}
\noindent
For instance, the lowest puzzle difficulty is 1, which occurs when the $target$
has a prefix of one zero and the device is expected to generate 2 hashes before
solving the puzzle. Similarly, the maximum difficulty is $(2^{255}-1)$, for a
$target$ = 1, when the device is expected to perform $\frac{2^{255}-1}{1}
\times 2 \approx 2^{256}$ hashes.
\begin{figure}[t]
\renewcommand{\baselinestretch}{0.7}
\begin{minipage}{0.4\textwidth}
\begin{algorithm}[H]
\begin{tabbing}
XXX\=X\=X\=X\=X\=X\= \kill
1. \small{Object \mbox{\bf{FraudSysService}}}\\
2.\> \small{K: key;}\\
3. \> \small{Function\ {\bf BuildCookie}($U$, $D$, $S$, $A$, $q$)}\\
4.\>\> \small{$\eta_D$ := getHashrate($U, D$);}\\
5.\>\> \small{$r$ := computeFraudScore($U, S, A$);}\\
6.\>\> \small{$\tau$ := fraud2Penalty($r$);}\\
7.\>\> \small{$\Delta$ := $\eta_D \times \tau / 2q$}\\
8.\>\> \small{$oldT$ := getTimeout(U);}\\
9.\>\> \small{$newT$ := $oldT + \tau$;}\\
10.\>\> \small{$\Gamma$ := HMAC($K, U, D, S, newT, \Delta, A$);}\\
11.\>\> \small{setTimeout($U$, $newT$);}\\
12.\>\> \small{return $\Gamma$, $\Delta$, $newT$;}\\
13.\> \small{Function\ {\bf VerifyPuzzle}($U$, $D$, $S$, $A$, $timeout$, $\Gamma$, $\sigma$: share[$q$])}\\
14.\>\> \small{\mbox{\bf{if}}\ ($\Gamma$ != HMAC($K, U, D, S, A, timeout, \Delta$)\ return -1;}\\
15.\>\> \small{$target$ := getTarget($\Delta$);}\\
16.\>\> \small{\mbox{\bf{for}}\ ($i$ := 0;\ $i< q$;\ $i$++)}\\
17.\>\>\> \small{\mbox{\bf{if}}\ ($H^2(\sigma[i]\ ||\ \Gamma) > target$)\ return -1;}\\
18.\>\> \small{waitUntil($timeout$); post $A$;}\\
19.\>\> \small{$\tau' := T_c - T$;}\\
20.\>\> \small{\mbox{\bf{if}}\ (($\eta_D := 2 \Delta/\tau') \ge \eta_{min}$)}\\
21.\>\>\> \small{updateHashrate($U$, $D$, $\eta_D$);}\\
22. \small{Object\ \mbox{\bf{UserDevice}}}\\
23.\> \small{Function\ {\bf SolvePuzzle}($\Gamma$, $\Delta$, $timeout$, $q$)}\\
24.\>\> \small{$target$ := getTarget($\Delta$);}\\
25.\>\> \small{$\sigma$ := new share[q]; $i$ := 0;}\\
26.\>\> \small{\mbox{\bf{while}}\ ($i < q$)\ \mbox{\bf{do}}}\\
27.\>\>\> \small{$nonce$ := getRandom();}\\
28.\>\>\> \small{\mbox{\bf{if}}\ ($H^2(nonce\ ||\ \Gamma) < target$)}\\
29.\>\>\>\> \small{$\sigma[i]$ := $nonce$;}\\
30.\>\>\>\> \small{$i$ := $i$+1;}\\
31.\>\> \small{return\ $U, D, S, A, timeout, \Gamma, \sigma$;}
\end{tabbing}
\caption{FraudSys puzzle creation, verification and computation components.}
\label{alg:fraudsys}
\end{algorithm}
\end{minipage}
\normalsize
\vspace{-5pt}
\end{figure}
\vspace{0.05in}\noindent
{\bf The FraudSys puzzle and cookies}.
To minimize the storage imposed on the FraudSys service (see above), we
leverage the cookie concept~\cite{B96}. Algorithm~\ref{alg:fraudsys}
illustrates the puzzle creation, verification and computation components. The
FraudSys service generates and stores a secret key $K$ (line 2). When a user
$U$ performs an activity $A$ from a device $D$ on a subject $S$ of the online service,
the online service calls the BuildCookie function of the FraudSys service
(lines 3-11).
BuildCookie retrieves the hashrate of the device $D$ from the record stored by
FraudSys for $U$ (line 4). It then computes the fraud score associated to the
activity (line 5) then converts it to a time penalty $\tau$ (line 6). We
describe this functionality in the next subsections. BuildCookie then uses a
modified Equation~\ref{eq:time:to:hashrate:puzzler} to compute the difficulty
$\Delta$ that the puzzle should have (line 7). $\Delta$ is $q$ times smaller
than in Equation~\ref{eq:time:to:hashrate:puzzler}, as the puzzle solution
consists of $q$ shares, see SolvePuzzle.
BuildCookie gets the current timeout $oldT$ of $U$, and updates it to $newT$ by
adding the penalty $\tau$ to it (lines 8-9). It then computes the puzzle cookie
$\Gamma$,
\[
\Gamma = HMAC_K (U, D, S, A, timeout, \Delta)
\]
as a keyed HMAC~\cite{BCK96} over the user and device id, subject, activity,
new timeout and puzzle difficulty (lines 9-10). BuildCookie sets $U$'s timeout
value to the updated $newT$ value (line 11), then returns the following puzzle
(line 12) to the online service that forwards it to device $D$ (see
Figure~\ref{fig:system}):
\[
\Pi = \Gamma, \Delta, timeout.
\]
The puzzle cookie ensures that an adversary that modifies the puzzle's
difficulty or timeout, will be detected: the adversary does not know the key
$K$, which is a secret of the FraudSys service. Puzzle cookies are unique with
high probability, due to collision resistance properties of the HMAC, whose
input is non-repeating.
\vspace{0.05in}\noindent
{\bf Solving the puzzle}.
When the device $D$ receives the puzzle, it needs to solve it: search for $q$
$nonce$ values that satisfy the inequality $H^2(nonce\ ||\ \Gamma) < target$,
for a $target$ corresponding to the difficulty $\Delta$. Specifically, $D$
invokes the SolvePuzzle function (lines 23-31), that needs to identify $q$ {\it
shares}, i.e., nonce values that satisfy the puzzle. $q$ is a system parameter.
The function first uses Equation~\ref{eq:delta:target} to retrieve the $target$
value corresponding to the difficulty $\Delta$ (line 24). Then, it generates
random $nonce$ values until it identifies $q$ values that satisfy the puzzle
condition (lines 25-30). SolvePuzzle returns the identified shares (in the
$\sigma$ array), which are then sent to the online service and forwarded to the
FraudSys server, along with the user, device and subject ids, activity, timeout
and cookie of the received puzzle (see line 12 and Figure~\ref{fig:system}).
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{./figures/coreview/2clique.eps}
\caption{Visualization of the co-review graph of a fraudulent Google Play app.
The nodes represent user accounts; edges connect nodes corresponding to
accounts with common, past review activities. The nodes in each of the 2
clusters correspond to accounts controlled by the same fraudster.}
\label{fig:coreview:fraud}
\vspace{-15pt}
\end{figure}
\vspace{0.05in}\noindent
\noindent
{\bf Verification of puzzle correctness}.
Upon receiving these values, the FraudSys server invokes the VerifyPuzzle
function (lines 13-21), to verify its correctness as follows:
(1) Reconstruct the puzzle cookie $\Gamma$ based on the received values and the
secret key $K$. Verify that this cookie is equal to the received $\Gamma$
value (line 14). This ensures that all values, including the $timeout$
have not been altered by an adversary; and
(2)
Verify that each of the $q$ shares satisfy the puzzle (lines 15-17).
If these verifications succeed, FraudSys waits until $timeout$ expires to
confirm the user action $A$, for posting by the online service (line 18).
It then uses the time required by the device to solve the puzzle, to
re-evaluate the hashrate of the device (lines 19-20). It updates the stored
hashrate only if the new value is above a minimally accepted hashrate value
(lines 20-21).
\subsection{The Fraud Detection Module}
\label{sec:fraudsys:fraud}
To assign a fraud score to a user activity in real-time, the fraud detection
module can only rely on the existing history of the user and of the subject on
which the activity is performed. We propose an approach that builds on the {\it
co-activity graphs} of subjects, where nodes correspond to user accounts that
performed activities on the subject, and edges connect nodes whose user
accounts have a history of activities that targeted the same subjects. Edge
weights denote the size of that history. Figure~\ref{fig:coreview:fraud} shows
the co-review (where activities are reviews) graph of a fraudulent Google Play
app, that received fake reviews from 2 fraud workers. Each cluster is formed by
accounts controlled by one of the workers.
The fraud detection module leverages the adversary model findings
($\S$~\ref{sec:model:adversary}) that a fraudster-controlled user account that
performs a new activity on a subject, is likely to be well connected to the
co-activity graph of the subject, or at least one of its densely connected
sub-graphs. Figure~\ref{fig:coaction:dynamic} illustrates this approach: Let
$U$ be a user account that performs an activity $A$ for a subject $S$ at time
$T$. Let $G=(V, E)$ be the co-activity graph of $S$ before time $T$. Let $G_T
= (V_T, E_T)$ be the new co-activity graph of $S$, that also includes $U$,
i.e., $V_T = V \cup U$. Given $U$, $S$ and $G$, FraudSys extracts the following
features, that model the relationship of $U$ with $S$:
\begin{figure}
\centering
\includegraphics[width=3.5in]{./figures/fraudsys/fraudprob.eps}
\caption{Fraud detection illustration: temporal evolution of the co-activity
graph of a subject. The nodes represent user accounts that have performed an
activity on the subject. Edges connect accounts with common past activities. As
a new user account posts an activity, FraudSys assigns the activity a fraud
score (the $r_1 .. r_4$ values), based on its connectivity to previous
activities. Yellow nodes are considered fraudulent ($r >
0.5$).
\label{fig:coaction:dynamic}}
\vspace{-15pt}
\end{figure}
\vspace{0.05in}\noindent
$\bullet$
{\bf Connectivity features}.
The percentage of nodes in $V$ to whom $U$ is connected. The average weight of
the edges between $U$ and the nodes in $V$. The average weight of those edges
divided by the average weight of the edges in $E$. This feature will indicate
if $U$ increases or decreases the overall connectivity of $G$. The number of
triangles in $G_T$ that have $U$ as a vertex. The average edge weight of those
triangles.
\vspace{0.05in}\noindent
$\bullet$
{\bf Best fit connectivity features}.
Since $U$ may be controlled by one of multiple fraudsters who target $S$, $U$
may be better connected to the subgraph of $G$ controlled by that fraudster.
Then, use a weighted min-cut algorithm to partition $G$ into components
$G_1,..,G_k$, such that any node in a component is more densely connected to
the nodes in the same component than to the nodes in any of the other
components. $G_1,..,G_k$ may contain user accounts controlled by different
fraudsters, see Figure~\ref{fig:coreview:fraud}. Identify the component $G_b$,
$b \in \{1,..k\}$ to which $U$ is the most tightly connected (according to
the above connectivity features).
Output the connectivity features between $U$ and $G_b$.
\vspace{0.05in}\noindent
$\bullet$
{\bf Account based features}.
The number of activities previously performed by $U$. The age of $U$: the time
between $U$'s creation and the time when activity $A$ is performed on $S$. The
{\it expertise} of $U$: the number of actions of $U$ for subjects similar to
$S$. Similarity depends on the online service, e.g., same category apps in
Google Play, pages with similar topics in Facebook.
\vspace{0.05in}\noindent
The Fraud Detection module trains a probabilistic supervised learning algorithm
on these features and uses the trained model to output the probability that a
given activity is fraudulent. We detail the performance of various algorithms,
over data that we collected from Google Play and Facebook, in
$\S$~\ref{sec:evaluation}.
\vspace{0.05in}\noindent
{\bf Per-fraudster timeout}.
We exploit the ability of the fraud detection module to identify accounts
controlled by the same fraudster, to further restrict fraud. Specifically,
instead of storing a $timeout$ timestamp for each user account, FraudSys can
store a single $timeout$ per detected fraudster. Thus, FraudSys will accumulate
penalties in a single, per-fraudster account. This facilitates
Claim~\ref{claim:fps}.
\subsection{The Fraud2Penalty Module}
\label{sec:fraudsys:all}
Given the fraud score $r$ of an activity of user $U$ (output by the Fraud
Detection module), performed from a device $D$ associated with the account of
$U$ (see the model section), the Puzzler module generates a puzzle whose
difficulty is a function of both $r$ and the computational capability of $D$.
We now describe the Fraud2Penalty module, that converts $r$ into a time
penalty.
\begin{figure}
\centering
\includegraphics[width=0.47\textwidth]{graphs/conversion.function/all_functions2.eps}
\vspace{-5pt}
\caption{Comparison of functions to convert fraud scores (x axis) to time
penalties (y axis). The logistic function (red dot-line) exhibits the required
exponential increase.}
\label{fig:conversion}
\vspace{-15pt}
\end{figure}
We have explored several functions to convert the fraud score $r$ of a user
activity to a time penalty. Let $minh$ and $maxh$, and $minf$ and $maxf$,
denote the minimum and maximum times imposed on the device from which an
honest, respectively fraudulent activity is performed. Let $thr$ denote the threshold fraud score above which we start to consider a user
activity as being fraudulent. We propose a conversion function that increases
linearly when $r < thr$, and exponentially when $r > thr$. Specifically, we propose a flexible generalization of
the logistic function (when $r > thr$), where the parameter
$k$ is the {\it growth rate}:
\begin{equation}
\begin{cases}
\frac{maxh - minh}{thr} r + minh & 0 \leq r \leq thr \\
\frac{maxf}{1 + (\frac{maxf-minf}{minf})e^{-k(r-thr)}} & thr \leq r \leq 1
\end{cases}
\label{eq:fraud2penalty}
\end{equation}
\noindent
We have compared this logistic increase function with other functions, with the
same linear increase in the honest region, but exponential ($(maxf -
minf)\frac{e^{r} - e^{thr}}{e^{1} - e^{thr}}
+ minf$) and logarithmic ($(minf - maxf)\frac{log r}{log(thr)}
+ maxf$) increase in the fraudulent regions. Figure~\ref{fig:conversion}
compares the logistic, exponential and logarithmic functions. It shows that
unlike the exponential and logarithmic functions, the logistic function
exhibits the desired rapid increase for fraud probability values above the
threshold value.
In $\S$~\ref{sec:evaluation} we detail parameter values for the logistic
conversion function,
\subsection{The Hashrate Inference Module}
\label{sec:fraudsys:hashrate}
\noindent
{\bf New device registration}.
When a user registers a new device, the device sends its specs to the online
service that forwards them to FraudSys. FraudSys leverages its
list of profiled devices (see Table~\ref{table:hashrate}) to retrieve the
hashrate of the profiled device with the most similar capabilities. FraudSys
stores the new device along with this initial hashrate estimate under the id of
the user that registers it (see the Puzzle module).
Given this hashrate and the
above time penalty, FraudSys uses Equation~\ref{eq:time:to:hashrate:puzzler}
to compute an initial puzzle difficulty.
\vspace{0.05in}\noindent
{\bf Hashrate correction}.
The initial hashrate estimate of FraudSys may be incorrect. In addition, as
discussed in the System Model, the user may be adversarial, thus attempt to
provide an inaccurate view of the puzzle solving capabilities of his device.
To address these problems, FraudSys employs an adaptive hashrate correction
process. Specifically, an adversary with a more capable device than advertised
(see e.g., Table~\ref{table:hashrate}) will solve the assigned puzzle faster.
The incentive for this is a shorter wait time for his activity to post on the
online service. If this occurs, FraudSys increases its device hashrate
estimate to reflect the observed shorter time required by the device to solve
the puzzle (see Algorithm~\ref{alg:fraudsys}, lines 19-20).
\section{FraudSys Properties}
\label{sec:properties}
\begin{claim}
A fraudster that performs a fraudulent activity with fraud score $r$ from a
device with hashrate $\eta$, is expected to compute $\frac{\eta \times maxf}{1
+ (\frac{maxf-minf}{minf})e^{-k(r-thr)}}$ double hashes.
\label{claim:work}
\end{claim}
\vspace{-12pt}
\begin{proof}
According to Equation~\ref{eq:fraud2penalty}, the time penalty assigned to a
fraudulent activity with score $r$ is $\tau = \frac{maxf}{1 +
(\frac{maxf-minf}{minf})e^{-k(r-thr)}}$. Then,
Equation~\ref{eq:time:to:hashrate:puzzler} ensures that the number of expected
hashes that the device needs to perform to solve the puzzle of
Equation~\ref{eq:puzzle} is $\eta \times \tau$, which concludes the proof.
\end{proof}
\vspace{-5pt}
Let $f$ be the number of fraud workers in the system (i.e., $f =
|\mathcal{F}|$), $\tau$ be the average temporal penalty assigned by FraudSys
to a fraudulent activity, and let $p$ be the expected payout for a single
fraudulent activity. We introduce then the following claim:
\vspace{-5pt}
\begin{claim}
FraudSys is a $(p/\tau, f/\tau)$-fraud preemption\\ system.
\label{claim:fps}
\end{claim}
\vspace{-13pt}
\begin{proof}
The best fit connectivity features of the Fraud Detection module (see
$\S$~\ref{sec:fraudsys:fraud}) enable FraudSys to detect activities performed
from accounts controlled by the same fraudster. This, coupled with an
extension of the $timeout$ concept applied at the fraudster level (see
$\S$~\ref{sec:fraudsys:fraud}) ensures a serialization of fraudster activities.
Then, the average number of fraudulent activities that a fraudster can post per
time unit in FraudSys is $1/\tau$. This implies that, per time unit, the
expected payout of a fraudster is $p/\tau$, and a subject can be the target of
at most $f/\tau$ fraudulent activities. This, according to the definition of
$\S$~\ref{sec:model:problem}, completes the proof.
\end{proof}
\vspace{-5pt}
\subsection{Security Discussion and Limitations}
The FraudSys puzzle not only ties the penalty computation to the user
activity, but also addresses pre-computation, replay and guessing attacks: the
adversary cannot predict the cookie value of its actions, thus cannot
pre-compute puzzles and cannot reuse old cookies. It also offloads significant
work from the FraudSys service, which no longer needs to keep track of puzzle
assignments.
\vspace{0.05in}\noindent
{\bf Device deception}.
An adversary with a specialized puzzle solving device (e.g., AntMiner) will be
assigned puzzles with large difficulty values (see, e.g.,
Table~\ref{table:hashrate}), thus consume the same amount of time as when using
a resource constrained device (e.g., a smartphone). The adversary can exploit
this observation to avoid the implications of Claim~\ref{claim:work}: register
a resource constrained device, but rely on a powerful back-end device to solve
the assigned puzzles faster. The adversary has two options. First, report the
solutions as soon as the back-end device retrieves them. In this case however,
the adversary leaks his true capabilities, as FraudSys will update the
adversary hashrate (Algorithm~\ref{alg:fraudsys}, line 20). Thus,
subsequently, his assigned puzzles will have a significantly higher difficulty
value. In a second strategy, the adversary estimates the time that his
front-end device would take to complete the puzzle, then waits the remaining
penalty time. In this case, the adversary incurs two penalties, the long
wait time and the underutilized back-end device investment.
\vspace{0.05in}\noindent
{\bf Adversary strategies: new user accounts}.
To avoid the implications of Claim~\ref{claim:fps}, the adversary registers new
user accounts. While new accounts are cheap, their freshness and lack of
history will enable the account based features of the Fraud Detection module to
label them as being likely fraudulent. As the adversary reuses such accounts,
the connectivity features start to play a more important role in labeling their
activities as fraudulent. Thus, the adversary has a small usable window of
small penalties for new accounts.
While new honest accounts may also be assigned larger penalties for their first
few activities, they will not affect the user experience: the user can continue
her online activities, while her device solves the assigned puzzle in the
background.
\section{Empirical Evaluation}
\label{sec:evaluation}
\subsection{Datasets}
\label{sec:data}
\begin{figure*}
\centering
\captionsetup[subfigure]{aboveskip=0.1in, belowskip=-0.05in}
\begin{subfigure}[b]{2.25in}
\includegraphics[width=2.25in]{graphs/gplay/honest.fraud.penalty/penalty.evolution.eps}
\caption{}
\label{fig:penalty:evolution}
\end{subfigure}
~
\captionsetup[subfigure]{aboveskip=0.1in, belowskip=-0.05in}
\begin{subfigure}[b]{2.25in}
\includegraphics[width=2.25in]{graphs/gplay/honest.fraud.penalty/fraud.penalty.distribution.5min.24h.eps}
\caption{}
\label{fig:gplay:fraud}
\end{subfigure}
~
\begin{subfigure}[b]{2.25in}
\includegraphics[width=2.25in]{graphs/gplay/honest.fraud.penalty/honest.penalty.distribution.5min.24h.eps}
\caption{}
\label{fig:gplay:honest}
\end{subfigure}
\vspace{-5pt}
\caption{
Stats over the Google Play data when $maxf$ = 24h, $minh$ = 2s, $maxh = minf$ = 5min.
(a) Evolution of average, 1st and 3rd quartile of the penalty imposed on the
$i$-th fraud activity of a fraudster for the same subject. It
shows a steep increase: the average penalty of the first three fraud activities
for a subject sums to 15.34h, while {\bf the average penalty of the $12$th activity
exceeds 24h}.
(b) Distribution of per-fraudster daily penalties, over data from 23
fraudsters: in 1,812 days out of 2,708 days, the penalty assigned to a single
fraudster exceeds 24 hours.
(c) Distribution of penalties assigned to an honest review. Only 14 out of
4,600 honest review instances received a penalty exceeding 5 minutes, but still
below 1 hour.}
\label{fig:gplay}
\vspace{-15pt}
\end{figure*}
We have collected the following datasets of fraudulent and honest behaviors
from Google Play and Facebook.
\vspace{0.05in}\noindent
{\bf Google Play: fraud behavior data}.
We have identified 23 workers in Freelancer, Fiverr and Upwork, with proven
expertise on performing fraud on Google Play apps. We have
contacted these workers and collected the ids of 2,664 Google Play accounts
controlled by them. We have also collected 640 apps heavily reviewed from
those accounts, with between 7 and 3,889 reviews, of which between 2\% and
100\% (median of 50\%) were written from accounts controlled by the workers.
These apps form our gold standard {\it fraud app} dataset. We have also
collected the 23,028 fake reviews written from the 2,664 fraudster controlled
accounts for the 640 apps. Figure~\ref{fig:coreview:fraud} shows the co-review
graph of one of these apps, that received fake reviews from 2 of the identified
23 workers.
\vspace{0.05in}\noindent
{\bf Google Play: honest behavior data}.
We have selected $925$ candidate apps that have been developed by Google
designated ``top developers''. We have removed the apps whose apks
(executables) were flagged as malware by VirusTotal. We have manually
investigated $601$ of the remaining apps, and selected a set of $200$ apps that
(i) have more than $10$ reviews and (ii) were developed by reputable media
outlets (e.g., NBC, PBS) or have an associated business model (e.g., fitness
trackers). We call these the gold standard {\it benign app dataset}.
We have identified 600 reviewers of these 200 benign apps and 140 reviewers of
the 640 fraud apps (see above), such that each has reviewed at least 10 paid
apps, i.e., paid to install the app, then reviewed it, and had at least 5 posts
on their associated Google Plus (social network) accounts. These 740 user
accounts form our gold standard {\it honest user dataset}. We have then
retrieved and manually vetted 854 reviews written by the 600 honest reviewers
for the 200 benign apps, and 207 reviews written by the 140 honest reviewers of
the 640 fraud apps. Each selected review is informative, containing both
positive and negative sentiment statements. We call the resulting dataset, the
{\it honest review dataset}, with 1,061 reviews.
\vspace{0.05in}\noindent
{\bf Facebook Like dataset}.
We have used a subset of the dataset from~\cite{BLLTZ16}, consisting of 15,694
Facebook pages, that each has received at least 30 likes. The pages were liked
from 13,147 user accounts, of which 6,895 are fraudster controlled, and 6,252
are honest. In total, these fraudsters have posted 274,297 fake likes, and the
honest accounts have posted 180,400 honest likes.
\subsection{Device Hashrate Profile}
\begin{table}
\setlength{\tabcolsep}{.16677em}
\centering
\small
\textsf{
\begin{tabular}{l | r | r r r}
\toprule
\textbf{Device} & \textbf{Hashrate} & \textbf{Diff (5s)} & \textbf{ (12hr)} & \textbf{ (7 day)}\\
\midrule
Nexus 4 & 6.53 KH/s& 16.32K& 141.04M& 1.97G\\
Nexus 5 & 13.26 KH/s & 33.15K& 286.41M& 4.00G\\
LG Leon LTE & 10.1 KH/s& 25.25K& 218.16M& 3.05G\\
NVS 295 & 1.7MH/s & 4.25M& 36.72G& 514.08G\\
Server & 80 MH/s & 200M & 1.72T & 24.19T\\
AntMiner & 4.72 TH/s & 11.8T& 101.95P& 1427P\\
\bottomrule
\end{tabular}
}
\caption{Hashrate profiling table for various device types (smartphone, tablet,
PC and Bitcoin miner), along with difficulty values for penalty times of 5s, 12
hours and 7 days.}
\label{table:hashrate}
\vspace{-20pt}
\end{table}
\begin{table}[h]
\vspace{-5pt}
\centering
\textsf{
\begin{tabular}{l | r r | r}
\toprule
\textbf{Strategy} & \textbf{FPR}\% & \textbf{FNR}\% & \textbf{Accuracy}\%\\
\midrule
k-NN & 1.41 & 4.45 & 97.92\\
SVM & 5.8 & 11.3 & 92.40\\
Random Forest & 3.44 & 6.46 & 95.69\\
\bottomrule
\end{tabular}
}
\caption{10-fold cross validation results of supervised learning algorithms in
fraud vs. honest Google Play review classification. k-NN achieves the lowest
FPR and FNR.}
\label{table:fraudsys:algorithms}
\vspace{-15pt}
\end{table}
We have profiled the hashrate of several devices, ranging from smartphones to a
Bitcoin mining hardware (AntMiner S7: ARMv7 CPU, 254 Mb of RAM, 135 BM1385
chips @ 700MHz). Since Bitcoin mining requires capabilities far exceeding
those of smartphones, we have implemented an Android app to evaluate the
hashrate of several Android devices.
Table~\ref{table:hashrate} shows the hashrate values for the profiled devices,
along with the corresponding difficulty ($\Delta$) values for puzzles required
to impose 5 second, 12 hour and 7 day time penalties on such devices. We
observe the significant gap between the hashrate of a smartphone (10-15 KH/s)
and a specialized device (4.72 TH/s). This motivates the need for the puzzles
issued by FraudSys to have different $\Delta$ values for various user devices.
FraudSys maintains a similar table in order to be able to build appropriate
puzzles for newly registered devices.
\subsection{Fraud Penalty Evaluation: Google Play}
\begin{figure*}
\centering
\captionsetup[subfigure]{aboveskip=0.1in, belowskip=-0.05in}
\begin{subfigure}[b]{2.25in}
\includegraphics[width=2.25in]{graphs/facebook/honest.fraud.penalty/fraud.penalty.distribution.5min.24h.eps}
\caption{}
\label{fig:facebook:fraud:penalty}
\end{subfigure}
~
\captionsetup[subfigure]{aboveskip=0.1in, belowskip=-0.05in}
\begin{subfigure}[b]{2.25in}
\includegraphics[width=2.25in]{graphs/facebook/honest.fraud.penalty/honest.penalty.distribution.5min.24h.eps}
\caption{}
\label{fig:facebook:honest:penalty}
\end{subfigure}
~
\begin{subfigure}[b]{2.25in}
\includegraphics[width=2.25in]{graphs/payout/compare.fraud.payout.eps}
\caption{}
\label{fig:payouts}
\end{subfigure}
\vspace{-5pt}
\caption{
(a) Penalty distribution for the fake Facebook likes. 84\% of the likes
received a penalty that exceeds 12 hours, and the average fake like penalty is
19.32 hours.
(b) Penalty distribution for the honest Facebook likes. 82.97\% of the honest
likes are assigned a penalty of under 5 min. The maximum penalty assigned
to an honest like is 70 minutes.
(c) Comparison of daily payouts provided by Bitcoin mining, writing fake
reviews in Google Play and posting fake likes in Facebook, under FraudSys.
{\bf Fraud does not pay off under FraudSys}: the fraud payout is less than half
the Bitcoin mining payout.}
\label{fig:facebook}
\vspace{-5pt}
\end{figure*}
\noindent
{\bf Supervised learning algorithm choice}.
We first used 10 fold cross-validation to evaluate the ability of the Fraud
Detection module to correctly classify the 23,028 fraudulent vs. 1,061 honest
reviews of the Google Play dataset previously described.
Table~\ref{table:fraudsys:algorithms} shows the false positive (FPR) and
negative (FNR) rates, as well as the accuracy achieved by the top 3 performing
supervised algorithms. k-NN has the lowest FPR and FNR, for an accuracy of
97.92\%. Thus, in the following experiments we use only k-NN.
\vspace{0.05in}\noindent
{\bf Parameter evaluation}.
We have used the fraud and honest review datasets described earlier, to compute
the temporal penalties imposed by FraudSys on fraudsters and honest users. We
have performed the following experiments. In each experiment, we use the data
of 22 fraud workers and 200 randomly chosen honest reviews (out of 1,061) to
train the supervised learning algorithm (k-NN) then test the model on the data
of the remaining fraud worker and on the remaining 861 honest reviews. Thus,
we have performed 23 experiments, one for each worker.
We set the $maxf$ parameter such that the average daily payout of a fraudster
is below the average Bitcoin mining payout with a last generation AntMiner
device. Thus, this ensures that even such a powerful adversary has more
incentive to do Bitcoin mining instead of search rank fraud. Specifically, the
above AntMiner's current (Jan. 2017) average daily payout is 0.0037 BTC. At
the current BTC to USD rate, this means \$3.67 per day~\footnote{Historically
speaking, the BTC to USD rate is increasing. The next generation AntMiner
coming up this year is expected to be 3 times more capable.}. In addition, we
have experimented with $maxf$ values ranging from 12 to 48 hours. The average
penalty assigned by FraudSys to a fraudulent review is 8.01 hours when
$maxf$=12h, 15.34h when $maxf$=24h, and 29.33h when $maxf$=48h.
Figure~\ref{fig:penalty:evolution} shows the median, first and third quartiles
for the time penalty (in hours) imposed on the $i$-th fraudulent activity
performed by a fraudster for a subject, when $maxf$= 24h: the 12th fake
activity receives a median penalty of 24h.
Thus, we set $maxf$=24h, which is sufficient for Google Play reviews: A
fraudster would be able to post on average less than 2 fake reviews per day,
thus, even with a reward of \$2 per fraud activity (see
Figure~\ref{fig:fraudsites}), achieve a payout of around \$3.15 per day, below
the Bitcoin mining payout. In addition, we have set $minh$ = 2s.
Figure~\ref{fig:timeline} shows the penalty timelines of two workers when
$minh$ = 2s, $maxh = minf$ = 5 min, $maxf$ = 24 hours, $thr = 0.5$, and $k=30$
(for a steep increase of time penalty with fraud score). We note that a $maxh$
= 5 min is not excessive: this penalty is not imposed on the user, but on his
device. The user experience remains the same in the online service.
Each vertical bar shows the daily temporal penalty assigned to a single worker,
over reviews posted from multiple accounts. The maximum daily penalty of the
two workers is 1,247 hours and 3,079 hours respectively. We observe that each
worker has many days with a daily penalty exceeding 24 hours.
Figure~\ref{fig:gplay:fraud} shows for $maxh = minf$ = 5min, the overall
distribution of daily penalties assigned by FraudSys, over all the 23 fraud
workers, in the above experiment. It shows that during most of the active days,
fraud workers are assigned a daily penalty exceeding 24 hours.
Figure~\ref{fig:gplay:honest} (also for $maxh = minf$ = 5min) shows the
distribution of per-review penalty assigned by FraudSys to honest reviews,
shown over 4,600 (23 $\times$ 200) honest reviews. Irrespective of the $maxh$
value, only 14 honest reviews were classified as fraudulent, but assigned a
penalty below 1 hour. We observed minimal changes in the distribution of
penalties of fraudulent reviews when $maxh = minf$ ranges from 5 to 15 minutes.
\subsection{Fraud Penalty Evaluation: Facebook}
We have performed a similar parameter analysis using the Facebook ``like''
dataset. Since this dataset lacks information about the fraudsters who control
the accounts that posted fake likes, we focus on the penalties assigned by
FraudSys to fake and honest likes.
Figure~\ref{fig:facebook:fraud:penalty} shows the distribution of penalties
assigned to fake likes and Figure~\ref{fig:facebook:honest:penalty} shows the
distribution of the honest likes. Compared to the results over the Google Play
data, we observe a higher FPR, i.e., more honest likes with fraud level
penalties. We posit that this is due to the fewer features that we can extract
for the Facebook likes, as, unlike for Google Play reviews, we lack the time of
the activity. Specifically, absence of like sequence information enables us to
only extract features based on the last ``snapshot'' of the page, and not the
current page snapshot when the like was posted.
However, 82.97\% of the honest likes receive a penalty of under 5 mins and the
maximum penalty assigned to an honest review is 70 mins. In addition, 84\% of
the fake likes receive a penalty that exceeds 12 hours, and the average penalty
for a fake like is 19.32 hours. Figure~\ref{fig:payouts} compares the daily
payouts received by an AntMiner equipped fraudster who writes fake reviews in
Google Play (at \$1 per fake review), posts fake likes (at \$2 per fake like),
or honestly uses his device to mine Bitcoins. It shows that under FraudSys,
fraud doesn't pay off: the Bitcoin mining payout is more than double the fraud
payout for either fake reviews or likes.
\section{Conclusion}
We have introduced the concept of real-time fraud preemption systems, named as the FraudSys, that seek
to restrict the profitability and impact of fraud in online systems. We
propose and develop stateless, verifiable computational puzzles, that impose
minimal overheads, but enable their efficient verification. We have developed
a graph based, real-time algorithm to assign fraud scores to user activities
and mechanisms to convert scores to puzzle difficulty values. We used data
collected from Google Play and Facebook to show that our solutions impose
significant penalties on fraudsters, and make fraud less productive than
Bitcoin mining.
\section{Acknowledgments}
This research was supported in part by NSF grants CNS-1527153, SES-1450619,
CNS-1422215, IUSE-1525601, and Samsung GRO.
\bibliographystyle{abbrv}
| 0e37657040e9cd398b519d50b9d1dbbbba577643 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{S1}
In the last decades, various investigations have shown that black hole physics plays a vital role in our understanding of gravity on both macroscopic and microscopic scales. Aspects of black holes such as horizon's global structure, Hawking radiation, thermodynamical properties and black hole information are key concepts related to the fundamental structure of space-time. For recent review see Ref. (\cite{carlip} and references therein). Examples of these new concepts and constructions are the Holographic Principle and its realization in AdS/CFT correspondence \cite{thooft,susskind,maldacena}, Ashtekar approach and loop quantum gravity \cite{rovelli}, Jacobson's thermodynamic formulation \cite{jacobson} as well as Verlinde's ENTOPIC approach to gravity \cite{verlinde}. In general, seeking for new black hole solutions is extremely relevant to set up any relativistic theory of gravity.
Recently, a lot of attention has been devoted to the teleparallel equivalent of General Relativity (TEGR) \cite{Ea,Ea1,Ea2,HS9}. This theory is an equivalent to General Relativity (GR) and can be generalized into a class of theories called $f(T)$ gravity \cite{FF7,FF8,BF9,CCMS}. TEGR and its generalizations are based on Weitzenb\"{o}ck connection instead of Levi-Civita connection and use a vielbein field, ${e^a}_{\mu}$, as a fundamental field variable, instead of the metric \cite{N10}. TEGR is equivalent to GR since its field equations as well as its test particle equation of motion are equivalent to that of GR. In TEGR and its generalizations the gravitational field is represented by torsion instead of curvature because the later vanishes in these theories. TEGR can be constructed as a gauge theory of the translation group, where the vielbein is the gauge field and the torsion is the field strength \cite{APVB,AP}. This theory is invariant under diffeomorphism and local Lorentz transformations \cite{M1,Mj}.
There is a considerable interest in generalizing TEGR in cosmological through adding higher-torsion terms. These terms can mimic a dark energy fluid in Friedman-Robertson-Walker (FRW) cosmologies without introducing exotic equations of state \cite{bamba,CC23,CC,FF11,Ngrg02,GSV,WNE1,RHTMM,GG2,JRH,GG3,CC8,CC9,CC10,AN,FLR,BFG,NNC,Km,KS,BMT,DWM}. Teleparallel gravity with higher-torsion terms depending only on the torsion scalar, $T$, are known as $f(T)$ theories. A part from these cosmological applications some interesting black holes solutions were discovered and studied in the context of $f(T)$ theories. One of the most studied models of these theories is $f(T)=T+\beta \,T^2$, where spherically symmetric approximate solution was found and analyzed \cite{SOE,Ncpl,IS,IS1,XD}, furthermore, the higher-torsion correction term was constrained in this theory using the Solar System data.
Exact solutions in higher curvature/ torsion theories are not always easy to find also in the framework of $f(T)$ \cite{Nprd3,CGSV13, Ngrg3}. Black holes with cosmological constant present some attractive features: for example various horizon topologies appear in contrast to the asymptotically flat case. In these cases, black hole horizons can be spherical, hyperbolic or planar: these features can lead to tori or cylindrical structures depending on the global identifications applied \cite{Mrb}. Asymptotically de Sitter and Anti-de-Sitter charged black holes and rotating black holes in four and higher dimensions have been intensively studied in the context of AdS/CFT and dS/CFT correspondences; see for example Refs. \cite{lemos,Aa02,Aa06,Aa07,Aa08,gubser}.
In this paper, we are going to present new charged asymptotically Anti-de-Sitter black hole solutions with flat horizons in $f(T)$ theories, where $f(T)=T+\beta T^2$, $\beta <0$, and $D \geq 4$. First of all, these black holes solutions have electric potentials which contains a monopole as well as a quadrupole term. In order to get an asymptotically Anti-de-Sitter (AdS) solution, we are forced to relate the electric monopole momentum and the quadrupole momentum of these solutions, therefore, these two momenta are inseparable. The second interesting feature is that, although this black hole solution is singular at the origin, $r=0$, its singularity is clearly milder than asymptotically Anti-de-Sitter charged black solution in GR or TEGR. For example in {\it D} dimensions, the Kretschmann scalar, derived from the Ricci tensor square, and the Ricci scalar are $K=R_{\mu \nu}R^{\mu \nu} \sim r^{-2(D-2)}$, $R\sim r^{-(D-2)}$ in contrast with, the known solutions in Einstein-Maxwell theory in both GR and TEGR which have $K =R_{\mu \nu}R^{\mu \nu}\sim r^{-2D}$ and $R \sim r^{-D}$. Furthermore, although these asymptotically AdS charged solutions have different $g_{tt}$ and $g^{rr}$ component for the metric, they have coinciding Killing and event horizons.
The outline of the paper is the following. In Section \ref{S1.1}, we briefly review the TEGR formalism through tensors definitions and the field equations, then we introduce the field equations of $f(T)$ gravity. In Section \ref{S3}, a vielbein field having flat horizon in $D$ dimensions is applied to the field equations of $f(T)$ gravity to obtain a general neutral black hole solution in {\it D} dimensions, which is asymptotically AdS. In Section \ref{S9}, a cylindrically symmetric vielbein is applied to the Einstein-Maxwell field equations in $f(T)$ gravity. We show how this solution can be reduced to an exact charged static black hole in AdS space. The interesting feature of this black hole is that it has monopole and the quadrupole momenta. This feature extends results presented in \cite{CGSV13}. In Section \ref{S5}, some relevant physical features of these black holes are reported. In Section \ref{S11}, we discuss the thermodynamics of the black holes presented in \ref{S9}. Finally, concluding remarks are reported in Section \ref{S12}.
\section{The TEGR geometry and $f(T)$ gravity}\label{S1.1}
TEGR is described by the pair $\{{\cal M},~e_{i}\}$, with $\cal M$ being a $D$-dimensional manifold and $e_{i}$ ($i=1,2,\cdots, D$) are vectors globally defined on the manifold $\cal M$. Vector fields $e_{i}$ are the parallel vectors. In $D$-dimension, the parallel vectors are named the {\it vielbein} fields and the derivative of the contravariant vielbein fields have to vanish
\begin{equation}\label{q1}
D_{\mu} {e_i}^\nu:=\partial_{\mu}
{e_i}^\nu+{\Gamma^\nu}_{\lambda \mu} {e_i}^\lambda= 0,
\end{equation}
where the differentiation is with respect to the Weitzenb\"{o}ck connection and $\partial_{\mu}:=\frac{\partial}{\partial x^{\mu}}$ and ${\Gamma^\nu}_{\lambda \mu}$ is a non-symmetric affine connection defined as \cite{Wr}
\begin{equation}\label{q2}
{\it {\Gamma^\lambda}_{\mu \nu} := {e_i}^\lambda~ \partial_\nu e^{i}{_{\mu}}}.
\end{equation}
In this geometry, the metric is given by
\begin{equation}\label{q3}
{\it g_{\mu \nu} := \eta_{i j} {e^i}_\mu {e^j}_\nu,}
\end{equation}
where $\eta_{i j}=(+,-,-,- \cdots)$ is {\it D} dimensional Minkowskian metric. The metricity condition is fulfilled as a consequence of Eq. (\ref{q1}). The torsion, ${T^\alpha}_{\mu \nu}$, and the contortion, $K^{\mu \nu}{}_\alpha$, tensors field are defined as
\begin{eqnarray}
\nonumber {T^\alpha}_{\mu \nu} & := &
{\Gamma^\alpha}_{\nu \mu}-{\Gamma^\alpha}_{\mu \nu} ={e_i}^\alpha
\left(\partial_\mu{e^i}_\nu-\partial_\nu{e^i}_\mu\right),\\
{K^{\mu \nu}}_\alpha & := &
-\frac{1}{2}\left({T^{\mu \nu}}_\alpha-{T^{\nu
\mu}}_\alpha-{T_\alpha}^{\mu \nu}\right). \label{q4}
\end{eqnarray}
We introduce the teleparallel torsion scalar of TEGR theory which is
\begin{equation}\label{Tor_sc}
T := {T^\alpha}_{\mu \nu} {S_\alpha}^{\mu \nu},
\end{equation}
where the skew symmetric tensor ${S_\alpha}^{\mu \nu}$ is defined as
\begin{equation}\label{q5}
{S_\alpha}^{\mu \nu} := \frac{1}{2}\left({K^{\mu\nu}}_\alpha+\delta^\mu_\alpha{T^{\beta
\nu}}_\beta-\delta^\nu_\alpha{T^{\beta \mu}}_\beta\right).
\end{equation}
Using Eq. (\ref{q4}) it is possible to re-express Eq. (\ref{q2}) as
\begin{equation}\label{contortion}
{\Gamma^\mu}_{\nu \rho }=\left \{_{\nu \rho}^\mu\right\}+{K^{\mu}}_{\nu \rho},
\end{equation}
where the first term is the Levi-Civita affine connection of GR while the second one is derived from the contortion.
It is natural to extend TEGR theory including higher torsion terms defining a Lagrangian $f(T)$ where $f$ is a function of the torsion invariant $T$:
\begin{equation}\label{q7}
{\cal L}=\frac{1}{2\kappa}\int |e|(f(T)-2\Lambda)~d^{D}x+\int |e|{\cal L}_{ em}~d^{D}x,
\end{equation}
where $\kappa$ is a dimensional constant defined as $\kappa =2(D-3)\Omega_{D-1} G_D$, with $G_D$ being the Newton gravitational constant in $D$-dimensions and
$\Omega_{D-1}$ the volume of $(D-1)$-dimensional unit sphere given by
the expression $\Omega_{D-1} = \frac{2\pi^{(D-1)/2}}{\Gamma((D-1)/2)}$, with the $\Gamma$-function
of the argument that depends on the dimension of spacetime\footnote{When $D = 4$, one can show that $2(D-3)\Omega_{D-1} = 8 \pi$.}. In Eq. (\ref{q7}), $ |e|=\sqrt{-g}=\det\left({e^a}_\mu\right)$ and ${\cal L}_{
em}=-\frac{1}{2}{ F}\wedge ^{\star}{F}$ is the Maxwell Lagrangian,
with $F = dA$, with $A=A_{\mu}dx^\mu$, is the electromagnetic
potential 1-form \cite{CGSV13}. The variation of Eq. (\ref{q7}) with respect to the vielbein field ${e^i}_\mu$ and the vector potential $A_{\mu}$ gives the following field equations \cite{BF9}
\begin{eqnarray}\label{q8}
& &{S_\mu}^{\rho \nu} \partial_{\rho} T f_{TT}+\left[e^{-1}{e^i}_\mu\partial_\rho\left(e{e_i}^\alpha
{S_\alpha}^{\rho \nu}\right)-{T^\alpha}_{\lambda \mu}{S_\alpha}^{\nu \lambda}\right]f_T
-\frac{f-2\Lambda}{4}\delta^\nu_\mu +\frac{1}{2}\kappa{{{\cal
T}^{{}^{{}^{^{}{\!\!\!\!\scriptstyle{em}}}}}}}^\nu_\mu=H^\nu{}_\mu \equiv0,\nonumber\\
&&\partial_\nu \left( \sqrt{-g} F^{\mu \nu} \right)=0\; ,
\end{eqnarray}
where $f := f(T)$, \ \ $f_{T}:=\frac{\partial f(T)}{\partial T}$, \ \ $f_{TT}:=\frac{\partial^2 f(T)}{\partial T^2}$ and ${{{\cal
T}^{{}^{{}^{^{}{\!\!\!\!\scriptstyle{em}}}}}}}^\nu_\mu$ is the
energy-momentum tensor of the electromagnetic field defined as
\[
{{{\cal
T}^{{}^{{}^{^{}{\!\!\!\!\scriptstyle{em}}}}}}}^\nu_\mu=F_{\mu \alpha}F^{\nu \alpha}-\frac{1}{4} \delta_\mu{}^\nu F_{\alpha \beta}F^{\alpha \beta}.\] Eq. (9) can be re-expressed as
\be \partial_\nu \Biggl[e{S}^{a \rho \nu} f_T\Biggr]=\kappa e
{e^a}_\mu \Biggl[t^{\rho \mu}+{{{\cal
T}^{{}^{{}^{^{}{\!\!\!\!\scriptstyle{em}}}}}}}^{\rho \mu}\Biggr],\ee
where $t^{\nu \mu}$ has the form \be t^{\nu
\mu}=\frac{1}{\kappa}\Biggl[4f_T {S^\alpha}^{\nu
\lambda}{T_{\alpha \lambda}}^{\mu}-g^{\nu \mu} f\Biggr].\ee Since ${S}^{a \nu \lambda}$ is a skew-symmetric tensor in the last pairs, then \be
\partial_\mu \partial_\nu\left[e{S}^{a \mu \nu} f_T\right]=0, \quad
\textrm{which \quad yields}
\quad \partial_\mu\left[e\left(t^{a \mu}+{{{\cal
T}^{{}^{{}^{^{}{\!\!\!\!\scriptstyle{em}}}}}}}^{a
\mu}\right)\right]=0. \ee Eq. (12) yields \be
\frac{d}{dt}\int_V d^{(D-1)}x \ e \ {e^a}_\mu \left(t^{0 \mu}+{{{\cal
T}^{{}^{{}^{^{}{\!\!\!\!\scriptstyle{em}}}}}}}^{0
\mu}\right)+ \oint_\Sigma \left[e \ {e^a}_\mu \ \left(t^{j
\mu}+{{{\cal T}^{{}^{{}^{^{}{\!\!\!\!\scriptstyle{em}}}}}}}^{j
\mu}\right)\right]=0.\ee Eq. (13) is the conservation law of the
energy-momentum tensor ${{{\cal
T}^{{}^{{}^{^{}{\!\!\!\!\scriptstyle{em}}}}}}}^{\lambda
\mu}$ and the quantity $t^{\lambda \mu}$. Thus, we can consider $t^{\lambda \mu}$
to describe the gravitational energy-momentum tensor in
$f(T)$ gravity \cite{US3}. Therefore,
the energy-momentum of $f(T)$ theory contained
in a (D-1)-dimensional volume $V$ takes the form \be P^a=\int_V d^{(D-1)}x
\ e \ {e^a}_\mu \left(t^{0 \mu}+{{{\cal
T}^{{}^{{}^{^{}{\!\!\!\!\scriptstyle{em}}}}}}}^{0
\mu}\right)=\frac{1}{\kappa}\int_V d^{(D-1)}x \partial_\nu\left[e{S}^{a 0
\nu} f_T\right].\ee From Eq. (14), we can return to the standard TEGR as soon as $f(T)=T$ \cite{MDTC}. Eq. (13) represents the conserved four-momentum for any asymptotic flat solution: in this work we discuss a class of asymptotically AdS solution. Therefore, it is natural to calculate conserved quantities relative to a pure AdS space. Otherwise, the conserved quantities is plagued by infinities because of the asymptotic behavior of the asymptotically AdS solution. For example the total mass/energy of an AdS black hole measured by a stationary observer at infinity might be understood as the difference in energy between the AdS black hole solution and AdS space itself. Therefore, in calculating conserved quantities, it is natural to subtract off the contributions coming from pure AdS space in the above conserved quantities of Eq. (13); this is why we have subscript "$r$" which stands for the regularized value of the physical quantity.
\[
\frac{d}{dt}\int_V d^{(D-1)}x \ e \ {e^a}_\mu \left(t_{r}^{0 \mu}+{{{\cal
T}_{r}^{{}^{{}^{^{}{\!\!\!\!\scriptstyle{em}}}}}}}^{0
\mu}\right)+ \oint_\Sigma \left[e \ {e^a}_\mu \ \left(t_{r}^{j
\mu}+{{{\cal T}_{r}^{{}^{{}^{^{}{\!\!\!\!\scriptstyle{em}}}}}}}^{j
\mu}\right)\right]=0.\]
\newpage
\section{Asymptotically AdS black holes}\label{S3}
We apply the field equations of extended teleparallel gravity $f(T)$, Eq. (\ref{q8}), to the flat $D$-dimensional spacetime horizon, which directly gives rise to the vielbein written in cylindrical coordinate ($t$, $r$, $\phi_1$, $\phi_2$,$\cdots$ $\phi_{D-2}$) as follows \cite{CGSV13}:
\begin{equation}\label{tetrad}
\hspace{-0.3cm}\begin{tabular}{l}
$\left({e_{i}}^{\mu}\right)=\left( \sqrt{N(r)}, \; \frac{1}{\sqrt{N_1(r)}}, \; r, \; r, \; r\;\cdots \right)$
\end{tabular}
\end{equation}
where $N(r)$ and $N_1(r)$ are two unknown functions of $r$. Substituting from Eq. (\ref{tetrad}) into Eq. (\ref{Tor_sc}), we evaluate the torsion scalar as\footnote{For abbreviation we will write $N(r)\equiv N$, \ \ $N_1(r)\equiv N_1$, \ \ $N'\equiv\frac{dN}{dr}$ and $N'_1\equiv\frac{dN_1}{dr}$ .}
\begin{equation}\label{df}
T=2(D-2)\frac{N'N_1}{rN}+(D-2)(D-3)\frac{N_1}{r^2}.
\end{equation}
Applying Eq. (\ref{tetrad}) to the field equation (\ref{q8}) when ${{{\cal
T}^{{}^{{}^{^{}{\!\!\!\!\scriptstyle{em}}}}}}}^\nu_\mu=0$ we get the following non-vanishing components:
\begin{eqnarray}\label{df1}
& & H^r{}_r= 2Tf_T+2\Lambda-f=0,\nonumber\\
& & H^{\phi_1}{}_{\phi_1}= H^{\phi_2}{}_{\phi_2}=\cdots \cdots =H^{\phi_{D-2}}{}_{\phi_{D-2}}= \frac{f_{TT} [r^2T+(D-2)(D-3)N_1]T'}{(D-2)r}+\frac{f_T}{2r^2{N}^2}\Biggl\{2r^2NN_1N''\nonumber\\
& &-r^2N_1N'^2+2(2D-5)rNN_1N' +r^2NN'N'_1+2(D-3)N^2[2(D-3)N_1+rN'_1]\Biggr\}-f+2\Lambda=0, \nonumber\\
& & H^t{}_t=\frac{2(D-2)N_1f_{TT} T'}{r}+\frac{(D-2)f_T}{r^2N}\Biggl\{2(D-3)NN_1+rN_1N'+rNN'_1\Biggr\}-f+2\Lambda=0.\nonumber\\
& &
\end{eqnarray}
Now we are going to find a general solution to the above differential equations using a specific form of $f(T)$, i.e., $f(T)=T+\beta T^2$.
For this specific form of $f(T)$, Eqs. (\ref{df1}) take the form
\begin{eqnarray}
& & H^r{}_r= T+3\beta T^2+2\Lambda=0,\nonumber\\
& & H^{\phi_1}{}_{\phi_1}= H^{\phi_2}{}_{\phi_2}=\cdots \cdots H^{\phi_{D-2}}{}_{\phi_{D-2}}= \frac{2\beta[r^2T+(D-2)(D-3)N_1]T'}{(D-2)r}+\frac{(1+2\beta T)}{2r^2{N}^2}\Biggl\{2r^2NN_1N''\nonumber\\
& &-r^2N_1N'^2+2(2D-5)rNN_1N'+r^2NN'N'_1+2(D-3)N^2[2(D-3)N_1+rN'_1]\Biggr\} -T-\beta T^2+2\Lambda=0,\nonumber\\
& &H^t{}_t= \frac{4(D-2)\beta N_1T'}{r}+\frac{(1+2\beta T)(D-2)}{r^2N}\Biggl\{2(D-3)NN_1+rN_1N'+rNN'_1\Biggr\}-T-\beta T^2+2\Lambda=0.\nonumber\\
&&
\end{eqnarray}
A general $D$-dimension solution of Eq. (18) is
\begin{eqnarray}
& & N(r)=-\frac{r^2}{6(D-1)(D-2)\beta}-\frac{m}{r^{D-3}}, \qquad \quad
N_1(r)=\frac{1}{N(r)}, \nonumber\\
\end{eqnarray}
where $m$ is the mass parameter and we choose $\Lambda=\frac{1}{24\beta} $ to get a unique solution\footnote{The cosmological constant for these solutions has two values $ \frac{-1 \pm \sqrt{1-24\alpha \, \Lambda}}{12 \alpha}$}.
\section{A new charged AdS black hole solution}\label{S9}
Using the {\it D} dimensional spacetime of Eq. (\ref{tetrad}) with a vector potential $A = V(r) dt$, the field equations have the following non-vanishing components:
\begin{eqnarray}
& & H^r{}_r= 2Tf_T+2\Lambda-f+\frac{2V'^2(r)N_1}{N}=0,\nonumber\\
& & H^{\phi_1}{}_{\phi_1}= H^{\phi_2}{}_{\phi_2}=\cdots \cdots =H^{\phi_{D-2}}{}_{\phi_{D-2}}= \frac{f_{TT} [r^2T+(D-2)(D-3)N_1]T'}{(D-2)r}+\frac{f_T}{2r^2{N}^2}\Biggl\{2r^2NN_1N''\nonumber\\
& & -r^2N_1N'^2+4(D-3)^2N^2N_1+2(2D-5)rNN_1N'+r^2NN'N'_1+2(D-3)rN^2N'_1\Biggr\}\nonumber\\
& & -f+2\Lambda-\frac{2V'^2(r)N_1}{N}=0, \nonumber\\
& & H^t{}_t= \frac{2(D-2)N_1f_{TT} T'}{r}+\frac{(D-2)f_T[2(D-3)NN_1+rN_1N'+rNN'_1]}{r^2N}-f+2\Lambda+\frac{2V'^2(r)N_1}{N}=0,\nonumber\\
& &
\end{eqnarray}
where $V'=\frac{dV}{dr}$ and as before we set $\Lambda=\frac{1}{24\beta} $.
The general {\it D}-dimensional solution of the above differential equations takes the form
\begin{eqnarray}
& & N(r)=\frac{r^2(D-3)^4c_2{}^4}{(D-1)(D-2)(2D-5)^2c_3{}^2}+\frac{c_1}{r^{D-3}}+\frac{3(D-3)c_2{}^2}{(D-2)r^{2(D-3)}}+\frac{2(D-3)c_2c_3}{(D-2)r^{3D-8}},\nonumber\\
& &
N_1(r)=\frac{1}{f(r)N(r)}, \qquad \textrm {where} \qquad f(r)=-\frac{(2D-5)^2c_3{}^2\left[1+\frac{(2D-5)c_3}{c_2(D-3)r^{D-2}}\right]^2}{6\beta (D-3)^4c_2{}^4},\nonumber\\
& & V(r)=\frac{c_2}{r^{D-3}}+\frac{c_3}{r^{2D-5}}.\end{eqnarray}
To get an asymptotically AdS or dS solution we have to set \begin{eqnarray} & & c_3{}^2=\frac{-6(D-3)^4c_2{}^4 \beta}{(2D-5)^2},\end{eqnarray}
otherwise the solution have no clear asymptotic behavior. As a result, the monopole momentum is related to the quadrupole momentum of the solution. In this case, one gets
\begin{eqnarray} & &N(r)=\frac{r^2}{6(D-1)(D-2)\left|\beta\right|}-\frac{m}{r^{D-3}}+\frac{3(D-3)q{}^2}{(D-2)r^{2(D-3)}}+\frac{2\sqrt{6\left|\beta\right|}(D-3)^3q{}^3}{(2D-5)(D-2)r^{3D-8}},\nonumber\\
& &
N_1(r)= \frac{1}{f(r)N(r)}, \qquad f(r)=\left[1+\frac{(D-3)q\sqrt{6\left|\beta\right|}}{r^{D-2}}\right]^2,\qquad V(r)=\frac{q}{r^{D-3}}+\frac{(D-3)^2q{}^2\sqrt{6\left|\beta\right|}}{(2D-5)r^{2D-5}},\nonumber\\
& &\end{eqnarray}
where we set $c_1=-m$, and $c_2=q$, which is the monopole momentum. The quadrupole moment is $Q= \frac{(D-3)^2q{}^2\sqrt{6 \left| \beta \right|}}{(2D-5)}$. As one can notice Eq. (22) tells us that $\beta$ have a negative value otherwise we get an unphysical solution.\\
It is important here to comment on the charged black solutions obtained in \cite{CGSV13}. Note that differential equations (20) are different from those derived in \cite{CGSV13} from many aspects:\vspace{0.2cm}\\
i) The disappearance of the term $f(T)$ from Eqs. (5$\cdot$13) and (5$\cdot$14).\vspace{0.2cm}\\
ii) The terms of the charges in Eqs. (5$\cdot$12), (5$\cdot$13) and (5$\cdot$14) are different from the present Eqs. (20).
Furthermore, the solution (23) generalizes the solution in \cite{CGSV13}. As it is clear from Eq. (23), the potential $V(r)$ depends on a monopole and quadrupole momenta and by setting $q=0$ both momenta vanish and we get a non-charged solution. On the other hand, in \cite{CGSV13}, the charged solution depends only on the monopole.
\section{The main features of the solution}\label{S5}
Let us now discuss some relevant features of the charged solution presented in the previous Section.\\
The metric of the vielbein (23) takes the form
\ba \A \A ds{}^2=\Biggl[r^2\Lambda_{ef}-\frac{m}{r^{D-3}}+\frac{3(D-3)q^2}{(D-2)r^{2(D-3)}}+\frac{2\sqrt{6\left|\beta\right|}(D-3)^3q^3}{(2D-5)(D-2)r^{3D-8}}
\Biggr]dt^2\nonumber\\
-\A\A\frac{dr^2}{\left[1+\frac{(D-3)q\sqrt{6\left|\beta\right|}}{r^{D-2}}\right]^2\Biggr[r^2\Lambda_{ef}-\frac{m}{r^{D-3}}+\frac{3(D-3)q^2}{(D-2)r^{2(D-3)}}
+\frac{2\sqrt{6\left|\beta\right|}(D-3)^3q^3}{(2D-5)(D-2)r^{3D-8}}\Biggr]} -r^2\sum_{i=1}^{D-2}d\phi^2_i,\nonumber\\
\A\A\ea
where $\Lambda_{ef}= \frac{1}{6(D-1)(D-2)\left|\beta\right|}$. Eq. (23) shows clearly that the metric of the charged solution is asymptotically AdS. Notice that there is no corresponding TEGR solution upon taking the limit $\beta \rightarrow 0$, which means this charged solution has no analogue in GR or TEGR. By taking the limit $q \rightarrow 0$, we get the AdS non-charged black holes presented in section (\ref{S3}). Notice that although these asymptotically AdS charged solutions have different $g_{tt}$ and $g^{rr}$ component for the metric, they have coinciding Killing and event horizons.
\underline{Singularity:}\vspace{0.2cm}\\
Here we derive physical singularities by calculating curvature and torsion invariants. Since the function $f(r)$ could have roots (when $q<0$), which we call $r_n$, one has to consider the behavior of curvature invariants close to these roots. By calculating the Kretschmann scalar as function of the radial coordinate $r$, we found that the scalar is well behaved. Now calculating the various curvature and torsion invariants, one obtains
\begin{eqnarray} \A \A R^{\mu \nu \lambda \rho}R_{\mu \nu \lambda \rho}= F_1(r)\,\left(\frac{1}{r^{2(D-2)}}\right), \qquad R^{\mu \nu}R_{\mu \nu}= F_2(r)\,\left(\frac{1}{r^{2(D-2)}}\right), \nonu
\A \A R= F_3(r)\,\left(\frac{1}{r^{(D-2)}}\right), \qquad T^{\mu \nu \lambda}T_{\mu \nu \lambda} = F_4(r)\, \left(\frac{1}{r^{(D-2)}}\right), \qquad T^\mu T_\mu = F_5(r)\, \left(\frac{1}{r^{(D-2)}}\right),
\nonu
\A \A T(r)=F_6(r)\, \left(\frac{1}{r^{(D-2)}}\right),\end{eqnarray}
where $R^{\mu \nu \lambda \rho}R_{\mu \nu \lambda \rho}$, $R^{\mu \nu}R_{\mu \nu }$, $R$, $T^{\mu \nu \lambda}T_{\mu \nu \lambda}$ $T^{\mu }T_{\mu }$ and $T$ are the Kretschmann scalar, the Ricci tensor square, the Ricci scalar, the torsion tensor square, the torsion square vector and the torsion scalar: $F_i(r)$ are polynomial functions in $r$.
The above invariants show that:\vspace{0.1cm}\\
a) There is a singularity at $r=0$ which is a curvature singularity. \vspace{0.1cm}\\
b) In the charged case, the torsion scalar has the form \be T=\frac{r^{(D-2)}-2q^2(D-3){\sqrt{6\left|\beta\right|}}}{6\left|\beta\right| r^{(D-2)}},\ee which shows that the scalar torsion has singularity at $r=0$. Close to $r=0$, the behavior of the Kretschmann scalar, the Ricci tensor square and the Ricci scalar for the charged solution is given by $K=R_{\mu \nu}R^{\mu \nu} \sim r^{-2(D-2)}$, $R=T^{\alpha \beta \gamma}T_{\alpha \beta \gamma}=T^{\alpha}T_{\alpha}=T\sim r^{-(D-2)}$ in contrast with the solutions of the Einstein-Maxwell theory in both GR and TEGR which have $K =R_{\mu \nu}R^{\mu \nu}\sim r^{-2D}$ and $R=T^{\alpha \beta \gamma}T_{\alpha \beta \gamma}=T^{\alpha}T_{\alpha}=T \sim r^{-D}$. This shows clearly that the singularity is much milder than the one obtained in GR and TEGR for the charged case. This result raises the question if these singularities are {\it weak singularities}, according to Tipler and Krolak \cite{tipler,krolak}, and if it is possible to extend geodesics beyond these regions. This topic will be discussed in forthcoming studies.
\underline{Energy:}\vspace{0.2cm}\\
Let us now calculate the energy related to the charged black holes given by Eqs. (23). Using Eq. (14), it is possible to derive the components of energy in the solution (19). We get:
\be S^{001}=\frac{(D-2)}{2r}. \ee
Substituting Eq. (27) into Eq. (14), we get the energy in the form
\be P^0=E= \frac{(D-2)\Omega_{D-2}[m-\Lambda_{ef}\,r^{(D-1)}]}{2\kappa }
=\frac{(D-2)[m-\Lambda_{ef}\,r^{(D-1)}]}{4(D-3)G_D}, \ee
where the value of $\kappa$ has been used in the second equation of Eq. (28). The value of energy of Eq. (28) is therefore divergent, so we have to use a regularization procedure to obtain a finite value. The regularized expression of Eq. (14) takes the form
\be P^a:=\frac{1}{\kappa}\int_V d^{D-2}x \left[e{S}^{a 0
0} f_T\right]-\frac{1}{\kappa}\int_V d^{D-2}x \left[e{S}^{a 0
0} f_T\right]_{AdS},\ee where {\it AdS} means calculated for pure Anti-de-Sitter space. Using (30) in solution (19), we get
\be E =\frac{(D-2)\Omega_{D-2}m}{2\kappa }= \frac{(D-2)m}{4(D-3)G_D},\ee
which is a finite value and clearly shows that the energy is depending on the coefficient of the higher order torsion terms. For the charged solutions given in (21), and by using the same procedure adopted for the non-charged case, we get
\ba \A \A E=\frac{(D-2)m}{4(D-3)G_D}-\frac{(D-3)q^2 }{2G_D r^{D-3}}+\frac{(D-3)^3\sqrt{6\left|\beta\right|}q^3 }{3(2D-5)G_D r^{2D-5}}+O\Biggl(\frac{1}{r^3}\Biggr).\ea
This shows the contributions of monopole and quadrupole potential energies to the total energy at large distances $r$. Notice only term depending on $\beta$ is the quadrupole term. Considering the limit $r\rightarrow \infty$, we get the total energy measured by a stationary observer at infinity.
\section{The first law of thermodynamics}\label{S11}
There is a great deal of work in analyzing the behavior of the horizon thermodynamics in modified theories of gravity. In a wide category of these theories, one gets solutions with horizons and can connect the temperature and entropy with the surface gravity and the area of the outer horizon. Let us now check the validity of the first law of thermodynamics in $f(T)$ gravity using the charged solution derived above.
To investigate the violation of the first law of thermodynamics of the black hole (23), let us follow the analysis performed by Miao et al. \cite{MLM}. In this work the authors use the Jacobson thermodynamics approach \cite{jacobson}, which has been generalized in \cite{akbar,adel+ahmed}, to formulate the first law through rewriting the non-symmetric field equations (9) into symmetric and skew symmetric parts as
\begin{eqnarray}\label{q88}
\A \A L_{(\mu\nu)}:=S_{\mu \nu \rho} \partial^{\rho} T f_{TT}+f_T \left[G_{\mu \nu} -\frac{1}{2}g_{\mu \nu}T\right]
+\frac{f-2\Lambda}{2}g_{\nu \mu} =\frac{\kappa {{{\cal
T}^{{}^{{}^{^{}{\!\!\!\!\scriptstyle{em}}}}}}}^\nu_\mu}{2},\nonu
\A \A L_{[\mu \nu]}:=S_{[\mu \nu] \rho} \partial^{\rho} T f_{TT}=0.\end{eqnarray}
Assuming an exact Killing vector, they have shown that for a heat flux $\delta Q$ passing through the black hole horizon, it is
\be \delta Q=\frac{\kappa}{2\pi}\left[\frac{f_T dA}{4}\right]^{d\lambda}_0+\frac{1}{\kappa}\int_H k^\nu f_{TT} \ T_{,\mu}(\xi^\rho S_{\rho \nu}{}^ \mu-\nabla_\nu \xi^\mu),\ee where $H$ stands for the black hole horizon which is equal to $(D - 2)$-dimensional boundary of the hypersurface at infinity.
The authors have shown that the first term in Eq. (33) can be rewritten as $T\delta S$ \cite{ MLM}. Thus, if the second term in Eq. (33) is not vanishing then there will be a violation of the first law of thermodynamics. Miao et al. \cite{MLM} have explained that the second term in Eq. (34) cannot be equal to zero. Therefore, if we want to satisfy the first law of thermodynamics, we must have either $f_{TT}=0$ which gives the TEGR (GR) theory, or $T=constant$. Indeed, AdS black hole solution (19) satisfies the fact that $T=constant$ and therefore, the first law of thermodynamics of this black hole is satisfied. However, the new charged solution (23) enforces the torsion scalar to be non-trivial, i.e., not constant. Therefore, according to Eq. (33), solution (23) violates the first law of thermodynamics. Explicitly calculating the violation term in (33), it is not vanishing and proportional to the electric charge $q$. Due to this feature, the first law of thermodynamics is violated.
\section{Conclusions}\label{S12}
In this work, we present a new charged solution in Maxwell-$f(T)$ gravitational theory for any dimension $D\geq 4$. The exact solution is achieved for $f(T)=T+\beta T^2$, where $\beta <0$ and possesses some interesting features. First of all, the solution has a monopole and quadrupole term which are related by requiring the metric is asymptotically AdS. This fact generalizes the result presented in \cite{CGSV13} where only the monopole term is present.
Secondly, we have studied the singularity of this black hole and have shown that all the invariants constructed from the curvature and torsion have a singularity at $r=0$. This singularity is milder than the one of a charged black hole in GR and TEGR. The asymptotic behavior of the Kretschmann invariant and the Ricci tensor squared, and the Ricci scalar have the form $K=R_{\mu \nu}R^{\mu \nu} \sim r^{-2(D-2)}$, $R\sim r^{-(D-2)}$ in contrast with, the known solutions in Einstein-Maxwell theory in both GR and TEGR. Also the non-charged solution derived in this study, Eq. (19), behaves as $K =R_{\mu \nu}R^{\mu \nu}\sim r^{-2D}$ and $R \sim r^{-D}$. Moreover, in spite that the charged black hole has different components of $g_{tt}$ and $g^{rr}$, both have a coinciding Killing and event horizons. We have calculated the total energy of the charged solution using the generalization of the energy-momentum tensor and have shown that the resulting form depends on the mass of the black hole. Finally, we have shown that the charged black hole violates the first law of thermodynamics according to the discussion given in \cite{MLM}.
From a genuine physical point of view, these kind of objects can contribute in the debate to establish on what the most reliable representation of gravity is, i.e. the curvature or torsion picture. As discussed in \cite{CCMS}, being GR and TEGR substantially equivalent, the ground of debate should be shifted to $f(R)$ and $f(T)$ models since these two theories are substantially inequivalent. As shown in \cite{diego}, also fundamental structures like gravitational waves are substantially different in $f(R)$ and in $f(T)$ formulations. A deep understanding of black hole features could be of great interest to solve the debate.
\subsection*{Acknowledgments}
We would like to thank E. N. Saridakis for useful discussions on the topic. S. Capozziello acknowledges the COST Action CA15117 (CANTATA).
This work is partially supported by the Egyptian Ministry of Scientific Research under project No. 24-2-12.
\newpage
| 9a4908e2baf9afd1eb871e15ab435a6251d0730f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The study of the structure of the Milky Way galaxy emerges
as a key tool in the analysis of the nature of its dark halo
and the most massive dwarf galaxies within it, through the manner
in which they interact with the visible gas, dust, and stars of the Galactic disk.
Astrometric observations at large scales support a cosmology
of cold dark matter (CDM) and dark energy, the so-called $\Lambda$CDM model~\citep{planck16}, and
an ``inside-out''
formation history of the cosmos~\citep{white78,frenk83} and
indeed of the Milky Way.
At small scales this model predicts
many more satellite galaxies in the Milky Way than have been
observed~\citep{kauffmann93,klypin99}.
Discoveries of faint Milky Way satellites continue
to be made (see, e.g., \citet{belokurov07}),
however, and the technical limitations of current observations
explain at least
some of the mismatch~\citep{tollerud08,walsh09,bullock10}
in the numbers of observed and expected satellite galaxies,
or subhalos.
Nevertheless, questions persist concerning the number, evolution, and mass distribution
of Milky Way subhalos (see \citet{kravtsov10} for a review),
as well as their stellar content~\citep{boylan-kolchin_big11}.
Yet more dwarf galaxies could potentially be detected through a search
for their tidal imprint on outer gaseous disks~\citep{chakrabarti11}, and low-mass subhalos
can be destroyed through interactions with the disk~\citep{donghia10}:
these disjoint ideas argue for the importance of the observational study we effect here, namely,
of the vertical structure of the Galactic disk within 2 kpc of the Sun in height and in-plane distance.
The hierarchical nature of structure formation, i.e., the
ongoing merging of clumps with time, suggests that dark matter could have surviving
phase-space structure, which could also be imprinted on the stars. This could occur, e.g.,
through the formation of stellar streams in the halo. In this paper we are interested
in identifying possible traces of halo-disk interactions and thus wish to search for
the appearance of stellar number
count distributions that break the spatial symmetries expected from the integral
of motions embedded in virialized probability distribution functions.
To do this we focus on an observational study of that
possibility through use of a sample of some 3.6 million K and M
dwarfs from the Sloan Digital Sky Survey (SDSS).
The appearance of asymmetric structures in the Milky Way has been
noted through HI gas, dust, and stellar tracers, particularly in
association with the warping of the disk~\citep{binney92}.
Various scenarios have been proposed for the appearance of warps, though analytic~\citep{nelson95}
and numerical~\citep{shen06} studies suggest a dynamical origin: warps can appear and disappear
through interactions of the disk with the halo and/or the satellites it contains
over time scales short with respect to the
age of the universe.
The stellar disk is more complicated and thus exhibits additional features.
Rings~\citep{newberg03,morganson16}, as well as ripples~\citep{price-whelan15,xu15},
have been noted, with the latter
at distances in excess of 10 kpc from the Galactic center,
where the disk is thinning out. In the solar neighborhood
vertical asymmetries have been noted in the number counts~\citep{widrow12,yanny13},
as well as in the radial velocities~\citep{widrow12},
which have been confirmed by other studies~\citep{carlin13,williams13}.
These asymmetries could be temporally recent as well and
be generated by non-axisymmetric features in the disk,
such as the spiral arms or galactic bar~\citep{debattista14,faure14,monari15,monari16},
or by interactions of the disk with the halo and its
embedded satellite galaxies~\citep{gomez13,widrow14,gomez15b,gomez16,laporte16}.
It is thought that the
asymmetries in the number
counts are more suggestive of a dynamical origin~\citep{laporte16},
although, on the other hand, it appears
that the spiral arms possess out-of-plane structure as well~\citep{camargo15}.
Only two of the Milky Way's satellites are known to be possibly
massive enough and close enough to
be able to perturb
the structure of the dark halo and Galactic disk: the Sagittarius dwarf spheroidal (dSph)
galaxy at 20 kpc and the Large Magellanic Cloud at 50 kpc from the
Sun~\citep{weinberg98,jiang99,bailin03,purcell11,gomez13}. Such massive objects are
believed to have formed at late cosmic times~\citep{boylan-kolchin11}.
Numerical simulations of the tidal interactions of the satellites reveal significant vertical
perturbations of the disk~\citep{gomez13,widrow14,gomez15b,gomez16,laporte16}.
However, simulations of such effects
do not yield the precise vertical asymmetries observed in the
existing data~\citep{laporte16}.
In this paper we revisit and expand the earlier studies of \citet{widrow12} and \citet{yanny13}
to scrutinize the vertical structure of the Galactic disk
and its variation across the Galactic plane. We
use a larger stellar photometric sample to
subdivide the disk of the Milky Way within $|z| < 2$ kpc in vertical distance from the
Galactic plane into bins of longitude $l$ and latitude $b$
to explore the extent to which the large-scale
asymmetries persist to smaller scales. We show that these asymmetries are more
pronounced toward the Galactic anticenter,
and we find similarities to the ripples
seen at further distances from the Sun by \citet{xu15} and \citet{price-whelan15}.
We provide quantitative locations and amplitudes of these asymmetries so that
those who model Milky Way structure can compare with them; associated with these
features are changes in the disk scale heights, north and south --- we report these as well.
Interestingly, numerical simulations of the
tidal interactions of the
disk-satellite-halo system reveal an induced spiral arm and
barred structure in the Galactic disk~\citep{purcell11,laporte16}.
This also motivates our observational study, and
we compare the variations we do observe with the known spiral arm structure of
the Galactic disk~\citep{camargo15}.
Our studies are also pertinent to the long-standing problem of
the determination of the matter density in the vicinity of the Sun~\citep{kapteyn22}, as
inferred from the measured kinematics of the local stars~\citep{oort32}.
In this so-called Oort problem, an assumed gravitationally relaxed population of stars
is used to trace the local gravitational potential and to infer the local matter
and dark matter densities, where we note \citet{binney08} and \citet{read14} for reviews.
The appearance of
vertical oscillations in the stellar number counts and
velocity distributions of the Milky Way~\citep{widrow12,yanny13} suggests
that the local stars may not be sufficiently gravitationally relaxed, incurring
additional uncertainty in the assessment of the local dark matter density~\citep{banik16},
a parameter key to
current efforts to detect dark matter directly~\citep{peter10}.
We believe that further observational study and analysis, as we help realize in this paper,
will be able to resolve the origin of the
vertical asymmetries and ultimately that of their impact
on the assessment of the local dark matter density.
Further studies with {\it Gaia}~\citep{gaia16a,gaia16b} will also be key~\citep{banik16}.
The SDSS~\citep{york2000} is a mature observational platform that gives us
access to a very large photometric sample of stars, observed with the same telescope under
carefully calibrated conditions. The uniformity of the photometric calibration is ensured
through multiple observations of the same stars, rendering it nearly free of
astrophysical assumptions~\citep{padmanabhan08}.
The SDSS is the first survey to provide significant
coverage in the south, allowing us to study and compare regions in
both the north and the south. We study red, main-sequence stars, particularly K and M dwarfs,
because main-sequence stars vary little in their intrinsic luminosity
for a given color and metallicity, and the population of faint red stars is less likely
to be infiltrated by non-main-sequence stars, such as giants.
Our largest analysis sample contains 3.6 million of such stars. By breaking the sky into regions,
we study the vertical structure of the Galaxy and search for changes across the footprint.
We use Galactic longitude and latitude with the Sun at the center
such that the longitude is measured counterclockwise
from the line toward the Galactic center
and latitude is measured from the plane of the Galactic equator
such that positive latitude is north, noting the standard transformation
from equatorial to Galactic coordinates~\citep{binneytext}.
As in \citet{widrow12} and \citet{yanny13},
we study the vertical symmetry of the Galactic disk
by comparing the number of stars observed north and south.
For continuity and simplicity we follow
\citet{widrow12} and \citet{yanny13}
in our modelling of
the vertical structure of the disk.
That is, we modify the solution of \citet{spitzer42} for
the vertical structure of an axially symmetric thin disk
to give a parameterization of the stellar density that has both a
thin disk and a thick disk, as inferred from observations~\citep{gilmore83}.
Apparently, the two disks have distinct
metallicities and scale lengths~\citep{bensby14,bensby13}, though we
refer the reader to \citet{bovy13} for further discussion;
we do not include these refinements, however, in our analysis.
We probe the local structure of the stellar disk
by determining
the parameters of our model across the footprint, or, specifically, as we
change the chosen range of longitude for fixed latitude.
In this paper we not only compare the structure, north and south,
with longitude, across the footprint,
but we also exploit the complete northern coverage of the SDSS to determine
(i) the rate at which the mass in the stars changes with
distance from the Galactic center, to compare with galactic models,
and (ii) how changing the selected latitude interval
for a fixed longitude interval impacts the shape of the vertical stellar number
count distribution. If there were no change in vertical structure across the Galactic plane
and if the vertical
metallicity
distribution were uniform, then the shape of the vertical distribution in number counts
would not change when the latitude window is changed, once corrections are made for the
geometric acceptance. We do, however, observe definite changes in shape; consequently, we have
a photometric proxy for changes in in-plane structure and/or metallicity.
We believe that such studies can be refined and sharpened with further observations from
the {\it Gaia} mission~\citep{gaia16a,gaia16b},
and not only through more photometric observations, because
many more spectra will also become available. Spectral information
makes it possible to measure the vertical gradient
in metallicity~\citep{hayden14}, and further studies with better
resolution across the Galactic plane should help reveal
the specific origin of the shape differences.
We begin by discussing our data selection and consider the various
systematic effects that
must be understood to determine the vertical distribution
of stellar counts. We then turn to a discussion of our fitting procedure
before presenting the results of our
north-south combined analysis and north-only analysis.
\section{Data Selection and Photometric Distance Assessment}
\label{photod}
This study uses data from the SDSS DR9~\citep{ahn12},
the Ninth Data Release of the SDSS, which was
also used in \citet{widrow12} and \citet{yanny13}. As in our earlier papers, the
errors are predominantly systematic and derive from the use of
a photometric ``parallax'' relation to relate the color of a
main-sequence star to its intrinsic luminosity and thus to determine its distance.
Particular sources of error include those associated with
stellar identification, as well as possible inadequacies in the application
of corrections for reddening and absorption due to dust. The first source of error
includes not only the possible confusion of dwarfs with giants, or the appearance
of unresolved binaries, but also the possibility of admixtures of stellar populations of
varying age and metallicity. We refer to
\citet{yanny13} for an extended discussion and for tests of our corrections for dust effects.
We follow the procedures described in \citet{yanny13}
and now summarize them briefly.
To minimize dust effects, we analyze data only above $30^\circ$ in absolute latitude, and
we correct the data for the reddening and absorption due to dust
by using the maps of \citet{schlegel98}.
In selecting the stars for our analysis,
we require $15 < r_0 < 21.5$, where $r_0$ is the apparent brightness of the star,
and $1.8 < (g-i)_{0} < 2.4$. Note that $g-i$ is the color defined as the ratio of the
intensities of the $g$-band and the $i$-band, which are parts of a five-band survey
where each band is associated with a range of frequencies, whereas
$(g-i)_0$ is the color once the effect of reddening from dust has been removed.
For our selection of $|b| > 30^\circ$, the color excess $E(B-V)$ is typically less
than 0.03 mag. This translates to a color correction in $g-r$ of
about 0.03 mag, and an error on that correction of
much less than that amount. Recall that a color error of 0.01 mag corresponds to
a distance error of $< 5\%$ and that the typical SDSS photometric color errors are
about 0.02 mag for fainter stars.
The majority of stars in our data selection are K- and M-type main-sequence stars,
which are the reddest and coldest. They
make up the majority of the main sequence, with M-type stars being the most prevalent.
By using
the reddest stars, we reduce the possibility of pollution
from non-main-sequence stars.
However, it is still possible for giants to infiltrate the sample.
Figure \ref{fig:giants} shows the results of a photometric test for the presence of giants.
This plot analyzes stars with $14.9 < r_0 < 15.4$ in a region of the sky where
giants are expected to occur; we thus expect our test sample to
have a much greater proportion
of giants than our analysis sample.
We have found
the number of giants in our photometric test sample to be a very small fraction
of the total number of stars and thus conclude that the giant admixture
in our data set is trivially small.
\begin{figure}[hp!]
\centering
\includegraphics[scale=0.65]{giantstest.pdf}
\caption[Plot of $(u-g)_{0}$ vs. $(g-i)_{0}$ to discern non-main-sequence stars]{
Plot of $(u-g)_{0}$ versus $(g-i)_{0}$ to discern non-main-sequence stars.
Stars in the north are denoted by small, filled, blue circles, whereas stars in the south
are denoted by large, open, red circles.
The plotted stars are all those in
the region $150^\circ<l<170^\circ$ and $51^\circ<\lvert b\rvert<65^\circ$
with $u$-band errors less than 0.05 mag and with $14.9<r_{0}<15.4$ ---
because this selection favors giants. The
trail of stars within the color-color region bounded by
$((u-g)_0, (g-i)_0)=(3,1.4)$ and $(4,2.2)$ as in \citet{yanny2009} and \citet{yanny13}
are the
identified giants; our selection admits the appearance of giants from the Sagittarius stream,
which is in the south.
For our analysis, we use $15<r_{0}<21.5$, which will contain an even smaller proportion of giants.
}
\label{fig:giants}
\end{figure}
In order to determine the distance to each of our selected stars, we use the
photometric parallax relations devised by \citet{juric08} for $(r-i)_{0}$ color
and by \citet{ivezic08} for $(g-i)_{0}$ color, along with the refinements
of \citet{yanny13} from fits to globular clusters
for the red stars we consider here --- we refer the reader to \citet{yanny13}
for a detailed discussion and comparison with \citet{widrow12}.
The photometric parallax method determines the distance to a star from its apparent brightness and
its intrinsic brightness --- the latter is inferred from its color and metallicity.
The relation we have used in $(r-i)_{0}$ color for the absolute magnitude
$M_{r}$ is that of \citet{juric08} but also includes a metallicity correction
$\Delta M_r$:
\begin{eqnarray}
M_r &=& \Delta M_r ([\hbox{Fe/H}]) + 3.2 + 13.30(r-i)_0 \nonumber \\
&& -11.50(r-i)_0^2+5.40(r-i)_0^3-0.70(r-i)_0^4 \,,
\label{pprmi0}
\end{eqnarray}
with
$\Delta M_r ([\hbox{Fe/H}])= -1.11 [\hbox{Fe/H}] - 0.18 [\hbox{Fe/H}]^2$ \citep{ivezic08}.
We have taken $[\hbox{Fe/H}]$ to be $-$0.3 universally
because we are unable to use the photometric metallicity assessment of \citet{ivezic08}
due to the absence of sufficiently precise
$u$-band information for the faint, red stars we consider in this paper.
Moreover, for calibration purposes, only the stellar clusters
M67 and NGC 2420 are sufficiently red and
out of the Galactic plane~\citep{yanny13}, and their metallicities are much larger
than expected for thick-disk stars. (In Section \ref{Nonlyscale} we report evidence for
decreasing metallicity in the K/M dwarfs as we sample them well above the Galactic plane.)
A similar color-magnitude equation, adapted from \citet{ivezic08},
has been used in $(g-i)_{0}$ color:
\begin{eqnarray}
M_r &=& \Delta M_r ([\hbox{Fe/H}]) -0.50 + 14.32(g-i)_0 \nonumber \\
&& -12.97(g-i)_0^2+6.127(g-i)_0^3-1.267(g-i)_0^4+0.0967(g-i)_0^5 \,,
\label{ppgmi0}
\end{eqnarray}
with the same metallicity correction.
Using the intrinsic brightness $M_r$ and the apparent brightness $r_0$ in magnitudes,
we calculate the distance $d[r_0,M_r]$ in kiloparsecs
using the relation $r_0 - M_r = -5 + 5 \log_{10} d[r_0,M_r]$.
We refine these distances using the analysis of \citet{yanny13}, so that,
for $(r-i)_0$ color, the distance $d$ is finally determined to be
\begin{equation}
d
= d[r_0, M_r( (r-i)_0,[{\rm Fe/H}])]
+ 0.1415 (r-i)_0 + 0.0436 \,,
\label{drmi0tune}
\end{equation}
whereas in $(g-i)_0$ color, it is
\begin{equation}
d
= d[r_0, M_r( (g-i)_0,[{\rm Fe/H}])]
+ 0.08982 (g-i)_0 - 0.0726 \,.
\label{dgmi0tune}
\end{equation}
For the faint, red stars we analyze in this paper
the color-color diagram, as shown in Figure 14 of \citet{yanny13}, in $(r-i)_0$ versus
$(g-i)_0$
shows
a tight correlation. We thus expect the two photometric parallax relations to work comparably well, and
this is borne out by both our findings and those of previous studies \citep{widrow12,yanny13}.
The color precision of the SDSS photometry ($\pm 0.02$ mag)
leads to typically $\pm 0.2$ mag precision in $M_r$
and thus $\pm 10\%$ distance errors~\citep{juric08}.
This assumes that the assigned metallicity is correct.
If an incorrect metallicity were used, then the distance errors would be larger.
Supposing, as usual, that the stars close to the plane (thin disk) have an [Fe/H] metallicity of
$-0.3$, but the stars far from the plane (thick disk) have an [Fe/H]
metallicity of $-0.8$, then one has an
additional error of up to $+0.4$ mag (-20\% in distance)
for a full mismatch of $\Delta {\rm [Fe/H]} =-0.5$~\citep{ivezic08}.
We thus expect the reported distances to become
gradually too large
as the population shifts from mostly thin-disk stars at $|z| < 0.5$ kpc
to that of mostly thick-disk stars at
$|z| > 1.5$ kpc because the distances have been overestimated
by 10\%-20\% as a consequence of our metallicity assumption.
In this paper we consider a sample of stars several times larger than those analyzed previously,
to the end of discerning and interpreting variations in the
stellar number distributions across the footprint. As a result, in the current analysis,
systematic errors associated with
the possible inadequacy of the dust corrections for reddening and extinction across the sky,
as well as with any nonuniformity in the photometric calibration itself,
become potentially
pertinent and need to be considered. We now address each of these effects in turn.
We have used the dust maps of \citet{schlegel98} to assess the effects of dust; they
are constructed from
observations of dust emission in the far infrared after correcting for its temperature.
Photometric~\citep{schlafly10} and
spectroscopic~\citep{schlafly11} tests have shown these maps to be
accurate for the higher-latitude ($|b| > 30^\circ$)
sightlines we employ in our study,
though this work has also revealed, working in $(g-r)_0$ color and averaging over large
angular scales, that the stars in the south are redder than those in the north.
Quantitatively the two assessments are not the same
(note Table 5 in \citet{schlafly11}), with the photometric assessment using blue-tip stars
being larger, yielding a result of
21.8 mmag redder in the south rather than $8.8\pm1.5$ mmag.
We believe that the common reddening difference is better attributed to an inadequacy in the
calibration of the $g_0$-band magnitude in the south, rather than to one
in the dust reddening
correction per se~\citep{betoule13,yanny13}, though metallicity variations in the photometric
sample can appear as well. Indeed \citet{schlafly11} suggests that the blue-tip population
could be fundamentally different in the north,
due to age and metallicity differences.
Subsequent work on different fronts supports this view.
Photometric studies with Pan-STARRS1 of stellar reddening
under assumptions of variation in
stellar metallicity and population across the Galactic disk
have recently been employed to construct an independent dust map, based on
dust absorption~\citep{schlafly14,green15}; these maps agree closely with the older
emission-based maps if compared out of the Galactic plane~\citep{schlafly14}.
using Pan-STARRS1 data, and its claimed stability of $(20, 10, 10, 10, 20)$ mmag in
$ugriz$ bands has been confirmed in the northern hemisphere by \citet{finkbeiner16}.
Additionally, they note small differences in the median magnitudes
of Pan-STARRS1 and SDSS data, north and south, of less than $\pm10\, {\rm mmag}$
in the rms for $|b|>20^\circ$ and $griz$ colors.
Were these shifts uniform over smaller
angular scales, then they would be of little relevance. However, their maps of the
differences between Pan-STARRS1 and SDSS data reveal that they can differ up to
$\pm 40$ mmag over $15'$ pixels. Nevertheless, the shifts in $(r-i)$ color
are smaller than those in $(g-i)$ color, even in localized regions~\citep{finkbeiner16}.
Thus we think that
if our conclusions are insensitive to the choice
of the photometric parallax scheme then they
should also be robust with respect to a refinement of the
photometric calibration.
Nevertheless, in later sections we consider how the errors
we discuss compare to the size of
the effects we discover.
\section{Functional Fits}
A geometric selection function is used in
order to correct for geometric effects
and to calculate a stellar density. The selection function determines
the effective volume for each bin of observed stars in height $z$ from the Galactic plane.
Each bin is divided by the associated selection function to determine the number density
at that $z$.
The selection function is
\begin{equation}
{\cal V}(z) =
\frac{1}{2}
\delta (l_2 - l_1) z^2 \left(\frac{1}{\sin^2 b_1} -
\frac{1}{\sin^2 b_2} \right) \,,
\label{selection}
\end{equation}
where $\delta$ is the width of the bin in kpc and $z$ is in
the middle of the bin and is given in kpc.
Note that $l_1 < l <l_2$, $b_1 < |b| <b_2$, and both quantities are in radians.
The coordinates
\begin{equation}
x = -8 + d \cos b \cos l \quad;\quad
y = d \cos b \sin l \quad;\quad
z = d \sin b
\end{equation}
are all measured from the Galactic center with the convention
that the Sun is located at (-8, 0, 0) kpc in this coordinate system.
We define distances from the Sun in the Galactic plane as $\Delta x$ and $\Delta y$.
Figure~\ref{fig:selectionfunction} shows a plot of $z$ versus $\Delta x$ or $\Delta y$,
demonstrating how different slices in latitude and $z$ yield a different projected area
on the $z-x$ or $z-y$ plane and impact the volume ${\cal V}(z)$,
Slices of different width in longitude
also affect ${\cal V}(z)$,
but the specific longitudes themselves do not.
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.5]{selectionfunction.pdf}
\caption[Selection function]{
Plot of $z$ vs. $\Delta x$ or $\Delta y$ to show
the different projected areas associated with various slices in latitude and in $z$.
}
\label{fig:selectionfunction}
\end{center}
\end{figure}
The stellar densities $n(z)\equiv n_{\rm raw}(z)/{\cal V}(z)$ are then plotted
and fit to the thin- and thick-disk model of \citet{widrow12},
a modification of the model of \citet{spitzer42} that adds a second disk:
\begin{equation}
n(z)=n_0 \left({\rm sech}^2\left(\frac{(z+z_{\odot})}{2H_1}\right)
+ f {\rm sech}^2\left(\frac{(z+z_{\odot})}{2H_2}\right)
\right) \,,
\label{fit}
\end{equation}
where $n_0$ is a normalization factor,
$H_1$ is the scale height of the thin disk, $H_2$ is the scale height of the thick disk,
and $f$ is the fraction of the stellar population that belongs to the
thick disk. The parameter $z_{\odot}$ is the height of the Sun
above the Galactic plane, though we emphasize that
$z_\odot$ does not correspond to a point-to-point distance because the Galactic plane
is a locally determined quantity: the Galactic disk need not be perfectly flat.
Our model assumes a two-disk structure
that is north-south-symmetric with respect to the Galactic plane.
Due to brightness saturation effects,
we analyze only those stars that possess a vertical height
$|z| > 0.35\,{\rm kpc}$
from the Galactic plane. Thus we believe that we are considering a mix of thin- and thick-disk
stars
across the SDSS footprint. It is pertinent to compare our choice with other recent models
of the structure
of the Galactic disk. The recent work of \citet{robin14} and \citet{bovy17}, for example, also
supports the use of a ${\rm sech}^2$ vertical distribution for thick-disk stars
in the Milky Way, though
the thin-disk model of \citet{czekaj14} employs Einasto ellipsoids (generalized exponentials).
In addition, the study of \citet{schwarzkopf2000} suggests, rather, that the
vertical distribution of stars in galaxies with a merger history
might be better described by a ${\rm sech}^1$ distribution.
Since our particular purpose is to
study the appearance of north-south (N/S) symmetry breaking across the
Galactic plane, we believe that our
conclusions do not rely on the particular N/S symmetric parameterization of the vertical
distributions that we use. If the stellar disk is flat over small variations in $x$ and $y$,
rather than merely flat on average, then one can expect $z_\odot$ to be universal for all lines
of sight. However, if the stellar disk has local warps or ripples, then the determined $z_\odot$ need not
be universal, and can serve as a proxy for such local variability.
Thus variations in the determination of
$z_{\odot}$ in the context of a fit to a distribution that is N/S symmetric
probe how well one
can meaningfully define the Galactic ``plane.''
To fit Eq.~(\ref{fit}) to the vertical distribution of star counts,
we optimize $\chi^2$, namely,
\begin{equation}
\chi^2 = \sum_{i=1}^{N_{\rm bin}} \left( \frac{\left( N_i - n(z_i) \right)}{\sigma_i} \right)^2\,,
\end{equation}
over $N_{\rm bin}$ bins, where $N_i$ stars are in bin $i$ centered on a height $z_i$
and $n(z_i)$ is the theoretical model of Eq.~(\ref{fit}).
The weights $\sigma_i$ are given by $\sqrt{N_i}$ of $N_i$ stars. We
use a combination of C and Python and
the fitting routines of
ROOT~\footnote{https://root.cern.ch}, a powerful data analysis framework.
At low $z$, the selection function becomes
ineffective if the stars cannot be seen because they are too bright.
A minimum $z$ cut is employed to remove this effect.
This cut was determined for each data selection
by fitting the model with trial cuts and minimizing the $\chi^2/{\rm dof}$ of the fit;
we have determined that
a minimum $z$ cut of $\lvert z \rvert>0.35\,{\rm kpc}$ can be used for all our fits.
Thus far we have addressed the vertical structure of the disk exclusively. However,
in certain stellar data sets, the variation in in-plane radial distance $R$,
noting $R\equiv\sqrt{x^2 + y^2}$,
can exceed
1.5 kpc. To investigate the consequences of such variations, we also employ fits in
which we bin our data set in $R$ as well and scale out the $R$ variation, noting
a similar procedure in \citet{banik16}, by modifying the selection
function of Eq.~(\ref{selection}) to include the factor
\begin{equation}
\exp((R - R_0)/(2.7 {\rm kpc})) \,,
\label{selectionR}
\end{equation}
where $R_0 = 8\,{\rm kpc}$ and $2.7\, {\rm kpc}$ is the central value of the radial scale
length (with an uncertainty of 4\%)
found by \citet{bovy13}. That is, after binning the data in $R$ and $z$, where
$R_i$ is the coordinate at the center of a bin in $R$, we construct
\begin{equation}
n_{\rm eff}(z) = \sum_{i} n_{\rm raw}(z,R_i) \exp((R_i-R_0)/(2.7\,{\rm kpc}))\,,
\end{equation}
after summing over all over bins in $R$ for fixed $z$, and then compute
$n(z)=n_{\rm eff}(z)/{\cal V}(z)$. As the result of this procedure we end up
with larger relative errors in $n(z)$, though we find no indications that such an
analysis is warranted. We find that using $n_{\rm eff}(z)$ in place of $n(z)$
does not affect our extracted parameters in a significant way and thus does
not affect our conclusions. Thus in what follows,
we do not employ a simultaneous analysis in $R$ and
$\phi$ outright, but rather bin our data in the vertical coordinate using the
selection function of Eq.~(\ref{selection}) to determine $n(z)$.
\begin{figure}[h!]
\begin{center}
\subfloat[]{\includegraphics[scale=0.33]{fig1a.pdf}}
\subfloat[]{\includegraphics[scale=0.33]{fig1b.pdf}}
\caption[Sloan Digital Sky Survey footprint for the north and south matched analysis]{
The SDSS footprint~\citep{aihara11}
as a map of $b$ vs. $l$. The blocked-in
regions represent the selections analyzed in the matched north and south study.
(a) A
region with $70^\circ<l<165^\circ$ and $51^\circ<\lvert b\rvert <65^\circ$.
(b) Selections over $45^\circ<l<180^\circ$
for which each wedge region in $l$ has a maximum range in $b$.
}
\label{fig:footprintmatched}
\end{center}
\end{figure}
\begin{figure}[h!]
\begin{center}
\includegraphics[scale=0.45]{polarplot_section_arms_zsun_crop.pdf}
\caption[$R$ and $\phi$ distribution of SDSS footprint]{
The sample in Figure \ref{fig:footprintmatched}(b), broken into wedges in
$l$ of roughly equal population, with those stars that have
$1.8 < (g-i)_0 < 2.4$ and $|z|< 2\,{\rm kpc}$
plotted in terms of their Galactocentric coordinates $R$ and $\phi$. The indicated
radial distances $R$ are in kpc, and ``180'' marks the azimuth
$\phi=180^\circ$ that extends through the location of the Sun to the Galactic center.
The different $l$ wedges, of which there are 14 in all, span different distances in the
$(R,\phi)$ plane because each
has a different latitude wedge.
The spiral arm structures inferred from distances to embedded clusters
determined by \citet{camargo15} are also indicated as dashed-dotted lines, with the inner arm
being that of Sagittarius-Carina and the outer arm being Perseus.
The Orion spur, in which the Sun is located, is not apparent, however.
The wedges correspond to those analyzed in Figure \ref{fig:histogramsmatchedset}.
For subsequent reference, we have also superimposed
the $z_\odot$ results of Figure \ref{fig:zsun}, with
``$\pm$'' denoting a $z_{\odot}$ greater (less) than the global average at 1$\sigma$ and
``$0$'' denoting no significant change.}
\label{fig:polarplot}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=0.27]{fig3bn.pdf}
\includegraphics[scale=0.27]{fig3cn.pdf}
\includegraphics[scale=0.27]{fig3dn.pdf}
\includegraphics[scale=0.27]{fig3en.pdf}
\includegraphics[scale=0.27]{fig3fn.pdf}
\includegraphics[scale=0.27]{fig3gn.pdf}
\includegraphics[scale=0.27]{fig3hn.pdf}
\includegraphics[scale=0.27]{fig3in.pdf}
\includegraphics[scale=0.27]{fig3jn.pdf}
\includegraphics[scale=0.27]{fig3kn.pdf}
\includegraphics[scale=0.27]{fig3ln.pdf}
\includegraphics[scale=0.27]{fig3mn.pdf}
\includegraphics[scale=0.27]{fig3nn.pdf}
\includegraphics[scale=0.27]{fig3on.pdf}
\caption[Stellar density as a function of vertical displacement z from the Sun]{
Stellar density as a function of vertical displacement $z$ from the Sun.
These plots employ the expanded ranges in $|b|$ for the north and south matched set (Figure \ref{fig:footprintmatched}(b)). Each histogram represents a specific region in $l$ and $|b|$ (as shown
in degrees in the upper-right corner of each image).
}
\label{fig:histogramsmatchedset}
\end{center}
\end{figure}
\begin{figure}[hp!]
\centering
\includegraphics[scale=0.50]{fig5-zsun.pdf}
\caption[$z_{\odot}$ as a function of longitude for north and south matched analysis]{
The Sun's height above the Galactic plane, $z_{\odot}$, in pc as a function of longitude
obtained from the best-fit models of the stellar density histograms.
The graph overlays the values obtained for the combined north and south analysis
for each region in latitude: with the uniform selection
as the blue squares (region shown in Figure \ref{fig:footprintmatched}(a)) and
the expanded selection as the red triangles (region shown in Figure \ref{fig:footprintmatched}(b)).
The blue and red lines, respectively,
show the average $z_{\odot}$
obtained for the uniform and expanded latitudes within the region $70^\circ \le l \le 165^\circ$.
}
\label{fig:zsun}
\end{figure}
\begin{figure}[hp!]
\begin{center}
\subfloat[]{\includegraphics[scale=0.5]{thindiskthick_fig6.pdf}}
\subfloat[]{\includegraphics[scale=0.5]{thickdiskthick_fig6.pdf}}
\caption[Thin and thick disk thicknesses as functions of longitude for north and south matched analysis]{
Thickness $H_1$ and $H_2$, respectively, of (a) the thin dick and (b) the thick disk in kpc,
as functions of longitude as
obtained from the best-fit models of
the stellar density histograms for the uniform latitude selections (shown in
Figure \ref{fig:footprintmatched}(a)) and the expanded latitude selections
(shown in Figure \ref{fig:footprintmatched}(b)).
The notation and conventions are identical to those in Figure \ref{fig:zsun};
the blue and red lines show the average thicknesses.
The thicknesses are correlated with the thick-disk
fraction $f$, so that $f$ tends to be smaller when the thicknesses are larger.
}
\label{fig:matchedthinthick}
\end{center}
\end{figure}
\subsection{North and South Combined Analysis}
\label{NScomb}
We begin by studying
regions selected such that the north and south have completely matched coverage.
The first selection has been designed for uniformity
by selecting stars with Galactic coordinates in $70^\circ < l < 165^\circ$
and $51^\circ < |b| <65^\circ$, which are then analyzed
in longitude increments of $5^\circ$, $10^\circ$, or $15^\circ$ so that each selection has
roughly the same number of stars (Figure \ref{fig:footprintmatched}(a)).
This uniform latitude cut creates a sample that can be directly compared as a function of
longitude without the added complication of changing latitude.
The second selection has been made to maximize the number of stars included in the analysis,
by dividing the sample into regions
that cover $5^\circ$, $10^\circ$, or $15^\circ$ of longitude, and choosing
the maximum range in latitude possible in each wedge
while maintaining complete coverage in the north and south (Figure \ref{fig:footprintmatched}(b)).
The $R$ and $\phi$ distributions of the $(l,b)$ wedges in Figure \ref{fig:footprintmatched}(b),
for the particular range of colors $1.8 < (g-i)_0 < 2.4$ that we analyze, are shown in Figure \ref{fig:polarplot}.
Using the previously discussed cuts and selections,
we apply the selection function in Eq.(\ref{selection}) and employ
the fitting function of Eq.~(\ref{fit}) with the photometric parallax method
in $(g-i)_0$ color as described in Section \ref{photod} to determine the distance
to each star. We require $0.35\,{\rm kpc} \le z \le 2.0\,{\rm kpc}$, where the
maximum value of $z$ is determined by color completeness~\citep{yanny13}.
We first analyzed
the region shown in Figure \ref{fig:footprintmatched}(a)
to determine what variations occur when
changing the selected longitude range and then repeated our analysis with the
expanded latitude sample of Figure \ref{fig:footprintmatched}(b). There are 1.05 and 3.59
million stars in total in the selections of Figures~\ref{fig:footprintmatched}(a)
and ~\ref{fig:footprintmatched}(b), respectively.
The changes in Galactic parameters for the two sets
are sufficiently small that
we have shown the stellar density profiles for the expanded latitude selections in
Figure \ref{fig:footprintmatched}(b).
Figure~\ref{fig:histogramsmatchedset}
shows the results, displaying the stellar number density as a function of vertical distance.
We find
the quality of fit ($\chi^2/{\rm dof}$) to range from 1.4 to 2 for the fits shown.
Although
the vertical stellar distribution is grossly smooth and symmetric, the
plots do show a definite north-south asymmetry of ${\cal O}(10\%)$ in the stellar
number counts, with the effect
becoming more marked at larger $l$, as noticed in \citet{yanny13}, though the
expanded latitude data set is four times larger.
The wave-like nature of the departures from the north-south-symmetric fit argues against
either metallicity variations or
a failure of the north-south photometric calibration in explaining this result; indeed
the possible sizes of such effects are too small to have an impact.
We discuss the residuals,
north and south, in Section \ref{cfNS}. We note that the quality of fit is poorest
in regions where the observed wave-like nature of the N/S asymmetry is most prominent.
In the remainder of this section we consider the
changes in the Galactic parameters across the regions described in
Figures~\ref{fig:footprintmatched}(a) and (b), to provide a context
for our analysis of the N/S variations.
Figures~\ref{fig:zsun}, \ref{fig:matchedthinthick}(a), and \ref{fig:matchedthinthick}(b)
show the fit parameters $z_{\odot}$, thin-disk thickness, and thick-disk thickness,
respectively, as functions of longitude.
Our analysis shows only modest variation in $z_{\odot}$ across
the footprint, both for the selections with a
uniform latitude-cut (which we henceforth refer to as ``uniform'')
and for the expanded latitude selections (which we henceforth refer to as ``expanded'').
The average $z_\odot$ values obtained for the uniform and expanded analyses,
respectively, are $z_{\odot} = 14.9 \pm 0.5$ pc and $z_{\odot} = 15.3 \pm 0.4$ pc.
The average for the uniform analysis is consistent
with the value of $z_{\odot} = 14.3 \pm 0.6$ pc found by \citet{yanny13}.
Fitting the expanded results with a straight line
reveals a $\chi ^2/{\rm dof}$ of 3.53,
so that there appears to be a genuine variation with longitude. However, there is little
difference in the average if the uniform or expanded latitude selection is
employed. Our values for $z_\odot$ are about 2$\sigma$ lower than the
recent literature using similar stars, noting $z_\odot \sim 25$ pc~\citep{juric08} and
$z_\odot = 27.5 \pm 6 \rm pc$~\citep{chen99,chen01}.
Other determinations of $z_\odot$, such as the value of
$26 \pm 3$ pc found in a study of Cepheid variables~\citep{majaess09}, or that
of \citet{joshi07}, using younger population tracers,
with values ranging
from 6 to 20 pc,
differ
from these results and support the existence of the
``environmental'' sensitivity we have found (see also \citet{bovy17}).
Figure~\ref{fig:matchedthinthick} shows the analogous results
for the disk thicknesses.
The differences between the averages for uniform and expanded thickness are
small, particularly for the thin disk, which is most pertinent to the north-south
asymmetry seen in Figure \ref{fig:histogramsmatchedset}, as will become clear in
Section \ref{cfNS}. Recall, too, that the
errors in photometric distances are at least $\pm 10\%$.
The evidence for variations with $l$ is also weaker than in the case of $z_\odot$,
though a change in the vertical scale height farther from the
Galactic center --- and hence at larger $l$ --- is expected~\citep{kent91,narayan02}.
Sensitivity to such variations
is evidently limited by our statistics,
so that we will return to this point in our
north-only analysis, for which we have a vastly larger number
of stars.
Although other analyses have revealed a range in the value of $z_\odot$, determined
from stars of differing spectral class in differing regions of the sky,
our analysis is the first to reveal variations in its value
across the Galactic plane with stars of the same spectral class.
Barring the existence of significant
in-plane metallicity gradients, which would be at odds with
the results of existing observational studies~\citep{hayden14},
we believe the variations we have found speak to a Galactic disk that possesses ripples
and thus is not globally flat.
We have confirmed that this feature, and indeed all the
features we have found thus far, also appear in our analyses
using photometric parallax with $(r-i)_0$ color.
We note, moreover, that a positive warp in the north,
toward $l\sim 90^\circ$, has also been observed~\citep{russeil03},
though this does not explain the variations in $z_\odot$ that we have observed.
The ripples could potentially arise from tidal effects on the disk, which appear
in numerical simulations of disk-satellite
encounters~\citep{gomez13,widrow14,gomez15b,gomez16,laporte16}.
However, it is also important to compare our results with known in-plane features that
vary across the SDSS footprint, and thus we consider
the Milky Way's spiral arm structure.
The dust obscuring the Milky Way's disk has limited our ability to study its spiral arms.
Questions persist as to their nature and origin~\citep{baba09,sellwood11,sellwood14},
as well as to their precise structure~\citep{georgelin76,levine06,hou09,lepine11,francis12,camargo13,camargo15,bobylev14,griv14,pettitt14,vallee14}.
The absence of bright, massive $^{13}$CO clouds in the interarm
space~\citep{roman-duval09} promotes the association of these apparently short-lived
features with the spiral arms themselves. As a result, young, red embedded clusters
of stars should
act as tracers of the spiral arms~\citep{camargo13,camargo15a,camargo15b,camargo15}.
\citet{camargo15a}
have challenged the notion of spiral arms as in-plane structures
with north-south-symmetric vertical
extent~\citep{cox02,monari16},
revealing embedded clusters with locations extending both above and
below the plane. These authors also note that cluster formation
can be associated with the halo as well~\citep{camargo15b,camargo16}.
Nevertheless, it is possible that the variations in $z_{\odot}$ are
related to an out-of-plane structure of the spiral arms.
Figure~\ref{fig:polarplot}
shows the spiral arm structure detailed in \citet{camargo15} overlaid on the
analysis sample of Figure \ref{fig:footprintmatched}(b) plotted in Galactocentric coordinates $R,\phi$, and
the analyzed wedges in longitude
have been marked
with a $+$ in regions where the value of $z_{\odot}$ is greater than the global average
and with a $-$ in regions where it is less than the global average.
While the work of \citet{camargo15a,camargo15b,camargo15}
shows evidence for some parts of the
Milky Way's stellar spiral arms being
located a few hundred parsecs above or
below the Galactic plane,
we do not see any
correlation between these out-of-plane
stellar spiral arms and variations in $z_\odot$ with $l$ or
any correlation with
regions of the Galactic plane where significant north-south
asymmetries are apparent at $|z| > 0.35$ kpc.
\subsection{North and South Comparative Analysis}
\label{cfNS}
We now turn to a detailed analysis of the north-south differences shown in
Figure \ref{fig:histogramsmatchedset}. Working in a coordinate system
with the Galactic plane as its origin (by shifting $z$ so that $z\rightarrow z +z_\odot$),
we show residuals calculated as (data-model)/model as a function of vertical distance
in Figure \ref{fig:residualoverlays}.
We overlay the residual in the north and that
in the south in each region of Figure \ref{fig:histogramsmatchedset} using the parameters
of its fit. This visualizes the north-south asymmetry of Figure \ref{fig:histogramsmatchedset} and
confirms the earlier results of \citet{widrow12} and \citet{yanny13}.
The residuals have wave-like features that
grow in amplitude and vertical extent
with increasing longitude, refining and extending what was observed by \citet{yanny13}.
Interestingly, at low $z$ the south always has a positive residual,
whereas the north always has a negative residual.
Note that although the figures go out to 2.5 kpc,
the fits themselves only extend to 2 kpc.
Repeating this analysis on regions with uniform latitude
reveals the same pattern; the changes due to latitude are small
compared to the effects seen in the figure. The data set with the
expanded latitude selection is used here because it allows for a larger range in longitude and better
statistics. We find that similar results emerge when we repeat our analysis using distances computed
using $(r-i)_0$ color.
\begin{figure}[hp!]
\begin{center}
\includegraphics[scale=0.267]{4550.pdf}
\includegraphics[scale=0.267]{5055.pdf}
\includegraphics[scale=0.267]{5560.pdf}
\includegraphics[scale=0.267]{6065.pdf}
\includegraphics[scale=0.267]{6570.pdf}
\includegraphics[scale=0.267]{7075.pdf}
\includegraphics[scale=0.267]{7585.pdf}
\includegraphics[scale=0.267]{8595.pdf}
\includegraphics[scale=0.267]{95105.pdf}
\includegraphics[scale=0.267]{105120.pdf}
\includegraphics[scale=0.267]{120135.pdf}
\includegraphics[scale=0.267]{135150.pdf}
\includegraphics[scale=0.267]{150165.pdf}
\includegraphics[scale=0.267]{165180.pdf}
\caption[Residuals of the stellar densities overlaying north and south]{
Residuals of the stellar densities, namely,
the difference between the observed
stellar density and the best-fit model divided by the best-fit model, are
shown with respect to
the magnitude of
the vertical distance
to the Galactic plane, $|z_{\rm GP}| = |z + z_\odot|$ with $z_\odot$ = 14.9 pc.
The plots overlay the residuals in the north (filled, black)
with those in the south (open, red).
Each panel corresponds to the histogram covering the same region
in $l$ and $b$ in Figure \ref{fig:histogramsmatchedset}, which is the result of
using the matched north and south regions with expanded latitude shown in
Figure \ref{fig:footprintmatched}(b).
}
\label{fig:residualoverlays}
\end{center}
\end{figure}
\begin{figure}[hp!]
\centering
\begin{center}
\subfloat[]{\includegraphics[scale=0.5]{thindiskthickness.pdf}}
\subfloat[]{\includegraphics[scale=0.5]{thickdiskthickness.pdf}}
\caption[
Plots of the thickness of the thin (a) and thick (b) disks
as functions of longitude comparing the north and the south]{
Plots of the thickness of the thin (a) and thick (b) disks
as functions of longitude obtained from best-fit
models on the stellar density histograms using $(g-i)_0$ color
for the uniform latitude regions shown in Figure \ref{fig:footprintmatched}(a). We compare
regions in the north (green triangles), the south (maroon diamonds), and the north and
south combined (blue squares).
}
\label{fig:thinthicknorthsouth}
\end{center}
\end{figure}
\begin{figure}[hp!]
\begin{center}
\subfloat[]{\includegraphics[scale=0.5]{thindiskthickness_color_calibration.pdf}}
\subfloat[]{\includegraphics[scale=0.5]{thickdiskthickness_color_calibration.pdf}}
\caption[Color calibration comparing the north and the south]{
Thicknesses of the thin (a) and thick (b) disks
as functions of longitude, repeating the analysis of
Figure \ref{fig:thinthicknorthsouth} and adopting its notation, but now testing for sensitivity
to color effects.
The filled symbols represent the analysis done with $(g-i)_{0}$ color, implementing
an overall shift of -16 mmag shift in
the $(g-i)_{0}$ color in the south,
whereas the open symbols represent the analysis performed using $(r-i)_{0}$ color.
}
\label{fig:colorcalibration}
\end{center}
\end{figure}
\begin{figure}[hp!]
\begin{center}
{\subfloat[]{\includegraphics[scale=0.27]{NS8to8p5phi175to180a.pdf}}
\subfloat[]{\includegraphics[scale=0.27]{NS8to8p5phi170to175a.pdf}}}
{\subfloat[]{\includegraphics[scale=0.27]{NS8to8p5phi180b.pdf}}
\subfloat[]{\includegraphics[scale=0.27]{NS8to8p5phi175b.pdf}}
\subfloat[]{\includegraphics[scale=0.27]{NS7p5to8phi175b.pdf}}}
\caption[R-phi distribution]{
We illustrate the origin of the N/S difference in the Galactic disk
parameters by plotting the raw number counts, in the north (black) and
south (red), using the samples of Figures \ref{fig:footprintmatched}(a) and (b),
restricted to those stars that satisfy $1.8 \le (g-i)_0 \le 2.4$.
We plot particular ranges in the Galactocentric coordinates
$R$ and $\phi$, as a function of the vertical distance determined
from the center of the Galactic plane. We use the
average value of $z_\odot = 14.9\, {\rm pc}$ that emerged from our (uniform)
analysis of Figure \ref{fig:footprintmatched}(a) for this purpose.
We have made the following selections:
in (a) $R\in [8.0, 8.5]\,{\rm kpc}$ and $\phi \in [175^\circ, 180^\circ]$,
in (b) $R\in [8.0, 8.5]\,{\rm kpc}$ and $\phi \in [170^\circ, 175^\circ]$, both from
Figure \ref{fig:footprintmatched}(a),
and
in (c) $R\in [8.0, 8.5]\,{\rm kpc}$ and $\phi \in [175^\circ, 180^\circ]$,
in (d) $R\in [8.0, 8.5]\,{\rm kpc}$ and $\phi \in [170^\circ, 175^\circ]$,
in (e) $R\in [7.5, 8.0]\,{\rm kpc}$ and $\phi \in [170^\circ, 175^\circ]$,
all from Figure \ref{fig:footprintmatched}(b).
}
\label{fig:reftable}
\end{center}
\end{figure}
\begin{figure}[hp!]
\begin{center}
\includegraphics[scale=0.60,angle=-90]{polarplot_arrange2_alt_fix.pdf}
\caption[R-phi distribution in footprint]{
We illustrate the N/S variations in the raw number counts with $R$ and $\phi$
using the
samples of Figure \ref{fig:footprintmatched}(b) in the red box and the notation and choices of
Figure \ref{fig:reftable}.}
\label{fig:Phibreaking}
\end{center}
\end{figure}
\tabletypesize{\footnotesize}
\begin{deluxetable}{l|l|lllll}
\tablecaption{Best-fit Results to the North only (n), South only (s), and
North-South Combined (c) Selections of the Uniform Latitude Sample.}
\tablehead{
\multicolumn{1}{c} {$l$ range (deg)} &
\multicolumn{1}{c} {} &
\multicolumn{1}{c} {$N_{0}$ $(\times 10^6)$} &
\multicolumn{1}{c} {$H_1$ (kpc)} &
\multicolumn{1}{c} {$H_2$ (kpc)} &
\multicolumn{1}{c} {$f$} &
\multicolumn{1}{c} {$\chi^2 /{\rm dof}$} \\
}
\startdata
70 $< l <$ 75& s& 6.11 $\pm$ 0.33& 0.211 $\pm$ 0.007& 0.688 $\pm$ 0.028& 0.112 $\pm$ 0.008& 1.18\\
& c& 4.83 $\pm$ 0.17& 0.239 $\pm$ 0.006& 0.773 $\pm$ 0.034& 0.108 $\pm$ 0.008& 1.19\\
& n& 3.86 $\pm$ 0.16& 0.278 $\pm$ 0.009& 1.004 $\pm$ 0.118& 0.083 $\pm$ 0.013& 0.96\\\hline
75 $< l <$ 85& s& 6.01 $\pm$ 0.24& 0.210 $\pm$ 0.005& 0.662 $\pm$ 0.019& 0.116 $\pm$ .006& 1.07\\
& c& 5.07 $\pm$ 0.13& 0.234 $\pm$ 0.004& 0.742 $\pm$ 0.022& 0.104 $\pm$ 0.006& 1.12\\
& n& 4.35 $\pm$ 0.13& 0.262 $\pm$ 0.006& 0.899 $\pm$ 0.059& 0.082 $\pm$ 0.008& 0.95\\\hline
85$ < l <$ 95& s& 5.71 $\pm$ 0.22& 0.213 $\pm$ 0.005& 0.662 $\pm$ 0.020& 0.113 $\pm$ 0.007& 1.23\\
& c& 4.84 $\pm$ 0.12& 0.239 $\pm$ 0.004& 0.735 $\pm$ 0.024& 0.100 $\pm$ 0.006& 1.32\\
& n& 4.19 $\pm$ 0.12& 0.271 $\pm$ 0.006& 0.933 $\pm$ 0.077& 0.069 $\pm$ 0.009& 1.14\\\hline
95$ < l <$ 105& s& 5.76 $\pm$ 0.23& 0.211 $\pm$ 0.005& 0.654 $\pm$ 0.021& 0.108 $\pm$ 0.007& 0.83\\
& c& 4.60 $\pm$ 0.11& 0.242 $\pm$ 0.004& 0.734 $\pm$ 0.026& 0.098 $\pm$ 0.007& 1.20\\
& n& 3.86 $\pm$ 0.10& 0.278 $\pm$ 0.006& 0.948 $\pm$ 0.088& 0.066 $\pm$ 0.009& 1.18\\\hline
105$ < l <$ 120& s& 6.17 $\pm$ 0.24& 0.196 $\pm$ 0.005& 0.595 $\pm$ 0.013& 0.119 $\pm$ 0.005& 1.10\\
& c& 4.74 $\pm$ 0.10& 0.235 $\pm$ 0.004& 0.695 $\pm$ 0.019& 0.099 $\pm$ 0.006& 1.48\\
& n& 4.01 $\pm$ 0.09& 0.276 $\pm$ 0.005& 1.028 $\pm$ 0.092& 0.052 $\pm$ 0.006& 1.13\\\hline
120$ < l <$ 135& s& 5.69 $\pm$ 0.20& 0.206 $\pm$ 0.005& 0.626 $\pm$ 0.017& 0.105 $\pm$ 0.006& 1.09\\
& c& 4.72 $\pm$ 0.10& 0.236 $\pm$ 0.004& 0.718 $\pm$ 0.022& 0.085 $\pm$ 0.005& 1.38\\
& n& 4.18 $\pm$ 0.10& 0.268 $\pm$ 0.005& 0.952 $\pm$ 0.074& 0.051 $\pm$ 0.006& 1.10\\\hline
135$ < l <$ 150*& s*& 0.67 $\pm$ 0.04& 0.588 $\pm$ 0.015& 0.193 $\pm$ 0.005& 9.580 $\pm$ 0.471& 1.20\\
& c& 4.76 $\pm$ 0.10& 0.232 $\pm$ 0.004& 0.686 $\pm$ 0.020& 0.088 $\pm$ 0.005& 1.63\\
& n& 4.00 $\pm$ 0.09& 0.273 $\pm$ 0.005& 1.006 $\pm$ 0.095& 0.045 $\pm$ 0.006& 1.21\\\hline
150$ < l <$ 165& s& 6.20 $\pm$ 0.27& 0.190 $\pm$ 0.005& 0.574 $\pm$ 0.012& 0.115 $\pm$ 0.005& 1.12\\
& c& 4.78 $\pm$ 0.11& 0.229 $\pm$ 0.004& 0.666 $\pm$ 0.019& 0.092 $\pm$ 0.006& 1.65\\
& n& 4.14 $\pm$ 0.10& 0.269 $\pm$ 0.004& 1.015 $\pm$ 0.097& 0.041 $\pm$ 0.005& 1.15\\\hline
\enddata
\tablecomments{ \label{table:f}
The asterisk denotes a case in which the fit switched
the thin and thick disks.}
\end{deluxetable}
Given the visual differences between the north and the south shown
in Figure \ref{fig:residualoverlays}, a quantitative comparison has been made by
performing fits on the north and south independently.
We employ the same thin- and thick-disk
model but for one modification.
The original model had difficulties fitting stars that are only in the north or only in the south;
this is solved by fixing the value of $z_{\odot}$
--- the other parameter values are determined by fitting.
We use a $z_{\odot}$ of $14.3\,{\rm pc}$,
the average value obtained by \citet{yanny13}, though our results are not
sensitive to that particular choice.
The results of these fits are shown in Figure \ref{fig:thinthicknorthsouth},
which shows the thicknesses of the thin and thick disks as functions of longitude.
In these fits we found it pertinent to employ the uniform sample of
Figure \ref{fig:footprintmatched}(a) in order to focus on the possibility of longitudinal
variations in a crisp way.
Separate fits to the northern and southern samples reveal a much greater
thickness in the north
for both the thin and thick disks, though it is
also the case that the fraction $f$ of the thick disk relative to the thin disk
is also smaller in the north.
We show an explicit illustration of this correlation in Table~\ref{table:f},
which lists the fit parameters for the two-disk component fits (note Eq.~(\ref{fit}))
to the matching north and south data samples in bins of $5^\circ-15^\circ$
bins in longitude (see Figure \ref{fig:footprintmatched}(a)), including
the normalization $N_0$, the thin-
and thick-disk scale heights $H_1$ and $H_2$, respectively,
the relative fraction $f$ of the thick- and thin-disk
populations, and the $\chi^2/{\rm dof}$
for each fit. We show the fit results
for the sample in the north fitted by itself, the sample in the south fitted
by itself, and for the combined north-south sample.
In the combined north-south fits, $f$ is approximately 10\%,
and the thin- and thick-disk scale heights
are 0.24 and 0.7 kpc, respectively, with a relatively good $\chi^2/{\rm dof}$.
When the fits are done separately for the north and south, however,
the scale heights for both the thin and thick disks are always
larger in the north, by about 20\%-30\%, with a corresponding decrease
in $f$, by about 10\%-50\%. This strong correlation
between $f$ and the scale heights has been noted several times previously
in the literature, and it has been speculated that
it is a consequence of a fitting degeneracy, rather than an indication of a
variation in the scale heights and $f$ with position or hemisphere.
We refer to Figure 9 of \citet{chen01}, as well as Figure 21
of \citet{juric08} for concrete
examples.
The combined values of the scale heights and thick-disk fraction are in good
agreement with \citet{juric08}, which analyzed observations
from both the northern and southern hemispheres, finding $H_1=0.24$ kpc and $H_2=0.8$
kpc with $f$ around 0.1; they, too, noted a
strong correlation between $f$ and the disk scale heights.
Given the good $\chi^2/{\rm dof}$ of the separate
north and south fits shown in Table \ref{table:f}
(to our knowledge such separate fits have not been done before),
and the poorer combined fits, we suggest, rather, that the difference in
scale heights and fractions with hemisphere is a real effect.
Its existence supports that of a large north-south
asymmetry in the vertical stellar distribution in
the vicinity of the Sun. Figure 10 shows this is not due to calibration errors.
The N/S differences shown in Figures \ref{fig:residualoverlays}--\ref{fig:colorcalibration}
are depicted in an $(l,b,z)$, Sun-centered coordinate system. One may also present the
same star-count data in an $(R,\phi,z)$ coordinate
system based on the Galactic center, and this is shown in
Figures \ref{fig:reftable} and \ref{fig:Phibreaking}.
Panels (a) and (b) of Figure \ref{fig:reftable} shows two azimuthal bins of star-count
data from the footprint of Figure 3(a) at radial distances up
to 0.5 kpc from the sun, and up to 3 kpc from the Galactic plane.
The pattern of asymmetries seen in several
panels of Figure \ref{fig:residualoverlays} is clearly present in panel (a), which combines
several $(l,b)$ bins into one nearby $(R,\phi)$ bin.
Panels (c)-(e) offer $(R,\phi)$ presentations of the
asymmetries of the data in the expanded footprint of Figure \ref{fig:footprintmatched}(b).
While the selection function of the data points with $|z|$ has not been applied, though
it has been in Figure \ref{fig:residualoverlays}, the data
samples are in all cases taken from matching north and south
regions, plotting $|z_{\rm GP}| = |z + z_\odot|$ with $z_\odot$ = 14.9 pc.
Note that the N/S thickness difference emerges as a gross feature close
to the Galactic plane in both the uniform and expanded latitude samples.
Figure \ref{fig:Phibreaking} offers the broadest look at the data
arranged in Galactic $(R,\phi)$ coordinates.
Here raw star counts, uncorrected by the selection function, are compared
in N/S matched samples.
In this figure we relax the $|z| < $ 2 kpc constraint for better
illustration. The circular inset at left
portrays the plane of the Galaxy marked with the azimuthal coordinate $\phi$ and
radial coordinate $R$, where
($R_\odot,\phi_\odot$) = (8 kpc, 180$^\circ$). We have broken our
coverage in the ($R$, $\phi$) plane into smaller regions indicated
by the small solid dots, and we have picked out
the dots (regions) with the most star counts, as indicated by the red box,
for explicit illustration.
The faint large gray arcs
indicate curves of constant $R$ in steps of 0.5 kpc. Each row of
overlapping inset panels selects a different $\phi$ range, as marked.
Each panel shows in black (red) the northern (southern) Galactic
hemisphere counts as a function of $|z_{\rm GP}|$ from 0-6 kpc,
where $|z_{\rm GP}|$ has been adjusted to account for a universal offset
of the Sun above the plane of $z_\odot$ = 14.9 pc. The fact that
the north and south counts converge at large $|z_{\rm GP}|$ suggests that
this offset is appropriate. While the fine asymmetrical structure
in $(l,b)$ is not visible, one still may clearly see, especially in
the panel for $8 < R < 8.5$ kpc, $175^\circ < \phi < 180^\circ$, that there are
vertical oscillations in the stellar density near the Sun.
Several other panels show an excess of counts in the south at
vertical distances up to 1 kpc from the plane.
We see explicitly that (i) there are more stars in the south
over most of the footprint and (ii) the feature of
Figure \ref{fig:reftable}(c), which drives the wave-like N/S asymmetry we have
found, as well as the N/S variation in Galactic parameters, also
breaks axial symmetry. It also has no analogue at slightly larger $R$.
We encourage further studies of the
combined stellar and dark matter mass distributions in the solar neighborhood
to delve into its origin.
It should be noted that a partial explanation of these differences could come from
a global calibration difference.
The photometric calibration of
the SDSS survey is done by studying adjacent regions across the footprint and
calibrating overlapping regions together~\citep{padmanabhan08}.
Since the entire footprint in the north is connected, it can be calibrated in a uniform way.
However, the northern and southern regions of the footprint do not overlap,
so that they cannot be calibrated in this manner.
Studies of the blue-tip stars suggest that the stars are systematically redder in the south~\citep{schlafly10,yanny13}.
To study the implications of this, a downward shift of -16 mmag for $(g-i)_{0}$ color
was implemented in the south,
which would redress the typical differences in reddening found by \citet{yanny13}.
By repeating the fits with
$(g-i)_{0}$ color
as well as repeating the fits using distances determined with $(r-i)_{0}$ color,
Figure \ref{fig:colorcalibration} is obtained.
The shift in color calibration
has a negligible effect on the thicknesses, as we would expect from our discussion in
Section \ref{photod}.
Although using the $(r-i)_{0}$ color relation
reveals slightly different thicknesses,
the north-south offsets still exist, encouraging the stance
that there is a real difference in the stellar densities between north and south.
\begin{figure}[hp!]
\begin{center}
\subfloat[]{\includegraphics[scale=0.33]{fig12a.pdf}}
\subfloat[]{\includegraphics[scale=0.33]{fig12b.pdf}}
\caption[Sloan Digital Sky Survey footprint for north only analysis]{
The SDSS footprint~\citep{aihara11} as a map of $b$ vs. $l$.
The blocked-in regions represent the sections this analysis studies.
(a) A longitude range of $35^\circ<l<210^\circ$ with latitude of $50^\circ< b < 90^\circ$.
(b) Two regions of longitude $35^\circ <l< 60^\circ$ and $165^\circ< l< 210^\circ$ with latitude $30^\circ< b< 90^\circ$.
}
\label{fig:footprintnorth}
\end{center}
\end{figure}
\subsection{North-only Analysis: Scale Height Changes Across the Footprint}
\label{Nonlyscale}
Since more stars have been observed in the north
we use these observations to
study variations in Galactic parameters across the footprint.
Once again, we have made two different types of selections.
Both divide the footprint into slices of
$5^\circ$, $10^\circ$, and $15^\circ$ slices in longitude,
but each uses different latitudes.
The first analysis uses a fixed latitude range of $50^\circ < b <90^\circ$
and longitude range $35^\circ < l < 210^\circ$ (Figure \ref{fig:footprintnorth}(a)), in which
we have 1.67 million stars in total.
The second analysis uses a fixed latitude range of $30^\circ < b <90^\circ$ but only
looks at the longitudes $35^\circ < l < 60^\circ$ and $165^\circ < l < 210^\circ$ (Figure \ref{fig:footprintnorth}(b)).
These regions are selected in order to have the largest latitude coverage possible, and
we analyze 2.22 million stars in total.
Since we use $ b >30^\circ$ in order to avoid dust effects,
this is the maximum range of latitude allowed by our analysis.
\begin{figure}[hp!]
\begin{center}
\subfloat[]{\includegraphics[scale=0.5]{fig13.pdf}}
\subfloat[]{\includegraphics[scale=0.5]{norththinradial.pdf}}
\caption[Thin disk thickness as a function of longitude and radial distance for the north only analysis]{
Thin-disk thickness as a function of longitude (a) and radial distance (b). The graphs overlay the values obtained from Figure \ref{fig:footprintnorth}(a) (blue squares) and the values obtained from Figure \ref{fig:footprintnorth}(b) (red diamonds).
In (b) horizontal errors of ${\cal O}(0.5-0.7 {\rm kpc})$ in $R$ have not been included.
}
\label{fig:thinnorth}
\end{center}
\end{figure}
\begin{figure}[hp!]
\begin{center}
\subfloat[]{\includegraphics[scale=0.5]{fig14.pdf}}
\subfloat[]{\includegraphics[scale=0.5]{norththickradial.pdf}}
\caption[Thick disk thickness as a function of longitude and radial distance for the north only analysis]{
Thick-disk thicknesses as a function of longitude (a) and radial distance (b).
The graphs overlay the values obtained from Figure \ref{fig:footprintnorth}(a) (blue squares) and the values obtained from Figure \ref{fig:footprintnorth}(b) (red diamonds).
In (b) horizontal errors of ${\cal O}(0.5-0.7 {\rm kpc})$ in $R$ have not been included.
}
\label{fig:thicknorth}
\end{center}
\end{figure}
\begin{figure}[hp!]
\begin{center}
\subfloat[]{\includegraphics[scale=0.5]{fig15a.pdf}}
\subfloat[]{\includegraphics[scale=0.5]{fig15b.pdf}}
\caption[Thin and thick disk thicknesses as functions of minimum latitude]{
Thicknesses of the thin (a) and thick (b) disks
as functions of minimum latitude for an analysis of the region $55^\circ<l< 60^\circ$
in the north with a latitude up to $90^\circ$.
}
\label{fig:latitudechange}
\end{center}
\end{figure}
Figures \ref{fig:thinnorth}
and \ref{fig:thicknorth} show the thicknesses of the thin and thick disks as functions of
longitude and the in-plane radial distance $R$ from the Galactic center
for both of the regions depicted in Figure \ref{fig:footprintnorth}.
The difference in thickness for regions
with the same longitudes but different latitudes reveals that the
scale heights are sensitive to the value of the lower latitude cut. We have checked
that this result persists once we repeat our fits to a data sample in which the $R$ dependence
has been scaled out, as per the modified selection function of Eq.~(\ref{selectionR}).
It can also be observed,
particularly for the analysis of Figure \ref{fig:footprintnorth}(b) shown in Figures \ref{fig:thinnorth} and \ref{fig:thicknorth},
that there is an increase in thickness, for both disks, with an increase in
$R$.
The value of $R$ used is the average value of $\sqrt{x^2+y^2}$
for the stars in that wedge that are included in the fit.
The red points, corresponding to the inclusion of stars with latitudes down
to $30^\circ$, sample distances further from the Galactic center at large $l$ and
those closer to the Galactic center at small $l$.
In a single-disk model,
in particular, if $\rho(z) = \rho_0 {\rm sech}^2 (z/(2 z_0))$, then the surface mass
density $\Sigma$ has the form $\Sigma = 4\rho_0 z_0$~\citep{spitzer42,binney08}.
Figure \ref{fig:thinnorth} shows an increase in scale height of about 10\% over a change in
$R$ from 7 to 9 kpc, but from our fits we determine the parameter $z_0$ to be more
nearly constant and, moreover,
we infer that $\Sigma$ is decreasing.
A change in the vertical scale height with $R$ has been noticed previously by \citet{kent91}
(and in galaxies other than the Milky Way by \citet{degrijs97}),
with an estimated scale height of $\sim 247$ pc at
$R=8$ kpc, in good agreement with our result for the thin-disk scale height.
Its origin has been discussed by \citet{narayan02}.
It is important to note that this analysis
covers the entire range of latitude from $30^\circ < b <90^\circ$; this provides us with
more statistics than our analyses limited to higher latitudes.
In all of our results, the thin- and thick-disk thicknesses appear to depend
on the lower limit of the latitudes included in the analysis.
To show this effect explicitly, Figure \ref{fig:latitudechange} shows how the thickness
depends on te minimum latitude chosen,
for a region with $55^\circ < l < 60^\circ$ and a maximum latitude of $90^\circ$.
We thus see that a higher minimum latitude cut
leads to a larger thickness, for both the thin and thick disks; the change with minimum latitude can be as large as 20\%.
Since we employ a selection function to eliminate geometric effects and the
sizes of the systematic errors that could account
for such a change are ruled out (see Section 2),
we expect that
this apparent change arises from one or more physical effects.
In order to understand this apparent change in thickness with latitude cut,
regions of a low minimum latitude cut ($30^\circ < b <50^\circ$) and a high minimum latitude cut ($50^\circ < b <90^\circ$)
are compared in Figure \ref{fig:latitudeoverlay}. This has been repeated
for all the regions shown in Figure \ref{fig:footprintnorth}(b). Although the shapes do not change grossly,
it can nevertheless be seen at high $z$ that the higher-latitude data points
have a greater stellar density than the lower-latitude data points.
This
illustrates how the measured thickness could increase with increasing minimum latitude cut.
The apparent increase in thickness could be the result of
vertically changing stellar populations or of a change in in-plane structure.
The manner in which a change in the selected latitude window changes
the sampling of the Galactic plane is
illustrated in Figure \ref{fig:selectionfunction}.
The different latitude selections also sample the stellar populations
at different average heights above the Galactic plane as shown in Table \ref{table:avgz}. Note that $\langle z_{\rm raw} \rangle$
determines the average in $z$ of the raw stellar number counts --- no selection function has been applied.
We observe that the approximate difference in $\langle z_{\rm raw} \rangle$ in the two latitude samples is about
$0.5$ kpc, with the difference increasing slightly, to about $0.6$ kpc, at large $l$. Since we expect the
metallicity of a selected star to be smaller as its value of $z$ grows larger, noting that this has been
established
in a recent spectroscopic study of red giants~\citep{hayden14}, we expect that the vertical changes in the metallicity of
the stellar populations should play a role to some extent.
\begin{figure}[hp!]
\begin{center}
\includegraphics[scale=0.27]{overlay-35-40.pdf}
\includegraphics[scale=0.27]{overlay-40-45.pdf}
\includegraphics[scale=0.27]{overlay-45-50.pdf}
\includegraphics[scale=0.27]{overlay-50-55.pdf}
\includegraphics[scale=0.27]{overlay-55-60.pdf}
\includegraphics[scale=0.27]{overlay-165-180.pdf}
\includegraphics[scale=0.27]{overlay-180-195.pdf}
\includegraphics[scale=0.27]{overlay-195-203.pdf}
\includegraphics[scale=0.27]{overlay-203-210.pdf}
\caption[Stellar density histogram overlay of high and low latitude]{
Stellar density histograms in the north for the regions in Figure \ref{fig:footprintnorth}(b) overlaying a region with latitude $30^\circ< b<50^\circ$ (filled, black) and a region with latitude $50^\circ< b<90^\circ$ (open, red).
}
\label{fig:latitudeoverlay}
\end{center}
\end{figure}
\tabletypesize{\footnotesize}
\begin{deluxetable}{c|cc|ccc}
\tablecaption{Quantifying the Effects of Latitude Selection in the North-only Analysis
\label{table:avgz}}
\tablehead{
\multicolumn{1}{c} {$l$ range (deg)} &
\multicolumn{1}{c} {$b$ range (deg)} &
\multicolumn{1}{c} { $\langle z_{\rm raw} \rangle$ (kpc)} &
\multicolumn{1}{c} { $z$ range (kpc) } &
\multicolumn{1}{c} { $\Delta$[Fe/H] (dex)} &
\multicolumn{1}{c} { $\Delta M_r$ (mmag)}
\\
}
\startdata
\hline
35 $< l <$ 40& 30 $< b <$ 50& 1.57 & 0.35 $< z <$ 0.5& 0.01 & -10\\
& 50 $< b <$ 90& 2.09 & 0.8 $< z <$ 1.3& -0.08 & 80 \\
& & & 1.7 $< z <$ 2.0& -0.49 & 450 \\
\hline
40 $< l <$ 45& 30 $< b <$ 50& 1.55 & 0.35 $< z <$ 0.5& -0.04 & 40 \\
& 50 $< b <$ 90& 2.10 & 0.8 $< z <$ 1.3& -0.1 & 100 \\
& & & 1.7 $< z <$ 2.0& -0.6 & 500 \\
\hline
45$ < l <$ 50& 30 $< b <$ 50& 1.54 & 0.35 $< z <$ 0.5& -0.03 & 30\\
& 50 $< b <$ 90& 2.08 & 0.8 $< z <$ 1.3& -0.05 & 50 \\
& & & 1.7 $< z <$ 2.0& -0.7 & 600 \\
\hline
50$ < l <$ 55& 30 $< b <$ 50& 1.53 & 0.35 $< z <$ 0.5& 0.05 & -50 \\
& 50 $< b <$ 90& 2.07 & 0.8 $< z <$ 1.3& -0.06 & 60 \\
& & & 1.7 $< z <$ 2.0& -0.53 & 480 \\
\hline
55$ < l <$ 60& 30 $< b <$ 50& 1.51 & 0.35 $< z <$ 0.5& 0.05 & -50 \\
& 50 $< b <$ 90& 2.07 & 0.8 $< z <$ 1.3& -0.06 & 60 \\
& & & 1.7 $< z <$ 2.0& -0.62 & 550 \\
\hline
165$ < l <$ 180& 30 $< b <$ 50& 1.29 & 0.35 $< z <$ 0.5& -0.3 & 280 \\
& 50 $< b <$ 90& 1.95 & 0.8 $< z <$ 1.3& -0.48 & 440 \\
& & & 1.7 $< z <$ 2.0& -1.32 & 1010 \\
\hline
180$ < l <$ 195& 30 $< b <$ 50& 1.29 & 0.35 $< z <$ 0.5& -0.44 & 410 \\
& 50 $< b <$ 90& 1.95 & 0.8 $< z <$ 1.3& -0.46 & 420 \\
& & & 1.7 $< z <$ 2.0& -1.4 & 1100 \\
\hline
195$ < l <$ 203& 30 $< b <$ 50& 1.30 & 0.35 $< z <$ 0.5& -0.5 & 500 \\
& 50 $< b <$ 90& 1.95 & 0.8 $< z <$ 1.3& -0.34 & 320 \\
& & & 1.7 $< z <$ 2.0& -1.26 & 977 \\
\hline
203$ < l <$ 210& 30 $< b <$ 50& 1.28 & 0.35 $< z <$ 0.5& -0.29 & 280 \\
& 50 $< b <$ 90& 1.92 & 0.8 $< z <$ 1.3& -0.38 & 360 \\
& & & 1.7 $< z <$ 2.0& -1.3 & 1000 \\
\hline
\enddata
\end{deluxetable}
We now quantify the latitude-dependent difference in shape in
each panel of Figure \ref{fig:latitudeoverlay}.
To do this, we have determined the metallicity shift required to minimize the shape differences
between the high- and low-latitude
samples. Applying the selection function to each latitude sample, we
multiplied the number of counts in each bin in the
high-latitude sample by an overall scaling factor such that the total counts for
each latitude sample are the same. This has been
done to permit a comparison of the vertical distribution of the stars
regardless of their net number.
We then examined the three different regions in $z$, in which we have seen
shape differences, separately.
In each one we executed a $\chi^2$ test to determine
whether the shapes of the
two latitude samples are the same, computing, namely,
\begin{equation}
\chi_{b}^2 = \sum_{i=1}^{N_{\rm bin}} \frac{\left(n_i^{>} - n_i^{<}\right)^2}{n_i^{<}} \,,
\end{equation}
where $n_i^>$ and $n_i^<$ are the numbers of stars in the (rescaled) high-latitude and
low-latitude samples in bin $i$ of $N_{\rm bin}$ equal-width bins. We then modified
the metallicity of the higher-latitude sample by steps of -0.01
until $\chi_b^2$ is minimized.
We repeated this for each region in $z$ and then repeated the entire process
for each slice of longitude.
The results of this analysis
are shown in Table \ref{table:avgz} as a metallicity shift $\Delta$[Fe/H] in dex and as a shift
in the intrinsic magnitude $\Delta M_r$ in mmag as per Eq.~(\ref{ppgmi0}).
We can see that the shifts in
$\Delta M_r$ are typically much larger, particularly at large $z$, than
the photometric calibration
shifts of less than 10 mmag discussed by \citet{finkbeiner16} and thus speak to a physical effect.
(We have also checked
that similar features emerge when we repeat our analysis in $(r-i)_0$ color.)
These shifts are largely, but not completely,
explained by the common paradigm of two Galactic disks ---
a thin disk with a [Fe/H] of -0.3 and a thick
disk with a [Fe/H] of -0.8 with different scale heights.
Different $(l,b)$ bins sample different ratios of thin- and thick-disk stars, and thus our
approximation that all stars have a [Fe/H] of -0.3 is more incorrect as $|z|$ moves further
above 0.5 kpc.
Interestingly, too, the manner in which the metallicity changes, that it
decreases substantially at large $z$,
is also consistent with the trend found in spectroscopic studies~\citep{hayden14},
though the change is also larger in
the large $l$ samples at large $z$ than the typical
shifts observed in those studies~\citep{hayden14}.
It is intriguing that a simple thick- and thin-disk paradigm
is not enough to explain the very large shifts seen for $1.7 < |z| < 2.0$ kpc, toward
the anticenter ($165^\circ < l < 210^\circ$),
and here a significantly different stellar population,
with lower metallicity than the thick disk may be present in significant amounts. This
is worth exploring further. It is also possible that the vertical structure changes
as one moves across the Galactic plane.
We anticipate that the rich spectroscopic data sets
to emerge from the {\it Gaia} mission~\citep{gaia16a,gaia16b}
will help to resolve the origin of the
effect we have found regarding the
change in the vertical distributions of stars with latitude.
We nevertheless view photometric studies of the sort we have pioneered here
as being of continuing utility, because they serve as an
efficient way of
discerning what the most interesting
regions of the sky might be for detailed spectroscopic studies.
\section{Summary and Future Prospects}
We have used a photometric sample of up to 3.6 million K and M dwarf stars from the SDSS to
study the structure of the Galactic disk in the vicinity of the Sun.
By selecting different regions of Galactic latitude and longitude, we have been able
to study changes in structure as a function of the stars' location within
and above the Galactic plane. As shown in Figure \ref{fig:histogramsmatchedset},
the stellar number count distributions are remarkably smooth across the footprint, though
we have discovered significant changes in their distribution nonetheless.
Our sample of red, main-sequence stars are the longest-lived stars, so that we would expect
its one-body distribution function to be the solution of a collisionless Boltzmann equation,
which describes the interaction of a star with a mean-field mass distribution.
We have presumed this solution to be both axially and mirror (north-south) symmetric in
adopting a ${\rm sech}^2 (z/2H_1)$ form for its spatial projection in $z$, $n(z)$, as
we have used in Eq.~(\ref{fit}). With this expectation,
we have analyzed the stars and their distribution through matched observations
in the north and the south,
as well as through observations made in the north only. The departures we have
found break these supposed symmetries, though we cannot definitely say whether this is the
result of recent dynamical interactions of the disk stars with external agents or
because of other internal effects, such as the presence of the spiral arms or the Galactic bar,
so that they do not hold precisely.
Our combined north-south analysis
reveals a difference in the stellar densities north and south, providing further evidence for
a vertical wave in number counts found by \citet{widrow12} and \citet{yanny13},
as well as circumstantial
evidence for dynamical symmetry breaking of the presumed integrals of motion.
We have presented, in Figure \ref{fig:residualoverlays}
and Table \ref{table:f}, quantitative measures of the
differences in star counts in matched areas above and below the plane of the Milky Way
as a function of distance in $(l,b)$ bins of ${\cal O}(10^\circ)$.
We find significant offsets
from north-south symmetry
in individual bins of up to 20\% in star counts, even after correcting
for the location of the Sun at some 14.9 pc above the plane. These differences cannot
be explained solely by misestimations of the metallicities of the stars, because
the differences
are too large and oscillate in $|z|$ with a vertical period of about $400$ pc.
We have also studied the evolution of the north-south
differences with longitude and find that they
become more pronounced as $l$ approaches $l \approx 180^\circ$.
Panels (c)-(d) in Figure \ref{fig:reftable} (and Figure \ref{fig:Phibreaking}) indicate
variations
in the azimuthal,
in addition to the vertical, direction.
Although we have observed vertical wave-like features (see Figure 8),
such are not apparent in the breaking of axial symmetry we observe.
A close look at Figure 12 reveals, rather, an impulse localized within
5 degrees
of the Galactic anticenter.
This is intriguingly reminiscent of the numerical study of the impact of the
Sagittarius dSph
galaxy with the Galactic disk by \citet{purcell11}.
A comparison of the disk thicknesses, north and south, also reveals
that the thicknesses are larger in the north, though this is also correlated with a
smaller thick-disk fraction in the north.
These results remain after correcting for possible color calibration offsets and
repeating the analysis with $(r-i)_{0}$ color.
In addition, we have used our comparative analysis
to determine the inferred location of the Sun above the Galactic plane
across the footpoint: we can say
that
this distance, $z_\odot$, changes across the footprint, speaking
possibly to ripples in the stellar density across the plane.
No significant correlation is seen between sightlines toward spiral arms where the
stellar spiral arms of the Galaxy appear out of the plane by up to several hundred parsecs
and the wave-like overdensities we have seen at distances of $> 500$ pc --- but the data
are very limited: such a correlation cannot be ruled out.
Studying the observations in the north exclusively admits the possibility of studying the effects of changing
latitude, allowing the analysis of observations up to $b=90^\circ$ with improved statistics. We have found that the
determined thicknesses, for both thin and thick disks,
do depend on the value of the minimum latitude analyzed.
An increase in thickness was also found with increasing lower latitude cut, which is
likely representative of vertically changing stellar
populations or of changes in the in-plane structure.
Finally we have compared the shape in the stellar distributions with vertical height for different
latitude windows, finding definite changes in shape that we think are reflective of either or
both stellar population changes and/or changes in the in-plane structure.
A variety of systematic effects have been considered in this study.
The effects of dust are minimized by employing dust corrections and restricting our analysis to
out-of-plane sightlines with $|b|>30^\circ$. We have also performed a photometric test
for giants and have found that their infiltration into our analysis sample is
completely negligible. The possibility of a color
calibration effect that could be different in the north and south has also been
studied explicitly and determined to be insignificant.
Our error in converting from stellar colors to distance is a combination of that in the
photometric
calibration in magnitude and color and of metallicity misestimation, noting that
we have used [Fe/H] of the thin disk
for a mixed population of thin- and thick-disk stars in the absolute
magnitude calculation.
The distance errors from calibration and color error have been
estimated to be approximately
10\% rms, or 0.2 in absolute magnitude~\citep{juric08}, and we have refined our
distance assessment using red globular clusters. As for the metallicity error,
we reiterate that
we fix [Fe/H] to be -0.3, i.e., we assume a 100\% thin-disk population
for our photometric metallicity correction. An additional error would
appear if we were, rather, sampling a 100\% thick-disk population with [Fe/H] of -0.8.
The empirical effects of sampling different ratios of thin- and thick-disk populations
stars in different $(l,b)$ bins are tabulated in Table \ref{table:avgz}.
We leave the exercise of
``inverting''
the results of this table
to deduce the thin-disk and thick-disk fractions in each bin
to a future work.
We note that the results are broadly consistent with a switchover from a thin-disk
to a thick-disk population as one moves from heights of $< 0.5$ kpc
to over 1.5 kpc above
or below the plane, with a few places where halo or other low-metallicity
stars, with [Fe/H] of -1.6,
may represent a significant fraction of the mix toward the Galactic
anticenter.
We ask the question of whether the larger differences of up to 20\%
in stellar density between
the north and south in various longitude bins seen in Figure \ref{fig:residualoverlays},
could be due to a significant
difference in the metallicity of the population of stars between the north and the
south, rather than a difference in overall star density at a given distance.
For instance, following the lower right panel in Figure \ref{fig:residualoverlays},
which has $165^\circ < l < 180^\circ$,
$46.6^\circ < |b| < 64.3^\circ$,
there is a 20\% excess in density of star counts at $|z|=0.7$ kpc in the
north versus the south.
To explain this difference as a difference in population metallicity
and not as an overall stellar density wave, one would need, for instance, a very large
difference in the metallicity
as a function of $|z|$ in the ratio of thin- to thick-disk stars between
north and south, and that large difference in metals is not compatible
with the relatively smooth change in metals content of stars seen around the
Galaxy~\citep{ivezic08,hayden14}.
Plotting the matched N/S stellar count data in Galactic
$(R,\phi,z)$ coordinates shows a 20\% excess of counts in the south
just beyond (0-0.5 kpc) the solar radius toward the anticenter at
a height of 0.4 kpc below the Galactic plane. This excess becomes
a 10\% deficit (or excess in the north) once one reaches a height
of 0.8 kpc above the plane (see Figure \ref{fig:Phibreaking}). It is interesting
to note that at lower vertical heights, the Orion spur is
in this same part of the sky. In other
directions, as in we see in Figure \ref{fig:residualoverlays},
at similar (sub-kpc) distances from the Sun, the
asymmetry mostly manifests itself as an excess of counts in
the south at about 0.4 kpc below the plane. In all subsamples
of the matched north-south star-count data, we report here
that the scale height of both the thin- and thick-disk stellar
populations appears to be systematically larger in the
north than the south, when the scale heights are fit separately.
This may indicate a slight displacement of the center plane
of the thick disk above the thin disk near the Sun or other
configurations suggesting that the disks are not fully in stationary
equilibrium.
These quantitative locations and amplitudes of over- and underdensities
should be useful in serving as a constraint on dynamical models of the Galaxy,
which would describe its dark matter
distribution, its satellites, and its stellar disk
as they interact and evolve over time. Insights
into the past history and distribution of matter can be inferred by a model that
reproduces these disk asymmetries. We note, too, that
comparing the vertical distribution of
number counts as a function of selected latitude may prove an efficient way of
locating possible stellar populations, or streams,
of ultralow metallicity for subsequent spectroscopic study.
The $\Lambda$CDM model speaks to dark matter with small-scale phase-space structure
and motivates the search for stellar streams.
Follow-up studies at yet higher resolution with our methods could look for
variations and asymmetries with greater sensitivity, though they
would likely require improved photometric calibrations and
reddening corrections.
These improvements already largely exist~\citep{green15,finkbeiner16},
so that we look forward to data from the {\it Gaia} era~\citep{gaia16a,gaia16b} with
confidence in the ability to further and refine the studies pioneered here.
\acknowledgements{
This manuscript evolved from D.F.'s 2016 May senior honors thesis at the University of Kentucky,
and she acknowledges the support of
The University of Kentucky Singletary Scholarship
as well as an Undergraduate Summer Research
Fellowship from the University of Kentucky during its completion.
D.F. and S.G. acknowledge partial support from the U.S. Department of Energy under
contract DE-FG02-96ER40989. S.G. thanks Jonathan Feng and his colleagues at the
University of California, Irvine for gracious hospitality
during the completion of
this work.
We thank the referee for several important suggestions that improved
the presentation of this paper, including presentation of the data in
Galactic coordinates.
Funding for the Sloan Digital Sky Survey has been provided by the
Alfred P. Sloan Foundation, the U.S. Department of Energy Office of
Science, and the Participating Institutions. SDSS acknowledges
support and resources from the Center for High-Performance Computing at
the University of Utah. The SDSS web site is http://www.sdss.org$\, .$
SDSS is managed by the Astrophysical Research Consortium for the Participating Institutions of the SDSS Collaboration including the Brazilian Participation Group, the Carnegie Institution for Science, Carnegie Mellon University, the Chilean Participation Group, the French Participation Group, Harvard-Smithsonian Center for Astrophysics, Instituto de AstrofΓsica de Canarias, The Johns Hopkins University, Kavli Institute for the Physics and Mathematics of the Universe (IPMU) / University of Tokyo, Lawrence Berkeley National Laboratory, Leibniz Institut f\"ur Astrophysik Potsdam (AIP), Max-Planck-Institut f\"ur Astronomie (MPIA Heidelberg), Max-Planck-Institut f\"ur Astrophysik (MPA Garching), Max-Planck-Institut f\"ur Extraterrestrische Physik (MPE), National Astronomical Observatories of China, New Mexico State University, New York University, University of Notre Dame, ObservatΓ³rio Nacional / MCTI, The Ohio State University, Pennsylvania State University, Shanghai Astronomical Observatory, United Kingdom Participation Group, Universidad Nacional AutΓ³noma de MΓ©xico, University of Arizona, University of Colorado Boulder, University of Oxford, University of Portsmouth, University of Utah, University of Virginia, University of Washington, University of Wisconsin, Vanderbilt University, and Yale University.
}
{\noindent
$^*$Present Address: {School of Physics, Georgia Institute of Technology, Atlanta, GA 30332, USA}}
\bibliographystyle{apj}
| f1350140b5b2375d85ce883661fd81e68361888f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\subsection{Acknowledgement}
This work was supported by the Canada Excellence Research Chairs program and the National Science and Engineering Research Council of Canada (NSERC). R.F. also acknowledges the support of the Banting postdoctoral fellowship of NSERC.
| 8b8dbbe82035345627057fb7040b2e4f3a1d93fa | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Abstract}
\textbf{Background}: The analyses of human daily interactive behaviors at the population level play a critical role to understand many dynamic processes, e.g., information diffusion, and disease transmission, taking place in the human society. The recent advance of information technologies such as mobile phone, RFID, wireless sensor, and our interest in, WiFi provides powerful measures to digitize the population interactions and facilitate quantitative investigations.\\
\textbf{Methods and Findings}: To explore large-scale population indoor interactions, we analyze \emph{18,715} users' WiFi access logs recorded in a Chinese university campus during \emph{3} months, and define two categories of human interactions, the event interaction (EI) and the temporal interaction (TI). The EI helps construct a transmission graph, and the TI helps build an interval graph. The dynamics of EIs show that their active durations are truncated power-law distributed, which is independent on the number of involved individuals. The transmission duration presents a truncated power-law behavior at the daily timescale with weekly periodicity. Besides, those \emph{`leaf'} individuals in the aggregated contact network may participate in the \emph{`super-connecting cliques'} in the aggregated transmission graph. Analyzing the dynamics of the interval graph, we find that the probability distribution of TIs' inter-event duration also displays a truncated power-law pattern at the daily timescale with weekly periodicity, while the pairwise individuals with burst interactions are prone to randomly select their interactive locations, and those individuals with periodic interactions have preferred interactive locations.\\
\textbf{Conclusion and Significance}: The dynamics of a large-scale population's indoor interactions shows the convolution of human activities' daily bursts and weekly rhythms. The `leaf' individuals in the aggregated contact network gathering to function as `hubs' in the transmission graph highlights the significant difference between aggregated and temporal structures, and the analyses of temporal dyads from the spatiotemporal perspective pave a new path to predict human interactive behaviors.
\section*{Introduction}
Nowadays, a great deal of digital technologies are unobtrusively embedded into the physical world of human daily activities: we communicate with each other through emails and/or mobile phones, pay the shopping with credit cards, use public transportation by transit cards, and surf via wired or wireless networks. These technical innovations not only reshape our social life, but also record tremendous digital trajectories of human activities, which provide proxy clues to understand human behavioral patterns \cite{TechnoSocialSystem,ComputeSocialScience,email,mobilephone,banknote,onlinedata,Pentland2010AmerScience,Pentland2009WEF,Pentland2008Book,EagleN2005PhD,EagleN2006PUC,EagleN2009PNAS,RFID2010PLoSONE,RFID2011PLoSONEHosptial,RFID2010JTB,RFID2011PLoSONESchool,WirelessSensor2010PNAS,Takaguchi11PRX}. For instance, Pentland et al. \cite{Pentland2010AmerScience,Pentland2009WEF,Pentland2008Book} constructed `sociometers' to capture human honest signals in the workplace to infer their productivity and creativity. Eagle et al. \cite{EagleN2005PhD,EagleN2006PUC,EagleN2009PNAS} used Bluetooth embedded in the mobile phones to infer friendships from close proximity interactions. Barrat and his colleagues \cite{RFID2010PLoSONE,RFID2011PLoSONEHosptial,RFID2010JTB,RFID2011PLoSONESchool} built a flexible framework based on the distributed active Radio Frequency Identification (RFID) technology to record human face-to-face interactions in different rendezvouses, e.g., conference, office, hospital, museum, and school. Salath\'{e} et al. \cite{WirelessSensor2010PNAS} developed wireless sensors to evaluate the respiratory disease transmission risk among the members of a high school through person-to-person contacts. Takaguchi and his colleagues analyzed two sets of face-to-face interaction logs sampled by infrared sensor devices to predict conversation patterns\cite{Takaguchi11PRX}. Since human physical close proximity has been recognized to play a crucial role to (biological and computer) virus transmission, word-of-mouth information diffusion, and human social ties\cite{EagleN2009PNAS,RFID2011PLoSONEHosptial,RFID2010JTB,RFID2011PLoSONESchool,WirelessSensor2010PNAS,Takaguchi11PRX,RFIDDisease,RFIDInformation}, these fruitful researches not only shed light on human interactive features in a small community, but also inspire us to put a step further to explore human indoor interactions in an environment with a large-scale `natural' population.
Recent years, WiFi(or known as IEEE 802.11), as a ubiquitous wireless data exchange technology, has become one of the brightest areas in the communication business. Actually, WiFi signals can be found at almost every corner of the urban areas, and the notion of `WiFi-city' becomes reality. A university is a snapshot of the modern society, so is the WiFi coverage, where the WiFi control system records the digital access logs of the authorized WiFi users during their activities in the campus. Such WiFi access records work the proxy of a large-scale population's interactive activities with time stamps \cite{EPL68002,UrbComp2012}, which deserve more extensive efforts to capture the temporal patterns of their dynamics.
The WiFi dataset involved in this study was collected at Handan campus of Fudan University in Shanghai, China, which contains \emph{18,715} WiFi users' individual behavioral trajectories during \emph{3} months of a fall semester. We define two categories of human indoor interactions: event interaction(EI) and temporal interaction(TI). From the perspective of temporal networks\cite{Temporalnetworks}, the EIs help construct a transmission graph to characterize the group interactive features, and the TIs build an interval graph to characterize the dyad interactive features. The dynamics of EIs show that their active durations are truncated power-law distributed and `size-free', independent on the number of involved individuals. The transmission duration presents a power-law behavior at the daily timescale with weekly periodicity. Besides, some \emph{`leaf'} individuals in the aggregated contact network participate in the \emph{`super-connecting cliques'} of the aggregated transmission graph. Analyzing the dynamics of the interval graph, we find that the probability distribution of TIs' inter-event duration also displays a truncated power-law pattern at the daily timescale with weekly periodicity, while the pairwise individuals with burst interactions are prone to randomly select their interactive locations, and those individuals with periodic interactions have preferred interactive locations.
\section*{Results}
\subsection*{Two Categories of Human Indoor Interactions and Temporal Networks}
Before transforming the WiFi data of access logs as the proxy of a large-scale population's indoor interactions, we state our main assumption \cite{opportunistic2007IEEE,EPL68002,UrbComp2012} in this paper: the WiFi devices seeing the same wireless access point(WAP) infers an indoor interaction among these devices' owners (see \emph{Methods} and \emph{Text SI} for the validity of the main assumption).
We first define two categories of human indoor interactions and temporal graphs. Figure \ref{schematic}(a) shows an instance which translates three WiFi users' access logs into the individual behavioral traces. The bold lines pertain to their online durations. In the observed epoch [$t_0,t_7$], these three users are uniquely recorded by this WAP.
\indent
\textbf{From Event Interactions To Transmission Graph}
\indent
With the aforementioned main assumption - the users 'seeing' the same WAP have an indoor interaction, Figure \ref{schematic}(b) illustrate the process of translating individual behavioral trajectories into event interactions(EIs) and constructing the corresponding transmission graph(TG). Event interaction is defined to characterize the group interactive dynamics and evolution features, which facilitates the analysis of epidemic spread or information diffusion. For instance, in the epoch of [$t_{1},t_{2}$], the online duration of \emph{A} and \emph{B} are overlapped, we therefore define an EI $E_{AB}^{t_1}$ characterizing the fact that \emph{A} and \emph{B} are assembled as a contact group.
Regarding event interactions as vertices \cite{TG1,TG2}, we define two vertices are linked given the following three rules are satisfied:
\begin{enumerate}
\item In the time series, a source EI (e.g., $E_{AB}^{t_1}$) is the closest EI prior to the sink EI ($E_{ABC}^{t_2}$).
\item At least one user coexists in the source and sink EIs.
\item When there are several sources before one sink, any set of the shared users between the given source and sink EIs never intersect with each other(set).
\end{enumerate}
The three rules ensure that the source EI has a potential to transfer some information or virus to the sink EI due to the intermediate role of shared individuals. As illustrated in Figure 1(b), $E_{AB}^{t_1}$ is active in the epoch [$t_{1},t_{2}$] and $E_{ABC}^{t_2}$ in [$t_{2},t_{3}$], while two EIs satisfy the above three rules, a directed edge links these two vertices in the transmission graph.
To quantify the temporal information in event interactions and the transmission graph, we introduce the following quantities:
\begin{itemize}
\item The size \emph{s} of a given EI is defined as the number of users involved.
\item If a given EI is active in the period of [$t^{begin},t^{end}$], its active duration is defined as \\$\Delta t^{EI}=t^{end}-t^{begin}$.
\item Given two connected EIs, $i, j$, linked in the TG, denote the edge weight $\omega_{ij}=(\delta, t^{observed})$, i.e., the shortest transmission duration from the source $E_i$ to the sink $E_j$ is $\delta$ when observed at the time $t^{observed}$. (Without special statement, $t^{observed}$ is considered as $t^{begin}$ of each source EI in this paper.)
\end{itemize}
Generally, the topological properties of a temporal graph are measured from the corresponding aggregated network on a given time window \cite{RFID2010PLoSONE,RFID2011PLoSONEHosptial,RFID2010JTB,RFID2011PLoSONESchool }. Here, we aggregate the transmission graphs on the whole observation period (\emph{3 } months) to generate the aggregated transmission graph (ATG), and define several quantities as follows:
\begin{itemize}
\item The frequency $n^{EI}$ gives the number of EI's active times. For instance, $E_{AB}$ in Figure \ref{schematic}(b) takes place in two time intervals, $[t_1,t_2]$ and $[t_3,t_4]$, so the frequency $n^{EI}_{AB}=2$.
\item The number of repeated transmission paths from source EI (e.g.,$E_i$) to sink EI ($E_j$) is frequency $n^{TP}_{E_iE_j}$.
\end{itemize}
\indent
\textbf{From Temporal Interactions To Interval Graph}
\indent
We further define temporal interactions to characterize the dyad interactions. With the main assumption mentioned above, \emph{A} and \emph{B} in Figure \ref{schematic}(a,c), who share a common duration, assemble into a temporal interaction $T_{AB}^{t_1}$ during the epoch [$t_{1},t_{4}$]. The set of temporal interactions(TIs) can be represented by an interval graph \cite{Temporalnetworks} directly, where users are the vertices, and TIs are the edges.
The following definitions of an interval graph are introduced to quantify the temporal information of dyad interactions:
\begin{itemize}
\item Given an active TI during [$t^{started},t^{finished}$], the active duration of the edge (TI) is \\$\Delta t^{TI}= t^{started}-t^{finished}$.
\item The frequency of edge $n^{TI}_{ij}$ gives the number of times that two connected users $i,j$ repeat their TIs.
\item Given the same dyad's two successive TIs, $T^{1}_{ij}$ and $T^{2}_{ij}$ ($i,j$ are two users), their active durations are [$t^{started}_{1},t^{end}_{1}$] and [$t^{started}_{2},t^{end}_{2}$], respectively. The inter-event duration is defined as the difference between the beginning time of these two successive TIs ($\Delta\tau=t^{started}_{2}-t^{started}_{1}$). Take \emph{A} and \emph{C} in Figure \ref{schematic}(c) as an example, $\Delta\tau_{AC}=t_{6}-t_{2}$.
\end{itemize}
The dyads' TIs have a non-decreasing time sequence if their frequency $n^{TI}_{ij}>1$. As shown in Figure \ref{schematic}(c), the TIs of \emph{A,C} have a non-decreasing active time sequence:[$t_2,t_3$],[$t_6,t_7$].
\indent
To summarize, as shown in Figure \ref{schematic}, event interactions are driven by the group access-related events, while temporal interactions are driven by the time sequences of dyad interactions. Each event interaction possesses an exclusive time interval on the WAP it takes place, whereas different temporal interactions may coexist with each other (e.g., $T_{AC}^{t_2}$ and $T_{BC}^{t_3}$ in Figure \ref{schematic}(c)).
\subsection*{Dynamics of event interaction}
Figure \ref{dynei}(a) shows that the probability distribution of EIs' active durations collected in 3 months is a truncated power law (the exponent of the power-law part approximates to 1), indicating that the long-lasting group interactions can hardly survive due to the frequent emergence of abundant short-active-duration events. In Figure \ref{dynei}(b), the number of event interactions with a given size (\emph{N(s)}) is exponential distributed, which further implies that the EIs having large sizes are rare. Therefore, it is natural to explore the dependence between the distribution of EIs' active durations $\Delta t^{EI}$ and a given size \emph{s}. Firstly, we study the impact of small sizes on EIs' active durations: the probability distributions of EIs' active durations $\Delta t^{EI}_{s}$ with size $s=2,3,4,5$ exhibit the truncated power-law behaviors, and their power-law parts present the similar exponents close to 1. However, their exponential cutoffs gradually decay with the growth of event size(Figure \ref{dynei}(c)). Besides, when the size of EIs is larger than 5, the influence of EIs' sizes to their active durations is subtle. Figure \ref{dynei}(d) reports the probability distributions of the EIs' active durations with a given size $s\geq 5$, which keeps almost invariant with the growth of size, i.e., the active durations of EIs are \emph{`size free'} in this case.
As shown in Figure \ref{dynei}, all EIs' active durations are less than 5 hours, which are much less than the circadian rhythms \cite{Circadian2007}. On the contrary, many transmission durations are obviously longer than 24 hours. In Figure \ref{dyntg}(a), the probability distribution of the transmission durations in the whole dataset falls into a bifold power-law with the turning point approximately equal to 24 hours, indicating the existence of circadian rhythms. We also calculate the integral days spent by each $\delta$ ([$\frac{\delta}{1440}$]) to examine whether some periodicity exists. In Figure \ref{dyntg}(b), the probability distribution of [$\frac{\delta}{1440}$] has two peaks at the 1st day and 7th day, i.e., the daily bursty behavior of successive event interactions coevolutes with weekly rhythms. We further perform the de-seasoning analysis to remove the circadian or weekly rhythms presented in Figure \ref{dyntg}(a,b) with two methods: I) natural de-seasoning method: only conserve the source and sink EIs taking place in the same day; II) artificial de-seasoning time-shuffled method in \cite{Temporalnetworks,Karsai11PRE,Jo12NJP,KivelaArxiv2012} to suit event interactions' definition (the details see \emph{Text SI}). As shown in Figure \ref{dyntg}(c), the probability distributions of the transmission durations 'filtered' by two de-seasoning methods both fall into the truncated power-laws, which indicates that the bursty behavior of event interactions takes place independently on the daily and weekly rhythms.
In addition, we examine the dependence of the source and sink EIs' active durations of each transmission path as shown in Table \ref{dependence}. Both the Pearson's coefficients and memory coefficients \cite{burstcoefficient} between them are very minor, indicating there does not exist the interdependence although the two groups of EIs share some individuals. From the definition of transmission graph, the transmission durations may be equal to the source EIs' active durations. However, their Pearson's coefficients and memory coefficients are also very small, hence the source EI' active duration is only the lower bound of the corresponding transmission duration, and there is no more dependence between them.
We construct the aggregated transmission graph (ATG) with the whole 3-month WiFi data set (the statistical properties of the ATG see Figure S1 and \emph{Text SI}). In Figure S1(a), most of the vertices only have small out-degrees ($k_{out}<10$) and in-degrees($k_{in}<10$), but there are 20 vertices with both high out-degrees and in-degrees to function as `super-connecting groups'(hubs) (see \emph{Text SI}). Recent studies reveal that super-connectors(hubs) in an aggregated network dominate the susceptible-infected-susceptible spreading processes due to the fact that they are self-sustained sources continually diffusing the virus to other neighbors \cite{SpreadingPRL2001,ABRMP,NewmanSIAM,PhysRep,AdvanInPhys,NewmanPRE2002,ThresholdPRL2010}. Therefore, the members of these `super-connecting groups' play a crucial role in the dynamics of disease transmission. However, traditional studies about the roles of individuals in the spreading processes focus on human contact network (for reviews see \cite{spreadingReview1,spreadingReview2}), where vertices are individuals and links are their contacts. We also aggregate the whole WiFi data set into a contact network(the detailed aggregation procedure refers to \emph{Text SI}) and find that the members of those `super-connecting groups' are `leaf' individuals in the contact network(red circle in Figure \ref{leafhub}(a)). Moreover, we further explore two other human indoor interaction data sets collected in a conference and an public exhibition\cite{RFID2010JTB}(see \emph{Methods and Materials}). Both these two high-resolution human face-to-face data sets(Figure \ref{leafhub}(b,c)) and our data set (Figure \ref{leafhub}(a)) illustrate that `leaf' individuals in the static human contact network gather to function as `hubs' in the transmission graph independently on different rendezvouses, indicating that the roles of many `leaf' individuals in a traditional static contact network may have been underestimated in a spreading process.
\subsection*{Dynamics of temporal interaction}
As shown in Figure \ref{dynti}(a), the probability distribution of TIs' active durations shows an exponential pattern(and the maximum value of $\Delta t^{TI}$ is less than 24 hours), which is different from the empirical results of human close proximity interactions recorded by Bluetooth \cite{EagleN2006PUC} or wireless sensors \cite{WirelessSensor2010PNAS}. This difference might derive from the fact that our data reflects the university daily teaching schedules and the human circadian rhythms, which constrain the individuals' resident periods in classrooms. Since the TIs' active durations can not reflect the population's interactive features beyond one day, we analyze the probability distribution $P(n^{TI})$ of TIs' frequencies in Figure \ref{dynti}(b). The outcome of power-law distribution indicates that although the number of the frequent contacting pairwise individuals is far less than that of the casual encounters, they can not be ignored.
The inter-event duration measures the time interval between two successive TIs of the same dyads (considering the start point of the two TIs). We also partition all the inter-event durations into two subsets according to the following criterion: whether or not two TIs of the same two users take place in the same day. As shown in Figure \ref{dynti}(c), the probability distribution of the inter-event durations of the successive TIs taking place in the same day exhibits a truncated power-law, which is similar to the findings in \cite{interContact2007,opportunistic2007IEEE,opportunistic2005,pocketSwitched2005,Eagle2007}. Moreover, Figure \ref{dynti}(d) reports the probability distribution of the inter-event duration of the successive TIs taking place in different days, where the red circles emphasize that there are many remarkably peaks indicating the weekly periodicity of the inter-event features. Therefore, the temporal interactions is in fact the convolution of burst behaviors on daily timescale and weekly rhythms.
To further explore the pairwise individuals' inter-event durations, we randomly choose two frequently interactive dyads. The two chosen series of temporal interactions are labeled as \emph{TI 1} and \emph{TI 2} with frequency $n^{TI}_1=56,n^{TI}_2=59$, respectively. In Figure \ref{spatiotemporal}(a), the time series \emph{TI 1} shows a typical burst characteristics, while the time series \emph{TI 2} presents more regular intervals between two active events. We analyze the cumulative probability distribution(CPD) of the inter-event durations of \emph{TI 1, TI 2} in Figure \ref{spatiotemporal}(b). The CPD of \emph{TI 1}'s inter-event durations falls into a power-law with the cutoff of one week (10,000 minutes), while the CPD of \emph{TI 2}'s inter-event durations presents an exponential form. In addition, there does not exist the dependence between TI's active durations and inter-event durations as shown in Figure S3, which are uncorrelated.
Since the addresses of WAPs are encoded in the temporal interactions of pairwise users, we define '\emph{spatial entropy}' to measure the spatial information of involved temporal interactions:
$$H_{ij}(n^{TI})=-\sum\limits_{\ell=1}\limits^{L}p_{\ell} \ln p_{\ell}$$ ($L \leq n^{TI}$ and $n^{TI} \geq 2$), where $p_{\ell}$ is the probability that the TIs of users $i,j$ simultaneously occur at the location '$\ell$' (equivalent to WAP '$\ell$'). The average spatial entropy with a given frequency is averaged as:
$$<H(n^{TI})>=\frac{\sum\limits_{ij\in \Psi } H_{ij}(n^{TI})}{N(n^{TI})}$$ where $\Psi$ represents the set of pairwise users with the same frequency ($n^{TI}$), and $N(n^{TI})$ is the size of $\Psi$.
Figure S4 reports the average spatial entropy $<H(n^{TI})>$ increases with the growth of frequency $n^{TI}$ at the population level. However, at the level of dyads, the TIs' spatial entropy are independent on their frequency $n^{TI}$, as illustrated by two aforementioned (\emph{TI 1, TI 2}) in Figure \ref{spatiotemporal}(a). We further observe the occurrence probability of \emph{TI 1, TI 2} in Figure \ref{spatiotemporal}(c). The low entropy of \emph{TI 2} ($H = 0.25$) indicates that two individuals have preferred interactive locations, and the predication of their future interactive location is possible. While the high entropy of \emph{TI 1}($H = 2.48$) means that the involved individuals randomly select their interactive locations. To summarize, we conclude that the pair of individuals with burst interactions (\emph{TI 1} in Figure \ref{spatiotemporal}(b)) randomly select their interactive locations (Figure \ref{spatiotemporal}(c)), and the pair of individuals with periodic interactions (\emph{TI 2}) present their preference on the interactive locations.
\section*{Materials and Methods}
\subsection*{Ethics Statement}
Although there is no official institutional review board (IRB) or ethics committee in Fudan University, this research achieves the approvement from the Informatization Office of Fudan University, which is the chief office responsible for the deployment of the WiFi wireless access point network and data collection at Fudan University. According to the local regulation rules of Fudan University `THE REGULATIONS FOR ADMINISTRATION OF FUDAN UNIVERSITY WIFI CAMPUS USERS' (available: http://www.xxb.fudan.edu.cn/s/52/t/104/42/e6/info17126.htm (Chinese Edition)), the part of the ninth provision, `Without licence, monitor other people's information with any monitor software is forbidden.', we have signed a confidentiality agreement and obtained the permission from the Informatization Office of Fudan University, the administrative unit of department level in Chinese Government. We declare that we have the responsibility to protect the privacy of the WiFi Campus users at Fudan University involved in this research, the data set involved is de-identified, and none privacy information of the WiFi users are available.
\subsection*{Study Setting}
The WiFi wireless access point(WAP) network involved in this study is deployed at Handan campus of Fudan University in Shanghai, China. All of the WAPs deployed in the Campus provide the dual-band(802.11g/n (2.4-GHz) and 802.11a (5-GHz)) free wireless accessing services to all the authorized campus members (e.g., students, teachers, office staffs and visiting scholars). These members' wireless electronic devices are generally automatically connected to the \emph{closest} WAP. Only when the WAP is overloaded, the device will be switched to another adjacent WAP. In the campus WiFi system of Fudan University, each WAP can serve around 50 users. Considering the percentage of students equipping wireless devices in the involved data set collected in the teaching buildings, it is sufficient to guarantee that generally every wireless device is served by the closest WAP. Furthermore, the number of devices co-served by one WAP is equal to the size of event interactions. As shown in Figure \ref{dynei}(b), the maximum number of co-served devices is less than 50, i.e., all the involved WAPs were not overloaded in the three months.
\subsection*{The Pre-processing of the Data Set}
The main data set used in our study is named as `FudanWiFi09' which records the users' access logs in the 2009-2010 fall semester (18/10/2009-9/1/2010). It contains the hardware information of users' electronic devices (the Media Access Control address, CPU trademark, et al.), the hardware information of the connected WAP, and the connecting/disconnecting time as well. Since the Media Access Control address, as the unique serial number, can uniquely identify different electronic devices and WAPs, we conserve them with the connecting/disconnecting time in `FudanWiFi09'.
From the statistics of Fudan University (http://www.fudan.edu.cn/new\_genview/genview.htm), there are totally 33,239 students, teachers and staffs in the campus, and `FudanWiFi09' contains the records of 22,050 individuals' 423,422 behavioral trajectories in the three-month period. Since we target the proxy data of those wireless devices to feature the users' indoor interactive behaviors, we only focus our research on the WiFi data recorded in those buildings which are open to public without safeguards for the WiFi users' devices (i.e., when a user leaves such a place, he/she should bring the device with himself/herself.) Therefore, the dataset of 18,715 individuals' 262,109 behavioral trajectories from all six teaching buildings with the WiFi coverage have been employed in the work of this paper.
\subsection*{Two Public Data Sets Used in the Study}
Two public data sets used in the study are both collected by the RFID technology available at the website of SocioPatterns (http://www.sociopatterns.org), and a related research refers to \cite{RFID2010JTB}. The first data set `HT09' (http://www.sociopatterns.org/datasets/infectious-sociopatterns-dynamic-contact-networks/) was collected during the ACM Hypertext 2009 conference, hosted by the Institute for Scientific Inter-change Foundation in Turin, Italy, from June 29th to July 1st. It involved about 100 conference participants to wear radio badges which monitored around 10,000 face-to-face interactions in the conference period of three days. The next data set `HGInfectious' (http://www.sociopatterns.org/datasets/hypertext-2009-dynamic-contact-network/) contains more than 230,000 face-to-face contacts among more than 14,000 visitors, which was collected during the art-science exhibition `INFECTIOUS: STAY AWAY' at the Science Gallery of Dublin, Ireland, from April 17th to July 17th, 2009.
These two data sets both have the similar tab-separated lists representing the active contacts during 20-second intervals of the data collection. Each line has the form `$t$ $i$ $j$', where $i$ and $j$ are the anonymous IDs of the persons in contact, and the interval, during which this contact was active, is $[t-20s, t]$. In the study, the two data sets are pre-processed by assembling the continues contacts of the same pair into one interaction, before translating them into the defined event interactions as shown in Figure \ref{schematic}.
\subsection*{The Average Spatial Distances of Two Users}
In the campus WiFi system of Fudan University, each WAP can serve around 50 users, which we conjecture it is sufficient to guarantee that every device is served by the closest WAP. Besides, the WAPs in adjacent floors have none overlapped covered regions because the building materials can dramatically attenuate the wireless signals, and each floor can be decomposed by the WAPs into the corresponding Voronoi tessellations. However, it is impossible to precisely locate the position of a user within a Voronoi \cite{Gonzalez2008Nature}, and the spatial distance between any two users inside one Voronoi cell can not be directly calculated. Therefore, we propose an approximative method to estimate the average spatial distance of any two users in Voronoi cells (The details see \emph{Text SI}), and Table \ref{distance} shows the average spatial distances of any two users and their standard derivation in all involved teaching buildings.
\section*{Acknowledgments}
We gratefully acknowledge the insightful comments from Prof. Alain Barrat and two anonymous reviewers, and the assistance from Bochun Wu, and the Informatization Office of Fudan University for the WiFi Data collection, and the Archives of Fudan University for providing the blueprint of all the teaching buildings. This work was partly supported by the National Key Basic Research and Development Program (No.2010CB731403), the NCET program (No.NCET-09-0317), and the National Natural Science Foundation of China.
| b86c34c42c0a3d069b46653263bece3fbc3114bf | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In the past decade, optical lattice clocks \cite{Ludlow2015,Katori2011} have made dramatic progress in accuracy and stability, surpassing their microwave counterparts to have the lowest fractional uncertainty of any frequency standard to date. Ensuring the continuation of this progress demands that the environmental perturbations affecting their accuracy are characterized to increasingly precise levels. Motivated by this challenge, we report on a new method using highly excited Rydberg states to provide an \textit{in situ} measurement of the DC electric field.
Uncharacterized electric fields can severely impact the accuracy of an atomic clock. For the $5s^2$~$^1$S$_0-5s5p$~$^3$P$_0$ clock transition in strontium, an electric field of 570 V/m yields a DC Stark shift of 1 Hz \cite{Middelmann2012}, or $2\times10^{-15}$ in fractional units, some three orders of magnitude above the lowest estimated total inaccuracy of a strontium optical lattice clock \cite{Nicholson2015}. Where dielectric surfaces are close to the atoms, shifts as large as 1$\times 10^{-13}$ have been observed \cite{Lodewyck2012}. While steps can be taken to reduce the residual electric field seen by the reference atoms, such as Faraday shielding \cite{Beloy2014}, or UV discharge of dielectric surfaces \cite{Pollack2010}, a characterization of the remaining field is necessary. This is typically done by direct spectroscopy of the clock transition with an externally applied electric field. With no residual electric field present, the quadratic nature of the perturbation implies the resulting induced frequency shift should be unchanged if the polarity of the applied field is reversed but the magnitude is left unchanged \cite{Matveev2011}. However, this method relies on the ability to apply sufficiently large and stable electric fields at the position of the atoms to induce a shift large enough to be quickly resolved during operation of the clock. For metallic vacuum chambers with minimal dielectric openings and no internal electrodes, producing such a shift is problematic. Furthermore, the applied field can charge dielectric materials such as the vacuum viewports \cite{Abel2011}, resulting in a time dependence of the effective applied field.
We circumvent these challenges by performing \emph{in-situ} electrometry using electromagnetically induced transparency (EIT) spectroscopy \cite{Fleischhauer2005} to measure the quadratic Stark shift of the Sr $5s75d$~$^1D_2$~$m_J = 0,\pm1,\pm2$ Rydberg states. Rydberg states of alkaline earth atoms are of growing interest for applications in quantum information \cite{Daley2008} and many body physics \cite{Mukherjee2011}, motivating their study by several groups \cite{Millen2010,DeSalvo2016}. The low-frequency polarizability scales with principal quantum number $n$ as $n^7$, making Rydberg states well suited for AC \cite{Sedlacek2012,Holloway2014,Fan2015} and DC \cite{Osterwalder1999, Thiele2015, Doughty1984} electrometry, with EIT spectroscopy being a particularly convenient measurement technique \cite{Abel2011, Mohapatra2007, Mohapatra2008,Tauschinsky2010}. The polarizability of our chosen Rydberg state is eight orders of magnitude larger than that of the clock transition, which reduces the required spectroscopic resolution from sub Hz, as needed when using to clock transition, to MHz when using Rydberg states to achieve the desired level of inaccuracy. It has also been proposed to use Rydberg states to measure ambient black-body radiation \cite{Ovsiannikov2011} which is responsible for the leading systematic uncertainty in many current Sr lattice clocks \cite{LeTargat2013, Falke2014, Poli2014, Takamoto2005}.
Using this spectroscopic method, we reduce the fractional uncertainty of the DC Stark shift of the clock transition to $2\times10^{-20}$. Furthermore the formation of Rydberg states in a system designed for the operation as an atomic clock opens the possibility to investigate proposals to use long range Rydberg interactions to generate squeezed states which exhibit reduced quantum projection noise \cite{Gil2014}.
\section{Theory: Single electron model and Stark maps }
\label{theory}
Alongside their large polarizability, another key advantage of Rydberg states for precision electrometry is that their Stark map - the variation of energy levels with the applied electric field - may be calculated to a very high degree of accuracy \cite{Zimmerman1979}. Even in divalent atoms such as strontium, where inter-electronic Coulomb interactions lead to perturbations of the Rydberg states \cite{Gallagher1994}, it can be shown that accurate wavefunctions \cite{Vaillant2012,Vaillant2014,Ye2012}, and Stark maps \cite{Millen2011,Lochead2013,Hiller2014} can be obtained without recourse to the complex atomic structure calculations required for the clock states \cite{Safronova2013}.
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{figure_theoryV2.pdf}
\caption{\label{fig:fig1} (a) Predicted Stark shift of the $5s47d$~$^1$D$_2$ $m_J$ = 0 state compared to experimental data (black dots) without any adjustable parameters. (b) Predicted Stark shift of the three $\vert m_J \vert$ components of the $5s75d$~$^1$D$_2$ state. The shaded area represents the uncertainty arising from the experimental determination of the zero-field energy (see text).}
\end{figure}
This simplification occurs because for Rydberg states, the effect of interelectronic interactions occurs primarily through the existence of spatially compact doubly-excited perturber states that overlap in energy with the Rydberg manifold. However since the static polarizability is dominated by the long-range character of the wavefunction \cite{Vaillant2015}, these states do not significantly alter the Stark maps. In previous work we have shown that an effective one-electron treatment that neglects inter-electronic effects gives Stark maps that are agreement with measurements for high-lying strontium Rydberg states \cite{Millen2011}.
Here we develop this approach to calculate Stark maps with the well-characterized uncertainty necessary to constrain the instability due to the electric field. The method is based on analytic expressions for the wavefunctions and dipole matrix elements generated using the Coulomb approximation \cite{Zimmerman1979}. The wavefunction is parametrized by a quantum defect, which is obtained by fitting to the experimentally measured zero-field energies.
To obtain the Stark map, Rydberg states within a range of $[n-3,n+5]$ from the target states and with $l \in [0,15]$ are included in the Stark Hamiltonian, which is then diagonalised numerically for each value of the field. At low electric field, the Stark shift of non-degenerate states is approximately quadratic, and a fit of the form $\Delta_\mathrm{E} = \frac{1}{2}\alpha_0 E^2$ yields the static polarizability $\alpha_0$.
In Fig.\ref{fig:fig1}(a) we show an example Stark map compared to experimental measurements taken in an atomic beam apparatus with a well-defined electrode geometry \cite{Hanley2017}. The data and the model are in quantitative agreement without any adjustable parameters. The predicted Stark map for the higher-lying state used to constrain the field in the lattice clock is shown in Fig.\ref{fig:fig1}(b). In both cases, the shaded band indicates the theoretical uncertainty in the Stark map. By far the dominant contribution to this uncertainty is the experimental uncertainty in the zero-field energies used to calculate the wavefunctions. In strontium, the current state-of-the-art absolute frequency measurements of Rydberg energy levels is $\pm30$ MHz for $S$ and $D$ states \cite{Beigang1982}, with measurements on the other series having much greater errors \cite{Rubbmark1978}. The zero-field energies for each series and the corresponding errors are obtained by fitting these experimental data with the Rydberg-Ritz formula \cite{Vaillant2012}. The shaded region corresponds to the extremal cases where the 1-sigma errors on each series are combined to give the extremal overall polarizabilities. Using ultracold atoms and frequency comb technology, it was recently shown that absolute Rydberg spectroscopy with 10 kHz uncertainty is possible \cite{Kliese2016}, opening the way to significant improvement in the Stark map uncertainty.
\section{ Experimental Approach }
We follow the standard approach for producing cold strontium samples \cite{katori1999, Loftus2004, Sorrentino2006}. We operate a blue magneto-optical trap (MOT) on the 461 nm transition for 650 ms followed by a broadband and single-frequency red MOT for 100 ms and 150 ms, respectively, which results in a sample of approximately $10^5$ $^{88}$Sr atoms at a temperature of around 1 $\upmu$K. Details of our apparatus can be found in \cite{Hill2014} and \cite{Hill2016}. Before implementing the EIT probe pulse, the cloud of atoms is released from the red MOT for 5 ms, giving time for the magnetic field to settle to the desired bias value, and for the atoms to expand ballistically to a lower density which was observed to improve the signal to noise level.
For the implementation of the EIT spectroscopy counter-propagating beams, one resonant with the $5s^2$~$^1$S$_0-5s5p$~$^1$P$_1$ transition at 461~nm and the other with tunable frequency at 413 nm, excite atoms to a chosen 5$snd$~$^1$D$_2$ Rydberg state. The resulting EIT signal is measured using `lock in' detection of the 461~nm probe beam absorption via modulation of the 413 nm pump beam intensity by an optical chopper. A typical absorption measurement showing the modulated EIT signal induced by the pump beam is shown in the top of Fig.\ref{fig:timetrace}. The probe beam is derived from a commercial frequency doubled diode laser system. Its power and waist, as defined by the 1/$e^2$ radius of the intensity profile, are 800~fW and 120 $\upmu$m, respectively. At this power and atomic number, the probe beam absorption is between $20 - 40\%$. A home built extended-cavity diode laser (ECDL) provides 8 mW of pump light which is focused to an 80 $\upmu$m waist at the atoms' position. The bottom inset in Fig.\ref{fig:timetrace} shows typical spectra taken at zero magnetic field with and without the applied external electric field.
The long term frequency stability of the pump and probe beam is maintained to within 10 kHz by locking to a transfer cavity referenced to the `clock' laser. In the case of the probe beam, the sub-harmonic at 922 nm is directly locked to the cavity. To stabilize the 413~nm pump laser frequency, a commercial Ti-sapphire laser, which is typically used to form the magic wavelength lattice at 813~nm, is first tuned to 826 nm and locked to the transfer cavity. This light is then frequency doubled by a LBO crystal \footnote{Purchased from Eksma Optics, Optolita UAB Mokslininku str. 11 LT-08412 Vilnius Lithuania} to produce 20 $\upmu$W as needed to generate a beat-note with the pump beam. The beat-note signal is mixed with a direct digital synthesiser (DDS) and the intermediate frequency is stabilized using a delay line offset lock scheme \cite{Schunemann1999} via fast feedback to the diode current and slow feedback to the ECDL piezo. The DDS provides the necessary tunability needed for scanning the pump laser frequency.
In order to spectroscopically resolve the Zeeman sub-levels of the Rydberg state, an external magnetic field between 100-300 $\upmu$T is applied orthogonal to the propagation direction of the pump and probe beams. In this low field regime, the Zeeman splitting of the intermediate $5s5p$~$^1$P$_1$ state is negligible compared to its natural linewidth. The probe beam is linearly polarised orthogonal to the quantization axis to enable a balanced access to all the $m_J$ levels within the Rydberg manifold given the fixed polarisation of the pump light. Any background residual magnetic field is nulled using electron-shelving spectroscopy on the narrow-linewidth $5s^2$~$^1$S$_0-5s5p$~$^3$P$_1$ transition at 689 nm \cite{Akatsuka2008}. We resolve a Doppler-broadened linewidth of 40 kHz which constrains any residual field to below 2 $\upmu$T.
To test the sensitivity of our method, an external plate electrode located directly opposite a radial DN40 viewport is used to apply a DC electric field in order to induce a Stark shift of the Rydberg states. Shielding from the metal vacuum chamber greatly attenuates the applied field at the atoms' position, meaning several kV potentials are needed to induce a substantial Stark shift. Such large potentials have the unfortunate effect of charging the dielectric viewport resulting in an exponential decay of the applied electric field strength at the atoms' position as inferred from a reduction in the Stark shift with time. To ameliorate this effect, we interleave measurements with opposite field polarity to avoid charging any external surfaces.
\begin{figure}[h]
\centering
\includegraphics[width=0.47\textwidth]{EIT_time_trace_and_spectra_inkscape_mod.pdf}
\caption{\label{fig:timetrace} \textit{top:} Modulation of the pump beam intensity (green) and its corresponding modulated EIT induced on the probe absorption signal (blue). The pump beam is resonant with the $n=53$ Rydberg state and both the external magnetic and electric fields set to zero. \textit{bottom:} An EIT spectrum is obtained by scanning the pump beam frequency and repeating such an absorption measurement. Example spectra are shown for zero applied magnetic field, with the applied electric field on (grey) and off (blue). }
\end{figure}
\section{ Rydberg Electrometry using EIT Spectroscopy }
As we do not have a method to directly measure the pump and probe frequencies, we have instead developed a method for measuring the applied electric fields that is based on the relative splitting of spectral lines, rather than the absolute detuning. In the absence of an electric field, the Zeeman sublevels split symmetrically with applied magnetic field \textbf{\textsl{B}}. However, in the presence of an electric field \textbf{\textsl{E}}, the Stark shift will result in an asymmetry in the spectrum, since the Stark shift depends on $|m_\mathrm{J}|$. Example spectra with and without an applied electric field are shown in Fig.\ref{fig:Efit}. In order to extract the line centers, five Fano profiles are fit to each spectrum corresponding to each of the five possible $m_J$ transitions. The asymmetric effect of an applied electric field on the observed Zeeman shift of the $m_\mathrm{J}$ is clearly visible in Fig.\ref{fig:Efit}.
To obtain the electric field from the spectroscopic data, the relative line positions are compared to a calculation of the combined Zeeman and Stark shift of each level. In the general case, the magnetic and electric field vectors are separated by an angle $\beta$, requiring transformation to a common basis. Choosing to work in the $\ket{J, m}$ basis defined by the magnetic field quantization axis, the matrix elements of the Zeeman Hamiltonian are give by
\begin{equation}
\bra{J, m_1}H_B\ket{J,m_2}=-m_1\upmu B \delta_{m_1m_2}
\end{equation}
\noindent where $\mu$, the magnitude of the magnetic dipole moment, is the Bohr magneton for a singlet state and $\delta$ is the Kronecker delta function. The Stark Hamiltonian, with eigenenergies $\Delta_E(m_J, E)$ that are computed as outlined in section \ref{theory}, is rotated by an angle $\beta$ by applying the appropriate Wigner D-matrix, $d^J_{m,m'}(\beta)$ for $J$ = 2. The matrix elements of this transformed Hamiltonian are given by
\begin{equation}
\bra{J, m_1}H_E\ket{J,m_2}= \sum_{m'} d^J_{m_1,m'}(\beta)d^J_{m_2,m'}(\beta)\Delta_E(m', E)
\end{equation}
\noindent Finally, the theoretical splitting is computed by diagonalizing the Hamiltonian $H = H_{\textrm{E}} + H_{\textrm{B}}$. Using this approach, we fit the experimentally observed energy splitting by varying the electric field strength \textit{E} and its angle $\beta$ relative to the applied magnetic field in the model. As the Stark shift is quadratic, our method only determines $\beta$ modulo $\pi$. The only other fitting parameter is an overall two photon detuning from the zero field resonance as we have no measure of the absolute frequency of the 413~nm laser.
Fig.\ref{fig:Efit} shows the relative energy splitting for various magnetic fields with and without an applied electric field. An external electrode set to 2 kV, the maximum allowed by the high voltage supply, generated the applied electric field. From a fit to the Zeeman splitting, the electric field at the position of the atoms is estimated to be $5.75\pm 0.11(\textrm{stat})\pm 0.16 (\textrm{sys}) \textrm{V} \textrm{m}^{-1}$. The fitting procedure also returned a value of $\beta = 0.47(1)\pi$ that is consistent with the axial magnetic field and radially applied electric field. An electric field of such magnitude would result in a fractional frequency shift of the clock transition equal to $2\times10^{-19}$. Given this is the largest field we can apply, it would have taken approximately a year of continuous operation to resolve this frequency shift given our fractional frequency instability, highlighting the utility of this method when applying large external fields is not possible.
Next the external field was switched off and the procedure was repeated. A fit to the resulting splitting revealed a residual electric field of
$1.52^{+0.62(\textrm{stat})}_{-.22}$$^{+0.05(\textrm{sys})}_{-.03} \textrm{V} \textrm{m}^{-1}$ most likely due to patch potential on the surrounding chamber. The uncertainty for this electric field value is comprised of both statistical error resulting from the fitting procedure and systematic error arising the uncertainty in the Stark map for $75^1$D$_2$. The quoted statistical error corresponds to a $68\%$ confidence interval as determined by the fitting procedure. A weak correlation observed between the uncertainty in $\beta$ and the electric field is taken into account in this estimate \cite{Press1988}, see appendix \ref{fitting} for further details. The systematic error on the electric field value due to the uncertainty for the Rydberg polarizability was calculated by repeating the fitting procedure with revised Stark maps offset from the theoretically predicted value by $\pm\sigma$. Translating this electric field and corresponding uncertainty to the DC Stark shift of the $^1$S$_0$-$^3$P$_0$ clock transition results in a fractional frequency shift of $-1.6^{+0.4}_{-1.6}\times10^{-20}$. The fractional uncertainty of the differential polarizability of the clock states is negligible compared to that of the electric field and therefore has been ignored in the quoted uncertainty.
\begin{figure}[h]
\centering
\includegraphics[width=0.48\textwidth]{Peak_Location_vs_E_field_with_spectra.pdf}
\caption{Inset $a$ shows the Zeeman splitting without an applied electric field for the Rydberg state n = 75 (the error bars are smaller than the symbols). The line centers are extracted from the EIT spectra, an example of which is shown in inset $c$ for the highest field case. Even without the an applied field, a fit to the splitting reveals a slight asymmetry resulting from the tensor nature of the Stark shift consistent with a residual electric field of 1.52~$\textrm{V} \textrm{m}^{-1}$. For the applied field case, shown in insets $(b)$ and $(d)$, the Stark shift is clearly visible and the fitting procedure returns an electric field of 5.75$~\textrm{V} \textrm{m}^{-1}$}
\label{fig:Efit}
\end{figure}
\section{Conclusion}
In conclusion, we believe that Rydberg electrometry constitutes a valuable technique for controlling systematic errors in optical lattice clocks. DC Stark shifts can in principle be separated from other systematic uncertainties using measurements on the clock transition alone, but this is time-consuming and requires the application of well-characterized external electric fields. In contrast, Rydberg states selectively enhance the spectroscopic sensitivity to stray electric fields by several orders of magnitude. The high spectroscopic resolution provided by EIT thus enables rapid quantitative measurements of the stray electric field. On the practical side, all that is required is a single additional laser to provide the pump beam, and existing lattice clock setups need not be modified to include electrodes. The constraint on the clock uncertainty that we obtained is compatible with the accuracy of the current generation of lattice clocks, and improved spectroscopy of the relevant Rydberg levels would see this reduced to negligible levels. Lastly, we note that the combination of Rydberg states and optical lattice clocks could also be applied to measurements of blackbody-induced systematic errors, and the creation of non-classical states.
\section{Acknowledgments}
The authors would like to thank Marco Schioppo for careful reading of our manuscript and Elizabeth Bridge for useful discussion regarding the feasibility of strontium Rydberg atoms for electrometry. This work was done under the auspices of UK NMO program. WB would like to acknowledge the support from the Marie Curie Initial Training Network FACT. PH and MPAJ acknowledge C Vaillant for useful discussions and the early version of the Stark map code. The experimental Stark map in Fig 1 was measured by R Hanley. They also acknowledge financial support from EPSRC grant
EP/J007021/, and EU grants FP7-ICT-2013-612862-HAIRS, H2020- FETPROACT- 2014-640378-RYSQ and H2020-MSCA-IF-2014-660028.
| 6b7da8d213c63f87e7c528f98ae6ca659fa5b903 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\subsection{\resolved\ channel}
\label{sec:backgrounds_resolved}
The $\Vjet$ background prediction is MC-based, but data-driven corrections are applied
to the MC prediction in order to improve the description of the jet kinematics.
A $\Vjet$ control region (CR) is defined identically to the signal region, except
that the region $65~\gev<\mjjless<95~\gev$ is vetoed, in order to remove most of the signal events.
One-dimensional reweighting functions of the variables $\pT(\jet_1)$ and $\Delta\phi(\jet\jet)$
are derived from this $\Vjet$ CR.
These reweighting functions have approximately $10\%$ effects on the shapes of the $\pT(\jet_1)$ and $\Delta\phi(\jet\jet)$
distributions. Data--MC comparisons in the $\Vjet$ CR are shown in Figure~\ref{fig:prefit_dataMC_resolved_reweighting},
before and after application of the reweighting functions.
All further results in this paper are shown with these two reweighting functions applied to the $\Vjet$
MC samples. The same reweighting functions are used for both the $\Wjet$ and $\Zjet$ processes.
It was checked that the reweighting functions obtained from the low-$\mjjless$ and high-$\mjjless$
portions of the $\Vjet$ control region are compatible.
\begin{figure}[htbp]
\begin{center}
\subfigure[]{
\includegraphics[width=0.47\textwidth,keepaspectratio]{prefit/pTJ1NoWeight.pdf}
}
\subfigure[]{
\includegraphics[width=0.47\textwidth,keepaspectratio]{prefit/dPhijjNoWeight.pdf}
}
\caption{
Comparisons between the data and the prediction in the $\Vjet$ control region of the \resolved\ channel.
The top panel shows the data and prediction
before applying the $\pT(\jet_1)$ and $\Delta\phi(\jet_1,\jet_2)$ kinematic reweighting
to the $\Vjet$ predictions.
The distributions shown are
(a) $\pT$ of the leading jet and
(b) $\Delta \phi$ between the leading jet and sub-leading jet.
Overflow is included in the last bin of the $\pT(\jet_1)$ plot.
The bottom panel shows the ratio of the SM prediction to the data before and after
applying the kinematic reweighting to the $\Vjet$ prediction.
The hatched bands indicate the statistical uncertainty in the predictions.
}
\label{fig:prefit_dataMC_resolved_reweighting}
\end{center}
\end{figure}
The top-quark background is modelled with MC simulation, and is cross-checked in a
validation region containing three small-$R$ jets, one of which is
$b$-tagged using the MV1 algorithm~\cite{ATLAS-CONF-2014-046,Aad:2015ydr}.
Good agreement is observed between the data and the MC simulation, so no corrections are applied to the prediction.
The background from $ZZ$ events is also modelled with MC simulation.
The data-driven multijet background estimate makes use of a multijet CR. The multijet CR
is formed by selecting events in data that pass the same selection requirements as for the signal
region, except that the lepton quality criteria are modified in order to produce a CR enriched
in non-prompt and fake leptons. Lepton candidates satisfying these modified criteria are called
``anti-identified'' lepton candidates. Anti-identified muon candidates must have a non-negligible impact
parameter, $\dzsig>4$, and satisfy looser isolation criteria than the signal muon candidates.
Anti-identified electrons must fail the ``tight'' but satisfy the ``medium'' cut-based
identification criteria from Ref.~\cite{Aaboud:2016vfy}, and are also
required to contain a hit in the
innermost layer of the pixel detector. In addition, the isolation criteria are modified
for anti-identified electron candidates, in order to enrich the sample in non-prompt and fake electrons.
The shapes of the kinematic distributions (such as $\mjj$, $\met$, and $\pTjj$) of the
multijet background are estimated from events in the multijet CR, after
subtracting the MC predictions of the non-multijet contributions to the CR. These non-multijet
contributions are about 20\% (50\%) of the total in the electron (muon) channel.
The overall multijet background
event yield is estimated from a fit to the $\met$ distribution of events that
pass the full signal region selection, except that the requirements on $\met$ and
$\Delta\phi(\jet_1,\MET)$ (and also $\Delta\eta(\jet,\jet)$ and $\MT$ for the muon channel)
are removed in order to enhance the number of multijet events. This selection is referred to
as the \textit{extended} signal region. In this $\met$ fit, the multijet $\met$ shape is estimated from an
extended multijet CR, defined analogously to the extended signal region, but requiring the lepton
to pass the anti-identified-lepton
selection. The $\met$ shapes of the other backgrounds are estimated using MC samples.
The multijet event yield obtained from this fit is then extrapolated to the signal region, using the ratio
of events in the multijet CR and the extended multijet CR, corrected for non-multijet contributions.
The multijet background estimates are performed separately for the electron and muon
channels. Only about $5\%$ of the total multijet background is in the
muon channel.
\begin{figure}[htbp]
\begin{center}
\subfigure[]{
\label{fig:templates_normalized_resolved}
\includegraphics[width=0.47\textwidth,keepaspectratio]{templates_resolved.pdf}
}
\subfigure[]{
\label{fig:templates_normalized_boosted}
\includegraphics[width=0.47\textwidth,keepaspectratio]{templates_boosted.pdf}
}
\caption{
The shapes of (a) the predicted $\mjjless$ distributions in the \resolved\ signal region
and (b) the predicted $\mJ$ distributions in the \boosted\ signal region, for the signal
(peaked near 80~\gev) and various background processes. The distributions are normalized to unity.
}
\label{fig:templates_normalized}
\end{center}
\end{figure}
The expected signal and background yields in the \resolved\ signal region are
given in Table~\ref{tab:yields}, and compared to the number of events observed in data.
The predictions for the $\mjjless$ distribution shapes of the signal
and backgrounds are shown in Figure~\ref{fig:templates_normalized_resolved}.
\begin{table}[htbp]
\centering
\begin{tabular}{l!!}
\toprule
\midrule
& \multicolumn{1}{c}{\resolved} & \multicolumn{1}{c}{\boosted} \\
\midrule
\midrule
Signal\\
\midrule
$WW$ & 2860 , 110 & 542 , 61 \\
$WZ$ & 730 , 30 & 128 , 15 \\
Total Expected Signal & 3590 , 140 & 670 , 75 \\
\midrule
Background\\
\midrule
$\Wjet$ & 136000 , 8600 & 10500 , 1300 \\
$\Zjet$ & 2750 , 340 & 245 , 32 \\
$\ttbar$ & 12980 , 520 & 1130 , 150 \\
Single top-quark & 3620 , 150 & 249 , 35 \\
Multijet & 3689 , 60 & 313 , 18 \\
$ZZ$ & 14 , 1 & \multicolumn{1}{c}{-} \\
Total Expected Background & 159000 , 8600 & 12400 , 1500 \\
\midrule
Total SM Expected & 162600 , 8700 & 13100 , 1600 \\
\midrule
Observed & \multicolumn{1}{c}{$164502$} & \multicolumn{1}{c}{$12999$} \\
\midrule
$S/B$ ($65~\gev<\mjjless<95~\gev$) & \multicolumn{1}{c}{$5.5\%$} & \multicolumn{1}{c}{$10.1\%$} \\
$S/\sqrt{B}$ ($65~\gev<\mjjless<95~\gev$) & \multicolumn{1}{c}{$11.1$} & \multicolumn{1}{c}{$7.1$} \\
\bottomrule
\bottomrule
\end{tabular}
\caption{ Expected number of signal and background events in the \resolved\ and \boosted\ signal regions,
prior to performing the $\mjjless$ and $\mJ$ fits. The quoted uncertainties only include
detector-related uncertainties and statistical uncertainties of the MC
samples and control regions. The number of events observed in data is also shown.
The signal predictions only correspond to $\qq$-initiated $WV$ production.
}
\label{tab:yields}
\end{table}
\subsection{\boosted\ channel}
\label{sec:backgrounds_boosted}
In the \boosted\ channel, the $\Wjet$, $\Zjet$, and top-quark
backgrounds are estimated using MC samples. The MC predictions for the
two largest backgrounds ($\Wjet$ and top-quark production) are
corrected by scale factors obtained from dedicated control regions.
The top-quark control region (top CR) is formed by events satisfying the signal
region selection, except that the presence of at least one small-$R$ $b$-tagged jet with
$\pT>25~\gev$ and $\DR(\jet,\fatjet)>1.0$ is required instead of applying the nominal veto on small-R jets. The jets are $b$-tagged using the MV1
algorithm~\cite{ATLAS-CONF-2014-046,Aad:2015ydr}, using a working
point with a $b$-tagging efficiency of about 70\% and a gluon/light-quark jet rejection factor of over 100
in $\ttbar$ events.
About $90\%$ of the events in this top CR originate from top-quark backgrounds.
There is a deficit in data in the top CR relative to the MC prediction, which is attributed
to a mismodelling of the top-quark backgrounds. A global scale factor of $\topSFboost$ for the top-quark
backgrounds is obtained from this CR, after subtracting the prediction for non-top-quark backgrounds.
The data in the top CR is shown in Figure~\ref{fig:boostedTopCR_mJ}, compared
to the SM prediction after application of the top-quark scale factor.
This scale factor is applied to
the top-quark background predictions in the signal region.
The control region for the $\Wjet$ background ($\Wjet$ CR) is obtained
by applying the standard signal region selection, but adding the requirement that
$\mJ<65~\gev$ or $\mJ>95~\gev$.
This additional $\mJ$ requirement removes almost all of the $WV$ signal events and
also a large fraction of the top-quark events.
About $85\%$ of the events in this CR originate from $\Wjet$ backgrounds.
The top-quark background prediction in the $\Wjet$ CR is scaled by the top-quark scale factor obtained above.
A data deficit is observed in the $\Wjet$ CR relative to the prediction. A global
scale factor of $\wjetSFboost$ is obtained for the $\Wjet$ background, after subtracting
the expected contributions from the other signal/background processes.
A comparison between the data and the prediction in the $\Wjet$ CR is shown in Figure~\ref{fig:boostedWjetCR_pTJ},
after application of the $\Wjet$
scale factor. The $\Wjet$ scale factor is applied to the $\Wjet$ prediction in the signal region.
The method for estimating the multijet background is similar to that used
in the \resolved\ channel. As in the \resolved\ channel, a multijet
CR is defined by requiring an ``anti-identified'' lepton candidate.
The shapes of the kinematic distributions are estimated from this CR using the same
method as in the \resolved\ channel. The non-multijet background contributions to the CR are about 6\%
of the total. The multijet event yield
is estimated from a fit to the $\met$ distribution, as in the \resolved\ channel, but
the only requirement that is removed for the definiton of the extended signal region/multijet CR
is the $\met>50~\gev$ requirement.
The multijet background is found to be negligible for the muon channel, so only the
contribution in the electron channel is considered for the final results.
The numbers of expected and observed events in the \boosted\ signal region
are summarized in Table~\ref{tab:yields}. The previously mentioned top-quark and $\Wjet$
scale factors are applied to the predictions. The contribution from $ZZ$ events is
expected to be very small in the \boosted\ channel, so it is neglected.
The nominal predictions for the $\mJ$ distribution shapes of the signal
and backgrounds are shown in Figure~\ref{fig:templates_normalized_boosted}.
\begin{figure}[htbp]
\begin{center}
\subfigure[]{
\label{fig:boostedTopCR_mJ}
\includegraphics[width=0.47\textwidth,keepaspectratio]{figures/akt10_M_topCR.pdf}
}
\subfigure[]{
\label{fig:boostedWjetCR_pTJ}
\includegraphics[width=0.47\linewidth,keepaspectratio]{figures/akt10_Pt_wjetsCR.pdf}
}
\caption{
Comparison between data and prediction in the \boosted\ channel for (a) $\mJ$ in the top CR, and (b) $\pTJ$
in the $\Wjet$ CR. A scale factor is applied to the top-quark background prediction in
the top CR and the $\Wjet$ CR, and a scale factor is applied to the $\Wjet$ background
prediction (which is part of the ``$\Vjet$'' histogram)
in the $\Wjet$ CR.
The hatched bands indicate the systematic uncertainty of the prediction.
For the $\Vjet$ component, only shape systematic uncertainties are included in the bands.
}
\label{fig:boostedCR}
\end{center}
\end{figure}
\subsection{\resolved\ fiducial phase space}
\label{sec:fiducialResolvedDef}
The \resolved\ fiducial phase space is defined to closely match the
experimental event selection. The phase-space definition requires
a $WV$ pair with the bosons decaying as $\Vtoqq$ and $\Wtolv$, where $\ell=e,\mu$. Events
containing other kinds of $WV$ decay channels (such as $\WWlep$ events or
$\WVtaunuqq$ with the $\tau$ decaying to $\ell+X$), are
not included in the fiducial phase-space definition. Such $WV$ events can still pass
the experimental event selection (where they are included in the signal
category), and they are accounted for in the $\Dfid$ definition.
Leptons selected in the fiducial region must have $\pT(\ell)>30~\gev$ and $|\eta(\ell)|<2.47$.
The four-momentum of the lepton is modified by adding to it the
four-momenta of all the photons within $\DR=0.1$, excluding
photons produced by hadron decays. Particle-level anti-$k_t$ $R=0.4$ jets are constructed
using as constituents all stable particles, excluding muons and neutrinos. Stable particles are
defined as those having a mean lifetime of $\tau>30$~ps.
The particle-level jets must have $\pT>25~\gev$ and $|\eta|<2.5$. Jets within
$\DR=0.2$ of a selected electron are rejected, and then leptons within $\DR=0.4$
of a remaining jet are rejected.
The true $\met$ in the event is defined as the magnitude of the vector $\pT$ sum of all
the neutrinos.
The event must have exactly one lepton and two $R=0.4$ jets matching the above definitions.
The remaining requirements for the fiducial phase space are summarized
in Table~\ref{tab:FVDefinition}, and are analogous to the experimental event selection,
but are defined using the lepton, $\met$, and particle-level jets described
in this section.
\begin{table}[tb]
\begin{center}
\begin{tabular}{l|c|c}
\hline \hline
& \resolved\ & \boosted\ \\
\hline
Lepton & \multicolumn{2}{c}{$N_{\ell}=1$ with $\pT>30~\gev$ and $|\eta|<2.47$} \\
& \multicolumn{2}{c}{ $\DR(\ell,\jet)>0.4$ } \\
\hline
$\Wtolv$ & $\pT(\ell\nu)>100~\gev$ & $-$ \\
& $\MT>40~\gev$ & $-$ \\
\hline
\MET & $\MET > 40~\gev$ & $\MET > 50~\gev$ \\
\hline
Jet & $\njet=2$ with $\pT>25~\gev$, $|\eta|<2.5$, & $\nfatjet=1$ with $\pT>200~\gev$, $|\eta|<2.0$, \\
& $\DR(\jet,e)>0.2$ & $\DR(\fatjet,\ell)>1.0$ \\
& & No small-$R$ jets with $\pT>25~\gev$, $|\eta|<4.5$, \\
& & $\DR(\jet,\fatjet)>1.0$, $\DR(\jet,e)>0.2$ \\
& $40<\mjjless<200~\gev$ & $50<\mJ<170~\gev$ \\
& $\pTjj>100~\gev$ & $-$ \\
& $\Delta\eta(\jet,\jet)<1.5$ & $-$ \\
\hline
Global & $\Delta\phi(\jet_1,\MET)>0.8$ & $-$ \\
\hline\hline
\end{tabular}
\caption{
Summary of the fiducial phase-space definitions. All the specified selection criteria are applied at the particle level as specified in the text. The notations ``$\jet$'' and ``$\fatjet$'' refer to $R=0.4$ and $R=1.0$ jets,
respectively, as explained in the text.
}
\label{tab:FVDefinition}
\end{center}
\end{table}
\subsection{\boosted\ fiducial phase space}
\label{sec:fiducialBoostedDef}
As in the \resolved\ channel, the fiducial phase-space definition requires
a $WV$ pair with $\Vtoqq$ and $\Wtolv$. Leptons, \met, and particle-level $R=0.4$ jets
are defined in the same way as in the \resolved\ channel,
except that two
sets of leptons and small-$R$ jets are considered:
\textit{central} leptons (small-$R$ jets) are required to have $|\eta|<2.47$ ($|\eta|<2.5$),
and \textit{extended} leptons and small-$R$ jets are required to have $|\eta|<4.5$.
Particle-level large-$R$ jets are defined by applying
the anti-$k_t$ algorithm with radius parameter $R=1.0$ to all stable particles, excluding muons and neutrinos.
No trimming is applied to these jets. The
large-$R$ jets are required to have $\pT>200~\gev$ and $|\eta|<2.0$.
Central (extended) small-$R$ jets that are within
$\DR=0.2$ of a central (extended) electron are rejected.
Then, central leptons are rejected if they are within $\DR=0.4$
of a remaining central small-$R$ jet.
Large-$R$ jets are rejected if they are within $\DR=1.0$ of any remaining central leptons.
Events are required to contain exactly one central lepton and one large-$R$ jet with the above definitions,
and events are discarded if they contain any extended small-$R$ jets with $\DR(\jet,\fatjet)>1.0$.
The event must also have $\met>50~\gev$, and the large-$R$ jet must have a mass greater than $50~\gev$.
The fiducial phase-space definition is summarized
in Table~\ref{tab:FVDefinition}.
\section{Introduction}
\label{sec:intro}
\input{Intro}
\section{Analysis overview}
\label{sec:overview}
\input{AnalysisOverview}
\section{ATLAS detector}
\label{sec:detector}
\input{Detector}
\section{Data and Monte Carlo samples}
\label{sec:samples}
\input{Samples}
\section{Event reconstruction}
\label{sec:objselection}
\input{ObjectSelection}
\section{Event selection}
\label{sec:eventSelection}
\input{ResolvedAnalysis}
\section{Background estimation}
\label{sec:backgrounds}
\input{BoostedAnalysis}
\section{Cross-section extraction}
\label{sec:xsecMethod}
\input{CrossSection}
\section{Systematic uncertainties}
\label{sec:systematics}
\input{Systematics}
\section{Cross-section results}
\label{sec:xsecResult}
\input{CrossSectionResults}
\section{Constraints on anomalous gauge couplings}
\label{sec:aTGC}
\input{ATGC}
\FloatBarrier
\section{Conclusion}
\label{sec:conclusion}
\input{Conclusion}
\section*{Acknowledgements}
\input{acknowledgements/Acknowledgements}
\printbibliography
\newpage
\input{atlas_authlist}
\clearpage
\subsection{\resolved\ channel}
\label{eventSelectResolved}
Events must have $\MET>40~\gev$ and a
transverse mass\footnote{The transverse mass is defined as $\MT \equiv \sqrt{(\MET+\pT(\ell))^2 - |\vecmet+\vec{\pT}(\ell)|^2}$,
where $\pT(\ell)$ is the transverse momentum of the lepton candidate.} $\MT>40~\gev$.
Events must contain exactly two small-$R$ jets.
The requirement of exactly two jets substantially
reduces the background from top-quark decays. The pseudorapidity separation
of the selected jets is required to satisfy
$\Delta\eta(\jet,\jet)<1.5$, in order to improve the signal-to-background
ratio.
In order to reduce the multijet background
not removed by the $\met>40~\gev$ requirement,
an azimuthal-angle difference between the $\MET$ direction and the
direction of the leading-\pt jet of \linebreak $|\Delta\phi(\jet_1,\MET)| > 0.8$
is required.
Also, both the $\Vqq$ and $\Wlv$ candidates must pass requirements on
their transverse
momenta: $\pTjj>100~\gev$ and $\pTWlep>100~\gev$, where
$\pTWlep\equiv |\vecmet+\vec{\pT}(\ell)|$.
These $\pT$ requirements enhance the separation between the signal and background
distributions in the dijet mass.
As described in Section~\ref{sec:xsecMethod}, the signal is extracted
using a maximum-likelihood (ML) fit
to the dijet mass ($\mjjless$) distribution. In the dijet mass calculation,
the mass of each individual jet is set to zero, which makes the
variable easier to model in the MC simulation.
Since the signal is extracted from a fit to $\mjjless$, only a loose
requirement is made on this variable: $40~\gev<\mjjless<200~\gev$.
\subsection{\boosted\ channel}
\label{eventSelectBoosted}
Events must contain exactly one large-$R$ jet with $\pT>200~\gev$
and $|\eta|<2.0$.
The backgrounds from top-quark decays are suppressed by rejecting events containing
any small-$R$ jets with $\pT>25~\gev$ and $|\eta|<4.5$ that are separated
from the large-$R$ jet by $\DR(\jet,\fatjet)>1.0$.
In order to suppress the multijet background, a requirement
of $\met>50~\gev$ is applied.
The trimmed mass of the large-$R$ jet, $\mJ$, must
be $50~\gev<\mJ<170~\gev$, and the signal
is measured from the ML fit to $\mJ$.
Since the \resolved\ and \boosted\ event selections are done independently,
some events pass both selections.
About $10\%$ of the signal MC events that pass the \resolved\ selection also
pass the \boosted\ selection, while about $50\%$ of the signal MC
events that pass the \boosted\ selection
also pass the \resolved\ selection.
| 3d08a24f8afbeb6be1c787e73f9e638e2c385ed5 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{Sect.1}
In the mathematical modeling of the nano-size semiconductor devices (e.g. HEMTs, MOSFETs, RTDs and superlattice devices ), the quantum effects (like particle tunneling through potential barriers and particle buildup in quantum wells) take place and can not be simulated by classical hydrodynamic models. Therefore, the quantum hydrodynamical (QHD) equations are important and dominative in the description of the motion of electrons or holes transport under the self-consistent electric field.
The QHD conservation laws have the same form as the
classical hydrodynamic equations (for simplicity, we treat the flow of electrons in the self-consistent electric field for unipolar devices):
\begin{subequations}\label{qhdcl}
\begin{numcases}{}
\pd{t}{}n+\pd{x_k}{}j_k=0, \label{qhdcl1}\\
\pd{t}{}j_l+\pd{x_k}{}(u_kj_l-P_{kl})=n\pd{x_l}{}\phi-\frac{j_l}{\tau_m},\quad l=1,2,3, \label{qhdcl2}\\
\pd{t}{}e+\pd{x_k}{}(u_ke-u_lP_{kl}+q_k)=j_k\pd{x_k}{}\phi+C_e, \label{qhdcl3}\\
\lambda^2\Delta\phi=n-D(\bm{x}), \label{qhdcl4}
\end{numcases}
\end{subequations}
where $n>0$ is the electron density, $\bm{u}=(u_1,u_2,u_3)$ is the velocity, $\bm{j}=(j_1,j_2,j_3)$ is the momentum density, $\bm{P}=(P_{kl})$ is the stress tensor, $\phi$ is the self-consistent electrostatic potential, $e$ is the energy density, $\bm{q}=(q_1,q_2,q_3)$ is the heat flux. Indices $k,l$ equal $1,2,3$, and repeated indices are summed over using the Einstein convention. Equation \eqref{qhdcl1} expresses conservation of electron number, \eqref{qhdcl2} expresses conservation of momentum, and \eqref{qhdcl3} expresses conservation of energy. The last terms in \eqref{qhdcl2} and \eqref{qhdcl3} represent electron scattering (the collision terms may include the effects of electron-phonon and electron-impurity collisions, intervalley and interband scattering), which is modeled by the standard relaxation time approximation with momentum and energy relaxation times $\tau_m>0$ and $\tau_e>0$. The energy relaxation term $C_e$ is given by
\begin{equation*}
C_e=-\frac{1}{\tau_e}\Bigg(\frac{1}{2}n|\bm{u}|^2+\frac{3}{2}n(\theta-\theta_{L})\Bigg),
\end{equation*}
where $\theta>0$ is the electron temperature and $\theta_{L}>0$ is the temperature of the semiconductor lattice in energy units. The transport equations \eqref{qhdcl1}$\sim$\eqref{qhdcl3} are coupled to Poisson's equation \eqref{qhdcl4} for the self-consistent electrostatic potential, where $\lambda>0$ is the Debye length, $D=N_d-N_a$ is the doping profile, $N_d>0$ is the density of donors, and $N_a>0$ is the density of acceptors.
The QHD equations \eqref{qhdcl1}$\sim$\eqref{qhdcl3} are derived as a set of nonlinear conservation laws by a moment expansion of the Wigner-Boltzmann equation \cite{W32} and an expansion of the thermal equilibrium Wigner distribution function to $O(\pc^2)$, where $\pc>0$ is the scaled Planck constant. However, to close the moment expansion at the first three moments, we must define, for example, $\bm{j}$, $\bm{P}$, $e$ and $\bm{q}$ in terms of $n$, $\bm{u}$ and $\theta$. According to the closure assumption \cite{G94}, up to order
$O(\pc^2)$, we define the momentum density $\bm{j}$, the stress tensor $\bm{P}=(P_{kl})$, the energy density $e$ and the heat flux $\bm{q}$ as follows:
\begin{gather*}
\bm{j}=n\bm{u},\qquad P_{kl}=-n\theta\delta_{kl}+\frac{\pc^2}{2}n\pd{x_k}{}\pd{x_l}{}\ln n,\\
e=\frac{3}{2}n\theta+\frac{1}{2}n|\bm{u}|^2-\frac{\pc^2}{4}n\Delta\ln n,\qquad \bm{q}=-\kappa\nabla\theta-\frac{3\pc^2}{4}n\Delta\bm{u},
\end{gather*}
with the Kronecker symbol $\delta_{kl}$ and the heat conductivity $\kappa>0$. The quantum correction to the stress tensor was first stated in the semiconductor context by Ancona and Iafrate \cite{AI89} and Ancona and Tiersten \cite{AT87}. Since
\begin{equation*}
\frac{\pc^2}{2}\mathrm{div}(n(\nabla\otimes\nabla)\ln n)=\pc^2n\nabla\Bigg(\frac{\Delta\sqrt{n}}{\sqrt{n}}\Bigg),
\end{equation*}
it can be interpreted as a force including the Bohm potential $\pc^2\Delta\sqrt{n}/\sqrt{n}$ \cite{FZ93}. The quantum correction to the energy density was first derived by Wigner \cite{W32}. The heat conduction term consists of a classical Fourier law $-\kappa\nabla\theta$ plus a new quantum contribution $-3\pc^2n\Delta\bm{u}/4$ which can be interpreted as a dispersive heat flux \cite{G95,JMM06}. For details on the more general quantum models for semiconductor devices, one can refer to the references \cite{J01,MRS90,ZH16book}.
Interestingly, most quantum terms cancel out in the energy equation \eqref{qhdcl3}. In fact, by substituting the above expressions for $C_e$, $\bm{j}$, $\bm{P}$, $e$ and $\bm{q}$ into \eqref{qhdcl}, a computation yields the multi-dimensional full quantum hydrodynamic (FQHD) model for semiconductors as follows.
\begin{subequations}\label{fqhd}
\begin{numcases}{}
n_t+\mathrm{div}(n\bm{u})=0, \label{fqhd1}\\
(n\bm{u})_t+\mathrm{div}(n\bm{u}\otimes\bm{u})+\nabla(n\theta)-\pc^2n\nabla\Bigg(\frac{\Delta\sqrt{n}}{\sqrt{n}}\Bigg)=n\nabla\phi-\frac{n\bm{u}}{\tau_m}, \label{fqhd2}\\
n\theta_t+n\bm{u}\cdot\nabla\theta+\frac{2}{3}n\theta\mathrm{div}\bm{u}-\frac{2}{3}\mathrm{div}(\kappa\nabla\theta)\notag\\
\qquad\qquad\qquad\qquad\qquad\qquad\quad\,-\frac{\pc^2}{3}\mathrm{div}(n\Delta\bm{u})=\frac{2\tau_e-\tau_m}{3\tau_m\tau_e}n|\bm{u}|^2-\frac{n(\theta-\theta_{L})}{\tau_e}, \label{fqhd3}\\
\lambda^2\Delta\phi=n-D(\bm{x}). \label{fqhd4}
\end{numcases}
\end{subequations}
Comparing with the classical full hydrodynamic (FHD) model, the new feature of the FQHD model is the Bohm potential term
\begin{equation*}
-\pc^2n\nabla\Bigg(\frac{\Delta\sqrt{n}}{\sqrt{n}}\Bigg)
\end{equation*}
in the momentum equation \eqref{fqhd2} and the dispersive velocity term
\begin{equation*}
-\frac{\pc^2}{3}\mathrm{div}(n\Delta\bm{u})
\end{equation*}
in the energy equation \eqref{fqhd3}. Both of them are called quantum correction terms (or dispersive terms) and belong to the third-order derivative terms of the system \eqref{fqhd}.
Recently, the study concerning the semiconductor quantum models and the related quantum systems has become popular. J\"ungel and Li \cite{JL04,JL04-1} investigated the one-dimensional unipolar isentropic QHD model with the Dirichlet-Neumann boundary condition and the flat doping profile. The authors proved the existence, uniqueness and exponential stability of the subsonic stationary solution for the quite general pressure-density function. Nishibata and Suzuki \cite{NS08} reconsidered this QHD model with isothermal simplification and the vanishing bohmenian-type boundary condition. The authors generalized J\"ungel and Li's results to the non-flat doping profile case and also discussed the semi-classical limit for both stationary and global solutions. Hu, Mei and Zhang \cite{HMZ16} generalized Nishibata and Suzuki's results to the bipolar case with non-constant but flat doping profile. Huang, Li and Matsumura \cite{HLM06} proved the existence, exponential stability and semi-classical limit of stationary solution of Cauchy problem for the one-dimensional isentropic unipolar QHD model. Li and Yong \cite{LY17} studied the nonlinear diffusion phenomena on the Cauchy problem of the one-dimensional isentropic bipolar QHD model. The authors proved the algebraic stability of the diffusion waves.
In multi-dimensional case, J\"ungel \cite{J98} first considered the unipolar stationary isothermal and isentropic QHD model for potential flows on a bounded domain. The existence of solutions was proved under the assumption that the electric energy was small compared to the thermal energy, where Dirichlet boundary conditions were addressed. This result was then generalized to bipolar case by Liang and Zhang \cite{LZ07}. Unterreiter \cite{U97} proved the existence of the thermal equilibrium solution of the bipolar isentropic QHD model confined to a bounded domain by variational analysis, and the semi-classical limit is carried out recovering the minimizer of the limiting functional. This result recently was developed by Di Michele, Mei, Rubino and Sampalmieri \cite{MMRS17} to a new model of the bipolar isentropic hybrid quantum hydrodynamics. Regarding the unipolar QHD model for irrotational fluid in spatial periodic domain, the global existence of the dynamic solutions and the exponential convergence to their equilibria were artfully proved by Li and Marcati in \cite{LM04}. Remarkably, the weak solutions with large initial data for the quantum hydrodynamic system in multiple dimensions were further obtained by Antonelli and Marcati in \cite{AM09,AM12}. Li, Zhang and Zhang \cite{LZZ08} investigated the large-time behavior of solutions to the initial value problem of the isentropic QHD model in the whole space $\mathbb{R}^3$ and obtained the algebraic time-decay rate, and further showed in \cite{ZLZ08} the semi-classical and relaxation limits of global solutions. Recently, Pu and Guo \cite{PG16} studied the Cauchy problem of a quantum hydrodynamic equations with viscosity and heat conduction in the whole space $\mathbb{R}^3$. The global existence around a constant steady-state and semi-classical limit of the global solutions were shown by the energy method. This result was developed by Pu and Xu \cite{PX17}, the authors obtained the optimal convergence rates to the constant equilibrium solution by the pure energy method and negative Sobolev space estimates.
However, all of these research results more or less have some limitations from both physical and mathematical points of view. Actually, in practical applications, the semiconductor quantum models should be treated under the following physically motivated settings which make the mathematical analysis more difficult:
\begin{itemize}
\item[(1)] The model system should be considered on a bounded domain $\Omega$ and be supplemented by physical boundary conditions.
\item[(2)] In realistic semiconductor devices, the doping profile will be a non-flat function of the spatial variable. For instance, it has two steep slops in $n^+-n-n^+$ diodes \cite{G94}. Therefore, we should assume the continuity and positivity only to cover the actual devices. Namely,
\begin{equation}\label{nonflat}
D\in C(\overline{\Omega}),\qquad \inf_{x\in\overline{\Omega}}D(x)>0.
\end{equation}
\item[(3)] To study the FQHD model which includes the quantum corrected energy equation is essentially significant for understanding the quantum transport of the hot carriers in semiconductor devices. The new feature is that one has to investigate both thermal and quantum effects in the more complex model system than the various simplified models.
\end{itemize}
In this paper, under the above physical principle settings (1)$\sim$(3), we will study the FQHD model \eqref{fqhd} in one space dimension with $\tau_m=\tau_e=\kappa=\lambda=1$. Namely, we consider
\begin{subequations}\label{1dfqhd}
\begin{numcases}{}
\md_t+\dl_x=0, \label{1dfqhd1}\\
\dl_t+\Bigg(\frac{\dl^2}{\md}+\md\wnd\Bigg)_x-\pc^2\md\Bigg[\frac{\big(\sqrt{\md}\big)_{xx}}{\sqrt{\md}}\Bigg]_x=\md\dws_x-\dl, \label{1dfqhd2}\\
\md\wnd_t+\dl\wnd_x+\frac{2}{3}\md\wnd\Bigg(\frac{\dl}{\md}\Bigg)_x-\frac{2}{3}\wnd_{xx}-\frac{\pc^2}{3}\Bigg[n\Bigg(\frac{j}{n}\Bigg)_{xx}\Bigg]_x=\frac{1}{3}\frac{\dl^2}{\md}-\md(\wnd-\theta_{L}), \label{1dfqhd3}\\
\dws_{xx}=\md-D(x), \qquad\forall t>0,\ \forall x\in\Omega:=(0,1),\label{1dfqhd4}
\end{numcases}
\end{subequations}
with the initial condition
\begin{equation}\label{ic}
(\md,\dl,\wnd)(0,x)=(\md_0,\dl_0,\wnd_0)(x),
\end{equation}
and the boundary conditions
\begin{subequations}\label{bc}
\begin{gather}
\md(t,0)=\md_{l},\qquad \md(t,1)=\md_{r},\label{bc-a}\\
\big(\sqrt{\md}\big)_{xx}(t,0)=\big(\sqrt{\md}\big)_{xx}(t,1)=0,\label{bc-b}\\
\wnd(t,0)=\wnd_{l},\qquad \wnd(t,1)=\wnd_{r},\label{bc-c}\\
\dws(t,0)=0,\qquad \dws(t,1)=\phi_r,\label{bc-d}
\end{gather}
\end{subequations}
where the boundary data $\md_{l},\md_{r},\wnd_{l},\wnd_{r}$ and $\phi_r$ are positive constants. The vanishing bohmenian-type boundary condition \eqref{bc-b} means that the quantum Bohm potential vanishes on the boundary, which is derived in \cite{G94,P99} and is also physically reasonable. The other boundary conditions in \eqref{bc} are called ohmic contact boundary condition. In order to establish the existence of a classical solution, we further assume the initial data $(\md_0,\dl_0,\wnd_0)$ is compatible with the boundary data \eqref{bc-a}$\sim$\eqref{bc-c} and $\md_t(t,0)=\md_t(t,1)=0$, namely,
\begin{gather}
\md_0(0)=\md_l,\quad \md_0(1)=\md_r,\quad \wnd_0(0)=\theta_l,\quad \wnd_0(1)=\theta_r,\notag\\
\dl_{0x}(0)=\dl_{0x}(1)=\big(\sqrt{\md_0}\big)_{xx}(0)=\big(\sqrt{\md_0}\big)_{xx}(1)=0.\label{compatibility}
\end{gather}
An explicit formula of the electrostatic potential
\begin{align}\label{efep}
\phi(t,x)=&\Phi[\md](t,x) \notag\\
:=&\int_0^x\int_0^y\big(\md(t,z)-D(z)\big)dzdy+\Bigg(\phi_r-\int_0^1\int_0^y\big(\md(t,z)-D(z)\big)dzdy\Bigg)x,
\end{align}
follows from \eqref{1dfqhd4} and \eqref{bc-d}.
In consideration of the solvability of the system \eqref{1dfqhd}, the following properties
\begin{subequations}\label{psc}
\begin{gather}
\inf_{x\in\Omega}\md>0,\qquad \inf_{x\in\Omega}\wnd>0,\label{pc}\\
\inf_{x\in\Omega} S[\md,\dl,\wnd]>0,\qquad \text{where}\ S[\md,\dl,\wnd]:=\wnd-\frac{\dl^2}{\md^2}\label{sc}
\end{gather}
\end{subequations}
attract our main interest. The condition \eqref{pc} means the positivity of the electron density and temperature. The other one \eqref{sc} is called the subsonic condition . Apparently, if we want to construct the solution in the physical region where the conditions \eqref{psc} hold, then the initial data \eqref{ic} must satisfy the same conditions
\begin{equation}\label{ipsc}
\inf_{x\in\Omega}\md_{0}>0,\qquad \inf_{x\in\Omega}\wnd_0>0, \qquad\inf_{x\in\Omega}S[\md_{0},\dl_{0},\wnd_0]>0.
\end{equation}
The strength of the boundary data, which is defined by
\begin{equation}\label{delta}
\delta:=|n_l-n_r|+|\theta_l-\theta_{L}|+|\theta_r-\theta_{L}|+|\phi_r|,
\end{equation}
plays a crucial role in the proofs of our main results in what follows.
The first aim in this paper is to investigate the existence, uniqueness and asymptotic stability of the stationary solution satisfying the following boundary value problem,
\begin{subequations}\label{1dsfqhd}
\begin{numcases}{}
\sdl_x=0, \label{1dsfqhd1}\\
S[\smd,\sdl,\swnd]\smd_x+\smd\swnd_x-\pc^2\smd\Bigg[\frac{\big(\sqrt{\smd}\big)_{xx}}{\sqrt{\smd}}\Bigg]_x=\smd\sdws_x-\sdl, \label{1dsfqhd2}\\
\sdl\swnd_x-\frac{2}{3}\sdl\swnd\big(\ln\smd\big)_x-\frac{2}{3}\swnd_{xx}-\frac{\pc^2}{3}\Bigg[\smd\Bigg(\frac{\sdl}{\smd}\Bigg)_{xx}\Bigg]_x=\frac{1}{3}\frac{\sdl^2}{\smd}-\smd(\swnd-\theta_{L}), \label{1dsfqhd3}\\
\sdws_{xx}=\smd-D(x), \qquad\forall x\in\Omega,\label{1dsfqhd4}
\end{numcases}
\end{subequations}
and
\begin{subequations}\label{sbc}
\begin{gather}
\smd(0)=\md_{l},\qquad \smd(1)=\md_{r},\label{sbc-a}\\
\big(\sqrt{\smd}\big)_{xx}(0)=\big(\sqrt{\smd}\big)_{xx}(1)=0,\label{sbc-b}\\
\swnd(0)=\wnd_{l},\qquad \swnd(1)=\wnd_{r},\label{sbc-c}\\
\sdws(0)=0,\qquad \sdws(1)=\phi_r.\label{sbc-d}
\end{gather}
\end{subequations}
The second aim in the present paper is to study the singular limit as the scaled Planck constant $\pc>0$ tends to zero in both the stationary problem \eqref{1dsfqhd}$\sim$\eqref{sbc} and the transient problem \eqref{1dfqhd}$\sim$\eqref{bc}. Formally, we let $\pc=0$ in the model system \eqref{1dfqhd} and its stationary counterpart \eqref{1dsfqhd}, respectively, we then obtain the following limit systems. The limit transient system can be written as
\begin{subequations}\label{1dfhd}
\begin{numcases}{}
\pcmd{0}_t+\pcdl{0}_x=0, \label{1dfhd1}\\
\pcdl{0}_t+\Bigg(\frac{(\pcdl{0})^2}{\pcmd{0}}+\pcmd{0}\pcwnd{0}\Bigg)_x=\pcmd{0}\pcdws{0}_x-\pcdl{0}, \label{1dfhd2}\\
\pcmd{0}\pcwnd{0}_t+\pcdl{0}\pcwnd{0}_x+\frac{2}{3}\pcmd{0}\pcwnd{0}\Bigg(\frac{\pcdl{0}}{\pcmd{0}}\Bigg)_x-\frac{2}{3}\pcwnd{0}_{xx}=\frac{1}{3}\frac{(\pcdl{0})^2}{\pcmd{0}}-\pcmd{0}(\pcwnd{0}-\theta_{L}), \label{1dfhd3}\\
\pcdws{0}_{xx}=\pcmd{0}-D(x), \qquad\forall t>0,\ \forall x\in\Omega,\label{1dfhd4}
\end{numcases}
\end{subequations}
and is supplemented by the same initial and ohmic contact boundary conditions with \eqref{ic} and \eqref{bc},
\begin{equation}\label{0ic}
(\pcmd{0},\pcdl{0},\pcwnd{0})(0,x)=(\md_0,\dl_0,\wnd_0)(x),
\end{equation}
and
\begin{subequations}\label{0bc}
\begin{gather}
\pcmd{0}(t,0)=\md_{l},\qquad \pcmd{0}(t,1)=\md_{r},\label{0bc-a}\\
\pcwnd{0}(t,0)=\wnd_{l},\qquad \pcwnd{0}(t,1)=\wnd_{r},\label{0bc-b}\\
\pcdws{0}(t,0)=0,\qquad \pcdws{0}(t,1)=\phi_r.\label{0bc-c}
\end{gather}
\end{subequations}
We call the limit system \eqref{1dfhd} as the full hydrodynamic (FHD) model for semiconductor devices. The limit stationary system is the stationary version of the FHD model \eqref{1dfhd}, it can be written by
\begin{subequations}\label{1dsfhd}
\begin{numcases}{}
\pcsdl{0}_x=0, \label{1dsfhd1}\\
S[\pcsmd{0},\pcsdl{0},\pcswnd{0}]\pcsmd{0}_x+\pcsmd{0}\pcswnd{0}_x=\pcsmd{0}\pcsdws{0}_x-\pcsdl{0}, \label{1dsfhd2}\\
\pcsdl{0}\pcswnd{0}_x-\frac{2}{3}\pcsdl{0}\pcswnd{0}\big(\ln\pcsmd{0}\big)_x-\frac{2}{3}\pcswnd{0}_{xx}=\frac{1}{3}\frac{(\pcsdl{0})^2}{\pcsmd{0}}-\pcsmd{0}(\pcswnd{0}-\theta_{L}), \label{1dsfhd3}\\
\pcsdws{0}_{xx}=\pcsmd{0}-D(x), \qquad\forall x\in\Omega,\label{1dsfhd4}
\end{numcases}
\end{subequations}
and is supplemented by the same ohmic contact boundary condition with \eqref{sbc},
\begin{subequations}\label{0sbc}
\begin{gather}
\pcsmd{0}(0)=\md_{l},\qquad \pcsmd{0}(1)=\md_{r},\label{0sbc-a}\\
\pcswnd{0}(0)=\wnd_{l},\qquad \pcswnd{0}(1)=\wnd_{r},\label{0sbc-b}\\
\pcsdws{0}(0)=0,\qquad \pcsdws{0}(1)=\phi_r.\label{0sbc-c}
\end{gather}
\end{subequations}
Throughout the rest of this paper, we will use the following notations. For a nonnegative integer $l\geq0$, $H^l(\Omega)$ denotes the $l$-th order Sobolev space in the $L^2$ sense, equipped with the norm $\|\cdot\|_l$. In particular, $H^0=L^2$ and $\|\cdot\|:=\|\cdot\|_0$. For a nonnegative integer $k\geq0$, $C^k(\overline{\Omega})$ denotes the $k$-times continuously differentiable function space, equipped with the norm $|f|_k:=\sum_{i=0}^k\sup_{x\in\overline{\Omega}}|\pd{x}{i}f(x)|$. The positive constants $C$, $C_1$, $\cdots$ only depend on $\md_{l}$, $\theta_L$ and $|D|_0$. If the constants $C$, $C_1$, $\cdots$ additionally depend on some other quantities $\alpha$, $\beta$, $\cdots$, we write $C(\alpha,\beta,\cdots)$, $C_1(\alpha,\beta,\cdots)$, $\cdots$. The notations $\mathfrak{X}_m^l$, $\mathfrak{Y}_m^l$ and $\mathfrak{Z}$ denote the function spaces defined by
\begin{gather*}
\mathfrak{X}_m^l([0,T]):=\bigcap_{k=0}^m C^k([0,T];H^{l+m-k}(\Omega)),\\
\mathfrak{Y}_m^l([0,T]):=\bigcap_{k=0}^{[m/2]} C^k([0,T];H^{l+m-2k}(\Omega)),\quad\text{for}\ m,l=0,1,2,\cdots,\\
\mathfrak{Z}([0,T]):=C^2([0,T];H^2(\Omega)).
\end{gather*}
The limit problems \eqref{1dfhd}$\sim$\eqref{0bc} and \eqref{1dsfhd}$\sim$\eqref{0sbc} have been studied by Nishibata and Suzuki \cite{NS09}. The authors obtain the existence, uniqueness and asymptotic stability of the stationary solution. The corresponding results are stated in the following lemmas.
\begin{lemma}[Existence and uniqueness of the limit stationary solution, \cite{NS09}]\label{lem1}
Let the doping profile and the boundary data satisfy conditions \eqref{nonflat} and \eqref{0sbc}. For arbitrary positive constants $n_l$ and $\theta_{L}$, there exist three positive constants $\delta_1$, $c$ and $C$ such that if $\delta\leq\delta_1$, then the BVP \eqref{1dsfhd}$\sim$\eqref{0sbc} has a unique solution $(\pcsmd{0},\pcsdl{0},\pcswnd{0},\pcsdws{0})$ satisfying the condition \eqref{psc} in the space $C^2(\overline{\Omega})\times C^2(\overline{\Omega})\times H^3(\Omega)\times C^2(\overline{\Omega})$. Moreover, the stationary solution satisfies the estimates
\begin{equation}\label{0se}
0<c\leq\pcsmd{0},\pcswnd{0},S[\pcsmd{0},\pcsdl{0},\pcswnd{0}]\leq C,\quad |\pcsdl{0}|+\|\pcswnd{0}-\theta_{L}\|_3\leq C\delta,\quad |(\pcsmd{0},\pcsdws{0})|_2\leq C.
\end{equation}
\end{lemma}
\begin{lemma}[Stability of the limit stationary solution, \cite{NS09}]\label{lem2}
Let the doping profile and the boundary data satisfy conditions \eqref{nonflat} and \eqref{0bc}. Assume that the initial data $(\md_0,\dl_0,\wnd_0)\in\big[H^2(\Omega)\big]^3$ and satisfies the conditions \eqref{compatibility} and \eqref{ipsc}. For arbitrary positive constants $n_l$ and $\theta_{L}$, there exist three positive constants $\delta_2$, $\gamma_1$ and $C$ such that if $\delta+\|(\md_0-\pcsmd{0},\dl_0-\pcsdl{0},\wnd_0-\pcswnd{0})\|_2\leq\delta_2$, then the IBVP \eqref{1dfhd}$\sim$\eqref{0bc} has a unique global solution $(\pcmd{0},\pcdl{0},\pcwnd{0},\pcdws{0})$ satisfying the condition \eqref{psc} in the space $\mathfrak{X}_2([0,\infty))\times\big[\mathfrak{X}_1^1([0,\infty))\cap H_{loc}^2(0,\infty;L^2(\Omega))\big]\times\big[\mathfrak{Y}_2([0,\infty))\cap H_{loc}^1(0,\infty;H^1(\Omega))\big]\times\mathfrak{Z}([0,\infty))$. Moreover, the solution verifies the additional regularity $\pcdws{0}-\pcsdws{0}\in\mathfrak{X}_2^2([0,\infty))$ and the decay estimate
\begin{multline}\label{0de}
\|(\pcmd{0}-\pcsmd{0},\pcdl{0}-\pcsdl{0},\pcwnd{0}-\pcswnd{0})(t)\|_2+\|(\pcdws{0}-\pcsdws{0})(t)\|_4\\
\leq C\|(\md_{0}-\pcsmd{0},\dl_{0}-\pcsdl{0},\wnd_{0}-\pcswnd{0})\|_2\,e^{-\gamma_1 t},\quad\forall t\in [0,\infty).
\end{multline}
\end{lemma}
Now, we are in the position to state the main results in this paper. Firstly, the existence and uniqueness of the quantum stationary solution is summarized in the following theorem.
\begin{theorem}[Existence and uniqueness of the quantum stationary solution]\label{thm1}
Suppose that the doping profile and the boundary data satisfy conditions \eqref{nonflat} and \eqref{sbc}. For arbitrary positive constants $n_l$ and $\theta_{L}$, there exist three positive constants $\delta_3$, $\pc_1(\leq1)$ and $C$ such that if $\delta\leq\delta_3$ and $0<\pc\leq\pc_1$, then the BVP \eqref{1dsfqhd}$\sim$\eqref{sbc} has a unique solution $(\pcsmd{\pc},\pcsdl{\pc},\pcswnd{\pc},\pcsdws{\pc})\in H^4(\Omega)\times H^4(\Omega)\times H^3(\Omega)\times C^2(\overline{\Omega})$ satisfying the condition \eqref{psc} and the uniform estimates
\begin{subequations}\label{145.1}
\begin{gather}
0<b^2\leq\pcsmd{\pc}\leq B^2,\quad 0<\frac{1}{2}\theta_L\leq\pcswnd{\pc}\leq\frac{3}{2}\theta_L, \label{m145.1a}\\
\|\pcsmd{\pc}\|_2+\|(\pc\pd{x}{3}\pcsmd{\pc},\pc^2\pd{x}{4}\pcsmd{\pc})\|+|\pcsdws{\pc}|_2\leq C,\label{m145.1b}\\
|\pcsdl{\pc}|+\|\pcswnd{\pc}-\theta_{L}\|_3\leq C\delta,\label{m145.1c}
\end{gather}
\end{subequations}
where the positive constants $B$ and $b$ are defined as follows
\begin{equation}\label{75.1}
B:=\frac{3}{2}\sqrt{n_l}\,e^{2|D|_0/\theta_{L}},\quad b:=\frac{1}{2}\sqrt{n_l}\,e^{-(B^2+2|D|_0)/\theta_{L}}.
\end{equation}
\end{theorem}
The asymptotic stability of the quantum stationary solution is stated in the next theorem.
\begin{theorem}[Stability of the quantum stationary solution]\label{thm2}
Assume that the doping profile and the boundary data satisfy conditions \eqref{nonflat} and \eqref{bc}. Let the initial data $(\md_0,\dl_0,\wnd_0)\in H^4(\Omega)\times H^3(\Omega)\times H^2(\Omega)$ and satisfies the conditions \eqref{compatibility} and \eqref{ipsc}. For arbitrary positive constants $n_l$ and $\theta_{L}$, there exist four positive constants $\delta_4$, $\pc_2$, $\gamma_2$ and $C$ such that if $0<\pc\leq\pc_2$ and $\delta+\|(\md_0-\pcsmd{\pc},\dl_0-\pcsdl{\pc},\wnd_0-\pcswnd{\pc})\|_2+\|(\pc\pd{x}{3}(\md_0-\pcsmd{\pc}),\pc\pd{x}{3}(\dl_0-\pcsdl{\pc}),\pc^2\pd{x}{4}(\md_0-\pcsmd{\pc}))\|\leq\delta_4$, then the IBVP \eqref{1dfqhd}$\sim$\eqref{bc} has a unique global solution $(\pcmd{\pc},\pcdl{\pc},\pcwnd{\pc},\pcdws{\pc})$ satisfying the condition \eqref{psc} in $\big[\mathfrak{Y}_4([0,\infty))\cap H_{loc}^2(0,\infty;H^1(\Omega))\big]\times\big[\mathfrak{Y}_3([0,\infty))\cap H_{loc}^2(0,\infty;L^2(\Omega))\big]\times\big[\mathfrak{Y}_2([0,\infty))\cap H_{loc}^1(0,\infty;H^1(\Omega))\big]\times\mathfrak{Z}([0,\infty))$. Moreover, the solution verifies the additional regularity $\pcdws{\pc}-\pcsdws{\pc}\in\mathfrak{Y}_4^2([0,\infty))$ and the decay estimate
\begin{align}
&\|(\pcmd{\pc}-\pcsmd{\pc},\pcdl{\pc}-\pcsdl{\pc},\pcwnd{\pc}-\pcswnd{\pc})(t)\|_2 \notag\\
&+\|(\pc\pd{x}{3}(\pcmd{\pc}-\pcsmd{\pc}),\pc\pd{x}{3}(\pcdl{\pc}-\pcsdl{\pc}),\pc^2\pd{x}{4}(\pcmd{\pc}-\pcsmd{\pc}))(t)\|+\|(\pcdws{\pc}-\pcsdws{\pc})(t)\|_4 \notag\\
&\leq C\Big(\|(\md_0-\pcsmd{\pc},\dl_0-\pcsdl{\pc},\wnd_0-\pcswnd{\pc})\|_2 \notag\\
&\qquad\qquad\quad+\|(\pc\pd{x}{3}(\md_0-\pcsmd{\pc}),\pc\pd{x}{3}(\dl_0-\pcsdl{\pc}),\pc^2\pd{x}{4}(\md_0-\pcsmd{\pc}))\|\Big)\,e^{-\gamma_2 t},\quad\forall t\in [0,\infty).\label{de}
\end{align}
\end{theorem}
It is naturally expected that the solution $(\pcmd{\pc},\pcdl{\pc},\pcwnd{\pc},\pcdws{\pc})$ of the quantum system \eqref{1dfqhd} approaches the solution $(\pcmd{0},\pcdl{0},\pcwnd{0},\pcdws{0})$ of the limit system \eqref{1dfhd} as $\pc$ tends to zero. To justify this expectation, we first consider the convergence of the stationary solutions. Precisely, we show that the quantum stationary solution $(\pcsmd{\pc},\pcsdl{\pc},\pcswnd{\pc},\pcsdws{\pc})$ of the BVP \eqref{1dsfqhd}$\sim$\eqref{sbc} converges to the limit stationary solution $(\pcsmd{0},\pcsdl{0},\pcswnd{0},\pcsdws{0})$ of the BVP \eqref{1dsfhd}$\sim$\eqref{0sbc} as $\pc$ tends to zero. Then, we further study the convergence of the global solutions. The former result is summarized in the following theorem.
\begin{theorem}[Semi-classical limit of the stationary solutions]\label{thm3}
Suppose that the same conditions in Lemma \ref{lem1} and Theorem \ref{thm1} hold. For arbitrary positive constants $n_l$ and $\theta_{L}$, there exist two positive constants $\delta_5$ and $C$ such that if $\delta\leq\delta_5$, then for all $0<\pc\leq\pc_1$ (where $\pc_1$ is given in Theorem \ref{thm1}) the following convergence estimate
\begin{subequations}\label{ss39.1}
\begin{equation}\label{ss39.1a}
\|\pcsmd{\pc}-\pcsmd{0}\|_1+|\pcsdl{\pc}-\pcsdl{0}|+\|\pcswnd{\pc}-\pcswnd{0}\|_2+\|\pcsdws{\pc}-\pcsdws{0}\|_3\leq C\pc,
\end{equation}
holds true. Furthermore,
\begin{equation}\label{ss39.1b}
\big\|\big(\pd{x}{2}(\pcsmd{\pc}-\pcsmd{0}),\pc\pd{x}{3}\pcsmd{\pc},\pc^2\pd{x}{4}\pcsmd{\pc},\pd{x}{3}(\pcswnd{\pc}-\pcswnd{0}),\pd{x}{4}(\pcsdws{\pc}-\pcsdws{0})\big)\big\|\rightarrow0,\quad\text{as}\ \pc\rightarrow0.
\end{equation}
\end{subequations}
\end{theorem}
The semi-classical limit of the transient problem is stated in the next theorem.
\begin{theorem}[Semi-classical limit of the global solutions]\label{thm4}
Assume that the same conditions in Lemma \ref{lem2} and Theorem \ref{thm2} hold. For arbitrary positive constants $n_l$ and $\theta_{L}$, there exist four positive constants $\delta_6$, $\gamma_3$, $\gamma_4$ and $C$ such that if
\begin{multline}\label{sic}
\pc+\delta+\|(\md_0-\pcsmd{\pc},\dl_0-\pcsdl{\pc},\wnd_0-\pcswnd{\pc})\|_2\\
+\big\|\big(\pc\pd{x}{3}(\md_0-\pcsmd{\pc}),\pc\pd{x}{3}(\dl_0-\pcsdl{\pc}),\pc^2\pd{x}{4}(\md_0-\pcsmd{\pc})\big)\big\|\leq\delta_6,
\end{multline}
then the following convergence estimates
\begin{subequations}\label{gs}
\begin{equation}\label{gs164.5}
\|(\pcmd{\pc}-\pcmd{0},\pcdl{\pc}-\pcdl{0},\pcwnd{\pc}-\pcwnd{0})(t)\|_1+\|(\pcdws{\pc}-\pcdws{0})(t)\|_3\leq Ce^{\gamma_3 t}\pc^{1/2},\quad\forall t\in[0,\infty),
\end{equation}
and
\begin{equation}\label{gs1}
\sup_{t\in[0,\infty)}\Big(\|(\pcmd{\pc}-\pcmd{0},\pcdl{\pc}-\pcdl{0},\pcwnd{\pc}-\pcwnd{0})(t)\|_1+\|(\pcdws{\pc}-\pcdws{0})(t)\|_3\Big)\leq C\pc^{\gamma_4}
\end{equation}
hold true.
\end{subequations}
\end{theorem}
Now, we illustrate the main ideas and the key technical points in the proofs of the above theorems. Firstly, we apply the Schauder fixed-point theorem to solve the stationary problem \eqref{1dsfqhd}$\sim$\eqref{sbc}. To this end, we heuristically construct a fixed-point mapping $\mathcal{T}$, see \eqref{110.1}, through a careful observation on the structure of the stationary FQHD model \eqref{1dsfqhd}. Roughly speaking, in order to deal with the Bohm potential term in the stationary momentum equation \eqref{1dsfqhd2}, we introduce a transformation $\rsmd:=\sqrt{\smd}$ and reduce the stationary momentum equation to a parameter-dependent semilinear elliptic equation of the second order with the nonlocal terms by using the vanishing bohmenian-type boundary condition \eqref{sbc-b}. In order to treat the dispersive velocity term in the stationary energy equation \eqref{1dsfqhd3}, the desired mapping $\mathcal{T}$ has to be defined by solving two carefully designed nonlocal problems $(P1)$ and $(P2)$ in turn, see \eqref{109.1} and \eqref{109.2}. The unique solvability of both $(P1)$ and $(P2)$ can be proved by using the Leray-Schauder fixed-point theorem and energy estimates. This ensure that the mapping $\mathcal{T}$ is well-defined. During the proof, the main difficulty is to establish the uniform (in $\pc$) estimates \eqref{eu} and \eqref{eQ}.
Secondly, the existence of the global-in-time solution and the asymptotic stability of the stationary solution can be proved by the standard continuation argument based on the local existence and the uniform a priori estimate. Similar to the stationary problem, we also introduce a transformation $\rmd:=\sqrt{\md}$ to conveniently deal with the Bohm potential term in the momentum equation \eqref{1dfqhd2}. The local existence result is proved by combining the iteration method with the energy estimates. The unique solvability of the linearized problem \eqref{a3.1}$\sim$\eqref{a3.3} used to design the iteration scheme is shown in Appendix by Galerkin method, where we have used the existence result in \cite{NS08} for a fourth order wave equation. The uniform a priori estimate is established by refined energy method. The proof is very complicated due to the non-flatness of the stationary density and the appearance of the dispersive velocity term in the perturbed energy equation \eqref{10.2c}. During the proof, we find that the spatial derivatives of the perturbations $(\prmd,\pdl,\pwnd,\pdws)$ can be bounded by the temporal derivatives of the perturbations $(\prmd,\pdl,\pwnd)$ with the help of the special structure of the perturbed system \eqref{10.2}, see \eqref{46.5}. Therefore, we only need to establish the estimates of the temporal derivatives of the perturbations $(\prmd,\pdl,\pwnd)$ by using the homogeneous boundary condition \eqref{pbc}. We also find the interplay of the dissipative-dispersive effects in the FQHD model. Roughly speaking, the Bohm potential term in the perturbed momentum equation \eqref{10.2b} contributes the quantum dissipation rate $\|\pc\pd{t}{k}\prmd_{xx}(t)\|$, see \eqref{115.1}. The dispersive velocity term in the perturbed energy equation \eqref{10.2c} contributes the extra quantum dissipation rate $\|\pc\pd{t}{k}\prmd_{tx}(t)\|$, see \eqref{115.2}. The dissipative property of the dispersive velocity term plays a crucial role to close the uniform a priori estimate \eqref{127.2}.
Finally, we justify the semi-classical limit for both the stationary solutions and global solutions by using the energy method and compactness argument. For stationary solutions, in order to overcome the difficulties arising from the non-flatness of the stationary density, we need to introduce the transformations $\pcszmd{\pc}:=\ln \pcsmd{\pc}$ and $\pcszmd{0}:=\ln \pcsmd{0}$. In addition, we also have to technically estimate a bad integral term $I_2$ during establishing the error estimate of the stationary temperature error variable $\pcpswnd$. Actually, we find that the quantum stationary current density $\pcsdl{\pc}$ and the limit stationary current density $\pcsdl{0}$ possess the same explicit formula \eqref{104.1} due to the vanishing bohmenian-type boundary condition. Based on this fact, we can successfully overcome the difficulty in estimating the integral term $I_2$, see \eqref{ss17.2}. For global solutions, we have to pay more attention on the influences of the quantum corrected energy equation \eqref{157.2}, see \eqref{152.1} and \eqref{161.6} for example, the computations are very complicated. In the proof, the semi-classical limit of the stationary solutions plays important role.
The paper is organized as follows. In Section \ref{Sect.2}, we prove the existence and uniqueness of the stationary solution. Section \ref{Sect.3} is devoted to the global existence and stability analysis. In Subsection \ref{Subsect.3.1}, we show the local existence. In Subsections \ref{Subsect.3.2}$\sim$\ref{Subsect.3.5}, we reformulate the problem and establish the uniform a priori estimate. Section \ref{Sect.4} is devoted to the verification of the semi-classical limit. In Subsection \ref{Subsect.4.1}, we discuss the stationary case. In Subsection \ref{Subsect.4.2}, we study the non-stationary case.
\section{Existence and uniqueness of the stationary solution}\label{Sect.2}
In this section, we show Theorem \ref{thm1}. The proof is based on the Schauder fixed-point theorem (see Corollary 11.2 in \cite{GT98}), the Leray-Schauder fixed-point theorem (see Theorem 11.3 in \cite{GT98}) and the energy method.
\begin{proof}[\textbf{Proof of Theorem \ref{thm1}.}]
Since the proof is complicated, we divide it into several steps for clarification.
\emph{Step 1. Reformulation of the problem .}
It is convenient to make use of the transformation $\rsmd:=\sqrt{\smd}$. Inserting this transformation into the system \eqref{1dsfqhd}, dividing the the equation \eqref{1dsfqhd2} by $\rsmd^2$ and integrating the resultant on $[0,x)$ and then using the boundary condition \eqref{sbc}, applying the Green formula to the equation \eqref{1dsfqhd4} together with the boundary condition \eqref{sbc-d}, via the necessary calculations, the above procedures yield the following BVP with a constant current density $\sdl$ (which will be determined later, see \eqref{104.1} below),
\begin{subequations}\label{103.2}
\begin{numcases}{}
\pc^2\rsmd_{xx}=h(\rsmd,\swnd), \label{103.2a}\\
\frac{2}{3}\swnd_{xx}-\sdl\swnd_{x}+\frac{2}{3}\sdl\swnd(\ln \rsmd^2)_x-\rsmd^2(\swnd-\theta_L)=g(\rsmd,\swnd;\pc), \quad x\in\Omega,\label{103.2b}
\end{numcases}
\end{subequations}
with boundary conditions
\begin{subequations}\label{103.3}
\begin{gather}
\rsmd(0)=\rmd_{l},\qquad \rsmd(1)=\rmd_{r},\label{103.3a}\\
\swnd(0)=\wnd_{l},\qquad \swnd(1)=\wnd_{r},\label{103.3b}
\end{gather}
\end{subequations}
where
\begin{subequations}
\begin{gather}
F(a_1,a_2,a_3):=\frac{a_2^2}{2a_1^2}+a_3+a_3\ln a_1,\quad\rmd_{l}:=\sqrt{\md_{l}}, \quad\rmd_{r}:=\sqrt{\md_{r}},\label{103.5}\\
\sdws(x)=G[\rsmd^2](x):=\int_0^1G(x,y)(\rsmd^2-D)(y)dy+\phi_rx,\quad G(x,y):=\begin{cases}x(y-1),\ x<y\\y(x-1),\ x>y\end{cases},\label{104.2}\\
h(\rsmd,\swnd):=\rsmd\bigg[F(\rsmd^2,\sdl,\swnd)-F(\md_{l},\sdl,\theta_l)-\sdws-\int_0^x\swnd_{x}\ln\rsmd^2dy+\sdl\int_0^x\rsmd^{-2}dy\bigg],\label{103.4}\\
g(\rsmd,\swnd;\pc):=-\frac{1}{3}\frac{\sdl^2}{\rsmd^2}+\frac{\pc^2}{3}\sdl\bigg(\frac{12\rsmd_x^3}{\rsmd^3}-\frac{14\rsmd_x\rsmd_{xx}}{\rsmd^2}+\frac{2\rsmd_{xxx}}{\rsmd}\bigg).\label{103}
\end{gather}
\end{subequations}
Next, taking value $x=1$ in the equation \eqref{103.2a} and using the boundary condition \eqref{sbc}, we obtain the current-voltage relation
\begin{equation}\label{ucvr}
F(\md_{r},\sdl,\theta_r)-F(\md_{l},\sdl,\theta_l)-\phi_r-\int_0^1\swnd_{x}\ln\rsmd^2dy+\sdl\int_0^1\rsmd^{-2}dy=0.
\end{equation}
Easy to see, the the equation \eqref{ucvr} is a quadratic equation on $\sdl$. Based on the subsonic condition \eqref{psc}, we can uniquely solve $\sdl$ provided $\rsmd$, $\swnd$ are given and the strength parameter $\delta$ is small enough. Precisely, the constant stationary current density $\sdl$ satisfies the following explicit formula
\begin{align}\label{104.1}
&\sdl=J[\rsmd^2,\swnd]:=2\Big(\bar{b}+\int_0^1\swnd_{x}\ln\rsmd^2dy\Big)K[\rsmd^2,\swnd]^{-1},\\
&K[\rsmd^2,\swnd]:=\int_0^1\rsmd^{-2}dy+\sqrt{\Big(\int_0^1\rsmd^{-2}dy\Big)^2+2\Big(\bar{b}+\int_0^1\swnd_{x}\ln\rsmd^2dy\Big)\Big(n_{r}^{-2}-n_{l}^{-2}\Big)},\notag\\
&\bar{b}:=\phi_r-\theta_r+\theta_l-\theta_r\ln n_r+\theta_l\ln n_l.\notag
\end{align}
It is obvious that the BVP \eqref{103.2}$\sim$\eqref{103.3} combined with the explicit formulas \eqref{104.1} and \eqref{104.2} is equivalent to the original BVP \eqref{1dsfqhd}$\sim$\eqref{sbc} under the transformation $\smd=\rsmd^2$ for positive smooth solution $(\rsmd,\swnd)$.
\emph{Step 2. Construction of the fixed-point mapping.} From now on, we focus on the unique solvability of the BVP \eqref{103.2}$\sim$\eqref{103.3}. The system \eqref{103.2} is a one-dimensional semilinear nonlocal elliptic system with a singular parameter $\pc\in(0,1]$ in the principal part of its first component equation \eqref{103.2a}. To solve it, we adopt the conventional framework based on the Schauder fixed-point theorem.
Observing the structure of the system \eqref{103.2}, we can construct the fixed-point mapping appropriately by the following procedure.
Firstly, we introduce a closed convex subset $\mathcal{U}[N_1,N_2]$ in the Banach space $C^2(\overline{\Omega})$ below, where $N_1$ and $N_2$ are positive constants to be determined later (see \eqref{N12} below),
\begin{equation}\label{120.1}
\mathcal{U}[N_1,N_2]:=\Big\{q\in C^2(\overline{\Omega})\,\Big|\, \|q-\theta_L\|_1\leq N_1\delta,\quad\|q_{xx}\|\leq N_2\delta,\quad q(0)=\theta_l,\ q(1)=\theta_r \Big\}.
\end{equation}
Next, we define the fixed-point mapping
\begin{align}\label{110.1}
\mathcal{T}:\ \mathcal{U}[N_1,N_2]&\longrightarrow H^3(\Omega)\notag\\
q&\longmapsto Q
\end{align}
by solving the following two problems in turn. For any fixed $q\in \mathcal{U}[N_1,N_2]$, we firstly solve the problem $(P1)$:
\begin{subequations}\label{109.1}
\begin{numcases}{(P1)\quad}
\pc^2u_{xx}=h(u,q), \quad x\in\Omega,\label{109.1a}\\
u(0)=w_l,\quad u(1)=w_r.
\end{numcases}
\end{subequations}
For problem $(P1)$, we claim the following fact, its proof will be given in the next step.
\begin{quotation}
\emph{Claim 1. For given $q\in \mathcal{U}[N_1,N_2]$, if $\delta$ and $\pc$ are small enough, then $(P1)$ has a unique solution $u=u[q]\in H^4(\Omega)$ satisfying the following uniform estimate with respect to $\pc$, that is,
\begin{subequations}\label{eu}
\begin{gather}
0<b\leq u(x)\leq B,\label{eu-a}\\
\|u\|_2+\|(\pc\pd{x}{3}u, \pc^2\pd{x}{4}u)\|\leq C,\label{eu-b}
\end{gather}
\end{subequations}
where the positive constants $b$ and $B$ are given by \eqref{75.1}, and the positive constant $C$ only depends on $n_l$, $\theta_L$ and $|D|_0$.}
\end{quotation}
Based on the Claim 1, for given function pair $(u,q)$, we further solve the problem $(P2)$:
\begin{subequations}\label{109.2}
\begin{numcases}{(P2)\quad}
\frac{2}{3}Q_{xx}-JQ_{x}+\frac{2}{3}J_*(\ln u^2)_x\theta_L+\frac{2}{3}J(\ln u^2)_x(Q-\theta_L)\notag\\
\qquad\qquad\qquad\qquad\qquad\qquad\ \,-u^2(Q-\theta_L)=g(u,q;\pc), \quad x\in\Omega,\label{109.2a}\\
Q(0)=\theta_l,\quad Q(1)=\theta_r,
\end{numcases}
\end{subequations}
where $J:=J[u^2,q]$ and $J_*:=2\big(\bar{b}+\int_0^1Q_x\ln u^2dx\big)K[u^2,q]^{-1}$. For problem $(P2)$, we also have a claim, its proof will be given in the Step 4.
\begin{quotation}
\emph{Claim 2. For given $(u,q)$ in Claim 1, if $\delta$ is small enough, then $(P2)$ has a unique solution $Q\in H^3(\Omega)$ satisfying the following uniform estimate with respect to $\pc$, that is,
\begin{subequations}\label{eQ}
\begin{align}
&\|Q-\theta_L\|_1\leq C_1\delta+C_2(b,B,N_1)\delta^2,\label{eQa}\\
&\|Q_{xx}\|\leq C_3(b,B,N_1)\delta,\label{eQb}\\
&\|Q_{xxx}\|\leq C_4(b,B,N_1,N_2)\delta,\label{eQc}
\end{align}
\end{subequations}
where the positive constant $C_1$ only depends on $n_l$, $\theta_L$ and $|D|_0$.}
\end{quotation}
\emph{Step 3. Proof of Claim 1.}
Now, we begin to solve $(P1)$. In order to avoid vacuum $u=0$, we consider a truncation problem $(tP)$ induced by $(P1)$:
\begin{subequations}\label{111.1}
\begin{numcases}{(tP)\quad}
\pc^2u_{xx}=h(u_{\alpha\beta},q), \quad x\in\Omega,\label{111.1a}\\
u(0)=w_l,\quad u(1)=w_r,\label{111.1b}
\end{numcases}
\end{subequations}
where
\begin{equation*}
u_{\alpha\beta}:=\max\big\{\beta,\min\{\alpha,u\}\big\},\quad 0<\frac{1}{2}b=:\beta<\alpha:=2B.
\end{equation*}
This problem can be solved by Leray-Schauder fixed-point theorem. To this end, we define a fixed-point mapping $\mathcal{T}_1: r\mapsto R$ over $H^1(\Omega)$ by solving the linear problem:
\begin{subequations}\label{112.1}
\begin{numcases}{}
\pc^2R_{xx}=h(r_{\alpha\beta},q), \quad x\in\Omega,\label{112.1a}\\
R(0)=w_l,\quad R(1)=w_r.\label{112.1b}
\end{numcases}
\end{subequations}
In fact, for given $q\in\mathcal{U}[N_1,N_2]$ and $r\in H^1(\Omega)$, the right-side $h(r_{\alpha\beta},q)\in H^1(\Omega)$. Thus, the linear BVP \eqref{112.1} has a unique solution $R=:\mathcal{T}_1r\in H^3(\Omega)$ by the standard theory of the elliptic equations. In addition, the mapping $\mathcal{T}_1$ is a continuous and compact mapping from $H^1(\Omega)$ into itself. Next, we show that there exists a positive constant $M_1$ such that $\|v\|_1\leq M_1$ for an arbitrary $v\in\big\{f\in H^1(\Omega)\,|\,f=\lambda\mathcal{T}_1f, \ \forall\lambda\in[0,1] \big\}$. We may assume $\lambda>0$ as the case $\lambda=0$ is trivial. It is sufficient to show that $\|v\|_1\leq M_1$ for $v$ satisfying the following problem
\begin{subequations}\label{113.3}
\begin{numcases}{}
\pc^2v_{xx}=\lambda h(v_{\alpha\beta},q), \quad x\in\Omega,\label{113.3a}\\
v(0)=\lambda w_l,\quad v(1)=\lambda w_r.\label{113.3b}
\end{numcases}
\end{subequations}
Performing the procedure
\begin{equation*}\label{153.2}
\int_0^1\eqref{113.3a}\times(v-\lambda\bar{w})dx,\quad\text{where}\ \bar{w}(x):=w_l(1-x)+w_rx,
\end{equation*}
and using the Young inequality, the mean value theorem, the formula \eqref{104.1} and the estimate $|h(v_{\alpha\beta},q)|\leq C$, where $C$ is a positive constant which only depends on $\alpha$, $\beta$, $n_l$, $\theta_L$, and $|D|_0$. If $\delta$ is small enough, then these computations yield the desired estimate
\begin{equation}\label{153.4}
\|v\|_1\leq C\bigg(1+\frac{1}{\pc}\bigg)=:M_1.
\end{equation}
Based on the estimate \eqref{153.4}, we can directly apply the Leray-Schauder fixed-point theorem to the mapping $\mathcal{T}_1$, and see that $\mathcal{T}_1$ has a fixed-point $u=\mathcal{T}_1u\in H^3(\Omega)$ which is a strong solution to the truncation problem $(tP)$.
Next, we can further provide a maximum principle argument for any strong solution $u$ to the truncation problem $(tP)$. Consequently, this result can help us to remove the truncation in $(tP)$ and show that the solution $u$ to the truncation problem $(tP)$ exactly is a solution to the problem $(P1)$.
We first establish the upper bound of $u_{\alpha\beta}$. Before doing this, we note that if $\delta$ is small enough, then $q\in\mathcal{U}[N_1,N_2]$ implies
\begin{equation}\label{154.2}
0<\frac{1}{2}\theta_L\leq q(x)\leq\frac{3}{2}\theta_L,\quad\|q_x\|\leq N_1\delta,\quad\|q_{xx}\|\leq N_2\delta.
\end{equation}
Now, we can establish the upper bound of $u_{\alpha\beta}$ by choosing the appropriate test functions in $H_0^1(\Omega)$. To this end, we define $\bar{n}:=\max\{n_l,n_r\}>0$, and perform the procedure
\begin{equation}\label{154.1}
\int_0^1-\eqref{111.1a}\times\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^kdx,\quad k=1,2,3,\cdots,\quad\text{where}\ (\cdot)_+:=\max\{0,\cdot\}.
\end{equation}
The computations in terms of this procedure yield that
\begin{equation}\label{56.1}
\int_0^1-\pc^2u_{xx}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^kdx=\int_0^1-h(u_{\alpha\beta},q)\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^kdx.
\end{equation}
The left-side of \eqref{56.1} can be estimated as follows by integration by parts,
\begin{align}
\eqref{56.1}_l&=\int_0^1\pc^2u_x\Bigg[\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^k\Bigg]_xdx \notag\\
&=\int_0^12\pc^2k\frac{[(u_{\alpha\beta})_x]^2}{u_{\alpha\beta}}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^{k-1}dx\geq0.\label{57.1}
\end{align}
Based on the expression \eqref{103.4}, the estimate \eqref{154.2} and the Young inequality, we can estimate the right-side of \eqref{56.1} as follows, where $\bar{J}:=J[u_{\alpha\beta}^{2},q]$ satisfying the estimate $|\bar{J}|\leq C(\alpha,\beta,N_1)\delta$. Namely,
\allowdisplaybreaks
\begin{align}
\eqref{56.1}_r=&\int_0^1-u_{\alpha\beta}\bigg[F(u_{\alpha\beta}^2,\bar{J},q)-F(\md_{l},\bar{J},\theta_l)-G[u_{\alpha\beta}^{2}]\notag\\
&\qquad\qquad\qquad\qquad-\int_0^xq_{x}\ln u_{\alpha\beta}^2dy+\bar{J}\int_0^xu_{\alpha\beta}^{-2}dy\bigg]\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^kdx\notag\\
=&\int_0^1-u_{\alpha\beta}\bigg[F(u_{\alpha\beta}^2,\bar{J},q)-q\ln\bar{n}+q\ln\bar{n}-F(\md_{l},\bar{J},\theta_l)-G[u_{\alpha\beta}^{2}]\notag\\
&\qquad\qquad\qquad\qquad\qquad\qquad\qquad\ -\int_0^xq_{x}\ln u_{\alpha\beta}^2dy+\bar{J}\int_0^xu_{\alpha\beta}^{-2}dy\bigg]\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^kdx\notag\\
=&-\int_0^1u_{\alpha\beta}q\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^{k+1}dx\notag\\
&+\int_0^1\bigg(\phi_rx-\int_0^1G(x,y)D(y)dy-\bar{J}\int_0^xu_{\alpha\beta}^{-2}dy+\frac{\bar{J}^2}{2n_l}\bigg)u_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^kdx\notag\\
&+\int_0^1\bigg(\theta_l\ln n_l-q\ln\bar{n}+\theta_l-q+\int_0^xq_x\ln u_{\alpha\beta}^2dy \bigg)u_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^kdx\notag\\
&+\int_0^1\bigg(\int_0^1\underbrace{G(x,y)}_{\leq0}u_{\alpha\beta}^2(y)dy-\frac{\bar{J}^2}{2u_{\alpha\beta}^4}\bigg)u_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^kdx\notag\\
\leq&-\int_0^1\frac{1}{2}\theta_Lu_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^{k+1}dx+\int_0^1\big(C(N_1)\delta+|D|_0\big)u_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^kdx\notag\\
&+\int_0^1C(N_1)\delta u_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^kdx+0\notag\\
\leq&-\int_0^1\frac{1}{2}\theta_Lu_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^{k+1}dx+\int_0^12|D|_0u_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^kdx \qquad\text{if}\ \delta\ll1\notag\\
=&-\int_0^1\frac{1}{2}\theta_Lu_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^{k+1}dx+\int_0^12|D|_0\frac{2}{\theta_L}\frac{\theta_L}{2}u_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^kdx\notag\\
=&-\int_0^1\frac{1}{2}\theta_Lu_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^{k+1}dx+\int_0^1\frac{1}{2}\theta_Lu_{\alpha\beta}\underbrace{\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^k\frac{4|D|_0}{\theta_L}}_{\text{by Young inequality}}dx\notag\\
\leq&-\int_0^1\frac{1}{2}\theta_Lu_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^{k+1}dx\notag\\
&\qquad\qquad\qquad+\int_0^1\frac{1}{2}\theta_Lu_{\alpha\beta}\Bigg[\frac{k}{k+1}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^{k+1}+\frac{1}{k+1}\bigg(\frac{4|D|_0}{\theta_L}\bigg)^{k+1}\Bigg]dx\notag\\
=&-\frac{1}{k+1}\frac{1}{2}\theta_L\int_0^1u_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^{k+1}dx+\frac{1}{k+1}\frac{1}{2}\theta_L\bigg(\frac{4|D|_0}{\theta_L}\bigg)^{k+1}\int_0^1\underbrace{u_{\alpha\beta}}_{\leq\alpha}dx\notag\\
\leq&\frac{\theta_L}{2(k+1)}\Bigg[-\int_0^1u_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^{k+1}dx+\alpha\bigg(\frac{4|D|_0}{\theta_L}\bigg)^{k+1}\Bigg]. \label{59.3}
\end{align}
Inserting \eqref{57.1} and \eqref{59.3} into \eqref{56.1}, we have the estimate
\begin{equation}\label{60.2}
\int_0^1\sqrt{\bar{n}}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^{k+1}dx\leq\int_0^1u_{\alpha\beta}\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+^{k+1}dx\leq\alpha\bigg(\frac{4|D|_0}{\theta_L}\bigg)^{k+1},
\end{equation}
which implies
\begin{equation}\label{61.2}
\bigg\|\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+\bigg\|_{L^{k+1}(\Omega)}\leq\bigg(\frac{\alpha}{\sqrt{\bar{n}}}\bigg)^{\frac{1}{k+1}}\frac{4|D|_0}{\theta_L},\quad k=1,2,3,\cdots.
\end{equation}
Let $k\rightarrow\infty$ in \eqref{61.2}, we immediately obtain
\begin{equation}\label{61.3}
\bigg\|\bigg(\ln\frac{u_{\alpha\beta}^2}{\bar{n}}\bigg)_+\bigg\|_{L^{\infty}(\Omega)}\leq\frac{4|D|_0}{\theta_L}.
\end{equation}
Note that $[\ln(u_{\alpha\beta}^2/\bar{n})]_+$ is nonnegative, then the estimate \eqref{61.3} implies that
\begin{equation}\label{62.1}
u_{\alpha\beta}\leq\sqrt{\bar{n}}e^{2|D|_0/\theta_L}\leq B,\quad\text{if}\ \delta\ll1.
\end{equation}
Using the similar argument, we can establish the lower bound of $u_{\alpha\beta}$. To this end, we define $\underbar{n}:=\min\{n_l,n_r\}>0$, and perform the procedure
\begin{equation}\label{159.1}
\int_0^1-\eqref{111.1a}\times\frac{1}{u_{\alpha\beta}}\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k-1}dx,\quad k=1,2,3,\cdots,\quad\text{where}\ (\cdot)_-:=\min\{0,\cdot\}.
\end{equation}
The computations in terms of this procedure yield that
\begin{equation}\label{63.1}
\int_0^1-\pc^2\frac{u_{xx}}{u_{\alpha\beta}}\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k-1}dx=\int_0^1-\frac{h(u_{\alpha\beta},q)}{u_{\alpha\beta}}\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k-1}dx.
\end{equation}
The left-side of \eqref{63.1} can be estimated as follows by integration by parts,
\begin{align}
\eqref{63.1}_l&=\int_0^1\pc^2u_x\Bigg[\frac{1}{u_{\alpha\beta}}\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k-1}\Bigg]_xdx \notag\\
&=\int_0^1\pc^2u_x\bigg(\frac{1}{u_{\alpha\beta}}\bigg)_x\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k-1}dx+\int_0^1\pc^2\frac{u_x}{u_{\alpha\beta}}\Bigg[\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k-1}\Bigg]_x\notag\\
&=-\int_0^1\pc^2\bigg[\frac{(u_{\alpha\beta})_x}{u_{\alpha\beta}}\bigg]^2\Bigg[\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k-1}-2(2k-1)\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k-2}\Bigg]dx\notag\\
&\geq0.\label{64.1}
\end{align}
The right-side of \eqref{63.1} can be estimated as follows,
\allowdisplaybreaks
\begin{align}
\eqref{63.1}_r=&\int_0^1-\bigg[F(u_{\alpha\beta}^2,\bar{J},q)-F(\md_{l},\bar{J},\theta_l)-G[u_{\alpha\beta}^{2}]\notag\\
&\qquad\qquad\qquad\qquad-\int_0^xq_{x}\ln u_{\alpha\beta}^2dy+\bar{J}\int_0^xu_{\alpha\beta}^{-2}dy\bigg]\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k-1}dx\notag\\
=&\int_0^1-\bigg[F(u_{\alpha\beta}^2,\bar{J},q)-q\ln\underbar{n}+q\ln\underbar{n}-F(\md_{l},\bar{J},\theta_l)-G[u_{\alpha\beta}^{2}]\notag\\
&\qquad\qquad\qquad\qquad\qquad\qquad-\int_0^xq_{x}\ln u_{\alpha\beta}^2dy+\bar{J}\int_0^xu_{\alpha\beta}^{-2}dy\bigg]\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k-1}dx\notag\\
=&-\int_0^1q\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k}dx\notag\\
&+\int_0^1\Bigg[\phi_rx+\int_0^1G(x,y)(u_{\alpha\beta}^2-D)(y)dy-\bar{J}\int_0^xu_{\alpha\beta}^{-2}dy+\frac{\bar{J}^2}{2n_l}\notag\\
&\qquad\qquad+\theta_l\ln n_l-q\ln\underbar{n}+\theta_l-q+\int_0^xq_x\ln u_{\alpha\beta}^2dy-\frac{\bar{J}^2}{2u_{\alpha\beta}^4}\Bigg]\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k-1}dx\notag\\
\leq&-\int_0^1\frac{1}{2}\theta_L\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k}dx-\int_0^1\big(C(N_1)\delta+B^2+|D|_0\big)\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k-1}dx\notag\\
\leq&-\int_0^1\frac{1}{2}\theta_L\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k}dx-\int_0^1\big(B^2+2|D|_0\big)\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k-1}dx\qquad\text{if}\ \delta\ll1 \notag\\
=&-\int_0^1\frac{1}{2}\theta_L\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k}dx-\int_0^1\big(B^2+2|D|_0\big)\frac{2}{\theta_L}\frac{\theta_L}{2}\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k-1}dx\notag\\
=&-\int_0^1\frac{1}{2}\theta_L\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k}dx+\int_0^1\frac{1}{2}\theta_L\underbrace{\Bigg[-\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-\Bigg]^{2k-1}\frac{2\big(B^2+2|D|_0\big)}{\theta_L}}_{\text{by Young inequality}}dx\notag\\
\leq&-\int_0^1\frac{1}{2}\theta_L\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k}dx\notag\\
&\quad+\int_0^1\frac{1}{2}\theta_L\Bigg\{\frac{2k-1}{2k}\Bigg[-\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-\Bigg]^{(2k-1)\cdot\frac{2k}{2k-1}}+\frac{1}{2k}\bigg[\frac{2\big(B^2+2|D|_0\big)}{\theta_L}\bigg]^{2k}\Bigg\}dx\notag\\
=&\frac{\theta_L}{4k}\Bigg\{-\int_0^1\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-^{2k}dx+\bigg[\frac{2\big(B^2+2|D|_0\big)}{\theta_L}\bigg]^{2k}\Bigg\}. \label{66.3}
\end{align}
Inserting \eqref{64.1} and \eqref{66.3} into \eqref{63.1}, we have the estimate
\begin{equation}\label{67.1}
\bigg\|\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-\bigg\|_{L^{2k}(\Omega)}\leq\frac{2\big(B^2+2|D|_0\big)}{\theta_L},\quad k=1,2,3,\cdots.
\end{equation}
Let $k\rightarrow\infty$ in \eqref{67.1}, we immediately obtain
\begin{equation}\label{67.2}
\bigg\|\bigg(\ln\frac{u_{\alpha\beta}^2}{\underbar{n}}\bigg)_-\bigg\|_{L^{\infty}(\Omega)}\leq\frac{2\big(B^2+2|D|_0\big)}{\theta_L}.
\end{equation}
Note that $[\ln(u_{\alpha\beta}^2/\underbar{n})]_-$ is nonpositive, then the estimate \eqref{67.2} implies that
\begin{equation}\label{68.1}
u_{\alpha\beta}\geq\sqrt{\underbar{n}}e^{-(B^2+2|D|_0)/\theta_L}\geq b,\quad\text{if}\ \delta\ll1.
\end{equation}
Combining \eqref{62.1} with \eqref{68.1}, we have $b\leq u_{\alpha\beta}\leq B$, which means $u_{\alpha\beta}=u$. This gives the existence result of the problem $(P1)$. Differentiating the equation \eqref{109.1a} and using the regularity $u\in H^3(\Omega)$ to obtain the desired regularity $u\in H^4(\Omega)$ in Claim 1.
Before proving the uniqueness result of the problem $(P1)$, we need to establish the uniform estimate \eqref{eu} with respect to $\pc\in(0,1]$ for any $H^4$-solution $u$ of $(P1)$. To this end, the estimate \eqref{eu-a} can be proved similarly as the above derivation of $u_{\alpha\beta}=u$.
Furthermore, performing the procedure
\begin{equation}\label{117.1}
\int_0^1\bigg[\eqref{109.1a}\times\frac{1}{u}\bigg]_x\times u_xdx,
\end{equation}
we have
\begin{equation}\label{117.1r}
\int_0^1\pc^2\bigg(\frac{u_{xx}}{u}\bigg)_xu_xdx=\int_0^1\bigg[2\bigg(q-\frac{J^2}{u^4}\bigg)\frac{u_x}{u}+q_x-\varphi_x+\frac{J}{u^2}\bigg]u_xdx,
\end{equation}
where $J:=J[u^2,q]$ and $\varphi(x):=G[u^2](x)$. The left-side of \eqref{117.1r} can be estimated by using integration by parts,
\begin{equation}\label{117.1r-l}
\eqref{117.1r}_l=-\int_0^1\pc^2\frac{(u_{xx})^2}{u}dx\leq0.
\end{equation}
Based on the estimates \eqref{eu-a}, \eqref{154.2} and $|J|\leq C(b,B,N_1)\delta$, the right-side of \eqref{117.1r} can be estimated as follows provided $\delta$ is small enough,
\begin{equation}\label{117.1r-r}
\eqref{117.1r}_r\geq\frac{\theta_L}{2B}\|u_x\|^2-C(B,\theta_L)\|(q_x,\varphi_x)\|^2-J\bigg(\frac{1}{\rmd_{r}}-\frac{1}{\rmd_{l}}\bigg)\geq\frac{\theta_L}{2B}\|u_x\|^2-C,
\end{equation}
where the positive constant $C$ only depends on $n_l$, $\theta_L$ and $|D|_0$. We have used the elliptic estimate $\|\varphi\|_2\leq C(\|u^2-D\|+\|\phi_rx\|)$ in the last inequality of \eqref{117.1r-r}. Inserting \eqref{117.1r-l} and \eqref{117.1r-r} into \eqref{117.1r}, we get
\begin{equation}\label{117.3}
\|u_x\|\leq C,
\end{equation}
where the positive constant $C$ only depends on $n_l$, $\theta_L$ and $|D|_0$ and is independent of $\pc\in(0,1]$.
Performing the procedure
\begin{equation}\label{118.0}
\int_0^1\bigg[\eqref{109.1a}\times\frac{1}{u}\bigg]_x\times\bigg(\frac{u_{xx}}{u}\bigg)_xdx,
\end{equation}
and using integration by parts, we get
\begin{multline}\label{80.1}
\int_0^1\pc^2\bigg[\bigg(\frac{u_{xx}}{u}\bigg)_x\bigg]^2dx+\int_0^12S\bigg(\frac{u_{xx}}{u}\bigg)^2dx\\
=-\int_0^1\bigg(\frac{2S}{u}\bigg)_x\frac{u_xu_{xx}}{u}dx-\int_0^1q_{xx}\frac{u_{xx}}{u}dx+\int_0^1(u^2-D)\frac{u_{xx}}{u}dx+\int_0^1\frac{2J}{u^4}u_xu_{xx}dx,
\end{multline}
where $S:=q-J^2/u^4$. The left-side of \eqref{80.1} can be estimated as
\begin{equation}\label{80.2}
\eqref{80.1}_l\geq\frac{\theta_L}{2B^2}\|u_{xx}\|^2.
\end{equation}
The right-side of \eqref{80.1} can be estimated by H\"older, Sobolev and Cauchy-Schwarz inequalities as
\begin{align}
\eqref{80.1}_r&\leq C|u_x|_0\|u_x\|\|u_{xx}\|+C(\|q_{xx}\|+\|u^2-D\|+\|u_x\|)\|u_{xx}\|\notag\\
&\leq C(\|u_x\|^2+2\|u_x\|\|u_{xx}\|)^{1/2}\|u_{xx}\|+C\|u_{xx}\|\notag\\
&\leq\mu\|u_{xx}\|^2+C_\mu(\|u_x\|^2+2\|u_x\|\|u_{xx}\|)+C\|u_{xx}\|\notag\\
&\leq\mu\|u_{xx}\|^2+C_\mu(1+\|u_{xx}\|)\label{81.3}
\end{align}
where we have used the estimate \eqref{117.3}. The positive constant $C_\mu$ only depends on $n_l$, $\theta_L$, $|D|_0$ and $\mu$, where $\mu$ is a small number and will be determined later. Inserting \eqref{80.2} and \eqref{81.3} into \eqref{80.1}, and let $\mu\ll1$, we obtain $\|u_{xx}\|^2\leq C(1+\|u_{xx}\|)$. Solving this inequality with respect to $\|u_{xx}\|$ to obtain the estimate
\begin{equation}\label{82.4}
\|u_{xx}\|\leq C,
\end{equation}
where the positive constant $C$ only depends on $n_l$, $\theta_L$ and $|D|_0$ and is independent of $\pc\in(0,1]$.
Substituting the uniform estimates \eqref{eu-a}, \eqref{117.3} and \eqref{82.4} in the equality \eqref{80.1}, we have the estimate $\pc\|(u_{xx}/u)_x\|\leq C$. Note that
\[
\pc\pd{x}{3}u=\pc u\bigg(\frac{u_{xx}}{u}\bigg)_x+\pc\frac{u_xu_{xx}}{u}.
\]
We immediately get the following uniform estimate
\begin{equation}\label{85.2}
\|\pc\pd{x}{3}u\|\leq C.
\end{equation}
Furthermore, applying $\pd{x}{2}$ to the equation \eqref{109.1a} and taking the $L^2$-norm of the resultant equality, we finally have the following uniform estimate
\begin{equation}\label{86.1}
\|\pc^2\pd{x}{4}u\|\leq C.
\end{equation}
From the estimates \eqref{eu-a}, \eqref{117.3}, \eqref{82.4}, \eqref{85.2} and \eqref{86.1}, we have established the desired uniform estimate \eqref{eu-b} with respect to $\pc\in(0,1]$ for any strong solution $u$ to $(P1)$.
Based on the uniform estimate \eqref{eu}, now we can prove the uniqueness of solution to $(P1)$ by the energy method. To this end, we assume that $u_1$ and $u_2$ are two solutions to $(P1)$. Let $z_i:=\ln u_i^2$, $J_i:=J[e^{z_i},q]$, $S_i:=q-J_i^2/e^{2z_i}$, $\varphi_i:=G[e^{z_i}]$, $i=1,2$. Taking the difference $J_1$ and $J_2$, and applying the mean value theorem and \eqref{eu-a} to the explicit formula \eqref{104.1}, we have
\begin{equation}\label{138.0}
|J_i|\leq C(b,B,N_1)\delta,\qquad |J_1-J_2|\leq C\delta\|z_x\|,
\end{equation}
where $z:=z_1-z_2$ and $C$ is a positive constant which only depends on $n_l$, $\theta_L$ and $|D|_0$. Due to the procedure
\begin{equation}
\bigg[\eqref{109.1a}\times\frac{1}{u_1}\bigg]_x-\bigg[\eqref{109.1a}\times\frac{1}{u_2}\bigg]_x
\end{equation}
and the transformation $u_i=e^{z_i/2}$, the difference $z$ satisfies
\begin{equation}\label{95.3}
-\bigg(\frac{J_1^2}{e^{2z_1}}-\frac{J_2^2}{e^{2z_2}}\bigg)z_{1x}+S_2z_x-\frac{\pc^2}{2}\bigg[z_{xx}+\frac{z_{1x}^2}{2}-\frac{z_{2x}^2}{2}\bigg]_x=(\varphi_1-\varphi_2)_x-\bigg(\frac{J_1}{e^{z_1}}-\frac{J_2}{e^{z_2}}\bigg).
\end{equation}
Multiplying \eqref{95.3} by $z_x$, integrating the resultant equality and using the boundary conditions
\begin{equation}\label{94.5}
z_i(0)=\ln\md_l,\quad z_i(1)=\ln\md_r,\quad \bigg(z_{ixx}+\frac{z_{ix}^2}{2}\bigg)(0)=\bigg(z_{ixx}+\frac{z_{ix}^2}{2}\bigg)(1)=0
\end{equation}
to obtain that
\begin{multline}\label{96.1}
\int_0^1S_2z_x^2dx+\int_0^1\frac{\pc^2}{2}z_{xx}^2dx+\int_0^1\underbrace{(e^{z_1}-e^{z_2})z}_{\geq0}dx\\
=\int_0^1\bigg(\frac{J_1^2}{e^{2z_1}}-\frac{J_2^2}{e^{2z_2}}\bigg)z_{1x}z_xdx-\int_0^1\frac{\pc^2(z_{1x}+z_{2x})}{4}z_xz_{xx}dx-\int_0^1\bigg(\frac{J_1}{e^{z_1}}-\frac{J_2}{e^{z_2}}\bigg)z_xdx.
\end{multline}
The left-side of \eqref{96.1} can be estimated as
\begin{equation}\label{96.2}
\eqref{96.1}_l\geq\frac{\theta_L}{4}\|z_x\|^2+\frac{\pc^2}{2}\|z_{xx}\|^2,
\end{equation}
where we have used the estimates \eqref{154.2}, \eqref{eu} and \eqref{138.0}. The right-side of \eqref{96.1} can be estimated by H\"older, Poincar\'e and Cauchy-Schwarz inequalities and by the estimates \eqref{eu} and \eqref{138.0} as
\begin{align}
\eqref{96.1}_r&\leq C\Big(|J_1-J_2|\|z_x\|+|J_2|\|z\|\|z_x\|\Big)+C\pc^2\|z_x\|\|z_{xx}\|\notag\\
&\leq C(B,b,N_1)\delta\|z_x\|^2+\frac{\pc^2}{4}\|z_{xx}\|^2+C\pc^2\|z_x\|^2\notag\\
&=\Big(C(B,b,N_1)\delta+C\pc^2\Big)\|z_x\|^2+\frac{\pc^2}{4}\|z_{xx}\|^2.\label{97.3}
\end{align}
Substituting \eqref{96.2} and \eqref{97.3} in \eqref{96.1}, we see from letting $\delta$ and $\pc$ small enough that $\|z\|^2\leq0$. Thus we have shown $u_1\equiv u_2$. This complete the proof of Claim 1.
\emph{Step 4. Proof of Claim 2.}
For given function pair $(u,q)$ in the Claim 1, we discuss the unique solvability of the problem $(P2)$ by Leray-Schauder fixed-point theorem and energy method again. To this end, we define a fixed-point mapping $\mathcal{T}_2: q_1\mapsto Q_1$ over $H^1(\Omega)$ by solving the linear problem,
\begin{subequations}\label{126.1}
\begin{numcases}{}
\frac{2}{3}Q_{1xx}-JQ_{1x}+\frac{2}{3}J_{1*}(\ln u^2)_x\theta_L+\frac{2}{3}J(\ln u^2)_x(q_1-\theta_L)\notag\\
\qquad\qquad\qquad\qquad\qquad\qquad\ \,-u^2(Q_1-\theta_L)=g(u,q;\pc), \quad x\in\Omega,\label{126.1a}\\
Q_1(0)=\theta_l,\quad Q_1(1)=\theta_r,
\end{numcases}
\end{subequations}
where $J:=J[u^2,q]$ and $J_{1*}:=2\big(\bar{b}+\int_0^1q_{1x}\ln u^2dx\big)K[u^2,q]^{-1}$. In fact, for given $(u,q)$ in the Claim 1 and $q_1\in H^1(\Omega)$, the linear problem \eqref{126.1} is uniquely solvable in $H^3(\Omega)$ owing to the standard theory of the elliptic equations. By using the standard argument, we can further show that the mapping $\mathcal{T}_2$ is continuous and compact from $H^1(\Omega)$ into itself. Hence, it is sufficient to show that there exists a positive constant $M_2$ such that $\|\Theta\|_1\leq M_2$ for any $\Theta\in\big\{f\in H^1(\Omega)\,|\,f=\lambda\mathcal{T}_2f, \ \forall\lambda\in[0,1] \big\}$. We may assume $\lambda>0$ as the case $\lambda=0$ is trivial. Namely, for the function $\Theta$ verifying
\begin{subequations}\label{se127.2}
\begin{numcases}{}
\frac{2}{3}\Theta_{xx}-J\Theta_{x}+\frac{2}{3}\lambda J_*(\ln u^2)_x\theta_L+\frac{2}{3}\lambda J(\ln u^2)_x(\Theta-\theta_L)\notag\\
\qquad\qquad\qquad\qquad\qquad\qquad\ \,-u^2(\Theta-\lambda\theta_L)=\lambda g(u,q;\pc), \quad x\in\Omega, \label{127.2a}\\
\Theta(0)=\lambda\theta_l,\quad \Theta(1)=\lambda\theta_r,\qquad\forall\lambda\in[0,1],\label{127.2b}\\
J:=J[u^2,q],\quad J_*:=2\bigg(\bar{b}+\int_0^1\Theta_x\ln u^2dx\bigg)K[u^2,q]^{-1},\notag
\end{numcases}
\end{subequations}
we need to show the estimate $\|\Theta\|_1\leq M_2$.
Substituting $\Theta_\lambda:=\Theta-\lambda\bar{\theta}$ into the equation \eqref{127.2a}, where $\bar{\theta}(x):=\theta_l(1-x)+\theta_rx$, multiplying the resultant equation by $\Theta_\lambda$ and integrating it by parts over the domain $\Omega$ give
\begin{align}
&\frac{2}{3}\|\Theta_{\lambda x}\|^2+b^2\|\Theta_\lambda\|^2\notag\\
\leq&-\frac{2}{3}\lambda\theta_LJ_*\int_0^1\Theta_{\lambda x}\ln u^2dx-\int_0^1\bigg[\frac{2}{3}\lambda J\ln u^2\big(\Theta_\lambda^2\big)_x+J\Theta_{\lambda x}\Theta_\lambda\bigg]dx\notag\\
&-\frac{2}{3}\lambda J\int_0^1\ln u^2\Big[(\lambda\bar{\theta}-\theta_L)\Theta_\lambda\Big]_xdx-\lambda\int_0^1\Big[J\bar{\theta}_x+u^2(\bar{\theta}-\theta_L)+g(u,q;\pc)\Big]\Theta_\lambda dx\notag\\
\leq&-\frac{4\lambda\theta_L}{3K[u^2,q]}\bigg(\bar{b}+\int_0^1\big(\Theta_{\lambda x}+\lambda\bar{\theta}_x\big)\ln u^2dx\bigg)\int_0^1\Theta_{\lambda x}\ln u^2dx+C(b,B,N_1)\delta\|\Theta_{\lambda}\|_1^2\notag\\
&+\mu\|\Theta_{\lambda}\|_1^2+C(\mu,b,B,N_1)\delta^2\big(\delta^2+\|\lambda\bar{\theta}-\theta_L\|^2\big)+C(\mu,b,B)\delta^2\notag\\
\leq&\underbrace{-\frac{4\lambda\theta_L}{3K[u^2,q]}\bigg(\int_0^1\Theta_{\lambda x}\ln u^2dx\bigg)^2}_{\leq0}+\big[\mu+C(b,B,N_1)\delta\big]\|\Theta_{\lambda}\|_1^2\notag\\
&+C(\mu,b,B,N_1)\delta^2\big(\delta^2+\|\lambda\bar{\theta}-\theta_L\|^2\big)+C(\mu,b,B,\theta_L)\delta^2\notag\\
\leq&\big[\mu+C(b,B,N_1)\delta\big]\|\Theta_{\lambda}\|_1^2+C(\mu,b,B,N_1)\delta^2\big(\delta^2+\|\lambda\bar{\theta}-\theta_L\|^2\big)\notag\\
&+C(\mu,b,B,\theta_L)\delta^2,\label{133.0}
\end{align}
where we have used the expression \eqref{103}, the estimates \eqref{eu} and $|J|\leq C(b,B,N_1)\delta$, and the Young inequality. Taking $\mu$ and $\delta$ small enough in \eqref{133.0}, we have
\begin{equation}\label{133.3}
\|\Theta_{\lambda}\|_1^2\leq C(b,B,\theta_L)\delta^2+C(b,B,N_1)\delta^2\big(\delta^2+\|\lambda\bar{\theta}-\theta_L\|^2\big),
\end{equation}
which immediately means that
\begin{align}
\|\Theta\|_1=&\|\Theta_\lambda+\lambda\bar{\theta}\|_1\leq\|\Theta_\lambda\|_1+\lambda\|\bar{\theta}\|_1\notag\\
\leq&\sqrt{C(b,B,\theta_L)+C(b,B,N_1)(1+\theta_L^2)}+2\theta_L=:M_2.\label{es134.1}
\end{align}
Thus, the mapping $\mathcal{T}_2$ has a fixed point $Q=\mathcal{T}_2Q\in H^3(\Omega)$ by Leray-Schauder fixed-point theorem and the elliptic regularity theory. Hence, we have shown the existence of the solution $Q$ to the problem $(P2)$.
The uniqueness of the solution $Q$ follows from the energy method. Let $Q_i\in H^1(\Omega)$, $i=1,2$ be two solutions to $(P2)$ corresponding to the same function pair $(u,q)$. Define $\bar{Q}:=Q_1-Q_2$, which satisfies
\begin{subequations}\label{135.2}
\begin{numcases}{}
\frac{2}{3}\bar{Q}_{xx}-J\bar{Q}_{x}+\frac{4\theta_L}{3K[u^2,q]}\bigg(\int_0^1\bar{Q}_x\ln u^2dx\bigg)(\ln u^2)_x\notag\\
\qquad\qquad\qquad\qquad\qquad\qquad\ \,+\frac{2}{3}J(\ln u^2)_x\bar{Q}-u^2\bar{Q}=0, \quad x\in\Omega,\label{135.2a}\\
\bar{Q}(0)=\bar{Q}(1)=0.\label{135.2b}
\end{numcases}
\end{subequations}
Multiplying the equation \eqref{135.2a} by $-\bar{Q}$ and integrating the resultant equality over $\Omega$. In a similar way as the derivation of \eqref{133.0}, we have
\begin{align}
\frac{2}{3}\|\bar{Q}_x\|^2+b^2\|\bar{Q}\|^2\leq&-\frac{4\theta_L}{3K[u^2,q]}\bigg(\int_0^1\bar{Q}_x\ln u^2dx\bigg)^2+C(b,B,N_1)\delta\|\bar{Q}\|_1^2\notag\\
\leq&C(b,B,N_1)\delta\|\bar{Q}\|_1^2.\label{136.0}
\end{align}
Taking $\delta$ small enough in \eqref{136.0}, we see that $\|\bar{Q}\|_1^2\leq0$. Thus we have proven $Q_1\equiv Q_2$.
On the other hand, letting $\Theta=Q$ and $\lambda=1$ in the estimate \eqref{133.3}, we obtain
\begin{align}
\|Q-\theta_L\|_1\leq&\|Q-\bar{\theta}\|_1+\|\bar{\theta}-\theta_L\|_1\notag\\
\leq&C(b,B,\theta_L)\delta+C(b,B,N_1)\delta\big(\delta+\|\bar{\theta}-\theta_L\|\big)+\|\bar{\theta}-\theta_L\|_1\notag\\
\leq&C(b,B,\theta_L)\delta+C(b,B,N_1)\delta^2\notag\\
=&C_1\delta+C_2(b,B,N_1)\delta^2,\label{133.5}
\end{align}
which exactly is the desired estimate \eqref{eQa}. Solving the equation $\pd{x}{k}\eqref{109.2a}$ with respect to $\pd{x}{k}Q_{xx}$ for $k=0,1$ and directly taking the $L^2$-norm, we get the desired estimates \eqref{eQb} and \eqref{eQc} with the aid of the estimates \eqref{eu}, \eqref{133.5} and $|J|, |J_*|\leq C(b,B,N_1)\delta$. Consequently, the proof of Claim 2 is completed.
\emph{Step 5. End of the proof.}
Firstly, based on the estimate \eqref{eQ} we can determine the constants $N_1$ and $N_2$ by letting
\begin{equation}\label{N12}
N_1:=2C_1,\quad N_2:=C_3(b,B,2C_1).
\end{equation}
If $\delta$ is small enough, that is,
\begin{equation*}
\delta\leq\frac{C_1}{C_2(b,B,2C_1)},
\end{equation*}
then we see from the estimate \eqref{eQ} that $\mathcal{T}$ maps $\mathcal{U}[N_1,N_2]$ into itself. Combining the estimates \eqref{eu} and \eqref{eQ} with the Sobolev compact embedding theorem, via a standard argument, we see that the mapping $\mathcal{T}$ is continuous in the norm of $C^2(\overline{\Omega})$ and the image $\mathcal{T}\big(\mathcal{U}[N_1,N_2]\big)$ is precompact in $C^2(\overline{\Omega})$. Therefore, applying the Schauder fixed-point theorem to the mapping $\mathcal{T}:\mathcal{U}[N_1,N_2]\rightarrow\mathcal{U}[N_1,N_2]$, we obtain a fixed-point $\swnd\in\mathcal{U}[N_1,N_2]$ of the mapping $\mathcal{T}$. According to the construction of the mapping $\mathcal{T}$ above, we can easily see that $(\rsmd:=u[\swnd],\swnd)$ is a desired solution to the BVP \eqref{103.2}$\sim$\eqref{103.3}.
In addition, the solution $(\smd,\sdl,\swnd,\sdws)$ to the original BVP \eqref{1dsfqhd}$\sim$\eqref{sbc} is constructed from the the solution $(\rsmd,\swnd)$ to the BVP \eqref{103.2}$\sim$\eqref{103.3}. In fact, we define a function $\smd:=\rsmd^2$, the constant $\sdl:=J[\rsmd^2,\swnd]$ and a function $\sdws:=G[\rsmd^2]$, where $J[\cdot,\cdot]$ and $G[\cdot]$ are given in \eqref{104.1} and \eqref{104.2}. Then, we see that $(\smd,\sdl,\swnd,\sdws)\in H^4(\Omega)\times H^4(\Omega)\times H^3(\Omega)\times C^2(\overline{\Omega})$ is a desired solution to the BVP \eqref{1dsfqhd}$\sim$\eqref{sbc}. Moreover, this stationary solution satisfies the condition \eqref{psc} and the estimate \eqref{145.1}, thanks to the estimates \eqref{eu} and \eqref{eQ}.
Finally, using the same methods in Step 3 and Step 4, we can prove the local uniqueness of the stationary solution $(\smd,\sdl,\swnd,\sdws)$ to the BVP \eqref{1dsfqhd}$\sim$\eqref{sbc} if the parameters $\delta$ and $\pc$ are small enough and the solution additionally satisfies \eqref{psc} and \eqref{145.1}. The computations are standard but tedious, we omit the details.
\end{proof}
\section{Asymptotic stability of the stationary solution}\label{Sect.3}
In this section, we show Theorem \ref{thm2} by applying the standard continuation argument based on the local existence result and the uniform a priori estimate. To simplify the notations, we remove the superscript $\pc$ and denote the solution $(\pcmd{\pc},\pcdl{\pc},\pcwnd{\pc},\pcdws{\pc})$ in Theorem \ref{thm2} as $(\md,\dl,\wnd,\dws)$.
\subsection{Local existence}\label{Subsect.3.1}
In this subsection, we discuss the existence of the local-in-time solution. The proof is based on the iteration method and the energy estimates.
It is also convenient to make use of the transformation $\rmd:=\sqrt{\md}$ in the IBVP \eqref{1dfqhd}$\sim$\eqref{bc}. Then, we derive the equivalent IBVP for $(\rmd,\dl,\wnd,\dws)$ as follows
\begin{subequations}\label{a2.2}
\begin{numcases}{}
2\rmd\rmd_t+\dl_x=0, \label{a2.2-1}\\
\dl_t+2S[\rmd^2,\dl,\wnd]\rmd\rmd_x+\frac{2\dl}{\rmd^2}\dl_x+\rmd^2\wnd_x-\pc^2\rmd^2\Bigg(\frac{\rmd_{xx}}{\rmd}\Bigg)_x=\rmd^2\dws_x-\dl, \label{a2.2-2}\\
\rmd^2\wnd_t+\dl\wnd_x+\frac{2}{3}\rmd^2\wnd\Bigg(\frac{\dl}{\rmd^2}\Bigg)_x-\frac{2}{3}\wnd_{xx}-\frac{\pc^2}{3}\Bigg[\rmd^2\Bigg(\frac{\dl}{\rmd^2}\Bigg)_{xx}\Bigg]_x=\frac{1}{3}\frac{\dl^2}{\rmd^2}-\rmd^2(\wnd-\theta_{L}), \label{a2.2-3}\\
\dws_{xx}=\rmd^2-D(x), \qquad\forall t>0,\ \forall x\in\Omega,\label{a2.2-4}
\end{numcases}
\end{subequations}
with the initial condition
\begin{equation}\label{a2.3}
(\rmd,\dl,\wnd)(0,x)=(\rmd_0,\dl_0,\wnd_0)(x),\quad \rmd_0:=\sqrt{\md_0},
\end{equation}
and the boundary conditions
\begin{subequations}\label{a2.4}
\begin{gather}
\rmd(t,0)=\rmd_{l},\qquad \rmd(t,1)=\rmd_{r},\label{a2.4-1}\\
\rmd_{xx}(t,0)=\rmd_{xx}(t,1)=0,\label{a2.4-2}\\
\wnd(t,0)=\wnd_{l},\qquad \wnd(t,1)=\wnd_{r},\label{a2.4-3}\\
\dws(t,0)=0,\qquad \dws(t,1)=\phi_r.\label{a2.4-4}
\end{gather}
\end{subequations}
In the following discussion, we borrow the ideas in the papers \cite{NS08,NS09} which have shown the local existence theorems for the isothermal QHD model and the FHD model. Also, see \cite{KNN99,KNN03} for the general hyperbolic-elliptic coupled systems.
Now we are in the position to state the local existence.
\begin{lemma}\label{lema2}
Suppose that the initial data $(\rmd_0,\dl_0,\wnd_0)\in H^4(\Omega)\times H^3(\Omega)\times H^2(\Omega)$ and the boundary data satisfy the compatible condition
\begin{gather}
\rmd_0(0)=\rmd_l,\quad \rmd_0(1)=\rmd_r,\quad \wnd_0(0)=\theta_l,\quad \wnd_0(1)=\theta_r,\notag\\
\dl_{0x}(0)=\dl_{0x}(1)=\rmd_{0xx}(0)=\rmd_{0xx}(1)=0\label{rcompatibility}
\end{gather}
and the condition
\begin{equation}\label{ripsc}
\inf_{x\in\Omega}\rmd_{0}>0,\qquad \inf_{x\in\Omega}\wnd_0>0, \qquad\inf_{x\in\Omega}S[\rmd_{0}^2,\dl_{0},\wnd_0]>0.
\end{equation}
Then there exists a constant $T_*>0$ such that the IBVP \eqref{a2.2}$\sim$\eqref{a2.4} has a unique solution $(\rmd,\dl,\wnd,\dws)\in\big[\mathfrak{Y}_4([0,T_*])\cap H^2(0,T_*;H^1(\Omega))\big]\times\big[\mathfrak{Y}_3([0,T_*])\cap H^2(0,T_*;L^2(\Omega))\big]\times\big[\mathfrak{Y}_2([0,T_*])\cap H^1(0,T_*;H^1(\Omega))\big]\times\mathfrak{Z}([0,T_*])$ satisfying
\begin{equation}\label{rpsc}
\inf_{x\in\Omega}\rmd>0,\qquad \inf_{x\in\Omega}\wnd>0, \qquad\inf_{x\in\Omega}S[\rmd^2,\dl,\wnd]>0.
\end{equation}
\end{lemma}
To show Lemma \ref{lema2}, we first study the linear IBVP for the unknowns $(\hat{\rmd},\hat{\dl},\hat{\wnd})$
\begin{subequations}\label{a3.1}
\begin{numcases}{}
2\rmd\hat{\rmd}_t+\hat{\dl}_x=0, \label{a3.1-1}\\
\hat{\dl}_t+2S[\rmd^2,\dl,\wnd]\rmd\hat{\rmd}_x+\frac{2\dl}{\rmd^2}\hat{\dl}_x+\rmd^2\hat{\wnd}_x-\pc^2\rmd^2\Bigg(\frac{\hat{\rmd}_{xx}}{\rmd}\Bigg)_x=\rmd^2\dws_x-\dl, \label{a3.1-2}\\
\rmd^2\hat{\wnd}_t+\dl\hat{\wnd}_x+\frac{2}{3}\Bigg(\frac{\dl}{\rmd^2}\Bigg)_x\rmd^2\hat{\wnd}-\frac{2}{3}\hat{\wnd}_{xx}-\frac{\pc^2}{3}\Bigg[\rmd^2\Bigg(\frac{\dl}{\rmd^2}\Bigg)_{xx}\Bigg]_x=\frac{1}{3}\frac{\dl^2}{\rmd^2}-\rmd^2(\hat{\wnd}-\theta_{L}), \label{a3.1-3}\\
\dws:=\Phi[\rmd^2], \qquad\forall t>0,\ \forall x\in\Omega,\label{a3.1-4}
\end{numcases}
\end{subequations}
with the initial condition
\begin{equation}\label{a3.2}
(\hat{\rmd},\hat{\dl},\hat{\wnd})(0,x)=(\rmd_0,\dl_0,\wnd_0)(x),
\end{equation}
and the boundary conditions
\begin{subequations}\label{a3.3}
\begin{gather}
\hat{\rmd}(t,0)=\rmd_{l},\qquad \hat{\rmd}(t,1)=\rmd_{r},\label{a3.3-1}\\
\hat{\rmd}_{xx}(t,0)=\hat{\rmd}_{xx}(t,1)=0,\label{a3.3-2}\\
\hat{\wnd}(t,0)=\wnd_{l},\qquad \hat{\wnd}(t,1)=\wnd_{r},\label{a3.3-3}
\end{gather}
\end{subequations}
where the function $\dws$ is defined by \eqref{efep}. Let the functions $(\rmd,\dl,\wnd)$ in the coefficients in \eqref{a3.1} satisfy
\begin{subequations}\label{a4.1}
\begin{gather}
(\rmd,\dl,\wnd)(0,x)=(\rmd_0,\dl_0,\wnd_0)(x),\quad\forall x\in\Omega, \label{a4.1-1}\\
\rmd\in\mathfrak{Y}_4([0,T])\cap H^2(0,T;H^1(\Omega)),\quad \dl\in\mathfrak{Y}_3([0,T])\cap H^2(0,T;L^2(\Omega)),\notag\\
\wnd\in\mathfrak{Y}_2([0,T])\cap H^1(0,T;H^1(\Omega)),\label{a4.1-2}\\
\rmd(t,x),\ \wnd(t,x),\ S[\rmd^2,\dl,\wnd](t,x)\geq m,\quad\forall (t,x)\in[0,T]\times\Omega,\label{a4.1-3}\\
\|\rmd(t)\|_4^2+\|\dl(t)\|_3^2+\|(\rmd_t,\wnd)(t)\|_2^2+\|\dl_t(t)\|_1^2+\|(\rmd_{tt},\wnd_t)(t)\|^2\notag\\
+\int_0^t\|(\rmd_{ttx},\dl_{tt},\wnd_{tx})(\tau)\|^2d\tau\leq M,\quad\forall t\in[0,T],\label{a4.1-4}
\end{gather}
\end{subequations}
where $T$, $m$ and $M$ are positive constants. We denote by $X(T;m,M)$ the set of functions $(\rmd,\dl,\wnd)$ satisfying \eqref{a4.1}, and we abbreviate $X(T;m,M)$ by $X(\cdot)$ without confusion. The property of $\dws$ is that
\[
\dws\in\mathfrak{Z}([0,T]),\quad \|\pd{t}{i}\dws(t)\|_2^2\leq M,\quad\forall t\in[0,T], \ i=0,1,2.
\]
Then the next lemma means that for suitably chosen constants $T$, $m$ and $M$, the set $X(\cdot)$ is invariant under the mapping $(\rmd,\dl,\wnd)\mapsto(\hat{\rmd},\hat{\dl},\hat{\wnd})$ defined by solving the linear IBVP \eqref{a3.1}$\sim$\eqref{a3.3}. We discuss the solvability of this linear problem in Appendix. Since the next lemma is proved similarly as in \cite{NS08,NS09}, we omit the proof.
\begin{lemma}\label{lema3}
Under the same assumptions in Lemma \ref{lema2}, there exist positive constants $T$, $m$ and $M$ with the following property: If $(\rmd,\dl,\wnd)\in X(\cdot)$, then the linear IBVP \eqref{a3.1}$\sim$\eqref{a3.3} admits a unique solution $(\hat{\rmd},\hat{\dl},\hat{\wnd})$ in the same set $X(\cdot)$.
\end{lemma}
Using Lemma \ref{lema3}, we can show Lemma \ref{lema2}.
\begin{proof}[\textbf{Proof of Lemma \ref{lema2}}]
We define the approximation sequence $\{(\rmd^k,\dl^k,\wnd^k)\}_{k=0}^\infty$ by letting $(\rmd^0,\dl^0,\wnd^0)=(\rmd_0,\dl_0,\wnd_0)$ and solving
\begin{subequations}\label{a5.1}
\begin{numcases}{}
2\rmd^k\rmd^{k+1}_t+\dl^{k+1}_x=0, \label{a5.1-1}\\
\dl^{k+1}_t+2S[(\rmd^k)^2,\dl^k,\wnd^k]\rmd^k\rmd^{k+1}_x+\frac{2\dl^k}{(\rmd^k)^2}\dl^{k+1}_x+(\rmd^k)^2\wnd^{k+1}_x\notag\\
\qquad\qquad\qquad\qquad\qquad\qquad\qquad\quad-\pc^2(\rmd^k)^2\Bigg(\frac{\rmd^{k+1}_{xx}}{\rmd^k}\Bigg)_x=(\rmd^k)^2\dws^k_x-\dl^k, \label{a5.1-2}\\
(\rmd^k)^2\wnd^{k+1}_t+\dl^k\wnd^{k+1}_x+\frac{2}{3}\Bigg(\frac{\dl^k}{(\rmd^k)^2}\Bigg)_x(\rmd^k)^2\wnd^{k+1}-\frac{2}{3}\wnd^{k+1}_{xx}\notag\\
\qquad\qquad\qquad\quad-\frac{\pc^2}{3}\Bigg[(\rmd^k)^2\Bigg(\frac{\dl^k}{(\rmd^k)^2}\Bigg)_{xx}\Bigg]_x=\frac{1}{3}\frac{(\dl^k)^2}{(\rmd^k)^2}-(\rmd^k)^2(\wnd^{k+1}-\theta_{L}), \label{a5.1-3}\\
\dws^k:=\Phi[(\rmd^k)^2], \qquad\forall t>0,\ \forall x\in\Omega,\label{a5.1-4}
\end{numcases}
\end{subequations}
with the initial condition
\begin{equation}\label{a5.2}
(\rmd^{k+1},\dl^{k+1},\wnd^{k+1})(0,x)=(\rmd_0,\dl_0,\wnd_0)(x),
\end{equation}
and the boundary conditions
\begin{subequations}\label{a5.3}
\begin{gather}
\rmd^{k+1}(t,0)=\rmd_{l},\qquad \rmd^{k+1}(t,1)=\rmd_{r},\label{a5.3-1}\\
\rmd^{k+1}_{xx}(t,0)=\rmd^{k+1}_{xx}(t,1)=0,\label{a5.3-2}\\
\wnd^{k+1}(t,0)=\wnd_{l},\qquad \wnd^{k+1}(t,1)=\wnd_{r},\label{a5.3-3}
\end{gather}
\end{subequations}
Thanks to Lemma \ref{lema3}, the sequence $\{(\rmd^k,\dl^k,\wnd^k)\}_{k=0}^\infty$ is well defined and contained in $X(\cdot)$. Consequently, $(\rmd^k,\dl^k,\wnd^k)$ satisfies the estimates \eqref{a4.1-3} and \eqref{a4.1-4}. Next, applying the standard energy method to the system satisfied by the difference $(\rmd^{k+1}-\rmd^k,\dl^{k+1}-\dl^k,\wnd^{k+1}-\wnd^k)$, we see that $\{(\rmd^k,\dl^k,\wnd^k)\}_{k=0}^\infty$ is the Cauchy sequence in $\mathfrak{Y}_2([0,T_*])\times\mathfrak{Y}_1([0,T_*])\times\big[\mathfrak{Y}_2([0,T_*])\cap H^1(0,T_*;H^1(\Omega))\big]$ for small enough $0<T_*\leq T$. In showing this fact, we obtain the estimates of the higher-order derivatives in the time variable $t$ and then rewrite them into those in the spatial variable $x$ by using the linear equations. Thus, there exists a function $(\rmd,\dl,\wnd)\in\mathfrak{Y}_2([0,T_*])\times\mathfrak{Y}_1([0,T_*])\times\big[\mathfrak{Y}_2([0,T_*])\cap H^1(0,T_*;H^1(\Omega))\big]$ such that $(\rmd^k,\dl^k,\wnd^k)\rightarrow(\rmd,\dl,\wnd)$ strongly in $\mathfrak{Y}_2([0,T_*])\times\mathfrak{Y}_1([0,T_*])\times\big[\mathfrak{Y}_2([0,T_*])\cap H^1(0,T_*;H^1(\Omega))\big]$ as $k\rightarrow\infty$. Moreover, it holds $(\rmd,\dl)\in\big[\mathfrak{Y}_4([0,T_*])\cap H^2(0,T_*;H^1(\Omega))\big]\times\big[\mathfrak{Y}_3([0,T_*])\cap H^2(0,T_*;L^2(\Omega))\big]$ by the standard argument (see \cite{NS08,NS09} for example). Define $\dws:=\Phi[\rmd^2]$ by the limit function $\rmd$ and the explicit formula \eqref{efep}, we see that $(\rmd,\dl,\wnd,\dws)$ is the desired solution to the IBVP \eqref{a2.2}$\sim$\eqref{a2.4}. Notice that this solution also satisfies \eqref{rpsc}.
\end{proof}
\subsection{A priori estimate}\label{Subsect.3.2}
To show the asymptotic stability of the stationary solution $(\rsmd,\sdl,\swnd,\sdws)$, we introduce the perturbations around the stationary solution $(\rsmd,\sdl,\swnd,\sdws)$ below
\begin{align}
\prmd(t,x)&:=\rmd(t,x)-\rsmd(x),&\pdl(t,x)&:=\dl(t,x)-\sdl,\notag\\
\pwnd(t,x)&:=\wnd(t,x)-\swnd(x),&\pdws(t,x)&:=\dws(t,x)-\sdws(x).\label{10.1}
\end{align}
Taking the difference between the transient system \eqref{1dfqhd} and the stationary system \eqref{1dsfqhd} via the following procedure
\begin{equation*}\label{10.0}
\eqref{1dfqhd1}-\eqref{1dsfqhd1},\quad\eqref{1dfqhd2}/\rmd^2-\eqref{1dsfqhd2}/\rsmd^2,\quad\eqref{1dfqhd3}-\eqref{1dsfqhd3},\quad\eqref{1dfqhd4}-\eqref{1dsfqhd4},
\end{equation*}
we can derive the perturbed system for the perturbations $(\prmd,\pdl,\pwnd,\pdws)$ as
\begin{subequations}\label{10.2}
\begin{numcases}{}
2(\prmd+\rsmd)\prmd_t+\pdl_x=0, \label{10.2a}\\
\bigg[\frac{\pdl+\sdl}{(\prmd+\rsmd)^2}\bigg]_t+\frac{1}{2}\Bigg\{\bigg[\frac{\pdl+\sdl}{(\prmd+\rsmd)^2}\bigg]^2-\bigg(\frac{\sdl}{\rsmd^2}\bigg)^2\Bigg\}_x+\pwnd\Big[\ln(\prmd+\rsmd)^2\Big]_x\notag\\
\quad+\swnd\Big[\ln(\prmd+\rsmd)^2-\ln\rsmd^2\Big]_x+\pwnd_x-\pc^2\bigg[\frac{(\prmd+\rsmd)_{xx}}{\prmd+\rsmd}-\frac{\rsmd_{xx}}{\rsmd}\bigg]_x\notag\\
\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad=\pdws_x-\bigg[\frac{\pdl+\sdl}{(\prmd+\rsmd)^2}-\frac{\sdl}{\rsmd^2}\bigg], \label{10.2b}\\
(\prmd+\rsmd)^2\pwnd_t-\frac{2}{3}\pwnd_{xx}+\frac{2}{3}\swnd\pdl_x-\frac{4\sdl\swnd}{3\rsmd}\prmd_x\notag\\
\qquad\qquad\quad-\frac{\pc^2}{3}\Bigg\{(\prmd+\rsmd)^2\bigg[\frac{\pdl+\sdl}{(\prmd+\rsmd)^2}\bigg]_{xx}-\rsmd^2\bigg(\frac{\sdl}{\rsmd^2}\bigg)_{xx}\Bigg\}_x=H(t,x), \label{10.2c}\\
\pdws_{xx}=(\prmd+2\rsmd)\prmd,\label{10.2d}
\end{numcases}
\end{subequations}
where the right-side term of the perturbed energy equation \eqref{10.2c} is defined by
\begin{equation}\label{11.1}
H(t,x):=\frac{4\swnd(\prmd+\rsmd)_x}{3(\prmd+\rsmd)}\pdl-\rsmd^2\pwnd+H_1(t,x),
\end{equation}
and
\begin{align}
H_1(t,x):=&\frac{4\pwnd(\prmd+\rsmd)_x}{3(\prmd+\rsmd)}\pdl-(\prmd+2\rsmd)(\pwnd+\swnd-\theta_L)\prmd\notag\\
&-(\pwnd+\swnd)_x\pdl-\sdl\pwnd_x-\frac{2\pdl_x}{3}\pwnd+\frac{4\sdl\swnd(\prmd+\rsmd)_x}{3\rsmd^2}\prmd+\frac{4\sdl(\prmd+\rsmd)_x}{3(\prmd+\rsmd)}\pwnd\notag\\
&+\frac{4\sdl\swnd(\prmd+2\rsmd)(\prmd+\rsmd)_x}{3\rsmd^2(\prmd+\rsmd)}\prmd+\frac{\pdl+2\sdl}{3(\prmd+\rsmd)^2}\pdl-\frac{\sdl^2(\prmd+2\rsmd)}{3\rsmd^2(\prmd+\rsmd)^2}\prmd.\label{11.0}
\end{align}
The initial and the boundary conditions to the system \eqref{10.2} are derived from \eqref{ic}, \eqref{bc} and \eqref{sbc} as
\begin{gather}
\prmd(0,x)=\prmd_0(x):=\rmd_0(x)-\rsmd(x),\quad\pdl(0,x)=\pdl_0(x):=\dl_0(x)-\sdl,\notag\\
\pwnd(0,x)=\pwnd_0(x):=\wnd_0(x)-\swnd(x),\label{pic}
\end{gather}
and
\begin{subequations}\label{pbc}
\begin{gather}
\prmd(t,0)=\prmd(t,1)=0,\label{pbc1}\\
\prmd_{xx}(t,0)=\prmd_{xx}(t,1)=0,\label{pbc2}\\
\pwnd(t,0)=\pwnd(t,1)=0,\label{pbc3}\\
\pdws(t,0)=\pdws(t,1)=0.\label{pbc4}
\end{gather}
\end{subequations}
Theorem \ref{thm1} and Lemma \ref{lema2} ensure the local existence of the solution $(\prmd,\pdl,\pwnd,\pdws)$ to the IBVP \eqref{10.2}$\sim$\eqref{pbc}. It is summarized in the next corollary.
\begin{corollary}\label{cor1}
Suppose that $(\prmd_0,\pdl_0,\pwnd_0)\in H^4(\Omega)\times H^3(\Omega)\times H^2(\Omega)$ and $(\prmd_0+\rsmd,\pdl_0+\sdl,\pwnd_0+\swnd)$ satisfies \eqref{rcompatibility} and \eqref{ripsc}. Then there exists a constant $T_*>0$ such that the IBVP \eqref{10.2}$\sim$\eqref{pbc} has a unique solution $(\prmd,\pdl,\pwnd,\pdws)\in\big[\mathfrak{Y}_4([0,T_*])\cap H^2(0,T_*;H^1(\Omega))\big]\times\big[\mathfrak{Y}_3([0,T_*])\cap H^2(0,T_*;L^2(\Omega))\big]\times\big[\mathfrak{Y}_2([0,T_*])\cap H^1(0,T_*;H^1(\Omega))\big]\times\mathfrak{Y}_4^2([0,T_*])$ with the property that $(\prmd+\rsmd,\pdl+\sdl,\pwnd+\swnd)$ satisfies \eqref{rpsc}.
\end{corollary}
To show the global existence of the solution, the key step is to derive the a priori estimate \eqref{127.2} for the local solution in Corollary \ref{cor1}. The next three subsections are devoted to the proof of Proposition \ref{prop1}, where the following notations are frequently used.
\begin{equation}\label{16.1}
\pcN:=\sup_{t\in[0,T]}\pcn,\quad \pcn:=\|(\prmd,\pdl,\pwnd)(t)\|_2+\|(\pc\pd{x}{3}\prmd,\pc\pd{x}{3}\pdl,\pc^2\pd{x}{4}\prmd)(t)\|.
\end{equation}
\begin{proposition}\label{prop1}
Let $(\prmd,\pdl,\pwnd,\pdws)$ be a solution to the IBVP \eqref{10.2}$\sim$\eqref{pbc} which belongs to $\big[\mathfrak{Y}_4([0,T])\cap H^2(0,T;H^1(\Omega))\big]\times\big[\mathfrak{Y}_3([0,T])\cap H^2(0,T;L^2(\Omega))\big]\times\big[\mathfrak{Y}_2([0,T])\cap H^1(0,T;H^1(\Omega))\big]\times\mathfrak{Y}_4^2([0,T])$. Then there exist positive constants $\delta_0$, $C$ and $\gamma$ such that if $\pcN+\delta+\pc\leq\delta_0$, then the following estimate holds for $t\in[0,T]$,
\begin{equation}\label{127.2}
\pcn+\|\pdws(t)\|_4\leq Cn_\pc(0)e^{-\gamma t},
\end{equation}
where $C$ and $\gamma$ are two positive constants independent of $\delta$, $\pc$ and $T$.
\end{proposition}
Using the Sobolev inequality, the estimate \eqref{145.1}, the perturbed system \eqref{10.2} and the notation \eqref{16.1}, we can derive some frequently used estimates in the next lemma. Since the proof is straightforward and tedious, we omit the details.
\begin{lemma}
Under the same assumptions as in Proposition \ref{prop1}, the following estimates hold for $t\in[0,T]$,
\begin{gather}
|\rsmd|_1+\big|\big(\pc^{1/2}\rsmd_{xx},\pc^{3/2}\rsmd_{xxx}\big)\big|_0\leq C,\quad |\sdl|+|\swnd-\theta_L|_2\leq C\delta,\label{17.1}\\
|(\prmd,\pdl,\pwnd)(t)|_1+\big|\big(\pc^{1/2}\prmd_{xx},\pc^{1/2}\pdl_{xx},\pc^{3/2}\prmd_{xxx},\prmd_t,\pdl_t\big)(t)\big|_0\leq C\pcN,\label{16.3+16.5}\\
\|\pd{t}{i}\pdws(t)\|_2\leq C\bigg[\|\pd{t}{i}\prmd(t)\|+\frac{i(i-1)}{2}\pcN\|\prmd_t(t)\|\bigg],\quad i=0,1,2,\label{17.2}\\
\|\pdws_{tx}(t)\|\leq\|\pdl(t)\|,\quad\|\pdws(t)\|_4\leq C\|\prmd(t)\|_2,\label{17.3}\\
\|\pd{x}{l}\pdl_x(t)\|\leq C\|\prmd_t(t)\|_l,\quad\|\pd{x}{l}\prmd_t(t)\|\leq C\|\pdl_x(t)\|_l,\quad l=0,1,2,\label{32.4+44.2}\\
\|\pdl_{tx}(t)\|\leq C\|(\prmd_t,\prmd_{tt})(t)\|,\quad\|\pdl_{txx}(t)\|\leq C\|(\prmd_{tt},\prmd_{tx},\prmd_{ttx})(t)\|,\label{38.2c+38.2d}
\end{gather}
where the positive constant $C$ is independent of $\delta$, $\pc$ and $T$.
\end{lemma}
\subsection{Basic estimate}\label{Subsect.3.3}
In this subsection, we derive the following basic estimate.
\begin{lemma}\label{be}
Suppose the same assumptions as in Proposition \ref{prop1} hold. Then there exist positive constants $\delta_0$, $c$ and $C$ such that if $\pcN+\delta+\pc\leq\delta_0$, it holds that for $t\in[0,T]$,
\begin{equation}\label{25.1}
\frac{d}{dt}\Xi(t)+c\Pi(t)\leq C\Gamma(t),
\end{equation}
where
\begin{equation}\label{25.2}
\Xi(t):=\int_0^1\bigg\{\bigg[\frac{1}{2\rmd^2}\pdl^2+\swnd\rmd^2\Psi\bigg(\frac{\rsmd^2}{\rmd^2}\bigg)+\pc^2\prmd_x^2+\frac{1}{2}\pdws_x^2\bigg]+\frac{3\rmd^2}{4\swnd}\pwnd^2-\alpha\bigg(\frac{\dl}{\rmd^2}-\frac{\sdl}{\rsmd^2}\bigg)\pdws_x\bigg\}dx,
\end{equation}
here $\alpha\in(0,1)$ is a small constant which will be determined later and $\Psi(s):=s-1-\ln s$ for $s>0$,
\begin{equation}\label{D}
\Pi(t):=\|(\prmd,\pc\prmd_x,\pdl,\pwnd,\pwnd_x)(t)\|^2,
\end{equation}
and
\begin{equation}\label{R}
\Gamma(t):=\big(\pcN+\delta+\pc^{3/2}\big)\|(\prmd_x,\pdl_x)(t)\|^2+\pc^3\|(\prmd_{xx},\pdl_{xx})(t)\|^2.
\end{equation}
Furthermore, if $\alpha$ is small enough, then the following equivalent relation holds true,
\begin{equation}\label{25.3}
c\|(\prmd,\pdl,\pwnd,\pc\prmd_x)(t)\|^2\leq\Xi(t)\leq C\|(\prmd,\pdl,\pwnd,\pc\prmd_x)(t)\|^2,
\end{equation}
where the constants $c$ and $C$ are independent of $\delta$, $\pc$ and $T$.
\end{lemma}
\begin{proof}
Firstly, multiplying the equation \eqref{10.2b} by $\pdl$ and applying Leibniz formula to the resultant equality together with the equations \eqref{10.2a} and \eqref{10.2d}, we obtain
\begin{equation}\label{18.1}
\pd{t}{}\bigg[\frac{1}{2\rmd^2}\pdl^2+\swnd\rmd^2\Psi\bigg(\frac{\rsmd^2}{\rmd^2}\bigg)+\pc^2\prmd_x^2+\frac{1}{2}\pdws_x^2\bigg]+\frac{1}{\rsmd^2}\pdl^2+\frac{2\rsmd_x}{\rmd}\pwnd\pdl+\pwnd_x\pdl=\pd{x}{}R_{1}(t,x)+R_2(t,x),
\end{equation}
where
\begin{equation}\label{18.3a}
R_1(t,x):=\pdws\pdws_{tx}+\pdws\pdl-\swnd\Big(\ln\rmd^2-\ln\rsmd^2\Big)\pdl+\pc^2\bigg[\bigg(\frac{\rmd_{xx}}{\rmd}-\frac{\rsmd_{xx}}{\rsmd}\bigg)\pdl+2\prmd_t\prmd_x\bigg],
\end{equation}
\begin{align}
R_2(t,x):=&-\frac{\pdl+2\sdl}{2\rmd^4}\pdl\pdl_x-\frac{1}{2}\Bigg[\bigg(\frac{\dl}{\rmd^2}\bigg)^2-\bigg(\frac{\sdl}{\rsmd^2}\bigg)^2\Bigg]_x\pdl\notag\\
&-\frac{2\prmd_x}{\rmd}\pwnd\pdl+\swnd_x\Big(\ln\rmd^2-\ln\rsmd^2\Big)\pdl+\frac{(\rmd+\rsmd)\dl}{\rmd^2\rsmd^2}\prmd\pdl+\pc^2\frac{\rsmd_{xx}}{\rsmd\rmd}\prmd\pdl_x.\label{18.3b}
\end{align}
Applying the estimates \eqref{17.1}$\sim$\eqref{16.3+16.5} and Cauchy-Schwarz inequality to \eqref{18.3b}, we have the following pointwise estimate
\begin{equation}\label{18.4b}
R_2(t,x)\leq C\big(\pcN+\delta+\pc^{3/2}\big)|(\prmd,\pdl,\prmd_x,\pdl_x)(t,x)|^2.
\end{equation}
In addition, multiplying the equation \eqref{10.2c} by $3\pwnd/(2\swnd)$ and applying Leibniz formula to the resultant equality, we obtain
\begin{equation}\label{19.1}
\pd{t}{}\bigg(\frac{3\rmd^2}{4\swnd}\pwnd^2\bigg)+\frac{3\rsmd^2}{2\swnd}\pwnd^2+\frac{1}{\swnd}\pwnd_x^2-\frac{2\rsmd_x}{\rmd}\pdl\pwnd-\pdl\pwnd_x=\pd{x}{}R_3(t,x)+R_4(t,x),
\end{equation}
where
\begin{equation}\label{19.2a}
R_3(t,x):=\frac{1}{\swnd}\pwnd\pwnd_x-\pdl\pwnd+\frac{\pc^2}{2\swnd}\bigg[\rmd^2\bigg(\frac{\dl}{\rmd^2}\bigg)_{xx}-\rsmd^2\bigg(\frac{\sdl}{\rsmd^2}\bigg)_{xx}\bigg]\pwnd,
\end{equation}
\begin{align}
R_4(t,x):=&\frac{2\prmd_x}{\rmd}\pdl\pwnd-\frac{3\pdl_x}{4\swnd}\pwnd^2+\frac{\swnd_x}{\swnd^2}\pwnd\pwnd_x+\frac{2\sdl}{\rsmd}\prmd_x\pwnd+\frac{3}{2\swnd}H_1(t,x)\pwnd\notag\\
&+\frac{\pc^2\swnd_x}{2\swnd^2}\bigg[\rmd^2\bigg(\frac{\dl}{\rmd^2}\bigg)_{xx}-\rsmd^2\bigg(\frac{\sdl}{\rsmd^2}\bigg)_{xx}\bigg]\pwnd-\frac{\pc^2}{2\swnd}\bigg[\rmd^2\bigg(\frac{\dl}{\rmd^2}\bigg)_{xx}-\rsmd^2\bigg(\frac{\sdl}{\rsmd^2}\bigg)_{xx}\bigg]\pwnd_x.\label{19.2b}
\end{align}
Applying the estimates \eqref{17.1}$\sim$\eqref{16.3+16.5} and Cauchy-Schwarz inequality to \eqref{19.2b} together with \eqref{32.1+33.1} for $k=0$, we have the following pointwise estimate
\begin{align}
R_4(t,x)\leq&C\big(\pcN+\delta+\pc^3\big)|(\prmd,\pdl,\prmd_x,\pdl_x)(t,x)|^2\notag\\
&+C\pc|(\pwnd,\pwnd_x)(t,x)|^2+C\pc^3|(\prmd_{xx},\pdl_{xx})(t,x)|^2.\label{20.2b}
\end{align}
Since the stationary density $\rsmd$ is non-flat, see \eqref{17.1}, we have to capture the dissipation rate of the perturbed density $\prmd$ in establishing the basic estimate. To this end, multiplying the equation \eqref{10.2b} by $-\pdws_x$ and applying Leibniz formula to the resultant equality, we obtain
\begin{equation}\label{22.1}
-\pd{t}{}\bigg[\bigg(\frac{\dl}{\rmd^2}-\frac{\sdl}{\rsmd^2}\bigg)\pdws_x\bigg]+\swnd(\rmd+\rsmd)\Big(\ln\rmd^2-\ln\rsmd^2\Big)\prmd+\frac{\rmd+\rsmd}{\rmd}\pc^2\prmd_x^2+\pdws_x^2=\pd{x}{}R_5(t,x)+R_6(t,x),
\end{equation}
where
\begin{equation}\label{22.2a}
R_5(t,x):=\swnd\Big(\ln\rmd^2-\ln\rsmd^2\Big)\pdws_x-\pc^2\bigg(\frac{\rmd_{xx}}{\rmd}-\frac{\rsmd_{xx}}{\rsmd}\bigg)\pdws_x+\pc^2\frac{\rmd+\rsmd}{\rmd}\prmd_x\prmd,
\end{equation}
\begin{align}
R_6:=&-\bigg(\frac{\dl}{\rmd^2}-\frac{\sdl}{\rsmd^2}\bigg)\pdws_{tx}+\frac{1}{2}\Bigg[\bigg(\frac{\dl}{\rmd^2}\bigg)^2-\bigg(\frac{\sdl}{\rsmd^2}\bigg)^2\Bigg]_x\pdws_x+\Big(\ln\rmd^2\Big)_x\pwnd\pdws_x\notag\\
&-\swnd_x\Big(\ln\rmd^2-\ln\rsmd^2\Big)\pdws_x+\pwnd_x\pdws_x+\bigg(\frac{\dl}{\rmd^2}-\frac{\sdl}{\rsmd^2}\bigg)\pdws_x+\pc^2\frac{(\rmd+\rsmd)\prmd}{\rmd^2}\prmd_x^2\notag\\
&-\pc^2\frac{\rsmd_{xx}(\rmd+\rsmd)}{\rmd\rsmd}\prmd^2+\pc^2\frac{(\rmd+\rsmd)\rsmd_x}{\rmd^2}\prmd_x\prmd-\pc^2\frac{(\rmd+\rsmd)_x}{\rmd}\prmd_x\prmd.\label{22.2b}
\end{align}
Similarly, we also have the pointwise estimate
\begin{align}
R_6(t,x)\leq&(\mu+C\delta)|\pdws_x(t,x)|^2+C\big(\pcN+\delta\big)|(\prmd,\prmd_x,\pdl_x)(t,x)|^2\notag\\
&+C_\mu|(\pwnd,\pwnd_x,\pdl)(t,x)|^2+C|(\pdl,\pdws_{tx})(t,x)|^2+C\big(\pcN+\pc\big)|(\prmd,\pc\prmd_x)(t,x)|^2,\label{22.3b}
\end{align}
In particular, applying the mean value theorem to the second term on the left-side of \eqref{22.1}, and using the estimates \eqref{17.1}$\sim$\eqref{16.3+16.5}, the second and the third terms on the left-side of \eqref{22.1} can be further treated as
\begin{equation}\label{23.0}
\text{(the 2nd and 3rd terms)}\geq c|(\prmd,\pc\prmd_x)(t,x)|^2,
\end{equation}
and the quantity in the first term on the left-side of \eqref{22.1} can be estimated as
\begin{equation}\label{23.2b}
\bigg|-\bigg(\frac{\dl}{\rmd^2}-\frac{\sdl}{\rsmd^2}\bigg)\pdws_x\bigg|\leq C|(\prmd,\pdl,\pdws_x)(t,x)|^2.
\end{equation}
Substituting \eqref{22.3b} and \eqref{23.0} into \eqref{22.1}, letting $\mu$ and $\pcN+\delta+\pc$ be small enough, we have
\begin{align}
-\pd{t}{}\bigg[\bigg(\frac{\dl}{\rmd^2}-\frac{\sdl}{\rsmd^2}\bigg)\pdws_x\bigg]+c|(\prmd,\pc\prmd_x)(t,x)|^2\leq&\pd{x}{}R_5(t,x)+C|(\pdws_{tx},\pdl,\pwnd,\pwnd_x)(t,x)|^2\notag\\
&+C\big(\pcN+\delta\big)|(\prmd_x,\pdl_x)(t,x)|^2.\label{23.1}
\end{align}
Finally, from the following procedure
\begin{equation*}
\int_0^1\Big[\eqref{18.1}+\eqref{19.1}+\alpha\eqref{23.1}\Big]dx,
\end{equation*}
where $\alpha$ is an arbitrary positive constant to be determined, we obtain
\begin{multline}\label{24.2}
\frac{d}{dt}\Xi(t)+\int_0^1\bigg(\frac{1}{\rsmd^2}\pdl^2+\frac{3\rsmd^2}{2\swnd}\pwnd^2+\frac{1}{\swnd}\pwnd_x^2\bigg)dx+c\alpha\|(\prmd,\pc\prmd_x)(t)\|^2\\
\leq\int_0^1\Big[R_2(t,x)+R_4(t,x)\Big]dx+C\alpha\|(\pdws_{tx},\pdl,\pwnd,\pwnd_x)(t)\|^2+C\alpha\big(\pcN+\delta\big)\|(\prmd_x,\pdl_x)(t)\|^2,
\end{multline}
where we have used the fact
\begin{equation}\label{18.4a+20.1b+23.2a}
\int_0^1\pd{x}{}\Big[R_1(t,x)+R_3(t,x)+\alpha R_5(t,x)\Big]=0,
\end{equation}
which follows from the boundary conditions \eqref{pbc}. Applying the estimates \eqref{17.1}, \eqref{17.3}, \eqref{18.4b}, \eqref{20.2b} and \eqref{23.2b} to the inequality \eqref{24.2}, and then letting $\alpha$ and $\pcN+\delta+\pc$ be sufficiently small. These procedures yield the desired estimates \eqref{25.1} and \eqref{25.3}.
\end{proof}
\subsection{Higher order estimates}\label{Subsect.3.4}
This subsection is devoted to the derivation of the higher order estimates. In order to use the homogeneous boundary condition \eqref{pbc}, we first establish the estimates of the temporal derivatives of the perturbations $(\prmd,\pdl,\pwnd)$. And then, we find that the spatial derivatives of the perturbations $(\prmd,\pdl,\pwnd,\pdws)$ can be bounded by the temporal derivatives of the perturbations $(\prmd,\pdl,\pwnd)$ with the help of the special structure of the perturbed system \eqref{10.2}. To justify the above mentioned computations, we need to use the mollifier arguments with respect to the time variable $t$ because the regularity of the local solution $(\prmd,\pdl,\pwnd)$ is insufficient. However, we omit these arguments since they are standard.
It is convenient to intoduce the notations
\begin{gather*}\label{39.1}
A_{-1}(t):=\|(\prmd,\pdl,\pwnd,\pwnd_x)(t)\|,\\
A_k(t):=A_{-1}(t)+\sum_{i=0}^k\|(\pd{t}{i}\prmd_t,\pd{t}{i}\prmd_x,\pc\pd{t}{i}\prmd_{xx})(t)\|,\quad k=0,1.
\end{gather*}
Differentiating \eqref{1dfqhd2} with respect to $x$ and multiplying the result by $1/\rmd$. Similarly, differentiating \eqref{1dsfqhd2} with respect to $x$ and multiplying the result by $1/\rsmd$. Furthermore, taking the difference between the two resultant equalities and substituting the equations \eqref{10.2a} and \eqref{10.2d} in the resultant equation. Then applying the operator $\pd{t}{k}$ for $k=0,1$ to the result, we obtain the equation
\begin{align}
2\pd{t}{k}\prmd_{tt}-&2\swnd\pd{t}{k}\prmd_{xx}+\pc^2\pd{t}{k}\prmd_{xxxx}+2\pd{t}{k}\prmd_t-\rsmd\pd{t}{k}\pwnd_{xx}-2\rsmd_{xx}\pd{t}{k}\pwnd\notag\\
=&\frac{2(\pdl+\sdl)}{(\prmd+\rsmd)^3}\pd{t}{k}\pdl_{xx}-\frac{2(\pdl+\sdl)^2}{(\prmd+\rsmd)^4}\pd{t}{k}\prmd_{xx}+2\pwnd\pd{t}{k}\prmd_{xx}+\prmd\pd{t}{k}\pwnd_{xx}\notag\\
&+\pc^2\frac{(k+1)\prmd_{xx}+2\rsmd_{xx}}{\prmd+\rsmd}\pd{t}{k}\prmd_{xx}+\pd{t}{k}P(t,x)+O_k(t,x),\quad k=0,1,\label{12.3}
\end{align}
where
\begin{align*}
P(t,x):=&-(\prmd+\rsmd)(\prmd+2\rsmd)\prmd-(\rsmd^2-D)\prmd-\frac{2(\prmd+\rsmd)_x^2(\pwnd+\swnd)}{(\prmd+\rsmd)\rsmd}\prmd\notag\\
&+\frac{2\rsmd_x^2}{\rsmd}\pwnd+4\rsmd_x\pwnd_x-2(\prmd+\rsmd)_x\pdws_x\notag\\
&+\swnd_{xx}\prmd+\frac{6(\prmd+\rsmd)_x^2(\pdl+\sdl)^2}{(\prmd+\rsmd)^5\rsmd^5}\big[\rsmd^5-(\prmd+\rsmd)^5\big]+\frac{6\rsmd_x^2(\pdl+2\sdl)}{\rsmd^5}\pdl\notag\\
&-\frac{2(\pdl+\sdl)^2\big[\rsmd^4-(\prmd+\rsmd)^4\big]}{(\prmd+\rsmd)^4\rsmd^4}\rsmd_{xx}-\frac{2(\pdl+2\sdl)\pdl}{\rsmd^4}\rsmd_{xx}-\frac{\pc^2\rsmd_{xx}^2}{(\prmd+\rsmd)\rsmd}\prmd\\
&+\frac{6(\prmd+2\rsmd)_x(\pdl+\sdl)^2}{\rsmd^5}\prmd_x+\frac{2(\pwnd+\swnd)}{\rsmd}\prmd_x^2+\frac{2(\prmd+2\rsmd)_x\pwnd}{\rsmd}\prmd_x+4(\pwnd+\swnd)_x\prmd_x\notag\\
&-\frac{2}{\prmd+\rsmd}\prmd_t^2+\frac{2}{(\prmd+\rsmd)^3}\pdl_x^2-\frac{8(\prmd+\rsmd)_x(\pdl+\sdl)}{(\prmd+\rsmd)^4}\pdl_x+2\bigg(\frac{2\rsmd_x\swnd}{\rsmd}-\sdws_x\bigg)\prmd_x,
\end{align*}
and
\begin{align*}
&O_0(t,x):=0,& O_1(t,x):=&-\frac{6(\pdl+\sdl)}{(\prmd+\rsmd)^4}\prmd_t\pdl_{xx}+\frac{2}{(\prmd+\rsmd)^3}\pdl_t\pdl_{xx}+\frac{8(\pdl+\sdl)^2}{(\prmd+\rsmd)^5}\prmd_t\prmd_{xx}\notag\\
&&&-\frac{4(\pdl+\sdl)}{(\prmd+\rsmd)^4}\pdl_t\prmd_{xx}+2\pwnd_t\prmd_{xx}+\prmd_t\pwnd_{xx}-\frac{\pc^2(\prmd_{xx}+2\rsmd_{xx})\prmd_{xx}}{(\prmd+\rsmd)^2}\prmd_t.
\end{align*}
According to \eqref{17.1}$\sim$\eqref{38.2c+38.2d}, we show the estimate
\begin{align}
\|\pd{t}{k}P(t)\|+\|O_k(t)\|\leq&C\|(\pd{t}{k}\prmd,\pd{t}{k}\pwnd,\pd{t}{k}\pwnd_{x})(t)\|\notag\\
&+C\big(\pcN+\delta+\pc^{1/2}\big)\|(\pd{t}{k}\prmd_{t},\pd{t}{k}\prmd_{x},\pd{t}{k}\pdl)(t)\|,\quad k=0,1,\label{26.1+26.4+27.2+27.3+29.1}
\end{align}
where $C$ is a positive constant independent of $\delta$, $\pc$ and $T$. In deriving the $L^2$-norm estimate of $\|\pd{t}{k}P(t)\|$, we have used the equation \eqref{1dsfqhd2} and the estimate \eqref{17.1} to deal with the coefficient of the last term in the expression of $P(t,x)$ as
\begin{equation*}
\Bigg|2\bigg(\frac{2\rsmd_x\swnd}{\rsmd}-\sdws_x\bigg)\Bigg|=\Bigg|2\bigg[\frac{2\sdl^2}{\rsmd^5}\rsmd_x-\swnd_x+\pc^2\bigg(\frac{\rsmd_{xx}}{\rsmd}\bigg)_x-\frac{\sdl}{\rsmd^2}\bigg]\Bigg|\leq C\big(\delta+\pc^{1/2}\big).
\end{equation*}
Applying the operator $\pd{t}{k}$ for $k=0,1$ to \eqref{10.2c}, we have
\begin{multline}\label{15.1}
(\prmd+\rsmd)^2\pd{t}{k}\pwnd_t-\frac{2}{3}\pd{t}{k}\pwnd_{xx}+\frac{2}{3}\swnd\pd{t}{k}\pdl_x-\frac{4\sdl\swnd}{3\rsmd}\pd{t}{k}\prmd_x\\
=\pd{x}{}\mathcal{V}_k(t,x)+\pd{t}{k}H(t,x)+L_k(t,x),\quad k=0,1,
\end{multline}
where
\begin{align}
\mathcal{V}_k(t,x):=&\frac{\pc^2}{3}\pd{t}{k}\Bigg\{(\prmd+\rsmd)^2\bigg[\frac{\pdl+\sdl}{(\prmd+\rsmd)^2}\bigg]_{xx}-\rsmd^2\bigg(\frac{\sdl}{\rsmd^2}\bigg)_{xx}\Bigg\}\notag\\
=&\frac{\pc^2}{3}\pd{t}{k}\pdl_{xx}-\frac{2\pc^2\sdl}{3\rsmd}\pd{t}{k}\prmd_{xx}+\pd{t}{k}\mathcal{K}(t,x),\quad k=0,1,\label{32.1+33.1}
\end{align}
\begin{align}
\mathcal{K}(t,x):=\frac{\pc^2}{3}\bigg[-\frac{4(\prmd+\rsmd)_x}{\prmd+\rsmd}&\pdl_x+\frac{6(\prmd+\rsmd)_x^2}{(\prmd+\rsmd)^2}\pdl-\frac{6\sdl(\prmd+2\rsmd)(\prmd+\rsmd)_x^2}{(\prmd+\rsmd)^2\rsmd^2}\prmd\notag\\
+&\frac{6\sdl(\prmd+2\rsmd)_x}{\rsmd^2}\prmd_x-\frac{2(\prmd+\rsmd)_{xx}}{\prmd+\rsmd}\pdl+\frac{2\sdl(\prmd+\rsmd)_{xx}}{(\prmd+\rsmd)\rsmd}\prmd\bigg],\label{32.2a}
\end{align}
\begin{equation}
L_0(t,x):=0,\quad L_1(t,x):=-2(\prmd+\rsmd)\prmd_t\pwnd_t.
\end{equation}
For convenience, we further calculate the $\pd{x}{}\mathcal{V}_k(t,x)$ as
\begin{equation}\label{34.1}
\pd{x}{}\mathcal{V}_0(t,x)=\frac{\pc^2}{3}\pdl_{xxx}-\frac{2\pc^2\sdl}{3\rsmd}\prmd_{xxx}+\underbrace{\frac{2\pc^2\sdl\rsmd_x}{\rsmd^2}\prmd_{xx}+\pd{x}{}\mathcal{K}(t,x)}_{=:\mathcal{K}_1(t,x)},
\end{equation}
and
\begin{equation}\label{37.2}
\pd{x}{}\mathcal{V}_1(t,x)=\frac{\pc^2}{3}\pdl_{txxx}-\frac{2\pc^2(\pdl+\sdl)}{3(\prmd+\rsmd)}\prmd_{txxx}+\mathcal{K}_2(t,x),
\end{equation}
where
\allowdisplaybreaks
\begin{align}
\mathcal{K}_2(t,x):=\frac{\pc^2}{3}\bigg[-&\frac{4(\prmd+\rsmd)_x}{\prmd+\rsmd}\pdl_{txx}+\frac{4(\prmd+\rsmd)_x\pdl_{xx}}{(\prmd+\rsmd)^2}\prmd_t-\frac{4\pdl_{xx}}{\prmd+\rsmd}\prmd_{tx}\notag\\
&+\frac{10(\prmd+\rsmd)_x^2}{(\prmd+\rsmd)^2}\pdl_{tx}-\frac{20(\prmd+\rsmd)_x^2\pdl_x}{(\prmd+\rsmd)^3}\prmd_t+\frac{20(\prmd+\rsmd)_x\pdl_x}{(\prmd+\rsmd)^2}\prmd_{tx}\notag\\
&-\frac{6(\prmd+\rsmd)_{xx}}{\prmd+\rsmd}\pdl_{tx}+\frac{6(\prmd+\rsmd)_{xx}\pdl_x}{(\prmd+\rsmd)^2}\prmd_t-\frac{6\pdl_x}{\prmd+\rsmd}\prmd_{txx}\notag\\%line3
&-\frac{12(\prmd+\rsmd)_x^3}{(\prmd+\rsmd)^3}\pdl_t+\frac{36(\pdl+\sdl)(\prmd+\rsmd)_x^3}{(\prmd+\rsmd)^4}\prmd_t-\frac{36(\pdl+\sdl)(\prmd+\rsmd)_x^2}{(\prmd+\rsmd)^3}\prmd_{tx}\notag\\
&+\frac{14(\prmd+\rsmd)_{x}(\prmd+\rsmd)_{xx}}{(\prmd+\rsmd)^2}\pdl_t-\frac{28(\pdl+\sdl)(\prmd+\rsmd)_{x}(\prmd+\rsmd)_{xx}}{(\prmd+\rsmd)^3}\prmd_t\notag\\
&\qquad\qquad\qquad+\frac{14(\pdl+\sdl)(\prmd+\rsmd)_{xx}}{(\prmd+\rsmd)^2}\prmd_{tx}+\frac{14(\pdl+\sdl)(\prmd+\rsmd)_x}{(\prmd+\rsmd)^2}\prmd_{txx}\notag\\
&-\frac{2(\prmd+\rsmd)_{xxx}}{\prmd+\rsmd}\pdl_t+\frac{2(\pdl+\sdl)(\prmd+\rsmd)_{xxx}}{(\prmd+\rsmd)^2}\prmd_t\bigg].\label{37.1}
\end{align}
According to \eqref{17.1}$\sim$\eqref{38.2c+38.2d}, we further show the estimates
\begin{gather}
\|H(t)\|\leq C\|(\pdl,\pwnd)(t)\|+C\big(\pcN+\delta\big)\|(\prmd,\pwnd_x)(t)\|,\label{29.2a}\\
\|\pd{t}{}H(t)\|+\|L_1(t)\|\leq C\|(\pdl_t,\pwnd_t)(t)\|+C\big(\pcN+\delta\big)\|(\prmd_t,\prmd_{tt},\prmd_{tx},\pwnd_{tx})(t)\|,\label{30.1a+30.2}\\
\|\mathcal{K}(t)\|\leq C\pc^{3/2}\|(\prmd,\prmd_t,\prmd_x,\pdl)(t)\|,\label{32.2b}\\
\|\pd{t}{}\mathcal{K}(t)\|\leq C\pc^{3/2}\|\prmd_t(t)\|+C\pc^2\|(\prmd_{tt},\prmd_{tx},\pdl_t)(t)\|+C\big(\pcN+\delta\big)\pc\|\pc\prmd_{txx}(t)\|,\label{33.2b}\\
\|\mathcal{K}_1(t)\|\leq C\pc^{1/2}\|(\prmd,\pdl)(t)\|+C\pc^{3/2}\|(\prmd_t,\prmd_x)(t)\|+C\pc^2\|(\prmd_{tx},\prmd_{xx})(t)\|,\label{34.2c}\\
\|\mathcal{K}_2(t)\|\leq C\pc^{1/2}\|(\prmd_t,\pdl_t)(t)\|+C\pc^{3/2}\|(\prmd_{tt},\prmd_{tx})(t)\|+C\pc\|(\pc\prmd_{ttx},\pc\prmd_{txx})(t)\|,\label{37.1b}
\end{gather}
where the positive constant $C$ is independent of $\delta$, $\pc$ and $T$.
The following lemma is important for the strategy in which we establish the a priori estimate \eqref{127.2}. This means that the spatial derivatives of the perturbations $(\prmd,\pdl,\pwnd)$ can be controlled by their temporal derivatives.
\begin{lemma}\label{lem5}
Under the same assumptions as in Proposition \ref{prop1}, the following equivalent relation holds for $t\in[0,T]$,
\begin{equation}\label{46.5}
c\big(A_1(t)+\|\pwnd_t(t)\|\big)\leq\pcn\leq C\big(A_1(t)+\|\pwnd_t(t)\|\big),
\end{equation}
where the two positive constants $c$ and $C$ are independent of $\delta$, $\pc$ and $T$.
\end{lemma}
\begin{proof}
We only show the right-side inequality in \eqref{46.5} because the left-side inequality in \eqref{46.5} can be established by the similar method and the corresponding computations are much easier than the right-side one. According to the definitions of the notations $\pcn$, $A_{-1}(t)$, $A_0(t)$ and $A_1(t)$, and using the estimates \eqref{17.1}$\sim$\eqref{38.2c+38.2d}, \eqref{26.1+26.4+27.2+27.3+29.1} with $k=0$, \eqref{29.2a} and \eqref{34.2c}, the equation \eqref{12.3} with $k=0$, the equation \eqref{15.1} with $k=0$ and the equality \eqref{34.1}, we have
\begin{align}
\pcn=&\|(\prmd,\pdl,\pwnd)(t)\|_2+\|(\pc\pd{x}{3}\prmd,\pc\pd{x}{3}\pdl,\pc^2\pd{x}{4}\prmd)(t)\|\notag\\
\leq&CA_1(t)+\|\pc^2\pd{x}{4}\prmd(t)\|+\|\pwnd_{xx}(t)\|+\|\prmd_{xx}(t)\|+\|\pc\pd{x}{3}\prmd(t)\|\notag\\
=&CA_1(t)+\|\pwnd_{xx}(t)\|+\|\prmd_{xx}(t)\|+\|\pc\pd{x}{3}\prmd(t)\|\notag\\
&+\Big\|\big[\eqref{12.3}_r|_{k=0}-\big(2\prmd_{tt}-2\swnd\prmd_{xx}+2\prmd_t-\rsmd\pwnd_{xx}-2\rsmd_{xx}\pwnd\big)\big](t)\Big\|\notag\\
\leq&CA_1(t)+\|\pwnd_{xx}(t)\|+\|\prmd_{xx}(t)\|+\|\pc\pd{x}{3}\prmd(t)\|\notag\\
&+C\|(\prmd_t,\prmd_{tt},\pdl_{xx},P,\pwnd,\pwnd_{x},\pwnd_{xx},\prmd_{xx})(t)\|\notag\\
\leq&C\big(A_1(t)+\|\pwnd_{xx}(t)\|+\|\prmd_{xx}(t)\|+\|\pc\pd{x}{3}\prmd(t)\|\big)\notag\\
=&C\big(A_1(t)+\|\prmd_{xx}(t)\|+\|\pc\pd{x}{3}\prmd(t)\|\big)\notag\\
&+C\bigg\|-\frac{3}{2}\bigg[-(\prmd+\rsmd)^2\pwnd_t-\frac{2\swnd}{3}\pdl_x+\frac{4\sdl\swnd}{3\rsmd}\prmd_x+\frac{\pc^2}{3}\pd{x}{3}\pdl-\frac{2\pc^2\sdl}{3\rsmd}\pd{x}{3}\prmd+\mathcal{K}_1+H\bigg](t)\bigg\|\notag\\
\leq&C\big(A_1(t)+\|\prmd_{xx}(t)\|+\|\pc\pd{x}{3}\prmd(t)\|\big)+C\big(A_1(t)+\|\pwnd_t(t)\|+\pc\|\pc\pd{x}{3}\prmd(t)\|\big)\notag\\
\leq&C\big(A_1(t)+\|\pwnd_t(t)\|+\|\prmd_{xx}(t)\|+\|\pc\pd{x}{3}\prmd(t)\|\big).\label{42.2}
\end{align}
Moreover, multiplying the equation \eqref{12.3} with $k=0$ by $-\prmd_{xx}$ and integrating by parts with using the boundary condition \eqref{pbc2}, we obtain
\begin{align}
&\theta_{L}\|\prmd_{xx}(t)\|^2+\|\pc\pd{x}{3}\prmd(t)\|^2\notag\\
\leq&-\int_0^1\Big[\eqref{12.3}_r|_{k=0}-(2\prmd_{tt}+2\prmd_t-\rsmd\pwnd_{xx}-2\rsmd_{xx}\pwnd)\Big]\prmd_{xx}dx\notag\\
\leq&\big[\mu+C\big(\pcN+\delta+\pc^{3/2}\big)\big]\|\prmd_{xx}(t)\|^2+C_\mu\|(\prmd_t,\prmd_{tt},\prmd_{tx},\pwnd,\pwnd_x,\pwnd_{xx},P)(t)\|^2\notag\\
\leq&\big[\mu+C\big(\pcN+\delta+\pc^{3/2}\big)\big]\|\prmd_{xx}(t)\|^2+C_\mu\big(A_1^2(t)+\|\pwnd_{xx}(t)\|^2\big)\notag\\
\leq&\big[\mu+C\big(\pcN+\delta+\pc^{3/2}\big)\big]\|\prmd_{xx}(t)\|^2+C_\mu\big(A_1^2(t)+\|\pwnd_t(t)\|^2+\pc^2\|\pc\pd{x}{3}\prmd(t)\|^2\big),\label{42.0}
\end{align}
Let $\mu$ and $\pcN+\delta+\pc$ small enough, the inequality \eqref{42.0} implies
\begin{equation}\label{42.1}
\|\prmd_{xx}(t)\|+\|\pc\pd{x}{3}\prmd(t)\|\leq C\big(A_1(t)+\|\pwnd_t(t)\|\big)
\end{equation}
Substituting \eqref{42.1} into \eqref{42.2}, we have $\pcn\leq C(A_1(t)+\|\pwnd_t(t)\|)$
\end{proof}
For convenience of later use, we estimate the $L^2$-norm of $\pd{t}{k}\pdl_t$ for $k=0,1$ in the next lemma.
\begin{lemma}
Under the same assumptions as in Proposition \ref{prop1}, the following estimates hold for $t\in[0,T]$,
\begin{equation}\label{P79-70.3a}
\|\pdl_t(t)\|\leq C\|(\prmd,\pdl,\pwnd,\pwnd_x,\prmd_x)(t)\|+C\big(\pcN+\delta\big)\|\prmd_t(t)\|+C\pc^{1/2}\big(A_1(t)+\|\pwnd_t(t)\|\big),
\end{equation}
and
\begin{subequations}\label{P79-72.2+P79-78.1a}
\begin{equation}\label{P79-72.2}
\pd{t}{}\pdl_t=\pc^2\rmd\prmd_{txxx}+Y_1(t,x),
\end{equation}
\begin{align}
\|Y_1(t)\|\leq&C\|(\prmd,\pdl,\pwnd,\pwnd_x,\prmd_t,\prmd_x,\prmd_{tx},\pwnd_{tx})(t)\|\notag\\
&+C\big(\pcN+\delta+\pc^{1/2}\big)\|(\prmd_{tt},\pc\prmd_{xx},\pc\prmd_{txx})(t)\|,\label{P79-78.1a}
\end{align}
\end{subequations}
where $Y_1(t,x)$ is given by \eqref{72.3} and the positive constant $C$ is independent of $\delta$, $\pc$ and $T$.
\end{lemma}
\begin{proof}
Solving the equation \eqref{10.2b} with respect to $\pdl_t$, we have
\begin{equation}\label{es68.1}
\pdl_t=\pc^2\rmd^2\bigg(\frac{\rmd_{xx}}{\rmd}-\frac{\rsmd_{xx}}{\rsmd}\bigg)_x+Y(t,x),
\end{equation}
where
\begin{align}
Y(t,x):=&\frac{2\dl}{\rmd}\prmd_t-\frac{\rmd^2}{2}\Bigg[\bigg(\frac{\dl}{\rmd^2}\bigg)^2-\bigg(\frac{\sdl}{\rsmd^2}\bigg)^2\Bigg]_x-\rmd^2\pwnd\big(\ln\rmd^2\big)_x\notag\\
&-\rmd^2\swnd\big(\ln\rmd^2-\ln\rsmd^2\big)_x-\rmd^2\pwnd_x+\rmd^2\pdws_x-\rmd^2\bigg(\frac{\dl}{\rmd^2}-\frac{\sdl}{\rsmd^2}\bigg).
\end{align}
Taking the $L^2$-norm of \eqref{es68.1} directly, and applying the estimates \eqref{17.1}, \eqref{16.3+16.5}, \eqref{17.3} and \eqref{42.1} to the resultant equality, we obtain the desired estimate \eqref{P79-70.3a}.
Next, differentiating the equation \eqref{es68.1} with respect to the time variable $t$, we get the equality \eqref{P79-72.2}, where $Y_1(t,x)$ is defined by
\begin{align}
Y_1(t,x):=&-\pc^2\rmd_x\prmd_{txx}-\pc^2\rmd_{xxx}\prmd_t+\frac{2\pc^2\rmd_{xx}\rmd_x}{\rmd}\prmd_t\notag\\
&-\pc^2\rmd_{xx}\prmd_{tx}+2\pc^2\rmd\prmd_t\bigg(\frac{\rmd_{xx}}{\rmd}-\frac{\rsmd_{xx}}{\rsmd}\bigg)_x+\pd{t}{}Y(t,x).\label{72.3}
\end{align}
Similarly, taking the $L^2$-norm of \eqref{72.3} directly, and applying the estimates \eqref{17.1}, \eqref{16.3+16.5}, \eqref{17.3}, \eqref{38.2c+38.2d}, \eqref{42.1} and \eqref{P79-70.3a} to the resultant equality, we have the desired estimate \eqref{P79-78.1a}.
\end{proof}
Now, we begin to derive the higher order estimates to complete the a priori estimate \eqref{127.2}. From the following lemma, we can see that the Bohm potential term in the momentum equation contributes the quantum dissipation rate $\|\pc\pd{t}{k}\prmd_{xx}(t)\|$.
\begin{lemma}
Suppose the same assumptions as in Proposition \ref{prop1} hold. Then there exist positive constants $\delta_0$, $c$ and $C$ such that if $\pcN+\delta+\pc\leq\delta_0$, it holds that for $t\in[0,T]$,
\begin{equation}\label{115.1}
\frac{d}{dt}\Xi_1^{(k)}(t)+c\Pi_1^{(k)}(t)\leq C\Gamma_1^{(k)}(t),\quad k=0,1,
\end{equation}
where
\begin{gather}
\Xi_1^{(k)}(t):=\int_0^1\Big[\big(\pd{t}{k}\prmd\big)^2+2\pd{t}{k}\prmd_t\pd{t}{k}\prmd\Big]dx,\quad \Pi_1^{(k)}(t):=\|(\pd{t}{k}\prmd_x,\pc\pd{t}{k}\prmd_{xx})(t)\|^2,\notag\\
\Gamma_1^{(k)}(t):=\|(\pd{t}{k}\prmd_t,\pd{t}{k}\prmd,\pd{t}{k}\pwnd,\pd{t}{k}\pwnd_x)(t)\|^2+\big(\pcN+\delta+\pc^{1/2}\big)A_k^2(t),\label{115.1a}
\end{gather}
and the constants $c$ and $C$ are independent of $\delta$, $\pc$ and $T$.
\end{lemma}
\begin{proof}
Multiplying the equation \eqref{12.3} by $\pd{t}{k}\prmd$ and integrating the resultant equality by parts over the domain $\Omega=(0,1)$ together with the homogeneous boundary conditions \eqref{pbc}, we get
\begin{equation}\label{48.1}
\frac{d}{dt}\Xi_1^{(k)}(t)+\int_0^1\Big[2\swnd\big(\pd{t}{k}\prmd_x\big)^2+\big(\pc\pd{t}{k}\prmd_{xx}\big)^2\Big]dx=\mathcal{I}_1^{(k)}(t),\quad k=0,1,
\end{equation}
where the integral term $\mathcal{I}_1^{(k)}(t)$ is defined by
\begin{align}
\mathcal{I}_1^{(k)}(t):=\int_0^1\bigg\{&2\big(\pd{t}{k}\prmd_t\big)^2-2\swnd_x\pd{t}{k}\prmd_x\pd{t}{k}\prmd-\Big(\rsmd_x\pd{t}{k}\pwnd_x\pd{t}{k}\prmd+\rsmd\pd{t}{k}\pwnd_x\pd{t}{k}\prmd_x\Big)\notag\\
&+2\rsmd_{xx}\pd{t}{k}\pwnd\pd{t}{k}\prmd+\bigg(\frac{6\rmd_x\dl}{\rmd^4}\pd{t}{k}\pdl_x\pd{t}{k}\prmd-\frac{2\pdl_x}{\rmd^3}\pd{t}{k}\pdl_x\pd{t}{k}\prmd-\frac{2\dl}{\rmd^3}\pd{t}{k}\pdl_x\pd{t}{k}\prmd_x\bigg)\notag\\
&-\bigg[\frac{8\rmd_x\dl^2}{\rmd^5}\pd{t}{k}\prmd_x\pd{t}{k}\prmd-\frac{4\dl\pdl_x}{\rmd^4}\pd{t}{k}\prmd_x\pd{t}{k}\prmd-\frac{2\dl^2}{\rmd^4}\big(\pd{t}{k}\prmd_x\big)^2\bigg]\notag\\
&-\Big[2\pwnd_x\pd{t}{k}\prmd_x\pd{t}{k}\prmd+2\pwnd\big(\pd{t}{k}\prmd_x\big)^2\Big]-\Big(\prmd_x\pd{t}{k}\pwnd_x\pd{t}{k}\prmd+\prmd\pd{t}{k}\pwnd_x\pd{t}{k}\prmd_x\Big)\notag\\
&\qquad\quad+\pc^2\frac{(k+1)\prmd_{xx}+2\rsmd_{xx}}{\rmd}\pd{t}{k}\prmd_{xx}\pd{t}{k}\prmd+\Big(\pd{t}{k}P+O_k\Big)\pd{t}{k}\prmd\bigg\}dx.\label{48.0}
\end{align}
According to the estimates \eqref{17.1}, \eqref{16.3+16.5}, \eqref{38.2c+38.2d}, \eqref{26.1+26.4+27.2+27.3+29.1} and \eqref{P79-70.3a}, and then using Cauchy-Schwarz, Young and H\"older inequalities, we have the estimates
\begin{equation}\label{52.1}
\int_0^1\Big[2\swnd\big(\pd{t}{k}\prmd_x\big)^2+\big(\pc\pd{t}{k}\prmd_{xx}\big)^2\Big]dx\geq c\Pi_1^{(k)}(t),
\end{equation}
and
\begin{align}
\mathcal{I}_1^{(k)}(t)\leq&2\|\pd{t}{k}\prmd_t(t)\|^2+C\delta\|(\pd{t}{k}\prmd_x,\pd{t}{k}\prmd)(t)\|^2+\mu\|\pd{t}{k}\prmd_x(t)\|^2+C_\mu\|(\pd{t}{k}\pwnd_x,\pd{t}{k}\prmd)(t)\|^2\notag\\
&+C\|(\pd{t}{k}\pwnd,\pd{t}{k}\pwnd_x,\pd{t}{k}\prmd)(t)\|^2+C\big(\pcN+\delta\big)\|(\pd{t}{k}\prmd,\pd{t}{k}\prmd_x,\pd{t}{k}\pdl_x)(t)\|^2\notag\\
&+C\pc^{1/2}\|(\pd{t}{k}\prmd,\pc\pd{t}{k}\prmd_{xx})(t)\|^2+C\|(\pd{t}{k}P,O_k,\pd{t}{k}\prmd)(t)\|^2\notag\\
\leq&\mu\|\pd{t}{k}\prmd_x(t)\|^2+C_\mu\|(\pd{t}{k}\prmd_t,\pd{t}{k}\prmd,\pd{t}{k}\pwnd,\pd{t}{k}\pwnd_x)(t)\|^2\notag\\
&+C\big(\pcN+\delta+\pc^{1/2}\big)A_k^2(t).\label{52.2}
\end{align}
Substituting \eqref{52.1} and \eqref{52.2} into \eqref{48.1} and taking $\mu$ small enough, we obtain the desired estimate \eqref{115.1}.
\end{proof}
Next, the following lemma is the most difficult part in establishing the higher order estimates. From this lemma, we can find that the dispersive velocity term in the energy equation contributes the extra quantum dissipation rate $\|\pc\pd{t}{k}\prmd_{tx}(t)\|$, see \eqref{100.2}. It plays a similar role like the additional dissipation rate $\|\pwnd_{tx}(t)\|$ contributed by the diffusion term in the energy equation, see \eqref{126.2}.
\begin{lemma}
Suppose the same assumptions as in Proposition \ref{prop1} hold. Then there exist positive constants $\delta_0$, $c$ and $C$ such that if $\pcN+\delta+\pc\leq\delta_0$, it holds that for $t\in[0,T]$,
\begin{equation}\label{115.2}
\frac{d}{dt}\Xi_2^{(k)}(t)+c\Pi_2^{(k)}(t)\leq C\Gamma_2^{(k)}(t),\quad k=0,1,
\end{equation}
where
\begin{multline*}
\Xi_2^{(k)}(t):=\int_0^1\Bigg\{\big(\pd{t}{k}\prmd_t\big)^2+\bigg(\wnd-\frac{\dl^2}{\rmd^4}\bigg)\big(\pd{t}{k}\prmd_x\big)^2+\frac{1}{2}\big(\pc\pd{t}{k}\prmd_{xx}\big)^2\\
-\frac{3\rmd^3}{2}\pd{t}{k}\pwnd\pd{t}{k}\prmd_t-k\bigg[\frac{9\rmd^5\pc}{8}\pwnd_t\big(\pc\prmd_{txx}\big)+\frac{3\rmd^4\pc^2}{8}\big(\pc\prmd_{txx}\big)^2\bigg]\Bigg\}dx,
\end{multline*}
\begin{equation*}
\Pi_2^{(k)}(t):=\|(\pd{t}{k}\prmd_t,\pc\pd{t}{k}\prmd_{tx})(t)\|^2,\quad \Gamma_2^{(k)}(t):=\big(\mu+\pcN+\delta+\pc^{1/2}\big)\|\pd{t}{k}\prmd_x(t)\|^2+\Upsilon^{(k)}(t),
\end{equation*}
\begin{equation*}
\Upsilon^{(0)}(t):=C_\mu\|(\prmd,\pdl,\pwnd,\pwnd_x)(t)\|^2+\big(\pcN+\delta+\pc^{1/2}\big)\|(\prmd_{tt},\prmd_{tx},\pc\prmd_{xx},\pc\prmd_{txx},\pwnd_t)(t)\|^2,
\end{equation*}
\begin{align}
\Upsilon^{(1)}(t):=C_\mu&\|(\pwnd_t,\pwnd_{tx})(t)\|^2+\|(\prmd,\pdl,\pwnd,\pwnd_x,\prmd_t,\prmd_x)(t)\|^2\notag\\
&+\big(\pcN+\delta+\pc^{1/2}\big)\|(\prmd_{tt},\pc\prmd_{xx},\pc\prmd_{txx})(t)\|^2,\label{115.2a}
\end{align}
and the constants $c$ and $C$ are independent of $\delta$, $\pc$ and $T$. Here $\mu$ is an arbitrary positive constant to be determined and $C_\mu$ is a generic constant which only depends on $\mu$.
\end{lemma}
\begin{proof}
Multiplying the equation \eqref{12.3} by $\pd{t}{k}\prmd_t$ and integrating the resultant equality by parts over the domain $\Omega$ together with the homogeneous boundary conditions \eqref{pbc}, we get
\begin{multline}\label{60.1}
\frac{d}{dt}\int_0^1\bigg[\big(\pd{t}{k}\prmd_t\big)^2+\bigg(\wnd-\frac{\dl^2}{\rmd^4}\bigg)\big(\pd{t}{k}\prmd_x\big)^2+\frac{1}{2}\big(\pc\pd{t}{k}\prmd_{xx}\big)^2\bigg]dx\\
+2\|\pd{t}{k}\prmd_t(t)\|^2-\int_0^1\rmd\pd{t}{k}\pwnd_{xx}\pd{t}{k}\prmd_tdx=\mathcal{I}_2^{(k)}(t),\quad k=0,1,
\end{multline}
where the integral term $\mathcal{I}_2^{(k)}(t)$ is given by
\begin{align}
\mathcal{I}_2^{(k)}(t):=\int_0^1\bigg\{&\bigg[-2\bigg(\wnd-\frac{\dl^2}{\rmd^4}\bigg)_x\pd{t}{k}\prmd_x\pd{t}{k}\prmd_t+\bigg(\wnd-\frac{\dl^2}{\rmd^4}\bigg)_t\big(\pd{t}{k}\prmd_x\big)^2\bigg]\notag\\
&+2\rsmd_{xx}\pd{t}{k}\pwnd\pd{t}{k}\prmd_t+\frac{2\dl}{\rmd^3}\pd{t}{k}\pdl_{xx}\pd{t}{k}\prmd_t+\pc^2\frac{(k+1)\prmd_{xx}+2\rsmd_{xx}}{\rmd}\pd{t}{k}\prmd_{xx}\pd{t}{k}\prmd_t\notag\\
&+\Big[\pd{t}{k}P(t,x)+O_k(t,x)\Big]\pd{t}{k}\prmd_t\bigg\}dx
\end{align}
and can be estimated by the standard method as follows
\begin{align}
\mathcal{I}_2^{(k)}(t)\leq&C(\pcN+\delta)\big(\|(\pd{t}{k}\prmd_t,\pd{t}{k}\prmd_x)(t)\|^2+\|\pwnd_t(t)\|_k^2\big)\notag\\
&+\mu\|\pd{t}{k}\prmd_t(t)\|^2+C_\mu\|\pd{t}{k}\pwnd(t)\|_1^2\notag\\
&+C(\pcN+\delta)\|(\pd{t}{k}\prmd_t,k\prmd_{tx})(t)\|^2\notag\\
&+C\pc^{1/2}\|(\pd{t}{k}\prmd_t,\pc\pd{t}{k}\prmd_{xx})(t)\|^2\notag\\
&+\mu\|\pd{t}{k}\prmd_t(t)\|^2+C_\mu\|(\pd{t}{k}\prmd,\pd{t}{k}\pwnd,\pd{t}{k}\pwnd_x)(t)\|^2\notag\\
&\qquad+C_\mu\big(\pcN+\delta+\pc^{1/2}\big)\|(\pd{t}{k}\pdl,\pd{t}{k}\prmd_t,\pd{t}{k}\prmd_x)(t)\|^2\notag\\
\leq&2\mu\|\pd{t}{k}\prmd_t(t)\|^2+C_\mu\|(\pd{t}{k}\prmd,\pd{t}{k}\pwnd,\pd{t}{k}\pwnd_x)(t)\|^2\notag\\
&+C_\mu\big(\pcN+\delta+\pc^{1/2}\big)\Big(\|(\pd{t}{k}\pdl,\pd{t}{k}\prmd_t,\pd{t}{k}\prmd_x,\pc\pd{t}{k}\prmd_{xx})(t)\|^2+\|\pwnd_t(t)\|_k^2\Big),\label{107.2r}
\end{align}
with the aid of the estimates \eqref{17.1}, \eqref{16.3+16.5}, \eqref{38.2c+38.2d} and \eqref{26.1+26.4+27.2+27.3+29.1}, and the H\"older, Young and Sobolev inequalities.
Now, we have to deal with the last integral term on the left-side of the equality \eqref{60.1}. It is the most difficult part in the proof due to the dispersive velocity term in the energy equation and the Bohm potential term in the momentum equation. Precisely, solving the equation \eqref{15.1} with respect to $\pd{t}{k}\pwnd_{xx}$ and substituting the result in the last integral term on the left-side of \eqref{60.1} give
\begin{align}
-\int_0^1\rmd\pd{t}{k}\pwnd_{xx}\pd{t}{k}\prmd_tdx=&\int_0^1\rmd\frac{3}{2}\bigg[-\rmd^2\pd{t}{k}\pwnd_t-\frac{2}{3}\swnd\pd{t}{k}\pdl_x+\frac{4\sdl\swnd}{3\rsmd}\pd{t}{k}\prmd_x\notag\\
&\qquad\qquad\qquad+\pd{x}{}\mathcal{V}_k(t,x)+\pd{t}{k}H(t,x)+L_k(t,x)\bigg]\pd{t}{k}\prmd_tdx\notag\\
=&-\int_0^1\frac{3}{2}\rmd^3\pd{t}{k}\pwnd_t\pd{t}{k}\prmd_tdx-\int_0^1\rmd\swnd\pd{t}{k}\pdl_x\pd{t}{k}\prmd_tdx\notag\\
&+\int_0^1\frac{2\rmd\sdl\swnd}{\rsmd}\pd{t}{k}\prmd_x\pd{t}{k}\prmd_tdx+\int_0^1\frac{3}{2}\rmd\pd{x}{}\mathcal{V}_k(t,x)\pd{t}{k}\prmd_tdx\notag\\
&+\int_0^1\frac{3}{2}\rmd\Big[\pd{t}{k}H(t,x)+L_k(t,x)\Big]\pd{t}{k}\prmd_tdx\notag\\
=&\mathfrak{T}^{(k)}_1(t)+\mathfrak{T}^{(k)}_2(t)+\mathfrak{T}^{(k)}_3(t)+\mathfrak{T}^{(k)}_4(t)+\mathfrak{T}^{(k)}_5(t). \label{64.2}
\end{align}
The integrals $\mathfrak{T}^{(k)}_2(t)$, $\mathfrak{T}^{(k)}_3(t)$ and $\mathfrak{T}^{(k)}_5(t)$ are relatively easier to be estimated than to deal with the integrals $\mathfrak{T}^{(k)}_1(t)$ and $\mathfrak{T}^{(k)}_4(t)$. Before treating them one by one, we first derive the following equality which follows from the equation \eqref{10.2a},
\begin{equation}\label{65.3}
\pd{t}{k}\prmd_{tt}=-\frac{1}{2\rmd}\pd{t}{k}\pdl_{tx}+\mathcal{B}_k(t,x),\quad k=0,1,
\end{equation}
where
\begin{equation*}\label{65.2}
\mathcal{B}_0(t,x):=\frac{1}{2\rmd^2}\prmd_t\pdl_x,\quad \mathcal{B}_1(t,x):=\frac{1}{\rmd^2}\prmd_t\pdl_{tx}-\frac{1}{\rmd^3}\prmd_t^2\pdl_x+\frac{1}{2\rmd^2}\prmd_{tt}\pdl_x,
\end{equation*}
satisfying the estimate
\begin{equation}\label{66.2}
\|\mathcal{B}_k(t)\|\leq C\pcN\|(\pd{t}{k}\prmd_t,\prmd_t)(t)\|.
\end{equation}
Now, we begin to estimate $\mathfrak{T}^{(k)}_l(t)$, $l=1,\cdots,5$. Firstly, using the estimates \eqref{17.1}, \eqref{16.3+16.5}, \eqref{29.2a} and \eqref{30.1a+30.2}, the equation \eqref{10.2a} and the Young inequality, via the standard computations, we have
\begin{align}
\mathfrak{T}^{(k)}_2(t)+\mathfrak{T}^{(k)}_3(t)+\mathfrak{T}^{(k)}_5(t)\geq&\int_0^1\rmd\swnd\big(2\rmd\pd{t}{k}\prmd_t+k2\prmd_t^2\big)\pd{t}{k}\prmd_tdx-C\delta\|(\pd{t}{k}\prmd_t,\pd{t}{k}\prmd_x)(t)\|^2\notag\\
&-\mu\|\pd{t}{k}\prmd_t(t)\|^2-C_\mu\|(\pd{t}{k}\prmd,\pd{t}{k}\pdl,\pd{t}{k}\pwnd,\pd{t}{k}\pwnd_x)(t)\|^2\notag\\
&-kC_\mu(\pcN+\delta)\|(\prmd_{tt},\prmd_{tx})(t)\|^2\notag\\
\geq&c\|\pd{t}{k}\prmd_t(t)\|^2-C\|(\pd{t}{k}\prmd,\pd{t}{k}\pdl,\pd{t}{k}\pwnd,\pd{t}{k}\pwnd_x)(t)\|^2\notag\\
&-C(\pcN+\delta)\|\pd{t}{k}\prmd_x(t)\|^2.\label{92.2+93.1+101.1}
\end{align}
In addition, we continue to estimate $\mathfrak{T}^{(k)}_1(t)$ by using \eqref{65.3}, \eqref{66.2} and the integration by parts.
\begin{align}
\mathfrak{T}^{(k)}_1(t)=&-\frac{d}{dt}\int_0^1\frac{3\rmd^3}{2}\pd{t}{k}\pwnd\pd{t}{k}\prmd_tdx+\int_0^1\frac{9\rmd^2\prmd_t}{2}\pd{t}{k}\pwnd\pd{t}{k}\prmd_tdx+\int_0^1\frac{3\rmd^3}{2}\pd{t}{k}\pwnd\pd{t}{k}\prmd_{tt}dx\notag\\
\geq&-\frac{d}{dt}\int_0^1\frac{3\rmd^3}{2}\pd{t}{k}\pwnd\pd{t}{k}\prmd_tdx-C\pcN\|(\pd{t}{k}\prmd_t,\pd{t}{k}\pwnd)(t)\|^2\notag\\
&+\int_0^1\frac{3\rmd^3}{2}\pd{t}{k}\pwnd\bigg[-\frac{1}{2\rmd}\pd{t}{k}\pdl_{tx}+\mathcal{B}_k(t,x)\bigg]dx\notag\\
\geq&-\frac{d}{dt}\int_0^1\frac{3\rmd^3}{2}\pd{t}{k}\pwnd\pd{t}{k}\prmd_tdx-C\pcN\|(\pd{t}{k}\prmd_t,\prmd_t,\pd{t}{k}\pwnd)(t)\|^2\notag\\
&+\int_0^1\frac{3\rmd^2}{4}\pd{t}{k}\pwnd_x\pd{t}{k}\pdl_tdx+\int_0^1\frac{3\rmd\rmd_x}{2}\pd{t}{k}\pwnd\pd{t}{k}\pdl_tdx,\quad k=0,1.\label{ge67.1}
\end{align}
Moreover, we have to separately deal with the last two terms on the right-side of \eqref{ge67.1} for $k=0$ and $k=1$. In fact,
\begin{subequations}\label{ge80.1+90.3}
\begin{equation}\label{ge80.1}
\int_0^1\frac{3\rmd^2}{4}\pwnd_x\pdl_tdx\geq-\mu\|\pdl_t(t)\|^2-C_\mu\|\pwnd_x(t)\|^2,\qquad\text{for}\ k=0,
\end{equation}
\begin{align}
\int_0^1\frac{3\rmd^2}{4}\pwnd_{tx}\pd{t}{}\pdl_tdx=&\int_0^1\frac{3\rmd^2}{4}\pwnd_{tx}\Big[\pc^2\rmd\prmd_{txxx}+Y_1(t,x)\Big]dx\notag\\
=&-\int_0^1\frac{3\rmd^3\pc^2}{4}\pwnd_{txx}\prmd_{txx}dx-\int_0^1\frac{9\rmd^2\rmd_x\pc}{4}\pwnd_{tx}(\pc\prmd_{txx})dx+\int_0^1\frac{3\rmd^2}{4}\pwnd_{tx}Y_1dx\notag\\
\geq&\int_0^1\frac{9\rmd^3\pc^2}{8}\bigg[-\rmd^2\pwnd_{tt}-\frac{2}{3}\swnd\pdl_{tx}+\frac{4\sdl\swnd}{3\rsmd}\prmd_{tx}+\pd{x}{}\mathcal{V}_1+\pd{t}{}H+L_1\bigg]\prmd_{txx}dx\notag\\
&-C\pc\|(\pc\prmd_{txx},\pwnd_{tx})(t)\|^2-\mu\|Y_1(t)\|^2-C_\mu\|\pwnd_{tx}(t)\|^2\notag\\
\geq&-\int_0^1\frac{9\rmd^5\pc^2}{8}\pwnd_{tt}\prmd_{txx}dx+\int_0^1\frac{9\rmd^3\pc^2}{8}\pd{x}{}\mathcal{V}_1\prmd_{txx}dx\notag\\
&-C\pc\|(\pdl_t,\prmd_t,\prmd_{tt},\prmd_{tx},\pc\prmd_{txx})(t)\|^2-\mu\|Y_1(t)\|^2-C_\mu\|\pwnd_{tx}(t)\|^2\notag\\
\geq&-\frac{d}{dt}\int_0^1\frac{9\rmd^5\pc}{8}\pwnd_t(\pc\prmd_{txx})dx-\frac{d}{dt}\int_0^1\frac{3\rmd^4\pc^2}{8}(\pc\prmd_{txx})^2dx\notag\\
&-C\pc\|(\pdl_t,\prmd_t,\prmd_{tt},\prmd_{tx},\pc\prmd_{ttx},\pc\prmd_{txx})(t)\|^2\notag\\
&-\mu\|Y_1(t)\|^2-C_\mu\|\pwnd_{tx}(t)\|^2,\quad\qquad\qquad\qquad\text{for}\ k=1,\label{90.3}
\end{align}
\end{subequations}
where we have used the equality \eqref{P79-72.2}, the equation \eqref{15.1} with $k=1$, the estimates \eqref{38.2c+38.2d} and \eqref{30.1a+30.2}, the Cauchy-Schwarz inequality, and the following computations,
\begin{align}
-\int_0^1\frac{9\rmd^5\pc^2}{8}\pwnd_{tt}\prmd_{txx}dx=&-\frac{d}{dt}\int_0^1\frac{9\rmd^5\pc}{8}\pwnd_t(\pc\prmd_{txx})dx+\int_0^1\frac{45\rmd^4\prmd_t\pc}{8}\pwnd_t(\pc\prmd_{txx})dx\notag\\
&+\int_0^1\frac{9\rmd^5\pc^2}{8}\pwnd_t\prmd_{ttxx}dx\notag\\
=&-\frac{d}{dt}\int_0^1\frac{9\rmd^5\pc}{8}\pwnd_t(\pc\prmd_{txx})dx+\int_0^1\frac{45\rmd^4\prmd_t\pc}{8}\pwnd_t(\pc\prmd_{txx})dx\notag\\
&-\int_0^1\frac{45\rmd^4\rmd_x\pc}{8}\pwnd_t(\pc\prmd_{ttx})dx-\int_0^1\frac{9\rmd^5\pc}{8}\pwnd_{tx}(\pc\prmd_{ttx})dx\notag\\
\geq&-\frac{d}{dt}\int_0^1\frac{9\rmd^5\pc}{8}\pwnd_t(\pc\prmd_{txx})dx-C\pc\|(\pc\prmd_{ttx},\pc\prmd_{txx},\pwnd_{tx})(t)\|^2,\label{a1-82.1}
\end{align}
and
\begin{align}
&\int_0^1\frac{9\rmd^3\pc^2}{8}\pd{x}{}\mathcal{V}_1\prmd_{txx}dx\notag\\
=&\int_0^1\frac{9\rmd^3\pc^2}{8}\bigg(\frac{\pc^2}{3}\pdl_{txxx}-\frac{2\pc^2\dl}{3\rmd}\prmd_{txxx}+\mathcal{K}_2\bigg)\prmd_{txx}dx\notag\\
\geq&-\int_0^1\frac{3\rmd^3\pc^2}{8}\Big(2\rmd\prmd_{ttxx}+4\rmd_x\prmd_{ttx}+4\prmd_{tx}^2+4\prmd_t\prmd_{txx}+2\rmd_{xx}\prmd_{tt}\Big)\prmd_{txx}dx\notag\\
&+\int_0^1\bigg(\frac{3\pc^4}{8}\rmd^2\dl\bigg)_x\prmd_{txx}^2dx-C\pc\|(\mathcal{K}_2,\pc\prmd_{txx})(t)\|^2\notag\\
\geq&-\frac{d}{dt}\int_0^1\frac{3\rmd^4\pc^2}{8}(\pc\prmd_{txx})^2dx-C\pc\|(\pdl_t,\prmd_t,\prmd_{tt},\prmd_{tx},\pc\prmd_{ttx},\pc\prmd_{txx})(t)\|^2,\label{a4-88.2}
\end{align}
with the aid of the equality \eqref{37.2}, the equation \eqref{10.2a}, the integration by parts, the estimates \eqref{17.1}, \eqref{16.3+16.5} and \eqref{37.1b}, and the Sobolev, H\"older and Cauchy-Schwarz inequalities. Similarly, we continue to estimate the last term on the right-side of \eqref{ge67.1} as follows
\begin{subequations}\label{91.1+91.2}
\begin{equation}\label{91.1}
\int_0^1\frac{3\rmd\rmd_x}{2}\pwnd\pdl_tdx\geq-\mu\|\pdl_t(t)\|^2-C_\mu\|\pwnd(t)\|^2, \quad\text{for}\ k=0,
\end{equation}
\begin{align}
\int_0^1\frac{3\rmd\rmd_x}{2}\pwnd_t\pd{t}{}\pdl_tdx=&\int_0^1\frac{3\rmd\rmd_x}{2}\pwnd_t\Big(\pc^2\rmd\prmd_{txxx}+Y_1\Big)dx\notag\\
\geq&-\int_0^1\bigg(\frac{3\rmd^2\rmd_x\pc^2}{2}\pwnd_t\bigg)_x\prmd_{txx}dx-\mu\|Y_1(t)\|^2-C_\mu\|\pwnd_t(t)\|^2\notag\\
\geq&-C\pc^{1/2}\|(\pc\prmd_{txx},\pwnd_{tx})(t)\|^2-\mu\|Y_1(t)\|^2-C_\mu\|\pwnd_t(t)\|^2, \quad\text{for}\ k=1.\label{91.2}
\end{align}
\end{subequations}
Next, we estimate the integral $\mathfrak{T}^{(k)}_4(t)$ by using the integration by parts and the equality \eqref{32.1+33.1},
\begin{align}
\mathfrak{T}^{(k)}_4(t)=&-\int_0^1\bigg(\frac{3}{2}\rmd\pd{t}{k}\prmd_t\bigg)_x\mathcal{V}_kdx\notag\\
=&-\int_0^1\frac{3}{2}\rmd\pd{t}{k}\prmd_{tx}\mathcal{V}_kdx-\int_0^1\frac{3}{2}\rmd_x\pd{t}{k}\prmd_t\mathcal{V}_kdx\notag\\
\geq&-\int_0^1\frac{3}{2}\rmd\pd{t}{k}\prmd_{tx}\bigg(\frac{\pc^2}{3}\pd{t}{k}\pdl_{xx}-\frac{2\pc^2\sdl}{3\rsmd}\pd{t}{k}\prmd_{xx}+\pd{t}{k}\mathcal{K}\bigg)dx\notag\\
&-C\pc\|(\pd{t}{k}\pdl,\pd{t}{k}\prmd,\pd{t}{k}\prmd_t,\pd{t}{k}\prmd_x,\pc\pd{t}{k}\prmd_{tx},\pc\pd{t}{k}\prmd_{xx})(t)\|^2\notag\\
\geq&-\int_0^1\frac{\rmd\pc^2}{2}\pd{t}{k}\prmd_{tx}\pd{t}{k}\pdl_{xx}dx\notag\\
&-C\big(\pcN+\delta+\pc^{1/2}\big)\|(\pd{t}{k}\pdl,\pd{t}{k}\prmd,\pd{t}{k}\prmd_t,\pd{t}{k}\prmd_x,\pc\pd{t}{k}\prmd_{tx},\pc\pd{t}{k}\prmd_{xx})(t)\|^2\notag\\
=&\int_0^1\frac{\rmd\pc^2}{2}\pd{t}{k}\prmd_{tx}\Big(2\rmd\pd{t}{k}\prmd_{tx}+2\rmd_x\pd{t}{k}\prmd_t+k4\prmd_t\prmd_{tx}\Big)dx\notag\\
&-C\big(\pcN+\delta+\pc^{1/2}\big)\|(\pd{t}{k}\pdl,\pd{t}{k}\prmd,\pd{t}{k}\prmd_t,\pd{t}{k}\prmd_x,\pc\pd{t}{k}\prmd_{tx},\pc\pd{t}{k}\prmd_{xx})(t)\|^2\notag\\
\geq&\int_0^1\rmd^2\big(\pc\pd{t}{k}\prmd_{tx}\big)^2dx\notag\\
&-C\big(\pcN+\delta+\pc^{1/2}\big)\|(\pd{t}{k}\pdl,\pd{t}{k}\prmd,\pd{t}{k}\prmd_t,\pd{t}{k}\prmd_x,\pc\pd{t}{k}\prmd_{tx},\pc\pd{t}{k}\prmd_{xx})(t)\|^2\notag\\
\geq&c\|\pc\pd{t}{k}\prmd_{tx}(t)\|^2-C\big(\pcN+\delta+\pc^{1/2}\big)\|(\pd{t}{k}\pdl,\pd{t}{k}\prmd,\pd{t}{k}\prmd_t,\pd{t}{k}\prmd_x,\pc\pd{t}{k}\prmd_{xx})(t)\|^2,\label{100.2}
\end{align}
where we have used the equation \eqref{10.2a} and the estimates \eqref{32.2b}$\sim$\eqref{33.2b}.
Finally, substituting \eqref{107.2r}, \eqref{64.2}, \eqref{92.2+93.1+101.1}, \eqref{ge67.1}, \eqref{ge80.1+90.3}, \eqref{91.1+91.2} and \eqref{100.2} into \eqref{60.1}, applying the estiamtes \eqref{P79-70.3a} and \eqref{P79-78.1a} to the resultant inequality, making $\mu$ and $\pcN+\delta+\pc$ small enough and rewriting the result as a unified form in $k=0,1$, we obtain the desired estimate \eqref{115.2}.
\end{proof}
In order to close the uniform a priori estimate, we continue to derive the higher order estimates of the perturbed temperature $\pwnd$. The dispersive velocity term in the energy equation makes the corresponding computations more complex.
\begin{lemma}
Suppose the same assumptions as in Proposition \ref{prop1} hold. Then there exist positive constants $\delta_0$, $c$ and $C$ such that if $\pcN+\delta+\pc\leq\delta_0$, it holds that for $t\in[0,T]$,
\begin{equation}\label{116.2}
\frac{d}{dt}\Xi_3(t)+c\Pi_3(t)\leq C\Gamma_3(t),
\end{equation}
where
\begin{gather}
\Xi_3(t):=\int_0^1\bigg(\frac{1}{3}\pwnd_x^2+\frac{2\swnd}{3}\pdl_x\pwnd\bigg)dx,\quad\Pi_3(t):=\|\pwnd_t(t)\|^2,\notag\\
\Gamma_3(t):=\mu\|\prmd_x(t)\|^2+C_\mu A_{-1}^2(t)+\big(\pcN+\delta+\pc\big)\Big(A_1^2(t)+\|\pwnd_{tx}(t)\|^2\Big),\label{116.2a}
\end{gather}
and
\begin{equation}\label{117.2}
\frac{d}{dt}\Xi_4(t)+c\Pi_4(t)\leq C\Gamma_4(t),
\end{equation}
where
\begin{gather}
\Xi_4(t):=\int_0^1\frac{\rmd^2}{2}\pwnd_t^2dx,\quad\Pi_4(t):=\|\pwnd_{tx}(t)\|^2,\notag\\
\Gamma_4(t):=\|(\prmd,\pdl,\pwnd,\pwnd_x,\prmd_x,\pwnd_t)(t)\|^2+\big(\pcN+\delta+\pc\big)\Big(A_1^2(t)+\|\pc\prmd_{ttx}(t)\|^2\Big),\label{117.2a}
\end{gather}
and the constants $c$ and $C$ are independent of $\delta$, $\pc$ and $T$. Here $\mu$ is an arbitrary positive constant to be determined and $C_\mu$ is a generic constant which only depends on $\mu$.
\end{lemma}
\begin{proof}
Multiplying the equation \eqref{15.1} with $k=0$ by $\pwnd_t$ and integrating the resultant equality by parts over $\Omega$ together with the boundary conditions \eqref{pbc}, we get
\begin{equation}\label{53.1}
\frac{d}{dt}\Xi_3(t)+\int_0^1\rmd^2\pwnd_t^2dx=\mathcal{I}_3(t),
\end{equation}
where the integral term $\mathcal{I}_3(t)$ is defined by
\begin{equation}
\mathcal{I}_3(t):=\int_0^1\bigg[-\bigg(\frac{2\swnd_x}{3}\pdl_t\pwnd+\frac{2\swnd}{3}\pdl_t\pwnd_x\bigg)+\frac{4\sdl\swnd}{3\rsmd}\prmd_x\pwnd_t+\pd{x}{}\mathcal{V}_0(t,x)\pwnd_t+H(t,x)\pwnd_t\bigg]dx,
\end{equation}
and can be estimated as
\begin{align}
\mathcal{I}_3(t)\leq&\int_0^1\pd{x}{}\mathcal{V}_0(t,x)\pwnd_tdx+(\mu+C\delta)\|\pdl_t(t)\|^2+C_\mu\|(\pwnd,\pwnd_x)(t)\|^2\notag\\
&+C\delta\|(\prmd_x,\pwnd_t)(t)\|^2+\mu\|\pwnd_t(t)\|^2+C_\mu\|H(t)\|^2\notag\\
\leq&\int_0^1\pd{x}{}\mathcal{V}_0(t,x)\pwnd_tdx+(\mu+\delta+\pc)\|\pwnd_t(t)\|^2\notag\\
&+\mu\|\prmd_x(t)\|^2+C_\mu A_{-1}^2(t)+C\big(\pcN+\delta+\pc\big)A_1^2(t),\label{55.2}
\end{align}
with the aid of the estimates \eqref{17.1}, \eqref{29.2a} and \eqref{P79-70.3a}, the H\"older and Young inequalities. Furthermore, applying the integration by parts and the equality \eqref{32.1+33.1} with $k=0$ to the first integral term on the right-side of \eqref{55.2}, we have
\begin{align}
\int_0^1\pd{x}{}\mathcal{V}_0(t,x)\pwnd_tdx=&-\int_0^1\mathcal{V}_0(t,x)\pwnd_{tx}dx\notag\\
=&-\int_0^1\bigg(\frac{\pc^2}{3}\pdl_{xx}-\frac{2\pc^2\sdl}{3\rsmd}\prmd_{xx}+\mathcal{K}(t,x)\bigg)\pwnd_{tx}dx\notag\\
\leq&C\pc\|(\pdl_{xx},\pc\prmd_{xx},\pwnd_{tx})(t)\|^2+\|\mathcal{K}(t)\|\|\pwnd_{tx}(t)\|\notag\\
\leq&C\pc\|(\prmd,\pdl,\prmd_t,\prmd_x,\prmd_{tx},\pc\prmd_{xx},\pwnd_{tx})(t)\|^2\notag\\
\leq&C\pc\big(A_1^2(t)+\|\pwnd_{tx}(t)\|^2\big),\label{54.2}
\end{align}
with the aid of the estimates \eqref{32.4+44.2} and \eqref{32.2b}, the H\"older and Cauchy-Schwarz inequalities. Substituting \eqref{54.2} into \eqref{55.2}, we obtain
\begin{equation}\label{55.3a}
\mathcal{I}_3(t)\leq(\mu+\delta+\pc)\|\pwnd_t(t)\|^2+C\Gamma_3(t).
\end{equation}
On the other hand, it is easy to see that
\begin{equation}\label{55.3b}
\int_0^1\rmd^2\pwnd_t^2dx\geq c\Pi_3(t).
\end{equation}
Substituting \eqref{55.3a} and \eqref{55.3b} into \eqref{53.1}, and letting $\mu$ and $\pcN+\delta+\pc$ sufficiently small, we have proved the desired estimate \eqref{116.2}.
Next, multiplying the equation \eqref{15.1} with $k=1$ by $\pwnd_t$ and integrating the resultant equality by parts over $\Omega$ together with the boundary conditions \eqref{pbc}, we get
\begin{equation}\label{ge56.1}
\frac{d}{dt}\Xi_4(t)+\frac{2}{3}\Pi_4(t)=\mathcal{I}_4(t),
\end{equation}
where the integral term $\mathcal{I}_4(t)$ is given by
\begin{align}
\mathcal{I}_4(t):=\int_0^1\bigg[&\rmd\prmd_t\pwnd_t^2+\bigg(\frac{2\swnd_x}{3}\pdl_t\pwnd_t+\frac{2\swnd}{3}\pdl_t\pwnd_{tx}\bigg)\notag\\
&+\frac{4\sdl\swnd}{3\rsmd}\prmd_{tx}\pwnd_t+\pd{x}{}\mathcal{V}_1(t,x)\pwnd_t+\big(\pd{t}{}H+L_1\big)(t,x)\pwnd_t\bigg]dx,\label{56.0}
\end{align}
and can be estimated as
\begin{align}
\mathcal{I}_4(t)\leq&\int_0^1\pd{x}{}\mathcal{V}_1(t,x)\pwnd_tdx+C\big(\pcN+\delta\big)\|(\prmd_{tx},\pwnd_t)(t)\|^2\notag\\
&+\mu\|\pwnd_{tx}(t)\|^2+C_\mu\|\pdl_t(t)\|^2+\big(\|\pd{t}{}H(t)\|+\|L_1(t)\|\big)\|\pwnd_t(t)\|\notag\\
\leq&\int_0^1\pd{x}{}\mathcal{V}_1(t,x)\pwnd_tdx+\big[\mu+C(\pcN+\delta)\big]\|\pwnd_{tx}(t)\|^2\notag\\
&+C_\mu\|(\pdl_t,\pwnd_t)(t)\|^2+C(\pcN+\delta)\|(\prmd_t,\prmd_{tt},\prmd_{tx})(t)\|^2,\label{59.2}
\end{align}
with the aid of the estimates \eqref{17.1}, \eqref{16.3+16.5} and \eqref{30.1a+30.2}, the H\"older and Young inequalities. Furthermore, applying the integration by parts and the equality \eqref{32.1+33.1} with $k=1$ to the first integral term on the right-side of \eqref{59.2}, we have
\begin{align}
\int_0^1\pd{x}{}\mathcal{V}_1(t,x)\pwnd_tdx=&-\int_0^1\mathcal{V}_1(t,x)\pwnd_{tx}dx\notag\\
=&-\int_0^1\bigg(\frac{\pc^2}{3}\pdl_{txx}-\frac{2\pc^2\sdl}{3\rsmd}\prmd_{txx}+\pd{t}{}\mathcal{K}(t,x)\bigg)\pwnd_{tx}dx\notag\\
\leq&C\pc\|(\pc\pdl_{txx},\pc\prmd_{txx},\pwnd_{tx})(t)\|^2+\|\pd{t}{}\mathcal{K}(t)\|\|\pwnd_{tx}(t)\|\notag\\
\leq&C\pc\big(\|\pwnd_{tx}(t)\|^2+\|\pdl_t(t)\|^2+A_1^2(t)+\|\pc\prmd_{ttx}(t)\|^2\big),\label{58.1}
\end{align}
with the aid of the estimates \eqref{38.2c+38.2d} and \eqref{33.2b}, the H\"older and Cauchy-Schwarz inequalities. Substituting \eqref{58.1} into \eqref{59.2} together with the estimate \eqref{P79-70.3a}, we obtain
\begin{align}\label{I4}
\mathcal{I}_4(t)\leq&\big[\mu+C(\pcN+\delta+\pc)\big]\|\pwnd_{tx}(t)\|^2+C(\pcN+\delta+\pc)\big(A_1^2(t)+\|\pc\prmd_{ttx}(t)\|^2\big)\notag\\
&+C_\mu\|(\pdl_t,\pwnd_t)(t)\|^2\notag\\
\leq&\big[\mu+C(\pcN+\delta+\pc)\big]\|\pwnd_{tx}(t)\|^2+C_\mu\Gamma_4(t).
\end{align}
Substituting \eqref{I4} into \eqref{ge56.1}, and letting $\mu$ and $\pcN+\delta+\pc$ sufficiently small, we have shown the desired estimate \eqref{117.2}.
\end{proof}
\subsection{Decay estimate}\label{Subsect.3.5}
Based on the basic estimate \eqref{25.1} and the higher order estimates \eqref{115.1}, \eqref{115.2}, \eqref{116.2} and \eqref{117.2}, it is not difficult to find that we have captured the strong enough dissipation mechanism to show the decay estimate \eqref{127.2} in Proposition \ref{prop1}.
\begin{proof}[\textbf{Proof of Proposition \ref{prop1}.}]
From the procedure
\begin{equation*}\label{118.1}
\eqref{25.1}+\beta\Big[\alpha\eqref{115.1}+\eqref{115.2}\Big]\Big|_{k=0}+\beta\Big[\eqref{116.2}+\beta\eqref{117.2}\Big]+\beta^3\Big[\alpha\eqref{115.1}+\eqref{115.2}\Big]\Big|_{k=1},
\end{equation*}
where $\alpha$ is the positive constant in \eqref{25.2} and $\beta$ is another positive constant, both of them will be determined later, we have the energy inequality
\begin{equation}\label{118.2}
\frac{d}{dt}\mathbb{E}(t)+\mathbb{D}(t)\leq0,\quad\forall t\in[0,T],
\end{equation}
where the total energy $\mathbb{E}(t)$ is defined by
\begin{equation}\label{119.2}
\mathbb{E}(t):=\Xi(t)+\beta\Big[\alpha\Xi_1^{(0)}(t)+\Xi_2^{(0)}(t)\Big]+\beta\Big[\Xi_3(t)+\beta\Xi_4(t)\Big]+\beta^3\Big[\alpha\Xi_1^{(1)}(t)+\Xi_2^{(1)}(t)\Big],
\end{equation}
and the total dissipation rate $\mathbb{D}(t)$ is given by
\begin{align}
\mathbb{D}(t):=&\Big[c\Pi(t)-C\Gamma(t)\Big]+\beta\Big\{\alpha\Big[c\Pi_1^{(0)}(t)-C\Gamma_1^{(0)}(t)\Big]+\Big[c\Pi_2^{(0)}(t)-C\Gamma_2^{(0)}(t)\Big]\Big\}\notag\\
&+\beta\Big\{\Big[c\Pi_3(t)-C\Gamma_3(t)\Big]+\beta\Big[c\Pi_4(t)-C\Gamma_4(t)\Big]\Big\}\notag\\
&+\beta^3\Big\{\alpha\Big[c\Pi_1^{(1)}(t)-C\Gamma_1^{(1)}(t)\Big]+\Big[c\Pi_2^{(1)}(t)-C\Gamma_2^{(1)}(t)\Big]\Big\}.\label{122.1}
\end{align}
Substituting the specific definitions \eqref{25.2}$\sim$\eqref{R}, \eqref{115.1a}, \eqref{115.2a}, \eqref{116.2a} and \eqref{117.2a} into \eqref{119.2} and \eqref{122.1}, and then taking $\alpha$, $\mu$, $\beta$ and $\pcN+\delta+\pc$ sufficiently small in the following order $0<\pcN+\delta+\pc\ll\beta^3\ll\beta^2\ll\beta\ll\mu\ll\alpha\ll1$, via the elaborate calculations, we obtain the estimates
\begin{equation}\label{121.2}
c\big(A_1^2(t)+\|\pwnd_t(t)\|^2\big)\leq\mathbb{E}(t)\leq C\big(A_1^2(t)+\|\pwnd_t(t)\|^2\big),
\end{equation}
and
\begin{align}
\mathbb{D}(t)\geq&c\big(A_1^2(t)+\|\pwnd_t(t)\|^2+\|(\pwnd_{tx},\pc\prmd_{ttx})(t)\|^2\big)\notag\\
\geq&c\big(A_1^2(t)+\|\pwnd_t(t)\|^2\big),\label{126.2}\\
\mathbb{D}(t)\leq&C\big(A_1^2(t)+\|\pwnd_t(t)\|^2+\|(\pwnd_{tx},\pc\prmd_{ttx})(t)\|^2\big),\notag
\end{align}
where the positive constants $c$ and $C$ are independent of $\delta$, $\pc$ and $T$.
Applying \eqref{121.2} and \eqref{126.2} to the inequality \eqref{118.2}, we see that there exists a positive constant $\gamma$ which is independent of $\delta$, $\pc$ and $T$ such that the following inequality holds,
\begin{equation}\label{127.1}
\frac{d}{dt}\mathbb{E}(t)+2\gamma\mathbb{E}(t)\leq0,\quad\forall t\in[0,T].
\end{equation}
Finally, applying Gronwall inequality to \eqref{127.1} and using the elliptic estimate \eqref{17.3} and the equivalent relations \eqref{121.2} and \eqref{46.5}, we have the desired decay estimate \eqref{127.2}.
\end{proof}
Once Proposition \ref{prop1} is proved, Theorem \ref{thm2} immediately follows.
\begin{proof}[\textbf{Proof of Theorem \ref{thm2}.}]
The existence of the global-in-time solution to the initial-boundary value problem \eqref{1dfqhd}$\sim$\eqref{bc} follows from the continuation argument with Corollary \ref{cor1} and Proposition \ref{prop1}. The decay estimate \eqref{de} is derived by the transformations $\md=\rmd^2$, $\smd=\rsmd^2$ and the estimate \eqref{127.2}.
\end{proof}
\section{Semi-classical limit}\label{Sect.4}
In this section, we prove Theorem \ref{thm3} in Subsection \ref{Subsect.4.1} and Theorem \ref{thm4} in Subsection \ref{Subsect.4.2}, respectively.
\subsection{Stationary case}\label{Subsect.4.1}
In this subsection, we discuss the semi-classical limit of the stationary solutions based on the existence and uniqueness results in Lemma \ref{lem1} and Theorem \ref{thm1}. Since both of the quantum stationary density $\pcsmd{\pc}$ and the limit one $\pcsmd{0}$ are non-flat, it is convenient to introduce the logarithmic transformations $\pcszmd{\pc}:=\ln \pcsmd{\pc}$ and $\pcszmd{0}:=\ln \pcsmd{0}$ in the following discussion. We also introduce the error variables as follows
\begin{equation}\label{ss7.1}
\pcpszmd:=\pcszmd{\pc}-\pcszmd{0},\quad \pcpsdl:=\pcsdl{\pc}-\pcsdl{0},\quad\pcpswnd:=\pcswnd{\pc}-\pcswnd{0},\quad\pcpsdws:=\pcsdws{\pc}-\pcsdws{0}.
\end{equation}
These quantities $(\pcszmd{\pc},\pcsdl{\pc},\pcswnd{\pc},\pcsdws{\pc})$, $(\pcszmd{0},\pcsdl{0},\pcswnd{0},\pcsdws{0})$ and $(\pcpszmd,\pcpsdl,\pcpswnd,\pcpsdws)$ satisfy the following properties
\begin{equation}\label{psbc}
\pcpszmd\in H_0^1(\Omega)\cap C^2(\overline{\Omega}),\quad \pcpswnd\in H_0^1(\Omega)\cap H^3(\Omega),\quad \pcpsdws\in H_0^1(\Omega)\cap C^2(\overline{\Omega}),
\end{equation}
\begin{equation}\label{psbcxx}
\bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_x)^2}{2}\bigg](0)=\bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_x)^2}{2}\bigg](1)=0,
\end{equation}
the estimates
\begin{equation}\label{ss9.3}
\ln c\leq\pcszmd{0}(x)\leq\ln C,\quad 0<c\leq\pcswnd{0}(x)\leq C,\quad|\pcsdl{0}|+\|\pcswnd{0}-\theta_{L}\|_3\leq C\delta,\quad |(\pcsmd{0},\pcsdws{0})|_2\leq C,
\end{equation}
\begin{gather}
2\ln b\leq\pcszmd{\pc}(x)\leq2\ln B,\quad 0<\frac{\theta_L}{2}\leq\pcswnd{\pc}(x)\leq \frac{3\theta_l}{2},\quad |\pcsdl{\pc}|+\|\pcswnd{\pc}-\theta_{L}\|_3\leq C\delta,\notag\\
\|\pcszmd{\pc}\|_2+\|(\pc\pd{x}{3}\pcszmd{\pc},\pc^2\pd{x}{4}\pcszmd{\pc})\|+|\pcsdws{\pc}|_2\leq C,\qquad\forall\pc\in(0,\pc_1], \label{ss9.2}
\end{gather}
and the equations
\begin{equation}\label{ss8.1a}
S[e^{\pcszmd{0}},\pcsdl{0},\pcswnd{0}]\pcszmd{0}_x+\pcswnd{0}_x=\pcsdws{0}_x-\pcsdl{0}e^{-\pcszmd{0}},
\end{equation}
\begin{equation}\label{ss8.1b}
S[e^{\pcszmd{\pc}},\pcsdl{\pc},\pcswnd{\pc}]\pcszmd{\pc}_x+\pcswnd{\pc}_x-\frac{\pc^2}{2}\Bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_{x})^2}{2}\Bigg]_x=\pcsdws{\pc}_x-\pcsdl{\pc}e^{-\pcszmd{\pc}},
\end{equation}
\begin{equation}\label{ss8.1c}
S[e^{\pcszmd{\pc}},\pcsdl{\pc},\pcswnd{\pc}]\pcszmd{\pc}_x-S[e^{\pcszmd{0}},\pcsdl{0},\pcswnd{0}]\pcszmd{0}_x+\pcpswnd_{x}-\pcpsdws_x-\frac{\pc^2}{2}\Bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_{x})^2}{2}\Bigg]_x=-\Big(\pcsdl{\pc}e^{-\pcszmd{\pc}}-\pcsdl{0}e^{-\pcszmd{0}}\Big),
\end{equation}
\begin{multline}\label{ss8.1d}
\pcsdl{\pc}\pcswnd{\pc}_x-\pcsdl{0}\pcswnd{0}_x-\frac{2}{3}\Big(\pcsdl{\pc}\pcswnd{\pc}\pcszmd{\pc}_x-\pcsdl{0}\pcswnd{0}\pcszmd{0}_x\Big)-\frac{2}{3}\pcpswnd_{xx}+\frac{\pc^2}{3}\pcsdl{\pc}\Big(\pcszmd{\pc}_{xxx}-2\pcszmd{\pc}_{xx}\pcszmd{\pc}_{x}\Big)\\
=\frac{1}{3}\Big[(\pcsdl{\pc})^2e^{-\pcszmd{\pc}}-(\pcsdl{0})^2e^{-\pcszmd{0}}\Big]-\Big[e^{\pcszmd{\pc}}(\pcswnd{\pc}-\theta_L)-e^{\pcszmd{0}}(\pcswnd{0}-\theta_L)\Big],
\end{multline}
\begin{equation}\label{ss8.1e}
\pcpsdws_{xx}=e^{\pcszmd{\pc}}-e^{\pcszmd{0}},\qquad\forall\pc\in (0,\pc_1]
\end{equation}
due to the boundary conditions \eqref{sbc} and \eqref{0sbc}, the estimates \eqref{0se} and \eqref{145.1}, and the equations \eqref{1dsfqhd} and \eqref{1dsfhd}.
\begin{proof}[\textbf{Proof of Theorem \ref{thm3}.}]
Firstly, we prove the convergence rate \eqref{ss39.1a} in $\pc\in (0,\pc_1]$. Note that if $\delta$ is small enough, we have known that the quantum stationary current density $\pcsdl{\pc}=J[e^{\pcszmd{\pc}},\pcswnd{\pc}]$ is defined by the explicit formula \eqref{104.1}. Furthermore, the limit stationary current density $\pcsdl{0}$ can also be written by the same formula \eqref{104.1} as $\pcsdl{0}=J[e^{\pcszmd{0}},\pcswnd{0}]$. Therefore, the following estimate
\begin{align}
|\pcpsdl|=|\pcsdl{\pc}-\pcsdl{0}|&\leq C\Big(\delta\|\pcpszmd\|+\|\pcpswnd_x\|\Big)\notag\\
&\leq C\Big(\delta\|\pcpszmd_x\|+\|\pcpswnd_x\|\Big).\label{ss10.2}
\end{align}
follows from the straightforward computations with the formula \eqref{104.1}, the estimates \eqref{ss9.3} and \eqref{ss9.2}.
Multiplying the equation \eqref{ss8.1c} by $\pcpszmd_x$ and integrating the resultant equality over the domain $\Omega$, we obtain
\begin{multline}\label{ss10.5}
\int_0^1\Big(S[e^{\pcszmd{\pc}},\pcsdl{\pc},\pcswnd{\pc}]\pcszmd{\pc}_x-S[e^{\pcszmd{0}},\pcsdl{0},\pcswnd{0}]\pcszmd{0}_x\Big)\pcpszmd_xdx-\int_0^1\pcpsdws_x\pcpszmd_xdx\\
=\frac{\pc^2}{2}\int_0^1\Bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_{x})^2}{2}\Bigg]_x\pcpszmd_xdx-\int_0^1\pcpswnd_{x}\pcpszmd_xdx-\int_0^1\Big(\pcsdl{\pc}e^{-\pcszmd{\pc}}-\pcsdl{0}e^{-\pcszmd{0}}\Big)\pcpszmd_xdx.
\end{multline}
By virtue of integration by parts, the boundary conditions \eqref{psbc} and \eqref{psbcxx}, the equation \eqref{ss8.1e}, the Young and H\"older inequalities, the mean value theorem and the estimates \eqref{ss9.3}, \eqref{ss9.2} and \eqref{ss10.2}, the left-side of \eqref{ss10.5} can be estimated as follows
\allowdisplaybreaks
\begin{align}
\eqref{ss10.5}_l&=\int_0^1\Big(\pcsS{\pc}\pcszmd{\pc}_x-\pcsS{0}\pcszmd{0}_x\Big)\pcpszmd_xdx+\int_0^1\pcpsdws_{xx}\pcpszmd dx\notag\\
&=\int_0^1\Big[\Big(\pcsS{\pc}-\pcsS{0}\Big)\pcszmd{\pc}_x+\pcsS{0}\pcpszmd_x\Big]\pcpszmd_xdx+\int_0^1\underbrace{\Big(e^{\pcszmd{\pc}}-e^{\pcszmd{0}}\Big)(\pcszmd{\pc}-\pcszmd{0})}_{\geq0} dx\notag\\
&\geq\frac{\theta_L}{4}\|\pcpszmd_x\|^2+\int_0^1\Big(\pcsS{\pc}-\pcsS{0}\Big)\pcszmd{\pc}_x\pcpszmd_xdx\notag\\
&\geq\frac{\theta_L}{4}\|\pcpszmd_x\|^2-\mu\|\pcpszmd_x\|^2-C_\mu\|\pcpswnd\|^2-C\delta\|(\pcpszmd_x,\pcpswnd_x)\|^2\notag\\
&\geq\frac{\theta_L}{8}\|\pcpszmd_x\|^2-C\|\pcpswnd\|_1^2,\label{ss11.1}
\end{align}
where we have used the notation $\pcsS{\pc}:=S[e^{\pcszmd{\pc}},\pcsdl{\pc},\pcswnd{\pc}]$ for any $\pc\in[0,\pc_1]$ and the following estimate
\begin{equation}\label{ss12.1-2-3}
|(\pcsS{\pc}-\pcsS{0})(x)|\leq |\pcpswnd(x)|+C\delta|\pcpsdl|+C\delta^2|\pcpszmd(x)|,\quad\forall x\in\Omega.
\end{equation}
Similarly, we further estimate the right-side of \eqref{ss10.5} as follows
\begin{align}
\eqref{ss10.5}_r=&\frac{\pc^2}{2}\int_0^1\Bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_{x})^2}{2}\Bigg]_x\pcpszmd_xdx-\int_0^1\pcpswnd_{x}\pcpszmd_xdx-\int_0^1\Big(\pcsdl{\pc}e^{-\pcszmd{\pc}}-\pcsdl{0}e^{-\pcszmd{0}}\Big)\pcpszmd_xdx\notag\\
\leq&-\frac{\pc^2}{2}\int_0^1\Bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_{x})^2}{2}\Bigg]\pcpszmd_{xx}dx+\|\pcpszmd_x\|\|\pcpswnd_{x}\|+C|\pcpsdl|\|\pcpszmd_x\|+C\delta\|\pcpszmd\|\|\pcpszmd_x\|\notag\\
\leq&C\pc^2\int_0^1\Big(|\pcszmd{\pc}_{xx}|+|\pcszmd{\pc}_{x}|^2\Big)|\pcpszmd_{xx}|dx+\big(C\delta+\mu\big)\|\pcpszmd_x\|^2+C_\mu\|\pcpswnd_{x}\|^2\notag\\
\leq&C\pc^2\underbrace{\Big(\|\pcszmd{\pc}_{xx}\|+1\Big)\|\pcpszmd_{xx}\|}_{\leq C}+\big(C\delta+\mu\big)\|\pcpszmd_x\|^2+C_\mu\|\pcpswnd_{x}\|^2\notag\\
\leq&C\pc^2+\big(C\delta+\mu\big)\|\pcpszmd_x\|^2+C_\mu\|\pcpswnd_{x}\|^2.\label{ss12.4-5-ss13.1}
\end{align}
Substituting \eqref{ss11.1} and \eqref{ss12.4-5-ss13.1} into \eqref{ss10.5}, and letting $\delta$ and $\mu$ small enough, we have
\begin{equation}\label{ss13.3}
\|\pcpszmd\|_1^2\leq C\|\pcpswnd\|_1^2+C\pc^2,
\end{equation}
where we have used the Poincar\'e inequality $\|\pcpszmd\|\leq C\|\pcpszmd_x\|$.
Next, multiplying the equation \eqref{ss8.1d} by $\pcpswnd$ and integrating the resultant equality over the domain $\Omega$, we obtain
\begin{align}\label{ss14.1}
-\frac{2}{3}\int_0^1\pcpswnd_{xx}&\pcpswnd dx+\int_0^1\Big[e^{\pcszmd{\pc}}(\pcswnd{\pc}-\theta_L)-e^{\pcszmd{0}}(\pcswnd{0}-\theta_L)\Big]\pcpswnd dx\notag\\
&=-\frac{\pc^2}{3}\pcsdl{\pc}\int_0^1\Big(\pcszmd{\pc}_{xxx}-2\pcszmd{\pc}_{xx}\pcszmd{\pc}_{x}\Big)\pcpswnd dx+\frac{2}{3}\int_0^1\Big(\pcsdl{\pc}\pcswnd{\pc}\pcszmd{\pc}_x-\pcsdl{0}\pcswnd{0}\pcszmd{0}_x\Big)\pcpswnd dx\notag\\
&\qquad\qquad-\int_0^1\Big(\pcsdl{\pc}\pcswnd{\pc}_x-\pcsdl{0}\pcswnd{0}_x\Big)\pcpswnd dx+\frac{1}{3}\int_0^1\Big[(\pcsdl{\pc})^2e^{-\pcszmd{\pc}}-(\pcsdl{0})^2e^{-\pcszmd{0}}\Big]\pcpswnd dx\\
&=I_1+I_2+I_3+I_4.\notag
\end{align}
By the same fashion used to derive the estimate \eqref{ss13.3}, the left-side of \eqref{ss14.1} can be estimated as
\begin{align}
\eqref{ss14.1}_l&=\frac{2}{3}\|\pcpswnd_{x}\|^2+\int_0^1\Big[\Big(e^{\pcszmd{\pc}}-e^{\pcszmd{0}}\Big)(\pcswnd{\pc}-\theta_L)+e^{\pcszmd{0}}\pcpswnd\Big]\pcpswnd dx\notag\\
&\geq\frac{2}{3}\|\pcpswnd_{x}\|^2+c\|\pcpswnd\|^2-C\delta\|(\pcpszmd,\pcpswnd)\|^2\notag\\
&\geq c\|\pcpswnd\|_1^2-C\delta\|\pcpszmd\|^2.\label{ss14.2-ss18.3}
\end{align}
We further estimate the integrals $I_i$ $(i=1,2,3,4)$ on the right-side of \eqref{ss14.1} one by one,
\begin{align}
I_1&=\frac{\pc^2}{3}\pcsdl{\pc}\int_0^1\pcszmd{\pc}_{xx}\pcpswnd_xdx+\frac{2\pc^2}{3}\pcsdl{\pc}\int_0^1\pcszmd{\pc}_{x}\pcszmd{\pc}_{xx}\pcpswnd dx\notag\\
&\leq \frac{1}{3}\pc^2|\pcsdl{\pc}|\|\pcszmd{\pc}_{xx}\|\|\pcpswnd_x\|+\frac{2}{3}\pc^2|\pcsdl{\pc}||\pcszmd{\pc}_{x}|_0\|\pcszmd{\pc}_{xx}\|\|\pcpswnd\|\notag\\
&\leq C\pc^2\|\pcpswnd\|_1\leq C\pc^2\|\pcswnd{\pc}-\theta_L+\theta_L-\pcswnd{0}\|_1\notag\\
&\leq C\Big(\|\pcswnd{\pc}-\theta_L\|_1+\|\pcswnd{0}-\theta_L\|_1\Big)\pc^2\notag\\
&\leq C\pc^2.\label{ss18.1}
\end{align}
It is easy to estimate $I_3+I_4$ by the standard computations, that is,
\begin{equation}\label{ss15.1-ss18.2}
I_3+I_4\leq C\delta\|(\pcpszmd,\pcpswnd)\|_1^2.
\end{equation}
However, we need to pay more attention to the integral $I_2$ due to the non-flatness of $\pcszmd{\pc}$,
\begin{align}
I_2&=\frac{2}{3}\int_0^1\Big(\pcsdl{\pc}\pcswnd{\pc}\pcszmd{\pc}_x-\pcsdl{0}\pcswnd{0}\pcszmd{0}_x\Big)\pcpswnd dx\notag\\
&=\frac{2}{3}\int_0^1\Big(\pcpsdl\pcswnd{\pc}\pcszmd{\pc}_x+\pcsdl{0}\pcpswnd\pcszmd{\pc}_x+\pcsdl{0}\pcswnd{0}\pcpszmd_x\Big)\pcpswnd dx\notag\\
&=\frac{2}{3}\int_0^1\Big[\pcpsdl\big(\pcswnd{\pc}-\theta_L+\theta_L\big)\pcszmd{\pc}_x+\pcsdl{0}\pcszmd{\pc}_x\pcpswnd+\pcsdl{0}\pcswnd{0}\pcpszmd_x\Big]\pcpswnd dx\notag\\
&=\frac{2\theta_L}{3}\pcpsdl\int_0^1\pcpswnd\pcszmd{\pc}_xdx+\frac{2}{3}\int_0^1\Big[\big(\pcswnd{\pc}-\theta_L\big)\pcszmd{\pc}_x\pcpsdl+\pcsdl{0}\pcszmd{\pc}_x\pcpswnd+\pcsdl{0}\pcswnd{0}\pcpszmd_x\Big]\pcpswnd dx\notag\\
&=-\frac{2\theta_L}{3}\pcpsdl\int_0^1\pcpswnd_x\pcszmd{\pc}dx+\frac{2}{3}\int_0^1\Big[\big(\pcswnd{\pc}-\theta_L\big)\pcszmd{\pc}_x\pcpsdl+\pcsdl{0}\pcszmd{\pc}_x\pcpswnd+\pcsdl{0}\pcswnd{0}\pcpszmd_x\Big]\pcpswnd dx\notag\\
&\leq-\frac{2\theta_L}{3}\pcpsdl\int_0^1\pcpswnd_x\pcszmd{\pc}dx\notag\\
&\qquad+\frac{2}{3}\Big(|\pcswnd{\pc}-\theta_L|_0|\pcszmd{\pc}_x|_0|\pcpsdl|\|\pcpswnd\|+|\pcsdl{0}||\pcszmd{\pc}_x|_0\|\pcpswnd\|^2+|\pcsdl{0}||\pcswnd{0}|_0\|\pcpszmd_x\|\|\pcpswnd\|\Big)\notag\\
&\leq-\frac{2\theta_L}{3}\pcpsdl\int_0^1\pcpswnd_x\pcszmd{\pc}dx+C\delta\|(\pcpszmd,\pcpswnd)\|_1^2\notag\\
&=-\frac{2\theta_L}{3}\big(\pcsdl{\pc}-\pcsdl{0}\big)\int_0^1\pcpswnd_x\pcszmd{\pc}dx+C\delta\|(\pcpszmd,\pcpswnd)\|_1^2\notag\\
&=-\frac{2\theta_L}{3}\Bigg[\frac{2\big(\bar{b}+\int_0^1\pcswnd{\pc}_{x}\pcszmd{\pc}dx\big)}{\pcsK{\pc}}-\frac{2\big(\bar{b}+\int_0^1\pcswnd{0}_{x}\pcszmd{0}dx\big)}{\pcsK{0}}\Bigg]\int_0^1\pcpswnd_x\pcszmd{\pc}dx+C\delta\|(\pcpszmd,\pcpswnd)\|_1^2\notag\\
&=-\frac{4\theta_L}{3}\Bigg[\frac{1}{\pcsK{\pc}}\int_0^1\Big(\pcswnd{\pc}_{x}\pcszmd{\pc}-\pcswnd{0}_{x}\pcszmd{0}\Big)dx+\bigg(\bar{b}+\int_0^1\pcswnd{0}_{x}\pcszmd{0}dx\bigg)\bigg(\frac{1}{\pcsK{\pc}}-\frac{1}{\pcsK{0}}\bigg)\Bigg]\int_0^1\pcpswnd_x\pcszmd{\pc}dx\notag\\
&\qquad+C\delta\|(\pcpszmd,\pcpswnd)\|_1^2\notag\\
&=-\frac{4\theta_L}{3}\Bigg[\frac{1}{\pcsK{\pc}}\int_0^1\Big(\pcpswnd_{x}\pcszmd{\pc}+\pcswnd{0}_{x}\pcpszmd\Big)dx-\bigg(\bar{b}+\int_0^1\pcswnd{0}_{x}\pcszmd{0}dx\bigg)\frac{\pcsK{\pc}-\pcsK{0}}{\pcsK{\pc}\pcsK{0}}\Bigg]\int_0^1\pcpswnd_x\pcszmd{\pc}dx\notag\\
&\qquad+C\delta\|(\pcpszmd,\pcpswnd)\|_1^2\notag\\
&=\underbrace{-\frac{4\theta_L}{3\pcsK{\pc}}\bigg(\int_0^1\pcpswnd_{x}\pcszmd{\pc}dx\bigg)^2}_{\leq0}\notag\\
&\qquad-\frac{4\theta_L}{3\pcsK{\pc}}\Bigg[\int_0^1\pcswnd{0}_{x}\pcpszmd dx-\frac{\big(\bar{b}+\int_0^1\pcswnd{0}_{x}\pcszmd{0}dx\big)}{\pcsK{0}}\Big(\pcsK{\pc}-\pcsK{0}\Big)\Bigg]\int_0^1\pcpswnd_x\pcszmd{\pc}dx\notag\\
&\qquad+C\delta\|(\pcpszmd,\pcpswnd)\|_1^2\notag\\
&\leq-\frac{4\theta_L}{3\pcsK{\pc}}\Bigg[\int_0^1\pcswnd{0}_{x}\pcpszmd dx-\frac{\pcsdl{0}}{2}\Big(\pcsK{\pc}-\pcsK{0}\Big)\Bigg]\int_0^1\pcpswnd_x\pcszmd{\pc}dx+C\delta\|(\pcpszmd,\pcpswnd)\|_1^2\notag\\
&\leq C\delta\Big(\|\pcpszmd\|+|\pcsK{\pc}-\pcsK{0}|\Big)\|\pcpswnd_x\|+C\delta\|(\pcpszmd,\pcpswnd)\|_1^2\notag\\
&\leq C\delta\|(\pcpszmd,\pcpswnd)\|_1^2,\label{ss17.2}
\end{align}
where we have adopted the notation $\pcsK{\pc}:=K[e^{\pcszmd{\pc}},\pcswnd{\pc}]$ (see formula \eqref{104.1}) for any $\pc\in[0,\pc_1]$ and the following estimates
\begin{equation}\label{Pss4B-ss10.2bc}
0<c\leq\frac{1}{\pcsK{\pc}}\leq C,\qquad |\pcsK{\pc}-\pcsK{0}|\leq C\Big(\|\pcpszmd\|+\|\pcpswnd_x\|\Big),\qquad\forall\pc\in[0,\pc_1]
\end{equation}
which follow from the straightforward but tedious computations. Inserting the estimates \eqref{ss14.2-ss18.3}$\sim$\eqref{ss17.2} into \eqref{ss14.1}, and letting $\delta\ll1$, we have
\begin{equation}\label{ss19.2}
\|\pcpswnd\|_1^2\leq C\delta\|\pcpszmd\|_1^2+C\pc^2.
\end{equation}
Moreover, substituting \eqref{ss19.2} into \eqref{ss13.3}, and letting $\delta$ small enough, we get
\begin{equation}\label{ss19.4}
\|\pcpszmd\|_1^2\leq C\pc^2,\quad\forall\pc\in(0,\pc_1].
\end{equation}
Combining \eqref{ss19.4} with \eqref{ss19.2}, \eqref{ss10.2} and the elliptic estimate $\|\pcpsdws\|_3\leq C\|\pcpszmd\|_1$, we obtain
\begin{equation}\label{ss20.1}
\|\pcpszmd\|_1+|\pcpsdl|+\|\pcpswnd\|_1+\|\pcpsdws\|_3\leq C\pc.
\end{equation}
Next, we solve $\pcpswnd_{xx}$ from the equation \eqref{ss8.1d} and directly take the $L^2$-norm of the resultant equality, the standard but tedious computations yield the following estimate
\begin{equation}\label{ss21.1}
\|\pcpswnd_{xx}\|\leq C\pc\|\pc\pcszmd{\pc}_{xxx}\|+C\pc^2\|\pcszmd{\pc}_{xx}\|+C\Big(\|\pcpszmd\|_1+|\pcpsdl|+\|\pcpswnd\|_1\Big)\leq C\pc.
\end{equation}
Adding \eqref{ss20.1} and \eqref{ss21.1} up, we have
\begin{equation}\label{ss22.2}
\|\pcpszmd\|_1+|\pcpsdl|+\|\pcpswnd\|_2+\|\pcpsdws\|_3\leq C\pc,\quad\forall\pc\in(0,\pc_1].
\end{equation}
By using the exponential transformations $\pcsmd{\pc}=e^{\pcszmd{\pc}}$, $\pcsmd{0}=e^{\pcszmd{0}}$ and the above estimate \eqref{ss22.2}, we have showed the algebraic convergence rate \eqref{ss39.1a}.
Now, we begin to show the convergence \eqref{ss39.1b}. Firstly, differentiating the equation \eqref{ss8.1d} once and solving $\pcpswnd_{xxx}$ from the resultant equation, and taking the $L^2$-norm of the expression of $\pcpswnd_{xxx}$, then these computations yield the following estimate
\begin{align}
\|\pcpswnd_{xxx}\|\leq&C\Big(\|\pc^2\pcszmd{\pc}_{xxxx}\|+\underbrace{\pc\|\pc\pcszmd{\pc}_{xxx}\|+\pc^2\|\pcszmd{\pc}_{xx}\|+\|\pcpszmd\|_1+|\pcpsdl|+\|\pcpswnd\|_2}_{\leq C\pc}+\|\pcpszmd_{xx}\|\Big)\notag\\
\leq&C\Big(\|\pcpszmd_{xx}\|+\|\pc^2\pcszmd{\pc}_{xxxx}\|\Big) +C\pc.\label{ss22.1c}
\end{align}
Adding the elliptic estimate $\|\pcpsdws_{xxxx}\|\leq C\|\pcpszmd_{xx}\|$ to the above estimate \eqref{ss22.1c}, we have
\begin{equation}\label{wnd3+dws4}
\|(\pcpswnd_{xxx},\pcpsdws_{xxxx})\|\leq C\Big(\|\pcpszmd_{xx}\|+\|\pc^2\pcszmd{\pc}_{xxxx}\|\Big) +C\pc,\quad\forall\pc\in(0,\pc_1].
\end{equation}
In order to complete the proof, we need to establish the convergence results $\|\pcpszmd_{xx}\|\rightarrow0$, $\|\pc\pcszmd{\pc}_{xxx}\|\rightarrow0$ and $\|\pc^2\pcszmd{\pc}_{xxxx}\|\rightarrow0$ as $\pc\rightarrow0$. To this end, we first show $\|\pcpszmd_{xx}\|$ converges to zero as $\pc$ tends to zero. From the boundedness \eqref{ss9.2} of $\|\pcszmd{\pc}\|_2$ and the strong convergence \eqref{ss22.2}, we have
\begin{equation}\label{ss23.1}
\pcszmd{\pc}_{xx}\rightharpoonup\pcszmd{0}_{xx}\quad\text{in}\ L^2(\Omega)\ \text{weakly as}\ \pc\rightarrow0.
\end{equation}
However, we need to improve the above weak convergence into strong convergence. For this purpose, differentiating the equation \eqref{ss8.1b} once, multiplying the resultant equality by $\pcszmd{\pc}_{xx}+(\pcszmd{\pc}_x)^2/2$ and integrating the result over the domain $\Omega$, the integration by parts yields
\begin{equation}\label{ss26.1}
\int_0^1\pcsS{\pc}(\pcszmd{\pc}_{xx})^2dx+\frac{\pc^2}{2}\int_0^1\bigg\{\bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_{x})^2}{2}\bigg]_x\bigg\}^2dx=\pcR{\pc},\quad\forall\pc\in(0,\pc_1],
\end{equation}
where
\begin{multline}\label{ss26.2}
\pcR{\pc}:=-\int_0^1\pcsS{\pc}\pcszmd{\pc}_{xx}\frac{(\pcszmd{\pc}_x)^2}{2}dx-\int_0^1\pcsS{\pc}_x\pcszmd{\pc}_x\bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_{x})^2}{2}\bigg]dx\\
+\int_0^1\bigg[\pcsdws{\pc}_{xx}-\Big(\pcsdl{\pc}e^{-\pcszmd{\pc}}\Big)_x-\pcswnd{\pc}_{xx}\bigg]\bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_{x})^2}{2}\bigg]dx,\qquad\forall\pc\in[0,\pc_1].
\end{multline}
Similarly, differentiating the equation \eqref{ss8.1a} once, multiplying the resultant equality by $\pcszmd{0}_{xx}+(\pcszmd{0}_{x})^2/2$ and integrating the result over $\Omega$, we have
\begin{equation}\label{ss26.4}
\int_0^1\pcsS{0}(\pcszmd{0}_{xx})^2dx=\pcR{0},\quad\text{where}\ \pcR{0}\ \text{is given by}\ \eqref{ss26.2}\ \text{with}\ \pc=0.
\end{equation}
By using the estimates \eqref{ss9.3}, \eqref{ss9.2}, \eqref{ss22.2}, the weak convergence result \eqref{ss23.1} and the standard computations, we obtain
\begin{align}
0\leq|\pcR{\pc}-\pcR{0}|\leq&C\Big(\|\pcpszmd\|_1+|\pcpsdl|+\|\pcpswnd\|_2+\|\pcpsdws_{xx}\|\Big)\notag\\
&\quad+\bigg|\int_0^1\underbrace{\bigg[\pcsS{0}\frac{(\pcszmd{0}_x)^2}{2}+\pcsS{0}\pcszmd{0}_x+\pcswnd{0}_{xx}+\pcsdws{0}_{xx}+\Big(\pcsdl{0}e^{-\pcszmd{0}}\Big)_x\bigg]}_{=:f^0\in L^2(\Omega)}\big(\pcszmd{\pc}_{xx}-\pcszmd{0}_{xx}\big)dx\bigg|\notag\\
\leq&C\pc+\Big|\big<f^0, \pcszmd{\pc}_{xx}-\pcszmd{0}_{xx}\big>_{L^2(\Omega)}\Big|\rightarrow0\quad\text{as}\ \pc\rightarrow0.\label{ss31.2}
\end{align}
Combining \eqref{ss26.4} with \eqref{ss31.2}, we have
\begin{equation}\label{ss31.3}
\lim_{\pc\rightarrow0}\pcR{\pc}=\int_0^1\pcsS{0}(\pcszmd{0}_{xx})^2dx.
\end{equation}
On the other hand, owing to the estimates \eqref{ss12.1-2-3}, \eqref{ss9.2}, the Sobolev inequality, we obtain
\begin{align}
0\leq\bigg|\int_0^1\big(\pcsS{\pc}-\pcsS{0}\big)(\pcszmd{\pc}_{xx})^2dx\bigg|\leq&\big|\big(\pcsS{\pc}-\pcsS{0}\big)\big|_0\|\pcszmd{\pc}_{xx}\|^2\notag\\
\leq&C\Big(|\pcpszmd|_0+|\pcpsdl|+|\pcpswnd|_0\Big)\notag\\
\leq&C\Big(\|\pcpszmd\|_1+|\pcpsdl|+\|\pcpswnd\|_1\Big)\notag\\
\leq&C\pc\rightarrow0\quad\text{as}\ \pc\rightarrow0.\label{ss32.2}
\end{align}
Combining the limits \eqref{ss31.3}, \eqref{ss32.2} with the equality \eqref{ss26.1}, we have
\begin{align}
\Bigg\{\limsup_{\pc\rightarrow0}\bigg[\int_0^1\pcsS{0}(\pcszmd{\pc}_{xx})^2dx\bigg]^{1/2}\Bigg\}^2&\leq\limsup_{\pc\rightarrow0}\int_0^1\pcsS{0}(\pcszmd{\pc}_{xx})^2dx\notag\\
&=\lim_{\pc\rightarrow0}\int_0^1\big(\pcsS{\pc}-\pcsS{0}\big)(\pcszmd{\pc}_{xx})^2dx+\limsup_{\pc\rightarrow0}\int_0^1\pcsS{0}(\pcszmd{\pc}_{xx})^2dx\notag\\
&=\limsup_{\pc\rightarrow0}\int_0^1\big(\pcsS{\pc}-\pcsS{0}+\pcsS{0}\big)(\pcszmd{\pc}_{xx})^2dx\notag\\
&=\limsup_{\pc\rightarrow0}\int_0^1\pcsS{\pc}(\pcszmd{\pc}_{xx})^2dx\notag\\
&\leq\limsup_{\pc\rightarrow0}\pcR{\pc}=\lim_{\pc\rightarrow0}\pcR{\pc}=\int_0^1\pcsS{0}(\pcszmd{0}_{xx})^2dx,\label{ss33.3}
\end{align}
where we have used the non-negativity of the second term on the left-side of the equality \eqref{ss26.1}. Motivated by \eqref{ss33.3}, we choose $\pcsS{0}$ as the weight to define a weighted-$L^2$ space as follows
\begin{equation}\label{ss34.1a}
L^2_{\pcsS{0}}(\Omega):=\Bigg\{f:\Omega\rightarrow\mathbb{R} \text{ is measurable}\ \Bigg|\ \int_0^1\pcsS{0}|f|^2dx<+\infty\Bigg\}
\end{equation}
with the inner product
\begin{equation}\label{ss34.1b}
\big<f,g\big>_{L^2_{\pcsS{0}}(\Omega)}:=\int_0^1\pcsS{0}fgdx
\end{equation}
and the associated norm
\begin{equation}\label{ss34.1c}
\|f\|_{L^2_{\pcsS{0}}(\Omega)}:=\Bigg(\int_0^1\pcsS{0}|f|^2\;dx\Bigg)^{1/2}.
\end{equation}
Since the weight function $\pcsS{0}$ is strictly positive and continuous, then the weighted-$L^2$ space $L^2_{\pcsS{0}}(\Omega)$ is a Hilbert space and the weak convergence \eqref{ss23.1} implies
\begin{equation}\label{ss34.2}
\pcszmd{\pc}_{xx}\rightharpoonup\pcszmd{0}_{xx}\quad\text{in}\ L^2_{\pcsS{0}}(\Omega)\ \text{weakly as}\ \pc\rightarrow0.
\end{equation}
Furthermore, we can rewrite the inequality \eqref{ss33.3} in terms of the norm defined in \eqref{ss34.1c} as
\begin{equation}\label{ss33.3wn}
\limsup_{\pc\rightarrow0}\|\pcszmd{\pc}_{xx}\|_{L^2_{\pcsS{0}}(\Omega)}\leq\|\pcszmd{0}_{xx}\|_{L^2_{\pcsS{0}}(\Omega)}.
\end{equation}
The weak convergence \eqref{ss34.2} together with \eqref{ss33.3wn} implies the strong convergence
\begin{equation}\label{ss34.3}
\pcszmd{\pc}_{xx}\rightarrow\pcszmd{0}_{xx}\quad\text{in}\ L^2_{\pcsS{0}}(\Omega)\ \text{strongly as}\ \pc\rightarrow0,
\end{equation}
which immediately implies
\begin{equation}\label{ss34.5}
\|\pcpszmd_{xx}\|\rightarrow0,\quad\text{as}\ \pc\rightarrow0.
\end{equation}
In addition, we prove $\|\pc\pcszmd{\pc}_{xxx}\|$ converges to zero as $\pc$ tends to zero. From the strong convergence \eqref{ss34.3}, we can directly deduce that
\begin{equation}\label{ss34.6}
\lim_{\pc\rightarrow0}\int_0^1\pcsS{0}(\pcszmd{\pc}_{xx})^2dx=\int_0^1\pcsS{0}(\pcszmd{0}_{xx})^2dx.
\end{equation}
Combining \eqref{ss32.2} with \eqref{ss34.6}, we have
\begin{equation}\label{ss35.1}
\lim_{\pc\rightarrow0}\int_0^1\pcsS{\pc}(\pcszmd{\pc}_{xx})^2dx=\int_0^1\pcsS{0}(\pcszmd{0}_{xx})^2dx.
\end{equation}
Letting $\pc\rightarrow0$ in the equality \eqref{ss26.1}, and using the limit results \eqref{ss31.3} and \eqref{ss35.1}, we can easily see that
\begin{equation}\label{ss35.4}
\pc\bigg\|\bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_{x})^2}{2}\bigg]_x\bigg\|\rightarrow0,\quad\text{as}\ \pc\rightarrow0.
\end{equation}
Therefore,
\begin{align}
0\leq\|\pc\pcszmd{\pc}_{xxx}\|=&\bigg\|\pc\bigg\{\bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_{x})^2}{2}\bigg]_x-\pcszmd{\pc}_{x}\pcszmd{\pc}_{xx}\bigg\}\bigg\|\notag\\
\leq&\pc\bigg\|\bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_{x})^2}{2}\bigg]_x\bigg\|+C\pc\rightarrow0,\quad\text{as}\ \pc\rightarrow0.\label{ss36.1}
\end{align}
Finally, we show $\|\pc^2\pcszmd{\pc}_{xxxx}\|$ converges to zero as $\pc$ tends to zero. Differentiating the equation \eqref{ss8.1b} once, solving the quantum term from the resultant equality and taking the $L^2$-norm of this quantum term, we obtain
\begin{equation}\label{ss36.2}
\frac{\pc^2}{2}\bigg\|\bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_{x})^2}{2}\bigg]_{xx}\bigg\|=\pcr{\pc},\quad\forall\pc\in(0,\pc_1],
\end{equation}
where
\begin{equation}\label{ss36.3}
\pcr{\pc}:=\Big\|\Big(\pcsS{\pc}\pcszmd{\pc}_x\Big)_x+\pcswnd{\pc}_{xx}-\pcsdws{\pc}_{xx}+\Big(\pcsdl{\pc}e^{-\pcszmd{\pc}}\Big)_x\Big\|,\quad\forall\pc\in[0,\pc_1].
\end{equation}
The standard computations yield
\begin{align}
0\leq\pcr{\pc}=\pcr{\pc}-\pcr{0}\leq&\Big\|\Big(\pcsS{\pc}\pcszmd{\pc}_x\Big)_x-\Big(\pcsS{0}\pcszmd{0}_x\Big)_x+\pcpswnd_{xx}-\pcpsdws_{xx}+\Big(\pcsdl{\pc}e^{-\pcszmd{\pc}}\Big)_x-\Big(\pcsdl{0}e^{-\pcszmd{0}}\Big)_x\Big\|\notag\\
\leq&C\Big(\|\pcpszmd\|_1+|\pcpsdl|+\|\pcpswnd_{xx}\|+\|\pcpsdws_{xx}\|\Big)+C\|\pcpszmd_{xx}\|\notag\\
\leq&C\pc+C\|\pcpszmd_{xx}\|\rightarrow0,\quad\text{as}\ \pc\rightarrow0,
\end{align}
where we have used $\pcr{0}=0$ which follows from the differentiation of the equation \eqref{ss8.1a}. Consequently,
\begin{align}
0\leq\|\pc^2\pcszmd{\pc}_{xxxx}\|=&\bigg\|\pc^2\bigg\{\bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_{x})^2}{2}\bigg]_{xx}-\bigg[\frac{(\pcszmd{\pc}_{x})^2}{2}\bigg]_{xx}\bigg\}\bigg\|\notag\\
\leq&\pc^2\bigg\|\bigg[\pcszmd{\pc}_{xx}+\frac{(\pcszmd{\pc}_{x})^2}{2}\bigg]_{xx}\bigg\|+\pc^2\bigg\|\bigg[\frac{(\pcszmd{\pc}_{x})^2}{2}\bigg]_{xx}\bigg\|\notag\\
=&2\pcr{\pc}+\pc^2\big\|\big(\pcszmd{\pc}_{xx}\big)^2+\pcszmd{\pc}_x\pcszmd{\pc}_{xxx}\big\|\notag\\
\leq&2\pcr{\pc}+\pc^2\Big(|\pcszmd{\pc}_{xx}|_0\|\pcszmd{\pc}_{xx}\|+|\pcszmd{\pc}_{x}|_0\|\pcszmd{\pc}_{xxx}\|\Big)\notag\\
\leq&2\pcr{\pc}+C\pc^2\Big(\|\pcszmd{\pc}_{xx}\|_1\|\pcszmd{\pc}_{xx}\|+\|\pcszmd{\pc}_{x}\|_1\|\pcszmd{\pc}_{xxx}\|\Big)\notag\\
\leq&2\pcr{\pc}+C\pc\rightarrow0,\quad\text{as}\ \pc\rightarrow0.\label{ss37.3}
\end{align}
From \eqref{wnd3+dws4}, \eqref{ss34.5}, \eqref{ss36.1} and \eqref{ss37.3}, we know that
\begin{equation}\label{ss38.1b}
\big\|\big(\pcpszmd_{xx},\pc\pcszmd{\pc}_{xxx},\pc^2\pcszmd{\pc}_{xxxx},\pcpswnd_{xxx},\pcpsdws_{xxxx}\big)\big\|\rightarrow0,\quad\text{as}\ \pc\rightarrow0.
\end{equation}
By using the exponential transformations $\pcsmd{\pc}=e^{\pcszmd{\pc}}$ and $\pcsmd{0}=e^{\pcszmd{0}}$ again, the strong convergence \eqref{ss38.1b} implies the convergence \eqref{ss39.1b}.
\end{proof}
\subsection{Non-stationary case}\label{Subsect.4.2}
In this subsection, we continue to discuss the semi-classical limit of the global solutions based on the existence and uniqueness results in Lemma \ref{lem2} and Theorem \ref{thm2}. We introduce the error variables as follows
\begin{equation}\label{131.1}
\pcpmd:=\pcmd{\pc}-\pcmd{0},\quad \pcpdl:=\pcdl{\pc}-\pcdl{0},\quad\pcpwnd:=\pcwnd{\pc}-\pcwnd{0},\quad\pcpdws:=\pcdws{\pc}-\pcdws{0}.
\end{equation}
Since the global solutions $(\pcmd{\pc},\pcdl{\pc},\pcwnd{\pc},\pcdws{\pc})$ and $(\pcmd{0},\pcdl{0},\pcwnd{0},\pcdws{0})$ satisfy the same initial and boundary conditions, the error variables $(\pcpmd,\pcpdl,\pcpwnd,\pcpdws)$ satisfy the following initial and boundary conditions
\begin{gather}
(\pcpmd,\pcpdl,\pcpwnd,\pcpdws)(0,x)=(0,0,0,0),\label{134.1}\\
(\pd{t}{k}\pcpmd,\pd{t}{k}\pcpwnd,\pd{t}{k}\pcpdws)(t,0)=(\pd{t}{k}\pcpmd,\pd{t}{k}\pcpwnd,\pd{t}{k}\pcpdws)(t,1)=(0,0,0),\quad k=0,1.\label{134.2}
\end{gather}
Moreover, subtracting \eqref{1dfhd} from \eqref{1dfqhd}, the error variables $(\pcpmd,\pcpdl,\pcpwnd,\pcpdws)$ also satisfy the equations
\begin{gather}
\pcpmd_t+\pcpdl_x=0,\label{136.1a}\\
\pcpdl_t+\pcpdl=\mathcal{H}_1(t,x)+\pc^2\pcmd{\pc}\Bigg[\frac{\big(\sqrt{\pcmd{\pc}}\big)_{xx}}{\sqrt{\pcmd{\pc}}}\Bigg]_x,\label{136.1b}\\
\pcmd{\pc}\pcpwnd_t-\frac{2}{3}\pcpwnd_{xx}+\pcmd{0}\pcpwnd=\mathcal{H}_2(t,x;\pc),\label{157.2}\\
\pcpdws_{xx}=\pcpmd,\label{136.1d}
\end{gather}
where
\begin{align}
\mathcal{H}_1(t,x):=&-\big(\pcpmd_x\pcwnd{\pc}+\pcmd{0}_x\pcpwnd\big)-\big(\pcpmd\pcwnd{\pc}_x+\pcmd{0}\pcpwnd_x\big)+\big(\pcpmd\pcdws{\pc}_x+\pcmd{0}\pcpdws_x\big)\notag\\
&+\Bigg[\bigg(\frac{\pcdl{\pc}}{\pcmd{\pc}}\bigg)^2\pcmd{\pc}_x-\bigg(\frac{\pcdl{0}}{\pcmd{0}}\bigg)^2\pcmd{0}_x\Bigg]-2\bigg(\frac{\pcdl{\pc}}{\pcmd{\pc}}\pcdl{\pc}_x-\frac{\pcdl{0}}{\pcmd{0}}\pcdl{0}_x\bigg),\label{136.1br}
\end{align}
and
\begin{align}
\mathcal{H}_2(t,x;\pc):=&-\pcwnd{0}_t\pcpmd-\big(\pcpdl\pcwnd{\pc}_x+\pcdl{0}\pcpwnd_x\big)-\frac{2}{3}\bigg[\pcmd{\pc}\pcwnd{\pc}\bigg(\frac{\pcdl{\pc}}{\pcmd{\pc}}\bigg)_x-\pcmd{0}\pcwnd{0}\bigg(\frac{\pcdl{0}}{\pcmd{0}}\bigg)_x\bigg]\notag\\
&+\frac{1}{3}\Bigg[\frac{\big(\pcdl{\pc}\big)^2}{\pcmd{\pc}}-\frac{\big(\pcdl{0}\big)^2}{\pcmd{0}}\Bigg]-\pcpmd\big(\pcwnd{\pc}-\theta_L\big)+\frac{\pc^2}{3}\Bigg[\pcmd{\pc}\Bigg(\frac{\pcdl{\pc}}{\pcmd{\pc}}\Bigg)_{xx}\Bigg]_x.\label{157.3}
\end{align}
Differentiating the equation \eqref{136.1b} with respect to $x$ and using the equation \eqref{136.1a}, we obtain the equation
\begin{align}\label{135.1}
\pcpmd_{tt}-\pcwnd{\pc}\pcpmd_{xx}+\pcpmd_t=&\pcpmd_x\pcwnd{\pc}_x+\big(\pcmd{0}_x\pcpwnd\big)_x+\big(\pcpmd\pcwnd{\pc}_x+\pcmd{0}\pcpwnd_x\big)_x\notag\\
&-\big(\pcpmd\pcdws{\pc}_x+\pcmd{0}\pcpdws_x\big)_x-\Bigg[\bigg(\frac{\pcdl{\pc}}{\pcmd{\pc}}\bigg)^2\pcmd{\pc}_x-\bigg(\frac{\pcdl{0}}{\pcmd{0}}\bigg)^2\pcmd{0}_x\Bigg]_x\notag\\
&+2\bigg(\frac{\pcdl{\pc}}{\pcmd{\pc}}\pcdl{\pc}_x-\frac{\pcdl{0}}{\pcmd{0}}\pcdl{0}_x\bigg)_x-\pc^2\Bigg\{\pcmd{\pc}\Bigg[\frac{\big(\sqrt{\pcmd{\pc}}\big)_{xx}}{\sqrt{\pcmd{\pc}}}\Bigg]_x\Bigg\}_x.
\end{align}
From the estimates \eqref{0de} and \eqref{de}, we can deduce the following estimates
\begin{gather}
\pcmd{0}(t,x),\ \pcwnd{0}(t,x),\ \pcS{0}(t,x):=S[\pcmd{0},\pcdl{0},\pcwnd{0}]\geq c>0,\notag\\
\|(\pcmd{0},\pcdl{0},\pcwnd{0},\pcdws{0})(t)\|_2+\|(\pcmd{0}_t,\pcdl{0}_t,\pcwnd{0}_t)(t)\|_1\leq C,\label{138.1}
\end{gather}
and
\begin{gather}
\pcmd{\pc}(t,x),\ \pcwnd{\pc}(t,x),\ \pcS{\pc}(t,x):=S[\pcmd{\pc},\pcdl{\pc},\pcwnd{\pc}]\geq c>0,\notag\\
\|(\pcmd{\pc},\pcdl{\pc},\pcwnd{\pc},\pcdws{\pc})(t)\|_2+\|(\pc\pd{x}{3}\pcmd{\pc},\pc\pd{x}{3}\pcdl{\pc},\pc^2\pd{x}{4}\pcmd{\pc})(t)\|+\|(\pcmd{\pc}_t,\pcdl{\pc}_t)(t)\|_1+\|\pcwnd{\pc}_t(t)\|\leq C,\label{138.2}
\end{gather}
where $c$ and $C$ are positive constants independent of $\pc$, $\delta$, $x$ and $t$.
\begin{proof}[\textbf{Proof of Theorem \ref{thm4}.}]
Based on Theorem \ref{thm3}, Lemma \ref{lem2} and Theorem \ref{thm2}, it is easy to see that the assumption \eqref{sic} guarantees the coexistence of the quantum and limit global solutions with the same initial and boundary data. Thus, the above facts \eqref{134.1}$\sim$\eqref{138.2} are available now.
Multiplying the equation \eqref{136.1b} by $\pcpdl$ and integrating the resultant equality over the domain $\Omega$, we have
\begin{align}
\frac{d}{dt}\int_0^1\frac{1}{2}\big(\pcpdl\big)^2dx+\big\|\pcpdl(t)\big\|^2&=\int_0^1\mathcal{H}_1\pcpdl dx+\pc^2\int_0^1\pcmd{\pc}\Bigg[\frac{\big(\sqrt{\pcmd{\pc}}\big)_{xx}}{\sqrt{\pcmd{\pc}}}\Bigg]_x\pcpdl dx \notag\\
&\leq C\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1^2+C\pc^2.\label{142.1}
\end{align}
In the derivation of the estimate \eqref{142.1}, we have used the Cauchy-Schwarz inequality, the elliptic estimate $\|\pcpdws(t)\|_2\leq C\|\pcpmd(t)\|$ and the following estimate
\begin{equation}
\|\mathcal{H}_1(t)\|\leq C\big\|\big(\pcpmd,\pcpmd_x,\pcpdl,\pcpdl_x,\pcpwnd,\pcpwnd_x,\pcpdws_x\big)(t)\big\|\label{140.1-141.1}
\end{equation}
to control the first term on the right-side of this equality, and we have also used the integration by parts, the boundary condition \eqref{bc-b} and the estimates \eqref{138.1}$\sim$\eqref{138.2} to bound the last term on the right-side as follows
\begin{align}
&\pc^2\int_0^1\pcmd{\pc}\Bigg[\frac{\big(\sqrt{\pcmd{\pc}}\big)_{xx}}{\sqrt{\pcmd{\pc}}}\Bigg]_x\pcpdl dx\notag\\
=&-\pc^2\int_0^1\frac{\big(\sqrt{\pcmd{\pc}}\big)_{xx}}{\sqrt{\pcmd{\pc}}}\big(\pcmd{\pc}\pcpdl\big)_xdx\notag\\
\leq&\pc^2\Bigg\|\Bigg[\frac{\big(\sqrt{\pcmd{\pc}}\big)_{xx}}{\sqrt{\pcmd{\pc}}}\Bigg](t)\Bigg\|\Big\|\big(\pcmd{\pc}\pcpdl\big)_x(t)\Big\|\notag\\
\leq&\pc^2\Bigg|\frac{1}{\sqrt{\pcmd{\pc}}}(t)\Bigg|_0\Big\|\big(\sqrt{\pcmd{\pc}}\big)_{xx}(t)\Big\|\Big(\big|\pcmd{\pc}_x(t)\big|_0\big\|\pcpdl(t)\big\|+\big|\pcmd{\pc}(t)\big|_0\big\|\pcpdl_x(t)\big\|\Big)\notag\\
\leq&C\pc^2.
\end{align}
Multiplying the equation \eqref{135.1} by $\pcpmd_t$ and integrating the resultant equality over the domain $\Omega$, we obtain
\allowdisplaybreaks
\begin{align}
&\frac{d}{dt}\int_0^1\frac{1}{2}\big(\pcpdl_x\big)^2dx-\int_0^1\pcwnd{\pc}\pcpmd_{xx}\pcpmd_tdx+\big\|\pcpdl_x(t)\big\|^2\notag\\
=&\int_0^1\Big[\pcpmd_x\pcwnd{\pc}_x+\big(\pcmd{0}_x\pcpwnd\big)_x-\big(\pcpmd\pcdws{\pc}_x+\pcmd{0}\pcpdws_x\big)_x\Big]\pcpmd_tdx\notag\\
&+\int_0^1\big(\pcpmd\pcwnd{\pc}_x+\pcmd{0}\pcpwnd_x\big)_x\pcpmd_tdx-\int_0^1\Bigg[\bigg(\frac{\pcdl{\pc}}{\pcmd{\pc}}\bigg)^2\pcmd{\pc}_x-\bigg(\frac{\pcdl{0}}{\pcmd{0}}\bigg)^2\pcmd{0}_x\Bigg]_x\pcpmd_tdx\notag\\
&+2\int_0^1\bigg(\frac{\pcdl{\pc}}{\pcmd{\pc}}\pcdl{\pc}_x-\frac{\pcdl{0}}{\pcmd{0}}\pcdl{0}_x\bigg)_x\pcpmd_tdx-\int_0^1\pc^2\Bigg\{\pcmd{\pc}\Bigg[\frac{\big(\sqrt{\pcmd{\pc}}\big)_{xx}}{\sqrt{\pcmd{\pc}}}\Bigg]_x\Bigg\}_x\pcpmd_tdx\notag\\
=&\Lambda_1+\Lambda_2+\Lambda_3+\Lambda_4+\Lambda_5,\label{142.2}
\end{align}
where we have used the equation \eqref{136.1a}. Next, we respectively estimate the second term on the left-side of the equation \eqref{142.2} and the integrals $\Lambda_l$, $l=1,2,\cdots,5$ on the right-side of \eqref{142.2} by using the integration by parts, the Sobolev inequality, the H\"older inequality, the Cauchy-Schwarz inequality, the equation \eqref{136.1a}, the boundary condition \eqref{134.2}, the equation \eqref{157.2} and the estimates \eqref{138.1}$\sim$\eqref{138.2} as follows
\begin{align}
-\int_0^1\pcwnd{\pc}\pcpmd_{xx}\pcpmd_tdx=&\int_0^1\big(\pcwnd{\pc}\pcpmd_t\big)_x\pcpmd_xdx\notag\\
=&\frac{d}{dt}\int_0^1\frac{1}{2}\pcwnd{0}\big(\pcpmd_x\big)^2dx-\int_0^1\frac{1}{2}\pcwnd{0}_t\big(\pcpmd_x\big)^2dx+\int_0^1\pcpwnd\pcpmd_{tx}\pcpmd_xdx\notag\\
&-\int_0^1\pcwnd{\pc}_x\pcpdl_x\pcpmd_xdx\notag\\
\geq&\frac{d}{dt}\int_0^1\frac{1}{2}\pcwnd{0}\big(\pcpmd_x\big)^2dx\notag\\
&-C\Big(\big|\pcwnd{0}_t\big|_0\big\|\pcpmd_x\big\|^2+\big|\pcpwnd\big|_0\big\|\pcpmd_{tx}\big\|\big\|\pcpmd_x\big\|+\big|\pcwnd{\pc}_x\big|_0\big\|\pcpdl_x\big\|\big\|\pcpmd_{x}\big\|\Big)\notag\\
\geq&\frac{d}{dt}\int_0^1\frac{1}{2}\pcwnd{0}\big(\pcpmd_x\big)^2dx-C\Big(\big\|\pcpmd_x\big\|^2+\big\|\pcpwnd\big\|_1\big\|\pcpmd_x\big\|+\big\|\pcpdl_x\big\|\big\|\pcpmd_{x}\big\|\Big)\notag\\
\geq&\frac{d}{dt}\int_0^1\frac{1}{2}\pcwnd{0}\big(\pcpmd_x\big)^2dx-C\big\|\big(\pcpmd_x,\pcpdl_x,\pcpwnd,\pcpwnd_x\big)(t)\big\|^2,\label{143.2}
\end{align}
and
\begin{equation}\label{145.1-2+152.2}
\Lambda_1\leq C\big\|\big(\pcpmd,\pcpmd_x,\pcpdl_x,\pcpwnd,\pcpwnd_x\big)(t)\big\|^2,
\end{equation}
and
\begin{align}
\Lambda_2=&-\int_0^1\big(\pcpmd_x\pcwnd{\pc}_x+\pcpmd\pcwnd{\pc}_{xx}+\pcmd{0}_x\pcpwnd_x+\pcmd{0}\pcpwnd_{xx}\big)\pcpdl_xdx\notag\\
\leq&-\int_0^1\pcmd{0}\pcpwnd_{xx}\pcpdl_xdx+C\big\|\big(\pcpmd,\pcpmd_x,\pcpdl_x,\pcpwnd_x\big)(t)\big\|^2\notag\\
=&\frac{3}{2}\int_0^1\pcmd{0}\big[\mathcal{H}_2(t,x;\pc)-\pcmd{\pc}\pcpwnd_t-\pcmd{0}\pcpwnd\big]\pcpdl_xdx+C\big\|\big(\pcpmd,\pcpmd_x,\pcpdl_x,\pcpwnd_x\big)(t)\big\|^2\notag\\
\leq&\mu\big\|\pcpwnd_t(t)\big\|^2+C_\mu\big\|\pcpdl_x(t)\big\|^2+C\|\mathcal{H}_2(t;\pc)\|^2+C\big\|\big(\pcpmd,\pcpmd_x,\pcpdl_x,\pcpwnd,\pcpwnd_x\big)(t)\big\|^2\notag\\
\leq&\mu\big\|\pcpwnd_t(t)\big\|^2+C_\mu\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1^2+C\pc^2,\label{152.1}
\end{align}
where we have used the estimate
\begin{align}
&\|\mathcal{H}_2(t;\pc)\|\notag\\
\leq&C\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1+C\Bigg\|\pc^2\Bigg[\pcmd{\pc}\Bigg(\frac{\pcdl{\pc}}{\pcmd{\pc}}\Bigg)_{xx}\Bigg]_x\Bigg\|\notag\\
=&C\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1\notag\\
&+C\Bigg\|\pc^2\Bigg[\pcdl{\pc}_{xxx}-2\frac{\pcmd{\pc}_x}{\pcmd{\pc}}\pcdl{\pc}_{xx}+4\bigg(\frac{\pcmd{\pc}_x}{\pcmd{\pc}}\bigg)^2\pcdl{\pc}_x-3\frac{\pcdl{\pc}_x}{\pcmd{\pc}}\pcmd{\pc}_{xx}-4\bigg(\frac{\pcmd{\pc}_x}{\pcmd{\pc}}\bigg)^3\pcdl{\pc}+5\frac{\pcdl{\pc}\pcmd{\pc}_x}{(\pcmd{\pc})^2}\pcmd{\pc}_{xx}-\frac{\pcdl{\pc}}{\pcmd{\pc}}\pcmd{\pc}_{xxx}\Bigg]\Bigg\|\notag\\
\leq&C\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1+C\pc\big\|\big(\pc\pcdl{\pc}_{xxx},\pcdl{\pc}_{xx},\pcdl{\pc}_x,\pcmd{\pc}_{xx},\pcdl{\pc},\pcmd{\pc}_{xx},\pc\pcmd{\pc}_{xxx}\big)(t)\big\|\notag\\
\leq&C\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1+C\pc\label{148.3-151.1}
\end{align}
in the derivation of \eqref{152.1}. Next, we continue to estimate
\begin{align}
\Lambda_3\leq&-\int_0^1\bigg(\frac{\pcdl{0}}{\pcmd{0}}\bigg)^2\pcpmd_{xx}\pcpmd_tdx+C\big\|\big(\pcpmd,\pcpdl\big)(t)\big\|_1^2\notag\\
=&\int_0^1\bigg[\bigg(\frac{\pcdl{0}}{\pcmd{0}}\bigg)^2\pcpmd_t\bigg]_x\pcpmd_{x}dx+C\big\|\big(\pcpmd,\pcpdl\big)(t)\big\|_1^2\notag\\
=&\frac{d}{dt}\int_0^1\frac{1}{2}\bigg(\frac{\pcdl{0}}{\pcmd{0}}\bigg)^2\big(\pcpmd_x\big)^2dx-\int_0^1\frac{1}{2}\bigg[\bigg(\frac{\pcdl{0}}{\pcmd{0}}\bigg)^2\bigg]_t\big(\pcpmd_x\big)^2dx-\int_0^1\bigg[\bigg(\frac{\pcdl{0}}{\pcmd{0}}\bigg)^2\bigg]_x\pcpdl_x\pcpmd_xdx\notag\\
&+C\big\|\big(\pcpmd,\pcpdl\big)(t)\big\|_1^2\notag\\
\leq&\frac{d}{dt}\int_0^1\frac{1}{2}\bigg(\frac{\pcdl{0}}{\pcmd{0}}\bigg)^2\big(\pcpmd_x\big)^2dx+C\big\|\big(\pcpmd,\pcpdl\big)(t)\big\|_1^2,\label{153.1}
\end{align}
and
\begin{align}
\Lambda_4\leq&2\int_0^1\frac{\pcdl{0}}{\pcmd{0}}\pcpdl_{xx}\pcpmd_tdx+C\big\|\big(\pcpmd,\pcpdl\big)(t)\big\|_1^2\notag\\
=&-2\int_0^1\frac{\pcdl{0}}{\pcmd{0}}\pcpmd_{tx}\pcpmd_tdx+C\big\|\big(\pcpmd,\pcpdl\big)(t)\big\|_1^2\notag\\
=&-\int_0^1\frac{\pcdl{0}}{\pcmd{0}}\big[\big(\pcpmd_t\big)^2\big]_xdx+C\big\|\big(\pcpmd,\pcpdl\big)(t)\big\|_1^2\notag\\
=&\int_0^1\bigg(\frac{\pcdl{0}}{\pcmd{0}}\bigg)_x\big(\pcpdl_x\big)^2dx+C\big\|\big(\pcpmd,\pcpdl\big)(t)\big\|_1^2\notag\\
\leq&C\big\|\big(\pcpmd,\pcpdl\big)(t)\big\|_1^2,\label{gs154.1}
\end{align}
and
\begin{align}
\Lambda_5=&\int_0^1\pc^2\pcmd{\pc}\Bigg[\frac{\big(\sqrt{\pcmd{\pc}}\big)_{xx}}{\sqrt{\pcmd{\pc}}}\Bigg]_x\pcpmd_{tx}dx\notag\\
=&\int_0^1\pc^2\Big[\sqrt{\pcmd{\pc}}\big(\sqrt{\pcmd{\pc}}\big)_{xxx}-\big(\sqrt{\pcmd{\pc}}\big)_{x}\big(\sqrt{\pcmd{\pc}}\big)_{xx}\Big]\pcpmd_{tx}dx\notag\\
\leq&C\pc^2\Big(\big\|\big(\sqrt{\pcmd{\pc}}\big)_{xxx}(t)\big\|+\big\|\big(\sqrt{\pcmd{\pc}}\big)_{xx}(t)\big\|\Big)\big\|\pcpmd_{tx}(t)\big\|\notag\\
\leq&C\pc.\label{155.1}
\end{align}
Substituting \eqref{143.2}$\sim$\eqref{152.1} and \eqref{153.1}$\sim$\eqref{155.1} into \eqref{142.2}, we have
\begin{multline}\label{156.2}
\frac{d}{dt}\int_0^1\bigg[\frac{1}{2}\pcS{0}\big(\pcpmd_x\big)^2+\frac{1}{2}\big(\pcpdl_x\big)^2\bigg]dx+\big\|\pcpdl_x(t)\big\|^2\\
\leq\mu\big\|\pcpwnd_t(t)\big\|^2+C_\mu\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1^2+C\pc.
\end{multline}
Multiplying the equation \eqref{157.2} by $\pcpwnd_t$ and integrating the resultant equality over the domain $\Omega$, we get
\begin{equation}\label{161.1}
\int_0^1\pcmd{\pc}\big(\pcpwnd_t\big)^2dx-\int_0^1\frac{2}{3}\pcpwnd_{xx}\pcpwnd_tdx+\int_0^1\pcmd{0}\pcpwnd\pcpwnd_tdx=\int_0^1\mathcal{H}_2(t,x;\pc)\pcpwnd_tdx.
\end{equation}
Similarly, we can use the standard computations to deal with the each term in \eqref{161.1} as follows
\begin{equation}\label{161.2}
\int_0^1\pcmd{\pc}\big(\pcpwnd_t\big)^2dx\geq 2c\big\|\pcpwnd_t(t)\big\|^2,
\end{equation}
and
\begin{equation}\label{161.3}
-\int_0^1\frac{2}{3}\pcpwnd_{xx}\pcpwnd_tdx=\int_0^1\frac{2}{3}\pcpwnd_{x}\pcpwnd_{xt}dx=\frac{d}{dt}\int_0^1\frac{1}{3}\big(\pcpwnd_{x}\big)^2dx,
\end{equation}
and
\begin{align}
\int_0^1\pcmd{0}\pcpwnd\pcpwnd_tdx=&\frac{d}{dt}\int_0^1\frac{1}{2}\pcmd{0}\big(\pcpwnd\big)^2dx-\int_0^1\frac{1}{2}\pcmd{0}_t\big(\pcpwnd\big)^2dx\notag\\
\geq&\frac{d}{dt}\int_0^1\frac{1}{2}\pcmd{0}\big(\pcpwnd\big)^2dx-C\big\|\pcpwnd(t)\big\|^2,\label{161.4}
\end{align}
and
\begin{align}
\int_0^1\mathcal{H}_2(t,x;\pc)\pcpwnd_tdx\leq&c\big\|\pcpwnd_t(t)\big\|^2+C\|\mathcal{H}_2(t;\pc)\|^2\notag\\
\leq&c\big\|\pcpwnd_t(t)\big\|^2+C\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1^2+C\pc^2,\label{161.5}
\end{align}
where we have used the estimate \eqref{148.3-151.1} again in the last inequality of \eqref{161.5}. Substituting \eqref{161.2}$\sim$\eqref{161.5} into \eqref{161.1}, we have
\begin{equation}\label{161.6}
\frac{d}{dt}\int_0^1\bigg[\frac{1}{2}\pcmd{0}\big(\pcpwnd\big)^2+\frac{1}{3}\big(\pcpwnd_{x}\big)^2\bigg]dx+c\big\|\pcpwnd_t(t)\big\|^2\leq C\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1^2+C\pc^2.
\end{equation}
Adding \eqref{142.1}, \eqref{156.2} and \eqref{161.6} up, and letting $\mu$ small enough, we obtain
\begin{equation}\label{163.1}
\frac{d}{dt}\mathcal{E}^\pc(t)+\underbrace{\big\|\pcpdl(t)\big\|^2+\big\|\pcpdl_x(t)\big\|^2+\frac{c}{2}\big\|\pcpwnd_t(t)\big\|^2}_{\geq0}\leq C\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1^2+C\pc,
\end{equation}
where
\begin{equation}\label{163.2}
\mathcal{E}^\pc(t):=\int_0^1\bigg[\frac{1}{2}\pcS{0}\big(\pcpmd_x\big)^2+\frac{1}{2}\big(\pcpdl\big)^2+\frac{1}{2}\big(\pcpdl_x\big)^2+\frac{1}{2}\pcmd{0}\big(\pcpwnd\big)^2+\frac{1}{3}\big(\pcpwnd_{x}\big)^2\bigg]dx.
\end{equation}
From the estimate \eqref{138.1}, it is easy to check the following equivalent relation
\begin{equation}\label{163.3}
c\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1^2\leq\mathcal{E}^\pc(t)\leq C\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1^2,\quad\forall t\in[0,\infty)
\end{equation}
by using the Poincar\'e inequality. Therefore, the inequality \eqref{163.1} implies
\begin{equation}\label{164.1}
\frac{d}{dt}\mathcal{E}^\pc(t)\leq 2\gamma_3\mathcal{E}^\pc(t)+C\pc,\quad\forall t\in[0,\infty),
\end{equation}
where the positive constant $\gamma_3$ is independent of $\pc$ and $t$. Applying the Gronwall inequality to \eqref{164.1}, we have
\begin{equation}\label{164.3}
\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1\leq Ce^{\gamma_3 t}\pc^{1/2},\quad\forall t\in[0,\infty)
\end{equation}
Combining \eqref{164.3} with the elliptic estimate $\|\pcpdws(t)\|_3\leq C\|\pcpmd(t)\|_1$, we get the desired estimate \eqref{gs164.5}.
Finally, for fixed $\pc\in(0,\delta_6)$, we define a time
\begin{equation}\label{165.1}
T_\pc:=-\frac{\ln\pc}{4\gamma_3}>0.
\end{equation}
For $t\leq T_\pc$, the estimate \eqref{gs164.5} yields that
\begin{equation}\label{165.2}
\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1+\big\|\pcpdws(t)\big\|_3\leq Ce^{\gamma_3 T_\pc}\pc^{1/2}=C\pc^{1/4}.
\end{equation}
For $t\geq T_\pc$, using the estimates \eqref{de}, \eqref{ss39.1a} and \eqref{0de}, we obtain
\begin{align}
&\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1+\big\|\pcpdws(t)\big\|_3\notag\\
\leq&\|(\pcmd{\pc}-\pcsmd{\pc},\pcdl{\pc}-\pcsdl{\pc},\pcwnd{\pc}-\pcswnd{\pc})(t)\|_1+\|(\pcdws{\pc}-\pcsdws{\pc})(t)\|_3\notag\\
&+\|(\pcsmd{\pc}-\pcsmd{0},\pcsdl{\pc}-\pcsdl{0},\pcswnd{\pc}-\pcswnd{0})\|_1+\|(\pcsdws{\pc}-\pcsdws{0})\|_3\notag\\
&+\|(\pcmd{0}-\pcsmd{0},\pcdl{0}-\pcsdl{0},\pcwnd{0}-\pcswnd{0})(t)\|_1+\|(\pcdws{0}-\pcsdws{0})(t)\|_3\notag\\
\leq&C\big(e^{-\gamma_2T_\pc}+\pc+e^{-\gamma_1T_\pc}\big)\notag\\
=&C\Big(\pc^{\frac{\gamma_2}{4\gamma_3}}+\pc+\pc^{\frac{\gamma_1}{4\gamma_3}}\Big)\notag\\
\leq&C\pc^{\gamma_4},\label{165.3}
\end{align}
where
\begin{equation}
\gamma_4:=\min\bigg\{\frac{\gamma_1}{4\gamma_3},\frac{\gamma_2}{4\gamma_3},\frac{1}{4}\bigg\}>0.
\end{equation}
Owing to \eqref{165.2} and \eqref{165.3}, we have
\begin{equation}\label{166.1}
\big\|\big(\pcpmd,\pcpdl,\pcpwnd\big)(t)\big\|_1+\big\|\pcpdws(t)\big\|_3\leq C\pc^{\gamma_4},\quad\forall t\in[0,\infty).
\end{equation}
Note that the right-side of \eqref{166.1} is independent of $t$, this immediately implies the estimate \eqref{gs1}.
\end{proof}
\section{Appendix}\label{A}
In this appendix, we study the unique solvability of the linear IBVP \eqref{a3.1}$\sim$\eqref{a3.3}.
Firstly, the parabolic equation \eqref{a3.1-3} with the initial condition $\hat{\wnd}(0,x)=\wnd_0(x)$ and the boundary condition \eqref{a3.3-3} has a unique solution $\hat{\wnd}\in\mathfrak{Y}_2([0,T])\cap H^1(0,T;H^1(\Omega))$ for given function $(\rmd,\dl)\in\big[\mathfrak{Y}_4([0,T])\cap H^2(0,T;H^1(\Omega))\big]\times\big[\mathfrak{Y}_3([0,T])\cap H^2(0,T;L^2(\Omega))\big]$. This fact is proved by the Galerkin method (see \cite{T88,Z90} for example).
Next, we only need to show the unique solvability of the following linear IBVP for given function $(\rmd,\dl,\wnd,\hat{\wnd})$, namely,
\begin{subequations}\label{a7.1}
\begin{numcases}{}
2\rmd\hat{\rmd}_t+\hat{\dl}_x=0, \label{a7.1-1}\\
\hat{\dl}_t+2S[\rmd^2,\dl,\wnd]\rmd\hat{\rmd}_x+\frac{2\dl}{\rmd^2}\hat{\dl}_x+\rmd^2\hat{\wnd}_x-\pc^2\rmd^2\Bigg(\frac{\hat{\rmd}_{xx}}{\rmd}\Bigg)_x=\rmd^2\dws_x-\dl, \label{a7.1-2}\\
\dws:=\Phi[\rmd^2], \qquad\forall t>0,\ \forall x\in\Omega:=(0,1),\label{a7.1-3}
\end{numcases}
\end{subequations}
with the initial condition
\begin{equation}\label{a7.2}
(\hat{\rmd},\hat{\dl})(0,x)=(\rmd_0,\dl_0)(x),
\end{equation}
and the boundary conditions
\begin{subequations}\label{a7.3}
\begin{gather}
\hat{\rmd}(t,0)=\rmd_{l},\qquad \hat{\rmd}(t,1)=\rmd_{r},\label{a7.3-1}\\
\hat{\rmd}_{xx}(t,0)=\hat{\rmd}_{xx}(t,1)=0.\label{a7.3-2}
\end{gather}
\end{subequations}
To this end, performing the procedure $\pd{x}{}\eqref{a7.1-2}/(-2\rmd)$ and inserting the transformation $U(t,x):=\hat{\rmd}(t,x)-\bar{w}(x)$ into the resultant system, where $\bar{w}(x):=w_l(1-x)+w_rx$, we can equivalently reduce the IBVP \eqref{a7.1}$\sim$\eqref{a7.3} to the IBVP of a fourth order wave equation satisfied by $U$,
\begin{gather}
U_{tt}+b_0\pd{x}{}U_t+b_1U_t+b_2U_x+b_3U_{xx}+a\pd{x}{4}U=f,\label{a8.1}\\
U(0,x)=\rmd_0(x)-\bar{w}(x),\quad U_t(0,x)=-\frac{\dl_{0x}}{2\rmd_0}(x),\label{a8.2}\\
U(t,0)=U(t,1)=U_{xx}(t,0)=U_{xx}(t,1)=0,\label{a8.3}
\end{gather}
where
\begin{align}
&b_0:=\frac{2j}{w^2},\qquad b_1:=\frac{1}{w}\bigg[\bigg(\frac{2j}{w}\bigg)_x+w_t\bigg],\qquad b_2:=-\frac{1}{w}\bigg[\bigg(\wnd-\frac{j^2}{w^4}\bigg)w\bigg]_x,\notag\\
&b_3:=-\bigg[\bigg(\wnd-\frac{j^2}{w^4}\bigg)+\frac{\pc^2}{2}\frac{w_{xx}}{w}\bigg],\qquad a:=\frac{\pc^2}{2},\notag\\
&f:=-\frac{1}{2w}\big(w^2\dws_x-j-w^2\hat{\wnd}_x\big)_x+\frac{1}{w}\bigg[\bigg(\wnd-\frac{j^2}{w^4}\bigg)w\bigg]_x\bar{w}_x.\label{a8.4}
\end{align}
Applying the Lemma A.1 (\cite{NS08}, P870) to the linear IBVP \eqref{a8.1}$\sim$\eqref{a8.3}, we see that this problem has a unique solution $U\in\mathfrak{Y}_4([0,T])$.
We proceed to construct the solution $(\hat{\rmd},\hat{\dl})$ to the IBVP \eqref{a7.1}$\sim$\eqref{a7.3} from $U$ as follows,
\begin{subequations}
\begin{gather}
\hat{\rmd}(t,x):=U(t,x)+\bar{w}(x),\label{a9.0}\\
\hat{j}(t,x):=-\int_0^x2w\hat{w}_t(t,y)dy+\hat{j}(t,0),\label{a9.1}\\
\hat{j}(t,0):=\int_0^t\bigg[-2\bigg(\wnd-\frac{j^2}{w^4}\bigg)w\hat{w}_x+\frac{4j}{w}\hat{w}_t-w^2\hat{\wnd}_x\notag\\
\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad\qquad+\pc^2w^2\bigg(\frac{\hat{w}_{xx}}{w}\bigg)_x+w^2\dws_x-j\bigg](\tau,0)d\tau+j_0(0).\label{a9.2}\notag
\end{gather}
\end{subequations}
By the standard argument (see \cite{NS08,NS09} for example), we can easily see that the function $(\hat{\rmd},\hat{\dl})\in\big[\mathfrak{Y}_4([0,T])\cap H^2(0,T;H^1(\Omega))\big]\times\big[\mathfrak{Y}_3([0,T])\cap H^2(0,T;L^2(\Omega))\big]$ is a desired solution to the linear IBVP \eqref{a7.1}$\sim$\eqref{a7.3}.
\section*{Acknowledgements}
The research of KJZ was supported in part by NSFC (No.11371082) and the Fundamental Research Funds for the Central Universities (No.111065201).
| 0e2a992fdcc23e00c04aa575eaaca1ef375099c7 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\subsection{The Adjoint Method} \label{sec:adjointM}
In this section, we discuss the details of the Adjoint Method to obtain the gradient of the SI problem (\ref{eq:opt}) that we formulated in Section \ref{sec:probDef}. Particularly, the Lagrangian function of this constrained optimization problem is given as
$$ \ccalL(c,s,w) = \ccalJ(c,s) + \inprod{w, \ccalM(c,s)} , $$
where $w \in V''$ is the adjoint variable. From reflexivity of the Hilbert space $V$, we get $V'' = V$. Then, referring to the definition of the AD model \eqref{eq:ADmodel}, we have
\begin{eqnarray*}
\inprod{w, \ccalM(c,s)} &=& \inprod{w, Ac - \ell(s)}_{V'' \times V'} \\
&=& \inprod{Ac - \ell(s), w}_{V' \times V} = a(c,w) - \ell(w;s) .
\end{eqnarray*}
Thus, we can rewrite the Lagrangian as
\begin{equation} \label{eq:Lag}
\ccalL(c,s,w) = \ccalJ(c,s) + a(c,w) - \ell(w;s) ,
\end{equation}
where $w \in V$ is the adjoint variable.
In what follows, we use the notion of a G\^ateaux derivative to differentiate the Lagrangian (\ref{eq:Lag}); see, e.g., \cite[sec. 9.4]{IFA1998R}.
\begin{definition}[G\^ateaux derivative] \label{def:Gateaux}
A functional $\ccalT: V \to \reals$ on a normed space $V$ is G\^ateaux-differentiable at $c \in V$ if there exists an operator $D_c\ccalT: V \to V'$ defined by
$$ \inprod{D_c \ccalT,h} \triangleq \inprod{ \ccalT'_c,h} \triangleq \frac{d}{d \epsilon} [ \ccalT (c+ \epsilon h)] \Big|_{\epsilon = 0} , $$
for all $h \in V$. We use the two notations $\inprod{D_c \ccalT,h}$ and $\inprod{ \ccalT'_c,h}$ interchangeably whenever one of them is clearer.
\end{definition}
The Adjoint Method consists of the following three steps that yield an organized procedure for the calculation of the desired gradient; see, e.g., \cite[sec. 4]{SIPEIUAM2003ANG}.
First, in order to satisfy the AD constraint in the SI problem \eqref{eq:opt}, we set the G\^ateaux derivative of the Lagrangian \eqref{eq:Lag} with respect to the adjoint variable $w$ and in an arbitrary direction $v$ equal to zero. The bilinear form $a(c,w)$ and the functional $\ell(w;s)$ are the terms in the Lagrangian that contain $w$.
G\^ateaux differentiating $a(c,w)$ with respect to $w$ we get
\begin{eqnarray*}
\inprod{D_w a(c,w),v} = \frac{d}{d \epsilon} a(c,w + \epsilon v) \Big|_{\epsilon = 0} = a(c,v) ,
\end{eqnarray*}
where we have used linearity of the bilinear operator $a(c, w)$ in each argument.
Similarly, G\^ateaux differentiating $\ell(w;s)$ with respect to $w$ we get $\inprod{D_w \ell (w;s), v} = \ell(v;s)$.
Therefore, the first equation of the Adjoint Method is given as
\begin{equation} \label{eq:AdM1}
\inprod{\ccalL'_w,v} = a(c,v) - \ell(v;s) = 0, \ \forall v \in V .
\end{equation}
Note that this equation is identical to the VBVP \eqref{eq:VBVP} and for the function $c$ satisfying this equation, i.e., $c = \ccalF(s)$, we get $\ccalL(c,s,w) = \bar{\ccalJ}(s)$. Consequently, we can differentiate the Lagrangian \eqref{eq:Lag} to get the desired derivative $\bar{\ccalJ}'_s$.
Since $c = \ccalF(s)$, in order to calculate $D_s \ccalL(c,s,w)$ we need the derivative $\ccalF'_s$. We can avoid this calculation by setting the G\^ateaux derivative of the Lagrangian \eqref{eq:Lag} with respect to the concentration $c$ equal to zero for any arbitrary direction $h$. The two terms containing $c$ are the objective functional $\ccalJ(c,s)$ and the bilinear form $a(c,w)$.
From Definition \ref{def:Gateaux}, the G\^ateaux derivative of $\ccalJ(c,s)$ with respect to $c$ can be calculated explicitly using equation \eqref{eq:obj} as
\begin{eqnarray} \label{eq:gradJc}
\inprod{\ccalJ'_c, h} = \int_{\Omega} h \, (c - c^m) \ \chi_E \ d\Omega.
\end{eqnarray}
Moreover, similar to the previous case the G\^ateaux derivative of the bilinear form $a(c, w)$ with respect to $c$ is given by
$ \inprod{D_c a(c,w), h} = a(h,w) = a^*(w,h) , $
where $a^*(w,h)$ is the adjoint operator of the bilinear form $a(h,w)$.
Therefore, the second equation of the Adjoint Method is given as
\begin{equation} \label{eq:AdM2}
\inprod{\ccalL'_c,h} = \inprod{\ccalJ'_c,h} + a^*(w,h) = 0, \ \forall h \in V .
\end{equation}
Because of the appearance of the adjoint operator, this equation is called the adjoint equation and the procedure of calculating the desired gradient is referred to as Adjoint Method. Given the concentration $c$ obtained from \eqref{eq:AdM1}, the solution of equation \eqref{eq:AdM2} yields the corresponding adjoint variable $w$.
From the definition of the Lagrangian \eqref{eq:Lag}, for the functions $c$ and $w$ satisfying equations \eqref{eq:AdM1} and \eqref{eq:AdM2}, we have $D_s \ccalL(c,s,w) = \bar{\ccalJ'_s}$. Thus, we can calculate the desired gradient of the objective functional $\bar{\ccalJ}(s)$ with respect to the source term $s$ in a given direction $q$ by G\^ateaux differentiating the Lagrangian \eqref{eq:Lag} as
$ \inprod{\ccalL'_s,q} = \inprod{\ccalJ'_s,q} - \inprod{\ell'_s(w;s),q} .$
Combining equations \eqref{eq:AdM1} and \eqref{eq:AdM2} with this equation, we summarize the Adjoint Method to calculate the gradient of $\bar{\ccalJ}(s)$ with respect to $s$ in a given direction $q$ as:
\begin{subequations} \label{eq:adjoint}
\begin{align}
& \text{AD-PDE:} \ \ \ \, a(c,v) - \ell(v;s) = 0 , \ \ \ \ \forall v \in V , \label{eq:adVBVP} \\
& \text{Adjoint Eq: } \inprod{\ccalJ'_c,h} + a^*(w,h) = 0 , \ \forall h \in V, \label{eq:adAdjoint} \\
& \text{Gradient:} \ \ \ \ \inprod{\ccalL'_s,q} = \inprod{\ccalJ'_s - \ell'_s(w;s),q} . \label{eq:adGrad}
\end{align}
\end{subequations}
\subsection{The Source Identification Problem} \label{sec:introSI}
The SI problem in AD transport systems is known as chemical plume tracing and odor localization in the robotic literature and has been investigated since the early $80$s. This task often entails three steps: detection, localization, and declaration, and most of the available algorithms focus on the localization stage \cite{ROL2008KR}. The algorithms differ depending on the dispersal mechanism, i.e., diffusion- or turbulence-dominated, and are specialized for the particular types of sensors used to take the measurements. They are often bio-inspired and try to mimic the behavior of different bacteria \cite{OMRN2002MND}, insects \cite{CRRCA2003RBSW}, or crabs \cite{BAASGTCP2012WVW}.
Generally, the main idea is to stay in the plume and move upwind, in the concentration ascent direction, or a combination of the two. In the literature, the former approach is called anemotaxis while the latter is called chemotaxis \cite{ROL2008KR}.
The authors in \cite{OMRN2002MND} propose a controller that combines anemotaxis and chemotaxis to localize a source. This method is compared to our algorithm in Section \ref{sec:sim}.
Arguing that gradient based methods can get trapped in local optima and plateaus, \cite{BREM2004DSR} proposes a biased random walk strategy for a robotic swarm to localize multiple point sources.
The authors in \cite{DRACPT2005ZSS} propose fluxotaxis which uses the mass conservation principle to trace a chemical plume using a robotic swarm.
In a more recent work \cite{DFOSLA2015SAPM}, a group of mobile agents are controlled to stay in a formation centered in the plume while they move upwind and localize an ethanol source. To localize multiple sources, the authors of \cite{VCPMOSL2013CMG} construct a statistical model of the discovered sources allowing the robots to find the next source by subtracting the effect of the previous ones.
When the transport phenomenon is turbulent, disconnected and non-smooth concentration patches appear in the medium. In this case, gradient-based approaches are often significantly inaccurate. The authors in \cite{infotaxis2007VVS} propose a gradient-less search strategy, called infotaxis, that maximizes the expected rate of local information gain. The authors in \cite{MSSLTM2016HHS} extend the infotaxis strategy to a multi-agent system. This approach shows a behavior that resembles that of a moth, i.e., casting and zigzagging, which amounts to an extensive exploration of the domain and can be energy-inefficient for a mobile sensor.
The above heuristic approaches to SI are often successful in practice but they also suffer from various drawbacks: First, they do not offer a systematic approach that can handle the localization task under a wide range of conditions. Instead, they are specialized for specific scenarios and sensors.
Second, these methods often can only localize a single point source or at best multiple point sources and provide no information about the intensity of the sources. Moreover, they declare localization when the robots physically reach the source while in fact it might be unsafe to approach the source in some applications.
Finally, often these heuristics are proposed for convex environments and they do not handle obstacles and non-convex domains easily.
These limitations can be addressed if the underlying physics is properly incorporated in the formulation, which leads to model-based SI methods. \footnote{Note that some of the heuristics above selectively utilize physical principles but our intention here is a dedicated, systematic formulation.}
These model-based SI methods are a special class of Inverse Problems (IPs) which have been studied for a long time; see, e.g., \cite{DIP2010H}. Methods to solve IPs rely on a mathematical model of the underlying transport phenomenon which often is a Partial Differential Equation (PDE) and, in the special case of SI problems, it is linear in the unknown source term.
The literature on model-based SI problems can be classified in different ways based on the state of the problem, the number of sources, and their shape.
Generally, transient transport phenomena are more challenging compared to the steady-state ones, but time-dependent measurements are more informative.
The localization of a single point source in steady-state is considered in \cite{SLSDENAD2005MGK} for a semi-infinite domain, whereas the authors in \cite{MBSTSLP2000AS} address the problem for transient transport relying on the \textit{a priori} knowledge of the possible point source locations.
SI in the presence of multiple point sources is considered using optimization-based methods.
For instance, the work in \cite{MSDLADPUWSN2009WSK} addresses the detection and localization of multiple such sources using a wireless sensor network.
More general problems that involve sources of arbitrary shapes in arbitrary domains are typically solved numerically using, e.g., the Finite Element Method (FEM).
Discretization of a steady-state PDE using the FEM, leads to a linear time-invariant system where the source term acts as a control input. From this perspective, the IP is similar to the problem of input reconstruction. However, one of the main assumptions in this problem is that the number of observations is no less than the dimension of the unknown input \cite{SEUIRRHSMO2012Z}. This assumption is violated in IPs which are typically ill-posed. To resolve this issue, regularization techniques can be used \cite{DIP2010H,meC4}.
The authors in \cite{VFEMSICDT2003ABGLW} use the FEM along with total variation regularization to solve the SI problem.
Similarly, in our previous work \cite{meC1}, we proposed the Reweighted Debiased $\ell_1$ algorithm, which is an iterative sparse recovery approach to the SI problem. Despite generality, numerical methods such as FEM become computationally demanding as the size of the domain grows.
Furthermore, unlike heuristic methods, none of the model-based SI methods discussed above, rely on robots to collect the measurements that are needed to solve the problem.
\subsection{Active Sensing} \label{sec:introOSP}
Optimal measurement collection has been long studied in the robotics literature to solve state estimation problems.
Given a probabilistic model of the measurement noise, information-theoretic indices, e.g., covariance \cite{meJ2}, Fisher Information Matrix (FIM) \cite{EEDI2016MSMM}, different notions of entropy \cite{PFBIAS2010RH}, mutual information \cite{MIPPASEP2014CLD,DASSSMRN2015ALP}, and information divergence \cite{SMPFT2011AL}, have been used for general robotic planning.
For example, given an information distribution, the authors in \cite{EEDI2016MSMM} propose an optimal controller to navigate the robot through an ergodic path. We investigate the performance of this planning method for SI in Section \ref{sec:sim}.
A common predicament in applying some of these methods for SI is the need for the posterior distribution of the unknown source parameters. Obtaining this distribution for SI problems requires solving stochastic IPs and is computationally expensive, see, e.g., \cite{SRM2010FMWB}. This makes the application of optimality indices that require calculation of the expected information gain, e.g., entropy, mutual information, and information divergence, intractable.
Typically in SI problems, the amount of information provided by a measurement depends on the value of the unknown parameters in addition to the measurement location. A common approach to address this point is to combine the path planning for optimal measurement collection with the solution of the SI problem in a feedback loop which leads to Active SI methods. To solve the planning problem, scalar measures of the FIM can be used, as for state estimation; see, e.g., \cite[Ch. 2]{OMMDPSI2004D}.
The difference is that in SI the unknown parameters cannot be obtained in closed form by a filter update, but instead they are obtained by the solution of an IP that is much more difficult to solve.
Specifically, the work in \cite{ASIGRPE2005CR} presents trajectory planning for an autonomous robot, utilizing the trace of the FIM, to identify parameters of a transient Advection-Diffusion model under an instantaneous gas release in an infinite domain.
Similarly, \cite{OSMSNDLSCS2008PU, OMSPPEDPS2008TC} propose continuous-time optimal control methods that utilize the determinant of the FIM for trajectory planning for IPs with a few unknowns; \cite{OSMSNDLSCS2008PU} considers the SI problem in transient state under the assumption that the noisy measurements are taken continuously, while \cite{OMSPPEDPS2008TC} is an extension of \cite{OSMSNDLSCS2008PU} for general IPs.
In a different approach, the authors of \cite{DPSUE2009BLTT} propose an adaptive SI algorithm to localize a single point source emphasizing on path planning in unknown, possibly non-convex, environments.
Common in the above literature on ASI is that the proposed methods avoid the solution of complex SI problems by either assuming very simple mathematical models for the SI problem that can be efficiently solved, or by assuming that the solution of the SI problem is provided \textit{a priori} and the goal of planning is to collect measurements that are then examined to find a source term that matches those measurements.
Therefore, these methods do not apply to general SI problems and for this reason they have also not been demonstrated in practice. To the best of our knowledge, our work is the first ASI method that considers active measurement collection for realistic SI problems.
\subsection{Proposed Method}
In this paper, we consider the problem of Active Source Identification in Advection-Diffusion (AD) transport systems in steady-state. As most path planning methods for state estimation, we propose a method that combines SI and path planning in a feedback loop. The difference is that here the estimation problem is not solved by a closed form filter update, but instead it requires the solution of a complex PDE-constrained optimization problem.
Particularly, given a set of noisy measurements, we formulate the SI problem as a variational regularized least squares optimization problem subject to the AD-PDE.
To obtain a tractable solution to this problem, we employ Proper Orthogonal Decomposition \cite{OFROMPOD2007A} to approximate the concentration field using a set of optimal basis functions.
Moreover, we model the source term using nonlinear basis functions, which decreases the dimension of the parameter space significantly, although at the expense of introducing nonlinearity.
Using these parameterizations, we approximate the functional formulation of the SI problem with a low dimensional, nonlinear, constrained optimization problem, which we solve iteratively utilizing the gradient and Hessian information that we explicitly provide. To initialize this nonlinear optimization problem, we rely on the point-source Sensitivity Analysis of the SI objective function \cite{SSASLSSLS2013SA}.
Assuming a small number of measurements are available to initialize the identification process, we determine a sequence of waypoints from where a mobile robot sensor can acquire further measurements by formulating a path planning problem that maximizes the minimum eigenvalue of the FIM of the unknown source parameters with respect to the noisy concentration measurements.
The integrated algorithm, alternates between the solution of the SI and path planning problems. In particular, with every new measurement the solution of the SI problem produces a new source estimate, which is used in the path planning problem to determine a new location from where a new measurement should be taken, and the process repeats.
By appropriately decomposing the domain, we show that the proposed algorithm can identify multiple sources in complex AD systems that live in non-convex environments.
\subsection{Contributions}
To the best of our knowledge, this is the first model-based ASI framework that has been successfully demonstrated in practice. As discussed in Section \ref{sec:introSI}, existing literature on ASI includes either model-free bio-inspired heuristic methods or model-based approaches that employ simplifying assumptions, e.g., point sources in infinite domains, to mitigate the complexity of solving real SI problems. Compared to the heuristic approaches, i.e., chemotaxis, amenotaxis, fluxotaxis, and infotaxis, our proposed model-based SI method combines all these bio-inspired behaviors systematically in a general identification framework. Specifically, the concentration readings are explicitly modeled in the least squares objective while the gradient information (chemotaxis), velocity information (anemotaxis), and the first principals (fluxotaxis) are rigorously encapsulated in the AD-PDE. Finally, the information content of the measurements (infotaxis) are incorporated in the solution of the planning problem. On the other hand, compared to model-based approaches that rely on oversimplified models, our method can solve more general and realistic problems. We have shown that our method outperforms existing approaches both in simulation and experimentally.
Like the path planning methods for state estimation, our method combines SI and path planning in a feedback loop. However, as discussed in Section \ref{sec:introOSP}, the difference is that unlike the estimation problem that is solved by a closed-form filter update, the SI problem requires the solution of a complex PDE-constrained optimization problem. The key ideas that enable a tractable solution to this problem are: (a) a suitable integration of model order reduction, point-source Sensitivity Analysis, and domain decomposition methods, (b) a nonlinear representation of the source term that reduces the dimension of the parameter space, and (c) an information theoretic metric to measure the value of measurements for identifying unknown source parameters. The result is a set of techniques, insights, and methodological advancements that show how to efficiently design a model-based SI method that can be implemented onboard robots. Form a technical standpoint, the proposed framework bridges the gap in the rich but disconnected literature on source localization and active sensing that was discussed before; see Figure \ref{fig:literature}.
\begin{figure}[t!]
\centering
\includegraphics[width=0.5\textwidth]{main.png}
\caption{A schematic representation of the relevant literature.} \label{fig:literature}
\end{figure}
A significant contribution of this work is that it is the first to demonstrate applicability of robotic model-based SI methods to real-world problems. Real-world SI problems present major practical challenges related to modeling and estimation of the flow properties, which serve as the data in the AD-PDE, instability of the AD-PDE itself, and the uncertainties that are present in the parameters and boundary conditions; see the discussion in Section \ref{sec:discussion}.
We show that our algorithm is robust to uncertainties and performs well despite various simplifying assumptions made to model the real world; see \cite{TMCFD1993W,TD2002RW} for more details about these assumptions.
A preliminary version of this work can be found in the conference paper \cite{meC2}. The conference version only discusses the SI problem in convex domains and many theoretical details are absent due to space limitations; these details are included in Section \ref{sec:SI}.
In addition, here we discuss the integrated ASI algorithm that combines SI and planning and we also consider non-convex environments. Finally, we present experimental results that illustrate our method for real-world SI problems.
The rest of this paper is organized as follows. In Section \ref{sec:SI} we introduce the SI problem and discuss the proposed approach to solve it. Section \ref{sec:path} is devoted to the formulation and discussion of the path planning algorithm.
We discuss the integration of the SI and path planning algorithms along with their extension to non-convex domains in Section \ref{sec:autoSI}.
Section \ref{sec:sim} contains the numerical simulations and experimental results and finally Section \ref{sec:concl} concludes the paper.
\section{Introduction} \label{sec:intro}
\input{Intro}
\section{Source Identification} \label{sec:SI}
\input{SIP}
\section{Mobile Robot Path Planning} \label{sec:path}
\input{path}
\section{Active Source Identification\\ in Complex Domains} \label{sec:autoSI}
\input{autoSI}
\section{Results} \label{sec:sim}
\input{sim}
\section{Conclusions} \label{sec:concl}
\input{concl}
\section*{Acknowledgements}
The authors would like to thank Eric Stach, Yihui Feng, and Yan Zhang for their help with the design of the experimental setup.
\appendices
\section{Adjoint Method} \label{sec:adjointM}
\input{adjointM}
\section{Numerical Solution of the\\Source Identification Problem} \label{sec:SInumeric}
\input{SInumeric}
\section{Sequential Semi-Definite Programming\\for the Next Best Measurement Problem} \label{sec:SSDP}
\input{SSDP}
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\bibliographystyle{ieeetr}
\subsection{Numerical Simulations} \label{sec:ppComp}
In this section, we study the performance of the ASI Algorithm \ref{alg:autoSI} in a large non-convex domain $\Omega$.
We assume that the air flows into the domain through the sides, i.e., constant velocity inlet boundary conditions. Then, an in-house fluid dynamics code is utilized to simulate the steady-state velocity in the domain as depicted in Figure \ref{fig:Ex_BI-flow}, where a FE mesh with $n = 15034$ points is used.
\begin{figure}[t!]
\centering
\includegraphics[width=0.5\textwidth]{flow.pdf}
\caption{Steady-state flow pattern for the non-convex domain. } \label{fig:Ex_BI-flow}
\end{figure}
We consider Peclet numbers $Pe = 2.5, Pe = 25$, and $Pe = 250$. Moreover to reduce the order of the model, we utilize $R = 597$ snapshots with $\eta = 0.97$ corresponding to $N = 154, N = 183$, and $N = 205$ basis functions for each Peclet number, respectively.
In our first simulation study, we compare the planning method presented in Section \ref{sec:path} to placement over a lattice and an ergodic placement method that uses the determinant of the FIM as the information metric.
Both of these planning approaches employ the algorithm developed in Section \ref{sec:SI} to solve the SI problem \eqref{eq:opt}
and, therefore, the purpose of this comparison is to showcase the relative performance of the planning method proposed here.
Specifically, given an information distribution, the ergodic planner aims at designing paths where the time spent by the robot at any region is proportional to the information at that region. This approach is proposed in \cite{MEDEDMAS2011MM} and used for active sensing in \cite{EEDI2016MSMM}. Specifically, we use the normalized determinant of the FIM \eqref{eq:FIM} as the information distribution in the ergodic planner, where we assign zero information value to the points on obstacles. We implement the controller proposed in \cite{MEDEDMAS2011MM} and, similar to the ASI Algorithm \ref{alg:autoSI}, we update the information distribution at every step as newer estimates of the source parameters become available.
To highlight the advantages of the proposed ASI framework against the heuristic methods discussed in Section \ref{sec:introSI}, we also compare our algorithm to a heuristic approach that drives the robot along the normalized concentration gradient ascent and upwind directions, as proposed in \cite{OMRN2002MND}. The robot uses the initial $\bbarm$ measurements to detect the plume and initializes its path from the highest measured concentration point. The velocity field is known exactly to the robot and the concentration gradient at each point is approximated by taking two additional measurements in orthogonal directions. Since the heuristic approach only provides a location estimate, it is compared to the other methods in terms of $e_{\text{loc}}$.
In the following simulations, we use $\bbarm = 28$ initial measurements for the ASI, ergodic, and heuristic methods and set the maximum number of steps to $m_{\max} = 42$. The lattice placement uses $m_{\max} = 42$ measurements obtained by sensors located on an equidistant grid. For the ergodic and heuristic methods, we use a first-order model for the dynamics of the robot.
The results are plotted in Figure \ref{fig:Ex_BI-LATvsEXP} where we average over $50$ randomly generated sources.
\begin{figure}[t!]
\centering
\includegraphics[width=0.5\textwidth]{ExBI.pdf}
\caption{Comparison of the ASI Algorithm \ref{alg:autoSI} with lattice and ergodic placements, and heuristic approach for three Peclet numbers. The figure depicts the average uncovered, false detection, intensity, and localization errors, respectively. The standard deviation bars and scatter plots are overlaid on the error bars. The success percents are also given on top of the first subfigure.} \label{fig:Ex_BI-LATvsEXP}
\end{figure}
It can be seen that the proposed planning Algorithm \ref{alg:pathPlan} outperforms the lattice method in all cases and performs more consistently in terms of standard deviation. Particularly, the success rate of the ASI algorithm, i.e., the number of instances that the algorithm finds an overlapping source estimate, is considerably higher for all three Peclet numbers.
Note that the average false detection error of the lattice placement for $Pe = 250$ is smaller than the ASI algorithm but since this method often fails to find an overlapping source estimate for this Peclet number, $e_{\text{fd}}$ only indicates that the falsely detected sources have smaller volumes than the true sources on average.
The performance of the ergodic approach is close to the proposed planning method since it uses a similar optimality index to collect the measurements. Note that as the domain becomes larger, the computation of the information distribution required by this approach becomes expensive rendering this planning method intractable.
Furthermore, since the performance of the ergodic method depends on the combination of exploration and exploitation \cite{EEDI2016MSMM}, we allow the robot to travel through obstacles and take a measurement every three steps so that it can reach more informative regions of the domain more often.
To the contrary, the behavior of the proposed ASI method indicates that given an initial set of measurements, necessary to detect the unknown sources, the most informative measurements are obtained close to the location of sources as opposed to points farther away. Therefore, the better performance of the ASI algorithm, i.e., its smaller and more consistent false detection error values, can be attributed to this fact.
Considering the last subfigure in Figure \ref{fig:Ex_BI-LATvsEXP}, we observe that the ASI algorithm provides more accurate localization for $Pe=2.5$ and $Pe=25$ but the heuristic method performs better for $Pe=250$.
The reason for this is that for very high Peclet numbers, for which advection is the primary means of transport, the reduced order AD model \eqref{eq:Model} becomes inaccurate resulting in poor localization for model-based methods; see Section \ref{sec:discussion}.
Nevertheless, the heuristic approach does not provide any information about the size or intensity of the source and in the case of multiple sources, it localizes at most one source or fails altogether.
In our second simulation study, we use the same settings as before and consider an AD transport with $Pe = 25$ and two sources, specifically, a circular source centered at $(2.5, 0.25)$ with radius of $0.08$ and intensity of $0.25$ and a rectangular source with parameters $\barbp = (0.2, 3.85, 3.95, 0.8, 0.95)$ creating the concentration filed given in Figure \ref{fig:ExBII_path}.
Since the two sources are not located in one convex domain, decomposing $\Omega$ into convex subdomains and following the procedure described in Section \ref{sec:autoSI} is necessary to recover both sources. We note that the ASI algorithm has no \textit{a priori} knowledge of the number of sources.
It solves the problem in $1951$sec in $22$ steps, which amounts to solving $22$ instances of the SI problem \eqref{eq:optFD}. Time required to solve the planning problem \eqref{eq:pathOptL} is negligible. The final error values are $e_{\text{un}} = 0.67$ and $e_{\text{fd}} = 0.61$ with SNR $= 19.02$ dB.
The waypoints of the robot are given in Figure \ref{fig:ExBII_path} by white stars. Note the accumulation of the measurements around the high concentration regions of the domain, i.e., the hot spots \cite{MIPPASEP2014CLD}.
Note also that to cover both sources simultaneously, the robot needs to move back and forth between them. We can minimize the travelled distance, by adding a penalty term in the planning problem \eqref{eq:pathOptL} to encourage more measurements before moving to the next source but this would be suboptimal from an information perspective. A more viable option is to use multiple robots, which is part of our future work.
The result of the SA initialization Algorithm \ref{alg:init} and the final solution with the true source overlaid on it are plotted in Figure \ref{fig:ExBII_source}.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.5\textwidth}
\includegraphics[width=\textwidth]{path1.pdf}
\caption{waypoints of the mobile robot} \label{fig:ExBII_path}
\end{subfigure}
\quad
\begin{subfigure}[b]{0.5\textwidth}
\includegraphics[width=\textwidth]{source.pdf}
\caption{estimated source} \label{fig:ExBII_source}
\end{subfigure}
\caption{Waypoints of the mobile robot overlaid on the concentration field and the contours of the SA initialization and estimated source. The yellow stars in Figure \ref{fig:ExBII_path} indicate the initial measurements while the white stars show the sequence of waypoints. In Figure \ref{fig:ExBII_source}, the SA initialization (top) and the final solution (bottom) are shown, where the white squares depict the support of the initial estimate and the yellow lines delineate the support of the true source.} \label{fig:Ex_BII-twin}
\end{figure}
Note that the performance of the ASI Algorithm \ref{alg:autoSI} only depends on the dimensionless Peclet number. The units of the other quantities are arbitrary as long as they are selected consistently. Particularly, given a unit for concentration $c$, the unit for source term $s$ is defined as concentration per unit time; cf. AD-PDE \eqref{eq:BVP}. See the next section for a specific example.
\subsection{Experimental Results} \label{sec:experiment}
In this section, we demonstrate the performance of the proposed ASI Algorithm \ref{alg:autoSI} experimentally for the identification of an ethanol source in air. Particularly, we consider a non-convex domain with dimensions $2.2 \times 2.2 \times 0.4$m$^3$, cf. Figure \ref{fig:exp-dom}.
\begin{figure}[t!]
\centering
\includegraphics[width=0.4\textwidth]{domain.jpg}
\caption{Non-convex domain of the experiment. The velocity inlet, velocity outlet, and desired source are located in the lower-right, lower-left, and upper-right corners, respectively.} \label{fig:exp-dom}
\end{figure}
We connect a fan to the domain through a duct that blows air into domain with an axial speed of $0.9$m/s and a tangential speed of $0.2$m/s, creating a turbulent flow. We utilize \textsc{ANSYS-Fluent} to obtain the desired flow properties namely, the velocity and diffusivity fields. Since turbulence is a $3$D phenomenon, we build a $3$D mesh of the domain with $1.94 \times 10^6$ elements.
We note that determining the velocity and diffusivity fields for turbulent flow is non-trivial.
Turbulent flow is recognized with high Reynolds numbers and is characterized by severe fluctuations in the flow properties. These fluctuations enhance the mixing in the flow and facilitate the transport of the concentration. This enhanced mixing is often encapsulated in an effective turbulent diffusivity which is proportional to effective turbulent viscosity $\mu$ and the proportionality constant is the Schmidt number Sc. Then, the total diffusivity is the sum of laminar $\kappa_0$ and turbulent diffusivities \cite{TD2002RW}. Mathematically,
\begin{equation} \label{eq:diff}
\kappa = \kappa_0 + \frac{\mu}{\rho \, \text{Sc}} ,
\end{equation}
where $\rho \in \reals_+$ denotes the density of the transport medium.
We construct a $2$D discretized AD model using the FEM with $n=12121$ grid points located on the plane of the robot concentration sensor at height of $0.27$m. Using the $2$D model instead of a $3$D model is an approximation since we ignore the transport in $x_3$-direction but considerably decreases the computational cost.
The turbulent flow properties at the nodes of the AD FE mesh are interpolated from the \textsc{ANSYS-Fluent} $3$D model and are given in Figure \ref{fig:exp-turbSol}. The corresponding Peclet number is Pe $= 213$.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.4\textwidth}
\includegraphics[width=\textwidth]{vel.pdf}
\caption{velocity field (m/s)} \label{fig:exp-vel}
\end{subfigure}
\quad
\begin{subfigure}[b]{0.4\textwidth}
\includegraphics[width=\textwidth]{diff.pdf}
\caption{diffusivity field (m$^2$/s)} \label{fig:exp-diff}
\end{subfigure}
\caption{Turbulent flow properties required for the solution AD-PDE. The fields are interpolated to the plane of concentration sensor located at a height of $0.27$m. Top figure shows the predicted velocity field generated by blowing air through the inlet using a fan. The bottom plot shows the corresponding total diffusivity field which is the sum of laminar and turbulent diffusivities, given by \eqref{eq:diff}, and lower bounded by $10^{-3}$m$^2$/s to stabilize the AD FE model.} \label{fig:exp-turbSol}
\end{figure}
Regarding the diffusivity field, the following points are relevant. For molecular diffusion of ethanol in air, $\kappa_0 \approx 1.1 \times 10^{-5}$m$^2$/s. Then, from Figure \ref{fig:exp-diff} it can be seen that the turbulent diffusivity is considerably larger than the laminar diffusivity. This contributes to the numerical stability of the AD model by decreasing the Peclet number.
To further increase the stability, we add an artificial diffusion lower bounding the total diffusivity \eqref{eq:diff} by $10^{-3}$m$^2$/s; see the discussion of Section \ref{sec:discussion} for more details.
Given the discretized model, we use $N = 900$ basis functions to construct the reduced model \eqref{eq:Model}.
The ethanol source is located at $x_1 = 1.8$m, $x_2 = 1.8$m, and $x_3 = 0.3$m, across from the velocity inlet and releases ethanol at a steady rate. To collect the measurements, we use a custom built differential drive mobile robot equipped with a \textsc{MiCS-5524} concentration sensor. In order to eliminate the effect of velocity field on the sensor readings, we place the sensor in a confined box and utilize an air pump to deliver air to the sensor with an approximately constant flow rate, cf. Figure \ref{fig:exp-robot}.
\begin{figure}[t!]
\centering
\includegraphics[width=0.2\textwidth]{robot-small.jpg}
\caption{Mobile robot used to collect the measurements: (i) the concentration sensor placed in a confined box to separate it from flow conditions, (ii) air pump, (iii) \textsc{OptiTrack} markers used for localization. The robot is remotely controlled by a computer via radio communication.} \label{fig:exp-robot}
\end{figure}
As mentioned in Section \ref{sec:introSI}, non-smooth concentration patches appear when the flow is turbulent. To ensure the detection of these patches, we need to allow enough instantaneous readings at each location. Furthermore, to minimize the effect of intervals of low concentration between the detections, we calculate the final concentration value as the average of the readings at the highest quartile. In the experiment, we record $1000$ instantaneous readings at each measurement location.
We perform the computations off-board and communicate the commands to the robot via radio communications. This allows us to use a very small robot minimizing the interference with the flow field. The localization needed for motion control of the robot, is provided by an \textsc{OptiTrack} motion capture system and a simple controller is implemented for line tracking. We utilize the \textsc{VisiLibity} toolbox to generate the geodesic path between each pair of waypoints given by the planning Algorithm \ref{alg:pathPlan}, taking into account the obstacle; see \cite{VisiLibity2008O}.
The robot collects $\bbarm = 16$ initial measurements; these measurements are shown in Figure \ref{fig:exp-meas}. Note that the pattern of readings are in agreement with the model prediction verifying the overall accuracy of the numerical solutions of the flow and the AD-PDE \eqref{eq:BVP}. Figure \ref{fig:exp-place} shows the waypoints where we set the maximum number of measurements to $m_{\max} = 25$.
\begin{figure}[t]
\centering
\begin{subfigure}[b]{0.48\textwidth}
\includegraphics[width=\textwidth]{measurements.pdf}
\caption{initial measurements} \label{fig:exp-meas}
\end{subfigure}
\quad
\begin{subfigure}[b]{0.4\textwidth}
\includegraphics[width=\textwidth]{placement.pdf}
\caption{waypoints} \label{fig:exp-place}
\end{subfigure}
\caption{Top figure shows the initial $\bbarm$ measurements overlaid on the concentration field predicted by AD-PDE \eqref{eq:BVP} for a hypothetical source located at the true location. Bottom figure shows the waypoints of the robot.} \label{fig:exp-}
\end{figure}
The final solution is $\bbp = (3140, 1.69, 1.76, 1.77, 1.85)$ resulting in a location error $e_{\text{loc}} = 0.03$. This small error is due to the different heights of the planes at which ethanol is released ($x_3 = 0.30$m) and measurements are taken ($x_3 = 0.27$m). This causes the peak to be somewhat displaced downstream. The intensity of the source is predicted to be approximately $3 \times 10^3$ppm/s for the estimated source area of $5.6 \times 10^{-3}$m$^2$.
The video of the SI process is given in \cite{meJ1_video}. It can be seen from the video that the estimation of the source location approaches the true value immediately after the initial measurements are collected and the rest of the measurements correct the solution for the newer information that becomes available. Particularly, as the measurements get closer to the source, the intensity of the estimated source spikes to account for much larger observed concentration measurements.
\subsection{Discussion} \label{sec:discussion}
An important predicament in application of model-based ASI Algorithm \ref{alg:autoSI} is handling advection-dominated problems. Given a transport medium, e.g., air, high advection translates to turbulent flow which is non-trivial to model and an active area of research. Currently viable approaches are based on the Reynolds-Averaged Navier-Stokes (RANS) models that provide time-averaged properties. These models often suffice for engineering applications but major assumptions used in their derivation, technicalities pertaining to mesh generation and boundary layer treatment, and oftentimes conflicting predictions from different models affect their predictive ability so that additional experimental studies may be necessary to validate them; see \cite{TMCFD1993W} for more details.
On the other hand, advection-dominated AD models are also challenging and an active area of research. The reason is usually the presence of numerical instabilities.
In the previous section, we added a constant artificial diffusion which is a common practice in the relevant literature \cite{LRSROMBE2017BBSK}. Nevertheless, this might in general lead to forward solutions that are inconsistent with the solution of the original AD-PDE \cite{FEMFP2003DH}.
Numerous more advanced stabilization techniques exist that artificially introduce diffusion in a consistent manner; see e.g. \cite{SUPGFCDF1982BH}. Note that very high Peclet numbers, e.g., $Pe \approx 10^3$, are reported in the literature for the forward solution of the AD-PDE \eqref{eq:BVP} but solving the Inverse Problem using the AD model is considerably more challenging.
Specifically, instability of the AD model adversely affects the POD method, the SA initialization, and consequently the nonlinear optimization problems \eqref{eq:opt} and \eqref{eq:pathOptL}. After extensive simulation and experimental studies, we have observed that our method works well for Peclet numbers up to approximately $Pe \approx 250$.
While here we employ the standard Galerkin scheme for simplicity, more sophisticated FEM could be employed to improve this bound.
Note that an important feature of the ASI Algorithm \ref{alg:autoSI} is that it is highly modular, meaning that different components, i.e., the formulation of the SI problem \eqref{eq:opt}, planning problem \eqref{eq:pathOptL}, model reduction Algorithm \ref{alg:POD}, and SA initialization Algorithm \ref{alg:init}, can be independently improved for better SI performance. For instance in Section \ref{sec:ppComp} we replaced the planning module with the ergodic placement for the purpose of comparison.
Finally, for very low Peclet numbers, i.e., diffusion dominated cases, high concentration regions are local, thus it is possible for the SA method to miss some sources if there are no measurements close enough to those sources. Moreover, in the case of multiple sources, if the intensities differ considerably, the SA technique typically detects the high intensity ones.
In these cases, using higher numbers of initial measurements $\bbarm$ and tuning the thresholding parameter $\alpha$ in the SA Algorithm \ref{alg:init} can improve the initialization. We also note that, the SA initialization can be combined with possible prior knowledge about the sources to improve the performance of the proposed ASI algorithm.
\subsection{First Order Information} \label{sec:Grad}
In Appendix \ref{sec:adjointM} we discussed the Adjoint Method to obtain the gradient of the SI problem \eqref{eq:opt} when the variable $s$ is a function that lives in the infinite dimensional function space $S$. Here, we employ the approximations $V_d$ and $S_d$ defined in Section \ref{sec:finiteDim} to obtain a finite dimensional counterpart of the Adjoint Method equations \eqref{eq:adjoint} that is needed to solve the finite dimensional SI problem \eqref{eq:optFD} numerically.
First, we substitute the finite dimensional representations \eqref{eq:finiteDim} into equation \eqref{eq:adVBVP} to get
\begin{align*}
& a(c_d,v_d) - \inprod{\ell(s_d),v_d} = 0, \ \forall v_d \in V_d, \\
& a( \sum_{k=1}^N c_k \psi_k , \sum_{i=1}^N v_i \psi_i ) - \inprod{\ell(s_d), \sum_{i=1}^N v_i \psi_i } = 0, \ \forall v_i \in \reals, \\
& \sum_{i=1}^N v_i \set{ \sum_{k=1}^N c_k a( \psi_k , \psi_i ) - \inprod{\ell(s_d), \psi_i } } = 0, \ \forall v_i \in \reals, \\
& \sum_{k=1}^N c_k a( \psi_k , \psi_i ) - \inprod{\ell(s_d), \psi_i } = 0 , \ \forall i \in \set{1, \dots, N} .
\end{align*}
Writing the equations for all $i \in \set{1, \dots, N}$ in matrix form, we obtain the following linear system of equations
\begin{equation} \label{eq:VBVPLS}
\bbA \bbc = \bbb (\bbp),
\end{equation}
where $\bbA \in \reals^{N \times N}$ and $\bbb$ is a fixed vector for a given $\bbp$. Using equation \eqref{eq:VBVPLS}, we define the finite dimensional model in equation (\ref{eq:optFD}) explicitly as
\begin{equation} \label{eq:Model}
\bbM(\bbc, \bbp) = \bbA \bbc - \bbb (\bbp) = \bb0 .
\end{equation}
As explained in Section \ref{sec:ADPDE}, the AD model \eqref{eq:VBVP} has a unique solution that translates to the invertibility of matrix $\bbA$ in \eqref{eq:Model}.
Similar to approximations \eqref{eq:finiteDim}, we can write $w_d = \bbpsi \, \bbw \text{ and } \, h_d = \bbpsi \, \bbh$ where $\bbw, \bbh \in \reals^N$. Substituting these definitions into the adjoint equation \eqref{eq:adAdjoint}, we get
\begin{align*}
& \inprod{\ccalJ'_c, \psi_i } + \sum_{k=1}^N w_k a^*( \psi_k , \psi_i ) = 0 , \ \forall i \in \set{1, \dots, N} ,
\end{align*}
where the derivative $\inprod{\ccalJ'_c, \cdot }$ is defined by equation \eqref{eq:gradJc}.
Again writing the equations for all $i \in \set{1, \dots, N}$ in matrix form, we obtain
\begin{equation} \label{eq:adjointLS}
\bbA^T \bbw = - \bbd,
\end{equation}
where the transpose sign appears in \eqref{eq:adjointLS} because of the adjoint operator in the equations.
Given values for the source parameters $\bbp$, the linear systems \eqref{eq:VBVPLS} and \eqref{eq:adjointLS} can be used to obtain the corresponding concentration $\bbc$ and adjoint variable $\bbw$. This information can then be used in \eqref{eq:adGrad} to calculate the desired gradient $\nabla_{\bbp} \bbarJ$ of the objective function $\bbarJ(\bbp) = \bar{\ccalJ}(s_d)$ with respect to $\bbp$.
In order to simplify the notation and without loss of generality, we assume a single source in a $2$D domain given by
$ s_d(\bbx) = \beta \, \phi(\bbx; \undbbx, \barbx) ,$
where $\undbbx = (\undx_1,\undx_2)$ and $\barbx = (\bbarx_1, \bbarx_2)$.
Substituting the approximations \eqref{eq:finiteDim} in the Lagrangian \eqref{eq:Lag}, we get
$ \ccalL(c_d,s_d,w_d) = \ccalJ(c_d,s_d) + a(c_d,w_d) - \ell(w_d;s_d) .$
To obtain the finite dimensional counterpart of equation \eqref{eq:adGrad}, we need to take the derivative of this Lagrangian with respect to the parameters $\bbp$ of the source term $s_d$. The terms that contain $s_d$ are $\ccalJ(c_d,s_d)$ and $\ell(w_d; s_d)$.
For the objective functional $\ccalJ(c_d,s_d)$ from equation \eqref{eq:obj}, the only part involving $s_d$ is the regularization term
$ \int_{\Omega} s_d \ d\Omega = \beta \ (\bbarx_1 - \undx_1) (\bbarx_2 - \undx_2) .$
From this expression we can calculate the derivatives of $\ccalJ(c_d,s_d)$ with respect to $\bbp$, e.g.,
$$ \frac{\partial \ccalJ}{\partial \undx_1} = - \tau \beta (\bbarx_2 - \undx_2) . $$
For the functional $\ell(w_d; s_d)$, substituting $s_d$ into the definition \eqref{eq:functional} we get
$ \ell(w_d;s_d) =
\int_{\undx_1}^{\bbarx_1} \int_{\undx_2}^{\bbarx_2} \beta \ w_d(\bbx) \ dx_2 \ dx_1 . $
The derivative with respect to $\beta$ is straightforward and for the other parameters we use the Leibniz rule, e.g.,
$$ \frac{\partial \ell}{\partial \undx_1} = - \beta \int_{\undx_2}^{\bbarx_2} w_d(\undx_1, x_2) \ dx_2 . $$
Then by equation \eqref{eq:adGrad}, combining the two derivatives for $\undx_1$ we get
$ \partial \bbarJ / \partial \undx_1 = \partial \ccalJ / \partial \undx_1 - \partial \ell / \partial \undx_1 .$
The other derivatives can be calculated exactly the same way and we get the following values for the desired gradient
\begin{align} \label{eq:grad}
& \frac{\partial \bbarJ}{\partial \beta} = \displaystyle \tau (\bbarx_1 - \undx_1) (\bbarx_2 - \undx_2) - \int_{\undx_1}^{\bbarx_1} \int_{\undx_2}^{\bbarx_2} w_d(\bbx) \ dx_2 \ dx_1 , \nonumber \\
& \frac{\partial \bbarJ}{\partial \undx_1} = \displaystyle - \tau \beta (\bbarx_2 - \undx_2) + \beta \int_{\undx_2}^{\bbarx_2} w_d(\undx_1, x_2) \ dx_2 , \nonumber \\
& \frac{\partial \bbarJ}{\partial \undx_2} = \displaystyle - \tau \beta (\bbarx_1 - \undx_1) + \beta \int_{\undx_1}^{\bbarx_1} w_d(x_1, \undx_2) \ dx_1 , \nonumber \\
& \frac{\partial \bbarJ}{\partial \bbarx_1} = \displaystyle \tau \beta (\bbarx_2 - \undx_2) - \beta \int_{\undx_2}^{\bbarx_2} w_d(\bbarx_1, x_2) \ dx_2 , \nonumber \\
& \frac{\partial \bbarJ}{\partial \bbarx_2} = \displaystyle \tau \beta (\bbarx_1 - \undx_1) - \beta \int_{\undx_1}^{\bbarx_1} w_d(x_1, \bbarx_2) \ dx_1 ,
\end{align}
where $\bbarJ(\bbp) = \bar{\ccalJ}(s_d)$ and $\bbp = (\beta, \undx_1, \undx_2, \bbarx_1, \bbarx_2)$.
The process for calculating the desired gradient $\nabla_{\bbp} \bbarJ$ given a set of values for the parameters $\bbp$ is described in Algorithm \ref{alg:Grad}.
\begin{algorithm}[t]
\caption{The Adjoint Method}
\label{alg:Grad}
\begin{algorithmic}[1]
\REQUIRE The vector of parameters $\bbp$ and the matrix $\bbA$;
\STATE Compute the r.h.s. vector $\bbb$ of equation \eqref{eq:VBVPLS};
\STATE Solve the linear system $\bbA \bbc = \bbb$ for coefficients $\bbc$;
\STATE Compute the r.h.s. vector $\bbd$ of equation \eqref{eq:adjointLS} using \eqref{eq:gradJc};
\STATE Solve the adjoint equation $\bbA^T \bbw = - \bbd$ for $\bbw$;
\STATE Compute the desired gradient $\nabla_{\bbp} \bbarJ$ using equation (\ref{eq:grad}).
\end{algorithmic}
\end{algorithm}
Note that if there are multiple sources, i.e., if $M > 1$, then we calculate the gradients for each basis function separately. This follows from the rule for differentiating sums. Moreover if $\Omega \subset \reals^3$, we can exactly follow the same steps to calculate the gradient.
\subsection{Second Order Information} \label{sec:secOrderInfo}
Including second order information in the optimization algorithm can make the solution of the SI problem \eqref{eq:optFD} more efficient and accurate. Such information can be in the form of the Hessian $\bbH = \nabla_{\bbp \bbp} \bbarJ$ of the objective function itself, or in the form of a Hessian-vector product $\bbH \bbv$, for some vector $\bbv$, that is used in the optimization algorithm; see, e.g., \cite[ch. 7]{NO2006NW}.
The procedure to calculate the Hessian-vector multiplication is an attractive choice for large-scale problems but we use it here since it provides an organized approach to incorporate the AD model \eqref{eq:Model} into the Hessian calculations.
Specifically, using the finite dimensional approximation of the Lagrangian (\ref{eq:Lag}) given as
\begin{equation} \label{eq:LagFiniteDim}
L(\bbc, \bbp, \bbw) = J(\bbc, \bbp) + \bbw^T \bbM(\bbc, \bbp) = J(\bbc, \bbp) + \bbw^T (\bbA \bbc - \bbb) ,
\end{equation}
we can devise a procedure to calculate the product $\bbH \bbv$ for a given vector $\bbv$.
The details of this derivation can be found in \cite{PAPO2006ABGHK} and it results in the following equation
\begin{equation} \label{eq:HessVec}
\bbH \, \bbv = \bbM_{\bbp}^T \, \bbh_4 + \nabla^2_{\bbp \bbc} L \ \bbh_1 + \nabla^2_{\bbp \bbp} L \ \bbv ,
\end{equation}
where the subscripts denote differentiation and the process to calculate the vectors $\bbh_1$ and $\bbh_4$ is explained in Algorithm \ref{alg:Hess}.
\begin{algorithm}[t]
\caption{Hessian-vector Multiplication}
\label{alg:Hess}
\begin{algorithmic}[1]
\REQUIRE The vector $\bbv$;
\REQUIRE The matrices $\bbA$, $\bbM_{\bbp}$, $\nabla^2_{\bbc \bbc} L$, and $\nabla^2_{\bbp \bbp} L$ from equations \eqref{eq:VBVPLS}, \eqref{eq:Mp}, \eqref{eq:Lcc}, and \eqref{eq:Lpp};
\STATE Compute $\bbh_2 = \bbM_{\bbp} \bbv$ using equation \eqref{eq:Mp};
\STATE Solve $\bbM_{\bbc} \bbh_1 = \bbh_2$ for $\bbh_1$ where $\bbM_{\bbc} = \bbA$;
\STATE Compute $\bbh_3 = \nabla^2_{\bbc \bbp} L \, \bbv + \nabla^2_{\bbc \bbc} L \ \bbh_1 $ using equations \eqref{eq:Lpc} and \eqref{eq:Lcc};
\STATE Solve $\bbM_{\bbc}^{T} \bbh_4 = - \bbh_3$ for $\bbh_4$;
\STATE Calculate $\bbH \, \bbv$ from equation \eqref{eq:HessVec}.
\end{algorithmic}
\end{algorithm}
In what follows, we discuss all the second order terms needed in Algorithm \ref{alg:Hess} starting with the derivative of the AD model \eqref{eq:Model} with respect to the parameters denoted by $\bbM_{\bbp}$.
Recalling equation \eqref{eq:VBVPLS} and using the Leibniz rule, row $i$ of matrix $\bbM_{\bbp} \in \reals^{N \times 5}$ is given as
\begin{align} \label{eq:Mp}
& \frac{\partial \bbM_i}{\partial \beta} = - \int_{\undx_1}^{\bbarx_1} \int_{\undx_2}^{\bbarx_2} \psi_i(\bbx) \ dx_2 \ dx_1 , \nonumber \\
& \frac{\partial \bbM_i}{\partial \undx_1} = \beta \int_{\undx_2}^{\bbarx_2} \psi_i(\undx_1, x_2) \ dx_2 , \nonumber \\
& \frac{\partial \bbM_i}{\partial \undx_2} = \beta \int_{\undx_1}^{\bbarx_1} \psi_i(x_1, \undx_2) \ dx_1 , \nonumber \\
& \frac{\partial \bbM_i}{\partial \bbarx_1} = - \beta \int_{\undx_2}^{\bbarx_2} \psi_i(\bbarx_1, x_2) \ dx_2 , \nonumber \\
& \frac{\partial \bbM_i}{\partial \bbarx_2} = - \beta \int_{\undx_1}^{\bbarx_1} \psi_i(x_1, \bbarx_2) \ dx_1 .
\end{align}
Using equation \eqref{eq:Model}, the derivative of the AD model with respect to $\bbc$ is given as
$ \bbM_{\bbc} = \bbA . $
In addition, since there are no terms containing the multiplication of the concentration and source parameters, $\bbc$ and $\bbp$, in $L(\bbc, \bbp, \bbw)$, we have
\begin{equation} \label{eq:Lpc}
\nabla^2_{\bbp \bbc} L = \nabla^2_{\bbc \bbp} L = \bb0 .
\end{equation}
Finally we need to calculate the second order derivatives of the Lagrangian with respect to each of $\bbc$ and $\bbp$. Note that from equation \eqref{eq:LagFiniteDim}, $\nabla^2_{\bbc \bbc} L = \nabla^2_{\bbc \bbc} J$ and the value $\inprod{\ccalJ'_c,\psi_i}$ is basically the directional derivative in the direction $\psi_i$ or the derivative $\partial J / \partial c_i$. Thus we can Gateaux differentiate equation \eqref{eq:gradJc} once more to get the element in row $i$ and column $j$ of $\nabla^2_{\bbc \bbc} L$ as
\begin{eqnarray} \label{eq:Lcc}
[\nabla^2_{\bbc \bbc} L]_{ij} = \int_{\Omega} \chi_E \, \psi_i \, \psi_j \ d\Omega ,
\end{eqnarray}
where $i, j \in \set{1, \dots, N}$. Note that this expression is independent of the parameters $\bbp$ and can be calculated offline.
In order to calculate $\nabla^2_{\bbp \bbp} L$ note that the terms $J(\bbc, \bbp)$ and $\bbw^T \bbb$ in the Lagrangian (\ref{eq:LagFiniteDim}) contribute to this derivative. The calculation for $\nabla^2_{\bbp \bbp} J$ can be done by differentiating the result of Section \ref{sec:Grad} for $\nabla_{\bbp} J$ once more to get
\begin{equation} \label{eq:Jpp}
\nabla^2_{\bbp \bbp} J = \tau \left[
\begin{array}{ccccc}
0 & \cdot & \cdot & \cdot & \cdot \\
- (\bbarx_2 - \undx_2) & 0 & \cdot & \cdot & \cdot \\
- (\bbarx_1 - \undx_1) & \beta & 0 & \cdot & \cdot \\
(\bbarx_2 - \undx_2) & 0 & - \beta & 0 & \cdot \\
(\bbarx_1 - \undx_1) & - \beta & 0 & \beta & 0
\end{array}
\right] .
\end{equation}
For the second term we have $\bbw^T \bbb = \ell(w_d; s_d) $, since the Lagrangians \eqref{eq:Lag} and \eqref{eq:LagFiniteDim} are equivalent.
Thus, we can differentiate the expression for $\nabla_{\bbp} \ell$ from Section \ref{sec:Grad} once more, using the Leibniz rule, to get $ \nabla^2_{\bbp \bbp} \ell$ as is shown in equation \eqref{eq:LppSec}.
\begin{figure*}[t!]
\small
\begin{align} \label{eq:LppSec}
& \nabla^2_{\bbp \bbp} \ell = \nonumber \\
& \left[
\begin{array}{ccccc}
0 & \cdot & \cdot & \cdot & \cdot \\
- \int_{\undx_2}^{\bbarx_2} w_d(\undx_1, x_2) \ dx_2 &
- \beta \int_{\undx_2}^{\bbarx_2} \frac{\partial w_d}{\partial \undx_1} (\undx_1, x_2) \ dx_2
& \cdot & \cdot & \cdot \\
- \int_{\undx_1}^{\bbarx_1} w_d(x_1, \undx_2) \ dx_1 & \beta w_d(\undx_1, \undx_2) &
- \beta \int_{\undx_1}^{\bbarx_1} \frac{\partial w_d}{\partial \undx_2} (x_1, \undx_2) \ dx_1
& \cdot & \cdot \\
\int_{\undx_2}^{\bbarx_2} w_d(\bbarx_1, x_2) \ dx_2 & 0 & - \beta w_d(\bbarx_1, \undx_2) &
\beta \int_{\undx_2}^{\bbarx_2} \frac{\partial w_d}{\partial \bbarx_1} (\bbarx_1, x_2) \ dx_2
& \cdot \\
\int_{\undx_1}^{\bbarx_1} w_d(x_1, \bbarx_2) \ dx_1 & - \beta w_d(\undx_1, \bbarx_2) & 0 & \beta w_d(\bbarx_1, \bbarx_2) &
\beta \int_{\undx_1}^{\bbarx_1} \frac{\partial w_d}{\partial \bbarx_2} (x_1, \bbarx_2) \ dx_1
\end{array}
\right]
\end{align}
\end{figure*}
\normalsize
Putting the two terms given by equations \eqref{eq:Jpp} and \eqref{eq:LppSec} together, we have
\begin{equation} \label{eq:Lpp}
\nabla^2_{\bbp \bbp} L = \nabla^2_{\bbp \bbp} J - \nabla^2_{\bbp \bbp} \ell .
\end{equation}
Notice that we basically have differentiated equation (\ref{eq:grad}) once more in this process.
The case of multiple sources only affects the terms $\bbM_{\bbp}$ and $\nabla^2_{\bbp \bbp} L$ given by equations (\ref{eq:Mp}) and (\ref{eq:Lpp}), respectively.
Since the source term $s_d$ defined in equation \eqref{eq:sourceFiniteDim} is the summation of nonlinear basis functions, for $\bbM_{\bbp}$ we need to append more columns using equation \eqref{eq:Mp} corresponding to each basis function. On the other hand, for $\nabla^2_{\bbp \bbp} L$ we have to add blocks of matrices given by equation \eqref{eq:Lpp} corresponding to each basis function to the diagonal of $\nabla^2_{\bbp \bbp} L$.
\subsection{Initialization} \label{sec:init}
Appropriate initialization is critical for the solution of nonlinear optimization problems, such as \eqref{eq:optFD}, since otherwise the solution can get trapped in undesirable local minima.
In this paper, we employ a result on the point-source Sensitivity Analysis (SA) of the SI cost functional, presented in \cite{SSASLSSLS2013SA}, for initialization of our method.
The idea is to determine the sensitivity of the objective functional $\ccalJ(c,s)$ to the appearance of a point source in $\Omega$, i.e., we calculate the derivative of the objective with respect to the point-source term.
The regions with highest sensitivity represent the potential areas where the support of the true source function $\bbars$ is nonzero.
Note that by linearity of the AD-PDE (\ref{eq:BVP}), we only need to consider the infinitesimal deviations of the point-source from zero for a source-free domain, i.e., we calculate the derivative for the constant source function $s=0$.
In \cite{SSASLSSLS2013SA} it is shown that the adjoint variable is a measure of the sensitivity of the cost functional to these infinitesimal changes. Thus given the set of measurements $E$ introduced in Section \ref{sec:probDef}, we can obtain an approximation to the source locations via a solution of the adjoint equation. Specifically, we solve
$ \bbA^T \barbw = -\barbd $
with
$ \bbard_i = \int_{\Omega} c^m \ \psi_i \ d\Omega $
for $i \in \set{1, \dots, N}$, to get the desired finite dimensional adjoint function as $\bbarw_d = \bbpsi \, \barbw$.
Then an approximate localization of the source is obtained through thresholding as
\begin{equation} \label{eq:init}
\hhatw_d (\bbx) \triangleq
\left\{
\begin{array}{ll}
\bbarw_d (\bbx) & \text{if } \bbarw_d (\bbx) \leq \alpha \, \bbarw^{\min}_d \\
0 & \text{o.w.}
\end{array}
\right.
\end{equation}
where $\bbarw^{\min}_d = \min_{\bbx \in \Omega} \bbarw_d (\bbx)$ and $\alpha \in (0,1)$.
The thresholding parameter $\alpha$ determines the size of the support of $\hhatw_d (\bbx)$ and thus, the number of compact regions that indicate candidate source locations.
In order to separate these compact regions, we utilize the Single Linkage Agglomerative Clustering (SLAC) algorithm; see, e.g., \cite{MSTSLCA1969GR}. Specifically, given the nodal values $\hhatbbw_d$ of $\hhatw_d (\bbx)$ over the FE-mesh, we cluster the nonzero nodal values into sets $C_k$ for $k \in \set{1, \dots, K}$.
Then, we initialize the SI problem \eqref{eq:optFD} by placing a basis function with a small area at the point with highest sensitivity, given by equation \eqref{eq:init}, in each cluster; see Algorithm \ref{alg:init} for details.
\begin{algorithm}[t]
\caption{Point-source Sensitivity Analysis}
\label{alg:init}
\begin{algorithmic}[1]
\REQUIRE The set of measurements $E$;
\REQUIRE The thresholding parameter $\alpha \in (0,1)$;
\STATE Compute the sensitivity function $\hhatw_d (\bbx)$ from equation \eqref{eq:init} and the set $Z = \set {\bbz_i \, | \, \hhatw_d (\bbz_i) \neq 0, 1 \leq i \leq n}$;
\STATE Divide the set of points $Z$ into $K$ clusters $C_k$ according to their distance using the SLAC algorithm;
\STATE For each cluster $C_k$, set the cluster center as
$$\barbz_k = \argmin_{\bbz_i \in C_k} \hhatw_d (\bbz_i) ;$$
\STATE Initialize the source term \eqref{eq:sourceFiniteDim} using bases $\phi_k(\bbx)$ with small areas centered at $\barbz_k$ and $\beta_k \propto \abs{ \hhatw_d (\barbz_k) }$ .
\end{algorithmic}
\end{algorithm}
\subsection{Advection-Diffusion Transport} \label{sec:ADPDE}
Let $\Omega \subset \reals^d$ denote the domain of interest ($1 \leq d \leq 3$), and assume the presence of sources is modeled by a nonnegative function, $s: \Omega \rightarrow \reals_+$.
\footnote{For the problem considered here, we assume that sources are strictly positive functions. In general, sources can also be negative in the case of sinks. Sinks can appear, e.g., in the presence of chemical reactions that consume a contaminant. The treatment of the problem in that case is similar.}
Let $c: \Omega \rightarrow \reals_+$ be the measurable quantity, such as concentration, generated by this source function. Moreover, let the velocity at which this quantity is transported via advection be $\bbu \in \reals^d$ and $\kappa \in \reals_+$ be the diffusivity of the medium. Under steady-state assumption and applying a zero-valued Dirichlet condition to the boundaries $\Gamma$ of the domain, we arrive at the following Boundary Value Problem (BVP) \cite[ch. 2]{STMTPE2005SJ}
\begin{subequations} \label{eq:BVP}
\begin{eqnarray}
- \nabla \cdot ( \kappa \nabla c ) + \nabla \cdot (c \, \bbu) - s = 0 &\text{ in }& \Omega, \label{eq:ADPDE} \\
c = 0 &\text{ on }& \Gamma. \label{eq:EBC}
\end{eqnarray}
\end{subequations}
We consider Dirichlet conditions for the sake of simplicity; more general boundary conditions can be considered without any additional complications \cite{IFA1998R}.
In order for the BVP (\ref{eq:BVP}) to have a solution we assume that $s \in L^2(\Omega)$, i.e., $s$ is square integrable over $\Omega$, and define the feasible set for the source term as
$ S = \set{ s \in L^2(\Omega) \ | \ s \geq 0 } $.
The BVP (\ref{eq:BVP}) can be equivalently represented in variational form as follows.
Consider the set $V \subset H^1_0(\Omega)$, i.e., the set of functions that themselves and their first weak derivatives are square integrable and have compact supports. Thus every $v \in V$ satisfies the boundary condition (\ref{eq:EBC}).
Multiplying equation (\ref{eq:ADPDE}) by the trial function $v \in V$, integrating over the domain, and using Green's theorem, we obtain the variational formulation of the Advection-Diffusion PDE as
\begin{equation} \label{eq:VBVP}
a(c,v) = \ell(v;s), \ \forall v \in V ,
\end{equation}
where $a: V \times V \to \reals$ is a non-symmetric continuous positive-definite bilinear form defined as
\begin{equation} \label{eq:binomial}
a(c,v) \triangleq \int_{\Omega} \kappa \nabla c \cdot \nabla v \ d\Omega + \int_{\Omega} v \, \bbu \cdot \nabla c \ d\Omega ,
\end{equation}
\normalsize
and $\ell(s): V \to \reals$ is a continuous linear functional defined as
\begin{equation} \label{eq:functional}
\ell(v;s) \triangleq \inprod{\ell(s),v} \triangleq \int_{\Omega} sv \ d\Omega ,
\end{equation}
where the notation $\inprod{\ell(s),v}$ indicates the operation of $\ell(s)$ on the function $v$.
Given $s \in S$, we define the linear functional $\ccalM(c;s): V \to \reals$ as
\begin{equation} \label{eq:ADmodel}
\ccalM(c;s) \triangleq Ac - \ell(s) ,
\end{equation}
where the operator $A: V \to V'$ is defined by $\inprod{Ac,v} = a(c,v), \ \forall v \in V$. The notation $V'$ denotes the dual space of $V$, i.e., the space of linear functionals acting on $V$. Using this definition, the VBVP \eqref{eq:VBVP} is equivalent to the operator equation $\ccalM(c;s) = 0$ where $\ccalM: V \times S \to V'$.
Note that the functions $c$ and $v$ in the VBVP \eqref{eq:VBVP} have to be differentiable once.
Moreover, it can be shown that for $s \in S$ the BVP (\ref{eq:BVP}) and VBVP (\ref{eq:VBVP}) are equivalent and we can use them interchangeably. For further theoretical details, see \cite[ch. 8, 9]{IFA1998R}.
\subsection{The Source Identification Problem} \label{sec:probDef}
In this section, we formulate the SI problem as a constrained optimization problem subject to the AD transport model \eqref{eq:VBVP}.
Specifically, consider $m$ stationary sensors deployed in the domain $\Omega$ that take measurements of the concentration $c$, and let $E \subset \Omega$ be the set of $m$ compactly supported measurement areas enclosing the sensor locations.
\footnote{Note that the compact measurement area around any given sensor can be made arbitrarily small so that this sensing model approximates point measurements.}
Define, further, the indicator function $\chi_E: \Omega \to \set{0,1}$ for the set $E$ as
\begin{equation} \label{eq:indicFun}
\chi_E(\bbx) \triangleq
\left\{
\begin{array}{ll}
1 & \bbx \in E \\
0 & \bbx \notin E
\end{array}
\right.
\end{equation}
and let $c^m: \Omega \rightarrow \reals_+$ be a function that assigns to $\bbx \in \Omega$ the noisy concentration measurement at that location, i.e.,
\begin{equation} \label{eq:measureFun}
c^m(\bbx) = \chi_E(\bbx) \, c(\bbx) \, ( 1 + \epsilon ),
\end{equation}
where $\epsilon \sim {\cal{N}} (0, \ \sigma^2)$ and the measurement noise is proportional to the signal magnitude.
Then, the SI problem that we consider in this paper consists of determining an estimate $s$ of the true source term $\bbars$, given a set $E$ of $m$ noisy measurements in the domain $\Omega$, so that the AD model $\ccalM(c; s)=0$ defined in \eqref{eq:ADmodel} predicts the measurements $c^m$ as close as possible in the least squares sense.
The main challenges in solving the SI problem arise due to the following two reasons. First, generally the number of measurements $m$ is considerably smaller than the number of parameters that are used to describe the unknown source term. Second, the measurements are contaminated with noise.
To address these two challenges, we follow a standard approach and formulate the SI problem as a regularized least squares optimization problem subject to the AD model \eqref{eq:VBVP}. Let
\begin{equation} \label{eq:Xnorm}
\norm{c - c^m}_{\chi_E}^2 \triangleq \int_{\Omega} (c - c^m)^2 \ \chi_E \ d\Omega
\end{equation}
be a measure of discrepancy between the measurements and concentration field predicted by the AD model and define the cost functional $\ccalJ(c,s):V \times S \to \reals_+$ to be optimized by
$$ \ccalJ(c,s) \triangleq \frac{1}{2} \norm{c - c^m}_{\chi_E}^2 + \tau \ccalR(s) . $$
In this equation, $\tau$ is the regularization parameter and $\ccalR(s)$ is a functional that specifies the characteristics of the source $s$ that is selected as the solution of the SI problem.
In this work, we select
$ \ccalR(s) \triangleq \norm{s}_{L^1} = \int_{\Omega} \abs{s} \ d\Omega = \int_{\Omega} s \ d\Omega , $
where the last equality holds since $s$ is nonnegative. This choice of regularization penalizes the size of the source term.
Optimization of the objective functional $\ccalJ(c,s)$ subject to the AD model \eqref{eq:VBVP} gives rise to the following problem
\begin{equation} \label{eq:opt}
\min_{ (c,s) \in V \times S} \ccalJ(c,s) \ \, \st \ccalM(c,s) = 0,
\end{equation}
where the functional $\ccalM(c,s)$ is defined by \eqref{eq:ADmodel} and
\begin{equation} \label{eq:obj}
\ccalJ(c,s) = \frac{1}{2} \int_{\Omega} (c - c^m)^2 \ \chi_E \ d\Omega + \tau \int_{\Omega} s \ d\Omega .
\end{equation}
To solve the SI problem \eqref{eq:opt}, the gradient of the cost functional $\ccalJ(c,s)$ is needed. We obtain this gradient using the so called Adjoint Method. This method allows us to solve \eqref{eq:opt} directly in the reduced space $S$ of source functions rather than in the full space $V \times S$ of the concentration and source functions.
This is possible by using the model $\ccalM(c,s) = 0$ to represent the concentration $c$ as a function of the source term $s$, i.e., $c = \ccalF(s)$ where $\ccalF: S \to V$.
\footnote{As discussed in Section \ref{sec:ADPDE} such a representation exists and is unique.}
Using this gradient information we can minimize the cost functional $\bar{\ccalJ}(s) = \ccalJ(\ccalF(s),s)$ and determine the source term $s$ that solves the original problem \eqref{eq:opt}. See Appendix \ref{sec:adjointM} for the details of the Adjoint Method.
\subsection{Finite Dimensional Approximation} \label{sec:finiteDim}
The variables $c$ and $s$ of the optimization problem \eqref{eq:opt} are functions that live in the infinite dimensional function
spaces $V$ and $S$, respectively. Therefore, in order to solve this problem numerically, it is necessary to approximate $V$ and $S$ by finite dimensional subspaces $V_d \subset V$ and $S_d \subset S$ determined by appropriate sets of basis functions. This approximation allows us to parametrize the concentration and source functions by a finite number of parameters that depend on the basis functions that constitute $V_d$ and $S_d$.
The key idea to obtain the finite dimensional subspace $V_d$ of the concentration function space $V$ is to use Proper Orthogonal Decomposition (POD) to reduce the order of the model.
The POD method is easy to implement and gives an optimal set of basis functions that can be readily used in our formulation to parameterize $c$. For a survey of popular model order reduction methods, see, e.g., \cite{ALDS2005A}.
At the same time, we use a nonlinear representation of the source term $s$ as a combination of compactly supported tower functions. This representation reduces the dimension of $S_d$ drastically, compared to classical approaches that utilize the Finite Element method.
\subsubsection{Model Order Reduction} \label{sec:MOR}
To reduce the order of a model using POD we need to solve the AD-PDE \eqref{eq:BVP} for all values of the unknown source term and build a set of basis functions that span the solution of the AD model.
We refer to the solutions as the snapshots of the problem.
Let $C = \set{c_i(\bbx)}_{i=1}^R$ denote a set of $R$ snapshots obtained by solving the AD-PDE \eqref{eq:BVP} for different realizations of the source term, i.e., each $c_i(\bbx) \in V$ corresponds to a given $s_i(\bbx) \in S$.
The objective of POD is to generate a set of optimal basis functions that maximize the averaged projection of the snapshots over these basis functions; see, e.g., \cite{PODRBFCPE2001AK}.
This optimization problem is equivalent to an eigenvalue problem for the covariance matrix $\bbC \in \reals^{R \times R}$ defined by
\begin{equation} \label{eq:PODcov}
\bbC_{ij} \triangleq \frac{1}{R} \int_{\Omega} c_i \, c_j \, d\Omega .
\end{equation}
The details of this procedure are presented in Algorithm \ref{alg:POD}, which yields $V_d = \text{span} \set{\psi_k}_{k=1}^N $ for $N < R$ where $\psi_k$ are the POD basis functions.
\begin{algorithm}[t]
\caption{Proper Orthogonal Decomposition}
\label{alg:POD}
\begin{algorithmic}[1]
\REQUIRE The set of snapshots $C = \set{c_i(\bbx)}_{i=1}^R$;
\STATE Construct the covariance matrix $\bbC$ using equation (\ref{eq:PODcov});
\STATE Solve the eigenvalue problem $\bbC \bbQ = \bbLambda \bbQ$ such that \label{line:Lambda}
$$\lambda_1 \geq \lambda_2 \geq \dots \geq \lambda_R \geq 0 \text{ and } \bbQ = [\bbq^1 \ \bbq^2 \ \dots \ \bbq^R] ; $$
\STATE The POD bases $\set{\psi_k}_{k=1}^R$ are given by
\begin{equation}
\psi_k = \sum_{i=1}^R q^k_i c_i .
\end{equation}
\STATE For $N < R$ the reduced order model $c_d$ is given as $c_d \in V_d = \text{span} \set{\psi_k}_{k=1}^N$.
\end{algorithmic}
\end{algorithm}
In line \ref{line:Lambda} of this algorithm $\bbLambda$ is the diagonal matrix of the eigenvalues.
As shown in \cite[thm. 1]{PODRBFCPE2001AK}, the $i$-th eigenvalue $\lambda_i$ of matrix $\bbC$ contains the average energy in the $i$-th mode. Moreover for a given number $N < R$ of basis functions, the POD bases have the maximum possible energy and are optimal.
Thus, for a given fraction $\eta$, we can select the number $N$ of required bases as the smallest number such that
\begin{equation} \label{eq:selecN}
\frac{\sum_{i=1}^N \lambda_i}{\sum_{i=1}^R \lambda_i} \geq \eta.
\end{equation}
\subsubsection{Parameterization} \label{sec:param}
Using the basis functions $\psi_k$ that constitute $V_d = \set{\psi_k}_{k=1}^{N}$ we can represent the functions $c$ and $v$ by a finite number of parameters, that can be used for numerical optimization. Specifically, we define
\begin{equation} \label{eq:finiteDim}
c_d = \bbpsi \, \text{ and } \, v_d = \bbpsi \, \bbv ,
\end{equation}
where $\bbpsi = [\psi_1 \ \dots \ \psi_N]$ and $\bbc, \bbv \in \reals^N$.
To parametrize the source function $s$ we follow a different approach. Specifically, we propose a nonlinear representation of this term as a combination of compactly supported tower functions. The motivation for this representation is that each compactly supported source area can be approximately described by a very small number of parameters corresponding to the intensity and shape of the source. In this paper we focus on rectangular sources, although other geometric shapes can also be used for this purpose.
In particular, let $M$ be the number of basis functions used to approximate the source term in the domain $\Omega \subset \reals^d$ and consider two parameters $\set{ \undbbx_j, \barbx_j }$ for each basis function, where $\undbbx_j, \barbx_j \in \reals^d$ and $j \in \set{ 1, \dots, M}$. We define the compactly supported tower functions as
\begin{equation} \label{eq:towerF}
\phi_j(\bbx; \undbbx_j, \barbx_j) \triangleq
\left\{
\begin{array}{ccl}
1 & & \text{if } \undbbx_j \leq \bbx \leq \barbx_j \\
0 & & \text{o.w.}
\end{array}
\right.
\end{equation}
where the inequalities are considered component-wise and $\undbbx_j \leq \barbx_j$; cf. Figure \ref{fig:towerF}. Then, for practical purposes we can approximate the desired source term $s \in S$ by
\begin{equation} \label{eq:sourceFiniteDim}
s_d(\bbx) = \sum_{j=1}^M \beta_j \phi_j(\bbx; \undbbx_j, \barbx_j) ,
\end{equation}
where we require $\beta_j \geq 0$ so that $s_d \in S$.
We denote by $ \bbp = (\beta_1, \undbbx_1, \barbx_1, \dots, \beta_M, \undbbx_M, \barbx_M)$ the vector of parameters associated with the source term $s_d$. Thus for $\Omega \in \reals^d$, $\bbp \in \reals^p$ where $p = M(2d+1)$.
\begin{figure}[t!]
\centering
\includegraphics[width=0.35\textwidth]{main2.png}
\caption{The support of a tower function in $2$D case defined by equation \eqref{eq:towerF}. } \label{fig:towerF}
\end{figure}
Substituting the approximations $\set{c_d,s_d}$ of the concentration and source terms $c$ and $s$ in equation \eqref{eq:opt}, we obtain a finite dimensional counterpart of the SI problem as
\begin{align} \label{eq:optFD}
& \min_{ \bbc, \bbp} J(\bbc, \bbp) \\
& \ \st \bbM(\bbc, \bbp) = \bb0, \nonumber \\
& \ \ \ \ \ \ \beta_j \geq0, \ \bbl \leq \undbbx_j \leq \barbx_j \leq \bbu , \nonumber
\end{align}
where $j \in \set{ 1, \dots, M}$ and $\bbl, \bbu \in \reals^d$ are the lower and upper bounds on the coordinates of the domain.
\footnote{We can include additional convex constraints to confine the feasible region that might contain the true source. The extension to non-convex domains is considered in Section \ref{sec:autoSI}.}
Moreover, the objective $J:\reals^{N \times p} \to \reals_+$ is defined as $J(\bbc, \bbp) = \ccalJ(c_d, s_d)$ and the finite dimensional model $\bbM:\reals^{N \times p} \to \reals^N$ is defined as $\bbM(\bbc; \bbp) = \ccalM(c_d; s_d)$.
The optimization problem \eqref{eq:optFD} can be solved by a variety of available nonlinear optimization algorithms. Any such algorithm requires the first and possibly second order information, i.e., the gradient and Hessian of the objective function, as well as a proper initialization since the problem is nonlinear.
In Appendix \ref{sec:SInumeric}, we derive explicit expressions for the gradient and Hessian of the objective function in reduced space $S_d$. Consequently, only the bound constraints in \eqref{eq:optFD} need to be considered explicitly for numerical optimization. In the same appendix, we also discuss the Sensitivity Analysis (SA) method for the initialization of problem \eqref{eq:optFD}.
\subsection{The Next Best Measurement Problem} \label{sec:NBMP}
In this section we discuss the details of the numerical solution for the path planning developed in Section \ref{sec:path}. Let $\bbF(\bbx) = \bbS^T [ \bbX^T \bbX + \bbpsi(\bbx)^T \bbpsi(\bbx) ] \bbS$ denote the FIM at step $m$, where to simplify notation we have dropped the subscripts. Then introducing an auxiliary variable $z$ we can rewrite the optimization problem \eqref{eq:pathOptL} as
\begin{align} \label{eq:pathOptEig}
& \max_{z, \bbx} \ z \nonumber \\
& \ \st \lambda_i (\bbF(\bbx)) > z, \ \forall i \in \set{1, \dots, p}, \nonumber \\
& \ \ \ \ \ \ \bbx \in \Omega ,
\end{align}
where $\lambda_i$ denotes the $i$-th eigenvalue of the FIM. Problem \eqref{eq:pathOptEig} can be equivalently written as
\begin{align} \label{eq:pathOptL2}
& \max_{z, \bbx} \ z \nonumber \\
& \ \st \bbF(\bbx) - z \, \bbI \succ \bb0 , \nonumber \\
& \ \ \ \ \ \ \bbx \in \Omega ,
\end{align}
where the notation $\succ$ denotes a matrix inequality.
The optimization problem \eqref{eq:pathOptL2} is a nonlinear Semi-Definite Program (SDP) that can be solved using nonlinear optimization techniques; see, e.g., \cite{NO2006R}. In this paper we employ the Sequential SDP (SSDP) method which is the extension of sequential quadratic programming; see, e.g., \cite{SSDPMAPROM2005FJV}.
Defining
\begin{subequations} \label{eq:funDef}
\begin{align}
& f(z, \bbx) \triangleq -z , \label{eq:funDeff} \\
& \bbB(z, \bbx) \triangleq (\bbarepsilon+z) \bbI - \bbS^T [ \bbX^T \bbX + \bbpsi(\bbx)^T \bbpsi(\bbx) ] \bbS , \label{eq:funDefB}
\end{align}
\end{subequations}
where $0 < \bbarepsilon \ll 1$ is a very small positive number added to eliminate the strict inequality constraint, we can rewrite problem \eqref{eq:pathOptL2} in standard form as
\begin{align} \label{eq:NLSDP}
& \min_{z, \bbx} \ f(z, \bbx) \nonumber \\
& \ \st \bbB(z, \bbx) \preceq \bb0 , \nonumber \\
& \ \ \ \ \ \ \bbx \in \Omega .
\end{align}
The Lagrangian corresponding to problem \eqref{eq:NLSDP} is given as
\begin{equation} \label{eq:pathLag}
L(z, \bbx, \bbLambda) = f(z, \bbx) + ( \bbB(z, \bbx), \bbLambda ) \ ,
\end{equation}
where $\bbLambda \geq \bb0$ is the Lagrange multiplier matrix and the inner-product of two $r \times t$ real matrices $\bbB$ and $\bbLambda$ is defined as
$$ ( \bbB, \bbLambda ) = \tr (\bbB^T \bbLambda) = \sum_{i=1}^{r} \sum_{j=1}^{t} b_{ij} \lambda_{ij} . $$
Note that $\bbB(z, \bbx): \reals^{d+1} \to \mbS^p$ in \eqref{eq:funDefB} is a negative-semidefinite symmetric matrix function.
Since the Karush-Kuhn-Tucker (KKT) optimality conditions of the nonlinear SDP \eqref{eq:NLSDP} are locally identical to the second-order approximation around any point $(\bbarz, \barbx, \barbLambda)$, we can solve a sequence of convex SDPs to build the solution of the nonlinear problem \eqref{eq:NLSDP} iteratively. Under certain conditions that are satisfied for the functions in \eqref{eq:funDef}, the SSDP approach converges to a local minimum of the nonlinear SDP \eqref{eq:NLSDP}; see, e.g., \cite{SSDPMAPROM2005FJV}.
Specifically, at each iteration $k$, we construct a second-order convex approximation of \eqref{eq:NLSDP} at point $(z_k, \bbx_k, \bbLambda_k)$ as
\begin{align} \label{eq:tangent}
& \min_{\bbd \in \reals^{d+1}} \ \nabla f(\bbv_k)^T \bbd + 0.5 \ \bbd^T \, \bbH_k \, \bbd \nonumber \\
& \ \ \ \st \ \ \bbB(\bbv_k) + D_{\bbv} \bbB(\bbv_k) [\bbd] \preceq \bb0 , \nonumber \\
& \ \ \ \ \ \ \ \ \ \ \bbx_k + \bbd_{\bbx} \in \Omega ,
\end{align}
where $\bbv_k = (z_k, \bbx_k )$ is the primal variable at iteration $k$ and $\bbd = (d_z, \bbd_{\bbx})$ is a vector of directions, where $d_z \in \reals$ and $\bbd_{\bbx} \in \reals^d$ are the directions corresponding to $z_k$ and $\bbx_k$, respectively. The matrix $\bbH_k$ is a positive semidefinite approximation of the Hessian $ \nabla_{\bbv \bbv}^2 L(z_k, \bbx_k, \bbLambda_k)$ of the Lagrangian \eqref{eq:pathLag} and $D_{\bbv} \bbB(\bbv_k) [\bbd]$ is the directional derivative of the matrix function \eqref{eq:funDefB} at point $\bbv_k$ and direction $\bbd$ that is used to linearize the matrix inequality constraint around the current iterate $\bbv_k$.
This quantity along with the Hessian of the Lagrangian are derived in Appendix \ref{sec:1st2ndDeriv}.
We assume that the domain $\Omega$ is convex and defined by a set of affine constraints so that the linear constraint $\bbx_k + \bbd_{\bbx} \in \Omega$ can be directly incorporated in the SDP \eqref{eq:tangent}. This assumption holds for the box constrained domain that we considered in the SI problem \eqref{eq:optFD}.
The solution of the SDP \eqref{eq:tangent} denoted by $\bbd_k \in \reals^{d+1}$ determines the descent direction for the nonlinear problem \eqref{eq:NLSDP}. Using this solution, we update the primal variables as
\begin{equation} \label{eq:iterSSDP}
\bbv_{k+1} = \bbv_k + \alpha_k \bbd_k ,
\end{equation}
where $\alpha_k$ is a step-size whose selection is explained in Appendix \ref{sec:stepSize}.
Note that by the last constraint in the SDP problem \eqref{eq:tangent}, we implicitly assume that the maximum step-size is equal to one, i.e., $\alpha_{\max}=1$.
We update the dual variable $ \bbLambda_{k+1}$ directly as the optimal dual of the tangent problem \eqref{eq:tangent}. The details of the SSDP to solve the optimization problem \eqref{eq:pathOptL} are presented next.
\subsection{First and Second Order Information} \label{sec:1st2ndDeriv}
To solve problem \eqref{eq:NLSDP} we need the gradient and Hessian information. For this, we first define the expressions $D_{\bbv} \bbB(\bbv) [\bbd]$ and $\nabla_{\bbv \bbv}^2 L(z, \bbx, \bbLambda)$ that appear in the convex second-order SDP \eqref{eq:tangent}.
For the first term, we have
\begin{equation} \label{eq:linearCons}
D_{\bbv} \bbB(\bbv) [\bbd] = \sum_{i=1}^{d+1} d_i \ \bbB^{(i)}(\bbv) \, ,
\end{equation}
where $d_i$ is the $i$-th element of the vector of directions $\bbd$ and
\begin{equation} \label{eq:deriv1}
\bbB^{(i)}(\bbv) = \frac{\partial}{\partial v_i} \bbB(\bbv) ,
\end{equation}
for $i \in \set{1, \dots, d+1}$.
The operator $D_{\bbv} \bbB(\bbv): \reals^{d+1} \to \mbS^p$ is linear in $\bbd$ and $ D_{\bbv} \bbB(\bbv) [\bbd] \in \mbS^p$. Therefore the corresponding constraint in the SDP \eqref{eq:tangent} is a linear matrix inequality.
For the second term, i.e., the Hessian $\nabla_{\bbv \bbv}^2 L(z, \bbx, \bbLambda)$ of the Lagrangian \eqref{eq:pathLag}, since the objective function $f(z, \bbx)$ defined by equation \eqref{eq:funDeff} is linear, we have
$$\nabla_{\bbv \bbv}^2 L(z, \bbx, \bbLambda) = \nabla_{\bbv \bbv}^2 ( \bbB(z, \bbx), \bbLambda ) \in \mbS^{d+1} ,$$
where
\begin{align} \label{eq:LagHess}
\nabla_{\bbv \bbv}^2 ( \bbB, \bbLambda ) =
\left[
\begin{array}{ccc}
( \bbB^{(1,1)}, \bbLambda ) & \dots & ( \bbB^{(1,d+1)}, \bbLambda ) \\
\vdots & \ddots & \vdots \\
( \bbB^{(d+1,1)}, \bbLambda ) & \dots & ( \bbB^{(d+1,d+1)}, \bbLambda )
\end{array}
\right] ,
\end{align}
and
\begin{equation} \label{eq:deriv2}
\bbB^{(i,j)}(\bbv) = \frac{\partial^2}{\partial v_i \partial v_j} \bbB(\bbv) .
\end{equation}
Recalling the definition of the matrix function $\bbB(z, \bbx)$, given in equation \eqref{eq:funDefB}, we calculate the required derivatives \eqref{eq:deriv1} and \eqref{eq:deriv2} for, e.g., the $2$D case in which $\bbx = (x_1, x_2)$.
These derivatives then are used in equation \eqref{eq:tangent} to build quadratic SDPs that we solve sequentially to find the local optimum of the nonlinear SDP \eqref{eq:NLSDP}.
For the first order derivatives used in equation \eqref{eq:linearCons}, we have
\begin{align*}
& \bbB^{(1)} = \frac{\partial \bbB}{\partial z} = \bbI, \\
& \bbB^{(2)} = \frac{\partial \bbB}{\partial x_1} = - \bbS^T \left[ (\frac{\partial \bbpsi}{\partial x_1})^T \bbpsi + \bbpsi^T \frac{\partial \bbpsi}{\partial x_1} \right] \bbS .
\end{align*}
The value for $\bbB^{(3)}$ is exactly the same as $\bbB^{(2)}$, except that the differentiation variable is $x_2$.
Similarly for the second-order derivatives used in equation \eqref{eq:LagHess}, we have
$$ \bbB^{(1,1)} = \bbB^{(2,1)} = \bbB^{(3,1)} = \bb0 , $$
$$ \bbB^{(2,2)} = - \bbS^T \left\{ (\frac{\partial^2 \bbpsi}{\partial x_1^2})^T \bbpsi + 2 (\frac{\partial \bbpsi}{\partial x_1})^T \frac{\partial \bbpsi}{\partial x_1} + \bbpsi^T \frac{\partial^2 \bbpsi}{\partial x_1^2} \right\} \bbS. $$
$\bbB^{(3,3)}$ can be calculated exactly the same way. Finally, for the cross-derivative we have
\begin{align*}
\bbB^{(3,2)} = - \bbS^T & \left\{ (\frac{\partial^2 \bbpsi}{\partial x_1 x_2})^T \bbpsi + (\frac{\partial \bbpsi}{\partial x_1})^T \frac{\partial \bbpsi}{\partial x_2} + \right. \\
& \left. (\frac{\partial \bbpsi}{\partial x_2})^T \frac{\partial \bbpsi}{\partial x_1} + \bbpsi^T \frac{\partial^2 \bbpsi}{\partial x_1 x_2} \right\} \bbS.
\end{align*}
After calculating the Hessian $\nabla_{\bbv \bbv}^2 L(z, \bbx, \bbLambda)$ of the Lagrangian \eqref{eq:pathLag}, we construct a positive-definite approximation $\bbH$ of it so that the SDP \eqref{eq:tangent} is strictly convex with a unique global minimizer $\bbd$.
Such an approximation of $\bbH$ can be obtained in different ways; see, e.g., \cite{MIMAO1998HC}.
Here, we add a multiple of the identity matrix so that the minimum eigenvalue is bounded from zero by a small amount $\delta$, i.e., we set
\begin{equation} \label{eq:convexification}
\bbH = \nabla_{\bbv \bbv}^2 L + \mu \, \bbI,
\end{equation}
where $\mu = \max(0, \delta - \lambda_{\min}(\nabla_{\bbv \bbv}^2 L))$. The positive-definite matrix $\bbH$ is the closest to the Hessian $\nabla_{\bbv \bbv}^2 L$ measured by the induced Euclidean norm. Note that since the Hessian is a low dimensional matrix, i.e., $\nabla_{\bbv \bbv}^2 L \in \mbS^{d+1}$, we can easily calculate its minimum eigenvalue.
\subsection{Step-Size Selection} \label{sec:stepSize}
Necessary for the solution of the nonlinear SDP \eqref{eq:NLSDP} is an effective line-search strategy that connects the successive solutions of the quadratic SDPs \eqref{eq:tangent}. In this paper, we utilize the results from \cite{GANSDP2004C} to select an appropriate step-size $\alpha_k$ for the iterations of the SSDP defined by equation \eqref{eq:iterSSDP}.
The final SSDP algorithm to solve the nonlinear SDP \eqref{eq:NLSDP} is presented in Algorithm \ref{alg:SSDP}.
\begin{algorithm}[t]
\caption{Sequential Semi-definite Programming}
\label{alg:SSDP}
\begin{algorithmic}[1]
\REQUIRE The POD bases $\bbpsi = [\psi_1, \dots, \psi_N]$ of Algorithm \ref{alg:POD};
\REQUIRE The parameters $\bbarepsilon$ and $\delta$ of equations \eqref{eq:funDef} and \eqref{eq:convexification};
\REQUIRE The parameters $\epsilon_1$, $\epsilon_2$, and $\epsilon_3$ of equation \eqref{eq:stop};
\REQUIRE The parameters $\bbargamma > 0$, $\rho \in (0,1)$, and $\omega \in (0,1)$;
\STATE Initialize the iteration index $k=0$;
\STATE Initialize the primal variable $\bbv_0$ using equation \eqref{eq:initialize} and the dual variable $\bbLambda_0$ with $\bbLambda_0 \succeq \bb0$; \label{line:initialize}
\STATE Initialize the penalty parameter as $\gamma_0 = \tr(\bbLambda_0) + \bbargamma$;
\WHILE{the algorithm has not converged}
\STATE Build the convex SDP \eqref{eq:tangent} at $(\bbv_k, \bbLambda_k)$ using equations \eqref{eq:linearCons}, \eqref{eq:LagHess}, and \eqref{eq:convexification} and solve it for $(\bbd_k, \bbLambda_{k+1})$;
\STATE Check the stopping criterion \eqref{eq:stop} for $\bbv_k$;
\STATE Set $\gamma_k = \gamma_{k-1}$ if $\set{ \gamma_{k-1} \geq \tr(\bbLambda_{k+1}) + \bbargamma }$, otherwise set it as $\gamma_k = \max \set{1.5 \gamma_{k-1}, \tr(\bbLambda_{k+1}) + \bbargamma}$;
\STATE Select $\alpha_k$ as the largest member of the geometric sequence $\set{1, \rho, \rho^2, \dots}$ such that
$$ \theta_{\gamma_k} (\bbv_k + \alpha_k \bbd_k) \leq \theta_{\gamma_k} (\bbv_k) + \omega \alpha_k \Delta_k , $$
where the penalty function $\theta_{\gamma}(\bbv)$ and $\Delta_k$ are defined in equations \eqref{eq:penFun} and \eqref{eq:upperB}, respectively; \label{line:Armijo}
\STATE Update the primal variable $\bbv_{k+1}$ by equation \eqref{eq:iterSSDP};
\STATE $k \leftarrow k+1$;
\ENDWHILE
\end{algorithmic}
\end{algorithm}
In this algorithm, we define the penalty function $\theta_{\gamma}(\bbv)$ for the selection of the step-size $\alpha_k$ as
\begin{equation} \label{eq:penFun}
\theta_{\gamma}(\bbv) = f(\bbv) + \gamma \, \lambda_{\max}(\bbB(\bbv))_+ ;
\end{equation}
where $\gamma > 0$ is the penalty parameter, $\lambda_{\max}(\bbB)_+ = \max \set{0, \lambda_{\max}(\bbB)}$, and the functions $f(\bbv)$ and $\bbB(\bbv)$ are defined in equation \eqref{eq:funDef}. The upper bound $\Delta_k$ on the directional derivative $\theta'_{\gamma_k}(\bbv_k; \bbd_k)$ of the penalty function in a direction $\bbd_k$ is given as
\begin{equation} \label{eq:upperB}
\Delta_k = - \bbd_k^T \bbH_k \, \bbd_k + \tr(\bbLambda_{k+1} \bbB(\bbv_k)) - \gamma_k \, \lambda_{\max}(\bbB(\bbv_k))_+ ,
\end{equation}
where $\bbH_k$ is the positive-definite approximation given by equation \eqref{eq:convexification} and we have included the index $k$ to emphasize that we use the dual variable $\bbLambda_{k+1}$ to select the step-size $\alpha_k$.
The upper bound $\Delta_k$ is used in order to satisfy the Armijo condition in the backtracking line-search corresponding to line \ref{line:Armijo} in Algorithm \ref{alg:SSDP}. For theoretical details see \cite{GANSDP2004C}.
Note that since the domain of interest $\Omega$ is represented by a set of affine constraints that require no further linearization, the constraint $\bbx \in \Omega$ does not appear in the penalty function \eqref{eq:penFun}. Essentially, the constraint $\bbx \in \Omega$ is never violated and thus we do not penalize it in \eqref{eq:penFun}.
\subsection{Initialization and Stopping} \label{sec:pathInit}
Since the eigenvalue optimization problem \eqref{eq:pathOptL} is nonlinear, appropriate initialization of Algorithm \ref{alg:SSDP} is critical to obtain a reasonable solution.
Moreover, addition of a new measurement, reshapes the objective function \eqref{eq:pathOptL} and makes it flat around that measurement location. In other words, adding more measurements in that vicinity does not provide more information about the unknown source parameters compared to farther locations.
Therefore without global knowledge of the objective function, the algorithm gets trapped in undesirable local minima where the objective function does not change no matter how many measurements are taken in that region.
In order to generate new informative measurements, we sample the objective function of the Next Best Measurement Problem \eqref{eq:pathOptL},
$ g(\bbx) = \lambda_{\min} [ \bbF(\bbp) + \bbS(\bbp)^T \bbpsi(\bbx)^T \bbpsi(\bbx) \bbS(\bbp) ] , $
over a coarse set of points $\bbz_i \in \reals^d$ from the FE mesh, where $i \in \set{1, \dots, Z}$ for some $Z \ll n$ and $n$ is the number of FE grid points.
Then to initialize the primal variable $\bbv_0 = (z_0, \bbx_0)$ in Algorithm \ref{alg:SSDP} for step $m + 1$ of the robot, we calculate the values of the objective function $g_i = g(\bbz_i)$ over this set of points and we set
\begin{equation} \label{eq:initialize}
z_0 = \max_i g_i \ \text{and} \ \bbx_0 = \argmax_{\bbz_i} g(\bbz_i) ,
\end{equation}
where $z_0$ is the auxiliary variable introduced in \eqref{eq:pathOptEig}.
Note that each evaluation of the function $g(\bbx)$ amounts to solving a minimum eigenvalue problem for a $p \times p$ matrix where $p$ is the number of unknown parameters. The computational cost of such sampling procedure is comparable to a single backtracking line-search step of Algorithm \ref{alg:SSDP} in line \ref{line:Armijo}.
Finally, to determine whether Algorithm \ref{alg:SSDP} has reached a local minimum we evaluate bounds on the gradient of Lagrangian \eqref{eq:pathLag}, the nonlinear matrix inequality constraint violation, and the complementarity condition as follows
\begin{align} \label{eq:stop}
& \norm{ \nabla_{\bbv} L(\bbv_k, \bbLambda_{k+1}) }_2 \leq \epsilon_1, \ \lambda_{\max}(\bbB(\bbv_k))_+ \leq \epsilon_2, \nonumber \\
& \ \ \ \ \ \ \ \ \ \ \ \ \ \ \ \abs{ \tr(\bbLambda_{k+1} \bbB(\bbv_k)) } \leq \epsilon_3 .
\end{align}
See \cite{GANSDP2004C} for theoretical results supporting this selection of stopping criteria.
| 18d179f8ff0865dea87dd584cc4bce58e0bf790d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{intro}
Hierarchical structure formation models predict that the constituents of large scale structures (LSS) should follow the self-similarity \citep{kaiser1986MNRAS,morandi2016MNRAS}. Though, extensive studies on galaxies and galaxy clusters have established their distinct characteristics, intermediate structures are not studied well and for long, they were speculated to be just a scaled down version of the clusters \citep{kaiser1986MNRAS,Vikhlinin2006ApJ}. But, a handful of recent observations and simulations indicate a discrepancy in energy and mass scaling in low mass systems (\cite{Gaspari2011MNRAS,Dave_2002ApJ,Bharadwaj2015,Planelles_2013MNRAS,Paul_2015fers.confE} etc.). In the structural hierarchy, there are intermediate objects with distinct physical properties, possibly loose groups or cabal, that can help us to understand the evolutionary trails of galaxy clusters from galaxies. While, clusters are the most massive bound structures, groups are more numerous and are home to a significant fraction of the entire galaxy population in the universe \citep{Mulchaey2000ARA&A,Kamatsu2002MNRAS}. Unlike clusters that emerges at the filamentary nodes, groups are also form inside the dark matter (DM) filaments connecting the clusters \citep{lietzen2012,tempel2014a}.
Groups are supposed to be virialized as their sound crossing times are much less than the Hubble time (0.1 - 0.5 $\rm{H_0^{-1}}$). But, in reality, groups are mostly found to be in non-virialized state \citep{Diaferio1993AJ,Diaz2010MNRAS}. Moreover, groups being inside the shallower gravitational potential of filaments, are expected to be strongly affected by mergers, shock heating, feedback from supernova and super-massive black holes, and galactic winds etc. \citep{Lovisari2015}. Mach number dependent cosmic ray content is shown to be larger by almost an order of magnitude in the groups than in the clusters \citep{Jubelgas2008A&A}. Fractional presence and the activity of AGN are also much higher in low-mass systems \citep{Gilmour2007MNRAS,Sivakoff2008ApJ}. Thus, group environment is significantly different than the clusters, leading to the expectation of having different physical properties. This also indicates that many of the scaling laws derived from the cluster properties would not be applicable for groups. Some of the recent researches by \citealp{Sun2012NJPh,Stott_2012MNRAS,Dave_2002ApJ,Gaspari2011MNRAS,Bharadwaj2015,Planelles_2013MNRAS} etc. have indicated such deviations. But, whom should we call as group? No universal definition or classifying characteristics are provided for the groups yet. Also, the above mentioned studies are mostly dealing with either a specific subset of data or are done with the radius and mass at over density ratio of 500. But to verify the self similar scaling, it will be physically more appropriate to work with unbiased complete set of data as well as verify the scaling laws with virialized objects \citep{kaiser1986MNRAS,Cole_1996MNRAS,Miniati_2015Natur} i.e. properties within over-density $\sim 200$, derived from the spherical collapse and dynamical equilibrium model (well known value is 178). So, our study will mostly focus on over-density of 200 to find out a generalised definition or classifying characteristics for ``galaxy groups" and ``galaxy clusters".
In this study, we intend to understand the distinct features of low mass objects in comparison to the clusters through cosmological hydrodynamic plus N-body simulations. We have theoretically modelled LSS, taking into account the effect of cooling (radiative processes, e.g. X-ray emission), heating (Supernova (SN) and star motions), star formation and star formation feedback physics that we will further call as `coolSF' simulations. AGN feedback has not been considered in this present work which may fine tune our results, mostly, at very low masses i.e. few times of $10^{13} \rm{M_\odot}$ and below as indicated by \cite{LeBrun_2014MNRAS,McCarthy2010MNRAS} etc. Cluster core properties can also be effected by AGN feedback by quenching star formation, thus resulting change in X-ray emissions and related scaling laws \citep{Rasia_2014ApJ,LeBrun_2014MNRAS}, therefore, becomes a subject of our future research. Above studies also show that both coolSF and coolSF+AGN feedback are producing far better results than the non-radiative simulations. Self similar parameters produced by coolSF and additional AGN feedback are by and large agree with the observations. With the data set they have compared, it is very difficult to conclude which one is fitting better. For some parameters though, either of the model is slightly over or under-predicting, and few specific parameters only are better produced by coolSF+AGN simulations.
In current study, coolSF model of our's itself is seen to be fairly able to reproduce the observations (detail study will be followed in the next sections), unlike the studies by \cite{LeBrun_2014MNRAS,McCarthy2010MNRAS}. We have thus used the coolSF model for creating our sample set of about 360 objects and an extensive work has been done to figure out a possible break away point or knee in cluster scaling laws those distinguish correctly a `group' from a cluster. The reasons behind the unique properties of groups have also been investigated.
Section~\ref{intro} deals with the introduction to the problem with its background motivations. The simulation details and sample selection etc. has been written in Section~\ref{sample}. Details of studied parameters and their results are given in Section~\ref{study-self}. For reliability of our simulations, we have presented a resolution study in Section~\ref{res-study}. Discussion on the distinct features and point of segregation of groups from the clusters has been written in Section~\ref{discus}. Finally, we have summarised our findings in Section~\ref{conclude}.
\section{Simulation details and sample selection}\label{sample}
Lack of studies on properties of low mass systems prompted us to model them using cosmological simulations. To create our sample set, basic simulations were performed with the adaptive mesh refinement (AMR), grid-based hybrid (N-body plus hydro-dynamic) code Enzo v.~2.2 \citep{Bryan_2014ApJS}. With introduction of 2 nested child grid and another 4 levels of AMR at the central (32 $\rm{Mpc})^3$ volume (i.e. total 6 levels of refinement), we have achieved a resolution of $\sim$ 30 kpc at the highest level for our studied sample set. For resolution study, we have simulated two other sets of data with a lower resolution (`LOWRES' here after) $\sim$ 60 kpc and a higher resolution (`HIGHRES' here after) $\sim$ 15 kpc than the reference set (`REFRES' here after) i.e. $\sim$ 30 kpc. As cosmological parameters, we have taken a flat $\Lambda$CDM background cosmology with $\Omega_\Lambda$ = 0.7257, $\Omega_m$ now = 0.2743, $\Omega_b$ = 0.0458, h = 0.702 and primordial power spectrum normalization $\sigma_8$ = 0.812 \citep{Komatsu_2009_APJS}.
In an AMR simulation, it is very much important to resolve the objects very well. Density has been used as the primary refinement criteria to resolve the cells in this study. If a cell has density 4 times the average density of the neighbouring cells, the cell is refined. Similarly, if a cell has 4 times higher dark matter mass than the average mass of its neighbouring cells, the cell is refined. Since, density is very low at the outskirts of the large scale objects, we also require a dynamical parameter that covers almost all the volume of the objects. It is therefore convenient to choose the shocks as second parameter to refine the cells as, in forming objects, shocks spans over most of the volume. This is also justified since, shock heating is a primary heating engine in large structures that helps objects resist the rapid collapse due to radiative cooling and decrease unphysical star formation activity. Proper shock capturing is thus very important for scaling studies. In this study, shock computing has been done using un-split velocity jump method of \citet{Skillman_2008ApJ} with a temperature floor of $\rm{T}^4$K which is found to give better results in AMR simulations \citep{Vazza_2011MNRAS}. Heating, cooling and feedback physics has a dominant effect on smaller structures. We have implemented radiative cooling due to X-ray, UV \& optical emissions and heating is due to stellar motions and Supernova \citep{Sarazin1987ApJ}. Star formation and feedback scheme of \citet{Cen1992ApJ} have been implemented with a feedback of 0.25 solar. This is the model of additional physics that we have named as `coolSF' runs.
After generating initial conditions using the Eisenstein \& Hut transfer function \citep{Eisenstein_1998ApJ}, hydro plus N-body simulations were performed from redshift z=60 to current redshift z=0. We have saved data outputs in the intermediate redshifts and a very frequent snapshots were taken below redshift 1. Physical parameters are computed on these snapshots using tools developed from yt \citep{Turk_2011ApJS}.
We have simulated 10 realizations of parts of our universe, each of volume $(128\;\rm{Mpc})^3$. Central $(32\;\rm{Mpc})^3$ of highly resolved volume of each simulation contains few 10s of groups and clusters. We have identified individual objects using yt halo finder tool. For identifying the objects, we have used virial radius to be their delimiter, where, virial refers to the quantity at over density of 200 to the critical density of the universe at that redshift. In the further text, over density should be understood as compared to the critical density only. Finally, about 360 objects have been selected in the mass range of $5\times10^{12}\; M_{\odot}$ to 2.5 $\times 10^{15}\; M_{\odot}$. Our mass resolution at the smallest child grid is $<$ $10^{9}\; M_{\odot}$ providing enough mass resolution even for the groups with mass $5 \times 10^{12}\; M_{\odot}$ . Also, with $\sim$ 30 kpc spatial resolution, systems above this mass, having virial radii above 500 kpc get adequately resolved in space.
\section{Study of self similarity in the selected sample set}\label{study-self}
The scaling laws of self similarity are derived from ideal spherical collapse model with gravity only situations. Dark Matter (DM) only studies that corroborate the results of existence of self similarity \citep{kaiser1986MNRAS,Navarro_1996ApJ,Vikhlinin2006ApJ} could not explain the observed $\rm{L_x}$-T scaling for smaller objects. This model has failed to account for the dynamical and transient events such as mergers, also for thermodynamics and other forms of energies in these systems. So, the introduction of baryons to the DM-only formalism changes the physical conditions significantly, causing substantial variations in the energetics of these objects \citep{Dai_2010ApJ,Bharadwaj2015}. Since, the overall environment and dynamical processes experienced by the smaller and larger systems are different (see Section~\ref{intro}), it is quite possible that smaller objects can deviate from the cluster scaling laws. So, to understand the energetics of groups and clusters, we certainly need to study the correlation scaling of different physical parameters, both thermal and non-thermal along with the hydrostatic ones. Further, transformation of thermal to non-thermal energy could be a strong parameter to check for, as this indicates the ongoing dynamical processes that control the energy budget of the systems, thus explaining the deviation from self-similarity.
\subsection{Self similarity in thermal properties}\label{self-sim}
Dark matter being the dominant matter component of the universe, gas would just follow the DM collapse. In such a situation, as derived by \cite{Peebles_1980lssu.book,Kitayama_1996ApJ,kaiser1986MNRAS}, relations among the observables of gas properties with the mass observables can in fact be formulated. Virial temperature (T) of the system can be related to the total virial mass, by considering the heating of the medium through conversion of potential energy of collapsing gas as $\rm{T\propto (\Delta \rho_{cr})^{1/3} M^{2/3}}$. Assuming thermal Bremsstrahlung X-ray emission to be $\propto\rho \rm{T^{1/2}}$, X-ray luminosity will scale to the virial temperature as $\rm{L_X \propto\rho^{2} T^{1/2} r^{3}}$ i.e. $\propto \rm{T^{2}}$ as $\rm{r^3 \propto M \propto T^{3/2}}$. These also indicate that two very closely related parameters i.e. mass and X-ray luminosity should have a very tight correlation of $\rm{L_X \propto M^{4/3}}$, when estimated from independent observations. Entropy (S) being a function of density and temperature, should also show self similar relation which is given by $\rm {S \propto T}$. So, the crucial and independent self similar scaling relations needs to be studied are $\rm{L_X}$-M, $\rm{L_X}$-T and S-T and parameters must be computed within virial radius, i.e. within over-density of $\sim$ 200.
In our simulations, mass and virial temperature has been computed upto the over-density radius of $\rm{r_{200}}$ i.e. usual virial radius of the systems. We have computed virial temperature as a function of mass and virial radius of the object i.e. $\rm{^1T_{vir} = \frac{G M \mu m_{p}}{(k_b R_{vir})}}$ (operational temperature definition 1). For a bound and self-gravitating system, the virial mass can be related to the velocity dispersion of the system by $\rm{M = \frac{3 R_{vir} \sigma^{2}}{G}}$. Since, velocity dispersion $\sigma$ can be computed independently in our simulations by filtering out the bulk motion, we get a second definition of temperature as $\rm{^2T_{vir} = \frac{3\mu m_{p} \sigma^2}{k_b}}$, which is the same as the X-ray temperature if $\beta_{spec}$ value is considered to be unity (observed average value \citep{Girardi_1996ApJ}) in $\rm{T_{x} = \frac{3\mu m_{p} \sigma^2}{k_b \beta_{spec}}} $. Finally, entropy (S), which is a function of temperature and density of the objects, has been computed using the relation $\rm{ \frac{k_b T_{vir}}{\mu m_p \rho^{\gamma}}}$.
Gravitational collapse, mergers during structure formation leads to adiabatic compression of the intra-cluster medium (ICM) and heating up the structures to as high as $10^8 \rm{K}$ \citep{Sarazin_2002ASSL,Mathis_2005MNRAS,Paul2011ApJ}. Such a medium then emit X-rays through thermal bremsstrahlung (For a review:\cite{Felten_1966ApJ}), Inverse Compton Scattering (ICS) of Cosmic Microwave Background Radiation etc. \citep{Costain_1972ApJ}. Values of X-ray luminosity, emissivity, and photo emissivity fields for a given photon energy range can be obtained using Cloudy~\citep{Ferland_1998PASP} code. For the present study, we have limited the energy of X-ray emission between 0.1 keV to 12.0 keV in order to comply with the range of existing X-ray telescopes. This range is also sufficient for calculating approximate bolometric luminosity as beyond this range X-ray flux from thermal gas in galaxy clusters falls by $10^{-3}$ \citep{Henriksen_1986ApJ} times i.e. the considered energy band covers almost 99\% of the bolometric flux of the cluster samples.
\subsubsection{Comparison of simulated parameters with observations}
\begin{figure*}
\includegraphics[width=9cm]{lx_t_r_500_all_obs_T1.pdf}\hspace{-0.8cm}
\includegraphics[width=9cm]{lx_t_r_500_all_obs_T2.pdf}
\caption{{\bf Panel 1:} X-ray luminosity plotted against virial temperature (definition 1, $\rm{^1T}$) computed upto radius $\rm{r_{500}}$. Observed data points from \citep{Mittal_2011A&A,Helsdon_2000MNRAS,Maughan_2012_MNRAS, Markevitch_1998ApJ,Osmond_2004_MNRAS,Xue_2000ApJ,Arnaud_1999MNRAS,Zhang_2011A&A} have been overplotted as coloured stars. {\bf Panel 2:} Same plot as above but, against virial Temperature (definition 2, $\rm{^2T}$).}\label{x-ray-obs}
\end{figure*}
Most of the X-ray observations are done upto over-density ratio of 500 (i.e. till $\rm{r_{500}}$). So, to validate our results using observations, we have created a data set by computing X-ray luminosity till $\rm{r_{500}}$. X-ray luminosity has been plotted against virial temperature with both $\rm{^1T_{vir}}$ and $\rm{^2T_{vir}}$ of the selected samples from our simulations in Figure~\ref{x-ray-obs}, panel 1\&2 respectively. Observed X-ray luminosity from similar objects found in different studies have been over-plotted in the same. Since, both virial temperatures have been derived from hydrostatic equations that are related to self similarity, we have chosen to plot X-ray luminosity against these temperatures. This is also relevant as almost all the observed objects that are over-plotted here are available against virial temperature only. It can be noticed that simulated mock data set nicely follow the observed data trend. A large scatter in $\rm{L_x}$-T plot can be noticed when plotted against $\rm{^2T_{vir}}$, which is a function of velocity dispersion. Since, we did not put any bias while selecting our sample set from our simulations, it is obvious that it will contain merging objects as well. Velocity dispersion being a strong function of dynamics of the objects, mergers can strongly influence the temperature calculated from this parameter. On the other hand, possibly mass being comparatively smoothly varying parameter, shows less scatter. So, to have a better control over computation of scaling relations, we have chosen $\rm{^1T_{vir}}$ for further temperature related calculations.
\begin{figure*}
\includegraphics[width=9cm]{Baryonic_fraction.pdf}\hspace{-0.8cm}
\includegraphics[width=9cm]{entropy_our_temp.pdf}
\caption{{\bf Panel 1:} Baryon fraction (in \%) has been plotted against mass ($\rm{M_{500}}$). Observed data points from \citet{Gonzalez_2013ApJ,Giodini_2009ApJ,Lagana_2011ApJ,Sanderson_2013MNRAS,Lagana_2013A&A,Andreon_2010MNRAS} have been overplotted as coloured stars. {\bf Panel 2:} Entropy against virial mass has been plotted \citep{Sun_2009ApJ,Pratt_2010A&A} and compared with the overplotted observed data.}\label{x-ray-obs1}
\end{figure*}
Further, we have plotted baryon fraction and entropy within $\rm{r}_{500}$ against mass ($\rm{M}_{500}$) for our simulated samples and over plotted the available set of observed data in Figure~\ref{x-ray-obs1}, panel 1\&2 respectively. Though, simulated baryon fraction seems to follow observed values well, entropy shows a deviation below $\sim 10^{14} \; M_{\odot}$. Lack of observed data points in low mass makes it difficult to understand the reason for the apparent deviation. So, overall, these three parameters fits well to the observations and validate our coolSF simulation model used in this study to a large extent. It can also be noticed that, both observed and simulated data set show strong indication of deviation from cluster scaling for low mass systems. So, further, in this paper, we will compute the point of break where cluster scaling deviates from the self similarity.
\subsubsection{Breaks in self similar laws}
\begin{figure*}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.3cm},clip]{lx_mass_scaling.pdf}\hspace{-0.4cm}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.3cm},clip]{lx_t_r_200_split_T.pdf}\hspace{-0.4cm}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.3cm},clip]{temp_entropy.pdf}\hspace{-0.4cm}
\caption{X-ray Luminosity vs virial mass and virial temperature (upto $\rm{r_{200}}$) has been plotted from our sample data set in Panel 1 \& 2 respectively. B\'ezier curve (Section~\ref{self-sim}) have been fitted and plotted over data in Panel 1. In Panel 3. entropy of these systems plotted against virial temperature. All the plots are fitted with two slopes with break points using method as specified in Section~\ref{self-sim}.}
\label{xray-mass}
\end{figure*}
Self similar scaling laws are derived for the hydrostatic equilibrium or virial equilibrium. From the spherical collapse models, the original formulation was done using virial radius at over-density of $\sim 200$ (more accurately, 178) \citep{kaiser1986MNRAS,Cole_1996MNRAS,Miniati_2015Natur}. Though, use of over-density of $\sim 200$ would be more appropriate, most of the previous studies are observed to be done with over density 500 or $\rm{r}_{500}$ which does not guaranty a hydrostatic equilibrium. So, to ensure correct formulation and appropriate calculations, we have computed our parameters at over-density 200.
Figure~\ref{xray-mass}, Panel~1 shows the X-ray luminosity against mass within the radius $\rm{r_{200}}$ obtained from our simulated data set. It is always difficult to understand actual trend from the data set when large fluctuations are present and is a multivalued function. To overcome this difficulty, we have taken the statistical average of the data at each point and fitted a cubic B\'ezier curve to smooth out the fluctuations to compute actual trend and the bending points in the curve. In Panel 1, the B\'ezier fitted curve seen to turn towards cluster self similarity at $\sim$ $5\times 10^{13}\;\rm{M_{\odot}}$ and falls into the cluster scale beyond $\sim 8\times 10^{13}\;\rm{M_{\odot}}$. Further, to estimate the accurate point of break in the scale, we have computed the fitting power laws using all the data points in such a manner that slopes get connected at the break point. Power law fitting in figure~\ref{xray-mass}, Panel~1 shows that structures with mass below $\sim 8\times 10^{13}\;\rm{M_{\odot}}$ follows a 9/4 scaling, whereas structures above this mass has a scaling of 4/3 i.e. $\rm{L_X \propto M^{4/3}}$ as expected from the self similarity of clusters discussed in the Section~\ref{self-sim}.
X-ray luminosity with virial temperature indicates a break at around $1.16\times10^7$ K i.e. 1 keV in B\'ezier curve. Power law when fitted to the hotter to cooler systems in Figure~\ref{xray-mass}, Panel~2, a break point at 1 keV has been observed. The hotter objects are found to follow a scaling of $\rm{L_x \propto T^3}$ exactly as observed in \cite{Allen_1998MNRAS}, whereas cooler objects show a scaling of $\rm{T^{3/2}}$. Further, entropy of the selected samples are plotted against virial temperature in Figure~\ref{xray-mass}, Panel 3. Slope of S-T curve has been found to be $\sim$1 for objects with temperature above 1 keV. But, if we consider objects below 1 keV, it becomes much flatter with a slope of 0.3. Fluctuations in the data below this temperature also indicate that they have non correlated entropy distribution. Strikingly, this also shows a power law type behaviour only above the temperature of 1 keV as in the case of X-ray luminosity.
\subsection{Baryon fraction evolution}\label{bar-frac}
Thermally interacting gas (baryons) that has very less fractional abundance in the universe, brings a huge change in observable energy budget of these massive structures. Hot baryons are a very crucial component that controls the X-ray emission as well as all kinds of non-thermal emissions. Thus, understanding the evolution of gas in these systems is very crucial for describing the hierarchical growth of structures in the universe. We have computed the fractions of baryons within the virial radius ($\rm{r_{200}}$) of the chosen structures spanning almost 3 order of mass (see Figure~\ref{baryon-frac}), Panel 1. The fitted line in Figure~\ref{baryon-frac} has two clear breaks, indicating a drastic change of properties in the systems across these points. The first knee of the curve is at $\sim 6-8\times 10^{12} \; \rm{M_{\odot}}$, while the other one is at around $6-8\times 10^{13}\; \rm{M_{\odot}}$. It has been observed that the baryon fraction in structures below mass $8\times 10^{13}\; \rm{M_{\odot}}$ has a very steep decrement and oscillatory in nature indicating a rapid change in its properties. Above this mass, baryon fraction almost gets stabilized at $\sim$ 14\% and a very flat slope has been observed.
\begin{figure*}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.2cm},clip]{Baryon_Fract_Mass_Bezier.pdf}\hspace{-0.4cm}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.2cm},clip]{TE_Mass_Bezier.pdf}\hspace{-0.4cm}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.2cm},clip]{Temp_Mass_Bezier.pdf}\\
\caption{Averaged baryon fraction, total energy and virial temperature computed upto $\rm{r_{200}}$ have been plotted against virial mass in the Panel 1, 2 \& 3 respectively and fitted with B\'ezier curve as mentioned in Section~\ref{self-sim}.}
\label{baryon-frac}
\end{figure*}
\subsection{Total Kinetic energy distribution}\label{tot-en}
Deciphering the kinetic energy in LSS needs a proper understanding of energy shared among baryons and DM. During structure formation, a large amount of binding energy converts to kinetic energy. Released binding energy is shared by the DM and baryons and changes the ICM dynamics significantly. Total kinetic energy in such a system is given by KE = $\rm{K_g + K_d}$ =$\frac{3}{2}\;\sigma^{2}_v \left(\rm{ M_g + M_d}\right) \;$, \citep{Diaz2010MNRAS}. Where, $\rm{M_g}$ and $\rm{M_d}$ are the gas and the dark matter mass respectively, while, ${\sigma}_v$ is the radial velocity dispersion assumed to be the same for the DM and the galaxies i.e baryons. But, as it has been observed in our study (Section~\ref{bar-frac}), baryon fraction across the knees varies largely and their thermal interaction should also be different, thus, approximation of same velocity dispersion for DM and baryon could lead to an erroneous result. We have thus computed baryon and DM velocity separately (i.e. $\sigma^{2}_b$ and $\sigma^{2}_{dm}$) and modified the above equation accordingly as $\rm{E}_{tot} = \left( \rm{M_g \sigma^{2}_b + M_d \sigma^{2}_{dm}} \right)$. Total energy of the system has been plotted in Figure~\ref{baryon-frac}, Panel 2, also indicate a similar dichotomy like in the other parameters studied in this paper. Strikingly, again the turning point is found to be at $\sim 8\times 10^{13}\; \rm{M_{\odot}}$.
\subsection{Self similarity in non-thermal properties}
Though, intra cluster medium (ICM) is dominated by thermal particles a non-negligible fraction of non-thermal cosmic-ray (CR) particles has also been observed. Shared ICM energy attains an equipartition at a scale larger than dissipation scale, if ICM gets enough relaxation time (i.e. $\epsilon_{therm} \sim \epsilon_{CR} \sim E_B \sim \epsilon_{turb}$ \citep{Longair1994}. The most effective mechanism active for particle acceleration in LSS is the diffusive shock acceleration (DSA, \cite{Drury1983RPPh}). The pool of energetic particles that exists in ICM due to AGN activity, star formation, supernova explosions etc. are injected at the shocks along with the thermally energised particles. Particles get accelerated by DSA to cosmic rays (CRs). CR flux is a strong function of Mach number and can be computed from the hydrodynamic parameters using the function obtained by \citet{Kang_2013ApJ} as $f_{CR}= \eta(\mathcal{M}) \times \frac{1}{2}\rho(\mathcal{M} c_{s})^3$. Where, cosmic ray acceleration fraction is a function of Mach number i.e. $\eta(\mathcal{M})$. From Figure~\ref{CRs-scale} Panel 1, we see that slope gets consistent only above the mass $8\times 10^{13}\; M_{\odot}$ and the value is $L_{CR} \propto M^{1.2}$ which has also been observed in other studies with cooling and feedback physics \citep{Pfrommer_2008MNRAS}. This scaling does not work for the lower masses. In the same figure, Panel 3, total velocity dispersion (DM plus baryon) against radius plot indicates that objects with virial radius 1 Mpc and above only show a power law relation with velocity dispersion or turbulent energy of the system, but smaller objects are not having any correlation with the velocity distribution.
\begin{figure*}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.2cm},clip]{lcr_mass_scaling.pdf}\hspace{-0.4cm}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.2cm},clip]{LCR_Mass_Bezier.pdf}\hspace{-0.4cm}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.2cm},clip]{Vrms_Rad_Bezier.pdf}\hspace{-0.4cm}
\caption{ In Panel 1, cosmic ray luminosity ($\rm{L_{CR}}$) has been plotted against virial mass and slopes are computed similar to Figure~\ref{xray-mass}. Similarly, B\'ezier curves are fitted to ($\rm{L_{CR}}$) vs mass and Total velocity dispersion (DM plus baryon) vs virial radius ($\rm{r_{200}}$) in Panel 2 \& 3 respectively.}
\label{CRs-scale}
\end{figure*}
\section{Numerical resolution study}\label{res-study}
Effect of numerical resolution on computed physical parameters is a crucial aspect to be considered in simulation studies. Though, the resolution of our simulations is not among the highest one's, our resolution tests show it is adequate for our study. The main runs that are used for this study are performed with the cosmological and simulation parameters described in section~\ref{sample} with 6 levels of total (uni-grid + AMR) refinement leading to a resolution of $\sim$30 kpc. Keeping all other parameters same, we have further simulated some of our objects with different AMR levels to achieve different levels of resolution and compared the physical parameters obtained from them. With 5, 6 and 7 levels of total refinement, we have reached upto $\sim$ 60, 30 and 15 kpc at the highest resolution level and used these three resolutions for testing the convergence of the results.
In fig~\ref{radial-res-stud1}, we have plotted radial variations of different physical parameters for a group ($\sim$ 10$^{13} \rm{M_{\odot}}$) and a cluster ($\sim$ 10$^{15} \rm{M_{\odot}}$) that are almost in relaxed phase. X-ray luminosity, gas temperature and entropy have been plotted against normalised virial radius for three different resolutions. It can be noticed that REFRES simulation is almost same as the HIGHRES resolution with very little deviation though, LOWRES data are little away. In panel 1, the X-ray luminosity has a little spatial variation that usually occurring due to resolution sensitivity of transient phenomena like the shocks. The overall value of the respective parameters though does not get effected much by the resolution. This results show that our simulated parameters are almost converging with resolution that we took as the reference set of simulations i.e. $\sim$30 kpc with 6 levels of refinement. For further confirmation, a general study has been done with our total sample set. In Fig~\ref{stat-res-stud2}, X-ray luminosity has been plotted against mass in wide range and at each point, standard error has been computed and plotted. It can be noticed that REFRES and HIGHRES data are significantly overlapping and mostly are within the error bars. We have also studied the break points for different resolutions and found almost no change in the break points as expected and discussed (See the Appendix, Figure~\ref{stat-res-stud4}~\&~\ref{stat-res-stud5}). Further, discussions about other parameters have been given in the Appendix~\ref{appen-res-study}. These studies confirm that resolution-wise our REFRES runs are adequate for our present work.
\begin{figure*}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.3cm},clip]{radial_prof_L_x.pdf}\hspace{-0.4cm}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.3cm},clip]{radial_prof_temperature.pdf}\hspace{-0.4cm}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.3cm},clip]{radial_prof_entropy.pdf}\hspace{-0.4cm}
\caption{X-ray luminosity, temperature and entropy have been plotted against normalized radius (normalized to $\rm{r_{200}}$) in the Panel 1,2 \& 3 respectively for a galaxy cluster ($\sim$ 10$^{15} \rm{M_{\odot}}$) and a group ($\sim$ 10$^{13} \rm{M_{\odot}}$) and for three resolutions namely LOWRES, REFRES and HIGHRES (colours as indicated in the legend). }
\label{radial-res-stud1}
\end{figure*}
\begin{figure*}
\includegraphics[width=4.6cm,trim={0.1cm 0.1cm 0.2cm 0.3cm},clip]{R200_LX_Mass.pdf}\hspace{-0.4cm}
\includegraphics[width=4.6cm,trim={0.1cm 0.1cm 0.2cm 0.3cm},clip]{R200_Entropy_Mass_full_data.pdf}\hspace{-0.4cm}
\includegraphics[width=4.6cm,trim={0.1cm 0.1cm 0.2cm 0.3cm},clip]{R200_Entropy_2Temp.pdf}\hspace{-0.4cm}
\includegraphics[width=4.2cm,trim={0.1cm 0.1cm 0.2cm 0.3cm},clip]{R200_LX_1Temp.pdf}\hspace{-0.4cm}
\caption{Statistical average value and standard error of X-ray Luminosity and entropy from our sample data set has been plotted against virial mass ($\rm{M_{200}}$) in the Panel 1 \& 2 respectively for all three resolutions. Similarly, X-ray luminosity and entropy has been plotted against temperature in the Panel 3 \& 4 respectively for all three resolutions.}
\label{stat-res-stud2}
\end{figure*}
\section{Discussions}\label{discus}
Our simulations with coolSF model has been compared with the available observed data (see Section~\ref{x-ray-obs}). It can be noticed that, previous works, done with SPH codes and using similar physics as coolSF, could not match the observed data {to that extent as ours} \citep{LeBrun_2014MNRAS,McCarthy2010MNRAS}. Better agreement of our data with the observed data can possibly be attributed to difference in reproduction of baryon fraction as well as entropy and better gas mixing in ENZO AMR (PPM) code over SPH code \citep{Hubber_2013MNRAS,Valdarnini_2011A&A,O'Shea_2005ApJS}. AMR schemes are also known for better shock capturing and appropriate shock heating. We have implemented shock refinement very carefully to resolve the whole cluster volume as explained in Section~\ref{sample}. All these factors may have contributed in production of apparently different results from coolSF in ENZO compared to SPH. So, with the chosen simulation parameters for this study, ENZO-AMR code has been able to adequately match the observations without AGN feedback model. Nevertheless, AGN feedback is important for correct core entropy computation and inclusion of AGN feedback and other additional physics may effect these results and thus becomes an interesting topic for future research.
Further, we have computed $L_X$ - T, $\rm{L_X}$ -M, S-T scaling laws for simulated LSS with mass spanning from $5\times10^{12}\; \rm{M_{\odot}}$ to 2.5$\times 10^{15}\; \rm{M_{\odot}}$. Scaling of total energy, baryonic fraction and non-thermal cosmic ray luminosity with mass have also been worked out. All these studies are consistently indicating a breaking away point at about $8 \times 10^{13} \; \rm{M_{\odot}}$. This has enabled us to conclude that the LSS consists of two populations, one is above and another is below this breaking point. Henceforth, we will call objects with mass less than $8\times 10^{13}\;\rm{M_{\odot}}$ as `Galaxy groups' or `galaxy cabal' and with higher mass as 'Galaxy clusters'
\begin{figure*}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.2cm},clip]{hist_cluster_group_normalized.pdf}\hspace{-0.4cm}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.2cm},clip]{histogram_baryonFrac_normalized.pdf}\hspace{-0.4cm}
\includegraphics[width=6cm,trim={0.1cm 0.1cm 0.2cm 0.2cm},clip]{lcr_te_r_200_spl_8x.pdf}\hspace{-0.4cm}
\caption{In Panel 1 \& 2, normalised histogram of frequency of occurrence have been plotted for virial ratio and baryon fraction for the `galaxy groups' (light gray) and `galaxy clusters' (black). Panel 3, shows the cosmic ray luminosity vs total energy plot fitted for `galaxy clusters' and `groups'.}
\label{virial-ratio}
\end{figure*}
Galaxy clusters are mostly found to be relaxed systems and considered to be virialized \citep{Sarazin2003PhPl}. For a self gravitating system like the clusters, virial ratio can be expressed as $\xi$ = $\frac{(U_{int} + U_{ext}-E_s)}{2*KE} $ \citep{Davis_2011MNRAS}. Where, $U_{int}$ is the potential energy of the studied system and $U_{ext}$ is due to the mass outside the system radius (here, $r_{200}$) but, whose tidal effect can be felt. $E_s$ is the surface pressure term and KE is the kinetic energy of the system. For a perfectly virialised object, the ratio $\xi$ should be unity. We have computed the virial ratio for all objects in our sample set and plotted two different normalised histograms for the `galaxy groups' and `galaxy clusters' in Figure~\ref{virial-ratio}, Panel~1. Median and statistical (Lorentzian) peak of the `galaxy clusters' came out to be very close to unity (1.026 \& 1.019 respectively), indicating perfect virialization for most of the clusters. But, the median and peak value of virial ratio for `galaxy groups' are 0.791 \& 0.797 respectively i.e. far away from unity. This shows, groups are unstable and in hydrostatic disequilibrium, unlike the clusters and strongly supports the break point calculated from our study.
Further, we have plotted baryon fraction histogram in Figure~\ref{virial-ratio}, Panel~2, which clearly shows bimodal distribution with two Lorentzian peaks at $\sim 10.18\%$ and $\sim 13.77\%$ and distinctly separated median at 10.14 \& 13.81 for groups and clusters respectively, just like the virial ratio distribution. We have also plotted the cosmic ray luminosity as a function of total kinetic energy of the systems. Groups and Clusters are observed to follow two different evolutionary track as they fall into two parallel fitting lines separated by at least an order in energy (Figure~\ref{virial-ratio} Panel 3).
\section{Summary}\label{conclude}
This research studies most of the possible scaling laws in thermal and non-thermal energies for the large scale objects in the framework of implemented baryon physics (see Section~\ref{sample}). The main takeaway points from this study are as follows.
$\#$ We could define clear distinguishing parameters for classifying `galaxy groups' and `clusters' for the first time. Strikingly, we found that cluster self similarity scales applied to the structures deviates away below a particular break away point in mass at $\sim 8\times 10^{13}\; M_{\odot}$ in all the studied parameters such as X-ray luminosity, temperature, baryon fractions and even in non-thermal cosmic ray luminosity.
$\#$ We also report that cluster properties deviates at temperature $1.16\times10^7$ K i.e. $\sim$1 keV and at a radius of $\sim$1 Mpc. So, for the first time, we are able to give a strong characteristic numbers to separate `galaxy groups' from the clusters and this study presents `galaxy groups' as a unique object in the structural hierarchy.
$\#$ CRs luminosity slope shown to be flatter in groups, indicating more non-thermal energy in them. This study also shows a very high level velocity dispersion i.e. turbulence in galaxy groups making them a test bed for the study of magnetisation and non-thermal emissions.
$\#$ From the virialization study it is established that `galaxy groups' are far away from the virialization and any estimation of physical parameters based on virial theorem would certainly go wrong.
$\#$ Baryon fraction of galaxy groups are very low, and have large fluctuations in the values among the groups making them very unstable in nature. This opens up the door for research on modelling galaxy groups differently.
\section*{Acknowledgements}
This project is funded by DST INSPIRE Faculty (IFA-12/PH-44) and DST-SERB Fast Track scheme for young scientists, Grant No. SR/FTP/PS-118/2011. We are thankful to the Inter-University Centre for Astronomy and Astrophysics (IUCAA) for providing the HPC facility. RSJ would like to thank IUCAA for providing the research facilities as a visiting student. Computations described in this work were performed using the publicly-available \texttt{Enzo} code (http://enzo-project.org) and data analysis is done with the yt-tools (http://yt-project.org/). Authors would also like to thank the editor and the anonymous referee for their critical comments that has definitely helped in improving the content of this paper.
| 001de813250da623ec53a2639e5102519935da0a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction\label{sec:intro}}
The collisional relaxation of a temperature anisotropy is a canonical example of energy transport in plasmas~\cite{ichi:70}. It serves as a test for transport theories, as well as playing an essential role in many experiments~\cite{hyat:87,beck:92,dubi:05,ande:09} and natural occurring plasmas~\cite{maru:11}. For instance, plasmas that are preferentially heated, or cooled, in one direction will form a temperature anisotropy. Magnetized plasmas often have different energy confinement times either along or against the magnetic field, and can form a temperature anisotropy as a result~\cite{ott:17}. It is often essential to know the rate at which this temperature anisotropy relaxes in order to accurately model these plasmas. Here, we use molecular dynamics (MD) simulations to test theories of the collisional temperature anisotropy relaxation rate. The work concentrates on the unmagnetized one-component plasma (OCP)~\cite{baus:80} and considers a broad range of coupling parameters. Although this concentrates on unmagnetized plasmas, the results are also expected to apply to magnetized plasmas as long as $\omega_c/\omega_p \ll 1$, where $\omega_c = eB/m$ is the gyrofrequency and $\omega_p = \sqrt{4\pi e^2n/m}$ is the plasma frequency~\cite{ichi:70,dong:13}. Examples of systems in which moderate to strong coupling arises and magnetization may lead to temperature anisotropies include magnetized ultracold neutral plasmas~\cite{zhan:08} and in magnetized inertial confinement fusion~\cite{gome:14}.
The initial distribution function is assumed to have the form
\begin{equation}
f = \frac{n}{\pi^{3/2} v_{T\parallel} v_{T\perp}^2} e^{-v_\parallel^2/v_{T\parallel}^2} e^{-v_\perp^2/v_{T\perp}^2} \label{eq:amax}
\end{equation}
where $n$ is the number density, $v_{T\parallel}^2 = 2k_BT_\parallel/m$ and $v_{T\perp}^2 = 2k_BT_\perp/m$. Here, $T_\parallel$ and $T_\perp$ are the parallel and perpendicular temperatures, which are related to the total temperature by $T = (T_\parallel + 2 T_\perp)/3$. In weak and moderately coupled plasmas, the rate of relaxation of the temperature anisotropy can be modeled from the $v_\parallel^2$ or $v_\perp^2$ moment of a kinetic equation $df/dt = C(f)$. This provides the anisotropy evolution equation
\begin{equation}
\frac{dT_\perp}{dt} = - \frac{1}{2} \frac{d T_\parallel}{dt} = - \nu (T_\perp - T_\parallel) \label{eq:nrl}
\end{equation}
where
\begin{equation}
\nu = \frac{m}{n} \int d^3v\, v_\perp^2 \frac{C(f)}{T_\perp - T_\parallel} = - \frac{1}{2} \frac{m}{n} \int d^3v\, v_\parallel^2 \frac{C(f)}{T_\perp - T_\parallel} \label{eq:nu_def}
\end{equation}
is the collisional equipartition rate and $C(f)$ is the collision operator.
A consequence of the underlying assumptions of kinetic theories of weak or moderately coupled plasmas is that kinetic energy is a conserved quantity throughout the relaxation $dT/dt=0$. However, this assumption becomes suspect at strong coupling where the potential energy of interactions in the system is high enough that the exchange between kinetic and potential energy may significantly influence the relaxation processes. The MD simulation results shown here reveal that, indeed, kinetic energy is well conserved at weak and moderate coupling, but that small dynamic oscillations in the kinetic energy, due to exchange with potential energy, are observed at strong coupling. Recent MD simulations have also shown that external energy perturbations can lead to the spontaneous development of temperature anisotropies~\cite{ott:17}.
Each of the theories that will be discussed assume kinetic energy conservation, and thus satisfy Eq.~(\ref{eq:nrl}). This is a set of two coupled ordinary differential equations that can be solved for $T_\parallel(t)$ and $T_\perp(t)$. For comparing different collision theories, it is convenient to note that the equations can be decoupled to a single equation by defining
\begin{equation}
\label{eq:Adef}
A \equiv \frac{T_\perp}{T_\parallel} - 1
\end{equation}
as a measure of the temperature anisotropy. Note that $T>0$ implies $A>-1$. In terms of $A$, Eq.~(\ref{eq:nrl}) is
\begin{equation}
\frac{dA}{dt} = - \nu (3A + 2 A^2) . \label{eq:dAdt}
\end{equation}
Thus, $A(t)$ along with $dT/dt = 0$ provides the solution for both the parallel $T_\parallel(t) = 3T/(3+2A(t))$ and perpendicular $T_\perp(t) = 3T (1+A(t))/(3+2A(t))$ temperatures.
Here, we compute $\nu$ from four different collision operators and compare the results with MD simulations. Simulations were carried out spanning a wide range of coupling strengths. The coupling strength of the OCP is described by the Coulomb coupling parameter~\cite{baus:80}
\begin{equation}
\Gamma \equiv \frac{e^2/a}{k_BT} ,
\end{equation}
where $e$ is the elementary electron charge, $T$ is the total temperature and $a=(3/4\pi n)^{1/3}$ is the average interparticle spacing. Two of the collision operators we compare with apply to weakly coupled plasmas ($\Gamma \ll 1$): the Landau equation~\cite{land:36} and the Lenard-Balescu equation~\cite{lena:60}. The distinguishing feature of these is that the Lenard-Balescu equation treats dynamic (velocity-dependent) screening, whereas the Landau equation is based on a static screening model. We find that both give very similar predictions for $\Gamma \lesssim 0.1$ and that both models rapidly break down for larger $\Gamma$ values. Thus, there is no significant contribution associated with dynamic screening for this process over the range of coupling strength in which these theories are applicable. Physically, this is because temperature anisotropy relaxation is a thermal processes associated with particles in the thermal part of the distribution, which do not have significant velocity-dependent components to their screening clouds. We also compare with two theories that extend these collision operators to strong coupling: effective potential theory (EPT)~\cite{baal:13,baal:14,baal:15} and generalized Lenard-Balescu theory (GLB)~\cite{tana:86,ichi:92,bene:12}. Again, GLB includes dynamic effects in the long distance component of the linear response function, whereas EPT is based on a static description. We find that each of these give similar results that are in agreement with MD from weak to moderate coupling, but that EPT is able to extend to significantly larger $\Gamma$ values than the GLB theory. Again, the conclusion is that dynamic screening does not significantly influence the anisotropy relaxation rate.
The main approximation in this analysis is that the distribution function maintains the anisotropic Maxwellian form from Eq.~(\ref{eq:amax}) throughout the relaxation, with only the magnitudes of $T_\parallel$ and $T_\perp$ evolving. It is known from other transport processes, such as self-diffusion of the OCP, that deviations from Maxwellian distributions can contribute corrections on the order of $20\%$ in the weakly coupled limit, but are typically insignificant in the strongly coupled limit~\cite{ferz:72,baal:14}. To the best of our knowledge, a perturbation expansion of the distribution function, such as the Chapman-Enskog~\cite{ferz:72,chap:70} or Grad~\cite{grad:58} expansions for hydrodynamic transport coefficients, has not yet been developed for temperature anisotropy relaxation. This aspect of the analysis will be discussed further in Sec.~\ref{sec:results}.
In addition to comparing with MD simulations, explicit expressions are derived for $\nu$ from each collision operator. Various limits of each are also considered, including early time, weak anisotropy and limits connecting the theories based on static or dynamic screening. The results quantify the accuracy of each approach, and the various reduced models in asymptotic limits. This provides useful information for modeling a variety of different plasmas with temperature anisotropies.
Finally, time-dependent temperature profiles from the MD simulations are used to highlight interesting features of correlation effects that arise at strong coupling, which are not described by any of the theories considered. These include delays in the temperature relaxation at early times, substantial oscillations in the temperature profiles in time, and small violations of the kinetic energy conservation assumption ($dT/dt=0$), indicating correlation between kinetic and internal energies during the relaxation process. These aspects are discussed in Sec.~\ref{sec:results}.
\section{Weakly Coupled Plasma Theories\label{sec:wc}}
\subsection{Landau theory}
Plasma transport theory is typically derived from the Landau kinetic equation~\cite{land:36}, which for the OCP is
\begin{equation}
C_\textrm{L} = - \frac{2\pi e^4 \ln \Lambda}{m^2} \frac{\partial}{\partial \mathbf{v}} \cdot \int d^3v^\prime \frac{u^2 \ensuremath{\mathcal{I}} - \mathbf{u} \mathbf{u}}{u^3} \cdot \biggl[ f(\mathbf{v}) \frac{\partial f(\mathbf{v}^\prime)}{\partial \mathbf{v}^\prime} - f(\mathbf{v}^\prime) \frac{\partial f(\mathbf{v})}{\partial \mathbf{v}} \biggr] \label{eq:c_landau}
\end{equation}
where $\mathbf{u} \equiv \mathbf{v} - \mathbf{v}^\prime$ and $\ln \Lambda$ is the Coulomb logarithm. In Landau's theory $\Lambda = b_\textrm{max}/b_\textrm{min}=\lambda_D/r_L$ is the ratio of the maximum and minimum length scales over which particles interact. The maximum is determined by the Debye screening cloud $\lambda_D = \sqrt{k_BT/(4\pi e^2n)}$ and the minimum by the distance of closest approach $r_L = e^2/(2k_BT)$. In terms of the Coulomb coupling parameter $\Lambda = 2/(\sqrt{3} \Gamma^{3/2})$. The standard plasma physics result for the temperature anisotropy relaxation rate follows from substituting the distribution from Eq.~(\ref{eq:amax}) into Eq.~(\ref{eq:nu_def}) with Eq.~(\ref{eq:c_landau}) as the collision operator. This leads to Eq.~(\ref{eq:nrl}) with the equipartition rate~\cite{ichi:70,huba:16}
\begin{equation}
\frac{\nu_{\textrm{L}}}{\bar{\nu}} = \frac{(1+\frac{2}{3}A)^{3/2}}{A^{2}} \biggl[ - 3 + (A+3) \frac{\arctan (\sqrt{A})}{\sqrt{A}} \biggr] \ln \Lambda \label{eq:nu_nrl}
\end{equation}
where
\begin{equation}
\bar{\nu} = 2 \sqrt{\frac{\pi}{m}} \frac{ne^4}{(k_BT)^{3/2}} = \omega_p \sqrt{\frac{3}{4\pi}} \Gamma^{3/2}
\end{equation}
is a reference collision rate. Equation~(\ref{eq:nu_nrl}) is valid for any sign of $A$, but for $A<0$ is often written in the equivalent form by replacing $\arctan (\sqrt{A})/\sqrt{A} = \textrm{arctanh} ( \sqrt{-A})/\sqrt{-A}$~\cite{huba:16}.
\begin{figure}[t]
\includegraphics[width=75mm]{fig1a.pdf}
\includegraphics[width=75mm]{fig1b.pdf}
\caption{Time evolution of $T_\parallel/T$ and $T_\perp/T$ computed from the Landau theory Eq.~(\ref{eq:nu_b0}) (solid lines), along with the first order (dash-dotted lines) and second order (dashed lines) terms of the small anisotropy expansion from Eqs.~(\ref{eq:AL_1}) and (\ref{eq:AL_2}) respectively: (a) $A_o = 0.2$ and (b) $A_o = 1$. }
\label{fg:T_t_landau}
\end{figure}
Equation~(\ref{eq:nu_nrl}) was derived from Eq.~(\ref{eq:c_landau}) without explicit reference to the magnitude of the anisotropy. In the weak anisotropy limit $A \ll 1$, Eqs.~(\ref{eq:dAdt}) and (\ref{eq:nu_nrl}) reduce to $dA/d\tilde{t} = -\frac{4}{5} A - \frac{68}{105} A^2 + \mathcal{O}(A^3)$, where $\tilde{t} \equiv t \bar{\nu} \ln \Lambda$. Thus, to first order in the anisotropy expansion the relaxation is exponential
\begin{equation}
A_1(t) = A_o \exp \biggl(- \frac{4}{5} \tilde{t} \biggr) \label{eq:AL_1}
\end{equation}
and at second order has the form
\begin{equation}
A_2(t) = \frac{21 A_o}{\exp (\frac{4}{5} \tilde{t}) (17A_o + 21)-17A_o} . \label{eq:AL_2}
\end{equation}
Figure~(\ref{fg:T_t_landau}) shows results of the temperature evolution for initial anisotropies of $A_o = 0.2$ and $A_o=1$ predicted from Eq.~(\ref{eq:nu_nrl}). Also shown are the small anisotropy expansions from Eqs.~(\ref{eq:AL_1}) and (\ref{eq:AL_2}). This shows that for a weak initial anisotropy, the simple expression obtained from the lowest order term in the expansion, Eq.~(\ref{eq:AL_1}), provides an accurate approximation of the full result. As $A$ increases to 1, deviations from the exponential form become important, but these are accurately modeled by Eq.~(\ref{eq:AL_2}).
\subsection{Lenard-Balescu theory}
An alternative to Landau's kinetic theory for weakly coupled plasmas is the Lenard-Balescu equation~\cite{lena:60}, which for the OCP is
\begin{equation}
C_\textrm{LB} = - \frac{2 e^4}{m^2} \frac{\partial}{\partial \mathbf{v}} \cdot \int d^3v^\prime \int d^3k \frac{\delta (\mathbf{k} \cdot \mathbf{u})}{|\hat{\varepsilon} (\mathbf{k}, \mathbf{k} \cdot \mathbf{v})|^2} \frac{\mathbf{k} \mathbf{k}}{k^4} \cdot \biggl[ f(\mathbf{v}) \frac{\partial f(\mathbf{v}^\prime)}{\partial \mathbf{v}^\prime} - f(\mathbf{v}^\prime) \frac{\partial f(\mathbf{v})}{\partial \mathbf{v}} \biggr] . \label{eq:c_lb}
\end{equation}
The primary distinction between this and Landau's collision operator is that the long-range screening is self-consistently accounted for via the linear dielectric response function $\hat{\varepsilon}(\mathbf{k}, \mathbf{k} \cdot \mathbf{v})$. In contrast to the static Debye-H\"{u}ckel screening assumed in the Landau equation, this is a generalized theory that can, in principle, account for velocity dependence (aka dynamics) of the screening.
Following the same procedure as the previous section by using the distribution from Eq.~(\ref{eq:amax}) in Eq.~(\ref{eq:nu_def}) with Eq.~(\ref{eq:c_lb}) as the collision operator provides
\begin{equation}
\nu_{\textrm{LB}} = \frac{n e^4}{\pi} \int d^3k \int_{-\infty}^\infty d\omega \frac{1}{k^4} \frac{1}{|\hat{\varepsilon}(\mathbf{k}, \omega)|^2} \frac{k_\parallel^2 k_\perp^2}{(k_\parallel^2 k_BT_\parallel + k_\perp^2 k_BT_\perp)^2} \exp \biggl( - \frac{m \omega^2}{k_\parallel^2 k_BT_\parallel + k_\perp^2 k_BT_\perp} \biggr) \label{eq:nu_b0}
\end{equation}
where
\begin{equation}
\hat{\varepsilon}(\mathbf{k},\omega) = 1 + \frac{\omega_p^2}{k^2} \int d^3v \frac{\mathbf{k} \cdot \partial f/\partial \mathbf{v}}{\omega - \mathbf{k} \cdot \mathbf{v}} = 1 - \frac{\omega_p^2}{k_\parallel^2 v_{T\parallel}^2 + k_\perp^2 v_{T\perp}^2} Z^\prime \biggl( \frac{\omega}{\sqrt{k_\parallel^2 v_{T\parallel}^2 + k_\perp^2 v_{T\perp}^2}} \biggr) . \label{eq:ep_hat}
\end{equation}
The expression following the second equality in Eq.~(\ref{eq:ep_hat}) follows from inserting the anisotropic Maxwellian distribution from Eq.~(\ref{eq:amax}), $Z$ is the plasma dispersion function~\cite{frie:61} and the ``prime'' denotes the first derivative with respect to the argument.
Lenard-Balescu theory contains an ultraviolet divergence because it does not self-consistently account for the close interaction of colliding particles. This physics is imposed by limiting the minimum interaction length scale to be the classical distance of closest approach, or equivalently, the maximum wavenumber to be $k_{\textrm{max}} = 1/b_{\textrm{min}}=2k_BT/e^2$. Representing $k$ in a spherical coordinate system
$\mathbf{k} = k \cos \theta \sin \phi \hat{x} + k \sin \theta \sin \phi \hat{y} + k \cos \phi \hat{z}$, Eq.~(\ref{eq:nu_b0}) can be expressed as
\begin{equation}
\frac{\nu_{\textrm{LB}}}{\bar{\nu}} = \biggl( \frac{1+\frac{2}{3}A}{1+A} \biggr)^{3/2} \frac{1}{\sqrt{\pi}} \int_{-\infty}^\infty dx e^{-x^2} \int_{-1}^1 dy \frac{y^2 (1-y^2)}{(y^2 \tilde{A} + 1)^{3/2}} \int_0^{\bar{k}_{\max}} \frac{d \bar{k}}{\bar{k}} \biggl| 1 + \frac{\Phi(x)}{\bar{k}^2 (y^2 \tilde{A} + 1)} \biggr|^{-2} . \label{eq:nu_mid2}
\end{equation}
Here, the following variables have been defined:
$y=\cos\phi$, $\bar{k} = k \lambda_{D\perp}$, $x = (\omega/\omega_p)/\sqrt{\bar{k}_\parallel^2 r + \bar{k}_\perp^2}$,
$r = T_\parallel/T_\perp=1/(1+A)$,
$\tilde{A} \equiv -A/(1+A)$, and $\Phi(x) \equiv - \frac{1}{2} Z^\prime (x/\sqrt{2})$, where $\lambda_{D\perp}^2 \equiv k_BT_\perp/(4\pi e^2 n)$ is a Debye length associated with the perpendicular temperature.
Equation~(\ref{eq:nu_mid2}) can be further reduced by carrying out the $\bar{k}$ integral, leading to
\begin{align}
\frac{\nu_{\textrm{LB}}}{\bar{\nu}} &= \biggl( \frac{1+\frac{2}{3}A}{1+A} \biggr)^{3/2} \frac{2}{\sqrt{\pi}} \int_{-\infty}^\infty dx e^{-x^2} \int_{0}^1 dy \frac{y^2 (1-y^2)}{(y^2 \tilde{A} + 1)^{3/2}} \biggl\lbrace \frac{1}{4} \ln \biggl( \frac{|\bar{k}_{\max}^2 (y^2 \tilde{A}+1) + \Phi|^2}{|\Phi|^2} \biggr) \label{eq:nu_b0_full} \\ \nonumber
& - \frac{1}{2} \frac{\Phi_R}{\Phi_I} \biggl[ \arctan \biggl( \frac{\bar{k}_{\max}^2 (y^2 \tilde{A} +1) + \Phi_R}{\Phi_I} \biggr) - \arctan \biggl( \frac{\Phi_R}{\Phi_I} \biggr) \biggr] \biggr\rbrace ,
\end{align}
where
\begin{equation}
\bar{k}_{\max} = \frac{\lambda_{D\perp}}{b_{\min}} = \frac{\sqrt{k_BT_\perp/(4\pi e^2n)}}{e^2/(2k_BT)} = \sqrt{\frac{1+A}{1+\frac{2}{3}A}} \frac{2}{\sqrt{3} \Gamma^{3/2}},
\end{equation}
$\Phi_R = \textrm{Re} \lbrace \Phi \rbrace$ and $\Phi_I = \textrm{Im} \lbrace \Phi \rbrace$. Equation~(\ref{eq:nu_b0_full}) follows directly from the collision operator, Eq.~(\ref{eq:c_lb}), assuming only that the distribution function remains an anisotropic Maxwellian throughout the evolution. This will be compared directly with the Landau equation result from Eq.~(\ref{eq:nu_nrl}), but first the following subsections discuss how the integrals in Eq.~(\ref{eq:nu_b0_full}) can be simplified via accurate approximations.
\subsubsection{Isotropic Screening Approximation \label{sec:isa}}
Note that the wavenumber ($\bar{k}$) and azimuthal ($y$) integrals in Eq.~(\ref{eq:nu_mid2}) would be decoupled if not for the anisotropy of the screening (i.e., take $\tilde{A}=0$ in the last term). This feature arises in the dielectric function of Eq.~(\ref{eq:ep_hat}) because the temperature anisotropy leads to a different screening length in the parallel and perpendicular directions. However, the distance of closest approach, which is also temperature dependent, must be added in an ad-hoc way in Lenard-Balescu theory, so it is not clear what the overall anisotropy dependence of this term should be. In any case, since the dependence of transport rates on the screening length is logarithmic in the weakly coupled limit, it has a negligible influence if the coupling parameter $\Gamma$ is sufficiently small. If the screening is assumed isotropic, one can connect the results from the Landau equation based theory, Eq.~(\ref{eq:nu_nrl}), with the results from the Lenard-Balescu equation based theory via the Coulomb logarithm. Taking $\tilde{A}=0$ in the screening term of Eq.~(\ref{eq:nu_mid2}), or equivalently in the terms in curly braces in Eq.~(\ref{eq:nu_b0_full}), we note that
\begin{equation}
2 \int_{0}^1 dy \frac{y^2 (1-y^2)}{(y^2 \tilde{A} + 1)^{3/2}} = \frac{1}{r^{3/2}} \frac{1}{A^2} \biggl[-3 + (A+3) \frac{\arctan (\sqrt{A})}{\sqrt{A}} \biggr] .
\end{equation}
In this isotropic screening limit, the Lenard-Balescu result from Eq.~(\ref{eq:nu_b0_full}) can be written in the same form as the Landau result from Eq.~(\ref{eq:nu_nrl})
\begin{equation}
\frac{\nu_{\textrm{LB,I}}}{\bar{\nu}} = \frac{(1+\frac{2}{3}A)^{3/2}}{A^2} \biggl[ -3 + (A+3) \frac{\arctan (\sqrt{A})}{\sqrt{A}} \biggr] \Xi_{\textrm{LB,I}} \label{eq:nu_b0_approx}
\end{equation}
but where the Coulomb logarithm is replaced by
\begin{equation}
\Xi_{\textrm{LB,I}} = \frac{1}{2} \frac{1}{\sqrt{\pi}} \int_{-\infty}^\infty dx e^{-x^2} \biggl\lbrace \ln \biggl( \frac{|\bar{k}_{\max}^2 + \Phi |}{|\Phi |} \biggr) - \frac{\Phi_R}{\Phi_I} \biggl[ \arctan \biggl( \frac{\bar{k}_{\max}^2 + \Phi_R}{\Phi_I} \biggr) - \arctan \biggl( \frac{\Phi_R}{\Phi_I} \biggr) \biggr] \biggr\rbrace . \label{eq:xi_d}
\end{equation}
This is an effective Coulomb logarithm that includes dynamic screening (via the $\Phi$ terms), as well as an $\mathcal{O}(1)$ term in addition to the logarithmic term. Section~\ref{sec:wc_comp} will show that the latter contribution is significant only when $\Gamma \gtrsim 0.1$, which also happens to be where the theory breaks down due to the onset of strong correlations.
\subsubsection{Static Screening Limit}
The primary difference between the Landau and Lenard-Balescu based theories is the role of dynamic screening. To quantify the role of dynamic screening within the Lenard-Balescu theory, Eq.~(\ref{eq:nu_b0_full}), we compare with the static screening limit. In this limit, $\Phi(x) = \Phi(0)$, where $\Phi_R(0) = 1$, $\Phi_I(0) = 0$ and the $x$ integral in Eq.~(\ref{eq:nu_b0_full}) can be evaluated explicitly, leading to
\begin{align}
\label{eq:nu_b0_s}
\frac{\nu_{\textrm{LB,s}}}{\bar{\nu}} &= 2 \biggl(\frac{1+\frac{2}{3}A}{1+A} \biggr)^{3/2} \int_{0}^1 dy \frac{y^2 (1-y^2)}{(y^2 (r-1) + 1)^{3/2}} \biggl[ \ln \sqrt{\bar{k}_{\max}^2 (y^2 (r-1)+1) + 1} \\ \nonumber
&- \frac{1}{2} \frac{\bar{k}_{\max}^2(y^2(r-1)+1)}{1 + \bar{k}_{\max}^2(y^2(r-1)+1)} \biggr] .
\end{align}
This is compared with the full expression from Eq.~(\ref{eq:nu_b0_full}) in Fig.~\ref{fg:nu_wc_comp} (discussion in Sec.~\ref{sec:wc_comp}).
\subsubsection{Debye-H\"{u}ckel Screening\label{sec:dh}}
Finally, to connect the Lenard-Balescu and Landau results, the Deybe-H\"{u}ckel screening limit is considered. The linear dielectric corresponding to Debye-H\"{u}ckel screening is $\hat{\varepsilon} = 1 + 1/(k \lambda_D)^2 = 1 + 3/(\bar{k}^2(2+r))$. Using this in Eq.~(\ref{eq:nu_b0}) leads to the same form for the relaxation rate as Eqs.~(\ref{eq:nu_nrl}) and (\ref{eq:nu_b0_approx}),
\begin{equation}
\label{eq:nu_b0_dh}
\frac{\nu_{\textrm{DH}}}{\bar{\nu}} = \frac{(1+\frac{2}{3}A)^{3/2}}{A^2} \biggl[ -3 + (A+3) \frac{\arctan (\sqrt{A})}{\sqrt{A}} \biggr] \Xi_\textrm{DH}
\end{equation}
but where
\begin{equation}
\Xi_\textrm{DH} = \ln \bigl( \sqrt{\Lambda^2 + 1} \bigr) - \frac{1}{2} \frac{\Lambda^2}{\Lambda^2 + 1} . \label{eq:xi_s}
\end{equation}
Note that for $\Lambda \gg 1$, $\Xi_{\textrm{DH}} \rightarrow \ln \Lambda + \mathcal{O}(1)$, returning the conventional result from Eq.~(\ref{eq:nu_nrl}).
\subsection{Comparison of weakly coupled plasma theories \label{sec:wc_comp}}
\begin{figure}[t]
\includegraphics[width=75mm]{fig2a.pdf}
\includegraphics[width=73mm]{fig2b.pdf}
\caption{All figures use $T_\perp/T_\parallel = 0.8$ ($A=-0.2$): (a) Relaxation rate computed from the Landau theory from Eq.~(\ref{eq:nu_nrl}) (green solid line) and the Lenard-Balescu theory from Eq.~(\ref{eq:nu_b0_full}) (black solid line). Also shown are various limits of the Lenard-Balescu result: the isotropic screening result from Eq.~(\ref{eq:nu_b0_approx}) (red dashed line), the static screening approximation from Eq.~(\ref{eq:nu_b0_s}) (magenta dash-dotted line), and the Debye-H\"{u}ckel screening model from Eq.~(\ref{eq:nu_b0_dh}) (blue dotted line). (b) Ratio of approximate solutions of the Lenard-Balescu result from Eqs.~(\ref{eq:nu_b0_approx}), (\ref{eq:nu_b0_s}) and (\ref{eq:nu_b0_dh}) to the full numerical solution from Eq.~(\ref{eq:nu_b0_full}).}
\label{fg:nu_wc_comp}
\end{figure}
Figure~\ref{fg:nu_wc_comp} shows a comparison of the results of the weakly coupled theories described in this section. Panel (a) shows that both the Landau result from Eq.~(\ref{eq:nu_nrl}) and the Lenard-Balescu result from Eq.~(\ref{eq:nu_b0_full}) give essentially indistinguishable predictions for $\Gamma \lesssim 0.1$ at $A=-0.2$ (a weak temperature anisotropy $T_\perp/T_\parallel = 0.8$). This is because dynamic screening leads to a $\mathcal{O}(1)$ correction to the Coulomb logarithm; only super-thermal particles exhibit significant dynamic screening, but this transport process is controlled by thermal particles. Since $\ln \Lambda$ becomes large in the weakly coupled regime, any contribution from dynamic screening is inconsequential. For $\Gamma \gtrsim 0.1$ the two results differ, and the predictions of the Landau theory becomes negative near $\Gamma \simeq 1$. However, we will see in the next section that neither of these theories are accurate for $\Gamma \gtrsim 0.1$ because correlation effects onset, which require a strongly coupled plasma theory to describe. Thus, there seems to be no advantage to using the comparatively complicated Lenard-Balescu theory to describe this process. Figure~\ref{fg:nu_wc_comp}b shows the ratio of each of the approximations of the Lenard-Balescu result with the full numerical result. The results of the static screening model, and the Debye-H\"{u}ckel screening model are within a few percent of the full results for $\Gamma \lesssim 0.1$; further emphasizing the conclusion that dynamic screening does not significantly influence this process. The isotropic screening assumption is excellent for $\Gamma \lesssim 1$, beyond which the Lenard-Balescu theory is not expected to be accurate. The strongly coupled regime requires a theory that can account for correlation effects, which is the topic of the next section.
\begin{figure}[t]
\sidecaption
\includegraphics[width=80mm]{fig3.pdf}
\caption{Relaxation rate from the Landau theory of Eq.~(\ref{eq:nu_nrl}) (solid line) and the Lenard-Balescu theory of Eq.~(\ref{eq:nu_b0_full}) (dashed lines) as a function of the anisotropy parameter $A$. The linear expansion of Eq.~(\ref{eq:nu_nrl}), $\nu/(\bar{\nu}\ln \Lambda) = 4/15 + 4/105 A$ (dash-dotted line) is also shown. The relaxation rate is given in units of $\bar{\nu}\ln \Lambda$, so the Landau theory takes a single curve, whereas results of the Lenard-Balescu theory are shown at four values of $\Gamma = 10^{-1}$, $10^{-2}$, $10^{-4}$ and $10^{-6}$. }
\label{fg:nu_A_comp}
\end{figure}
Figure~\ref{fg:nu_wc_comp} shows that the influence of dynamic screening is negligible when the anisotropy is small. However, it may be expected to become more important as the anisotropy increases. Figure~\ref{fg:nu_A_comp} shows how the relaxation rate predicted by each theory depends on the anisotropy parameter $A$. The relaxation rate is plotted in units of $\bar{\nu} \ln \Lambda$, so the Landau result from Eq.~(\ref{eq:nu_nrl}) provides a single curve. The Taylor expansion of this curve up to linear order, $\nu/(\bar{\nu}\ln \Lambda) = 4/15 +(4/105)A$, shows that, as expected, the expansion holds in a narrow neighborhood near $A=0$. The accuracy of the expansion extends much further on the $A>0$ side than for $A<0$. The dashed lines show the predictions of the Lenard-Balescu result from Eq.~(\ref{eq:nu_b0_full}) at $\Gamma = 10^{-1}$, $10^{-2}$, $10^{-4}$ and $10^{-6}$. Comparison of these curves shows that, indeed, dynamic screening has a larger influence as the anisotropy increases or the coupling strength increases. For example, at $A=10$ and $\Gamma = 10^{-2}$ there is an approximately 10\% difference between the Landau and Lenard-Balescu predictions. Two notes should be kept in mind when interpreting this result.
First, the Lenard-Balescu theory treats the long-distance behavior of Coulomb interactions, which corrects the infrared divergence of Landau theory and incorporates temperature anisotropy in the screening length. However, it still contains an ultraviolet divergence, which is resolved by limiting the distance of closest approach to be the thermal Landau length $r_L = e^2/(2k_BT)$. The temperature arising in this length scale is assumed to be the total temperature, but it is not clear if it should also account for temperature anisotropy.
Second, the potential role of dynamic screening should be analyzed keeping in mind that both theories are making an assumption that the distribution function remains an anisotropic Maxwellian throughout the evolution. Expectations from other transport processes, such as diffusion, suggest that the error in this approximation is on the order of 20\% at weak coupling~\cite{ferz:72,baal:14}. Some evidence for this from comparison with MD simulations will also be shown in Sec.~\ref{sec:results}. Thus, the error associated with approximations related to the distribution function are often more significant than changes in the transport rates associated with anisotropy in the screening.
\section{Strongly Coupled Plasma Theories\label{sec:sc}}
Both the Landau and Lenard-Balescu based theories from the previous section address the weakly coupled limit. In this section, two theories are evaluated that attempt to extend these methods into the strongly coupled regime. Each of these is based on phenomenological assumptions, rather than rigorous asymptotic expansions, so the subsequent comparison with MD serves as a test of these assumptions.
\subsection{Effective Potential Theory}
Like the Landau approach, the effective potential theory (EPT)~\cite{baal:13,baal:14,baal:15} is based on the Boltzmann collision operator, which for the OCP is
\begin{equation}
C_\textrm{B} = \int d^3v^\prime \int d\Omega\, \sigma\, u [f(\hat{\mathbf{v}}) f(\hat{\mathbf{v}}^\prime) - f(\mathbf{v}) f(\mathbf{v}^\prime)], \label{eq:c_boltz}
\end{equation}
where $\sigma$ is the differential scattering cross section, the hat denotes the velocity vector after a binary collision: $\hat{\mathbf{v}} = \mathbf{v} + \Delta \mathbf{v}$ and $\mathbf{u} \equiv \mathbf{v} - \mathbf{v}^\prime$. Unlike the other theories discussed in this paper, the anisotropy relaxation rate has not previously been calculated from this theory, so we provide details of the calculation here. Starting with the $v_\parallel^2$ moment of Eq.~(\ref{eq:c_boltz}) provides
\begin{equation}
\frac{dT_\parallel}{dt} = \frac{m}{n} \int d^3v\ v_\parallel^2 C_B(f) = \frac{m}{n} \int d^3v \int d^3v^\prime \lbrace \Delta (v_\parallel^2) \rbrace f(\mathbf{v}) f (\mathbf{v}^\prime) \label{eq:ept_1}
\end{equation}
where the curly brackets are defined as $\lbrace \chi \rbrace = \int d\Omega \sigma u \Delta \chi$~\cite{baal:12}, for any function $\chi$, $\Delta (v_\parallel^2) = \hat{v}_\parallel^2 - v_\parallel^2$, and momentum conservation implies $\Delta \mathbf{v} = \Delta \mathbf{u}/2$, so $\Delta (v_\parallel^2) = v_\parallel \Delta u_\parallel + \frac{1}{4} (\Delta u_\parallel)^2$.
Assuming that the distribution function has the anisotropic Maxwellian form described by Eq.~(\ref{eq:amax}), the three velocity integrals for the variable $\mathbf{v}$ in Eq.~(\ref{eq:ept_1}) can be computed analytically leading to
\begin{equation}
\frac{dT_\parallel}{dt} = \frac{n}{\pi^{3/2} v_{T\perp}^2 v_{T\parallel}} \frac{m}{2^{9/2}} \int d^3u\, u (u^2 - 3 u_\parallel^2) \bar{\sigma}^{(2)}(u) \exp \biggl(- \frac{1}{2} \frac{u_\perp^2}{v_{T\perp}^2} \biggr) \exp \biggl( - \frac{1}{2} \frac{u_\parallel^2}{v_{T\parallel}^2} \biggr) \label{eq:ept_2}
\end{equation}
in which
\begin{equation}
\bar{\sigma}^{(l)} = 2\pi \int_0^\infty db\, b[1 - \cos^l (\pi - 2 \Theta)] \label{eq:sig_l}
\end{equation}
is the $l^\textrm{th}$ momentum scattering cross section and $\theta = \pi - 2\Theta$ is the scattering angle, where
\begin{equation}
\Theta = b \int_{r_o}^\infty dr\, r^{-2} [1-b^2/r^2 - (e\phi(r)/k_BT)/\xi^2]^{-1/2} . \label{eq:theta}
\end{equation}
Obtaining Eq.~(\ref{eq:ept_2}) from Eq.~(\ref{eq:ept_1}) makes use of the relations $\lbrace \Delta \mathbf{u} \rbrace = -u \mathbf{u} \bar{\sigma}^{(1)} (u)$ and $\lbrace \Delta \mathbf{u} \Delta \mathbf{u} \rbrace = u[2 \mathbf{u} \mathbf{u} \bar{\sigma}^{(1)} + \frac{1}{2} (u^2 \ensuremath{\mathcal{I}} - 3 \mathbf{u} \mathbf{u}) \bar{\sigma}^{(2)} (u) ]$, which imply $u_\parallel \lbrace \Delta u_\parallel \rbrace + \frac{1}{2} \lbrace \Delta (u_\parallel^2) \rbrace = \frac{1}{4} u (u^2 - 3u_\parallel^2) \bar{\sigma}^{(2)} $. Applying a spherical coordinate system $\mathbf{u} = u[\cos \theta^\prime\, \sin\phi^\prime \hat{x} + \sin \theta^\prime \sin \phi^\prime \hat{y} + \cos \phi^\prime \hat{z}]$ both the $\theta^\prime$ and $\phi^\prime$ integrals can be carried out. This leads to Eq.~(\ref{eq:nrl}) where the anisotropy relaxation rate is
\begin{equation}
\frac{\nu_\textrm{EPT}}{\bar{\nu}} = \chi \frac{3 \sqrt{\pi}}{16} \frac{(1+\frac{2}{3}A)^{3/2}}{\sqrt{\alpha} A^{5/2}} \int_0^\infty d\xi\, \xi^2 e^{-\alpha \xi^2} \frac{\bar{\sigma}^{(2)}}{\sigma_o} \biggl[ \frac{2}{3} \xi^2 \alpha A \ensuremath{\textrm{erf}} (\xi \sqrt{\alpha A}) - \psi (\xi^2 \alpha A) \biggr] . \label{eq:nu_ept}
\end{equation}
Here, $\sigma_o \equiv \pi e^4/(2k_BT)^{2}$, $\xi^2 \equiv \frac{1}{2} u^2/v_{T}^2 = u^2/(4k_BT/m)$, $\alpha \equiv T/T_\perp = \frac{1}{3} (3+A)/(1+A)$
and
\begin{equation}
\psi(x) \equiv \ensuremath{\textrm{erf}}( \sqrt{x}) - \frac{2}{\sqrt{\pi}} \sqrt{x} e^{-x}
\end{equation}
is the Maxwell integral. The right side of Eq.~(\ref{eq:nu_ept}) has been multiplied by a factor $\chi$, which is an aspect of EPT that will be explained below.
The expression in Eq.~(\ref{eq:nu_ept}) follows directly from inserting the anisotropic distribution function in the Boltzmann collision operator. The EPT extends this Boltzmann-based result by relaxing (but not eliminating) two of its underlying assumptions: binary collisions and molecular chaos. First, the main premise of the theory is that binary collisions between particles do not occur in isolation, but rather in a ``sea'' of background particles. This ``relaxes'' Boltzmann's assumption by accounting for aspects of many-body physics via the interaction potential. The appropriate potential is one which takes into account the fields associated with nearby charges. In particular, the potential of mean force is the potential obtained when taking two particles at fixed positions and averaging over the positions of all other particles~\cite{hill:60}
\begin{eqnarray}
\mathbf{F}_{12} &=& \int \biggl[ - \nabla_{\mathbf{r}_1} U(\mathbf{r}_1, \mathbf{r}_2, \ldots , \mathbf{r}_N) \biggr] \frac{e^{-U/k_BT}}{\mathcal{Z}} d\mathbf{r}_3 \ldots d\mathbf{r}_N \label{eq:f12} \\ \nonumber
&=& k_BT \nabla_{\mathbf{r}_1} \ln g(|\mathbf{r}_1 - \mathbf{r}_2|) \equiv - \nabla_{\mathbf{r}_1} \phi (\mathbf{r}_1 - \mathbf{r}_2) .
\end{eqnarray}
Here $\mathcal{Z} = \int \exp(-U/k_BT) d\mathbf{r}_N$ is the configurational integral and $U \equiv \sum_{i,j} v(|\mathbf{r}_i - \mathbf{r}_j|)$. This depends only on the bare interaction potential of the particles $v_{ij}$, which is the Coulomb potential. Thus, once $g(r)$ is determined, the interaction potential is modeled as $\phi(r) = - k_BT \ln[g(r)]$.
Second, the factor $\chi$ in Eq.~(\ref{eq:nu_ept}) accounts for an increased collision rate due to the excluded volume surrounding particles interacting via a repulsive (Coulomb) potential~\cite{baal:15}. This excluded volume is sometimes referred to as the Coulomb hole. It decreases the volume of space that particles can occupy, resulting in an increased scattering rate. Accounting for this relaxes Boltzmann's molecular chaos assumption somewhat because it introduces an aspect of correlation in the initial positions of scattering particles. We model this factor using a modified version of Enskog's theory of hard sphere gases that has been adapted to plasmas~\cite{baal:15}. In this theory, the collision frequency enhancement factor is
\begin{equation}
\chi \simeq 1 + 0.6250 b \rho + 0.2869 (b\rho)^2 \label{eq:chi}
\end{equation}
where $b\rho \simeq \pi n \bar{\sigma}^3/3$ and $\bar{\sigma}$ is the effective particle diameter. This is determined from the location where $g(r=\bar{\sigma}) = 0.87$. For the OCP at $\Gamma \lesssim 1$, $\chi \simeq 1$, while for $\Gamma \simeq 1-100$, $\chi \simeq 1.2-1.4$. This 20-40\% increase in the collision frequency can be seen in figure 7 of \cite{baal:15}.
Equations~(\ref{eq:sig_l})--(\ref{eq:nu_ept}) along with $\phi = -k_BT \ln [g(r)]$ and (\ref{eq:chi}) provide a closed set of equations based on one input: $g(r)$. To compute $g(r)$ we use the hypernetted chain (HNC) approximation~\cite{hans:06}
\begin{subequations}
\label{eq:hnc}
\begin{align}
g(r) &= \exp[- v(r)/k_BT + h(\mathbf{r}) - c(\mathbf{r}) + b(\mathbf{r})] \\
\hat{h} (\mathbf{k}) &= \hat{c}(\mathbf{k}) [1 + n\hat{h}(\mathbf{k})]
\end{align}
\end{subequations}
where $v(\mathbf{r})/k_BT = \Gamma a/r$ is the bare Coulomb potential, $h(\mathbf{r}) \equiv g(\mathbf{r}) - 1$ and ``hats'' denote Fourier transforms in the spatial coordinate. In Eq.~(\ref{eq:hnc}), $b(\mathbf{r})$ is a bridge function that that increases the accuracy of the HNC approximation at high $\Gamma$. For this term, we apply the model of Iyetomi \textit{et al}\cite{iyet:92}. The HNC approximation has long been benchmarked for the Coulomb OCP, and it has been shown to provide accurate input for the EPT over the range of coupling strengths considered here~\cite{baal:13,baal:14}.
Finally, we note that by modeling the interaction potential using an equilibrium theory, we have implicitly assumed that the effective interaction potential is spherically symmetric. For a large enough temperature anisotropy this approximation should be expected to break down. Next, we consider the weak anisotropy limit and the weakly coupled limit of the EPT theory.
\subsubsection{Weak anisotropy expansion}
If the temperature anisotropy is small, $|A| \ll 1$, expanding Eq.~(\ref{eq:nu_ept}) gives
\begin{equation}
\frac{\nu}{\bar{\nu}} = \frac{2}{15} \Xi^{(2,2)} + A \biggl( \frac{2}{63} \Xi^{(2,3)} - \frac{2}{45} \Xi^{(2,2)} \biggr) + \mathcal{O}(A^{3/2}) \label{eq:nu_ept_sa}
\end{equation}
in which the generalized Coulomb logarithms
\begin{equation}
\Xi^{(l,k)} = \frac{\chi}{2} \int_0^\infty d\xi\, \xi^{3k+3} e^{-\xi^2} \bar{\sigma}^{(l)}/\sigma_o
\end{equation}
have been identified in the analogous way as Refs.~\cite{baal:12,baal:13}. Equation~(\ref{eq:nu_ept_sa}) returns the same result as the small anisotropy expansions of the Landau theory from Eq.~(\ref{eq:nu_nrl}) if the appropriate Coulomb logarithm is substituted for $\ln \Lambda$.
This can be obtained by applying the weakly coupled limit of the generalized Coulomb logarithms from \cite{baal:14}, $\Xi^{(l,k)} = l \Gamma(k) \ln \Lambda$, where $\Gamma(k)$ is the Gamma function.
\subsubsection{Weakly coupled limit}
To connect Eq.~(\ref{eq:nu_ept}) with the standard result from weakly coupled plasma theory, Eq.~(\ref{eq:nu_nrl}), consider taking the weak coupling limit of Eq.~(\ref{eq:nu_ept}). In this limit, $\bar{\sigma}^{(l)} = \lambda_D^2 4\pi l \ln (\Lambda \xi^2)/(\Lambda \xi^2)^2$ \cite{baal:14}, and the $\xi$ dependence of the Coulomb logarithm is a small correction because $\Lambda$ is large, so $\bar{\sigma}^{(l)}/\sigma_o \simeq 4 l \ln \Lambda/\xi^4$. With this, Eq.~(\ref{eq:nu_ept}) is
\begin{equation}
\frac{\nu}{\bar{\nu}} = \frac{3 \sqrt{\pi}}{2} \ln \Lambda \frac{(1+\frac{2}{3}A)^{3/2}}{A^{5/2}} \biggl[ \frac{2}{3} A \int_0^\infty dx\, e^{-x^2} \ensuremath{\textrm{erf}} (x \sqrt{A}) - \int_0^\infty dx\, e^{-x^2} \frac{\psi(x^2A)}{x^2} \biggr] .
\end{equation}
The first of these integrals can be written in terms of the $\arctan$ function $\int_0^\infty dx\, e^{-x^2} \ensuremath{\textrm{erf}}(x \sqrt{A}) =\arctan (\sqrt{A})/\sqrt{\pi}$. The second integral can be evaluated by first integrating by parts, then identifying a similar integral that can be evaluated in terms of $\arctan$. The result returns Eq.~(\ref{eq:nu_nrl}).
In the weak anisotropy limit $|A| \ll 1$, $\nu/\bar{\nu} \rightarrow [4/15 + (4/105) A + \mathcal{O}(A^{3/2})] \ln \Lambda$. Alternatively, this can be obtained from Eq.~(\ref{eq:nu_ept_sa}) by applying the weakly coupled limit of the generalized Coulomb logarithms.
\subsection{Generalized Lenard-Balescu Theory\label{sec:glb}}
The second theory that we consider is Ichimaru's generalization of the Lenard-Balescu equation~\cite{ichi:92}
\begin{equation}
C_\textrm{I} = - \frac{2 e^4}{m^2} \frac{\partial}{\partial \mathbf{v}} \cdot \int d^3v^\prime \int d^3k \frac{[1-G(k)] \delta (\mathbf{k} \cdot \mathbf{u})}{|\hat{\varepsilon}_\textrm{I} (\mathbf{k}, \mathbf{k} \cdot \mathbf{v})|^2} \frac{\mathbf{k} \mathbf{k}}{k^4} \biggl[ f (\mathbf{v}) \frac{\partial f(\mathbf{v}^\prime)}{\partial \mathbf{v}^\prime} - f(\mathbf{v}^\prime) \frac{\partial f(\mathbf{v})}{\partial \mathbf{v}} \biggr] . \label{eq:c_ichi}
\end{equation}
Here, $G(k)$ is the local field correction (LFC) which accounts for static correlations. For the OCP~\cite{ichi:92}
\begin{equation}
1 - G(k) = \frac{(ka)^2}{3 \Gamma} \biggl( \frac{1}{S(k)} - 1 \biggr)
\end{equation}
where
\begin{equation}
S(k) = 1 + \frac{3}{ka^3} \int_0^\infty dr\, r \sin (kr) [g(r) - 1]
\end{equation}
is the static structure factor. This will be determined by the HNC approximation described in Eq.~(\ref{eq:hnc}). Inserting the anisotropic distribution function from Eq.~(\ref{eq:amax}) leads to the evolution equation Eq.~(\ref{eq:nu_def}) with the relaxation rate given by
\begin{equation}
\nu_{\textrm{I}} = \frac{n e^4}{\pi} \int d^3k \int_{-\infty}^\infty d\omega \frac{1-G(k)}{k^4 |\hat{\varepsilon}_\textrm{I}(\mathbf{k}, \omega)|^2} \frac{k_\parallel^2 k_\perp^2}{(k_\parallel^2 k_BT_\parallel + k_\perp^2 k_BT_\perp)^2} \exp \biggl( - \frac{m \omega^2}{k_\parallel^2 k_BT_\parallel + k_\perp^2 k_BT_\perp} \biggr) \label{eq:nu_ichi}
\end{equation}
where
\begin{equation}
\hat{\varepsilon}_\textrm{I}(\mathbf{k},\omega) = 1 + \frac{\omega_p^2 [1-G(k)]}{k^2} \int d^3v \frac{\mathbf{k} \cdot \partial f/\partial \mathbf{v}}{\omega - \mathbf{k} \cdot \mathbf{v}} = 1 - \frac{\omega_p^2 [1-G(k)]}{k_\parallel^2 v_{T\parallel}^2 + k_\perp^2 v_{T\perp}^2} Z^\prime \biggl( \frac{\omega}{\sqrt{k_\parallel^2 v_{T\parallel}^2 + k_\perp^2 v_{T\perp}^2}} \biggr) \label{eq:ep_ichi}
\end{equation}
is the dielectric function.
Unlike the Lenard-Balescu equation, Eq.~(\ref{eq:c_ichi}) does not suffer an ultraviolet divergence because the local field correction accounts for the close interaction physics. For computational simplicity, one of the $k$ integrals can be computed in an analogous way to what was done in Eq.~(\ref{eq:nu_mid2}) providing a similar result that includes the LFC
\begin{equation}
\frac{\nu_{\textrm{I}}}{\bar{\nu}} = \biggl( \frac{1+\frac{2}{3}A}{1+A} \biggr)^{3/2} \frac{1}{\sqrt{\pi}} \int_{-\infty}^\infty dx e^{-x^2} \int_{-1}^1 dy \frac{y^2 (1-y^2)}{(y^2 \tilde{A} + 1)^{3/2}} \int_0^{\infty} d \bar{k} \frac{ [1-G(\bar{k})}{\bar{k}} \biggl| 1 + \frac{\Phi(x)[1-G(\bar{k})]}{\bar{k}^2(y^2 \tilde{A} + 1)} \biggr|^{-2} . \label{eq:glb_mid2}
\end{equation}
Limiting values of Eq.~(\ref{eq:glb_mid2}) that are similar to what is listed for the Lenard-Balescu equation in Secs.~\ref{sec:isa}--\ref{sec:dh} can also be taken, providing similar results but where the LFC enters the expressions for the generalized Coulomb logarithm in place of the cutoff $b_{\min}$. The full expression is compared with the EPT predictions in Sec.~\ref{sec:results}.
\section{MD Simulation Description\label{sec:md}}
\begin{figure}[t]
\includegraphics[width=75mm]{fig4a.pdf}
\includegraphics[width=75mm]{fig4b.pdf}
\caption{(a) Parallel and perpendicular temperature profiles in time computed from two independent MD runs (blue and red lines), as well as the result obtained from the average of $25$ independent runs (black line). In these simulations $\Gamma = 0.2$ and $A=-0.2$. (b) Difference in the anisotropic temperatures in time from MD simulations (solid lines) and exponential fits over a time interval (dashed lines). Here, $T_{\parallel,o} = 1.15$.}
\label{fg:md_analysis}
\end{figure}
The MD simulations were performed as follows.
$N$ charged particles interacting through the pure Coulomb interaction in a uniform, neutralizing background were placed in a cubic box of volume $V$.
Periodic conditions were imposed on all boundaries.
Particle trajectories were determined by solving Newton's equations of motion with the velocity Verlet integrator \cite{FrenkelSmit}.
The force on an ion that results from its interaction with the ions in the simulation box and with those in the periodically replicated cells were calculated using the Ewald summation technique.
For numerical efficiency, the Ewald sum was calculated with a parallel implementation of the particle-particle-particle-mesh ($\rm P^3M$) method that simultaneously provides high resolution for individual encounters combined with rapid, mesh-based, long range force calculations \cite{HockneyEastwood}.
Time was normalized to the plasma frequency $\omega_p$.
The integration time step $\delta t$ and the number of particles $N$ given below were chosen in order to conserve energy to $<10^{-5}$, and to ensure high enough collision ability in the simulation cell (the calculations are more demanding at small coupling due to long collision mean-free path).
Specifically, for $\Gamma\leq 0.1$, we used $N=10^5$ and $\delta t=10^{-3}/\omega_p$; for $0.1<\Gamma<1$, $N=50000$ and $\delta t=0.01/\omega_p$; for $\Gamma\geq 1$, $N=5000$ and $\delta=0.01/\omega_p$.
Each simulation consisted of an equilibration phase of length $t_{\textrm{eq}}=N_{\textrm{eq}}\delta t$ followed by a relaxation phase of length $t_{\textrm{run}}=N_{\textrm{run}}\delta t$, with $t_{\textrm{eq}}=1000/\omega_p$ and $t_{\textrm{run}}=300/\omega_p$.
The initial particle positions at time $t=-t_{\textrm{eq}}$ were assigned randomly in the simulation box, with a small region surrounding each particle excluded to avoid initial explosion.
The initial particle velocities ${\bf v}_i$ were sampled from a Maxwell-Boltzmann distribution at the desired temperature $T$, i.e. at the desired value of $\Gamma$.
During the equilibration phase, velocity scaling \cite{FrenkelSmit} was used to maintain the desired $\Gamma$ value.
At $t=0$, velocity scaling was turned off and the particle velocities were rescaled to the desired initial parallel and perpendicular temperatures as follows
\begin{eqnarray}
v_{i,x}(t=0^+)=v_{i,x}(0^-)\sqrt{\frac{T_\perp}{T}}\,,\,
v_{i,y}(0^+)=v_{i,y}(0^-)\sqrt{\frac{T_\perp}{T}}\,,\,
v_{i,z}(0^+)=v_{i,z}(0^-)\sqrt{\frac{T_\parallel}{T}} .
\end{eqnarray}
The system was then left to evolve freely, i.e. in the microcanonical ensemble, for a duration $t_{\textrm{run}}$.
During this period the instantaneous parallel and perpendicular temperatures, defined as follows in terms of the particles kinetic energies
\begin{eqnarray}
T_\parallel(t)\equiv\frac{2}{Nk_B}\sum_{i=1}^N{\frac{1}{2}mv_{i,z}^2(t)}\quad,\quad
T_\perp(t)\equiv\frac{1}{Nk_B}\sum_{i=1}^N{\frac{1}{2}m\left(v_{i,x}^2(t)+v_{i,y}^2(t)\right)},
\end{eqnarray}
were monitored.
In order to sample the initial statistical distribution function, Eq.~(\ref{eq:amax}), and to compare the MD simulations to the predictions of kinetic theories, $N_{\textrm{sim}}=25$ independent runs were performed with different initial conditions and averaged.
As illustrated in Fig.~\ref{fg:md_analysis}, the averaging smooths out the fluctuations in the kinetic energies inherent to each microcanonical dynamics.
To obtain the relaxation rate $\nu$, we assume that the MD temperatures evolve according to the rate equations (\ref{eq:nrl}).
The latter implies $\frac{d}{dt}\Delta T=-3\nu \Delta T$ with $\Delta T=T_\parallel-T_\perp$.
In general, as predicted by kinetic theory, $\nu$ depends on the time $t$ through its dependence upon $T_\parallel$ and $T_\perp$.
If one assumes that the dependence is weak enough that $\nu$ is a constant on a short enough time scale $\Delta t$ beyond an initial time $t_0$ , then $\Delta T(t)=\Delta T(t_0)e^{-3\nu (t-t_0)}$ on this time scale (more sophisticated treatments that take into account the temperature dependence of $\nu$ are possible (e.g. \cite{dimo:08}) but we find that they do not affect the results in any significant manner).
The MD simulations confirm that $T_\parallel-T_\perp$ indeed decays exponentially for coupling parameters $\Gamma\!<\!1$ but only following a short transient period of duration $t_0$ of order $\omega_{p}^{-1}$; see Fig.~\ref{fg:md_analysis}b. For $\Gamma\geq 1$, as discussed in the next section, the temperature evolution is never exponential and the notion of relaxation rate defined based on the rate equations (\ref{eq:nrl}) does not apply.
The initial transient period describes the dependence on initial correlations, which are discarded in the kinetic theories discussed before.
In practice, the relaxation rates $\nu$ are obtained by fitting the MD data to the analytical solution $\Delta T(t_0)e^{-3\nu (t-t_0)}$ over the time interval $[t_0,t_0+\Delta t]$ with the time $t_0$ chosen right after the early transient behavior (the data shown below were obtained with $t_0=1/\omega_p$) and with $t_0+\Delta t$ chosen where the exponential behavior switches slope as a consequence of the dependence of $\nu$ on the time $t$. Figure~\ref{fg:md_analysis}b illustrates this procedure.
\section{Results \label{sec:results}}
\begin{figure}[t]
\sidecaption
\includegraphics[width=95mm]{fig5.pdf}
\caption{Parallel and perpendicular temperature profiles in time computed from MD simulations (solid lines) at $\Gamma$ = 0.05, 0.5, 5 and 25 and initial temperatures of $\bar{T}_\parallel = T_\parallel/T = 1.15$ and $\bar{T}_\perp = T_\perp/T = 0.923$ ($A=-0.2$). Dashed lines show profiles predicted from the EPT approximation. }
\label{fg:md_profiles}
\end{figure}
Figure~\ref{fg:md_profiles} shows parallel and perpendicular temperature profiles at four values of $\Gamma$ from MD simulations that were initiated with temperatures of $T_\parallel/T = 1.15$ and $T_\perp/T = 0.923$ ($A=-0.2$). Also shown are the profiles predicted by the EPT theory of Eq.~(\ref{eq:nu_ept}). The comparison shows good agreement with the predicted monotonically decreasing profiles at the lower $\Gamma$ values.
As the coupling strength increases the rate of the relaxation remains well modeled, but features emerge in the MD data that are not predicted by the model. One feature is that the profiles from MD become non-monotonic, exhibiting oscillations. Such oscillations are common in time correlation functions at strong coupling. For instance, they can be observed in the velocity autocorrelation function, which when integrated provides the macroscopic diffusion coefficient~\cite{dali:06}.
Dubin~\cite{dubi:05} has previously discussed how the temperature anisotropy relaxation rate can also be considered as the time integral of an energy correlation function.
Another interesting feature of the MD data at strong coupling is a short initial delay, compared to the theoretical prediction, before relaxation onsets. This is also common feature of other correlation functions at strong coupling and is associated with initial conditions. For instance, the velocity autocorrelation function at short times behaves as $Z(t)/Z(0) = 1 - \Omega_E^2 t^2$, where $\Omega_E$ is the Einstein frequency~\cite{hans:06}. For an OCP, $\Omega_E = \omega_p/\sqrt{3}$. At weak coupling the relaxation time is much longer than $\omega_p^{-1}$, so this initial stage is so short that it is inconsequential in the overall relaxation profile. For instance, Fig.~\ref{fg:md_profiles}a shows that the relaxation time is approximately 200$\omega_p^{-1}$ at $\Gamma = 0.05$. However, at strong coupling the relaxation rate is only a few $\omega_p^{-1}$, so the short time rolloff on the $\Omega_E^{-1}$ timescale is a more pronounced feature of the overall relaxation profile. The data shown in Fig.~\ref{fg:md_profiles} suggests that similar arguments can also be applied to the temperature anisotropy relaxation process. All of the theories discussed above are Markovian, and as a consequence do not include information about these initial conditions. They all predict exponential profiles at short times, which leads to a linear-in-time term at the relaxation onset rather than a quadratic one.
\begin{figure}[t]
\sidecaption
\includegraphics[width=75mm]{fig6.pdf}
\caption{Temperature anisotropy relaxation rate obtained from MD simulations (circles), and the predictions of the Landau theory from Eq.~(\ref{eq:nu_nrl}) (green line), the Lenard-Balescu theory from Eq.~(\ref{eq:nu_b0_full}) (magenta dash-dotted line), the EPT theory from Eq.~(\ref{eq:nu_ept}) (black line) and the generalized Lenard-Balescu theory from Eq.~(\ref{eq:glb_mid2}) (blue dashed line). }
\label{fg:md_compare}
\end{figure}
Figure~\ref{fg:md_compare} shows the results of the relaxation rate obtained from the MD data using the fitting procedure described in Sec.~\ref{sec:md}. The initial temperature anisotropy for this data set was the same $A=-0.2$ as from the data shown in Fig.~\ref{fg:md_profiles}. Predictions of the four theories described in Sections~\ref{sec:wc} and \ref{sec:sc} are shown for comparison. As expected, the weakly coupled theories compare well with the MD data at sufficiently small coupling strength ($\Gamma \lesssim 0.2$). As described in Sec.~\ref{sec:wc}, there is apparently no significant advantage to accounting for dynamic screening at these conditions. Both strongly coupled theories extend the standard plasma theories into the strongly coupled regime, but the generalized Lenard-Balescu theory from Sec.~\ref{sec:glb} predicts a negative collision frequency for $\Gamma \gtrsim 20$ (negative values are not shown on this log-log scale). This feature of the theory has also been noted in computing the viscosity coefficient in \cite{dali:14}. Although MD simulations were carried out at higher $\Gamma$ values than shown in this figure, the oscillatory nature of the temperature profiles made the fitting procedure unreasonable at sufficiently high coupling strength.
Figures~\ref{fg:md_profiles}a and \ref{fg:md_compare} show that for $\Gamma \lesssim 0.1$ the relaxation rate predicted by all of the theories is slightly larger than what is observed in the MD simulations (for example, by approximately 20\% at $\Gamma = 0.05$). This is likely associated with the assumption made in the theoretical analysis that the distributions maintain the anisotropic Maxwellian form. It is well known in other transport processes that distortions of the distribution function contributes an order unity correction to the transport rates at weak coupling. This is accounted for in hydrodynamic theories, such as Chapman-Enskog or Grad, through the higher order terms of the Sonine polynomial expansions of the distribution function. However, the magnitude of these terms is known to rapidly diminish as $\Gamma \gtrsim 1$. For instance, it was shown in \cite{baal:14} that for self-diffusion of the OCP the approximately 20\% contribution of the second order term of the expansion vanishes for $\Gamma > 1$ (see figure 9 of that reference). Figure~\ref{fg:md_compare} shows evidence for the same effect in the temperature anisotropy relaxation rate. To the best of our knowledge, such an expansion procedure has not yet been applied to temperature anisotropy relaxation.
\begin{figure}[t]
\sidecaption
\includegraphics[width=160mm]{fig7.pdf}
\caption{Time dependent temperature profiles with a highly anisotropic initial condition of $T_\perp/T_\parallel = 10$ ($A=9$). Panels (a)--(d) show the parallel ($\bar{T}_\parallel = T_\parallel/T$) and perpendicular ($\bar{T}_\perp = T_\perp/T$) temperatures from MD simulations (black solid lines) and predictions from EPT (red dashed lines). Panels (e)--(h) show the total kinetic temperature as a function of time in terms of its initial value ($T(t)/T_o$).}
\label{fg:large_a_profiles}
\end{figure}
Finally, Fig.~\ref{fg:large_a_profiles} shows temperature relaxation profiles from MD simulations at $\Gamma = 0.1$, 1, 10 and 100 conducted with a large initial temperature anisotropy of $T_\perp/T_\parallel = 10$ ($A=9$). The comparison with the EPT predictions is similar to what was observed at smaller initial anisotropy. The accuracy is, perhaps, slightly less than was observed from the small anisotropy cases in Fig.~\ref{fg:md_profiles} near $\Gamma \simeq 1$, but the generally good agreement over this entire range of coupling strength provides strong evidence that the theory is robust even at a very large initial anisotropy. The predicted rate compares well even at $\Gamma = 100$. Dynamic screening apparently is not significant even at $T_\perp/T_\parallel =10$. Panels e-h show the total temperature as a function of time for the same simulations. These show that at weak coupling the total temperature, and thus total kinetic energy, does not fluctuate more than approximately 0.1\% throughout the evolution of the system. The higher $\Gamma$ values show that the temperature fluctuates in time by as much as 0.5\%. Recall that each curve is obtained from 25 independent simulations. The total energy of the system was confirmed to be conserved to better than 1 part in $10^6$ throughout each simulation. These oscillations are not numerical artifacts, but rather represent oscillations in the exchange between kinetic and potential energy associated with correlations as the system relaxes. None of the theories discussed address potential energy of the system, and thus do not model this effect.
\section{Conclusions}
This comparison between two common approaches to plasma kinetic theory and MD simulations has suggested a few general conclusions with regard to the temperature anisotropy relaxation, and has also revealed a few remaining gaps in current understanding. The data comparison suggests that dynamic screening, which is modeled in the Lenard-Balescu and generalized Lenard-Balescu theories, does not significantly influence the relaxation rate. The Boltzmann-based approaches (Landau or EPT) can be accurately applied in most situations of practical interest, rather than Lenard-Balescu based theories which are comparatively difficult to evaluate. Comparison with MD revealed that the EPT provides an accurate approach to modeling the anisotropy relaxation rate over a similar range of coupling strength as has been encountered for other processes, such as diffusion~\cite{baal:15}. However, none of the theories capture kinetic energy oscillations (associated with kinetic-potential energy exchange) or early time delay (associated with initial conditions) in the time-dependent temperature profiles that were observed in the MD simulations. These effects are associated with strong correlations. Finally, one additional limitation of current models is the assumption that the distribution maintains the specified anisotropic Maxwellian form throughout the evolution. Comparison with MD reveals that improvements to predicted relaxation rates may be expected from accounting for deviations from this assumed distribution.
\begin{acknowledgement}
This material is based upon work supported by LDRD project 20150520ER at Los Alamos National Laboratory. The work of SDB was also supported by the Air Force Office of Scientific Research under award number FA9550-16-1-0221 and by the National Science Foundation under Grant No. PHY-1453736.
\end{acknowledgement}
\section*{Greek symbols -- {\tt w-greek.sty}}
\begin{table}
\begin{tabular}{*8l}
\X\alpha &\X\theta &\X o &\X\tau \\
\X\beta &\X\vartheta &\X\pi &\X\upsilon \\
\X\gamma &\X\iota &\X\varpi &\X\phi \\
\X\delta &\X\kappa &\X\rho &\X\varphi \\
\X\epsilon &\X\lambda &\X\varrho &\X\chi \\
\X\varepsilon &\X\mu &\X\sigma &\X\psi \\
\X\zeta &\X\nu &\X\varsigma &\X\omega \\
\X\eta &\X\xi \\
\\
\X\itGamma &\X\itLambda &\X\itSigma &\X\itPsi \\
\X\itDelta &\X\itXi &\X\itUpsilon &\X\itOmega \\
\X\itTheta &\X\itPi &\X\itPhi
\end{tabular}
\caption{Slanted greek letters}
\end{table}
\begin{table}
\begin{tabular}{*8l}
\X\upalpha &\X\uptheta &\X\upo &\X\uptau \\
\X\upbeta &\X\upvartheta &\X\uppi &\X\upupsilon \\
\X\upgamma &\X\upiota &\X\upvarpi &\X\upphi \\
\X\updelta &\X\upkappa &\X\uprho &\X\upvarphi \\
\X\upepsilon &\X\uplambda &\X\varrho &\X\upchi \\
\X\varepsilon &\X\upmu &\X\upsigma &\X\uppsi \\
\X\upzeta &\X\upnu &\X\upvarsigma &\X\upomega \\
\X\upeta &\X\upxi \\
\\
\X\Gamma &\X\Lambda &\X\Sigma &\X\Psi \\
\X\Delta &\X\Xi &\X\Upsilon &\X\Omega \\
\X\Theta &\X\Pi &\X\Phi
\end{tabular}
\caption{Upright greek letters}
\end{table}
\newpage
\begin{table}
\begin{tabular}{*8l}
\textrm{B}\alpha &\textrm{B}\theta &\textrm{B} o &\textrm{B}\tau \\
\textrm{B}\beta &\textrm{B}\vartheta &\textrm{B}\pi &\textrm{B}\upsilon \\
\textrm{B}\gamma &\textrm{B}\iota &\textrm{B}\varpi &\textrm{B}\phi \\
\textrm{B}\delta &\textrm{B}\kappa &\textrm{B}\rho &\textrm{B}\varphi \\
\textrm{B}\epsilon &\textrm{B}\lambda &\textrm{B}\varrho &\textrm{B}\chi \\
\textrm{B}\varepsilon &\textrm{B}\mu &\textrm{B}\sigma &\textrm{B}\psi \\
\textrm{B}\zeta &\textrm{B}\nu &\textrm{B}\varsigma &\textrm{B}\omega \\
\textrm{B}\eta &\textrm{B}\xi \\
\\
\textrm{B}\itGamma &\textrm{B}\itLambda &\textrm{B}\itSigma &\textrm{B}\itPsi \\
\textrm{B}\itDelta &\textrm{B}\itXi &\textrm{B}\itUpsilon &\textrm{B}\itOmega \\
\textrm{B}\itTheta &\textrm{B}\itPi &\textrm{B}\itPhi
\end{tabular}
\caption{Boldface variants of slanted greek letters}
\end{table}
\begin{table}
\begin{tabular}{*8l}
\P\upalpha &\P\uptheta &\P\upo &\P\uptau \\
\P\upbeta &\P\upvartheta &\P\uppi &\P\upupsilon \\
\P\upgamma &\P\upiota &\P\upvarpi &\P\upphi \\
\P\updelta &\P\upkappa &\P\uprho &\P\upvarphi \\
\P\upepsilon &\P\uplambda &\P\varrho &\P\upchi \\
\P\varepsilon &\P\upmu &\P\upsigma &\P\uppsi \\
\P\upzeta &\P\upnu &\P\upvarsigma &\P\upomega \\
\P\upeta &\P\upxi \\
\\
\textrm{B}\Gamma &\textrm{B}\Lambda &\textrm{B}\Sigma &\textrm{B}\Psi \\
\textrm{B}\Delta &\textrm{B}\Xi &\textrm{B}\Upsilon &\textrm{B}\Omega \\
\textrm{B}\Theta &\textrm{B}\Pi &\textrm{B}\Phi
\end{tabular}
\caption{Boldface variants of upright greek letters}
\end{table}
\end{document}
\endinput
| 6b84d84bd40fd00eaa7c0c13f917a902721dd9d2 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
A dissipative soliton is a stable, strongly localized structure forming inside a nonlinear dissipative system under suitable conditions \cite{Akhmediev}. Its applications range from optics, condensed-matter physics, cosmology to biology and medicine. Dissipative solitons arise in an open nonlinear system, far from equilibrium, and a continuous supply of energy is essential for them. More specifically, pulse-like dissipative solitons form inside a nonlinear active medium as a result of double balance between the medium's nonlinearity and its dispersion and between the gain and loss mechanisms that change pulse energy. Owing to this dual balance, the parameters of a dissipative soliton, such as its amplitude, width, chirp, and phase, do not depend on the initial conditions.
Active optical waveguides provide a fertile ground for observing optical dissipative solitons (ODSs) by launching short optical pulses inside them. In practice, however, such ODSs are sensitive to perturbations such as higher-order dispersion and self-steepening that become non-negligible for femtosecond pulses. Another important nonlinear effect for such short pulses is the intrapulse Raman scattering (IRS) that leads to a continuous red-shift of the pulse spectrum. In this paper we study the effects of IRS and other perturbations on the ODS dynamics through a variational approach technique \cite{Bondeson}. The variational technique is a standard method used extensively for both the dissipative \cite{Kaup} and non-dissipative \cite{Anderson} soliton systems. Its application is straightforward for conservative (non-dissipative) systems by choosing a suitable Lagrangian density \cite{Anderson}. The Lagrangian needs to be modified in case of dissipative systems such that it consists of a conservative part and a dissipative part \cite{Cerda}. Construction of a Rayleigh dissipation function is an alternative method to handle the dissipative effects \cite{Royjlt}. In all cases, The Lagrangian density is reduced by integrating over time. This reduction process requires a suitable \textit{ansatz}. The variation technique makes the assumption that the functional form of the ansatz remains intact in presence of a small perturbation but all parameters appearing in the anstaz (amplitude, width, position, phase, frequency etc.) may evolve with propagation. The reduced variational problem, followed by the Ritz optimization, leads to a set of coupled ordinary differential equation (ODE) that governs the evolution of individual pulse parameters under the influence of the perturbation~\cite{GPAbook1}.
A proper choice of the ansatz is critical for success of any variational approach. For example, soliton perturbation theory uses the hyperbolic-secant profile of a Kerr soliton as its ansatz with considerable success~\cite{GPAbook1}. However, this form will not be suitable for ODSs as they represent chirped optical pulses. In this work we adopt the Pereira--Stenflo solution~\cite{PS} of the Ginzburg--Landau equation (GLE) and show that the choice of this solution as an ansatz to the variational problem is much superior compared to the choice of a Kerr soliton. We examine the dynamics of various pulse parameters and predict accurately both the magnitude of the spectral red-shift of the ODS initiated by the IRS and corresponding changes in its speed. We also show that the ODS undergoes a slight blue-shift when self-steepening acts as a perturbation. The characteristic shift in the ODS location by the third-order dispersion (TOD) is also captured by the variational treatment presented here. As a special case, we consider the ODS formation inside an active silicon waveguide where free carriers are generated through multi-photon absorption and examine the perturbing effects of free carriers on an ODS\@. To verify the accuracy of our variational results, we compare them to the full numerical solution of the GLE and find a reasonable agreement between the two. We also propose some closed-form solutions which may prove more convenient to use in practice.
\section{THEORY}
\label{Ginzburg--Landau Equation}
To be realistic and to take into account several practical perturbations, we choose a silicon-based, active, nano-photonic waveguide \cite{Agazzi} and study the formation and evolution of ODSs in such a system. In such a waveguide. the leading loss mechanism comes from two-photon absorption (TPA) when pumped at a wavelength below 2.2~$\mu$m. As a consequence of TPA, free carriers are generated inside the waveguide that introduce additional loss so-called free-carrier absorption (FCA) and also change the refractive index \cite{Dieter, Tomita} through a phenomenon called free-carrier dispersion (FCD). In our model we take account these effects by coupling the carrier dynamics with the complex GLE that governs the pulse dynamics \cite{Lin,Roy-M-B}. This equation is a kind of nonlinear Schr\"{o}dinger equation with complex coefficients representing growth and damping \citep{PS, Anderson_Scr}. Its classical solution is known as the Pereira--Stenflo soliton \cite{PS} and it constitutes a specific example of dissipative solitons.
The extended GLE describing evolution of optical pulses inside a silicon-based active waveguide can be written in the following normalized form \cite{Roy-M-B, GPAbook2},
\begin{align} \label{gl}
i\frac{\partial u}{\partial \xi }-\frac{1}{2}sgn\left( {{\beta }_{2}} \right)\frac{{{\partial }^{2}}u}{\partial {{\tau }^{2}}}-i\left( {{g}_{0}}+{{g}_{2}}\frac{{{\partial }^{2}}}{\partial {{\tau }^{2}}} \right)u+i\alpha u \nonumber \\
+\left( 1+iK \right){{\left| u \right|}^{2}}u -i{{\delta }_{3}}\frac{{{\partial }^{3}}u}{\partial {{\tau }^{3}}} -\tau_R u \frac{\partial {\left| u \right|}^{2}}{\partial \tau } \nonumber \\
+i s\frac{\partial{({\left| u \right|}^{2} u)}}{\partial \tau} +\left( \frac{i}{2}-\mu \right){{\phi }_{c}}u=0
\end{align}
where the free-carrier effects are included through the normalized density parameter $\phi_c$ that satisfies the rate equation \cite{Lin},
\begin{equation} \label{ansatz}
d\phi_c/d\tau=\theta|u|^4-\tau_c \phi_c .
\end{equation}
The time and distance variables are normalized as $\tau=t/t_0$ and $\xi=z/L_D$, where $t_0$ is the initial pulse width and $L_D=t_0^2/|\beta_2 (\omega_0)|$ is the dispersion length, $\beta_2(\omega_0)$ being the group-velocity dispersion coefficient at the carrier frequency $\omega_0$.
The preceding equations contain multiple dimensionless parameters. The TOD, IRS and self-stepping parameters are normalized as $\delta_3=\beta_3/(3!|\beta_2| t_0)$, $\tau_R=T_R/t_0$ and $s=1/(\omega_0 t_0)$, where $T_R$ is the first moment of the Raman response function \cite{GPAbook1}. The field amplitude ($A$) is rescaled as, $A = u\sqrt{P_0}$, where peak power, $P_0=|\beta_2 (\omega_0 )|/(t_0^2 \gamma_R)$, $\gamma_R=k_0 n_2/A_{eff}$ and $n_2\approx(4\pm 1.5)\times 10^{-18} \ m^2 W^{-1}$ is the Kerr-nonlinear coefficient of silicon. The dimensionless TPA coefficient is given as, $K=\gamma_I/\gamma_R=\beta_{TPA}\lambda_0/(4\pi n_2)$, where, $\beta_{TPA}\approx 8\times 10^{-12} \ m W^{-1}$ and $\gamma_I=\beta_{TPA}/(2A_{eff})$. The linear loss coefficient is normalized as $\alpha=\alpha_l L_D$. The free-carrier density $N_c$ is related to $\phi_c$ as $\phi_c=\sigma N_c L_D$ where $\sigma\approx 1.45\times 10^{-21}\ m^2$ is the FCA cross section of silicon at $\lambda_0=1.55~\mu m$ \cite{Rong}. The generation of free carriers is regulated by the parameter $\theta=\beta_{TPA} |\beta_2 |\sigma/(2\hbar\omega_0 A_{eff}^2 t_0 \gamma_R^2)$ \cite{Lin-Z-P}. The parameter $\mu=2\pi k_c/(\sigma \lambda_0)$ is the FCD coefficient with $k_c\approx 1.35\times 10^{-27}\ m^3$ \cite{Dinu}. The carrier recombination time $t_c$ is scaled as $\tau_c=t_0/t_c$. The gain $G$ and the gain dispersion coefficient ($g_2$) are normalized as $g = GL_D$ and $g_2= g(T_2/t_0)^2$, where dephasing time is $T_2$. The spectral wings of the pulse experience less gain due to a finite gain bandwidth related to $g_2$.
In the absence of TOD ($\delta_3=0$), IRS ($\tau_R=0$), self-steepening ($s=0$) and free carriers (i.e., $\phi_c=0$), Eq.\ (1) reduces to the standard GLE, which is known to have the stable ODS solution in the following form \cite{PS, GPAbook2, Desurvire}:
\begin{equation} \label{ansatz}
u\left( \xi, \tau \right)={{u}_{0}}{{\left[ \text{sech}\left( \eta \tau \right) \right]}^{\left( 1+ia \right)}}{{e}^{i\text{ }\!\!\Gamma\!\!\text{ }\xi }},
\end{equation}
where the four parameters $u_0,\ \eta,\ a$ and $\Gamma$ are given by~\cite{GPAbook2}:
\begin{subequations} \label{ansatz_part}
\begin{align}
|u_0|^{2}& = \frac{(g_{0}-\alpha)}{K}\left[ 1-\frac{sgn(\beta_{2})a/2 + g_{2}} {{g}_{2} ({{a}^{2}}-1 )- sgn(\beta_2)a} \right], \\
\eta^{2} &= \frac{({{g}_{0}}-\alpha )}{ \left[{{g}_{2}}\left( {{a}^{2}}-1 \right)-sgn\left( {{\beta }_{2}} \right)a \right]}, \\
\Gamma &= \frac{{{\eta }^{2}}}{2}\left[ sgn\left( {{\beta }_{2}} \right)\left( {{a}^{2}}-1 \right)+4a{{g}_{2}} \right], \\
a &= \frac{H-\sqrt{{{H}^{2}}+2{{\delta }^{2}}}}{\delta}.
\end{align}
\end{subequations}
Here, $H=-[(3/2)sgn(\beta_2 )+3g_2 K]$ and $\delta=-[2g_2-sgn(\beta_2)K]$. The preceding solution was first obtained in 1977 and is known as the Pereira--Stenflo soliton~\cite{PS}
\section{VARIATIONAL ANALYSIS}
\label{perturbative}
The ODS solution exists only when four terms in Eq.\ \eqref{gl} related to TOD ($\delta_3=0$), IRS ($\tau_R=0$), self-steepening ($s=0$) and free carriers ($\phi_c=0$) are neglected. The important question is how these terms affect the ODS solution. One can study their impact by solving Eq.\ \eqref{gl} numerically. However, this approach hinders any physical insight. In this section we treat the four terms as small perturbations and study their impact through a variational analysis. The Variational method has been used with success in the past for many pulse-propagation problems \cite{Bondeson,Kaup,Anderson,Cerda,Royjlt}. It requires a suitable \textit{ansatz} for the pulse shape and makes the assumption that the functional form of the pulse shape remains intact in presence of small perturbations but its parameters appearing in the ansatz (amplitude, width, position, phase, frequency etc.) may evolve with propagation. For our problem, it is natural that we choose the Pereira--Stenflo solution in Eq.\ \eqref{ansatz} as our ansatz since it is the exact solution of Eq.\ \eqref{gl} in the absence of perturbations induced by TOD, IRS, self-steepening ($s=0$) and free-carrier generation. We thus choose the following ansatz:
\begin{align}\label{nls0}
u\left( \xi ,~\tau \right)={{u}_{0}}\left( \xi \right){{\left[ \text{sech}\left\{ \eta \left( \xi \right)\left( \tau -{{\tau }_{p}}\left( \xi \right) \right) \right\} \right]}^{\left\{ 1+ia\left( \xi \right) \right\}}} \nonumber \\
exp\left[ i\left\{ \phi \left( \xi \right)-\text{ }\Omega\left( \xi \right)\left( \tau -{{\tau }_{p}}\left( \xi \right) \right) \right\} \right],
\end{align}
where the six parameters $u_0,~\eta,~\tau_p,~\phi,~a$ and $\Omega$ are now assumed to depend on $\xi$. We first write Eq.~\eqref{gl} in the form of a perturbed nonlinear Schr\"{o}dinger equation~\cite{Anderson, GPAbook1}:
\begin{equation} \label{nls1}
i\frac{\partial u}{\partial\xi}+\frac{1}{2}\frac{\partial^{2}u}
{\partial\tau^{2}}+ |u|^{2}u = i\epsilon(u),
\end{equation}
where we have chosen the dispersion to be anomalous $(\beta_2<0)$ and define $\epsilon(u)$ as,
\begin{align} \label{nls2}
\epsilon(u) = \delta_3\frac{\partial^3u}{\partial{{\tau}^{3}}}-i\tau_R & \frac{\partial{|u|}^{2}}{\partial\tau} -s\frac{\partial ({|u |}^{2} u)}{\partial \tau} -\left(\frac{1}{2}+i\mu\right)\phi_c u \nonumber \\
&~~~~~ -K|u|^{2}u +{g}_{0}u +{{g}_{2}}\frac{{{\partial }^{2}}}{\partial {{\tau }^{2}}}u.
\end{align}
We then follow a standard procedure \cite{GPAbook1} and introduce the Lagrangian density appropriate for Eq.\ \eqref{nls1} and integrate over $\tau$ using the ansatz in Eq.\ \eqref{nls0} to obtain the following reduced Lagrangian:
\begin{gather}
L = \frac{2{{u}_{0}}^{2}}{\eta }\left( \frac{\partial \phi }{\partial \xi }+\Omega \frac{\partial {{\tau }_{p}}}{\partial \xi } \right)-\frac{a {{u}_{0}}^{2}}{{{\eta }^{2}}}\frac{\partial \eta }{\partial \xi }+C\frac{{{u}_{0}}^{2}}{\eta }\frac{\partial a}{\partial \xi } \nonumber \\
+\frac{\eta {{u}_{0}}^{2}}{3}\left( 1+{{a}^{2}} \right)+\frac{{{u}_{0}}^{2}}{\eta }\left( {{\Omega }^{2}}-\frac{2}{3}{{u}_{0}}^{2} \right) \nonumber \\
+i\int_{-\infty }^{\infty}(\epsilon{u}^{*}-\epsilon^* u)\,d\tau,
\end{gather}
where $C=[\ln(4)-2]$. The next step is to use the Euler-Lagrange equation for each pulse parameter to obtain a set of coupled ODEs for the six parameters that describe the overall soliton dynamics \cite{GPAbook1,Hasegawa-K}. These equations govern the evolution of pulse energy $(E=\int_{-\infty}^{\infty} |u|^{2}\,d\tau)$, temporal position $\tau_p$, frequency shift $\Omega$, amplitude $\eta$, frequency chirp $a$, and phase $\phi$ and have the form
\begin{align}
\frac{dE}{d\xi }&=\frac{d}{d\xi }\left( \frac{2{{u}_{0}}^{2}}{\eta} \right) =2{\rm Re}\int\limits_{-\infty }^{\infty }{\epsilon {{u}^{*}}}\,d\tau, \label{var1}\\
\frac{d{{\tau}_{p}}}{d\xi }& = -\Omega +\frac{\eta }{{{u}_{0}}^{2}} \int\limits_{-\infty }^{\infty }{\left( \tau -{{\tau }_{p}} \right){\rm Re}\left( \epsilon {{u}^{*}} \right)\,d\tau }, \label{var2}\\
\frac{d\Omega }{d\xi} &= \frac{{{\eta }^{2}}}{{{u}_{0}}^{2}}\int\limits_{-\infty }^{\infty }{\tanh\left[ \eta \left( \tau -{{\tau }_{p}} \right) \right]{\rm Re}\left[ \left( a+i \right)\epsilon {{u}^{*}} \right]\,d\tau }, \label{var3}\\
\frac{d\eta }{d\xi} &= \frac{2{{\eta }^{2}}}{{{u}_{0}}^{2}}\int\limits_{-\infty }^{\infty }{\ln\left[ {\rm sech}\left\{ \eta \left( \tau -{{\tau }_{p}} \right) \right\} \right]{\rm Re}\left( \epsilon {{u}^{*}} \right)\,d\tau } \nonumber \\
&\hspace{.5cm} -C\frac{\eta^{2}}{2u_{0}^{2}}E_\xi+\frac{2{{\eta}^{3}}a}{3}, \label{var4}
\end{align}
\begin{align}
\frac{da}{d\xi}& = -\frac{a\eta }{2{{u}_{0}}^{2}}E_\xi +\frac{2}{3}{{u}_{0}}^{2}-\frac{2}{3}{{\eta }^{2}}\left( 1+{{a}^{2}} \right) +\frac{\eta }{{{u}_{0}}^{2}}{\rm Im}\int\limits_{-\infty }^{\infty }{\epsilon {{u}^{*}}}\,d\tau \nonumber \\
&\hspace{0cm} -\frac{2{{\eta }^{2}}}{{{u}_{0}}^{2}}{\rm Im}\int\limits_{-\infty }^{\infty }{\left( \tau -{{\tau }_{p}} \right)tanh\left[ \eta \left( \tau -{{\tau }_{p}} \right) \right]\left( 1-ia \right)\epsilon {{u}^{*}}\,d\tau }, \label{var5}\\
\frac{d\phi}{d\xi} &= \frac{a}{2\eta}\eta_\xi -\frac{C}{2}a_\xi -\Omega \,{\tau_p}_\xi + \frac{2}{3}u_0^2 -\frac{1}{6}\eta^2(1+a^2) -\frac{1}{2}\Omega^2 \nonumber \\ &\hspace{0.5cm} +\frac{\eta}{2u_0^2}{\rm Im}\int\limits_{-\infty }^{\infty}{\epsilon {{u}^{*}}}\,d\tau, \label{var6}
\end{align}
where Re and Im stand for real and imaginary parts. The final step is to evaluate all the integrals using $\epsilon(u)$ given in Eq.\ \eqref{nls2}. It results in the following set of six coupled differential equations:
\begin{align}
\frac{dE}{d\xi }&=\frac{2}{3}(3g_0-K\eta E)E- \frac{2}{3}g_2[(1+a^2)\eta^2+3\Omega^2]E \nonumber \\&\hspace{0.5cm} -\frac{1}{6}\theta \eta E^3 , \label{var7} \\
\frac{d{{\tau }_{p}}}{d\xi }&=-\left(1 +2{{g}_{2}}a \right)\Omega -\frac{7}{72}\theta E^2 +\delta_3\left[(1+a^2)\eta^2 + 3\Omega^2 \right] \nonumber \\&\hspace{0.5cm} +\frac{1}{2}s \eta E , \label{var8}\\
\frac{d\Omega }{d\xi }&=-\frac{4}{3}{{g}_{2}}\left( 1+{{a}^{2}} \right)\Omega {{\eta }^{2}} -\frac{4}{15}\tau_R E{{\eta }^{3}} + \frac{4}{15}s a E \eta^3 \nonumber \\&\hspace{0.5cm} + \frac{2}{15}\left(\mu -\frac{a}{2}\right)\theta \eta^2 E^2 , \label{var9}\\
\frac{d\eta }{d\xi }&=\frac{2}{3} (a-EK)\eta^2-\frac{4}{9}(2-a^2)g_2\eta^3 -\frac{1}{6}C\theta \eta^2 E^2 \nonumber \\ &\hspace{0.5cm} -4\delta_3 a \Omega \eta^3,\label{var10}\\
\frac{da}{d\xi }&=\frac{1}{3}(1+aK)E\eta -\frac{2}{3} (1+ag_2)(1+a^2)\eta^2 -\frac{1}{6}\theta a \eta E^2 \nonumber \\ &\hspace{0.5cm} + \frac{1}{3}s \Omega \eta E +4\delta_3 \Omega \eta^2(1+ a^2), \label{var11}\\
\frac{d\phi}{d\xi}&=\frac{a}{2\eta}\eta_\xi -\frac{C}{2}a_\xi -\Omega{\tau_p}_\xi + \frac{1}{3}\eta E -\frac{1}{6}\eta^2(1+a^2)-\frac{1}{2}\Omega^2 \nonumber \\ &\hspace{-0.2cm} -\frac{1}{6}\mu \theta \eta E^2 +\frac{1}{3}s\Omega E \eta + \delta_3 \left[(1+a^2)\Omega \eta^2 + \Omega^3 \right].
\label{var12}
\end{align}
These equations provide considerable physical insight since they show which perturbations affect a specific pulse parameter. For example, the Raman parameter $\tau_R$ appears only in the equation for the frequency shift $\Omega$ and the term containing it has a negative sign. This immediately shows that the IRS leads to a spectral red-shift of the ODS\@. In contrast, the self-steepening parameter $s$ appears in the frequency equation in a term with the positive sign and shows that self-steepening will reduce the Raman-induced spectral red-shift. This kind of physical insight is very valuable in interpreting the numerical results. It is noteworthy that the phase $\phi$ does not appear in any equation except the last one. This indicates that the numerical value of the soliton's phase does not affect any of its other parameters. For this reason, we ignore the phase equation in the following discussion. In the next section we discuss the effects of various perturbations on the evolution of the ODS parameters and also compare variational results with the results obtained from direct simulation of Eq.\ \eqref{gl}.
\section{Full Numerical Simulations}
\begin{figure*}[tb!]
\begin{center}
\epsfig{file=1_a.jpg,trim=4.5in 0.05in 6.05in 0.0in,clip=true, width=59mm}
\epsfig{file=1_b.jpg,trim=5in 0.05in 5.7in 0.0in,clip=true, width=58mm}
\epsfig{file=1_c.jpg,trim=5in 0.05in 5.7in 0.0in,clip=true, width=58mm}
\vspace{0.5em}
~~~~~~~~(a)~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(b)~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(c)
\\
\epsfig{file=1_d.jpg,trim=5in 0.00in 5.7in 0.0in,clip=true, width=59mm}
\epsfig{file=1_e.jpg,trim=5in 0.00in 5.7in 0.0in,clip=true, width=59mm}
\epsfig{file=1_f.jpg,trim=5in 0.00in 5.7in 0.0in,clip=true, width=59mm}
\vspace{0.5em}
~~~~~~~~(d)~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(e)~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(f)
\caption{(Color online) Temporal (top) and spectral (bottom) evolution of an ODS in three cases under a single perturbation. (a, d) IRS acts alone with $\tau_R=0.1$; (b, e) self-steepening acts alone with $s=0.1$; (c, f) FCD acts alone with $\theta= 0.0044$. Other parameters used in the simulations are: $K=0.01,~g_0=0.01,~g_2=0.01$ and $\alpha=0$. The input (dotted trace) and output pulse shapes are also shown in the top panel.}\label{variational_all}
\end{center}
\end{figure*}
Before discussing the variational results, we solve Eq.~\eqref{gl} numerically and present the results for a realistic silicon active waveguide. More specifically, the individual and collective effects of various perturbation on the evolution of an ODS are discussed in this section. Since the temporal shape of the ODS is distorted rapidly in presence of TOD, which violets the basic assumption behind the variational technique, initially we study pulse dynamics by setting $\delta_3=0$ in Eq.~\eqref{gl}. We solve this equation with the standard split-step Fourier method \cite{GPAbook1} by taking the input pulse in the form of a Pereira--Stenflo soliton with the parameters given in Eq.\ (\ref{ansatz_part}). The parameter values used were $K=0.01,~g_0=0.01$, $g_2=0.01$, $\tau_R=0.1$, $s=0.1$, $\theta=0.0044$, and $\mu=3.7741$. The values of $\theta$ and $\mu$ are calculated by adopting the realistic values of device parameters.
Figure~\ref{variational_all} shows the temporal (top row) and spectral (bottom row) evolutions of the perturbed ODS in three cases: (a, d) only IRS, (b, e) only self-steepening, and (c, f) only free carriers perturb the ODS\@. As expected, IRS leads to a spectral red-shift and slows down the ODS considerably. However the red-shift saturates after some distance of propagation (around $\xi=20$). In the time domain, the ODS continues to shift because of a change in its speed induced by the red-shift. Our simulations confirm that the pulse width is also affected by the Raman term. In the case of self-steepening, the shape of the pulse remains almost intact and ODS slows down a bit even though its spectrum undergoes a small blue-shift. The influence of free carriers is more dramatic because of FCD that leads to a larger blue-shift with an acceleration of the pulse, consistent with the previously reported results~\cite{Roy-M-B}.
\section{Results of Variational Analysis}
\begin{figure*}[t!]
\begin{center}
\epsfig{file=2_a.pdf,trim=3.0in 0in 3.7in 0in,clip=true, width=60mm}
\epsfig{file=2_b.pdf,trim=3.0in 0in 3.7in 0in,clip=true, width=60mm}
\vspace{0.5em}
~~~~~~~~~(a)~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(b)
\\
\epsfig{file=2_c.pdf,trim=3.0in 0in 3.7in 0in,clip=true, width=60mm}
\epsfig{file=2_d.pdf,trim=3.0in 0in 3.7in 0in,clip=true, width=60mm}
\vspace{0.5em}
~~~~~~~~~(c)~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(d)
\\
\epsfig{file=2_legend.pdf,trim=-0.9in 5in 1.7in 0in,clip=true, width=110mm}
\vspace{0.5em}
\caption{(Color online) (a) Temporal position ($\tau_p$) and (b) frequency shift ($\Omega$) as a function of distance in four different cases listed at bottom. Changes in the peak intensity ($|u|^2_0$) and pulse width ($\tau_w$) in the same four cases are shown in parts (c) and (d) respectively. Solid lines show the variational predictions whereas circles represent the corresponding numerical data. Other parameters used in the simulations are: $K=0.01,~g_0=0.01,~g_2=0.01$ and $\alpha=0$.}
\label{variational_mm}
\end{center}
\end{figure*}
In this section we solve the coupled differential equations obtained with the variational approach [Eq.~\eqref{var7}-\eqref{var12}] and compare their predictions with the numerical simulations in Fig.~\ref{variational_mm}. The four parts of this figure compare changes in the pulse position $\tau_p$, spectral shift $\Omega$, peak intensity $|u|^2_0$, and the pulse width $\tau_w$. The red, blue, and green curves in each case correspond to the three cases shown in Fig.\ \ref{variational_all} when only physical process perturbs the ODS\@. The black curves show the case when all three perturbations are present simultaneously. In all cases, the solid lines show variational result and solid circles show the numerical predictions of Eq.~\eqref{gl}. The agreement between the variational and numerical results is remarkably good under so many diverse situations, indicating the suitability of our variational approach for perturbed ODSs.
\begin{figure}[tb!]
\begin{center}
\epsfig{file=3_a.pdf,trim=3.0in 0in 3.7in 0in,clip=true, width=44mm}\epsfig{file=3_b.pdf,trim=3.0in 0in 3.7in 0in,clip=true, width=44mm}
\vspace{0.5em}
~~~~~~~(a)~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(b)
\caption{(Color online) (a) Changes in pulse energy as a function of $\xi$ in the four cases of Fig.~\ref{variational_mm}. (b) Changes in the IRS-induced red shift and FCD-induced blue shift as predicted by the approximated analytic expressions.}
\label{closed_form}
\end{center}
\end{figure}
The set of coupled differential equations becomes more useful if we decouple them with suitable approximations. If we assume variations of $\eta$ and $a$ are relatively small (which is true for propagation distance $\xi<40$) and treat them as constants, we can integrate Eq.~\eqref{var9} analytically. The spectral red-shift owing to IRS can then be written in a close form as,
\begin{equation} \label{nls4}
\Omega(\xi)\approx -\Omega_R(1-e^{-\rho \xi}),
\end{equation}
where $\Omega_R=\tau_R E_{av}\eta/[5g_2(1+a^2)]$ and $\quad \rho=4g_2(1+a^2)\eta^2/3$. This equation shows how the red-shift increases with $\xi$ initially but saturates to a final value of $-\Omega_R$ when $\xi$ is large enough that $\rho\xi\gg 1$. Here changes in the total energy $E$ are approximated by its average over the distance at which $\Omega$ is calculated. Under the same assumptions, we can integrate Eq.~\eqref{var8} for the temporal shift analytically to obtain
\begin{equation} \label{nls5}
\tau_p(\xi) \approx \Omega_R(1+2g_2a)[\xi-\rho^{-1}(1-e^{-\rho\xi})].
\end{equation}
This equation shows that once the red-shift satuarates ($\rho\xi\gg 1$), $\tau_p$ varies linearly with $\xi$; this is clearly evident in Fig.~\ref{variational_mm}(a).
In the same way we can derive an approximate analytic expression for the spectral blue-shift induced by FCD\@. The results has the same form as for IRS, and the blue-shift is,
\begin{equation} \label{nls6}
\Omega(\xi)\approx\Omega_{FC}(1-e^{-\rho\xi}),
\end{equation}
where the saturated value becomes $\Omega_{FC}=(\mu-a/2)\theta E_{av }^2/[10g_2(1+a^2)]$. The temporal shift due to the FC can also be approximated as,
\begin{equation} \label{nls7}
\tau_p(\xi)\approx -\Omega_{FC}(1+2g_2a)[(1+\chi_{FC})\xi-\rho^{-1}(1-e^{-\rho\xi})].
\end{equation}
Where $\chi_{FC}=35g_2(1+a^2)/[36(\mu-a/2)(1+2g_2a)]$. The preceding results use the concept of average pulse energy to account for energy variations inside the waveguide in an average sense. In Fig.~\ref{closed_form}(a) we plot energy variations under different perturbations for the 4 cases shown in Fig.~\ref{variational_mm}. Depending on the distance and the mechanism involved, pulse energy may be reduced by more than 50\%. In part (b) we plot the spectral shifts under when IRS and free carriers act as perturbations and compare the full numerical results with the approximate analytical expressions derived above. The agreement with simulations is reasonable in the case of FCD when we use the average energy (red dashed curves). If we use the initial value of pulse energy, agreement is good at short distances but becomes increasingly poor for longer distances (blue dashed curve). In the Raman case, the red dashed curve disagrees initially with numerical results but merges asymptotically to the saturated value predicted by the full calculation. The mismatch at the initial stage occurs because we assumed $\eta$ to be constant, which is not the case. We emphasize that our closed form expressions help us to understand the pulse dynamics qualitatively. However, the inclusion of all variations of $a$, $\eta$ and $E$ is essential for accurate results.
As a final test of the set of ODEs derived variationally, we solve them under zero perturbation. If the derived ODEs are correct, they should provide the exact Pereira--Stenflo soliton when all perturbations are switched off. Figure \ref{fig_further}(a) shows that this is indeed the case. The simulated temporal profile at $\xi=200$ overlaps exactly with the variational temporal profile when there is no perturbation. We stress that the use of ODS at the input end is essential while solving Eq.\ (1). In Fig.~\ref{fig_further}(b) we compare the IRS-induced red-shifts obtained using a standard soliton and the ODS at the input end. It is evident that the variational results are consistent with the data obtained using the Pereira--Stenflo soliton as an input but not when a sech-profile of a standard soliton is used for solving Eq.\ (1).
\section{Impact of TOD}
\begin{figure}[tb!]
\begin{center}
\epsfig{file=4_a.pdf,trim=3.0in 0in 3.7in 0in,clip=true, width=44mm}\epsfig{file=4_b.pdf,trim=3.0in 0in 3.7in 0in,clip=true, width=44mm}
\vspace{0.5em}
~~~~~~~(a)~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(b)
\caption{(Color online) (a) Comparison of output intensity profile at $\xi=200$ between full numerical simulation (red dots) and variational prediction (blue trace). (b) Frequency shift as a function of distance for $\tau_R=0.1$. Blue circles and red squares correspond to sech-shape pulse and ODS inputs respectively. The solid green line shows variational prediction of the frequency shift. Parameters used in the simulations are: $K=0.01,~g_0=0.01,~g_2=0.0001$ and $\alpha=0$.}
\label{fig_further}
\end{center}
\end{figure}
So far, we have ignored the TOD perturbations. However, our variational analysis includes the TOD effects through the $\delta_3$ parameter. Indeed, the temporal position $\tau_p$ of the ODS depends explicitly on $\delta_3$ in Eq.~\eqref{var8}. If we ignore all other perturbations and set $\Omega=\theta=s=0$ in this equation, we get a simple relation $\tau_p(\xi) \approx \delta_3(1+a^2)\eta^2\xi$, provided both $\eta$ and $a$ remain nearly constant. It shows that TOD shifts the soliton position linearly with distance, a well-known result for the standard solitons. To see if this linear behavior persists for an ODS, we solve Eq.\ (1) by taking TOD as the only perturbation ($\tau_R=s=\theta=0$).
Figure~\ref{variational}(a) shows the evolution of ODS under TOD acting as the sole perturbation using $\delta_3=0.1$. We observe that the ODS nearly preserves its shape with only small variations in the pulse width (mild breathing). The pulse shape at $\xi=200$ is plotted on top in Fig.~\ref{variational}(a), and it shows a small temporal shift from the initial ODS position. We compare this temporal shift (red dots) with the variational prediction (solid line) in Fig.~\ref{variational}(b). The two agree reasonably well for up to $\xi=100$ with increasing departure for longer distances. This agreement is expected only for relatively low values of $\delta_3$. Indeed, significant distortions of the pulse shape are observed for high values of $\delta_3$. Under weak TOD perturbation, the ODS maintains its overall shape over relatively long distances, and the variational analysis works reasonably well in that situation.
\begin{figure}[tb!]
\begin{center}
\epsfig{file=5_a.jpg,trim=5.1in 0.3in 5.6in 0.0in,clip=true, width=40mm}
\epsfig{file=5_b.pdf,trim=3.3in -0.1in 3.8in 0.0in,clip=true, width=41mm}
~~~~~~~~~~~~~~(a)~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~(b)
\caption{(Color online) (a) Temporal evolution of ODS under the TOD perturbation.
Input (dotted trace) and output pulse shapes are shown on top. (b) Changes in ODS position with distance as predicted by the variational technique (solid blue line) are compared with numerical data (red circles). The parameters used in the simulations are: $K=0.01,~g_0=0.001,~g_2=0.001$ and $\delta_3=0.1$.} \label{variational}
\end{center}
\end{figure}
\section{CONCLUSIONS}
By exploiting the standard variational technique, we study the dynamics of a perturbed dissipative soliton excited inside an active semiconductor waveguide. The pulse evolution is governed by an extended GLE containing additional terms that have their origin in higher-order effects such as TOD, self-steepening, Raman scattering, and free-carrier generation. We treat these terms as small perturbations and carry our variational analysis after choosing a dissipative soliton as our ansatz. Being an exact solution of the unperturbed GLE, this chirped soliton maintains its shape inside the active waveguide with slow evolution of its parameters with distance. The variational treatment provides with a set of coupled ordinary differential equations. We have shown that solution of this set of equations predicts quite well how the individual pulse parameters will evolve with distance. we solve the GLE numerically using the split-step Fourier method and show that the variational predictions agree well with full numerical simulations. We also propose simple analytical solutions for the Raman-induced spectra red-shifts and the corresponding temporal shift of the pulse peak. With suitable approximations, our closed-form expressions should prove useful in practice. In summary, our semi-analytical treatment provides significant insights in understanding the complex dynamics of perturbed dissipative solitons.
\section*{Acknowledgements}
This work is supported by SRIC, Indian Institute of Technology, under the project ISIRD. A.S. acknowledges MHRD, India for a research fellowship.
| 6c6d6286ecc8e9c8f9ea5535d07e73cbde509b3c | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In field theory, conformal maps are fundamental for our understanding of
spacetime. Moreover, the existence of a conformal vector field on a manifold
can provide valuable information, which can go up to full classification
results, \cite{Alekseevsky}, \cite{Camargo}, \cite{Caminha}, \cite{Kuhnel
survey},\ on the metric structure.
Among the applications of (pseudo-)Finsler geometry, field-theoretical ones
are the most numerous, e.g., \cite{Asanov}, \cite{Barletta}, \cit
{Bogoslovsky}, \cite{Gibbons}, \cite{Javaloyes}, \cite{Kouretsis}, \cit
{Kostelecky}, \cite{Stavrinos}, \cite{Vacaru}. But these applications
typically require metrics to be of Lorentzian signature. And, while on
conformal maps between positive definite Finsler spaces there exists quite a
rich literature, \cite{Aikou}, \cite{Asanjarani}, \cite{Cheng}, \cit
{Bidabad}, \cite{Hashiguchi}, \cite{Matsumoto}, \cite{Youssef}, \cite{Zhang
, in pseudo-Finsler spaces, the situation is completely different. Apart
from a very few papers dedicated to the particular case of isometries, \cit
{Herrera}, \cite{Torrome-isometries} or to a particular metric, \cite{Pavlov
, to the best of our knowledge, even basic questions related to conformal
transformations have not been tackled yet.
\bigskip
Conformal groups of pseudo-Finsler metrics have a much more complicated -
and more interesting - structure than both pseudo-Riemannian and Finslerian
conformal groups. To prove this statement, we present in Section \re
{BM_example} some classes of examples of flat (locally Minkowski)
pseudo-Finsler spaces whose conformal symmetries depend on arbitrary
functions. Comparatively, in dimension $n\geq 3,$ conformal symmetries of a
pseudo-Euclidean can only be similarities, inversions and compositions
thereof, \cite{Haantjes}, while the only conformal symmetries of a
non-Euclidean flat Finsler space are similarities, \cite{Matveev}. This
hints at the fact that extending results from either pseudo-Riemannian or
Finsler geometry to pseudo-Finsler spaces can be far from straightforward -
and some of these results might very well fail when passing to
pseudo-Finsler spaces.
\bigskip\
In the following sections, we focus on two topics:
1. \textit{The behavior of geodesics under conformal mappings.} Here, we
prove that several results in pseudo-Riemannian geometry (which are
fundamental for general relativity)\ can still be extended to pseudo-Finsler
spaces:
- In dimension greater than 1, any mapping between two pseudo-Finsler
structures which is both conformal and projective is a similarity. In other
words, Weyl's statement (e.g., \cite{Cheng}) that \textit{projective and
conformal properties of a metric space univocally determine its metric up to
a dilation factor} remains true in pseudo-Finsler spaces.
- Lightlike geodesics are preserved, up to re-parametrization, under
arbitrary conformal mappings.
- A conservation law for conformal vector fields along lightlike geodesics.
\bigskip
2. \textit{Conformal vector fields}. In positive definite Finsler spaces,
the technique of averaged Riemannian metrics allows one to prove profound
results regarding conformal transformations, by reducing the corresponding
problems to their Riemannian counterparts, \cite{Matveev}. But,
unfortunately, this technique is not available in pseudo-Finsler spaces, as
noticed in \cite{Torrome-isometries}.
Still, dealing with conformal vector fields, we can find a partial
substitute for this method. Given a pseudo-Finslerian metric tensor $g$ on
some manifold $M,$ an associated Riemannian metric is a pseudo-Riemannian
metric $g^{\xi }:=g\circ \xi ,$ where $\xi $ is a vector field on $M.$
Associated Riemannian metrics have a series of appealing properties (e.g.,
smoothness, same signature as $g$) and behave well under conformal
transformations of $g;$ more precisely, we show (Lemma \re
{conformal_vector_fields_lemma}) that,\ if $\xi $ is a conformal vector
field for a pseudo-Finsler metric $g,$ then $\xi $ is also a conformal
vector field for $g^{\xi }.$ This way, some results in pseudo-Riemannian
geometry become available in the more general context of Finsler metrics. As
an example, we extend to pseudo-Finsler spaces two results on Killing vector
fields in \cite{Sanchez}.
Also, we prove that any essential conformal vector field of a pseudo-Finsler
metric has to be lightlike at least at a point.
\bigskip
The paper is organized as follows. Section 2 presents some preliminary
notions and results. Section 3 deals with the basic conformality notions and
examples of pseudo-Finslerian conformal maps. Section 4 is devoted to the
behavior of geodesics under conformal transformations. In Sections 5 and 6,
we discuss pseudo-Finslerian conformal vector fields.
\section{Pseudo-Finsler spaces. Finsler spacetimes}
Let $M$ be a $\mathcal{C}^{\mathcal{\infty }}$-smooth, connected manifold of
dimension $n$ and $(TM,\pi ,M),$ its tangent bundle. We denote by
(x^{i})_{i=\overline{0,n-1}}$ the coordinates of a point $x\in M$ in a local
chart $(U,\varphi )$ and consider local charts $(\pi ^{-1}(U),\Phi ),$ $\Phi
=(x^{i},y^{i})_{i=\overline{0,n-1}}$ on $TM$ induced by the choice of the
natural basis $\left\{ \partial _{i}\right\} $ in each tangent space. Commas
$_{,i}$ will denote differentiation with respect to $x^{i}$ and dots
_{\cdot i},$ differentiation with respect to $y^{i}.$ The set of sections of
any fibered manifold $E$ over $M$ will be denoted by $\Gamma (E).$
\bigskip
Consider a non-empty open submanifold$\ $ $A\subset TM,$ with $\pi (A)=M$
and $0\not\in A.$ We assume that each $A_{x}:=T_{x}M\cap A,$ $x\in M,$ is a
positive conic set, i.e., $\forall \alpha >0,$ $\forall y\in A_{x}:\alpha
y\in A_{x}.$ Then the triple $(A,\pi _{|A},M),$ where $\pi _{|A}$ is the
restriction of $\pi $ to $A,$ is a fibered manifold over $M.$ For $x\in M$,
elements $y\in A_{x}$ are called \textit{admissible vectors }at $x.$
\begin{definition}
(\cite{Bejancu}): Fix a natural number $0\leq q<n.$ A smooth function
L:A\rightarrow \mathbb{R}$ defines a \textit{pseudo-Finsler structure }
(M,A,L)$ on $M$ if, at any point $(x,y)\in A$ and in any local chart $(\pi
^{-1}(U),\Phi )$ around $(x,y):$
1)$\ L(x,\alpha y)=\alpha ^{2}L(x,y),$ $\forall \alpha >0$;
2) the matrix\ $g_{ij}(x,y)=\dfrac{1}{2}\dfrac{\partial ^{2}L}{\partial
y^{i}\partial y^{j}}(x,y)$ has $q$ negative and $n-q$ positive eigenvalues.
\end{definition}
The \textit{Finsler Lagrangian (Finslerian energy) }$L$ can always be
prolonged by continuity to the closure $\bar{A}.$ In particular, we can set
L(x,0)=0$.
\bigskip
\textbf{Particular cases.}
1)\ If $q=0,$ then the Finsler structure $(M,A,L)$ is called \textit
positive definite.} If $A=TM\backslash \{0\},$ then it is called \textit
smooth. }A smooth and positive definite pseudo-Finsler structure is a
\textit{Finsler structure}.
2)\ A pseudo-Finsler space $(M,A,L)$ with $q=n-1,$ is called a \textit
Lorentz-Finsler space} or a \textit{Finsler spacetime}.
In a Finsler spacetime, $ds^{2}=L(x,dx)$ is interpreted as spacetime
interval - and it allows the introduction of the basic causality notions.
For any point $x\in M,$ an admissible vector $y\in A_{x}$\ will be called:\
\textit{timelike}, if $L(x,y)>0,$ \textit{spacelike}, if $L(x,y)<0$ and
\textit{null }or \textit{lightlike,} if $L(x,y)=0.$ Accordingly, a curve
c:[a,b]\rightarrow M,$ $t\mapsto c(t)$ is called timelike (respectively,
null, spacelike) if its tangent vector $\dot{c}$ is everywhere timelike
(respectively, null, spacelike)\footnote
This terminology will be actually used not only in Finsler spacetimes, but
in pseudo-Finsler spaces of arbitrary signature.}.
3)\ A pseudo-Finsler space $(M,A,L)$ is (\textit{pseudo)-Riemannian}, if, in
any local chart, $g_{ij}=g_{ij}(x)$ and \textit{flat (locally Minkowski) }if
around any point of $A,$ there exists a local chart in which
g_{ij}=g_{ij}(y)$ only.
A curve on $M$ is called \textit{admissible} if its tangent vector is
everywhere admissible. In the following, we will assume that all the curves
under discussion are admissible. The arc length of a curve $c:t\in \lbrack
a,b]\mapsto (x^{i}(t))$ on $M$ is calculated as $l(c)=\underset{a}{\overset{
}{\int }}F(x(t),\dot{x}(t))dt,$ where the \textit{Finslerian norm}
F:A\rightarrow \mathbb{R}$ is defined as: $F=\sqrt{\left\vert L\right\vert }.
$
The correspondence $\left( x,y\right) \mapsto g_{(x,y)}$, wher
\begin{equation}
g_{(x,y)}=g_{ij}(x,y)dx^{i}\otimes dx^{j} \label{metric_tensor}
\end{equation
defines a mapping $g:A\rightarrow T_{2}^{0}M,$ (where $T_{2}^{0}M=T^{\ast
}M\otimes T^{\ast }M$), called the \textit{pseudo-Finslerian metric tensor}
attached to $L.$ A pseudo-Finsler metric $g$ can thus be regarded as a
section of the pullback bundle $\pi _{|A}^{\ast }(T_{2}^{0}M).$
In any local chart $(\pi ^{-1}(U),\Phi )$, there hold the equalities
\begin{equation}
L_{\cdot i}=2y_{i},~~\ y_{i\cdot j}=g_{ij}, \label{L_i}
\end{equation
where $y_{i}=g_{ij}y^{j}.$
On $A^{o}:=\{(x,y)\in A~|~L(x,y)\not=0\},$ it makes sense the \textit
angular metric
\begin{equation}
h=g-\dfrac{1}{4L}p\otimes p:A^{o}\rightarrow T_{2}^{0}M
\label{angular metric global}
\end{equation
where $p:=\dfrac{\partial L}{\partial y^{i}}dx^{i}.$ Using (\ref{L_i}), this
is written locally as
\begin{equation}
h=h_{ij}dx^{i}\otimes dx^{j},~\ \ \ \ \ h_{ij}=g_{ij}-\dfrac{y_{i}y_{j}}{L}.
\label{angular metric def}
\end{equation
The functions $h_{ij}$ and their contravariant versions
h^{ij}=g^{ik}g^{jl}h_{kl}$ obey:
\begin{equation}
h_{ij}y^{i}=0,~\ \ h^{ij}y_{i}=0. \label{angular metric prop}
\end{equation}
\bigskip
Geodesics of $(M,A,L)$ are described (e.g., \cite{Anto}, \cite{Bucataru}),
by the equations
\begin{equation}
\dfrac{d^{2}x^{i}}{dt^{2}}+2G^{i}(x,\dot{x})=0, \label{geodesic_eqn}
\end{equation
where the geodesic coefficient
\begin{equation}
2G^{i}(x,y)=\dfrac{1}{2}g^{ih}(L_{\cdot h,j}y^{j}-L_{,h})
\label{geodesic_spray}
\end{equation
are defined for $\left( x,y\right) \in A.$ The \textit{canonical nonlinear
connection\ }$N$ will be understood as a connection on the fibered manifold
A$, in the sense of \cite{Sardanashvili}, pp. 30-32, i.e., as a splittin
\begin{equation*}
TA=HA\oplus VA,
\end{equation*
where $VA=\ker d\pi _{|A}$ is called the vertical subbundle and $HA,$ the
horizontal subbundle of the tangent bundle $(TA,\pi _{|A},A)$. The local
adapted basis will be denoted by $(\delta _{i},\dot{\partial}_{i}),$ where
\delta _{i}:=\dfrac{\partial }{\partial x^{i}}-G_{~i}^{j}\dfrac{\partial }
\partial y^{j}},$ $\dot{\partial}_{i}=\dfrac{\partial }{\partial y^{j}}$ and
its dual basis, by $(dx^{i},\delta y=dy^{i}+G_{~j}^{i}dx^{j}),$ wher
\begin{equation}
G_{~j}^{i}=G_{~\cdot j}^{i}. \label{nonlinear_conn_coeffs}
\end{equation
Every vector field $X\in \mathcal{X}(M)$ can thus be uniquely decomposed as
X=hX+vX,$ where $hX:=X^{i}\delta _{i}\in \Gamma (HA)$ and $vX:=Y^{i}\dot
\partial}_{i}\in \Gamma (VA).$
By $^{h}:\Gamma (A)\rightarrow \Gamma (HA),$ $v=v^{i}\partial _{i}\mapsto
v^{h}:=v^{i}\delta _{i}$ and $^{v}:\Gamma (A)\rightarrow \Gamma (VA),$
v=v^{i}\partial _{i}\mapsto v^{v}:=v^{i}\dot{\partial}_{i},$ we will mean
the corresponding horizontal and vertical lifts of vector fields.
\bigskip
The \textit{dynamical covariant derivative}, \cite{Bucataru}, p. 34,
determined by the canonical nonlinear connection $N$ becomes, in a
pseudo-Finsler space $(M,A,L)$, a mapping $\nabla :\Gamma (VA)\rightarrow
\Gamma (VA),$ $X\mapsto \nabla X$ on the vertical subbundle $VA;$ it is
given in any local chart by
\begin{equation}
\nabla X_{(x,y)}:=(S(X^{j})+G_{~j}^{i}X^{j})_{(x,y)}\dot{\partial}_{i},~\ \
\forall (x,y)\in A, \label{dynamical covariant derivative}
\end{equation
where $X=X^{i}\dot{\partial}_{i}$ and $S:=y^{k}\delta _{k}.$ The operator
\nabla $ acts on functions $f:TM\rightarrow \mathbb{R}$ as: $\nabla f=S(f)$,
it is additive and obeys the Leibniz rule with respect to multiplication
with functions.
The complete lift $\xi ^{\mathbf{c}}=\xi ^{i}\partial _{i}+\xi _{,j}^{i}y^{j
\dot{\partial}_{i}$ of an admissible vector field $\xi \in \Gamma (A)$ can
be expressed in terms of $\nabla $ as
\begin{equation}
\xi ^{\mathbf{c}}=\xi ^{h}+\nabla (\xi ^{v}). \label{complete_lift_nabla}
\end{equation
From the 2-homogeneity in $y$ of the geodesic coefficients $2G^{i},$ it
follows that, along geodesics $c:[a,b]\rightarrow M,$ $t\mapsto (x^{i}(t))$
of $(M,A,L),$ we have, \cite{Bucataru}, p. 108: $\nabla \dot{x}^{i}=0;$
equivalently
\begin{equation}
(\nabla \dot{c}^{v})_{(c(t),\dot{c}(t))}=0. \label{geodesic_eqn_nabla}
\end{equation}
The canonical nonlinear connection $N$ is metrical, that is, for the
vertical lift $g^{v}=g_{ij}\delta y^{i}\otimes \delta y^{j}:\Gamma
(VA)\times \Gamma (VA)\rightarrow \mathbb{R}$ of the metric $g,$ there holds
(\cite{Bucataru}, p. 98), at any $(x,y)\in A:$
\begin{equation}
\nabla g^{v}=0, \label{metricity_nabla}
\end{equation
where $(\nabla g^{v})(X,Y)=\nabla (g^{v}(X,Y))-g^{v}(\nabla
X,Y)-g^{v}(X,\nabla Y),$ $\forall X,Y\in \Gamma (VA).$
Another known property which will be used in the following is that $L$ is
constant along horizontal curves, \cite{Crampin}, that is,
\begin{equation}
X(L)=0,~\ \ \forall X\in \Gamma (HA). \label{horizontal_derivs_L}
\end{equation}
\section{Basic notions and examples}
\subsection{\label{basic notions}Conformal maps and conformal vector fields}
The notion of conformal map between Finsler spaces is extended in a
straightforward way to pseudo-Finsler spaces; we have to just take care to
the domains of definition of the involved metric tensors.
\begin{definition}
A diffeomorphism $f:M\rightarrow M^{\prime }$ is called a \textit{conformal
map} between two pseudo-Finsler spaces $(M,A,L)$ and $\left( M^{\prime
},A^{\prime },L^{\prime }\right) $ if there exists a function $\sigma
:M\rightarrow \mathbb{R}$ such that
\begin{equation}
L^{\prime }\circ df_{|A}=e^{\sigma }L. \label{conformal_def_L}
\end{equation}
\end{definition}
In Finsler spacetimes, conformal maps preserve the light cones $L=0.$ For
positive definite Finsler spaces, transformations (\ref{conformal_def_L})
coincide with angle-preserving transformations, \cite{Anto}.
A conformal map is a \textit{similarity }if $\sigma =const.$ and an \textit
isometry }if $\sigma =1.$
\bigskip
Denoting by $\tilde{A}:=A\cap \left( df^{-1}\right) (A^{\prime })$ the set
where (\ref{conformal_def_L}) makes sense, (\ref{conformal_def_L}) reads
\begin{equation}
L^{\prime }(f(x),df_{x}(y))=e^{\sigma (x)}L(x,y),~\ \ \forall (x,y)\in
\tilde{A}. \label{conformal_explicit}
\end{equation}
\bigskip
\textbf{Convention.}\textit{\ }In the following, we will assume that $\pi
\tilde{A})=M$ (in particular, this implies that $\tilde{A}$ is a fibered
manifold over $A$). Under this assumption, there will be no loss of
generality if we consider that $A^{\prime }=\left( df\right) (A);$ in the
contrary case, we will restrict our discussion to the sets $\tilde{A}$ and
\left( df\right) (\tilde{A})=A^{\prime }\cap df(A)$ respectively and
re-denote them by $A$ and $A^{\prime }$.We will denote the restriction
df_{|A}:A\rightarrow A^{\prime }$ simply by $df.$
\bigskip
With the notatio
\begin{equation}
\tilde{L}:=L^{\prime }\circ df, \label{L_tilde_notation}
\end{equation
and with the above convention, (\ref{conformal_def_L}) becomes
\begin{equation}
\tilde{L}(x,y)=e^{\sigma }L(x,y),~~\forall (x,y)\in A; \label{L_tilde}
\end{equation
this is equivalent to
\begin{equation}
\tilde{g}(x,y)=e^{\sigma }g(x,y),~~\forall (x,y)\in A. \label{g_tilde_g}
\end{equation}
\bigskip
Assume that $f:M\rightarrow M^{\prime }$ is given with respect to two
arbitrary local charts on $M$ and $M^{\prime }$ as: $\tilde{x}^{i}=\tilde{x
^{i}(x^{j})$; the differential $df:A\rightarrow A^{\prime },\left(
x,y\right) \mapsto (\tilde{x},\tilde{y})$ is locally expressed as: $\tilde{x
^{i}=\tilde{x}^{i}(x^{j}),~\tilde{y}^{i}=\dfrac{\partial \tilde{x}^{i}}
\partial x^{j}}y^{j},$ therefore, differentiating (\ref{L_tilde_notation})
twice with respect to $y^{i},$ we find
\begin{equation}
\tilde{g}_{ij}(x,y)=\dfrac{\partial \tilde{x}^{k}}{\partial x^{i}}\dfrac
\partial \tilde{x}^{l}}{\partial x^{j}}g_{kl}^{\prime }(\tilde{x},\tilde{y
),~\ \ \forall (x,y)\in A. \label{g_tilde_local}
\end{equation
In coordinate-free writing, this is
\begin{equation}
\tilde{g}=T_{2}^{0}f\circ g^{\prime }\circ df, \label{g_tilde_global}
\end{equation
where $T_{2}^{0}f:T_{2}^{0}M^{\prime }\rightarrow T_{2}^{0}M$ is the mapping
naturally induced by $f$ on the respective tensor powers (giving the
multiplication by the Jacobian matrix of $f$ in (\ref{g_tilde_local})); we
will write this also as
\begin{equation}
\tilde{g}:=\left( df\right) ^{\ast }g^{\prime }. \label{pullback_notation}
\end{equation}
\bigskip
On a pseudo-Finsler space $(M,A,L)$, an admissible vector field $\xi \in
\Gamma (A)$ is called \textit{conformal} if its 1-parameter group $\left\{
\varphi _{\varepsilon }\right\} _{\varepsilon \in I}$ consists of conformal
transformations, i.e., for any $\varepsilon \in I:$
\begin{equation}
L\circ d\varphi _{\varepsilon }=e^{\sigma _{\varepsilon }}L,
\label{1-parameter group}
\end{equation
where $\sigma _{\varepsilon }:M\rightarrow M$ are smooth functions.
Assume that $\xi \in \Gamma (A)$ is a conformal vector field. Since
d\varphi _{\varepsilon }$ is generated by the complete lift $\xi ^{\mathbf{c
},$ we get, by differentiating (\ref{1-parameter group}) at $\varepsilon =0$
\begin{equation}
\mathcal{L}_{\xi ^{\mathbf{c}}}L=\dfrac{d}{d\varepsilon }|_{\varepsilon
=0}(e^{\sigma _{\varepsilon }}L)=\mu L, \label{Lie_deriv_L}
\end{equation
where $\mu :=\dfrac{d\sigma _{\varepsilon }}{d\varepsilon }|_{\varepsilon
=0}.$
In particular, if $\sigma _{\varepsilon }=1$ for all $\varepsilon ,$ i.e.,
\xi $ is a \textit{Killing vector field} for $L$, then:\ $\mathcal{L}_{\xi ^
\mathbf{c}}}L=0.$
\bigskip
\textbf{Examples. }If $L=L(y):T\mathbb{R}^{n}\rightarrow \mathbb{R}$ is
locally Minkowski, then:
1)\ The \textit{radial vector field} $\xi (x)=x^{i}\partial _{i}$ is a
conformal vector field. This can be checked easily, as $\xi ^{\mathbf{c
}=x^{i}\partial _{i}+x_{,j}^{i}y^{j}\dot{\partial}_{i}=x^{i}\partial
_{i}+y^{i}\dot{\partial}_{i}$ and, using the homogeneity of degree 2 of $L,$
we obtain:
\begin{equation*}
\mathcal{L}_{\xi ^{\mathbf{c}}}L=x^{i}L_{,i}+y^{i}L_{\cdot i}=0+2L=2L.
\end{equation*
The flow of $\xi $ consists of the \textit{dilations (homotheties)} $\varphi
_{\varepsilon }:\left( x^{i}\right) \mapsto (e^{\varepsilon }x^{i}).$
2) Any \textit{constant vector field} $\xi _{0}$ is a Killing vector field
for $L=L(y).$ This follows from: $\xi _{0}^{\mathbf{c}}=\xi _{0}^{i}\partial
_{i}$ and:
\begin{equation*}
\mathcal{L}_{\xi _{0}^{\mathbf{c}}}L=\xi _{0}^{i}L_{,i}=0.
\end{equation*
The vector field $\xi _{0}$ generates the translations $\left( x^{i}\right)
\mapsto (x^{i}+\varepsilon \xi _{0}^{i})$.
\subsection{\label{BM_example} Conformal maps between locally Minkowski
spaces}
In Euclidean spaces, Liouville's Theorem states that any conformal
transformation relating two domains of $\mathbb{R}^{n},$ $n>2,$ is a
similarity or the composition between a similarity and an inversion; passing
to pseudo-Euclidean spaces, one has to only add to the picture, \cit
{Haantjes}, compositions of two inversions.
In Finsler spaces, the situation is even more rigid; it was proven in \cit
{Matveev} that any conformal map between two non-Euclidean locally Minkowski
Finsler spaces is a similarity. Taking all these into account, one could
reasonably expect that conformal groups of pseudo-Finsler spaces could not
be too rich.
Still, as we will show in the following, one can create whole \textit
families} of pseudo-Finsler metrics with conformal symmetries which are not
only non-similarities, but they depend on arbitrary functions. This gives an
affirmative answer, in indefinite signature, to an old and famous question
raised by M. Matsumoto, \cite{Matveev}, namely \textit{whether there exist
two locally Minkowski structures which are conformal to each other.}
\bigskip
For $\dim M=4,$ a first example is actually known from \cite{Pavlov}. This
example can be extended to any dimension, as follows.
\textbf{Example 1: }\textit{Conformal symmetries of Berwald-Moor metrics.
Consider, on $M=\mathbb{R}^{n},$ $n>1:$
\begin{equation*}
A=\left\{ (x^{i},y^{i})_{i=\overline{0,n-1}}~|~y^{0}y^{1}....y^{n-1}\not=
\right\} \subset TM\backslash \{0\}
\end{equation*
and the $n$-dimensional Berwald-Moor pseudo-Finsler function (\cit
{Matsumoto}, pp. 155-156) on $A:
\begin{equation}
L(y)=\varepsilon \left\vert y^{0}y^{1}....y^{n-1}\right\vert ^{\tfrac{2}{n}},
\label{BM_Lagrangian}
\end{equation
where\footnote
The sign $\varepsilon $ is meant to ensure the existence of spacelike
vectors. Our $L$ is, up to a sign, the square of the one in \cite{Matsumoto}
(the latter would be, in our notations, $F$).} $\varepsilon
:=sign(y^{0}y^{1}....y^{n-1})$.
\bigskip
For an arbitrary diffeomorphism of the for
\begin{equation}
f:\mathbb{R}^{n}\rightarrow \mathbb{R}^{n},~\ \ x=\left(
x^{0},x^{1},...,x^{n-1}\right) \mapsto
(f^{0}(x^{0}),f^{1}(x^{1}),...,f^{n-1}(x^{n-1})),
\label{conformal transf BM}
\end{equation
the Jacobian determinant $J(x):=\dfrac{df^{0}}{dx^{0}}\dfrac{df^{1}}{dx^{2}
....\dfrac{df^{n-1}}{dx^{n-1}}$ is always nonzero, hence, there is no loss
of generality if we assume that $J(x)>0$, $\forall x\in \mathbb{R}^{n}.$ We
find:
\begin{equation}
\tilde{L}(y)=L(df(y))=J(x)^{\tfrac{2}{n}}L(y),~\ \forall y\in A,
\label{transformed _BM metric}
\end{equation
i.e., $\tilde{L}$ is also defined on $A.$ Moreover, $f$ is a conformal map,
with conformal factor
\begin{equation}
\sigma (x)=\dfrac{2}{n}\ln J(x). \label{conformal factor BM}
\end{equation
The Finsler function (\ref{transformed _BM metric}) is locally Minkowski;
more precisely, the coordinate transformation on $\pi ^{-1}(U)$ induced by:
(x^{i})=f^{-1}(x^{i^{\prime }})$ brings $\tilde{L}$ to the form $\tilde{L
(y)=(y^{0^{\prime }}y^{1^{\prime }}...y^{n-1^{\prime }})^{\tfrac{2}{n}}.$
Yet, $\sigma (x)$ is not only non-constant, but it depends on $n$ arbitrary
functions.
\bigskip
Berwald-Moor metrics are not the only such examples. Here is a much more
general class of flat pseudo-Finsler metrics on $\mathbb{R}^{n},$ $n\geq 2,$
which admit nontrivial conformal symmetries.
\textbf{Example 2: }\textit{Weighted product Finsler functions. }Consider $M
\mathbb{R}^{k}\times \mathbb{R}^{n-k}$ and a pseudo-Finsler metric function
L:A\rightarrow \mathbb{R}$ (with $A\subset A_{1}\times A_{2},$ $A_{1}\subset
T\mathbb{R}^{k},$ $A_{2}\subset T\mathbb{R}^{n-k}$), of the form
\begin{equation}
L=L_{1}^{\alpha }L_{2}^{1-\alpha }, \label{weighted direct product}
\end{equation
where $L_{1}:A_{1}\rightarrow \mathbb{R}$ and $L_{2}:A_{2}\rightarrow
\mathbb{R}$ are pseudo-Finsler functions and $\alpha \in (0,1).$
Assume that $f_{1}:\mathbb{R}^{k}\rightarrow \mathbb{R}^{k},$ $\left(
x^{0},...,x^{k-1}\right) \mapsto (\tilde{x}^{0},...,\tilde{x}^{k-1})$ is a
conformal transformation with non-constant factor $\sigma =\sigma (x),$ such
that $\tilde{L}_{1}=L_{1}\circ df$ is locally Minkowski - and let $L_{2}$ be
completely arbitrary. Then, the transformation
\begin{equation}
f:\mathbb{R}^{n}\rightarrow \mathbb{R}^{n},~f:=(f_{1},id_{\mathbb{R}^{n-k}})
\label{conformal_transf_product}
\end{equation
leads to: $\tilde{L}(y):=L(df(y))=e^{\alpha \sigma (x)}L(y),$ $\forall
y=(y^{0},...,y^{k-1},y^{k},..,y^{n-1})\in \mathbb{R}^{n},$ i.e., $f$ is a
conformal symmetry (which is not\textit{\ }a similarity) of $L.$ The
obtained Finsler function $\tilde{L}$ is obviously locally Minkowski - and
it depends on the choice of the function $f_{1}.$
\bigskip
\textbf{Particular cases} for the choice of $L_{1}$ in (\ref{weighted direct
product}) include:
a) \textit{The case} $k=1.$ In this case, $L_{1}=\lambda (y^{0})^{2},$ for
some $\lambda \in \mathbb{R}$ and therefore, \textit{any} diffeomorphis
\begin{equation*}
f_{1}:\mathbb{R\rightarrow R},x^{0}\mapsto f_{1}(x^{0}),
\end{equation*
serves the purpose, since: $L_{1}(df_{1}(y^{0}))=L_{1}(\dot{f
_{1}(x^{0})y^{0})=\dot{f}_{1}^{2}(x^{0})L_{1}(y^{0}).$
b) \textit{The }$k$\textit{-dimensional Minkowski metric:
\begin{equation*}
L_{1}(y^{0},...,y^{k-1})=\left( y^{0}\right) ^{2}-\left( y^{1}\right)
^{2}-....-(y^{n})^{2};
\end{equation*
then, $f_{1}$ can be, e.g., an inversion.
c) \textit{The }$k$\textit{-dimensional Berwald-Moor metric} can also be
chosen as $L_{1}$. In this case, $f_{1}$ can be any mapping of the form (\re
{conformal transf BM}).
\section{Behavior of geodesics under conformal maps}
\textbf{1. Projective-and-conformal mappings. }A diffeomorphism
f:M\rightarrow M$ between is called a \textit{projective map }if geodesics
of $L$ coincide, up to re-parametrization, with geodesics of $\tilde{L
:=L\circ df.$ In a completely similar manner to the positive definite case
\cite{Anto}, pp. 110-111), it follows that the mapping $f$ is projective if
and only if there exists a 1-homogeneous scalar function $P:A\rightarrow
\mathbb{R}$ such that, in any local chart,
\begin{equation}
2\tilde{G}^{i}\left( x,y\right) =2G^{i}\left( x,y\right) +P\left( x,y\right)
y^{i},~\ \forall \left( x,y\right) \in A. \label{projective_rel_G}
\end{equation}
Assume that the projective map $f$ is also conformal, with conformal factor
e^{\sigma }.$ Then, a direct calculation using (\ref{geodesic_spray}) shows
that
\begin{equation*}
2\tilde{G}^{i}=2G^{i}+\dfrac{1}{2}g^{ih}\left( \sigma _{,k}y^{k}L_{\cdot
h}-\sigma _{,h}L\right) ;
\end{equation*
using (\ref{L_i}), this is:
\begin{equation}
2\tilde{G}^{i}=2G^{i}+\sigma _{,k}y^{k}y^{i}-\dfrac{1}{2}g^{ih}\sigma _{,h}L.
\label{rel_spray_coefficients}
\end{equation}
\bigskip
Based on the properties of the angular metric tensor (\ref{angular metric
global}), we can extend to arbitrary signature a result known in positive
definite Finsler spaces from \cite{Cheng}, \cite{Szilasi}:
\begin{theorem}
If a mapping $f:M\rightarrow M,$ relating two pseudo-Finsler structures on a
manifold $M$ with $\dim M\geq 2$ is both conformal and projective, then $f$
is a similarity.
\end{theorem}
\begin{proof}
Denote by $(M,A,L)$ and $(M,A,\tilde{L})$ the two Finsler structures; that
is, $\tilde{L}=L\circ df$. As $f$ is both conformal and projective,
equalities (\ref{rel_spray_coefficients}) and (\ref{projective_rel_G}) are
both satisfied. Therefore, at any $\left( x,y\right) \in A$ and in any local
chart around $(x,y),
\begin{equation}
\sigma _{,k}y^{k}y^{i}-\dfrac{1}{2}g^{ik}\sigma _{,k}L=Py^{i}. \label{aux0}
\end{equation
Now, fix an arbitrary $x\in M$ and an arbitrary open region of $A_{x}$ where
$L\not=0$; on such a region, it makes sense the angular metric tensor (\re
{angular metric global}). Contracting (\ref{aux0}) with $h_{ij}$ and using
\ref{angular metric prop}), it remains: $h_{ij}g^{ik}\sigma _{,k}L=0.$
Taking into account that $h_{ij}g^{ik}=\delta _{j}^{k}-\dfrac{y^{k}y_{j}}{L},
$ this becomes
\begin{equation}
L\sigma _{,j}-\sigma _{,k}y^{k}y_{j}=0. \label{aux1}
\end{equation
Differentiating with respect to $y^{i},$ we find, by (\ref{L_i})
\begin{equation*}
2y_{i}\sigma _{,j}-\sigma _{,i}y_{j}-\sigma _{,k}y^{k}g_{ij}=0.
\end{equation*
Now, contract both hand sides \ of the above equality with $h^{ij}.$ Using
again (\ref{angular metric prop}), we get rid of the first and of the second
term. Further, noticing that $h^{ij}g_{ij}=n-1,$ we obtain: $\left(
n-1\right) \sigma _{,k}y^{k}=0.$ But, by hypothesis, $n=\dim M\geq 2$,
therefore
\begin{equation*}
\sigma _{,h}y^{h}=0,
\end{equation*
which, by differentiation with respect to $y^{k},$ gives that: $\sigma
_{,k}(x)=0.$ As the point $x$ was arbitrarily chosen, we obtain $\sigma
(x)=const.$, q.e.d.
\end{proof}
\bigskip
\textbf{Remark.}\ Substituting $\sigma =const.$ into (\ref{aux0}), we obtain
$P=0.$ That is, if two pseudo-Finsler metrics $L$ and $\tilde{L}$ are both
conformally and projectively related, then, $2\tilde{G}^{i}=2G^{i}$ -
meaning that their parametrized geodesics coincide.
\bigskip
\textbf{2. Conformal changes and null geodesics. }Generally, conformal maps
do not preserve geodesics. Still, for null geodesics, we can extend a
remarkable result from the semi-Riemannian case :
\begin{proposition}
Null geodesics of two conformally related pseudo-Finsler metrics coincide up
to parametrization.
\end{proposition}
\begin{proof}
Denote by $L$ and $\tilde{L}$ the two conformally related pseudo-Finsler
metrics. Taking into account that, along null geodesics, $L=0$ and
substituting into (\ref{rel_spray_coefficients}), we find that, along these
curves, $2\tilde{G}^{i}=2G^{i}+\sigma _{,k}y^{k}y^{i}.$ Setting $P:=\sigma
_{,k}y^{k}$, we get: $2\tilde{G}^{i}=2G^{i}+Py^{i},$ which means that null
geodesics of the two spaces coincide up to re-parametrization.
\end{proof}
\bigskip
Another result in pseudo-Riemannian geometry, \cite{Kuhnel survey}, which
can be extended to pseudo-Finsler spaces is:
\begin{proposition}
\label{conservation law}Let $\xi \in $ $\Gamma (A)$ be a conformal vector
field for a pseudo-Finsler space $(M,A,L).$ Along any lightlike geodesic
c:[a,b]\rightarrow M$, $t\mapsto c(t)$ the quantity $g_{(c(t),\dot{c}(t))}
\dot{c}(t),\xi (t))$ is conserved.
\end{proposition}
\begin{proof}
Take an arbitrary lightlike geodesic $c$ on $M$ and denote by $C=(c,\dot{c
), $ the lift of $c$ to $TM.$ Under the above made assumption that $c$ is
admissible, we can write $C:[a,b]\rightarrow A.$
Denote, for simplicity: $g:=g_{(c(t),\dot{c}(t))},$ $\nabla X:=(\nabla
X)_{(c(t),\dot{c}(t))}$ for $X\in \Gamma (VA)$ and $\nabla f:=\nabla
f_{(c(t),\dot{c}(t))}$ for smooth functions on $M.$ As $c$ is a geodesic, we
have, by (\ref{geodesic_eqn_nabla}), $\dot{C}=\dot{x}^{i}\delta _{i};$ hence
\begin{equation*}
\dfrac{df}{dt}=\dot{C}f=\dot{x}^{i}\delta _{i}f=\nabla f,~\ \ \ \ \forall
f:TM\rightarrow \mathbb{R}.
\end{equation*
Applying the above equality to
\begin{equation}
f:=g(\dot{c},\xi )=g^{v}(\dot{c}^{v},\xi ^{v}), \label{rel_g_to_gv}
\end{equation
we get: $\dfrac{d}{dt}(g(\dot{c},\xi ))=\nabla (g^{v}(\dot{c}^{v},\xi ^{v}))$
and therefore,
\begin{equation}
\dfrac{d}{dt}(g(\dot{c},\xi ))=\left( \nabla g^{v}\right) (\dot{c}^{v},\xi
^{v})+g^{v}((\nabla \dot{c}^{v},\xi ^{v})+g^{v}(\dot{c}^{v},\nabla \xi ^{v}).
\label{deriv_g_csi_c}
\end{equation
The first term in the right hand side is zero by (\ref{metricity_nabla}).
The second one is also zero since $c$ is a geodesic. It remains to evaluate
g(\dot{c}^{v},\nabla \xi ^{v}).$
Since $\xi $ is a conformal vector field, it obeys:\ $\mathcal{L}_{\xi ^
\mathbf{c}}}L=\mu L$ for some function $\mu .$ But, by hypothesis, $c$ is
lightlike, i.e., $L$ vanishes along $C.$ It follows
\begin{equation}
\mathcal{L}_{\xi ^{\mathbf{c}}}L=0~\text{\ on \ }C. \label{zero_Lie_deriv}
\end{equation
Further, using (\ref{complete_lift_nabla}) for the Lie derivative $\mathcal{
}_{\xi ^{\mathbf{c}}}L=$ $\xi ^{\mathbf{c}}(L),$ relation (\re
{zero_Lie_deriv}) becomes
\begin{equation*}
\xi ^{h}(L)+(\nabla \xi ^{v})(L)=0.
\end{equation*
The term $\xi ^{h}(L)$ vanishes by (\ref{horizontal_derivs_L}), which leads
to $(\nabla \xi ^{v})(L)=0.$ In coordinates, this is: $(\nabla \xi
^{i})L_{\cdot i}=0.$ Taking into account (\ref{L_i}), we can write it as:
2g_{ij}y^{j}\nabla \xi ^{i}=0.$ Along $C,$ this is equivalent to
\begin{equation*}
g^{v}(\dot{c}^{v},\nabla \xi ^{v})=0.
\end{equation*
Substituting the latter relation into (\ref{deriv_g_csi_c}), we get:\
\dfrac{d}{dt}g(\dot{c},\xi )=0,$ q.e.d.
\end{proof}
\section{Associated Riemannian metrics a useful lemma}
Consider a pseudo-Finsler space $(M,A,L)$, with metric tensor
g:A\rightarrow T_{2}^{0}M.$ For any admissible vector field $\xi \in \Gamma
(A),$ the mappin
\begin{equation}
g^{\xi }:=g\circ \xi :M\rightarrow T_{2}^{0}M \label{g^v}
\end{equation
defines a pseudo-Riemannian metric on $M,$ called an \textit{associated
(pseu\-do-)\-Rie\-man\-ni\-an metric }or, \cite{Stavrinos}, an \textit
osculating (pseudo)-Riemannian metric}.
\bigskip
Here are some immediate properties of metrics $g^{\xi },$ $\xi \in \Gamma
(A) $:
1. $g^{\xi }$ is defined and smooth on the whole base manifold $M$ (even if
g$ cannot be defined on the entire $TM\backslash \{0\}$).
2. $g^{\xi }$ has the same signature as $g.$
3. If, in particular, $g=g(x)$ is pseudo-Riemannian, then, all the metrics
g^{\xi },$ $\xi \in \Gamma (A)$ coincide (up to projection onto $M$) with
g, $ i.e., $g^{\xi }\circ \pi =g.$
4. If $g,\tilde{g}:A\rightarrow T_{2}^{0}M$ are conformally related, with
conformal factor $\sigma =\sigma (x),$ the
\begin{equation}
\tilde{g}^{\xi }=e^{\sigma }g^{\xi },~\ \forall \xi \in \Gamma (A).
\label{conformal_g_csi}
\end{equation}
\bigskip
Let us analyze conformal point transformations associated with (\re
{conformal_g_csi}). With this aim, consider an arbitrary pseudo-Finsler
metric tensor $g$ and - for the moment - an arbitrary diffeomorphism
f:M\rightarrow M.$ Assume, as above, that $df(A)=A,$ denote $df:=df_{|A}$
and conside
\begin{eqnarray}
\tilde{\xi} &:&=df\circ \xi \circ f^{-1}:M\rightarrow A, \label{csi_tilde}
\\
\tilde{g} &:&=(df)^{\ast }g=T_{2}^{0}f\circ g\circ df:A\rightarrow
T_{2}^{0}M,
\end{eqnarray
the corresponding deformations of $\xi $ and $g.$ Noticing that the pullback
$f^{\ast }(g^{\tilde{\xi}})$ of the pseudo-Riemannian metric $g^{\tilde{\xi}}
$ can be written as: $f^{\ast }(g^{\tilde{\xi}})=T_{2}^{0}f\circ g^{\tilde
\xi}}\circ f,$ we get
\begin{equation*}
f^{\ast }(g^{\tilde{\xi}})=T_{2}^{0}f\circ (g\circ \tilde{\xi})\circ
f=T_{2}^{0}f\circ g\circ df\circ \xi =\tilde{g}\circ \xi ,
\end{equation*
i.e., the left hand side of (\ref{conformal_g_csi}) is:
\begin{equation}
\tilde{g}^{\xi }=f^{\ast }(g^{\tilde{\xi}}). \label{f_star_g_v}
\end{equation
\bigskip In particular, if $f$ is a conformal map, then
\begin{equation}
f^{\ast }(g^{\tilde{\xi}})=e^{\sigma }g^{\xi }. \label{f_star_g_conformal}
\end{equation}
Using (\ref{f_star_g_conformal}), we obtain:
\begin{lemma}
\label{conformal_vector_fields_lemma}If $\xi :M\rightarrow A$ is a conformal
vector field for a pseudo-Finsler metric structure $(M,A,L)$, with
1-parameter group $\left\{ \varphi _{\varepsilon }\right\} ,$ then:
(i) $\xi $ it is a conformal vector field for the pseudo-Riemannian metric
g^{\xi }.$
(ii) The conformal factor relating the pseudo-Finsler metrics $g$ and
\tilde{g}=(d\varphi _{\varepsilon })^{\ast }g$ is the same as the conformal
factor relating $g^{\xi }$ and $\varphi _{\varepsilon }^{\ast }(g^{\xi }).$
\end{lemma}
\begin{proof}
\textit{(i) }Set:$\ \tilde{g}:=\left( d\varphi _{\varepsilon }\right) ^{\ast
}g,$ $\tilde{\xi}:=d\varphi _{\varepsilon }\circ \xi \circ \varphi
_{\varepsilon }^{-1}.$ By (\ref{f_star_g_conformal}), we have: $\varphi
_{\varepsilon }^{\ast }(g^{\tilde{\xi}})=e^{\sigma }g^{\xi }.$ But, the
vector field $\xi $ is invariant under its own flow, that is, $\tilde{\xi
=\xi $. We find:
\begin{equation}
\varphi _{\varepsilon }^{\ast }(g^{\xi })=e^{\sigma }g^{\xi },
\label{g_tilde_phi_epsilon}
\end{equation
that is, $\xi $ is a conformal vector field for $g^{\xi }.$
\textit{(ii)} The statement follows from (\ref{g_tilde_phi_epsilon}).
\end{proof}
\section{Conformal and Killing vector fields}
Here is another property in pseudo-Riemannian geometry, \cite{Kuhnel survey
, which can be extended to pseudo-Finsler spaces:
\begin{proposition}
If a conformal vector field $\xi :M\rightarrow A$ for a pseudo-Finsler space
$(M,A,L)$ is nowhere lightlike, then, $\xi $ is a Killing vector field for a
conformally related pseudo-Finsler structure.
\end{proposition}
\begin{proof}
As $\xi $ is a conformal vector field for $L,$ we have, at any $(x,y)\in A:
\begin{equation}
(\mathcal{L}_{\xi ^{\mathbf{c}}}L)\left( x,y\right) =\mu L(x,y).
\label{csi_y}
\end{equation
Using the hypothesis that $L$ is nowhere lightlike, the quantity $\alpha
(x):=L(x,\xi (x))$ does not vanish. Set
\begin{equation*}
\tilde{L}(x,y):=\dfrac{1}{\alpha (x)}L(x,y):A\rightarrow \mathbb{R}.
\end{equation*
Taking the Lie derivative of $\tilde{L}$: $(\mathcal{L}_{\xi ^{\mathbf{c}}
\tilde{L})=\mathcal{L}_{\xi ^{\mathbf{c}}}(\dfrac{1}{\alpha })L+\dfrac{1}
\alpha }\mathcal{L}_{\xi ^{\mathbf{c}}}(L)$ and noticing that $\mathcal{L
_{\xi ^{\mathbf{c}}}(a)=-\mu \alpha ,$ $\mathcal{L}_{\xi ^{\mathbf{c}}}L=\mu
L,$ we get:
\begin{equation*}
(\mathcal{L}_{\xi ^{\mathbf{c}}}\tilde{L})=-\dfrac{1}{\alpha ^{2}}\mu \alpha
L+\dfrac{1}{\alpha }\mu L(y)=0,
\end{equation*
i.e., $\xi $ is a Killing vector field for $\tilde{L}$.
\end{proof}
\textbf{Remark. }A conformal vector field for a pseudo-Finsler metric $L$ is
called \textit{essential }if it is not a Killing vector field for any
conformally related metric to $L.$ That is: any essential pseudo-Finslerian
conformal vector field must be lightlike at least at a point.
\bigskip
Passing to Killing vector fields, let us mention the following results due
to Sanchez, \cite{Sanchez}, in Riemannian geometry:
\begin{proposition}
\label{Sanchez1}, \cite{Sanchez}: Let $(M,g)$ be a Lorentzian manifold with
a non-spacelike (at any point)\ Killing vector field $\xi $. If $\xi _{p}=0$
for some $p\in M,$ then $\xi $ vanishes identically.
\end{proposition}
\begin{theorem}
\label{Sanchez 2}, \cite{Sanchez}: If $\xi $ is a Killing vector field on a
Lorentzian manifold $(M,g),$ admitting an isolated zero at some point $p\in
M,$ then, the dimension of $M$ is even and $\xi $ becomes timelike,
spacelike and null on each neighborhood of $p.$
\end{theorem}
Now, using Lemma \ref{conformal_vector_fields_lemma}, the extensions to
pseudo-Finsler spaces of the above results become simple corollaries:
\begin{proposition}
Let $(M,A,L)$ be a Finsler spacetime, with a non-spacelike (at any point)
Killing vector field $\xi $. If $\xi =0$ at some point $p\in M,$ then $\xi $
vanishes identically.
\end{proposition}
\begin{proof}
Since $\xi $ is a Killing vector field for $L,$ it follows from Lemma \re
{conformal_vector_fields_lemma} that $\xi $ is a Killing vector field for
the pseudo-Riemannian metric $g^{\xi }.$ But, since the signature of $g^{\xi
}$ coincides with the one of $L,$ $g^{\xi }$ is Lorentzian. The statement
now follows from Proposition \ref{Sanchez1}.
\end{proof}
\begin{theorem}
If $\xi $ is a Killing vector field for a Finsler spacetime $(M,A,L),$
admitting an isolated zero at some point $p\in M,$ then, the dimension of $M$
is even and $\xi $ becomes timelike, spacelike and null on each neighborhood
of $p.$
\end{theorem}
\begin{proof}
Assume $\xi $ is a Killing vector field for $(M,A,L),$ with an isolated zero
at some $p\in M.$ Then, $\xi $ is also a Killing vector for the Lorentzian
metric $g^{\xi }$ on $M$ and
\begin{equation*}
L(\xi )=g^{\xi }(\xi ,\xi ),
\end{equation*
which means that $\xi $ is timelike (respectively, null, spacelike)\ for $L$
iff it is timelike (resp., null, spacelike)\ for $g^{\xi }.$ The result now
follows from Theorem \ref{Sanchez 2}.
\end{proof}
\bigskip
\textbf{Acknowledgment. }The work was supported by a local grant of the
\textit{Transilvania }University of Brasov.
| 260121f0d870436fd5171f1d2f08852304474fa6 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
A common technique to reduce the impact of systematic uncertainties, in particular in precision measurements in high energy physics, is to constrain the range of their variations by the data. A simultaneous fit of these variations and the parameters to be measured can be performed based on prior knowledge of the uncertainties and suitable distributions to constrain them. This technique can reduce the total uncertainty in many cases significantly, as in Refs.~\cite{Khachatryan:2137231,Sirunyan:2017uhy,Khachatryan:2016yzq}, and can be used to measure several parameters simultaneously (see e.g. Refs.~\cite{Kieseler:2015jzh,Aad:2014jra}).
However, it leads to non-negligible correlations between all fitted parameters. These are particularly important when at least one measurement that uses this technique contributes to a combination.
The most consistent approach to such combination would be to define a \textit{combined likelihood} based on the original models, including all systematic variations, and the original data the models were fit to.
Also other commonly used combination techniques and the corresponding software tools would require this information~\cite{LYONS1988110,Nisius1,Nisius2,Alekhin:2014irh}, which is publicly available only in very rare cases and often unrecoverable. This poses a serious problem for a consistent combination involving measurements obtained with simultaneous nuisance parameter fits.
The method described here provides a solution for this problem, since it is based on the central results and their covariance or Hessians, only.
It allows separating constraints and correlations imposed by the previously fitted data from those that stem from prior knowledge of the systematic variations.
Therefore, the combination can be performed accounting for correlations between the measurements as well as for correlations and constraints within each individual measurement.
The dedicated software tool ``Convino'' is also presented in this article. It is specifically developed to perform combinations based on the method described here and provides a simple text-based user interface that can be used without knowledge of any programming language.
Assumptions on correlations can be varied in an automated way.
Moreover, partially correlated measurements of different quantities (e.g. bins of a differential distribution) can be combined simultaneously accounting for all correlations.
In addition to the text-based interface, a C\kern-0.1em\raisebox{.2ex}+\kern-0.1em\raisebox{.2ex}+{} interface is provided to define the input to the combination. This interface can either read basic C\kern-0.1em\raisebox{.2ex}+\kern-0.1em\raisebox{.2ex}+{} standard library data types or ROOT~\cite{Brun:1997pa} histogram and graph classes, which are commonly used in high-energy-physics analyses.
The combination method is described in Section~\ref{sec:math}. It is validated against the combined likelihood approach in Section~\ref{sec:valid}. The effect of neglecting correlations within the same measurement is studied in Section~\ref{sec:negcorr}.
The installation and the user interface of the Convino program is described in Section~\ref{sec:install}.
\section{Combination method}
\label{sec:math}
The combination is performed using a $\chi^2$ minimisation. The $\chi^2$ is defined as
\begin{equation}
\label{eq:fullchi}
\chi^2 = \sum_{\alpha} \left(\chi^2_{s,\alpha} + \chi^2_{u,\alpha} \right) + \chi^2_p
\end{equation}
It is composed of three terms: the term $\chi^2_{s,\alpha}$ represents the results of each measurement $\alpha$ and its statistical uncertainties. It follows a Neyman or Pearson $\chi^2$ definition, with the statistical uncertainty being fixed for each measurement or being scaled with the combined value, respectively. A measurement can aim to determine a set of quantities, e.g. bins of a differential cross section, where the Pearson definition is more applicable. Alternatively a single quantity can be measured, e.g. the mass of a particle from a fit of an invariant mass peak position, where the Neyman definition is presumably better suited to describe the measurement.
In both cases, the measured quantities are referred to as estimates in the following.
The additional term $\chi^2_{u,\alpha}$ describes the correlations between the systematic uncertainties and constraints on them from the data for each measurement $\alpha$. The last term, $\chi^2_p$, incorporates prior knowledge of the systematic uncertainties and correlation assumptions between uncertainties of the measurements to be combined.
\
In a typical measurement that exploits simultaneous constraints on the uncertainties from the data, the sources of uncertainties are uncorrelated prior to the fit to the data and the knowledge about their variations is modelled by independent penalty terms in the original likelihood.
The goal of the {me-thod} described here is to find an approximation for this likelihood that allows disentangling the independent penalty terms from the constraints and correlations between them, which are typically introduced by the data\footnote{The same procedure can be applied if correlations are present prior to the fit to the data}.
Therefore, the first central assumption is that the original likelihood for a measurement $\alpha$ can be approximated by:
\begin{equation}
\label{eq:evchi2}
\chi^2_\alpha = (\chi^2_{s,\alpha} + \chi^2_{u,\alpha}) + \sum_i (P_i^\alpha(\lambda_i) )^2 \text{,}
\end{equation}
where $(P_i^\alpha(\lambda_i) )^2$ represents the penalty term for a systematic uncertainty $i$ parametrised by a continuous parameter $\lambda_i$, such that $\lambda_i=1$ corresponds to a $1\sigma$ variation.
The terms $\chi^2_{s,\alpha}$ and $\chi^2_{u,\alpha}$ are defined as:
\begin{eqnarray}
\chi^2_{s,\alpha} & = & \sum_{\mu\nu} {M}_{\mu\nu}^\alpha \frac{\xi_{\mu}^\alpha \xi_{\nu}^\alpha} { \tau_{\mu}^\alpha \tau_{\nu}^\alpha} \text{ and} \\
\chi^2_{u,\alpha} & = & \sum_{ij} \lambda_i D_{ij}^\alpha \lambda_j \text{, with}
\end{eqnarray}
\begin{equation}
\xi_{\mu}^\alpha = x_{\mu}^\alpha - \bar{X}_\mu \text{ and}
\end{equation}
\begin{equation}
\bar{X}_\mu = \bar{x}_\mu \prod_i (\lambda_i K_{\mu i}^\alpha /x_{\mu}^\alpha +1 ) + \sum_i \lambda_i k_{\mu i}^\alpha
\text{.}
\end{equation}
Here, $x_{\mu}^\alpha$ is the estimate $\mu$ obtained in measurement $\alpha$ and $\bar{x}_\mu$ the combined value to be determined. The relation between both is given by $\tau_{\mu}^\alpha = ({\bar{X}_\mu/x_{\mu}^\alpha})^{0.5}$ for the Pearson $\chi^2$ definition. In case of the Neyman $\chi^2$, all $\tau_{\mu}^\alpha = 1$. The matrix $M$ represents the inverted statistical covariance of the estimates.
The parameters $k_{\mu i}^\alpha$ and $K_{\mu i}^\alpha$ model the effect of the systematic variations on the estimates for absolute or relative uncertainties, respectively. Here, a relative uncertainty is an uncertainty that scales with the measured value such as e.g. the luminosity uncertainty in a cross-section measurement, while absolute uncertainties have a constant value irrespective of the central result.
In principle, also other classes of dependencies can be incorporated through additional terms in $\xi^\alpha_\mu$.
The correlations between the uncertainties and the constraints that stem from the fit to the data are described by the matrix $D^\alpha$. Throughout this paper, indices $\mu$ and $\nu$ are used for estimates, while systematic uncertainties are denoted with indices $i$ and $j$.
The procedure to obtain the parameters of $\chi^2_\alpha$ from the measurements to be combined are discussed in the following - firstly for results obtained through a simultaneous nuisance parameter fit and secondly for the specific case of orthogonal uncertainties.
\subsection{Measurements obtained with simultaneous fits}
The second central assumption of the method is that the parameters of $\chi^2_\alpha$ can be determined from the Hessian of measurement $\alpha$ evaluated at the best-fit values, $\tilde{H}^\alpha_\text{in}$.
The entries of the Hessian can be ordered such that the matrix can be split in the following sub-matrices:
\begin{equation}
\tilde{H}^\alpha_\text{in} = \begin{pmatrix}
\tilde{D} & {\kappa}^T \\
{\kappa} & \tilde{M}
\end{pmatrix}^\alpha \text{,}
\end{equation}
where $\tilde{M}$ describes the relation between the estimates $x_\mu^\alpha$ and $x_\nu^\alpha$. The relation between systematic variations and the estimates is described by $\kappa$. The matrix $\tilde{D}$ quantifies the relation between the systematic variations.
All parameters of $ \chi^2_\alpha$ are determined by calculating analytically the Hessian of $\chi^2_\alpha$, ${H}^\alpha (\vec{0})$, and identifying the resulting terms with their counterparts of the input $\tilde{H}^\alpha_\text{in}$. Here $\vec{0}$ means $\lambda_i =0$ and $x_\mu^\alpha - \bar{x}_\mu = 0\ \forall\ i,\ \mu$.
The components are calculated as follows:
\begin{equation}
{H}^\alpha_{\mu\nu} (\vec{0}) = \frac{1}{2} \left(\frac{\partial^2 }{ \partial \Delta x_\mu^\alpha \partial \Delta x_\nu^\alpha} \chi^2_\alpha \right)\biggr\rvert_{\vec{0}}
= {M}_{\mu\nu} \text{,}
\end{equation}
\begin{equation}
{H}^\alpha_{\mu i} (\vec{0}) = \frac{1}{2} \left( \frac{\partial^2 }{ \partial \Delta x_\mu^\alpha \partial \lambda_i} \chi^2_\alpha \right)\biggr\rvert_{\vec{0}}
= \sum_\nu {M}_{\mu\nu} ( - \hat{k}_{\nu i}^\alpha) \text{,}
\end{equation}
with $\hat{k}_{\nu i}^\alpha = K_{\nu i}^\alpha + k_{\nu i}^\alpha$.
The matrix ${M}$ can be directly identified with $\tilde{M}$.
Since $\tilde{M}$ stems from a measurement of a physics quantity, $\tilde{M}$ is positive definite and therefore invertible. Thus, the parameters $\hat{k}_{\nu i}^\alpha$ can be determined as:
\begin{equation}
\hat{k}_{\nu i}^\alpha = - \sum_\mu ((M^\alpha)^{-1})_{\mu\nu}\ \kappa_{\mu i}^\alpha \text{.}
\end{equation}
Since a variation $i$ is either relative or absolute, $\hat{k}_{\nu i}^\alpha$ equals either $K_{\nu i}^\alpha$ or $k_{\nu i}^\alpha$, with the other parameter being 0.
The terms describing the analytic relations between the systematic uncertainties are calculated as:
\begin{equation}
\begin{aligned}
{H}^\alpha_{i j} (\vec{0}) = & \frac{1}{2} \left( \frac{\partial^2 }{ \partial \lambda_i \partial \lambda_j} \chi^2_\alpha \right)\biggr\rvert_{\vec{0}} \\
= &
D_{ij}^\alpha + \delta_{ij} \frac{1}{2} \frac{\partial^2 }{ \partial \lambda_i ^2}(P^\alpha_i)^2\biggr\rvert_{\lambda_i =0} + \sum_{\mu \nu} M_{\mu \nu}^\alpha \hat k_{\nu i}^\alpha\hat k_{\mu j}^\alpha \text{.} \\
\end{aligned}
\end{equation}
Only Gaussian penalty terms describing the prior knowledge of the uncertainties
are considered\footnote{In principle also other penalty terms (e.g. log-normal priors) can be accounted for. However, these lead to a non-constant $D^\alpha$.} ($P_i^\alpha(\lambda_i) = \lambda_i$), which simplify the equations, since in this case
\begin{equation}
\frac{1}{2}\frac{\partial^2 }{ \partial \lambda_i ^2}(P^\alpha_i)^2 \biggr\rvert_{\lambda_i =0} = 1 \text{.}
\end{equation}
In consequence $D^\alpha$ becomes:
\begin{equation}
D_{ij}^\alpha = \tilde{D}_{ij}^\alpha - \delta_{ij} - \sum_{\mu \nu} M_{\mu \nu}^\alpha \hat k_{\nu i}^\alpha\hat k_{\mu j}^\alpha \text{,}
\end{equation}
such that all parameters of Eq.~\ref{eq:evchi2} are defined.
\subsection{Measurements with orthogonal uncertainties}
\label{sec:orthunc}
For orthogonal uncertainties, the same initial $\chi^2$ described in Eq.~\ref{eq:evchi2} is used. However, the calculation of its parameters does not necessarily require the full Hessian. Instead, the calculation of the parameters simplifies to:
\begin{equation}
\hat{k}_{\mu i}^\alpha = \frac{\sigma_{\mu i}^\alpha}{\sigma_{\mu \text{ total}}^\alpha} \text{,}
\end{equation}
with $\sigma_{\mu i}^\alpha$ being the contribution of uncertainty $i$ to the total uncertainty $\sigma_{\mu \text{ total}}^\alpha$ of estimate $x_\mu^\alpha$. The matrix $D^\alpha$ is 0, the terms of $M$ are calculated as:
\begin{equation}
M_{\mu \nu} = \frac{ \rho_{\mu\nu} }{\sigma_\mu \sigma_\nu} \text{.}
\end{equation}
Here, $\rho_{\mu\nu}$ is the statistical correlation between estimate $\mu$ and $\nu$, and $\sigma_\mu$ and $\sigma_\nu$ are the corresponding statistical uncertainties.
\
For the orthogonal uncertainties as well as for measurements obtained by simultaneous fits, the constraints from the prior knowledge of the uncertainties are implemented in Eq.~\ref{eq:fullchi} through the term:
\begin{equation}
\label{eq:globalConstr}
\chi^2_{p} = \sum_{ij} P_i(\lambda_i) (C^{-1})_{ij} P_i(\lambda_j) \text{,}
\end{equation}
with $C$ being the matrix describing the correlation assumptions between the systematic uncertainties. In case no correlations are assumed, the term simplifies to:
\begin{equation}
\chi^2_{p}(\text{no corr}) = \sum_{i} P^2_i(\lambda_i) \text{.}
\end{equation}
Only Gaussian penalty terms are considered in the following, such that $P_i(\lambda_i) = \lambda_i$.
For a combination, $C$ will be of the structure
\begin{equation}
C = \begin{pmatrix}
\mathds{1} & A & \cdots \\
A & \mathds{1} & \cdots \\
\vdots & \vdots & \ddots \\
\end{pmatrix}\text{,}
\end{equation}
with matrices $A$ describing the correlation assumptions, and $\mathds{1} $ being the identity matrix.
\subsection{Technical implementation}
\label{sec:tech}
The final minimisation of Eq.~\ref{eq:fullchi} is performed using the Minuit algorithms~\cite{James:1975dr}. The total uncertainty on each combined value is determined by scanning $\chi^2 = \chi^2_\text{min} +1$ using the Minos algorithm.
These algorithms as implemented in ROOT~6 as ``TMinuit2'' are employed.
The correlations that are assumed between systematic uncertainties can vary between -1 and 1. These extremes are special cases for which the correlation matrix $C$ becomes non-invertible. In practice, a correlation of $C_{ij}=\pm1$ means that parameters $i$ and $j$ describe the same variation.
In such cases, an entry $C_{ij}=\pm1$ is replaced by $C_{ij}=\pm(1-10^{-3})$.
The difference to $\pm1$ is almost negligible.
For illustration, $2\times 2$ parameters are chosen with $C_{ij}\approx \pm1$. The affected part of the $\chi^2$, $\chi^2_{F}$, can be simplified to
\begin{eqnarray}
\chi^2_{F} & = & \frac{1}{1-C_{ij}^2} (\lambda_i^2 + \lambda_j^2 \mp 2 C_{ij} \lambda_i \lambda_j) \\
& \approx & \frac{1}{1-C_{ij}^2} ( \lambda_i \mp \lambda_j)^2
\end{eqnarray}
and corresponds to $( \lambda_i \mp \lambda_j)^2 \cdot 10^6$ for $C_{ij} = \pm (1- 10^{-3})$. Given that a variation of $\lambda = \pm 1$ corresponds to only a fraction of the total uncertainty on each estimate, the effect of the approximation $C_{ij} = \pm (1- 10^{-3})$ is negligible.
\section{Validation}
\label{sec:valid}
The validation is based on \textit{{pseu\-do-\-mea\-su\-re\-me\-nt}s}. Each {pseu\-do-\-mea\-su\-re\-me\-nt} is a binned likelihood fit with steerable central value and bin-wise uncertainties.
This has the advantage, that the full likelihood of each {pseu\-do-\-mea\-su\-re\-me\-nt} is known and can be adjusted to different scenarios. Therefore, it is possible to compare the results obtained with the method proposed here to the ones obtained using the combined likelihood as reference. Since the latter in principle contains arbitrarily more parameters, small deviations are expected.
The validation is first performed with respect to the statistical bias, only. Secondly, the modelling of systematic uncertainties is tested with respect to correlations between the uncertainties of the {pseu\-do-\-mea\-su\-re\-me\-nt}s, and the modelling of relative uncertainties.
For each {pseu\-do-\-mea\-su\-re\-me\-nt} a Poisson likelihood is chosen to determine the central result $\bar{x}_\mu$. The bin-wise uncertainties are randomly generated and modelled by the parameters $\lambda_i$.
In the case that an uncertainty corresponds to an absolute variation, its effect on each bin is generated independently.
For more than one bin ($N_\text{bins}>1$) this results in correlations between the uncertainties after the fit, as well as in constraints on their variations.
The likelihood for {pseu\-do-\-mea\-su\-re\-me\-nt} $\alpha$ is defined as:
\begin{equation}
\label{eq:pseudoLH}
L^\alpha = \prod_\mu \prod_i^{N_\text{bins}^\alpha} \mathcal{P}\left( X_\mu^\alpha, \bar{X}_{\mu i}^\alpha \right)
\cdot \prod_i \tilde{P}^\alpha_i(\lambda_i) \text{,}
\end{equation}
with $\mathcal{P}$ being the Poisson likelihood and $\tilde{P}^\alpha_i(\lambda_i)$ the Gaussian penalty terms modelling the prior knowledge of each uncertainty. The parameters $X_\mu^\alpha$ and $\bar{X}^\alpha_{\mu i}$ are given as:
\begin{eqnarray}
X_\mu^\alpha & = & \frac{x_\mu^\alpha}{N_\text{bins}^\alpha} \text{ and} \\
\bar{X}_{\mu i}^\alpha & = & \frac{\bar{x}_\nu}{N_\text{bins}^\alpha}\prod_j \left( \frac{K_{\nu j}^\alpha \lambda_j}{x_\nu}^\alpha + 1 \right) + \sum_j k_{\nu ij}^\alpha \lambda_j \text{,}
\end{eqnarray}
where $K_{\nu j}^\alpha$ describes the magnitude of global relative variations and $k_{\nu ij}^\alpha$ absolute shape variations, different for each bin $i$. The value of $x_\mu^\alpha$ is the input to each {pseu\-do-\-mea\-su\-re\-me\-nt} and corresponds to the number of events that would be observed in a real measurement. The elements of the matrices $K^\alpha$ and $k_i^\alpha$ are chosen to describe different validation scenarios. Finally, the fit to determine $\bar{x}_\mu^\alpha$ is performed and the resulting Hessian is recorded.
The combined likelihood for several {pseu\-do-\-mea\-su\-re\-me\-nt}s is given by:
\begin{equation}
L_\text{comb} = \left( \prod_\alpha \frac{L^\alpha}{\prod_i \tilde{P}_i^\alpha} \right) \cdot \phi (\lambda_0, ..., \lambda_N) \text{,}
\end{equation}
where $\phi$ models the prior knowledge of $N$ systematic uncertainties and the correlation assumptions between them, analogue to $\chi^2_p$ in Eq.~\ref{eq:globalConstr}. For every validation step, the difference $\Delta \bar{x}$ between the result obtained with the method proposed in this document and using the combined likelihood is recorded.
The compatibility of both approaches is is quantified by $\Delta \bar{x}/ \sigma_{\bar{x}}$, which is the difference between their central results normalised to the total uncertainty of the combined-likelihood combination.
\subsection{Statistical bias}
\label{sec:statbias}
To evaluate the statistical bias, the impact of systematic uncertainties on each {pseu\-do-\-mea\-su\-re\-me\-nt} is set to 0, corresponding to $K=0$ and $k=0$. Only one quantity, $\bar{x}$, is determined from two estimates $x^a$ and $x^b$, chosen as:
\begin{eqnarray}
x^a & = & s \cdot 100 \text{,} \\
x^b & = & x^a + \gamma \sqrt{x^a} \text{,}
\end{eqnarray}
with $s$ being a scaling factor and $\gamma$ describing the compatibility between the estimates. The latter is chosen to be either $\gamma=10$ to describe two very incompatible measurements or $\gamma=3$ for a more realistic scenario where both estimates still agree well enough to enter a combination.
Two {pseu\-do-\-mea\-su\-re\-me\-nt}s are generated for each choice of $s$ and $\gamma$ and are combined either using a Pearson or Neyman $\chi^2$ definition.
For both choices, the uncertainties on the combined results agree very well with the ones obtained using the direct combination based on $L_\text{comb}$. The bias of the central value is shown in Figure~\ref{fig:chibias} relative to the uncertainty of the combined value. It behaves as expected: it is smaller but of opposite sign for the Pearson $\chi^2$ definition and is reduced with smaller statistical uncertainties and better compatibility between the results.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.4\textwidth]{Convino_p_ney.pdf}
\caption{Difference between the combined values using a direct Poisson-likelihood combination and the method proposed here with Neyman and Pearson $\chi^2$ definition relative to the total uncertainty. The estimates to be combined differ by about $\gamma\sigma$ and are displayed as a function of the first estimate's relative statistical uncertainty. The second estimate's statistical uncertainty scales accordingly. \label{fig:chibias}}
\end{center}
\end{figure}
\subsection{Systematic uncertainties}
\label{sec:sysunc}
The effect of absolute systematic uncertainties is evaluated by combining two {pseu\-do-\-mea\-su\-re\-me\-nt}s, with randomly chosen elements of the matrices $k_\nu^\alpha$. An upper threshold $t$ is defined, such that for each element $i,\ j$:
\begin{equation}
| k_{\nu i j}^\alpha | \leq t \cdot X_{\mu}^\alpha \text{,}
\end{equation}
limiting the contribution of systematic uncertainties.
Two bins, two systematic uncertainties, and one $x_\mu^\alpha$ per {pseu\-do-\-mea\-su\-re\-me\-nt} are considered. The sign of $k_{\nu i j}^\alpha$ is chosen to be constant for each systematic uncertainty.
The estimates $x^a$ and $x^b$ for measurement $a$ and $b$ are set to:
\begin{eqnarray}
x^a & = & 30000 \text{ and}\\
x^b & = & 30600
\end{eqnarray}
to reduce the effect of statistical uncertainties. The resulting statistical uncertainty of 0.6\% does not account for the difference of 2\% between both values, such that the modelling of the systematic uncertainties will affect the combination significantly.
For large systematic variations the maximisation of Eq.~\ref{eq:pseudoLH} with Minuit can become numerically unstable. This is the case when the variation becomes as large as the nominal entry, $X_\mu^\alpha$, in at least one of the bins. Therefore, the Poisson likelihood is approximated with a Gaussian form, which is valid for low statistical uncertainties such as in this test. Thus, $L^\alpha$ becomes:\\
\begin{equation}
\label{eq:npseudoLH}
L^\alpha = \prod_{\mu\nu} \prod_i^{N_\text{bins}^\alpha}
\exp\left[
- S^\alpha_{\mu\nu} \frac{(\bar{X}_{\mu i}^\alpha - X_\mu^\alpha)(\bar{X}_{\nu i}^\alpha - X_\nu^\alpha)}{2 (\bar{X}_{\mu i}\bar{X}_{\nu i})^{1/2} }
\right] \text{.}
\end{equation}
The matrix $S^\alpha$ allows modelling direct statistical correlations between $\bar{X}_{\mu i}^\alpha$ and $\bar{X}_{\nu i}^\alpha$. Here, $S$ is set to $\mathds{1}$.
In total $2 \times 20,000$ {pseu\-do-\-mea\-su\-re\-me\-nt}s are generated, each with a different random choice of the uncertainties.
The total relative uncertainty, $\sigma_x/x$, on the estimate of {pseu\-do-\-mea\-su\-re\-me\-nt} $a$ is shown in Figure~\ref{fig:errb} for different values of the threshold $t$. Depending on $t$, the uncertainty varies from moderate values to more than 100\%. The same applies to {pseu\-do-\-mea\-su\-re\-me\-nt} $b$ (not displayed). The average constraints on the systematic uncertainties reach from about 90\% ($t=0.01$) to 50\% ($t=1.0$) with respect to their initial $1\sigma$ variation.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.4\textwidth]{sys_contr.pdf}
\caption{Relative total uncertainty of the {pseu\-do-\-mea\-su\-re\-me\-nt} $a$ for different values of the threshold $t$. \label{fig:errb}}
\end{center}
\end{figure}
In a first validation step, each uncertainty of one {pseu\-do-\-mea\-su\-re\-me\-nt} is assumed to be highly correlated with exactly one uncertainty of the other {pseu\-do-\-mea\-su\-re\-me\-nt} by assigning a correlation factor $c=0.99$. A total of 20,000 combinations are performed.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.4\textwidth]{err_dev.pdf}
\caption{Ratio of the uncertainties on the combined value obtained using the method proposed here and the direct likelihood combination, shown for different values of the upper threshold for systematic uncertainties $t$. \label{fig:errb2}}
\end{center}
\end{figure}
The ratio $\Delta^r \sigma$ between the uncertainty on the combined value obtained with the method proposed here and by maximising $L_\text{comb}$ is shown in Figure~\ref{fig:errb2} as a function of $t$. Asymmetric uncertainties on the combined value are accounted for and are equally well described.
With an increasing contribution of the systematic uncertainties, the $\Delta^r \sigma$-distribution becomes slightly broader, but does not indicate any numerically relevant mismodelling.
The resulting values for $\Delta \bar{x}/ \sigma_{\bar{x}}$ are illustrated in Figure~\ref{fig:deltaxbarrel}.
For $t=0.01$, the statistical uncertainties are non negligible. This leads to a small bias towards lower values introduced by the choice of Eq.~\ref{eq:npseudoLH} and discussed in the Section~\ref{sec:statbias}. However, for all choices of $t$ and all pseudo experiments, the differences between the direct likelihood approach and the method described here are well below 5\% of the total uncertainty and can therefore be considered negligible.
%
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.4\textwidth]{reldev.pdf}
\caption{
Difference between the combined result $\bar{x}$ obtained with the method proposed here and using a direct likelihood combination relative to the total uncertainty on $\bar{x}$, shown for different values of the upper threshold for systematic uncertainties $t$.\label{fig:deltaxbarrel}}
\end{center}
\end{figure}
Moreover, the dependence on the assumed correlation between the uncertainties of both {pseu\-do-\-mea\-su\-re\-me\-nt}s is studied, as well as possible biases with respect to the number of bins in each {pseu\-do-\-mea\-su\-re\-me\-nt}.
Figure~\ref{fig:syscorr} shows the dependence of $\Delta \bar{x}/ \sigma_{\bar{x}}$ on the choice for the correlation coefficients $c$ for $t=1$.
The modelling worsens slightly when $|c|$ decreases, but is below about 3\% with respect to the total uncertainty on the combined value for all 20,000 pseudo experiments.
\begin{figure}[h!]
\begin{center}
\includegraphics[width=0.4\textwidth]{reldev_syscorr.pdf}
\caption{
Difference between the combined results $\bar{x}$ from the method proposed here and using a direct likelihood combination relative to the total uncertainty on $\bar{x}$. The distribution is shown for different values of the correlation $c$ between the systematic uncertainties of both {pseu\-do-\-mea\-su\-re\-me\-nt}s.
\label{fig:syscorr}}
\end{center}
\end{figure}
Also, the total uncertainty remains well modeled with only a very moderate increase of combinations with $|\Delta^r \sigma |$ slightly different from 1, as shown in Figure~\ref{fig:syscorr2}. The same conclusion can be drawn when the procedure described here is repeated for a different number of bins in each {pseu\-do-\-mea\-su\-re\-me\-nt} (not shown here). All results for 2, 4, 20, and 100 bins show a good modelling of the combined likelihood approach with respect to the central values and the total uncertainties.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.4\textwidth]{errdev_syscorr.pdf}
\caption{
Ratio of the uncertainties on the combined value obtained using the method proposed here and the direct likelihood combination. The distribution is shown for different values of the correlation $c$ between the systematic uncertainties of both {pseu\-do-\-mea\-su\-re\-me\-nt}s.
\label{fig:syscorr2}}
\end{center}
\end{figure}
In general, the central result and its uncertainty are very well modelled for a large range of relative contributions from systematic uncertainties, correlations among them, and the chosen number of bins in each {pseu\-do-\-mea\-su\-re\-me\-nt}.
Very small deviations of the order of a few per-cent with respect to the total uncertainty are observed.
These are expected as a result of reducing the binned information of the initial measurement likelihood to one or many estimates and a corresponding Hessian.
The method described here shows a similar stability for different choices of $x^a$ and $x^b$, and of the number of uncertainties. Moreover, it is also valid for multiple estimates within one {pseu\-do-\-mea\-su\-re\-me\-nt} without statistical correlations between them. The case of statistical correlations is discussed separately in Section~\ref{sec:dircorr}.
\subsection{Modeling of statistical correlations}
\label{sec:dircorr}
The correct modelling of statistical correlations between the estimates within a measurement is tested by generating two {pseu\-do-\-mea\-su\-re\-me\-nt}s $a$ and $b$ similar to Section~\ref{sec:sysunc}, each with two estimates $x^a_1$ and $x^a_2$ or $x^b_1$ and $x^b_2$, respectively. The corresponding correlation matrices $S^a$ and $S^b$ are randomly chosen to have off-diagonal elements with an absolute value of $d\pm 0.1$. In total, 10000 combinations are performed for each choice of $d=\{0,\ 0.3,\ 0.9\}$, $t=\{0.01,\ 0.50\}$, and $c=\{0.00, 0.99\}$. The values for $x^\alpha_\mu$ are chosen to be $x^a_1 = 30000 $, $x^b_1 = 30600$, $x^a_2 = 20000$, and $x^b_2 = 20500$.
The resulting values for $\Delta \bar{x}_1/ \sigma_{\bar{x},1}$ and $\Delta^r\sigma_1$ are illustrated in Figure~\ref{fig:dircorr} and~\ref{fig:dircorr1a} for $c=0$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.4\textwidth]{reldev_dtscan.pdf}
\caption{
Difference between the combined value of $\bar{x}_1$ obtained with the method described here and using a direct likelihood combination relative to the total uncertainty of $\bar{x}_1$. The distribution is shown for different values of the scale $t$ for systematic uncertainties and the statistical correlation between the estimates $d$.
The combination assumes no correlation between the uncertainties of both {pseu\-do-\-mea\-su\-re\-me\-nt}s.
\label{fig:dircorr}}
\end{center}
\end{figure}
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.4\textwidth]{errdev_dtscan.pdf}
\caption{
Ratio of the uncertainties on $\bar{x}_1$ obtained with the method proposed in this document and using a direct likelihood combination. The distribution is shown for different values of the scale for systematic uncertainties $t$ and the statistical correlation between the estimates $d$. The combination assumes no correlation between the uncertainties of both {pseu\-do-\-mea\-su\-re\-me\-nt}s.
\label{fig:dircorr1a}}
\end{center}
\end{figure}
No significant mismodelling of the statistical correlation between estimates of the same measurement can be observed.
The dependence on $c$ is similar to the one discussed in Section~\ref{sec:sysunc} and not shown here. Also the result of the combination of $x_2$ shows identical behaviour and is therefore not depicted either. Different choices for $x^\alpha_\mu$ were tested and confirm a good modelling with respect to $d$.
\subsection{Relative uncertainties}
The modelling of relative uncertainties is studied by generating two {pseu\-do-\-mea\-su\-re\-me\-nt}s, each of them with one parameter to be combined, one relative uncertainty, and two absolute uncertainties. The relative uncertainty applies to all bins in the same way and will therefore not receive constraints. In consequence, it will be dominant. Thus, the total uncertainty of each {pseu\-do-\-mea\-su\-re\-me\-nt} will differ from the dependence on $t$ previously illustrated in Figure~\ref{fig:deltaxbarrel}. A total of $2 \times 5000$ {pseu\-do-\-mea\-su\-re\-me\-nt}s are generated.
Figure~\ref{fig:relerrrelunc} shows the relative uncertainty of {pseu\-do-\-mea\-su\-re\-me\-nt} $a$, including one relative uncertainty, as a function of $t$. For $t$ larger than 0.15, the direct likelihood combination shows instabilities in some cases, likely related to the Gaussian penalty terms, while log-normal terms would be more suitable for large relative uncertainties.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.4\textwidth]{errcontr_relunc.pdf}
\caption{Relative total uncertainty of the {pseu\-do-\-mea\-su\-re\-me\-nt} $a$ for different values of the threshold $t$ for systematic uncertainties. All {pseu\-do-\-mea\-su\-re\-me\-nt}s comprise one relative and two absolute uncertainties.
\label{fig:relerrrelunc}}
\end{center}
\end{figure}
As shown in Figures~\ref{fig:relerrrelunc2} and~\ref{fig:relerrrelunc3}, also when combining {pseu\-do-\-mea\-su\-re\-me\-nt}s with contributions from relative uncertainties, central values and uncertainties are well modelled, assuming the uncertainties of one {pseu\-do-\-mea\-su\-re\-me\-nt} to be uncorrelated with the uncertainties of the other. The same holds true for high correlations between the {pseu\-do-\-mea\-su\-re\-me\-nt}s.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.4\textwidth]{reldev_relunc.pdf}
\caption{Difference between the combined value of $\bar{x}$ obtained with the method proposed here and using a direct likelihood combination relative to the total uncertainty of $\bar{x}$, shown for different values of the threshold $t$ for systematic uncertainties. All {pseu\-do-\-mea\-su\-re\-me\-nt}s comprise one relative and two absolute uncertainties.
\label{fig:relerrrelunc2}}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.4\textwidth]{errdev_relunc.pdf}
\caption{Ratio of the uncertainties on $\bar{x}$ obtained with the method described in this document and using a direct likelihood combination, shown for different values of the threshold $t$ for systematic uncertainties. All {pseu\-do-\-mea\-su\-re\-me\-nt}s comprise one relative and two absolute uncertainties.
\label{fig:relerrrelunc3}}
\end{center}
\end{figure}
Additionally, the method is validated using exactly one estimate per {pseu\-do-\-mea\-su\-re\-me\-nt} and one large relative uncertainty of +15\%. The input estimates are set to $30000 + \beta$, where $\beta$ is a randomly generated value between 0 and 750.
The uncertainty is assumed to be fully correlated between the {pseu\-do-\-mea\-su\-re\-me\-nt}s. This results in asymmetric uncertainties on each {pseu\-do-\-mea\-su\-re\-me\-nt} and the combined value. Moreover, for this particular choice of uncertainties, the combined value can be larger than the highest input estimate.
When comparing the direct likelihood combination to the method proposed here, also in this case no bias with respect to the central value or its uncertainties can be observed.
In summary, the combination method described here does not require the original data and the full fit model, but sufficiently describes the initial measurement for a large variety of possible central values, binning choices and uncertainties. In consequence, the combination results are numerically equivalent to a using the full likelihood information, in particular in case of dominant systematic uncertainties.
\
\section{Neglecting correlations}
\label{sec:negcorr}
For comparison, two {pseu\-do-\-mea\-su\-re\-me\-nt}s $a$ and $b$ with two bins, two systematic uncertainties, and the same parameters described in Section~\ref{sec:sysunc} are combined neglecting correlations between uncertainties within the same {pseu\-do-\-mea\-su\-re\-me\-nt}, but still considering strong correlations between {pseu\-do-\-mea\-su\-re\-me\-nt}s $a$ and $b$. This approximates the situation in which the BLUE method~\cite{LYONS1988110,Nisius1,Nisius2} can be used for the combination.
The correlations within one {pseu\-do-\-mea\-su\-re\-me\-nt} are removed by inverting $(D^\alpha + \mathds{1})$ in Eq.~\ref{eq:evchi2}, removing the off-diagonal elements of the resulting covariance matrix, and replacing $D^\alpha$ by the inverse of this covariance matrix minus $\mathds{1}$. By choosing the Neyman $\chi^2$ definition in addition, this makes this test equivalent to the BLUE method.
As shown in Figure~\ref{fig:nocorr}, this approximation can lead to wrong individual combination results with respect to the central value when the contribution of systematic uncertainties becomes non-negligible.
\begin{figure}[hbtp]
\begin{center}
\includegraphics[width=0.4\textwidth]{reldev_nocorr.pdf}
\caption{
Difference between the combined results neglecting correlations between uncertainties within a measurement and using a direct likelihood combination relative to the total uncertainty on $\bar{x}$.
The distribution is shown for different values of the upper threshold for systematic uncertainties $t$. For comparison with the method proposed here, see Figs.~\ref{fig:errb2} and~\ref{fig:deltaxbarrel}.
\label{fig:nocorr}}
\end{center}
\end{figure}
Also, the uncertainty on the combined value can be severely mismodelled if the correlations within one measurement are neglected, as displayed in Figure~\ref{fig:nocorr2}. The total uncertainty can be underestimated or strongly overestimated, in particular if it is dominated by systematic uncertainties.
Therefore, it is crucial to model these correlations consistently when performing a combination of results obtained in simultaneous fits of systematic uncertainties and the quantity to be determined.
\begin{figure}[hbtp]
\begin{center}
\includegraphics[width=0.4\textwidth]{err_dev_nocorr.pdf}
\caption{
Ratio of the uncertainties on the combined value obtained neglecting correlations between uncertainties within a measurement and using a direct likelihood combination. The distribution is shown for different values of the upper threshold for systematic uncertainties $t$. For comparison with the method proposed here, see Figs.~\ref{fig:errb2} and~\ref{fig:deltaxbarrel}.
\label{fig:nocorr2}}
\end{center}
\end{figure}
\
\input{software_man.tex}
\section{Summary}
\label{sec:summary}
The combination method presented in this document allows combining measurements obtained with simultaneous nuisance parameter fits consistently, taking into account the constraints from the data as well as correlations between systematic uncertainties within each measurement.
In contrast to the optimal case of a direct likelihood combination, based on the product of the individual likelihoods of each measurement, the method does not require the full knowledge of the original data and the fit models. This information would also be required by other commonly used combination methods, however, it is publicly available only in rare cases. It is shown that not accounting for correlations between uncertainties within the same measurement can lead to non-negligible deviations from the combined likelihood approach with respect to the combined value and its uncertainty.
The method described here does not introduce such deviations and relies on the central results and their covariances or Hessians, only, which makes it applicable to a significantly larger variety of combinations.
An extensive validation is performed using {pseu\-do-\-mea\-su\-re\-me\-nt} with varying contributions of statistical and systematic uncertainties, correlation assumptions, binning choices, and {pri-or} statistical correlations. All obtained results and uncertainties are numerically equivalent to a direct likelihood combination. Only for measurements strongly limited by statistical precision, the same known caveats as in other $\chi^2$ or least-squares-based approaches (e.g. the BLUE method) apply.
In addition, the Convino program is presented. It is developed to perform combinations using the method described here and provides a text-based and a C\kern-0.1em\raisebox{.2ex}+\kern-0.1em\raisebox{.2ex}+{} user interface. The text-based user interface provides an automatic scan of correlation assumptions and creates the corresponding figures for graphical representation.
\section*{Acknowledgements}
I would like to thank O. Behnke for fruitful discussions and comments alongside M. Aldaya Martin, C. Diez Pardos, M. Mulders, and A. Giammanco for editorial suggestions.
\bibliographystyle{ieeetr}
\section{Program Installation and User Interface}
\label{sec:install}
The method described in Section~\ref{sec:math} is implemented in the dedicated Convino program for the combination of experimental results.
The source code can be found at \path{https://github.com/jkiesele/Convino/releases}. It can be compiled using \texttt{make} with gcc version~$4.9$ or newer, or clang~8.0.0 or newer (OSX) and ROOT~6 installed on the system. Other versions might be sufficient but are not tested.
The measurements and the configuration for the combination are contained in human-readable text files.
Alternatively, a C\kern-0.1em\raisebox{.2ex}+\kern-0.1em\raisebox{.2ex}+{} library is provided with the software package, providing an interface to {C\kern-0.1em\raisebox{.2ex}+\kern-0.1em\raisebox{.2ex}+{}} standard-library or ROOT classes, the latter commonly used in high-energy physics.
Both interfaces are described in the following, starting with the text-based interface. The discussion of the text-based interface serves as reference for the description of the C\kern-0.1em\raisebox{.2ex}+\kern-0.1em\raisebox{.2ex}+{} interface.
\subsection{Text-based Interface}
\label{sec:tbui}
The ``convino" executable can be found in the base directory after compiling. It prints usage information and a list of options if the \verb+-h+ option is specified. Other options are:
\begin{itemize}
\item[] \verb+-s+ perform correlation scan
\item[] \verb+-p+ save scan plots as .pdf in addition to a .root file
\item[] \verb+-d+ switch on debug printout
\item[] \verb+--neyman+ uses a Neyman $\chi^2$ instead of the Pearson $\chi^2$
\item[] \verb+--prefix+ defines a prefix for all output files and directories
\end{itemize}
In addition to the options, a text file is passed to the executable. It is referred to as \textit{base file} in the following and is described in Section~\ref{sec:basefile}.
Each measurement comprising one or a set of estimates is described in a \textit{measurement file}. Well documented examples for both types of files are provided in the \texttt{examples} directory and should be consulted alongside this manual.
\subsubsection{Measurement File}
Each measurement file consists of blocks. Each block describes estimates or uncertainties. They are defined by a Hessian, a correlation matrix together with constraints, or a set of orthogonal uncertainties. The latter should be provided in the following format:
\begin{verbatim}
[not fitted]
sys_a1 sys_b1 sys_c1 stat
estimate_a1 5 6.1 2 6
estimate_b1 3 1 4 2
[end not fitted]
\end{verbatim}
The uncertainties \texttt{sys\_XX} on the estimates \texttt{esti\-mate\_XX} are given in absolute values. The keyword \texttt{stat} is reserved for the statistical uncertainty.
The uncertainties and their effect on the estimates in a measurement using a simultaneous nuisance parameter fit technique are described either by a Hessian or a correlation matrix. The Hessian must be written in the following form:
\begin{verbatim}
[hessian]
sys_a3 1944.6
sys_b3 -1349. 1154.4
estimate_a3 -0.525 0.5398 3.25e-4
estimate_b3 -0.708 0.2706 0 4.89e-4
[end hessian]
\end{verbatim}
while the correlation matrix has to include additional information about the constraints on the parameters. These constraints are given in units of $1\sigma$ variations for the systematic uncertainties, such that a value of 1 corresponds to no reduction and lower values indicate a constraint from the fit to the data. For estimates, the constraints are given in absolute units and correspond to the total uncertainties. In both cases, they are defined in parentheses, such that the correlation matrix is of the format:
\begin{verbatim}
[correlation matrix]
sys_a2 (1) 1
sys_b2 (1) -0.2 1
estimate_a2 (10.6) 0.547945 0.1059 1
estimate_b2 (12.8) 0.147945 0.4305 0 1
[end correlation matrix]
\end{verbatim}
If uncertainties have been described in form of a Hessian or correlation matrix, additional contributions from orthogonal uncertainties can be provided in the \texttt{[not fitted]} block. These uncertainties must not have any correlation with the uncertainties defined in the Hessian or the correlation matrix.
The next block of the measurement file describes the type of each uncertainty.
\begin{verbatim}
[systematics]
sys_a2 = absolute
sys_b2 = relative
[end systematics]
\end{verbatim}
The type can be either \texttt{ab\-so\-lute} or \texttt{re\-la\-tive}. The default is \texttt{ab\-so\-lute} and does not need to be specified explicitly.
The last block defines which of the parameters are estimates, and their nominal values:%
\begin{verbatim}
[estimates]
n_estimates = 2
name_0 = estimate_a2
value_0 = 780
name_1 = estimate_b2
value_1 = 280
[end measurements]
\end{verbatim}
Here, \texttt{n\_es\-ti\-mates} gives the number of estimates.
\subsubsection{Base File}
\label{sec:basefile}
The first block of the base file defines the number of measurement files (\texttt{nFiles}) to be considered for the combination and the corresponding file names. An example is given below:
\begin{verbatim}
[input]
nFiles = 2
file0 = exampleMeasurement1.txt
file1 = exampleMeasurement2.txt
[end input]
\end{verbatim}
The files must be in the same directory as the base file.
The second block defines the observables, the estimates should be combined to:
\begin{verbatim}
[observables]
combined_a = estimate_a1 + estimate_a2
combined_b = estimate_b1 + estimate_b2
[end observables]
\end{verbatim}
Here, \texttt{es\-ti\-mate\_a1} and \texttt{es\-ti\-mate\_a2} should be combined to \texttt{com\-bined\_a}, and similarly for \texttt{es\-ti\-mate\_b1} and \texttt{es\-ti\-mate\_b2}. The number of estimates that should be combined to a single quantity is not limited, as well as the number of combined values. This makes it possible to combine simultaneously e.g. a large amount of bins from differential cross sections from various channels and experiments. However, in this case, the C\kern-0.1em\raisebox{.2ex}+\kern-0.1em\raisebox{.2ex}+{} interface is probably more practical.
The last block describes the correlations that should be assigned using the following syntax.
\begin{verbatim}
[correlations]
sys_b1 = (0.2) sys_c2
sys_c1 = (-0.3) sys_d2
[end correlations]
\end{verbatim}
Here, a correlation coefficient of 0.2 is assigned between \texttt{s\-ys\_b1} and \texttt{s\-ys\_c2} and -0.3 between \texttt{s\-ys\_c1} and \texttt{s\-ys\_d2}.
The correlation assumptions between the parameters can be scanned in an automated way. In this case, the following syntax is used to define the scan ranges:
\begin{verbatim}
[correlations]
sys_b1 = (0.2 & -0.1 : 0.4) sys_c2
[end correlations]
\end{verbatim}
Here, \texttt{s\-ys\_b1} has a nominal correlation of 0.2 to \texttt{s\-ys\_c2} . The correlation is scanned from -0.1 to 0.4.
If several correlation coefficients should be scanned simultaneously, they have to be specified in a single line:
\begin{verbatim}
[correlations]
sys_b1=(0.2&-0.1:0.4)sys_c2+(-0.3&0.2:-0.3)sys_d2
[end correlations]
\end{verbatim}
In this case, the scan range for a single coefficient can start from positive values to negative values to allow accounting for anti-correlations between the parameters that are scanned simultaneously.
Correlation matrices are positive definite by definition, a correlation matrix $C$ with large off-diagonal entries might lose this property if ill-posed assumptions are made, such as:
\begin{equation}
\label{eq:wrongcorr}
C = \begin{pmatrix}
1 & .99 & 0 \\
.99 & 1 & 0.5 \\
0 & 0.5 & 1 \\
\end{pmatrix}\text{.}
\end{equation}
In this case, the program exits and it is strongly advised to revise the plausibility of the correlation assumptions.
The results of the combination are saved in the output file \path{result.txt}, or \path{<prefix>_result.txt} in case a prefix is specified.
The output file contains the original input correlations, the combination results, the minimum $\chi^2$, and pulls and constraints on all parameters.
The output of the scan, including all correlation matrices, is saved in the file \path{scan_result.txt}. The corresponding figures are saved as \texttt{TGraph\-Asymm\-Errors} classes in the file \path{scanPlots.root}. If pdf-file output was enabled, the resulting Figures can be found in the directory \path{scan_results}.
Examples of such Figures obtained with the example configuration are shown in Figure~\ref{fig:scan} and~\ref{fig:scan2}.
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.4\textwidth]{combined_a_sys_a1_sys_b2.pdf}
\end{center}
\caption{Combined value for \texttt{com\-bined\_a}
for a scan of the correlation coefficient for \texttt{sys\_a1}
and \texttt{sys\_b2}.
The open marker shows the result obtained with the nominal assumption and its uncertainty.The shaded area represents the uncertainty associated to the scanned dependence, indicated by a continuous line. All values are obtained with the example configuration. \label{fig:scan}
}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=0.4\textwidth]{combined_b_sys_a1_sys_b2_detailed.pdf}
\end{center}
\caption{Combined values for \texttt{com\-bined\_b}
(upper panel), and minimum $\chi^2$ (lower panel) for a scan of the correlation coefficient for \texttt{s\-ys\_a1}
and \texttt{s\-ys\_b2}.
The open marker shows the result obtained with the nominal assumption and its uncertainty. The shaded area represents the uncertainty associated to the scanned dependence, indicated by a continuous line. All values are obtained with the example configuration. \label{fig:scan2}
}
\end{figure}
\subsection{C++ Interface}
The C\kern-0.1em\raisebox{.2ex}+\kern-0.1em\raisebox{.2ex}+{} interface is optimized for the combination of differential distributions and provides three basic classes which will be described in the following: the class \texttt{meas\-ure\-ment}, which is analogous to a measurement file discussed in the previous Section, the class \texttt{com\-bi\-ner} to perform the combination, and a class \texttt{com\-bi\-na\-tion\-Re\-sult} that collects the output of the combination. The \texttt{meas\-ure\-ment} class and the \texttt{com\-bi\-na\-tion\-Re\-sult} class provide interfaces to C\kern-0.1em\raisebox{.2ex}+\kern-0.1em\raisebox{.2ex}+{} standard library \texttt{std::vec\-tor\-<dou\-ble>} or alternatively to ROOT histograms and graphs.
An example of the usage is provided in \texttt{bin/\-differ\-ential\-Ex\-ample.cpp}. Any cpp file that will be placed in the \path{bin} directory will be compiled automatically when running \path{make}.
Alternatively, the compilation of the Convino{} package will create the library \path{libconvino.so} that can be linked against. The header files can be found in the \path{include} directory.
Each class is documented in the corresponding header file. Therefore, the documentation here is limited to the general usage.
\subsubsection{Measurement class}
The measurement class provides the possibility to define a set of estimates, their statistical correlations and systematic uncertainties.
Each object can only contain one set of estimates at once. In case the information is read from a ROOT \texttt{TH1} histogram, each measurement class object can contain only one nominal histogram.
For a measurement with orthogonal uncertainties, the following procedure should be applied:
the nominal values are set using the function \texttt{set\-Meas\-ured}.
Systematic uncertainties can be added in a second step to the measurement object with \texttt{add\-Sys\-te\-matics}.
The type of each uncertainty is defined using the function using \texttt{set\-Para\-meter\-Type} after all uncertainties have been added. Here, it is recommended to use the parameter name to identify the correct uncertainty.
In a last step, statistical correlations between the estimates can be set using the method \texttt{set\-Es\-ti\-mate\-Cor\-re\-lation}.
If a measurement comprises correlated uncertainties, the corresponding measurement object should be configured using the function \texttt{set\-Hes\-sian}, which defines the uncertainties and estimates at once. Additional orthogonal uncertainties can be added in a subsequent step using \texttt{add\-Sys\-te\-ma\-tics}.
\subsubsection{Combiner class}
Once the individual \texttt{meas\-ure\-ment} objects are defined, they are added to a \texttt{com\-bi\-ner} object using the function \texttt{add\-Mea\-sure\-ment}.
For the following combination, it is assumed that the entries of each measurement in the same bin or with the same vector index should be combined. It is not possible to combine a number of estimates from one measurement object with a different number of estimates from another.
The correlation assumptions are defined with \texttt{set\-Syst\-Cor\-rel\-ation}. It is advised to use the uncertainties names as input for unambiguous identification.
The combination is initiated by calling the method \texttt{com\-bi\-ne}, which returns a \texttt{com\-bi\-na\-tion\-Re\-sult} class object.
\subsubsection{CombinationResult class}
The \texttt{com\-bi\-na\-tion\-Re\-sult} class is a container for all information regarding the inputs to the combination, the correlation matrices, and the combined values as well as the post-combination correlation matrices, pulls and constraints.
If differential distributions are combined, the result can be fed back to a ROOT \texttt{TH1} object or a \texttt{TGraph\-Asymm\-Errors} using the functions \texttt{fill\-TH1} or \texttt{fill\-TGraph\-Asymm\-Errors}.
| 456b3657615e3e77cd205f32ced11ad562ea2d25 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\subsection{1. Introduction}
The applications of gamma-ray are ubiquitous in our daily life, such as container security initiative\cite{lun2008institutional}, gamma-knife surgery\cite{ganz2012gamma}, nuclear medical imaging\cite{eisen1999cdte} and food storage\cite{LADO2002433}. While in the vastness of the universe, photons, ranging from several MeV to tens of TeV\cite{lamb1997point,abdo2010fermi,aharonian2004crab}, results from various different processes, such as energetic cosmic ray\cite{kulsrud1969effect,hunter1997egret}, luminous pulsars\cite{romani1996gamma} and gamma-ray burst\cite{gamma_ray_burst_RMP,gamma_ray_burst_theories}. The information of gamma-ray burst was firstly published by the results of Vela satellites\cite{klebesadel1973observations} and then were quickly verified by data from the Soviet satellites\cite{mazets1974cosmic}. The ability of cosmic sources to emit such intense gamma-rays indicates that investigating this extreme environment is a promising route to discover new physics which are impossible in earth-bound laboratories.
An alternative method of generating violent emission of gamma-rays is through the interaction of petawatt ($10^{15}$W) lasers and plasmas in the laboratory. Several multi-PW laser facilities, such as Extreme Light Infrastructure (ELI)\cite{ELI} and Exawatt Center for Extreme Light Studies (XCELS)\cite{XCELS}, are expected to operate at intensities beyond 10$^{23}$W/cm$^{2}$ in next few years. Under $\sim$10$^{23}$W/cm$^{2}$, various theoretical schemes have been put forward for multi-MeV photon sources with tens of percent for the total conversion efficiency, such as
reinjected electron synchrotron radiation\cite{Brady2012PRL}, skin-depth emission\cite{Ridgers2012PRL}, radiation reaction facilitating gamma-ray\cite{Nakamura2012PRL} and sandwich target design\cite{Stark2016PRL}.
Nevertheless, none of them has the ability to extend the energy of gamma photon up to several GeV, which is highly desirable to explore the laboratory astrophysics\cite{RevModPhys.78.755,RRD_bulanov2015}. Recently, exploiting the interplay between pair cascades\cite{bell2008possibility} and anomalous radiative trapping\cite{gonoskov2014anomalous}, ultrabright GeV photon source can be achieved in laser-dipole waves\cite{gonoskov2017ultrabright}. However, the scheme of dipole wave field\cite{gonoskov2012dipole,gonoskov2013probing,gonoskov2014anomalous} requires multi beams focused into a tiny point symmetrically, which is still an experimental challenge nowadays. Here we report an alternative all-optical scheme to realize the brilliant GeV gamma-ray emission via irradiating only one multi-PW circularly polarized (CP) pulse on a compound target in QED regime. This all optical backscatter scheme is already available in experiment for relative lower intensity circumstance\cite{phuoc2012all,chen2013mev,powers2014quasi,sarri2014ultrahigh,khrennikov2015tunable}.
\subsection{2. Theoretical model for coupling effect}
In the realm of nonlinear QED, electrons are able to emit a huge amount of kinetic energy in the form of high-energy photons $\gamma_{ph}$, as a result of absorbing a certain number $n$ of laser photons $\gamma_l$, $e^- + n\gamma_l \rightarrow e^- + \gamma_{ph}$. The invariant parameters $\eta=(e\hbar/m_e^3c^4)|F_{\mu\nu}p^\nu|=E_{RF}/E_{Sch}$ and $\chi=(e\hbar^2/2m_e^3c^4)|F_{\mu\nu}k^\nu|$ characterize the discrete photon emission process, where $e$ the electron charge, $m_e$ the electron rest mass, $\hbar$ the Planck constant, $c$ the light velocity in vacuum, $F_{\mu\nu}$ the field tensor and $p^\nu$ ($k^\nu$) the electron's (photon's) four-momentum. $E_{RF}$ denotes the electric field in the electron's rest frame and $E_{Sch}=m_e^2c^3/e\hbar\approx1.3\times10^{18}Vm^{-1}$ is the characteristic field of Schwinger limit\cite{schwinger1951gauge}. When $\eta\lesssim1$:(1) The radiation process should be described by probabilistic quantum emission rather than continuous one. (2) The corresponding quantum weaken correction for radiation is inevitable\cite{kirk2009pair,QED_domian_di2012}. When an electron beam co-propagates with the laser pulse, the electric force offset by the magnetic field effect results in $\eta\approx0$, which is undesired for high-energy photon emission\cite{RRD_bulanov2015,QED_domian_di2012}. However, if the laser counter-propagates with the electron beam, it leads to an enhancement as $\eta\approx 2\gamma E_L/E_{Sch}$, where $\gamma$ is the relativistic Lorentz factor of electron and $E_L$ is the polarized laser field. This colliding configuration can not only lower down the threshold of QED cascade from seed electrons\cite{grismayer2016laser}, but also facilitate the generation of $\gamma$-ray explosion\cite{nerush2011laser,gong2016high} and pair plasma\cite{bell2008possibility,zhu2016dense,jirka2016electron,chang2015generation}.
\begin{figure*}[tbp]
\includegraphics[keepaspectratio=true,height=60mm]{fig1.jpg}
\caption{Scheme of the ultra-brilliant GeV gamma-ray source with helical structure. (a) and (b) show light being reflected before and after respectively.}
\label{fig_schematic}
\end{figure*}
To exploit the counter-propagating configuration, in this letter, a CP femtosecond pulse was irradiated on a compound target (in Fig.\ref{fig_schematic}) consisted of a near-critical-density (NCD) plasma slab and a solid foil. Here the solid foil plays the role as a plasma mirror\cite{phuoc2012all,chen2013mev,powers2014quasi,sarri2014ultrahigh,khrennikov2015tunable} to spontaneously reflect the driven light to trigger the subsequent Compton backscattering. Generally, when a CP pulse of 10$^{19-21}$W/cm$^{2}$ propagates in the NCD target, the ionized electrons can be transversely expelled from central area to form a plasma channel\cite{pukhov1999channel,pukhov2002strong}. Some injected electron can experience a direct laser acceleration process and a collimated energetic electron bunch can be produced when its oscillation frequency in the channel field is close to the light frequency witnessed by the electron\cite{liu2013generating,hu2015dense,arefiev2012parametric}. However, under the higher intensity of $\sim$10$^{23}$W/cm$^{2}$, the injected electrons are mostly expelled from the central region and a hollow channel is merely filled with laser radiation\cite{ji2014radiation}. More interestingly, a great amount of electrons will be trapped back into the channel if radiation reaction (RR) is taken into account\cite{ji2014radiation,RR2016Chang}, where transverse ponderomotive force is properly balanced by the radiation recoil.
It should be noted that the interaction between laser and NCD plasma is very complicated, where the filamentation instability\cite{honda2000collective}, hosing instability\cite{huang2017relativistic} or non optimal laser-plasma matching\cite{mourou2002design} can destroy the laser propagating and the channel's shape. Here a relatively large spot radius and the small plasma density are adopted to avoid these detrimental influence and guarantee the stable channel. To understand the underlying mechanism of RR impact on this scheme, the single electron model is utilized to depict the interaction with laser transverse field $E_L$ and self-generated fields in the plasma channel. Based on previous work\cite{pukhov2002strong,liu2015quasimonoenergetic,hu2015dense}, self-generated fields in the channel include radial electrostatic field $\mathbf{E}_{Sr}=k_Er\hat{e}_r$, longitudinal electric field $\mathbf{E}_{S\parallel}$ and quasistatic azimuthal magnetic field $\mathbf{B}_{S\theta}=-k_Br\hat{e}_\theta$, where $k_E$ and $k_B$ can be seen as constant and are related to the plasma density. The time derivative of the ponderomotive phase $\psi$ can be written as
\begin{equation}
\begin{aligned}\label{eq1}
\frac{d\psi}{dt}=\omega_{\beta}-\omega_L=\sqrt{\frac{e}{\gamma m_e}(v_\parallel \langle k_B\rangle+\langle k_E\rangle)}-(1-v_\parallel/v_{ph})\omega_0.
\end{aligned}
\end{equation}
Here $\omega_{\beta}=\sqrt{e(v_\parallel k_B+k_E)/(\gamma m_e)}$ is the electron betatron frequency and $v_\parallel$ ($v_\perp$) the electron longitudinal (transverse) velocity. $\omega_L=(1-v_\parallel/v_{ph})\omega_0$ is the Doppler-shifted laser frequency witnessed by electron, where $\omega_0$ is the laser frequency and $v_{ph}=c/\sqrt{1-\omega_p^2/(\gamma\omega_0^2)}$ is the laser phase velocity\cite{decker1995group}. $\omega_p$ is the plasma frequency. The $\psi$ is relative phase between the electron rotation and the periodic laser field. The time derivative of the electron Lorentz factor is expressed as
\begin{equation}
\begin{aligned}\label{eq2}
\frac{d\gamma}{dt}=\frac{-e\mathbf{E}\cdot\mathbf{v}-\mathbf{f_{rad}}\cdot\mathbf{v}}{m_ec^2}=-\frac{e( v_\perp E_L cos\psi+v_\parallel \langle E_\parallel\rangle)}{m_ec^2}-\epsilon_{rad}\omega_0\beta^2a_s^2\eta^2G(\eta).
\end{aligned}
\end{equation}
Here $E_L$ is the light electric field amplitude. Since the stochasticity of photon emission is difficult to be simplified into a precise formula, the discontinuous influence is neglected in the single model and the quantum corrected RR force $\textbf{f}_{rad}\approx-G(\eta)\epsilon_{rad}m_ec\omega_0\vec{\beta}a_{s}^2\eta^2$ is used in Eq.(\ref{eq2}) to qualitatively analyze the RR influences, where $G(\eta)\approx(1+12\eta+31\eta^2+3.7\eta^3)^{-4/9}$ is the quantum weaken factor\cite{kirk2009pair}. The impacts issued from the discrete stochasticity in RR is beyond the scope of this manuscript and these are worth discussing in the future work. $\epsilon_{rad}=4\pi r_e/3\lambda_0$ is the dimensionless ratio, where $r_e=e^2/m_ec^2\approx2.8\times10^{-15}m$ is the classical electron radius and $\lambda_0$ is the laser wavelength. $\vec{\beta}=\vec{v}/c$ is the normalized electron velocity and $a_{s}=eE_{Sch}/m_e\omega_0c$ is the normalized Schwinger field. The parameters in above equations depend on time and are probably in especially complicated form so that the average values denoted by $\langle\ \rangle$ are used. From Eqs.(1)-(2), it can be found that the phase space ($\psi$,$\gamma$) has a fixed point\cite{jordan2007nonlinear,hirsch2012differential} at ($\psi_0$,$\gamma_0$) = ($\cos^{-1}\frac{-\epsilon_{rad}\omega_0\beta^2a_s^2\eta^2G(\eta)m_ec^2-e v_\parallel \langle E_\parallel\rangle}{e v_\perp E_L},\frac{e(v_\parallel \langle k_B\rangle+\langle k_E\rangle)}{m_e(1-v_\parallel/v_{ph})^2\omega_0^2}$). To determine the system dynamic property from Eqs.(1)-(2) in ($\psi$,$\gamma$) space, the perturbation expansion nearby ($\psi_0$,$\gamma_0$) of Eqs.(1)-(2) was made and quadratic terms were dropped to approach the characteristic Jacobian matrix $\mathbf{Ja}$\cite{jordan2007nonlinear,hirsch2012differential}:
\begin{equation}
\begin{aligned}
\mathbf{Ja}\approx\begin{pmatrix} 0 & -\frac{1}{2}\sqrt{\frac{e}{\gamma^3m_e}(v_\parallel \langle k_B\rangle+\langle k_E\rangle)} \\ ev_\perp E_Lsin\psi & -\epsilon_{rad}m_ec^2\omega_0\beta^2a_{s}^2\frac{\partial G(\eta)\eta^2}{\partial\gamma} \end{pmatrix}_{\psi_0,\gamma_0}.
\end{aligned}
\label{eq3}
\end{equation}
Without RR effect, the trace and determinant of Jacobian matrix are tr(Ja)$=$0 and det(Ja)$>$0 when the right lower RR term is canceled, which manifests that ($\psi_0$,$\gamma_0$) is a center without any source or sink property\cite{jordan2007nonlinear,hirsch2012differential}. On the contrary, with RR effect included, at fixed point tr(Ja)$<$0 and det(Ja)$>$0 indicates that its behaviour converts from center to spiral sink attractor\cite{gonoskov2014anomalous,ESIRKEPOV20152044,gong2016radiation,kirk2016radiative}. The sink attractor emerging illustrates a large fraction of the radiation trapped electrons tends to possess the same relative phase $\psi_0$ with respect to laser electric field and the helical density structure is an intrinsic rotary manner of the electric field of CP laser. Due to electron moving in the same direction as the pulse, the electric field $E_L$ counteracts the force from laser magnetic field $B_L$ leading to $\eta\approx\gamma|\mathbf{E}_L+\mathbf{v}\times\mathbf{B}_L|/E_{Sch}\approx0$ and tr(Ja)$\sim$0. Notwithstanding, the strong self-generated magnetic field $B_{s\theta}\approx n_eR/(2\varepsilon_0c)$ (here $\varepsilon_0$ the permittivity of vacuum, $n_e$ the RR trapped electron density and $R$ the channel radius) approaching the order of driven laser field\cite{Stark2016PRL} gives $\eta\approx\gamma|\mathbf{E}_L+\mathbf{v}\times(\mathbf{B}_L+\mathbf{B}_{s\theta})|/E_{Sch}\approx\frac{\gamma B_{s\theta}}{E_{Sch}}$, which results in tr(Ja)$\approx-2\epsilon_{rad}\beta^2e^2B_{s\theta}^2/m_e\omega_0<$0 and enables the attractor effect on achieving such a helical electron bunch (HEB). The nearby electrons are attracted to possess the identical Lorentz factor $\gamma_0$=$e(v_\parallel \langle k_B\rangle+\langle k_E\rangle)/m_e(1-v_\parallel/v_{ph})^2\omega_0^2$. The total angular momentum (AM) along the longitudinal x-axis, i.e. $L=yp_z-zp_y$, acquired by the HEB can also be estimated as
\begin{equation}
{\centering \ L \approx -\int\Sigma_{i}er_\perp E_Lcos\psi_idt \ \ \ i=1,2,3...
}
\label{eq4}
\end{equation}
here r$_{\perp}$ is the electron transverse radius and the index i refers to the i-th electron. From Eq.(4) we can see that the laser could transfer its spin angular momentum (SAM) to HEB only when most of electrons possess the same ponderomotive phase $\psi_i$, otherwise ensemble average leads to $\sum_{i}cos\psi_i\approx$0. Therefore, coupling effects among RR trapping, self-generated magnetic field and spiral attractor in phase space, enhance the net AM gain and realize the HEB. Eventually the discrete photon emission\cite{ritus1985quantum,neitz2013stochasticity,blackburn2014quantum} is triggered through the inverse Compton scattering (ICS) between the HEB and reflected light, where prolific high-energy photons inheriting a large fraction of electrons' energy and AM are generated.
\begin{figure}[tbp]
\includegraphics[keepaspectratio=true,height=80mm]{fig2.jpg} \caption{Distribution of electron density in $\psi$-$\gamma$ space at t=50T$_0$ with RR (a) and without RR (b), respectively. (c) Normalized amplitude of $\mathbf{B}_{s\theta}$ averaged over the channel in the plane z=0 at time t=30,45,60 T$_0$, where solid (dash) line denotes the circumstance with (without) RR. (d) presents the number of the electrons inside the channel and total AM of electrons in x direction as a function of the interaction time $t$, where solid (dash) corresponds to the case with (without) RR.}
\label{fig_attractor}
\end{figure}
\subsection{3. Particle-in-cell (PIC) simulation results}
The feasibility and robustness of this scheme are demonstrated by using the self-consistent three dimension PIC code EPOCH\cite{arber2015contemporary}.
A Monte Carlo probabilistic model\cite{duclous2010monte,ridgers2014modelling} has been successfully implemented, which is based on QED corrected synchrotron cross sections and coupled with the subsequent reduction of the electron momentum. Each particle is assigned an optical depth ($\tau$) at which it emits according to $P=1-e^{-\tau}$, where $P\in$[0,1] is chosen at random to consider the quantum correction in the emission processes as well as the straggling. The rates of photon production, $d\tau_\gamma/dt=(\sqrt{3}\alpha_fc\eta)/(\lambda_c\gamma)\int_{0}^{\eta/2}d\chi F(\eta,\chi)/\chi$, are then solved until the optical depth is reached, when the emission event occurs\cite{duclous2010monte}. Here, $\alpha_{fc}$ is the fine structure constant, $\lambda_c=\hbar/(m_ec)\approx3.9\times10^{-13}m$ is the Compton wavelength and $F(\eta,\chi)$ is the quantum synchrotron spectrum\cite{duclous2010monte}.
The incident 1.2$\times$10$^{23}$W/cm$^{2}$ CP pulse propagates along X direction with a profile of $a$=$a_0e^{-(t-t_0)^4/\tau_0^4}e^{-(y^2+z^2)/r_0^2}sin(\omega_0t)$, where $\tau_0$=$5T_0$ denotes the intensity with a full width at half maximum (FWHM) of 25.6fs (T$_0$$\approx$3.3fs is the laser period) and $a_0$=$eE_L/m_e\omega_0c$$\approx$$300$ is the normalized amplitude of the laser field. $r_0$=$5\lambda_0$ is the spot size ($\lambda_0$=1.0$\mu m$). The simulation box is 80$\lambda_0 \times$ 40$\lambda_0 \times$ 40$\lambda_0$ in X $\times$ Y $\times$ Z direction, which has been uniformly divided into 3200 $\times$ 800 $\times$ 800 cells. A hydrogen slab with initial density of $n_e=2n_c$ locates between 10$\lambda_0$ to 60$\lambda_0$ and aluminum foil of $n_e=700n_c$ is placed from 60$\lambda_0$ to 80$\lambda_0$, where $n_c=m_e\omega_0^2/4\pi e^2$ is critical density\cite{gibbon2004short}. The hydrogen slab and aluminum foil contain 4 and 16 macroparticles per cell (for both species), respectively. For reference, there is no obvious difference in our results when we double the number of macroparticle per cell.
\begin{figure*}[tbp]
\includegraphics[keepaspectratio=true,height=100mm]{fig_insert.png}
\caption{(a) and (b) correspond to the distributions of electron density $n_e$ for the case without and with RR, where the absolute value of laser electric field $|E_y|$ is also figured in grey with a transparency of 60\%. The distributions of longitudinal field $E_x$ generated in the plasma channel are shown in (c)(d) as well.}
\label{fig_insert}
\end{figure*}
The electron density distributions in $\gamma-\psi$ space at t=50T$_0$ for the cases with and without RR are presented in Fig.\ref{fig_attractor}(a) and (b). Lorentz factor at the fixed point obtained from Eqs.(1)-(2) as $\gamma_0=\frac{e(v_\parallel\langle k_B\rangle+\langle k_E\rangle)}{m_e(1-v_{\parallel}/v_{ph})^2\omega_0^2}\approx\frac{(v_\parallel/c)(n_e/n_c)}{2[1-v_\parallel/c\sqrt{1-n_e/(a_0n_c)}]^2}$ where $\langle k_B\rangle\approx\frac{en_e}{2\epsilon_0}$,$\omega_0=\sqrt{\frac{n_ce^2}{\epsilon_0m_e}}$ and $v_{ph}\simeq\frac{c}{\sqrt{1-n_e/(a_0n_c)}}$ are taken into account and $\langle k_E\rangle$ is neglected as the transverse static electric field is relatively weak compared with self-generated magnetic field. Substituting $n_e=2n_c$, $a_0=300$ and $v_\parallel=0.9863c$ (from simulation parameters and results) into above equation leads to $\gamma_0=3416$. Considering $\epsilon_{rad}=1.18\times10^{-8}(\frac{1\mu m}{\lambda_0})$, $\beta\approx1$, $a_s\approx4.1\times10^5$, $\eta\approx\frac{\gamma_0B_{s\theta}}{a_s}\approx0.165$, $G(\eta)\approx1$, $\langle E_\parallel\rangle\approx0.015E_L$ and $v_\perp=0.165c$, the relative phase is deduced as $\psi_0=\arccos\frac{-\epsilon_{rad}\omega_0\beta^2a_s^2\eta^2G(\eta)m_ec^2-ev_\parallel\langle E_\parallel\rangle}{ev_\perp E_L}\approx2.24$. When RR force is switched on, most of electrons possess a relative phase $\psi$=2.3 in Fig.\ref{fig_attractor}(b) which is in good agreement with our theoretically derived attractor point ($\psi_0, \gamma_0$)=(2.24, 3416). Since neither RR trapping nor attractor emerging occurs, the number density of electron in Fig.\ref{fig_attractor}(a) is relatively small compared to RR case and it does not behave like the attractor modulated distribution. The self-generated azimuthal magnetic field B$_{s\theta}$ averaged over the channel cross plane z=0 is plotted in Fig.\ref{fig_attractor}(c) with maximum $\approx$0.6MT (normalized value equals 60$m_e\omega_0/e\approx$0.2B$_L$, where B$_L$ is the laser magnetic amplitude) at t=65T$_0$, which demonstrates that RR recoil enhances the $B_{s\theta}$ generation due to the more trapped electron current along longitudinal axis. This kind of self-generated magnetic field in channel can not only enhance the gamma photon emission\cite{Stark2016PRL}, but also help accelerate ions in the rear surface of target\cite{bulanov2015helium}, which has already been verified in experiment under lower laser intensity with shock-compressed gas target\cite{helle2016laser}. The temporal evolution of electron number inside the plasma channel and their total AM $L=\sum_i y_ip_{zi}-z_ip_{yi}$ are recorded in Fig.\ref{fig_attractor}(d) for both cases. It is found that RR not only boosts the electron accumulation inside the channel but also facilitates the AM transfer to HEB, which is in good agreement with the theoretical prediction of Eq.(4). The RR force prevents electrons from being expelled transversely, resulting in a increase of electrons from 172 nano-Coulombs(nC) to 291 nC at t=65T$_0$. The enhancement of electron current strengthens the B$_{s\theta}$, which gives a positive feedback on spiral attractor merging in phase space and effectively favors angular momentum transformation from laser's SAM to HEB's AM.
The electron density distributions for the case with and without RR are shown in Fig.3(a)(b). Here the emergence of helical spatial structure depends on the RR impact, which accords with the attractor facilitating electron density modulation with the similar frequency as laser electric field in Eq.(\ref{eq3}). When RR is switched off, a ball of electrons are injected into the tail of plasma channel and can be accelerated by the longitudinal electric field $E_x$. The distributions of $E_x$ are plotted in Fig.\ref{fig_insert}(c)(d) for with RR case or not. Since the quantity of electron in the channel for RR case is much higher than that for no RR, the sheild effect weakens the accelerating field in RR case when compared to the no RR one.
\begin{figure*}[tbp]
\includegraphics[keepaspectratio=true,height=65mm]{fig3.jpg}
\caption{(a) Volume distribution of the photon energy density where only photon with energy higher than 10MeV is recorded for less computation costs. (b)(c) Final photon angular-spectral distribution for energy higher than 1GeV and 100MeV respectively.}
\label{fig_photon}
\end{figure*}
Since the ponderomotive force of CP pulse avoids the longitudinal oscillation at twice the optical frequency\cite{gibbon2004short}, plasma in the second layer cannot be heated violently and the driven light is substantially reflected. Under colliding configuration, the parameter $\eta\approx2\gamma E_L/E_{Sch}\gtrsim1$ indicates that the discrete incoherent photon emission\cite{QED_domian_di2012} gives a more appropriate description compared with the coherent electromagnetic wave radiation derived from the Li\'{e}nard-Wiechert retarded potential\cite{jackson1999classical}. The volume snapshot of the photon energy density at t=70T$_0$ is exhibited in Fig.\ref{fig_photon}(a) where photon beam inherits spatial helical structure and transverse size of the source is about 1.5$\mu$m. The gamma-ray flash duration is $\sim$16fs roughly equal to half of the laser because the driven pulse and trapped electrons completely overlap inside the channel. The angular-spectral distribution calculated by accumulating the forward photons at t=70T$_0$ over the entire simulation region is shown in Figs.\ref{fig_photon}(b) and (c). Most of energetic photons are highly collimated and predominantly located within an emission polar angle $\phi\leq$15$^\circ$ ($\phi\leq$30$^\circ$) for energies higher than 1GeV (100MeV). In a 0.1\% bandwidth (BW) around 1GeV we have 1.05$\times$10$^8$ photons, implying the brightness of 1.7$\times$10$^{23}$ photons/s/mm$^2$/mrad$^2$/0.1\%BW for the GeV gamma-ray emission. The corresponding source brilliances at 100 MeV and 10 MeV are 2.3$\times$10$^{24}$ and 1.5$\times$10$^{25}$ photons/s/mm$^2$/mrad$^2$/0.1\%BW, respectively. The comparasion among different photon source is illustrated in Fig.\ref{fig_brill}. Our ICS scheme predominantly aims at high brilliance around GeV. Another dipole wave field can achieve the brightest gamma photon emission with 9$\times$10$^{24}$ photons/s/mm$^2$/mrad$^2$/0.1\%BW at GeV\cite{gonoskov2017ultrabright}, but the dipole wave needs to be realized through symmetrically colliding multi pulses, which is still a challenge in experiment. Here our scheme shooting one laser pulse onto double layer target is the most efficient method to generate brilliant GeV gamma ray source\cite{Brady2012PRL,Ridgers2012PRL,Nakamura2012PRL,Stark2016PRL} and it is more experimentally accessible.
\begin{figure*}[tbp]
\includegraphics[keepaspectratio=true,height=60mm]{fig_brill.png}
\caption{Comparison of the peak brilliance of our proposed ICS source with the other existing photon source, e.g., Synchrotron, XFEL and Dipole-cascade\cite{gonoskov2017ultrabright}.}
\label{fig_brill}
\end{figure*}
\subsection{4. Discussion and conclusion}
In Fig.\ref{fig_discussion}(a), the exponential decay spectrum of photon covers higher energy range from 1MeV to several GeV with a cutoff energy at 2.9GeV and that of the electron before(t=60T$_0$) and after(t=70T$_0$) ICS process are presented. The nonlinear QED regime predicts that most photons are emitted with an energy $h\nu_{ph}\approx$0.44$\eta\gamma m_ec^2$\cite{bell2008possibility,kirk2009pair} which carries a large fraction of electron's kinetic energy. It is obvious that the amount of high-energy electron is drastically curtailed with the cutoff-energy declining from 3.9GeV to 2.5GeV and simultaneously most of energy is converted to gamma photons. The temporal evolutions of the particle energy are illustrated in Fig.\ref{fig_discussion}(b), where 14.5\%;4.2\%;0.108\% of the total laser energy is transferred into the gamma-ray photon with energy above 1MeV;100MeV;1GeV. For energies above 100MeV and 1GeV, the photons are emitted almost exclusively by ICS process during 65T$_0<$t$<$70T$_0$. Based on power radiated by a single electron, $P_{rad}=(4\pi m_ec^3/3\lambda_c)\alpha_fc\eta^2G(\eta)$\cite{duclous2010monte,ridgers2014modelling}, the instantaneous radiation power of this regime can be estimated as
\begin{equation}
{\centering P_{rad}\approx\left\{
\begin{aligned}
N_e\frac{4\pi\alpha_fm_ec^3}{3\lambda_c}(\frac{\gamma B_{s\theta}}{E_{Sch}})^2G(\frac{\gamma B_{s\theta}}{E_{Sch}}) &\ \ \ & t<t_{ref}, \\
N_e\frac{4\pi\alpha_fm_ec^3}{3\lambda_c}(\frac{2\gamma E_L}{E_{Sch}})^2G(\frac{2\gamma E_L}{E_{Sch}}) &\ \ \ & t\geq t_{ref}.
\end{aligned}
\right.
}
\label{eq5}
\end{equation}
Here t$_{ref}$=65T$_0$ is the time of light reflecting and $\eta$ is approximated by $\gamma B_{s\theta}/E_{Sch}$ at $t<t_{ref}$ and $2\gamma E_L/E_{Sch}$ at $t\geq t_{ref}$, respectively. The length of NCD plasma $l=50\mu m$ is not comparable with the laser depletion length $L_{depletion}\approx c\tau_0a_0n_c/n_e$=750$\mu m$\cite{pukhov2002strong,lu2007generating}, as a result a large part of laser energy is reflected and backscatter with the electron bunch. In addition, when laser propagates in the NCD plasma, both of its intensity and spot size will change due to the self-focusing, self-modulation, etc. The radius of the self-generated channel is defined by the balance of the ponderomotive and charge separation fields. Here, we choose the laser spot almost the same as the radius of such channel. That results in no significant change of the laser transverse size during the propagation in near critical plasma and we assume they are constant in estimation of Eq.(5). Eq.(5) predicts the radiation power P$_{rad}\approx$0.63PW (t$<$t$_{ref}$) and P$_{rad}\approx$19.2PW (t$\geq$t$_{ref}$) which qualitatively agrees with the simulation results in Fig.\ref{fig_discussion}(b), implying the nonlinear QED ICS based gamma-ray source power of the same order as the infrared incident laser.
\begin{figure}[tbp]
\includegraphics[keepaspectratio=true,height=80mm]{fig4.jpg}
\caption{(a) The energy spectra of electrons at t=60,70T$_0$ and photons at t=70T$_0$. (b) The laser energy conversion to the electrons (black), protons (green) and gamma-ray photons ($>$1MeV in solid blue, $>$100MeV in dash blue and $>$1GeV in solid red). The photon with energy greater than 1GeV, rendering in red, corresponds to the right red axis. The orange solid line plots the theoretical radiation prediction from eq.(4). (c) Temporal evolution of the total AM of electrons, protons and photons ($>$1MeV). (d) The laser energy conversion to $\gamma$-photons with different plasma densities. Here the value of $\gamma_{ph}>$1GeV is times by 20 and horizontal (density) axis is on logarithmic scale.}
\label{fig_discussion}
\end{figure}
The transfer of axial AM from the laser to the particles is plotted in Fig.\ref{fig_discussion}(c). The oscillation of electron and proton AM is due to charged particles interplaying with the laser electromagnetic field. The different sign of electric charge causes the opposite oscillation direction in electron and proton. Since the spiral attractor results in the fixed relative phase between electron velocity and laser electric field, the overall AM of electron rises gradually before backscattering with the reflected pulse. However, photons do not interplay with laser field and their AM has a moderate growth before the ICS. The photons are predominantly emitted from electron modulated by the spiral attractor during 65T$_0<$t$<$70T$_0$ so that a sharp photon AM increase and a pronounced electron AM drop occur in ICS process. In terms of quantum mechanics, the angular momentum carried by a photon of CP laser is $\sigma=\pm1$ for spin motion. The total angular momentum absorbed from laser is approximately expressed as $L_{l}=\delta\frac{W_{l}}{\hbar\omega_0}\sigma\hbar$=1.70$\times$10$^{-12}\delta$ kg$\ast$m$^2$/s, where $W_{l}$ is the whole laser energy and $\delta$ is the absorbing ratio. During ICS process, AM is more efficiently transferred from electron to gamma-ray and eventually the AM of photons reaches 8.2$\times$10$^{-14}$ kg$\ast$m$^2$/s, 4.8\% of the total laser SAM. In addition, a parameter scan has been carried out to investigate energy conversion efficiency for a wide range density 0.2-20n$_c$ of first layer plasma with thickness of 50$\mu$m in Fig.\ref{fig_discussion}(d) and find that there is an optimal condition $n_e\sim n_c$ for realizing the twisted GeV gamma-ray emission. The disadvantage for relatively rarefied plasma (n$_e$=0.2n$_c$) is lack in trapped helical electron amount so that insufficient electron quantity accounts for deficient gamma-ray production, while for relatively dense circumstance (n$_e\gtrsim$10n$_c$) driven laser tends to deplete most of their energy in the first slab and without any remnants to trigger ICS process.
In conclusion, we have shown how the ultra-intense and ultra-bright GeV gamma-ray flash can be achieved by irradiating a prospective 1.2$\times$10$^{23}$W/cm$^{2}$ laser on a compound plasma target in nonlinear QED regime. The initial energetic HEB results from the coupling effects among RR trapping, self-generated magnetic field and emergency of spiral attractor in $\gamma$-$\psi$ space. The helical gamma-ray flash inherits a considerable AM and energy of the parent electron beam through Compton backscattering between HEB and the reflected driven pulse. The final photon source, with unprecedented power of 20 PW and brightness of 1.7$\times$10$^{23}$ photons/s/mm$^2$/mrad$^2$/0.1\% (at 1 GeV), might enable significant development of application in particles physics and laboratory astrophysics. Our scheme is also feasible in the laboratory system where cluster jets\cite{fukuda2009energy} or nano-tube foams\cite{ma2007directly} can be utilized for NCD plasma generation and a solid foil acts as a plasma mirror to reflect laser. Such parameters of the gamma-ray sources will be achieved with the next generation of multi-PW laser facilities in the future.
\section*{Acknowledgements}
The work has been supported by the National Basic Research Program of China (Grant No.2013CBA01502), NSFC (Grant Nos.11535001) and National Grand Instrument Project (2012YQ030142). The PIC code Epoch was in part funded by the UK EPSRC grants EP/G054950/1, EP/G056803/1, EP/G055165/1 and EP/M022463/1. J.Q. Yu wants to thank the Project (2016M600007,2017T100009) funded by China Postdoctoral Science Foundation. Our simulations were carried out in Max Planck Computing and Data Facility and Shanghai Super Computation Center. The author Z.Gong acknowledges fruitful discussions with Prof. S.V. Bulanov and H.X. Chang.
| cb5554107d12ad92507682ac24dd37b7199a5952 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Data analyses of LHC data consist of workflows that utilize a diverse set of software tools to produce physics results. The tools range from large software frameworks like Gaudi\cite{Gaudi2001} to single-purpose scripts written by individual analyzers or analysis teams. The analysis steps that lead to a particular physics result are often not reproducible without significant assistance from the original authors. This severely limits the capability to re-execute the original analysis or to re-use its analysis procedures in new contexts. An important application for such re-use is the systematic re-interpretation of a given analysis with respect to alternative models of new physics\cite{Cranmer2011}. Therefore, it is desirable to have a system to archive analysis code as well as the analysis procedure in a manner, that enables both re-execution and re-use. This document presents work on workflow capture that addresses these issues in a platform and language-agnostic manner.
\subsection{Short anatomy of analysis workflows}
The driving paradigm of LHC analyses is the selection of events within the experiments' dataset and, typically, comparing those events to expectations derived using both data-driven techniques and Monte-Carlo simulations. Since every collision event (whether real or simulated) is independent of the others, the data analysis problem becomes \emph{embarrassingly parallel}. Consequently, the most common task in a LHC analysis is the parallel processing of events by algorithms that transform the event data into higher-level representations (e.g. from raw detector data to reconstructed `analysis objects`) or perform event selection or otherwise reduce the dataset size, for example by selectively storing only partial event information (`thinning').
The main reconstruction transformations are often handled either on a collaboration-wide or physics working group level and use centrally managed and documented code with fixed release schedules and procedures. Transform configurations, such as the used executable and its command line options, are managed centrally as well (e.g. with databases such as AMI\cite{AMI2010}). Therefore, these operations are comparatively easy to preserve and reproduce.
On the other hand, custom code developed by the end-user analysis team is often much harder to reproduce due to the diversity of tools, workflows and computing environments that are used by an individual analysis team. In the case of the ATLAS experiment, a very wide spectrum of analysis frameworks have been used during Run-1 to analyze events in the ``post-AOD'' stage, i.e. after central reconstruction. This ranged from large and complex frameworks such as \texttt{SFrame} or \texttt{Athena}, handling not only the main event-loop, but also managing calibration tool instantiation and data-handling, to pure ROOT-based programs such as \texttt{TTree::MakeClass}- and \texttt{TTree::MakeSelector}-based codes. Since Run-2, ATLAS has seen a increasing level of homogenization in analysis codes, where many groups use one of two high-level analysis frameworks within which they develop the custom routines needed for the analysis at hand.
Once all data (real and simulated) is sufficiently reduced, usually a statistical analysis is performed, in which the observed data is compared to the expectations given by the physics model under study. Here, a range of statistics packages such as HistFitter and HistFactory or loosely-structured scripts, that utilize RooFit/RooStats directly, is used to extract the relevant physics results such as interval estimates on model parameters. This can include precision measurements of Standard Model observables or exclusion limits on parameters of models of physics beyond the Standard Model.
\subsection{Analysis preservation for re-use}
In the context of analysis preservation, the entire analysis can viewed as an abstract function that maps data and the model hypothesis to the analysis results:
\begin{equation}
\textrm{result} = f_{\textrm{analysis}}(\textrm{data}, \textrm{model})
\end{equation}
Ideally, one would like to preserve this map in a completely parametrized form, $f_\mathrm{analysis}(\cdot,\cdot)$ independent of the specific data and model on which it has been applied to obtain the result at hand. Realistically, however, the analysis is tightly coupled to the recorded data it was developed against, due to various reasons such as file formats and re-processing versions. The model-dependence, on the other hand, can often times be factored out more easily, especially for analyses that search for physics phenomena beyond the Standard Model, where the Beyond the Standard Model (BSM) contribution is estimated separately from the Standard Model backgrounds. An analysis preservation approach that is designed to be model-independent would thus enable both re-interpretation and statistical combinations of multiple analyses after the initial publication.
In order for such to achieve such a parametrized preservation, two separate types of information need to be captured:
\begin{enumerate}
\item a descriptions of the individual parametrized analysis steps such as event selection steps or the subsequent statistical analysis
\item a description of the workflow that logically links these individual steps in order to arrive at the analysis result data
\end{enumerate}
In this document we introduce schemas to capture this information in flat JSON data as well as a framework to read back that information and re-execute such a preserved analysis.
\section{Capturing parametrized activities}
An appropriate model to capture the different steps of an analysis is the data model employed by the W3C PROV standard\cite{w3c-prov-dm}, in which the basic ingredients are \emph{entities} and \emph{activities} to track data provenance. Activities act on existing entities and generate new ones. In the context of an HEP analysis, an entity is often a set of files (such as a dataset) or a data product derived from them, while an activity is most often the execution of a piece of software that takes entities (i.e. data) as input and generates new entities as outputs, e.g. by writing a new set of files. These operations can be parametrized by a few variables such that the activity appears as a function of the parameters $p_1,p_2,\cdots,p_n$ and some notion of an input state $\sigma$ (this could for example be a filesystem directory), which may be modified as a side-effect of the function.
\begin{equation}
\mathrm{output} = f_{\mathrm{step}}(p_1,p_2,\cdots,p_n, \sigma),
\label{eq:singlestep}
\end{equation}
It is useful to partition the return value of this function into a tuple of an output state after processing $\sigma'$ and a separate record of human- and machine-readable \emph{result data}, $\mathbf{r}$, that provides additional machine-readable data, possibly describing the side effects such as filesystem paths of files generated during this step. This separation allows for a convenient definition of workflows later on, as each step identifies and publishes the relevant data fragments (i.e. entities) it produces.
\begin{equation}
(\mathbf{r},\sigma') = f_{\mathrm{step}}(\mathbf{p}, \sigma),
\label{eq:singlestep_v2}
\end{equation}
An activity is thus an abstract interface that transforms parameters into result data while modifying an external state. As an interchange format for both the input parameters $\mathbf{p}$ and result data $\mathbf{r}$ JSON is a suitable choice.
The information required to fully capture such parametrized activities may be partitioned into three basic pieces:
\begin{description}
\item[process] a parametrized description with which one can produce a fully-defined activity description -- the (``job'') -- based on concrete parameters (such as a templated command line)
\item[\emph{environment}] a description of the environment in which this job is to be executed. This may for example include a description of the necessary software to run the above process
\item[publisher] a description of how to extract the relevant information or data fragments subsequent to the execution of the job
\end{description}.
For each of these pieces, multiple concrete implementations are possible. Irrespective of the implementation, the basic procedure for execution (given some execution backend) is shown in algorithm~\ref{alg:packtivity}.
\\
\begin{algorithm}[H]
\footnotesize
\KwIn{$\mathbf{p}$, $\sigma$}
\KwOut{$\mathbf{r}$, $\sigma'$}
\SetKwFunction{Proc}{Process}
\SetKwFunction{Bknd}{Backend}
\SetKwFunction{Pblsh}{Publisher}
\Begin{
job $\leftarrow$ \Proc{$\mathbf{p}$}\;
$\sigma' \leftarrow$ \Bknd{$\sigma'$,job,environment}\;
$\mathbf{r} \leftarrow$ \Pblsh{job,$\sigma'$}\;
\Return{$\mathbf{r}$},$\sigma'$
}
\caption{Activity($\mathbf{p}$,$\sigma$)}
\label{alg:packtivity}
\end{algorithm}
\subsection{Packtivity}
To capture such `packaged activities` -- or `packtivities' -- a extensible set of JSON schemas have been developed to describe the three interfaces -- process, environment, publisher -- identified in the previous section. The choice is motivated by the simplicity and ubiquity of the JSON format, which makes it suitable for long-term and implementation-independent archival. Sub-schemas will generally be identified by a interface-wide `property` but implementation specific `property value`. An examples of full packtivity definitions are provided in listing~\ref{mnt:fullpacktivity}. A number of JSON schemas are collected under the \verb+yadage-schemas+ package available at GitHub and via PyPI\cite{yadage-schemas}
\begin{listing}[!ht]
{\footnotesize
\begin{Verbatim}[commandchars=\\\{\}]
\PYG{l+lScalar+lScalarPlain}{process}\PYG{p+pIndicator}{:}
\PYG{l+lScalar+lScalarPlain}{process\PYGZus{}type}\PYG{p+pIndicator}{:} \PYG{l+s}{\PYGZsq{}string\PYGZhy{}interpolated\PYGZhy{}cmd\PYGZsq{}}
\PYG{l+lScalar+lScalarPlain}{cmd}\PYG{p+pIndicator}{:} \PYG{l+s}{\PYGZsq{}DelphesHepMC}\PYG{n+nv}{ }\PYG{l+s}{\PYGZob{}delphes\PYGZus{}card\PYGZcb{}}\PYG{n+nv}{ }\PYG{l+s}{\PYGZob{}outputroot\PYGZcb{}}\PYG{n+nv}{ }\PYG{l+s}{\PYGZob{}inputhepmc\PYGZcb{}}\PYG{n+nv}{ }\PYG{l+s}{\PYGZam{}\PYGZam{}}\PYG{n+nv}{ }\PYG{l+s}{root2lhco}\PYG{n+nv}{ }\PYG{l+s}{\PYGZob{}outputroot\PYGZcb{}}\PYG{n+nv}{ }\PYG{l+s}{\PYGZob{}outputlhco\PYGZcb{}\PYGZsq{}}
\PYG{l+lScalar+lScalarPlain}{publisher}\PYG{p+pIndicator}{:}
\PYG{l+lScalar+lScalarPlain}{publisher\PYGZus{}type}\PYG{p+pIndicator}{:} \PYG{l+s}{\PYGZsq{}frompar\PYGZhy{}pub\PYGZsq{}}
\PYG{l+lScalar+lScalarPlain}{outputmap}\PYG{p+pIndicator}{:}
\PYG{l+lScalar+lScalarPlain}{lhcofile}\PYG{p+pIndicator}{:} \PYG{l+lScalar+lScalarPlain}{outputlhco}
\PYG{l+lScalar+lScalarPlain}{rootfile}\PYG{p+pIndicator}{:} \PYG{l+lScalar+lScalarPlain}{outputroot}
\PYG{l+lScalar+lScalarPlain}{environment}\PYG{p+pIndicator}{:}
\PYG{l+lScalar+lScalarPlain}{environment\PYGZus{}type}\PYG{p+pIndicator}{:} \PYG{l+s}{\PYGZsq{}docker\PYGZhy{}encapsulated\PYGZsq{}}
\PYG{l+lScalar+lScalarPlain}{image}\PYG{p+pIndicator}{:} \PYG{l+lScalar+lScalarPlain}{lukasheinrich/root\PYGZhy{}delphes}
\end{Verbatim}
}
\caption{An example packtivity manifest}
\label{mnt:fullpacktivity}
\end{listing}
\subsubsection{Process descriptions}
A form of capturing parametrized process information, that is accessible and convenient for analysis teams to produce, are templated string of multi-line scripts or single-line command lines from which concrete job manifests are formed by interpolating these strings based on input parameters provided as JSON documents. In the example shown in listing~\ref{mnt:fullpacktivity}, the process has five replacement fields that need to be provided by the input JSON document, for example by having a top-level object with appropriately named properties.
\subsubsection{Environment descriptions}
Describing the software environment or run-time that the job formed by the `process' requires is a trade-off between completeness and convenience and could range from specifying merely a specific software release number to a full virtual machine image that includes both hardware and software virtualization. In practice, using Linux container technologies such as Docker have proven to be a useful middle-ground.
In HEP contexts, a large amount of software is installed centrally in the global, read-only filesystem CVMFS, thus it may not be feasible for all application to provide a standalone CVMFS-independent installation. However, as CVMFS is a versioned filesystem, in principle it is possible to mount it at the version that the original analysis executed against. Similarly, some applications may require additional run-time data such as secrets used for VO authentication. Generally, the number of such external dependencies should be kept at a minimum for preservation purposes.
\subsubsection{Publisher descriptions}
The execution of the process in a given environment typically results in the modification of an external state which may, for example, be provided by mounting an external shared filesystem such as CephFS into the containers. However in order to describe multi-step analysis workflows, it is necessary to have a semantic description of what the relevant data fragments of this activity are. As laid out in the previous section, JSON is a suitable format. The processes themselves do not necessarily produce JSON data, so that it is helpful to include the notion of an external `publishing manifest' into the packtivity definition, that a implementation can use in order to to derive a JSON object.
Often, the result JSON data can be derived by simple inspection of the input parameters such as in listing~\ref{mnt:fullpacktivity}. In this case it suffices to provide a mapping from input parameters to output keys. In other scenarios the result data may only be fully formed after the process execution, for example if the process generates a dynamic number of files that all need to be published. In this case other publishers may be defined, such as using glob patterns or regular expressions.
\subsection{Reference implementation}
The python-based \verb+packtivity+\cite{packtivity} package provides an implementation of algorithm~\ref{alg:packtivity} for a number of different backends. Both synchronous and asynchronous backends such as Celery or IPython Clusters are provided for the currently defined process-types and environments. Both python language bindings as well as command line executables are provided to run packtivities.
\section{Parametrized workflow model}
The packaging of the individual processing activities captures a lot of the information needed to re-produce or re-use a given analysis. The command line interface or the language bindings can be used in code to execute pre-packaged activities in a suitable order. It is, however, desirable to capture this workflow logic in a similarly declarative fashion as the packtivities themselves, such that its execution may be automated.
A suitable data model for the description of workflows is the directed acyclic graph (DAG). In such a graph, nodes represent individual activities, while directed edges denote dependency relations of activities. This allows to capture non-linear workflows and enables the distribution and orchestration of analysis workflows across distributed systems.
As noted, for re-use applications, it is important that the workflows are parametrized. This may in turn introduce some parameter-dependency on not only the parameters for individual packtivities, but also the topology of the workflow graph. A number of DAG-based workflow systems such as DAGMan or the Common Workflow Language \cite{HTCondorDAGMan,Amstutz2016} exist, however those tools introduce limitations that hinder the definition of workflow in those parameter-dependent scenarios. Therefore, a number an extensible workflow definition system has been developed that grants first-class status to parametrized DAGs.
\subsection{Workflow stages}
The central tenet of parametrized and dynamic workflows is that, instead of archiving DAG descriptions that fully fix the topology, one should rather store sufficient information into a \emph{workflow template} $T$, from which it is possible construct these DAGs -- \emph{workflow instances} $W$ -- during run-time once sufficient information (such as parameter values) is available. As such, the workflow template T is made up from a set of \emph{workflow stages} $s_i$ and $T = \{s_1,s_2\dots s_n\}$. A stage represents a piece logic that add nodes and edges to the instance DAG.
The time at which this operation may be applied can be dependent on the state of the instance. For example it may require that some node within the instance graph has already been processed by a backend and its result data is known. Therefore a stage is defined by two pieces of information
\begin{itemize}
\item a definition of dependencies which the workflow instance $W$ must fulfill for the stage to be applicable. This is conveniently expressed as a DAG-valued \emph{predicate function} $d: W\to \{\mathrm{true},\mathrm{false}\}$
\item a DAG-valued function that, given a workflow instance, returns an updated workflow instance with additional nodes and edges or newly defined stages. This may be expressed as a DAG-valued \emph{scheduler function} $f: W\to W'$
\end{itemize}
In order to process a parametrized workflow defined by such a template, a simple algorithm, shown in algorithm~\ref{alg:adage} may be followed in order to continuously monitor a workflow instance. The algorithm applies the stages' scheduler functions as soon as its predicate is fulfilled and submit individual nodes to a packtivity backend.
\\
\begin{algorithm}[H]
\footnotesize
\KwIn{Workflow Template, Initialization JSON}
\KwOut{Complete set of Workflow data entities}
initialize new workflow instance (empty DAG) wflow;
\While{there exist unapplied stages}{
\ForEach{stage in unapplied stages}{
\If{stage applicable (predicate returns True)}{
expand DAG with stage as per scheduler function
}
}
\ForEach{node in DAG}{
\If{node has not been submitted to backend and all incoming edges succeeded}{
submit node to backend
}
}
}
wait until all nodes processed by backend
\caption{Basic Yadage Workflow Engine}
\label{alg:adage}
\end{algorithm}
\subsection{Yadage}
\subsubsection{Workflow Definition}
As for packtivities, the \verb+yadage-schemas+ package includes JSON schemas to define workflow stages. From experience complex workflows may be defined using a small number of stage-types. Currently stages that schedule one or more nodes of a singled packtivity (each with different parameters) are defined. In the case of multi-node stages, a number of scattering patterns maybe used. An example stage definition is shown in listing~\ref{mnt:stage}. The dependencies are listed by naming other stages. The current stage is then considered applicable if all nodes by the dependent stage are successfully processed. The scheduler function $f$, is defined under the \verb+scheduler+ property and includes instructions on how to access the result data of dependent nodes added to the Graph by dependent stages in order to define the parameters of the packtivity to be scheduled. As seen in the listing, JSON References are used in order to reference packtivity definitions via URIs.
\begin{listing}[!ht]
{\footnotesize
\begin{Verbatim}[commandchars=\\\{\}]
\PYG{l+lScalar+lScalarPlain}{name}\PYG{p+pIndicator}{:} \PYG{l+lScalar+lScalarPlain}{delphes}
\PYG{l+lScalar+lScalarPlain}{dependencies}\PYG{p+pIndicator}{:} \PYG{p+pIndicator}{[}\PYG{n+nv}{pythia}\PYG{p+pIndicator}{]}
\PYG{l+lScalar+lScalarPlain}{scheduler}\PYG{p+pIndicator}{:}
\PYG{l+lScalar+lScalarPlain}{scheduler\PYGZus{}type}\PYG{p+pIndicator}{:} \PYG{l+s}{\PYGZsq{}singlestep\PYGZhy{}stage\PYGZsq{}}
\PYG{l+lScalar+lScalarPlain}{step}\PYG{p+pIndicator}{:} \PYG{p+pIndicator}{\PYGZob{}}\PYG{n+nv}{\PYGZdl{}ref}\PYG{p+pIndicator}{:} \PYG{l+s}{\PYGZsq{}delphes.yml\PYGZsq{}}\PYG{p+pIndicator}{\PYGZcb{}}
\PYG{l+lScalar+lScalarPlain}{parameters}\PYG{p+pIndicator}{:}
\PYG{l+lScalar+lScalarPlain}{outputroot}\PYG{p+pIndicator}{:} \PYG{l+s}{\PYGZsq{}\PYGZob{}workdir\PYGZcb{}/output.root\PYGZsq{}}
\PYG{l+lScalar+lScalarPlain}{outputlhco}\PYG{p+pIndicator}{:} \PYG{l+s}{\PYGZsq{}\PYGZob{}workdir\PYGZcb{}/output.lhco\PYGZsq{}}
\PYG{l+lScalar+lScalarPlain}{delphes\PYGZus{}card}\PYG{p+pIndicator}{:} \PYG{l+s}{\PYGZsq{}delphes/cards/delphes\PYGZus{}card\PYGZus{}ATLAS.tcl\PYGZsq{}}
\PYG{l+lScalar+lScalarPlain}{inputhepmc}\PYG{p+pIndicator}{:} \PYG{p+pIndicator}{\PYGZob{}}\PYG{n+nv}{stages}\PYG{p+pIndicator}{:} \PYG{n+nv}{pythia}\PYG{p+pIndicator}{,} \PYG{n+nv}{output}\PYG{p+pIndicator}{:} \PYG{n+nv}{hepmcfile}\PYG{p+pIndicator}{\PYGZcb{}}
\end{Verbatim}
}
\caption{An Example Yadage Stage Manifest}
\label{mnt:stage}
\end{listing}
\subsection{Workflow composition}
An important feature of workflows defined via the yadage schemas is composability that is not dependent on coordination between workflow authors. HEP workflows can involve many different stages in order to transform real or simulated events from detector or even particle-level data all the way to a final analysis result. Parts of these workflows may be primarily defined by different groups. For example, the workflow to describe the generation of Monte Carlo events based on a certain model may be defined primarily by physics working groups, while the reconstruction chain to transform generated events into fully reconstructed events (`AOD' data) usually is the responsibility of a central reconstruction group. Finally the downstream analysis of those reconstructed events is done on the analysis-team level. Each of these macro-parts of the workflow may be a multi-stage workflow themselves, such that the ability to compose them without modification into a larger workflow is desirable. In the case of the currently defined yadage stages, this is easily achieved by modifying the stage schedule 'workflows' instead of individual packtivities. In this case, the scheduling function does not add nodes or edges, but rather adds newly defined stages corresponding to the sub-workflows to the instances list of stages, that subsequently will be applied in a scoped fashion to exclude the possibility of e.g. naming collisions/ambiguities.
\subsubsection{Reference implementation}
Workflows declared using the above schemas may be executed using the \verb+yadage+ package\cite{yadage} that implements the basic scheduling and submission algorithm outlined above and transparently is able to use any packtivity backend implementation. Further, it implements a number of convenience features such as caching/memoization of individual packtivity results and visualization capabilities.
\section{Applications}
\subsection{Run-I reinterpretation}
Workflows defined via the schemas outlined above have been used in a number of different contexts. ATLAS has published multiple re-interpretations of analyses prepared in Run-1. In these campaigns, a number of analyses designed to investigate particular supersymmetric scenarios have been re-interpreted to derive a more comprehensive assessment of the ATLAS experiments' sensitivity to supersymmetry such as under the 19-dimensional `phenomenological MSSM' (pMSSM) or a more restricted five-dimensional scan targeting electroweak sparticle production\cite{Aad:2015baa,Aaboud:2016wna,ATLAS-CONF-2016-033}.
\subsection{CERN Analysis Preservation Portal}
The CERN Analysis Preservation Portal (CAP)\cite{Chen2016} aims to preserve analysis information in the form of a digital library. Besides assembling metadata of analyses, such as the involved researchers and institutes, it also seeks to archive more technical information, such as code repositories and software environments. The workflow and activity schemas described here have been deeply integrated into the CAP system. Thanks to the choice of JSON schemas, the can be treated as native data in the context of the Invenio Digitial Library framework\cite{Invenio} facilitating, for example, discoverability and composability of workflow pieces.
\subsection{RECAST}
As noted, LHC experiments already engage in re-interpretation campaigns. However, the current approach requires a high level of coordination between analysis groups for each re-interpretation. RECAST is a framework in which the re-interpretation is streamlined by utilizing workflows that are archived in a re-usable manner. Originally proposed in 2010, a prototype backend implementation has been recently been developed and deployed at CERN. RECAST allows interested parties outside of the LHC collaboration to suggest for re-interpretation, by providing model information such as parameter cards in the SLHA format to the experiments using a web-based interface or, alternatively, REST and Python APIs. Upon review by the experiments, the experiment may decide to re-run an archived analysis based on these new model inputs and provide a response in the form of likelihood information (e.g. $CL_s$ values)
The prototype backend consists of a `control-center' web-service that is accessible using VO-filtered CERN Single-Sign-On. This web-service displays incoming requests and allows operators to launch a re-execution of an analysis based on the RECAST request. The actual workflow execution is then handed off to a distributed system. Here, yadage and packtivity are heavily used to both define workflows and drive their execution. In the course of this development, a packtivity backend has been implemented that allows the scheduling of HEP container workloads on a Kubernetes Cluster deployed on the CERN OpenStack infrastructure using OpenStack Magnum. The integration with cloud-native tools allows for convenient scaling characteristics and a highly distributed workflow execution that may be monitored in real-time using the web-interface of the `control-center'.
\section{Summary}
We have presented a framework to define parametrized workflows in order to preserve high-energy physics analyses in a format that allows collaboration members to re-execute the original analysis in the context of new physics models. The framework defines a set of portable JSON schemas that describe both the individual processing steps and and workflow logic to orchestrate multiple steps. We also presented language-agnostic algorithms to re-execute analyses based on these descriptions. Furthermore, reference implementations of these algorithms distributed as python-based packages (\verb+packtivity+ and \verb+yadage+) were presented. Workflows definitions and the reference implementation have been used in the past for re-interpretation campaigns within the ATLAS collaboration and are deeply integrated in the CERN Analysis Preservation Portal and RECAST projects.
\section{Acknowledgements}
Cranmer and Heinrich are both supported through NSF ACI-1450310, additionally Cranmer is supported by PHY-1505463.
\section*{References}
| 6b15a5688286e994067043120098c85af8ee41b2 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Hyperparameter values---regularization strength, model family choices like depth of a neural network or which nonlinear functions to use, procedural elements like dropout rates, stochastic gradient descent step sizes, and data preprocessing choices---can make the difference between a successful application of machine learning and a wasted effort. To search among many hyperparameter values requires repeated execution of often-expensive learning algorithms, creating a major obstacle for practitioners and researchers alike.
In general, on iteration (evaluation) $k$, a hyperparameter searcher suggests a $d$-dimensional hyperparameter configuration $x_k \in \mathcal{X}$ (e.g., $\mathcal{X}=\mathbb{R}^d$ but could also include discrete dimensions), a worker trains a model using $x_k$, and returns a validation loss of $y_k \in \mathbb{R}$ computed on a hold out set.
In this work we say a hyperparameter searcher is \textbf{open loop} if $x_k$ depends only on $\{x_{i}\}_{i=1}^{k-1}$; examples include choosing $x_k$ uniformly at random \citep{randombergstra}, or $x_k$ coming from a low-discrepancy sequence (c.f., \cite{iaco2015low}).
We say a searcher is \textbf{closed loop} if $x_k$ depends on both the past configurations and validation losses $\{ (x_i,y_i) \}_{i=1}^{k-1}$; examples include Bayesian optimization \citep{practical_bo} and reinforcement learning methods \citep{zoph2016neural}.
Note that open loop methods can draw an infinite sequence of configurations before training a single model, whereas closed loop methods rely on validation loss feedback in order to make suggestions.
While sophisticated closed loop selection methods have been shown to empirically identify good hyperparameter configurations faster (i.e., with fewer iterations) than open loop methods like random search, \textbf{two trends have rekindled interest in embarrassingly parallel open loop methods}: 1) modern deep learning model are taking longer to train, sometimes up to days or weeks, and 2) the rise of cloud resources available to anyone that charge not by the number of machines, but by the number of CPU-hours used so that 10 machines for 100 hours costs the same as 1000 machines for 1 hour.
This paper explores the landscape of open loop methods, identifying tradeoffs that are rarely considered, if at all acknowledged.
While random search is arguably the most popular open loop method and chooses each $x_k$ independently of $\{x_{i}\}_{i=1}^{k-1}$, it is by no means the only choice.
In many ways uniform random search is the least interesting of the methods we will discuss because we will advocate for methods where $x_k$ depends on $\{x_i\}_{i=1}^{k-1}$ to promote \textbf{diversity}.
In particular, we will focus on drawing $\{x_i\}_{i=1}^k$ from a \textbf{$k$-determinantal point process (DPP)} \citep{dpps_for_ml}. We introduce a sampling algorithm which allows
DPPs to support real, integer, and categorical dimensions, any of which may have a tree structure, and we describe connections between DPPs and Gaussian processes (GPs).
In synthetic experiments, we find our diversity-promoting open-loop method outperforms other open loop methods. In practical hyperparameter optimization experiments, we find that it significantly outperforms other approaches in cases where the hyperparameter values have a large effect on performance. Finally, we compare against a closed loop Bayesian optimization method, and find that sequential Bayesian optimization takes, on average, more than ten times as long to find a good result, for a gain of only 0.15 percent accuracy on a particular hyperparameter optimization task.
\section{Related Work}
While this work focuses on open loop methods, the vast majority of recent work on hyperparameter tuning has been on closed loop methods, which we briefly review.
\subsection{Closed Loop Methods}
Much attention has been paid to sequential model-based optimization techniques such as Bayesian optimization \citep{bo_algos, practical_bo} which sample hyperparameter spaces adaptively. These techniques first choose a point in the space of hyperparameters, then train and evaluate a model with the hyperparameter values represented by that point, then sample another point based on how well previous point(s) performed. When evaluations are fast, inexpensive, and it's possible to evaluate a large number of points (e.g. $k = \Omega(2^d)$ for $d$ hyperparameters) these approaches can be advantageous, but in the more common scenario where we have limited time or a limited evaluation budget, the sequential nature of closed loop methods can be cumbersome. In addition, it has been observed that many Bayesian optimization methods with a moderate number of hyperparameters, when run for $k$ iterations, can be outperformed by sampling $2k$ points uniformly at random \citep{hyperband}, indicating that even simple open loop methods can be competitive.
Parallelizing Bayesian optimization methods has proven to be nontrivial, though many agree that it's vitally important. While many algorithms exist which can sample more than one point at each iteration \citep{parallel_bo_ucb,parallel_tradeoffs,batch_bo, parallel_thompson_bo}, the sequential nature of Bayesian optimization methods prevent the full parallelization open loop methods can employ. Even running two iterations (with batches of size $k/2$) will take on average twice as long as fully parallelizing the evaluations, as you can do with open loop methods like grid search, sampling uniformly, or sampling according to a DPP.
One line of research has examined the use of $k$-DPPs for optimizing hyperparameters in the context of parallelizing Bayesian optimization \citep{bo_dpp,bo_dpp_struct}. At each iteration within one trial of Bayesian optimization, instead of drawing a single new point to evaluate from the posterior, they define a $k$-DPP over a relevance region from which they sample a diverse set of points. They found their approach to beat state-of-the-art performance on a number of hyperparameter optimization tasks, and they proved that generating batches by sampling from a $k$-DPP has better regret bounds than a number of other approaches.
They show that a previous batch sampling approach which selects a batch by sequentially choosing a point which has the highest posterior variance \citep{parallel_bo_ucb} is just approximating finding the maximum probability set from a $k$-DPP (an NP-hard problem \citep{dpps_for_ml}), and they prove that sampling (as opposed to maximization) has better regret bounds for this optimization task.
We use the work of \citet{bo_dpp} as a foundation for our exploration of fully-parallel optimization methods, and thus we focus on $k$-DPP sampling as opposed to maximization.
So-called configuration evaluation methods have been shown to perform well by adaptively allocating resources to different hyperparameter settings \citep{freeze_thaw,hyperband}. They initially choose a set of hyperparameters to evaluate (often uniformly), then partially train a set of models for these hyperparameters. After some fixed training budget (e.g., time, or number of training examples observed), they compare the partially trained models against one another and allocate more resources to those which perform best. Eventually, these algorithms produce one (or a small number) of fully trained, high-quality models.
In some sense, these approaches are orthogonal to open vs.~closed loop methods, as the diversity-promoting approach we advocate can be used as a drop-in replacement to the method used to choose the initial hyperparameter assignments.
\subsection{Sampling proportional to the posterior variance of a Gaussian process}\label{sec:gaussian_process}
GPs have long been lauded for their expressive power, and have been used extensively in the hyperparameter optimization literature. \citet{gp_dpp} show that drawing a sample from a $k$-DPP with kernel $\mathcal{K}$ is equivalent to sequentially sampling $k$ times proportional to the (updated) posterior variance of a GP defined with covariance kernel $\mathcal{K}$. This sequential sampling is one of the oldest hyperparameter optimization algorithms, though our work is the first to perform an in-depth analysis. Additionally, this has a nice information theoretic justification: since the entropy of a Gaussian is proportional to the log determinant of the covariance matrix, points drawn from a DPP have probability proportional to $\exp(\text{information gain})$, and the most probable set from the DPP is the set which maximizes the information gain. With our MCMC algorithm presented in Algorithm \ref{alg:mixed_mcmc}, we can draw samples with these appealing properties from any space for which we can draw uniform samples. The ability to draw $k$-DPP samples by sequentially sampling points proportional to the posterior variance grants us another boon: if one has a sample of size $k$ and wants a sample of size $k+1$, only a single additional point needs to be drawn, unlike with the sampling algorithms presented in \citet{dpps_for_ml}. Using this approach, we can draw samples up to $k=100$ in less than a second on a machine with 32 cores.
\subsection{Open Loop Methods}
As discussed above, recent trends have renewed interest in open loop methods. While there exist many different batch BO algorithms, analyzing these in the open loop regime (when there are no results from function evaluations) is often rather simple. As there is no information with which to update the posterior mean, function evaluations are hallucinated using the prior or points are drawn only using information about the posterior variance. For example, in the open loop regime, \citet{parallel_thompson_bo}'s approach without hallucinated observations is equivalent to uniform sampling, and their approach with hallucinated observations (where they use the prior mean in place of a function evaluation, then update the posterior mean and variance) is equivalent to sequentially sampling according to the posterior variance (which is the same as sampling from a DPP). Similarly, open loop optimization in SMAC \citep{parallel_smac} is equivalent to first Latin hypercube sampling to make a large set of diverse candidate points, then sampling $k$ uniformly among these points.
Recently, uniform sampling was shown to be competitive with sophisticated closed loop methods for modern hyperparameter optimization tasks like optimizing the hyperparameters of deep neural networks \citep{hyperband}, inspiring other works to explain the phenomenon \citep{ahmed2016we}.
\citet{randombergstra} offer one of the most comprehensive studies of open loop methods to date, and focus attention on comparing random search and grid search.
A main takeaway of the paper is that uniform random sampling is generally preferred to grid search\footnote{Grid search uniformly grids $[0,1]^d$ such that $x_k = (\tfrac{i_1}{m}, \tfrac{i_2}{m}, \dots, \tfrac{i_d}{m})$ is a point on the grid for $i_j = 0,1,\dots,m$ for all $j$, with a total number of grid points equal to $(m+1)^d$.} due to the frequent observation that some hyperparameters have little impact on performance, and random search promotes more diversity in the dimensions that matter.
Essentially, if points are drawn uniformly at random in $d$ dimensions but only $d' < d$ dimensions are relevant, those same points are uniformly distributed (and just as diverse) in $d'$ dimensions.
Grid search, on the other hand, distributes configurations aligned with the axes so if only $d'<d$ dimensions are relevant, many configurations are essentially duplicates.
However, grid search does have one favorable property that is clear in just one dimension.
If $k$ points are distributed on $[0,1]$ on a grid, the maximum spacing between points is equal to $\frac{1}{k-1}$.
But if points are uniformly at random drawn on $[0,1]$, the expected largest gap between points scales as $\frac{1}{\sqrt{k}}$.
If, by bad luck, the optimum islocated in this largest gap, this difference could be considerable; we attempt to quantify this idea in the next section.
\section{Measures of spread}
Quantifying the spread of a sequence $\mathbf{x} = (x_1,x_2,\dots,x_k)$ (or, similarly, how well $\mathbf{x}$ covers a space) is a well-studied concept. In this section we introduce discrepancy, a quantity used by previous work, and dispersion, which we argue is more appropriate for optimization problems.
\subsection{Discrepancy}
Perhaps the most popular way to quantify the spread of a sequence is star discrepancy. One can interpret the star discrepancy as a multidimensional version of the Kolmogorov-Smirnov statistic between the sequence $\mathbf{x}$ and the uniform measure; intuitively, when $\mathbf{x}$ contains points which are spread apart, star discrepancy is small.
We include a formal definition in Appendix~\ref{sec:star_discrep}.
Star discrepancy plays a prominent role in the numerical integration literature, as it provides a sharp bound on the numerical integration error through the the Koksma-Hlawka inequality (given in Appendix~\ref{sec:koksma-hlawka}) \citep{koksma-hlawka}. This has led to wide adoption of low discrepancy sequences, even outside of numerical integration problems. For example, \citet{randombergstra} analyzed a number of low discrepancy sequences for some optimization tasks and found improved optimization performance over uniform sampling and grid search. Additionally, low discrepancy sequences such as the Sobol sequence\footnote{\citet{randombergstra} found that the Niederreiter and Halton sequences performed similarly to the Sobol sequence, and that the Sobol sequence outperformed Latin hypercube sampling. Thus, our experiments include the Sobol sequence (with the Cranley-Patterson rotation) as a representative low-discrepancy sequence.} are used as an initialization procedure for some Bayesian optimization schemes \citep{practical_bo}.
\subsection{Dispersion}\label{sec:dispersion}
Previous work on open loop hyperparameter optimization focused on low discrepancy sequences \citep{randombergstra, norandomnocry}, but optimization performance---how close a point in our sequence is to the true, fixed optimum---is our goal, not a sequence with low discrepancy. As discrepancy doesn't directly bound optimization error, we turn instead to dispersion
\begin{align*}
d_k(\mathbf{x}) = \sup_{x\in [0,1]^d} \min_{1\leq i \leq k} \rho (x, x_i),
\end{align*}
where $\rho$ is a distance (in our experiments $L_2$ distance). Intuitively, the dispersion of a point set is the radius of the largest Euclidean ball containing no points; dispersion measures the worst a point set could be at finding the optimum of a space.
Following \cite{niederreiter1992random}, we can bound the optimization error as follows. Let $f$ be the function we are aiming to optimize (maximize) with domain $\mathcal{B}$, $m(f)=\sup\limits_{x\in \mathcal{B}} f(x)$ be the global optimum of the function, and $m_k(f;\mathbf{x})=\sup\limits_{1\leq i\leq k} f(x_i)$ be the best-found optimum from the set $\mathbf{x}$. Assuming $f$ is continuous (at least near the global optimum), the modulus of continuity is defined as
\begin{align*}
\omega(f; t) = \sup_{\substack{x,y\in \mathcal{B} \\ \rho(x,y) \leq t}} \left|f(x) - f(y)\right|,\text{ for some } t\geq 0 .
\end{align*}
\begin{theorem}\label{theorem:opt_bound}
\citep{niederreiter1992random}
For any point set $\mathbf{x}$ with dispersion $d_k(\mathbf{x})$, the optimization error is bounded as
\begin{align*}
m(f) - m_k(f;\mathbf{x}) \leq \omega(f; d_k(\mathbf{x})).
\end{align*}
\end{theorem}
Dispersion can be computed efficiently (unlike discrepancy, $D_k(\mathbf{x})$, which is NP-hard \citep{zhigljavsky2007stochastic}), and we give an algorithm in Appendix~\ref{sec:dispersion_algo}.
Dispersion is at least $\Omega(k^{-1/d})$, and while low discrepancy implies low dispersion ($d^{-1/2}d_k(\mathbf{x}) \leq \frac{1}{2}D_k(\mathbf{x})^{1/d}$), the other direction does not hold.\footnote{Discrepancy is a global measure which depends on all points, while dispersion only depends on points near the largest ``hole''.}
Therefore we know that the low-discrepancy sequences evaluated in previous work are also low-dispersion sequences in the big-$O$ sense, but as we will see they may behave quite differently. Samples drawn uniformly are not low dispersion, as they have rate $(\ln(k)/k)^{1/d}$ \citep{zhigljavsky2007stochastic}.
Optimal dispersion in one dimension is found with an evenly spaced grid, but it's unknown how to get an optimal set in higher dimensions.\footnote{In two dimensions a hexagonal tiling finds the optimal dispersion, but this is only valid when $k$ is divisible by the number of columns and rows in the tiling.} Finding a set of points with the optimal dispersion is as hard as solving the circle packing problem in geometry with $k$ equal-sized circles which are as large as possible. Dispersion is bounded from below with $d_k(\mathbf{x})\geq \big( \Gamma(d/2+1)\big)^{1/d}\pi^{-1/2}k^{-1/d}$, though it is unknown if this bound is sharp.
\subsection{Comparison of Open Loop Methods}
In Figure~\ref{fig:dispersion} we plot the dispersion of the Sobol sequence, samples drawn uniformly at random, and samples drawn from a $k$-DPP, in one and two dimensions. To generate the $k$-DPP samples, we sequentially drew samples proportional to the (updated) posterior variance (using an RBF kernel, with $\sigma=\sqrt{2}/k$), as described in Section~\ref{sec:gaussian_process}. When $d=1$, the regular structure of the Sobol sequence causes it to have increasingly large plateaus, as there are many ``holes'' of the same size.\footnote{By construction, each individual dimension of the $d$-dimensional Sobol sequence has these same plateaus.} For example, the Sobol sequence has the same dispersion for $42\leq k \leq 61$, and $84 \leq k \leq 125$.
Samples drawn from a $k$-DPP appear to have the same asymptotic rate as the Sobol sequence, but they don't suffer from the plateaus. When $d=2$, the $k$-DPP samples have lower average dispersion and lower variance.
One other natural surrogate of average optimization performance is to measure the distance from a fixed point, say $\frac{1}{2} \mathbf{1} = (\frac{1}{2},\dots,\frac{1}{2})$ or from the origin, to the nearest point in the length $k$ sequence. Our experiments (in Appendix~\ref{sec:center_origin}) on these metrics show the $k$-DPP samples bias samples to the corners of the space, which can be beneficial when the practitioner defined the search space with bounds that are too small.
Note, the low-discrepancy sequences are usually defined only for the $[0,1]^d$ hypecrube, so for hyperparameter search which involves conditional hyperparameters (i.e. those with tree structure) they are not appropriate.
In what follows, we study the $k$-DPP in more depth and how it performs on real-world hyperparameter tuning problems.
\begin{figure*}[h]
\centering
\subfloat[Dispersion, with $d=1$. ]{\includegraphics[width=0.5\textwidth]{dispersion_1d.pdf}\label{fig:disp_1d}}
\hfill
\subfloat[Dispersion, with $d=2$.]{\includegraphics[width=0.5\textwidth]{dispersion_2d.pdf}\label{fig:disp_2d}}
\caption{Dispersion of the Sobol sequence, samples from a $k$-DPP, and uniform random samples (lower is better). These log-log plots show when $d=1$ that Sobol suffers from regular plateaus of increasing size, while when $d=2$ the $k$-DPP samples have lower average dispersion and lower variance.
\label{fig:dispersion}}
\end{figure*}
\section{Method}
We begin by reviewing DPPs and $k$-DPPs.
Let $\mathcal{B}$ be a domain from which we would like to sample a finite subset. (In our use of DPPs, this is the set of hyperparameter assignments.) In general, $\mathcal{B}$ could be discrete or continuous; here we assume it is discrete with $N$ values, and we define $\mathcal{Y}=\{1,\ldots,N\}$ to be a a set which indexes $\mathcal{B}$ (this index set will be particularly useful in Algorithm \ref{alg:discrete_mcmc}). In Section \ref{sec:new_mcmc} we address when $\mathcal{B}$ has continuous dimensions.
A DPP defines a probability distribution over $2^\mathcal{Y}$ (all subsets of $\mathcal{Y}$) with the property that two elements of $\mathcal{Y}$ are more (less) likely to both be chosen the more dissimilar (similar) they are. Let random variable $\boldsymbol{Y}$ range over finite subsets of $\mathcal{Y}$.
There are several ways to define the parameters of a DPP. We focus on $\mathbf{L}$-ensembles, which define the probability that a specific subset is drawn (i.e., $P(\boldsymbol{Y} = \mathcal{A})$ for some $\mathcal{A} \subset \mathcal{Y}$) as:
\begin{align}
P(\boldsymbol{Y} = \mathcal{A}) = \frac{\text{det}(\mathbf{L}_\mathcal{A})}{\text{det}(\mathbf{L}+I)}.
\end{align}
As shown in \cite{dpps_for_ml}, this definition of $\mathbf{L}$ admits a decomposition to terms representing the \emph{quality} and \emph{diversity} of the elements of $\mathcal{Y}$. For any $y_i, y_j \in \mathcal{Y}$, let:
\begin{align}
\mathbf{L}_{i,j}=q_iq_j\mathcal{K}(\boldsymbol{\phi}_i,\boldsymbol{\phi}_j),
\end{align}
where $q_i > 0$ is the quality of $y_i$, $\boldsymbol{\phi}_i \in \mathbb{R}^d$ is a featurized representation of $y_i$, and $\mathcal{K}: \mathbb{R}^d \times \mathbb{R}^d \to [0,1]$ is a similarity kernel (e.g. cosine distance). (We will discuss how to featurize hyperparameter settings in Section~\ref{se:construction}.)
Here, we fix all $q_i = 1$; in future work, closed loop methods might make use of $q_i$ to encode evidence about the quality of particular hyperparameter settings to adapt the DPP's distribution over time.
\subsection{Sampling from a k-DPP}
DPPs have support over all subsets of $\mathcal{Y}$, including $\emptyset$ and $\mathcal{Y}$ itself. In many practical settings, one may have a fixed budget that allows running the training algorithm $k$ times, so we require precisely $k$ elements of $\mathcal{Y}$ for evaluation. $k$-DPPs are distributions over subsets of $\mathcal{Y}$ of size $k$. Thus,
\begin{align}
P(\boldsymbol{Y}=\mathcal{A} \mid |\boldsymbol{Y}| = k)= \frac{\text{det}(\mathbf{L}_\mathcal{A})}{\sum_{\mathcal{A}^\prime \subset \mathcal{Y}, |\mathcal{A}'| = k }\text{det}(\mathbf{L}_{\mathcal{A}^\prime})}.
\end{align}
Sampling from $k$-DPPs has been well-studied. When the base set $\mathcal{B}$ is a set of discrete items, exact sampling algorithms are known which run in $\mathcal{O}(Nk^3)$ \cite{dpps_for_ml}. When the base set is a continuous hyperrectangle, a recent exact sampling algorithm was introduced, based on a connection with Gaussian processes (GPs), which runs in $\mathcal{O}(dk^2 + k^3)$ \cite{gp_dpp}. We are unaware of previous work which allows for sampling from $k$-DPPs defined over any other base sets.
\subsection{Sampling k-DPPs defined over arbitrary base sets} \label{sec:new_mcmc}
\citet{dpp_discrete_mcmc} present a Metropolis-Hastings algorithm (included here as Algorithm~\ref{alg:discrete_mcmc}) which is a simple and fast alternative to the exact sampling procedures described above.
However, it is restricted to discrete domains. We propose a generalization of the MCMC algorithm which preserves relevant computations while allowing sampling from any base set from which we can draw uniform samples, including those with discrete dimensions, continuous dimensions, some continuous and some discrete dimensions, or even (conditional) tree structures (Algorithm~\ref{alg:mixed_mcmc}). To the best of our knowledge, this is the first algorithm which allows for sampling from a $k$-DPP defined over any space other than strictly continuous or strictly discrete, and thus the first algorithm to utilize the expressive capabilities of the posterior variance of a GP in these regimes.
\algrenewcommand\algorithmicrequire{\textbf{Input:}}
\algrenewcommand\algorithmicensure{\textbf{Output:}}
\begin{algorithm*}[htb]
\caption{Drawing a sample from a discrete $k$-DPP \citep{dpp_discrete_mcmc} }
\label{alg:discrete_mcmc}
\begin{algorithmic}[1]
\Require{$\mathbf{L}$, a symmetric, $N\times N$ matrix where $\mathbf{L}_{i,j}=q_iq_j\mathcal{K}(\boldsymbol{\phi}_i,\boldsymbol{\phi}_j)$ which defines a DPP over a finite base set of items $\mathcal{B}$, and $\mathcal{Y}=\{1,\ldots,N\}$, where $\mathcal{Y}_i$ indexes a row or column of $\mathbf{L}$}
\Ensure{$\mathcal{B}_\mathbf{Y}$ (the points in $\mathcal{B}$ indexed by $\mathbf{Y}$)}
\State Initialize $\mathbf{Y}$ to $k$ elements sampled from $\mathcal{Y}$ uniformly
\While{not mixed}
\State uniformly sample $u\in \mathbf{Y}, v\in \mathcal{Y}\setminus \mathbf{Y}$
\State set $\mathbf{Y}^\prime=\mathbf{Y}\cup \{v\}\setminus \{u\}$
\State $p \gets \frac{1}{2}min(1,\frac{\text{det}(\mathbf{L}_{\mathbf{Y}^\prime})} {\text{det}(\mathbf{L}_{\mathbf{Y}})})$
\State with probability $p$: $\mathbf{Y} = \mathbf{Y}^\prime$
\EndWhile
\State Return $\mathcal{B}_\mathbf{Y}$
\end{algorithmic}
\end{algorithm*}
Algorithm~\ref{alg:discrete_mcmc} proceeds as follows: First, initialize a set $\mathbf{Y}$ with $k$ indices of $\mathbf{L}$, drawn uniformly. Then, at each iteration, sample two indices of $\mathbf{L}$ (one within and one outside of the set $\mathbf{Y}$), and with some probability replace the item in $\mathbf{Y}$ with the other.
When we have continuous dimensions in the base set, however, we can't define the matrix $\mathbf{L}$, so sampling indices from it is not possible. We propose Algorithm~\ref{alg:mixed_mcmc}, which samples points directly from the base set $\mathcal{B}$ instead (assuming continuous dimensions are bounded), and computes only the principal minors of $\mathbf{L}$ needed for the relevant computations on the fly.
\begin{algorithm*}[htb]
\caption{Drawing a sample from a $k$-DPP defined over a space with continuous and discrete dimensions}
\label{alg:mixed_mcmc}
\begin{algorithmic}[1]
\Require{A base set $\mathcal{B}$ with some continuous and some discrete dimensions, a quality function $\boldsymbol{\Psi}:\mathbf{Y}_i \to q_i$, a feature function $\boldsymbol{\Phi}:\mathbf{Y}_i \to \boldsymbol{\phi}_i$}
\Ensure{$\boldsymbol{\beta}$, a set of $k$ points in $\mathcal{B}$}
\State Initialize $\boldsymbol{\beta}$ to $k$ points sampled from $\mathcal{B}$ uniformly
\While{not mixed}
\State uniformly sample $u\in \boldsymbol{\beta}, v\in \mathcal{B}\setminus \boldsymbol{\beta}$
\State set $\boldsymbol{\beta}^\prime=\boldsymbol{\beta}\cup \{v\}\setminus \{u\}$
\State compute the quality score for each item, $q_i=\boldsymbol{\Psi}(\boldsymbol{\beta}_i),\forall i$, and $q_i^\prime = \boldsymbol{\Psi} (\boldsymbol{\beta}_i^\prime),\forall i$
\State construct $\mathbf{L}_{\boldsymbol{\beta}}=[ q_i q_j \mathcal{K}( \boldsymbol{\Phi} (\boldsymbol{\beta}_i), \boldsymbol{\Phi} (\boldsymbol{\beta}_j))], \forall i,j$
\State construct $\mathbf{L}_{\boldsymbol{\beta}^\prime}=[q_i^\prime q_j^\prime \mathcal{K}(\boldsymbol{\Phi}(\boldsymbol{\beta}_i^\prime), \boldsymbol{\Phi}(\boldsymbol{\beta}_j^\prime))], \forall i,j$
\State $p \gets \frac{1}{2}min(1,\frac{\text{det}(\mathbf{L}_{\boldsymbol{\beta}^\prime})} {\text{det}(\mathbf{L}_{\boldsymbol{\beta}})})$
\State with probability $p$: $\boldsymbol{\beta} = \boldsymbol{\beta}^\prime$
\EndWhile
\State Return $\boldsymbol{\beta}$
\end{algorithmic}
\end{algorithm*}
Even in the case where the dimensions of $\mathcal{B}$ are discrete, Algorithm \ref{alg:mixed_mcmc} requires less computation and space than Algorithm \ref{alg:discrete_mcmc} (assuming the quality and similarity scores are stored once computed, and retrieved when needed). Previous analyses claimed that Algorithm \ref{alg:discrete_mcmc} should mix after $\mathcal{O}(N\log(N))$ steps. There are $\mathcal{O}(N^2)$ computations required to compute the full matrix $L$, and at each iteration we will compute at most $O(k)$ new elements of $L$, so even in the worst case we will save space and computation whenever $k\log(N)<N$. In expectation, we will save significantly more.
\subsection{Constructing $\mathbf{L}$ for hyperparameter optimization} \label{se:construction}
Let $\boldsymbol{\phi}_i$ be a feature vector for $y_i \in \mathcal{Y}$, a modular encoding of the attribute-value mapping assigning values to different hyperparameters, in which fixed segments of the vector are assigned to each hyperparameter attribute (e.g., the dropout rate, the choice of nonlinearity, etc.).
For a hyperparameter that takes a numerical value in range $[h_{\text{min}}, h_{\text{max}}]$, we encode value $h$ using one dimension ($j$) of $\boldsymbol{\phi}$ and project into the range $[0,1]$:
\begin{align}
\mathbf{\boldsymbol{\phi}}[j] &= \frac{h - h_{\text{min}}}{h_{\text{max}} - h_{\text{min}}}
\end{align}
This rescaling prevents hyperparameters with greater dynamic range from dominating the similarity calculations. A categorical-valued hyperparameter variable that takes $m$ values is given $m$ elements of $\boldsymbol{\phi}$ and a one-hot encoding. Ordinal-valued hyperparameters can be encoded using a unary encoding. (For example, an ordinal variable which can take three values would be encoded with [1,0,0], [1,1,0], and [1,1,1].) Additional information about the distance between the values can be incorporated, if it's available. In this work, we then compute similarity using an RBF kernel, $\mathcal{K}=\exp\left(-\frac{||\boldsymbol{\phi}_i-\boldsymbol{\phi}_j||^2}{2\sigma^2}\right)$, and hence label our approach $k$-DPP-RBF. Values for $\sigma^2$ lead to models with different properties; when $\sigma^2$ is small, points that are spread out interact little with one another, and when $\sigma^2$ is large, the increased repulsion between the points encourages them to be as far apart as possible.
\subsection{Tree-structured hyperparameters}
Many real-world hyperparameter search spaces are tree-structured. For example, the number of layers in a neural network is a hyperparameter, and each additional layer adds at least one new hyperparameter which ought to be tuned (the number of nodes in that layer). For a binary hyperparameter like whether or not to use regularization, we use a one-hot encoding. When this hyperparameter is ``on,'' we set the associated regularization strength as above, and when it is ``off'' we set it to zero. Intuitively, with all other hyperparameter settings equal, this causes the off-setting to be closest to the least strong regularization. One can also treat higher-level design decisions as hyperparameters \citep{hyperopt_sklearn}, such as whether to train a logistic regression classifier, a convolutional neural network, or a recurrent neural network. In this construction, the type of model would be a categorical variable (and thus get a one-hot encoding), and all child hyperparameters for an ``off'' model setting (such as the convergence tolerance for logistic regression, when training a recurrent neural network) would be set to zero.
\begin{figure*}[h]
\centering
\includegraphics[width=\linewidth]{three_lr.pdf}
\caption{Average best-found model accuracy by iteration when training a convolutional neural network on three hyperparameter search spaces (defined in Section~\ref{sec:learn_rate_exp}), averaged across 50 trials of hyperparameter optimization, with $k=20$.
\label{fig:three_learning_rate_acc}}
\end{figure*}
\section{Hyperparameter Optimization Experiments}
In this section we present our hyperparameter optimization experiments. Our experiments consider a setting where hyperparameters have a large effect on performance: a convolutional neural network for text classification \citep{cnn_for_text}. The task is binary sentiment analysis on the Stanford sentiment treebank \citep{stanford_sentiment}. On this balanced dataset, random guessing leads to 50\% accuracy. We use the CNN-non-static model from \citet{cnn_for_text}, with skip-gram \citep{word2vec} vectors. The model architecture consists of a convolutional layer, a max-over-time pooling layer, then a fully connected layer leading to a softmax. All $k$-DPP samples are drawn using Algorithm \ref{alg:mixed_mcmc}.
\subsection{Simple tree-structured space}\label{sec:learn_rate_exp}
We begin with a search over three continuous hyperparameters and one binary hyperparameter, with a simple tree structure: the binary hyperparameter indicates whether or not the model will use $L_2$ regularization, and one of the continuous hyperparameters is the regularization strength. We assume a budget of $k=20$ evaluations by training the convolutional neural net. $L_2$ regularization strengths in the range $[e^{-5}, e^{-1}]$ (or no regularization) and dropout rates in $[0.0, 0.7]$ are considered. We consider three increasingly ``easy'' ranges for the learning rate:
\begin{itemize}
\item Hard: $[e^{-5}, e^5]$, where the majority of the range leads to accuracy no better than chance.
\item Medium: $[e^{-5}, e^{-1}]$, where half of the range leads to accuracy no better than chance.
\item Easy: $[e^{-10}, e^{-3}]$, where the entire range leads to models that beat chance.
\end{itemize}
Figure~\ref{fig:three_learning_rate_acc} shows the accuracy (averaged over 50 runs) of the best model found after exploring 1, 2, \ldots, $k$ hyperparameter settings. We see that $k$-DPP-RBF finds better models with fewer iterations necessary than the other approaches, especially in the most difficult case. Figure~\ref{fig:three_learning_rate_acc} compares the sampling methods against a Bayesian optimization technique using a tree-structured Parzen estimator \citep[BO-TPE;][]{bo_algos}. This technique evaluates points sequentially, allowing the model to choose the next point based on how well previous points performed (a closed loop approach). It is state-of-the-art on tree-structured search spaces (though its sequential nature limits parallelization). Surprisingly, we find it performs the worst, even though it takes advantage of additional information. We hypothesize that the exploration/exploitation tradeoff in BO-TPE causes it to commit to more local search before exploring the space fully, thus not finding hard-to-reach global optima.
Note that when considering points sampled uniformly or from a DPP, the order of the $k$ hyperparameter settings in one trial is arbitrary (though this is not the case with BO-TPE as it is an iterative algorithm).
In all cases the variance of the best of the $k$ points is lower than when sampled uniformly, and the differences in the plots are all significant with $p<0.01$.
\subsection{Optimizing within ranges known to be good}\label{sec:known_good}
\citet{sensitivity} analyzed the stability of convolutional neural networks for sentence classification with respect to a large set of hyperparameters, and found a set of six which they claimed had the largest impact: the number of kernels, the difference in size between the kernels, the size of each kernel, dropout, regularization strength, and the number of filters. We optimized over their prescribed ``Stable'' ranges for three open loop methods and one closed loop method; average accuracies with 95 percent confidence intervals from 50 trials of hyperparameter optimization are shown in Figure~\ref{fig:known_good}, across $k= 5,10,15,20$ iterations. We find that even when optimizing over a space for which all values lead to good models, $k$-DPP-RBF outperforms the other methods.
Our experiments reveal that, while the hyperparameters proposed by \citet{sensitivity}, can have an effect, the learning rate, which they do not analyze, is at least as impactful.
\begin{figure}[t]
\centering
\includegraphics[width=250pt]{known_good_range.pdf}
\caption{Average best-found model accuracy by iteration when training a convolutional neural network on the ``Stable'' search space (defined in Section~\ref{sec:known_good}), averaged across 50 trials of hyperparameter optimization, with $k=5,10,15,20$, with 95 percent confidence intervals. The $k$-DPP-RBF outperforms uniform sampling, TPE, and the Sobol sequence.
\label{fig:known_good}}
\end{figure}
\subsection{Wall clock time comparison with Spearmint}
Here we compare our approach against Spearmint \citep{practical_bo}, perhaps the most popular Bayesian optimization package. Figure \ref{fig:wall_time} shows wall clock time and accuracy for 25 runs on the ``Stable'' search space of four hyperparameter optimization approaches: $k$-DPP-RBF (with $k=20$), batch Spearmint with 2 iterations of batch size 10, batch Spearmint with 10 iterations of batch size 2, and sequential Spearmint\footnote{When in the fully parallel, open loop setting, Spearmint simply returns the Sobol sequence.}. Each point in the plot is one hyperparameter assignment evaluation. The vertical lines represent how long, on average, it takes to find the best result in one run. We see that all evaluations for $k$-DPP-RBF finish quickly, while even the fastest batch method (2 batches of size 10) takes nearly twice as long on average to find a good result. The final average best-found accuracies are 82.61 for $k$-DPP-RBF, 82.65 for Spearmint with 2 batches of size 10, 82.7 for Spearmint with 10 batches of size 2, and 82.76 for sequential Spearmint. Thus, we find it takes on average more than ten times as long for sequential Spearmint to find its best solution, for a gain of only 0.15 percent accuracy.
\begin{figure}[t]
\centering
\includegraphics[width=250pt]{wall_clock_time.pdf}
\caption{Wall clock time (in seconds, x-axis) for 25 hyperparameter trials of hyperparameter optimization (each with $k=20$) on the ``Stable'' search space define in Section~\ref{sec:known_good}. The vertical lines represent the average time it takes too find the best hyperparameter assignment in a trial.
\label{fig:wall_time}}
\end{figure}
\section{Conclusions}
We have explored open loop hyperparameter optimization built on sampling from a $k$-DPP. We described how to define a $k$-DPP over hyperparameter search spaces, and showed that $k$-DPPs retain the attractive parallelization capabilities of random search. In synthetic experiments, we showed $k$-DPP samples perform well on a number of important metrics, even for large values of $k$. In hyperprameter optimization experiments, we see $k$-DPP-RBF outperform other open loop methods. Additionally, we see that sequential methods, even when using more than ten times as much wall clock time, gain less than 0.16 percent accuracy on a particular hyperparameter optimization problem. An open-source implementation of our method is available.
\section{Submission of conference papers to ICLR 2019}
ICLR requires electronic submissions, processed by
\url{https://openreview.net/}. See ICLR's website for more instructions.
If your paper is ultimately accepted, the statement {\tt
{\textbackslash}iclrfinalcopy} should be inserted to adjust the
format to the camera ready requirements.
The format for the submissions is a variant of the NIPS format.
Please read carefully the instructions below, and follow them
faithfully.
\subsection{Style}
Papers to be submitted to ICLR 2019 must be prepared according to the
instructions presented here.
Authors are required to use the ICLR \LaTeX{} style files obtainable at the
ICLR website. Please make sure you use the current files and
not previous versions. Tweaking the style files may be grounds for rejection.
\subsection{Retrieval of style files}
The style files for ICLR and other conference information are available on the World Wide Web at
\begin{center}
\url{http://www.iclr.cc/}
\end{center}
The file \verb+iclr2019_conference.pdf+ contains these
instructions and illustrates the
various formatting requirements your ICLR paper must satisfy.
Submissions must be made using \LaTeX{} and the style files
\verb+iclr2019_conference.sty+ and \verb+iclr2019_conference.bst+ (to be used with \LaTeX{}2e). The file
\verb+iclr2019_conference.tex+ may be used as a ``shell'' for writing your paper. All you
have to do is replace the author, title, abstract, and text of the paper with
your own.
The formatting instructions contained in these style files are summarized in
sections \ref{gen_inst}, \ref{headings}, and \ref{others} below.
\section{General formatting instructions}
\label{gen_inst}
The text must be confined within a rectangle 5.5~inches (33~picas) wide and
9~inches (54~picas) long. The left margin is 1.5~inch (9~picas).
Use 10~point type with a vertical spacing of 11~points. Times New Roman is the
preferred typeface throughout. Paragraphs are separated by 1/2~line space,
with no indentation.
Paper title is 17~point, in small caps and left-aligned.
All pages should start at 1~inch (6~picas) from the top of the page.
Authors' names are
set in boldface, and each name is placed above its corresponding
address. The lead author's name is to be listed first, and
the co-authors' names are set to follow. Authors sharing the
same address can be on the same line.
Please pay special attention to the instructions in section \ref{others}
regarding figures, tables, acknowledgments, and references.
\section{Headings: first level}
\label{headings}
First level headings are in small caps,
flush left and in point size 12. One line space before the first level
heading and 1/2~line space after the first level heading.
\subsection{Headings: second level}
Second level headings are in small caps,
flush left and in point size 10. One line space before the second level
heading and 1/2~line space after the second level heading.
\subsubsection{Headings: third level}
Third level headings are in small caps,
flush left and in point size 10. One line space before the third level
heading and 1/2~line space after the third level heading.
\section{Citations, figures, tables, references}
\label{others}
These instructions apply to everyone, regardless of the formatter being used.
\subsection{Citations within the text}
Citations within the text should be based on the \texttt{natbib} package
and include the authors' last names and year (with the ``et~al.'' construct
for more than two authors). When the authors or the publication are
included in the sentence, the citation should not be in parenthesis (as
in ``See \citet{Hinton06} for more information.''). Otherwise, the citation
should be in parenthesis (as in ``Deep learning shows promise to make progress towards AI~\citep{Bengio+chapter2007}.'').
The corresponding references are to be listed in alphabetical order of
authors, in the \textsc{References} section. As to the format of the
references themselves, any style is acceptable as long as it is used
consistently.
\subsection{Footnotes}
Indicate footnotes with a number\footnote{Sample of the first footnote} in the
text. Place the footnotes at the bottom of the page on which they appear.
Precede the footnote with a horizontal rule of 2~inches
(12~picas).\footnote{Sample of the second footnote}
\subsection{Figures}
All artwork must be neat, clean, and legible. Lines should be dark
enough for purposes of reproduction; art work should not be
hand-drawn. The figure number and caption always appear after the
figure. Place one line space before the figure caption, and one line
space after the figure. The figure caption is lower case (except for
first word and proper nouns); figures are numbered consecutively.
Make sure the figure caption does not get separated from the figure.
Leave sufficient space to avoid splitting the figure and figure caption.
You may use color figures.
However, it is best for the
figure captions and the paper body to make sense if the paper is printed
either in black/white or in color.
\begin{figure}[h]
\begin{center}
\fbox{\rule[-.5cm]{0cm}{4cm} \rule[-.5cm]{4cm}{0cm}}
\end{center}
\caption{Sample figure caption.}
\end{figure}
\subsection{Tables}
All tables must be centered, neat, clean and legible. Do not use hand-drawn
tables. The table number and title always appear before the table. See
Table~\ref{sample-table}.
Place one line space before the table title, one line space after the table
title, and one line space after the table. The table title must be lower case
(except for first word and proper nouns); tables are numbered consecutively.
\begin{table}[t]
\caption{Sample table title}
\label{sample-table}
\begin{center}
\begin{tabular}{ll}
\multicolumn{1}{c}{\bf PART} &\multicolumn{1}{c}{\bf DESCRIPTION}
\\ \hline \\
Dendrite &Input terminal \\
Axon &Output terminal \\
Soma &Cell body (contains cell nucleus) \\
\end{tabular}
\end{center}
\end{table}
\section{Default Notation}
In an attempt to encourage standardized notation, we have included the
notation file from the textbook, \textit{Deep Learning}
\cite{goodfellow2016deep} available at
\url{https://github.com/goodfeli/dlbook_notation/}. Use of this style
is not required and can be disabled by commenting out
\texttt{math\_commands.tex}.
\centerline{\bf Numbers and Arrays}
\bgroup
\def1.5{1.5}
\begin{tabular}{p{1in}p{3.25in}}
$\displaystyle a$ & A scalar (integer or real)\\
$\displaystyle {\bm{a}}$ & A vector\\
$\displaystyle {\bm{A}}$ & A matrix\\
$\displaystyle {\tens{A}}$ & A tensor\\
$\displaystyle {\bm{I}}_n$ & Identity matrix with $n$ rows and $n$ columns\\
$\displaystyle {\bm{I}}$ & Identity matrix with dimensionality implied by context\\
$\displaystyle {\bm{e}}^{(i)}$ & Standard basis vector $[0,\dots,0,1,0,\dots,0]$ with a 1 at position $i$\\
$\displaystyle \text{diag}({\bm{a}})$ & A square, diagonal matrix with diagonal entries given by ${\bm{a}}$\\
$\displaystyle {\textnormal{a}}$ & A scalar random variable\\
$\displaystyle {\mathbf{a}}$ & A vector-valued random variable\\
$\displaystyle {\mathbf{A}}$ & A matrix-valued random variable\\
\end{tabular}
\egroup
\vspace{0.25cm}
\centerline{\bf Sets and Graphs}
\bgroup
\def1.5{1.5}
\begin{tabular}{p{1.25in}p{3.25in}}
$\displaystyle {\mathbb{A}}$ & A set\\
$\displaystyle \mathbb{R}$ & The set of real numbers \\
$\displaystyle \{0, 1\}$ & The set containing 0 and 1 \\
$\displaystyle \{0, 1, \dots, n \}$ & The set of all integers between $0$ and $n$\\
$\displaystyle [a, b]$ & The real interval including $a$ and $b$\\
$\displaystyle (a, b]$ & The real interval excluding $a$ but including $b$\\
$\displaystyle {\mathbb{A}} \backslash {\mathbb{B}}$ & Set subtraction, i.e., the set containing the elements of ${\mathbb{A}}$ that are not in ${\mathbb{B}}$\\
$\displaystyle {\mathcal{G}}$ & A graph\\
$\displaystyle \parents_{\mathcal{G}}({\textnormal{x}}_i)$ & The parents of ${\textnormal{x}}_i$ in ${\mathcal{G}}$
\end{tabular}
\vspace{0.25cm}
\centerline{\bf Indexing}
\bgroup
\def1.5{1.5}
\begin{tabular}{p{1.25in}p{3.25in}}
$\displaystyle {a}_i$ & Element $i$ of vector ${\bm{a}}$, with indexing starting at 1 \\
$\displaystyle {a}_{-i}$ & All elements of vector ${\bm{a}}$ except for element $i$ \\
$\displaystyle {A}_{i,j}$ & Element $i, j$ of matrix ${\bm{A}}$ \\
$\displaystyle {\bm{A}}_{i, :}$ & Row $i$ of matrix ${\bm{A}}$ \\
$\displaystyle {\bm{A}}_{:, i}$ & Column $i$ of matrix ${\bm{A}}$ \\
$\displaystyle {\etens{A}}_{i, j, k}$ & Element $(i, j, k)$ of a 3-D tensor ${\tens{A}}$\\
$\displaystyle {\tens{A}}_{:, :, i}$ & 2-D slice of a 3-D tensor\\
$\displaystyle {\textnormal{a}}_i$ & Element $i$ of the random vector ${\mathbf{a}}$ \\
\end{tabular}
\egroup
\vspace{0.25cm}
\centerline{\bf Calculus}
\bgroup
\def1.5{1.5}
\begin{tabular}{p{1.25in}p{3.25in}}
$\displaystyle\frac{d y} {d x}$ & Derivative of $y$ with respect to $x$\\ [2ex]
$\displaystyle \frac{\partial y} {\partial x} $ & Partial derivative of $y$ with respect to $x$ \\
$\displaystyle \nabla_{\bm{x}} y $ & Gradient of $y$ with respect to ${\bm{x}}$ \\
$\displaystyle \nabla_{\bm{X}} y $ & Matrix derivatives of $y$ with respect to ${\bm{X}}$ \\
$\displaystyle \nabla_{\tens{X}} y $ & Tensor containing derivatives of $y$ with respect to ${\tens{X}}$ \\
$\displaystyle \frac{\partial f}{\partial {\bm{x}}} $ & Jacobian matrix ${\bm{J}} \in \mathbb{R}^{m\times n}$ of $f: \mathbb{R}^n \rightarrow \mathbb{R}^m$\\
$\displaystyle \nabla_{\bm{x}}^2 f({\bm{x}})\text{ or }{\bm{H}}( f)({\bm{x}})$ & The Hessian matrix of $f$ at input point ${\bm{x}}$\\
$\displaystyle \int f({\bm{x}}) d{\bm{x}} $ & Definite integral over the entire domain of ${\bm{x}}$ \\
$\displaystyle \int_{\mathbb{S}} f({\bm{x}}) d{\bm{x}}$ & Definite integral with respect to ${\bm{x}}$ over the set ${\mathbb{S}}$ \\
\end{tabular}
\egroup
\vspace{0.25cm}
\centerline{\bf Probability and Information Theory}
\bgroup
\def1.5{1.5}
\begin{tabular}{p{1.25in}p{3.25in}}
$\displaystyle P({\textnormal{a}})$ & A probability distribution over a discrete variable\\
$\displaystyle p({\textnormal{a}})$ & A probability distribution over a continuous variable, or over
a variable whose type has not been specified\\
$\displaystyle {\textnormal{a}} \sim P$ & Random variable ${\textnormal{a}}$ has distribution $P$\\% so thing on left of \sim should always be a random variable, with name beginning with \r
$\displaystyle \mathbb{E}_{{\textnormal{x}}\sim P} [ f(x) ]\text{ or } \mathbb{E} f(x)$ & Expectation of $f(x)$ with respect to $P({\textnormal{x}})$ \\
$\displaystyle \mathrm{Var}(f(x)) $ & Variance of $f(x)$ under $P({\textnormal{x}})$ \\
$\displaystyle \mathrm{Cov}(f(x),g(x)) $ & Covariance of $f(x)$ and $g(x)$ under $P({\textnormal{x}})$\\
$\displaystyle H({\textnormal{x}}) $ & Shannon entropy of the random variable ${\textnormal{x}}$\\
$\displaystyle D_{\mathrm{KL}} ( P \Vert Q ) $ & Kullback-Leibler divergence of P and Q \\
$\displaystyle \mathcal{N} ( {\bm{x}} ; {\bm{\mu}} , {\bm{\Sigma}})$ & Gaussian distribution %
over ${\bm{x}}$ with mean ${\bm{\mu}}$ and covariance ${\bm{\Sigma}}$ \\
\end{tabular}
\egroup
\vspace{0.25cm}
\centerline{\bf Functions}
\bgroup
\def1.5{1.5}
\begin{tabular}{p{1.25in}p{3.25in}}
$\displaystyle f: {\mathbb{A}} \rightarrow {\mathbb{B}}$ & The function $f$ with domain ${\mathbb{A}}$ and range ${\mathbb{B}}$\\
$\displaystyle f \circ g $ & Composition of the functions $f$ and $g$ \\
$\displaystyle f({\bm{x}} ; {\bm{\theta}}) $ & A function of ${\bm{x}}$ parametrized by ${\bm{\theta}}$.
(Sometimes we write $f({\bm{x}})$ and omit the argument ${\bm{\theta}}$ to lighten notation) \\
$\displaystyle \log x$ & Natural logarithm of $x$ \\
$\displaystyle \sigma(x)$ & Logistic sigmoid, $\displaystyle \frac{1} {1 + \exp(-x)}$ \\
$\displaystyle \zeta(x)$ & Softplus, $\log(1 + \exp(x))$ \\
$\displaystyle || {\bm{x}} ||_p $ & $L^p$ norm of ${\bm{x}}$ \\
$\displaystyle || {\bm{x}} || $ & $L^2$ norm of ${\bm{x}}$ \\
$\displaystyle x^+$ & Positive part of $x$, i.e., $\max(0,x)$\\
$\displaystyle \bm{1}_\mathrm{condition}$ & is 1 if the condition is true, 0 otherwise\\
\end{tabular}
\egroup
\vspace{0.25cm}
\section{Final instructions}
Do not change any aspects of the formatting parameters in the style files.
In particular, do not modify the width or length of the rectangle the text
should fit into, and do not change font sizes (except perhaps in the
\textsc{References} section; see below). Please note that pages should be
numbered.
\section{Preparing PostScript or PDF files}
Please prepare PostScript or PDF files with paper size ``US Letter'', and
not, for example, ``A4''. The -t
letter option on dvips will produce US Letter files.
Consider directly generating PDF files using \verb+pdflatex+
(especially if you are a MiKTeX user).
PDF figures must be substituted for EPS figures, however.
Otherwise, please generate your PostScript and PDF files with the following commands:
\begin{verbatim}
dvips mypaper.dvi -t letter -Ppdf -G0 -o mypaper.ps
ps2pdf mypaper.ps mypaper.pdf
\end{verbatim}
\subsection{Margins in LaTeX}
Most of the margin problems come from figures positioned by hand using
\verb+\special+ or other commands. We suggest using the command
\verb+\includegraphics+
from the graphicx package. Always specify the figure width as a multiple of
the line width as in the example below using .eps graphics
\begin{verbatim}
\usepackage[dvips]{graphicx} ...
\includegraphics[width=0.8\linewidth]{myfile.eps}
\end{verbatim}
or
\begin{verbatim}
\usepackage[pdftex]{graphicx} ...
\includegraphics[width=0.8\linewidth]{myfile.pdf}
\end{verbatim}
for .pdf graphics.
See section~4.4 in the graphics bundle documentation (\url{http://www.ctan.org/tex-archive/macros/latex/required/graphics/grfguide.ps})
A number of width problems arise when LaTeX cannot properly hyphenate a
line. Please give LaTeX hyphenation hints using the \verb+\-+ command.
\subsubsection*{Acknowledgments}
Use unnumbered third level headings for the acknowledgments. All
acknowledgments, including those to funding agencies, go at the end of the paper.
| 0f8ba7b84098cfdcf517375f2fe909835709745e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\renewcommand{\theequation}{I.\arabic{equation}}
\setcounter{equation}{0}
An excellent and detailed explanation of Huygens' principle for undergraduate students, together with the optical-mechanical analogy and the Hamilton-Jacobi method, can be found in the monograph by Arnold \cite{Arno89}. Students are introduced to a generalization of Huygens' principle, viz. the Huygens-Fresnel superposition principle, in the study of general physics (see, e.g., \cite{Save82}), and this principle is presented in greater detail in the study of theoretical physics (see, e.g., \cite{Land71}). The method of Green's functions (GF)
which has found numerous applications in a large variety of different fields
is discussed in the first volume of a two-volume monograph by Bjorken and Drell \cite{Bjor64,Bjor65}, where, in particular, the superposition principle is used in~\S\S~21~-~22 to derive the equation for the Green's function.
Further development of concepts related to the superposition principle
has led to the emergence in quantum theory of the path integral formalism,
an excellent overview of which can be found in the monograph by Dittrich and Reuter \cite{DITT2001}. This material is intended for advanced students studying quantum field theory.
A detailed presentation of the superposition principle for electromagnetic fields, its rationale and its generalizations, based on Kirchhoff's integral theorem,
\cite{Kirch1883}
is given in the monograph by Born and Wolf \cite{Born99}. This monograph is intended for postgraduate students and researchers specializing in the theory of the propagation of electromagnetic waves and light phenomena.
Thus, it is clear that the superposition principle is closely related to the GF method which, in turn, lies at the heart of quantum field theory and the diagram technique. In the literature, this relationship is typically mentioned only in passing, while the mathematical aspects, modifications, and physical meaning of the generalized schemes of superposition are treated as matters beyond dispute.
A rigorous
formulation of the
superposition principle is based on Kirchhoff's integral theorem. The generalizations to which it leads are used also in the theory of interacting fields.
In this paper, we attempt to specify the precise place of the superposition principle in classical and quantum field theory and discuss its relationship with the GF method and Kirchhoff's integral theorem.
Surprisingly, the answers to the main questions can be obtained by analyzing the dynamics of the one-dimensional oscillator. The oscillator problem from the viewpoint of Kirchhoff's integral theorem, as well as its connections with the superposition principle and the GF method, is discussed in the next section.
In Sect. III, we consider a free massive scalar field. For massive fields, the superposition scheme includes an integral over three-dimensional space. Both in the limit of zero mass and for monochromatic fields, the canonical superposition scheme, in which the summation of the sources of secondary waves is limited to a two-dimensional surface, arises.
The statement of Kirchhoff's theorem depends on the asymptotic conditions imposed on the propagator at $t\to \pm \infty $. In quantum field theory, the Feynman asymptotic conditions are used. Emphasis is therefore placed on the versions of the theorem that satisfy the Feynman asymptotic conditions. In Sect. IV, we discuss a charged scalar field in an external electromagnetic field, prove the appropriate version of Kirchhoff's integral theorem, and demonstrate that in an external electromagnetic field, the superposition schemes are not fundamentally modified.
In nonlinear theories the superposition principle holds in relation to the secondary waves.
In Sect. V, we consider a class of nonlinear scalar field theories.
The physical meaning of Kirchhoff's integral theorem is discussed,
including its connections with the GF method and the superposition principle.
Vectorial generalizations of Kirchhoff's integral theorem for retarded Green's function
are discussed Appendix.
The conclusions section summarizes the discussion.
\section{The Huygens-Fresnel superposition principle and Kirchhoff's integral theorem in the oscillator problem}
\renewcommand{\theequation}{II.\arabic{equation}}
\setcounter{equation}{0}
A free scalar field obeys the Klein-Gordon equation:
\begin{equation}
(\Box +m^{2} )\phi _{0} ^{} (x)=0.
\label{1}
\end{equation}
Of interest are the general features of solutions of the wave equation,
which extend to its nonlinear modifications.
The main consequences of Kirchhoff's theorem and the physical content of the Fresnel-Huygens superposition principle can be explained using the example of the one-dimensional oscillator; thus, we begin by considering the evolution of a one-dimensional harmonic oscillator. This problem can also be regarded as a problem of the evolution of a free scalar field in momentum space.
\subsection{Harmonic oscillator}
We write the equation in the form
\begin{equation}
\left(\frac{d^{2} }{dt^{2} } +m^{2} \right)\phi _{0} (t)=0.
\label{2}
\end{equation}
Here, $m$ is the frequency of the oscillator and $\phi _{0} (t)$ is its coordinate. If $\phi _{0} (t)$ is a spatially homogeneous field in the Klein-Gordon equation, then $m$ is the mass of the particle.
\subsubsection{Complete orthonormal basis functions}
A complete set of solutions to Eq.~(\ref{2}) is formed by the two functions
\begin{equation}
f^{(+)} (t)=\frac{e^{-imt} }{\sqrt{2m} } \, \, \, \, \textrm{and} \, \, \, \, f^{(-)} (t)=\frac{e^{imt} }{\sqrt{2m} } .
\label{3}
\end{equation}
The normalization and completeness conditions are expressed in terms of the Wronskian. If $\varphi $ and $\chi $ are two functions, then their Wronskian is equal to
\begin{equation}
W[\varphi ,\chi ]=\det \left\| \begin{array}{cc} {\varphi } & {\chi } \\ {\dot{\varphi }} & {\dot{\chi }} \end{array}\right\| =\varphi \dot{\chi }-\dot{\varphi }\chi .
\end{equation}
The notation
\[\varphi \stackrel{\leftrightarrow}{\partial }_{t} \chi =W[\varphi ,\chi ]\]
is often used. The normalization and orthogonality of the basis functions are represented as follows:
\begin{equation}
iW[f^{(\pm )*} ,f^{(\pm )} ]=\pm 1 \, \, \, \, \textrm{and} \, \, \, \, W[f^{(\pm )*} ,f^{(\mp )} ]=0. \label{4}
\end{equation}
If the functions for which we compute the Wronskian are solutions of Eq.~(\ref{2}), then the Wronskian is independent of time. Let $\phi_{0} (t)$ be a solution of Eq.~(\ref{2}). We define the following time-independent complex numbers:
\begin{equation}
a=iW[f^{(+)*} ,\phi _{0} ] \, \, \, \, \textrm{and} \, \, \, \, a^{*} =-iW[f^{(-)*} ,\phi _{0} ]. \label{5}
\end{equation}
After quantization, the values $a$ and $a^*$ become annihilation and creation operators.
The completeness condition takes the form
\begin{equation}
\phi _{0} (t)=f^{(+)} (t)iW[f^{(+)*} ,\phi _{0} ]-f^{(-)} (t)iW[f^{(-)*} ,\phi _{0} ]. \label{6}
\end{equation}
This equation also allows for the decomposition of the solution into its positive- and negative-frequency components:
\begin{equation}
\phi _{0}^{} (t)=\phi _{0}^{(+)} (t)+\phi _{0}^{(-)} (t), \label{7}
\end{equation}
where
\begin{equation}
\phi_{0}^{(\pm )} (t)=\pm f^{(\pm )} (t)iW[f^{(\pm )*} ,\phi _{0} ].\end{equation}
Equation (\ref{6}) is valid not only in the linear vector space spanned by the basis functions (\ref{3}), but also for any function evaluated at time $t$. The right-hand side of Eq.~(\ref{6}) for an arbitrary function $\chi(t)$ has the form
\begin{equation*}
\textrm{r.h.s.}
=i\left( f^{(+)}(t)f^{(+)\ast }(t)-f^{(-)}(t)f^{(-)\ast
}(t)\right) \dot{\chi}(t)
-i\left( f^{(+)}(t)\dot{f}^{(+)\ast }(t)-f^{(-)}(t)\dot{f
^{(-)\ast }(t)\right) \chi(t).
\end{equation*}
Using the explicit form of $f^{(\pm)}(t)$, one can see that $\textrm{r.h.s.} = \chi(t)$. Although this property appears fortuitous, it is rather fundamental.
Let us consider the Poisson bracket relations
\begin{eqnarray}
\{\phi _{0}(t),\phi _{0}(t)\} &=&0, \label{PB1} \\
\{\phi _{0}(t),\pi _{0}(t)\} &=&1, \label{PB2}
\end{eqnarray}
where $\pi_{0}(t) = \dot{\phi}_{0}(t)$ is the canonical momentum. A simple calculation using Eq.~(\ref{6}) gives
\begin{eqnarray}
\{\phi _{0}(t^{\prime }),\phi _{0}(t)\}
&=&f^{(+)}(t)i\{\phi _{0}(t^{\prime }),W[f^{(+)\ast },\phi
_{0}]\}-f^{(-)}(t)i\{\phi _{0}(t^{\prime }),W[f^{(-)\ast },\phi _{0}]\} \nonumber \\
&=&f^{(+)}(t)i\{\phi _{0}(t^{\prime }),f^{(+)\ast }(t^{\prime })\pi
_{0}(t^{\prime })-\dot{f}^{(+)\ast }(t^{\prime })\phi _{0}(t^{\prime
})\} \nonumber \\
&-&f^{(-)}(t)i\{\phi _{0}(t^{\prime }),f^{(-)\ast }(t^{\prime })\pi
_{0}(t^{\prime })-\dot{f}^{(-)\ast }(t^{\prime })\phi _{0}(t^{\prime })\} \nonumber \\
&=& i\left(
f^{(+)}(t)f^{(+)\ast}(t^{\prime})-f^{(-)}(t)f^{(-)\ast}(t^{\prime})
\right), \label{bracket1} \\
\{\phi _{0}(t^{\prime }),\pi _{0}(t)\}
&=& i\left(
\dot{f}^{(+)}(t)f^{(+)\ast }(t^{\prime }) - \dot{f}^{(-)}(t) f^{(-)\ast }(t^{\prime})
\right). \label{bracket2}
\end{eqnarray}
By virtue of Eqs.~(\ref{PB1}) and (\ref{PB2}),
\begin{eqnarray}
f^{(+)}(t)f^{(+)\ast}(t)-f^{(-)}(t)f^{(-)\ast}(t)
&=&0, \label{Fund1} \\
f^{(+)}(t)\dot{f}^{(+)\ast }(t) - f^{(-)}(t)\dot{f}^{(-)\ast }(t)
&=& i, \label{Fund2} \\
\dot{f}^{(+)}(t)f^{(+)\ast }(t) - \dot{f}^{(-)}(t)f^{(-)\ast }(t) &=& -i. \label{Fund3}
\end{eqnarray}
Identity $\textrm{r.h.s.} = \chi(t)$ is, therefore, a consequence of the completeness condition (\ref{6}) for functions $\phi_0(t)$, which are solutions of Eq.~(\ref{2}), and the Poisson bracket relations for the canonical variables.
\subsubsection{The Green's functions}
A Green's function is defined by the equation
\begin{equation}
\left(\frac{d^{2} }{dt^{2} } +m^{2} \right)\Delta_X (t)=-\delta (t). \label{8}
\end{equation}
By performing the Fourier transform in time, we obtain the Green's function in frequency space:
$\Delta_X (\omega )= (\omega ^{2} -m^{2})^{-1}$.
For the inverse Fourier transformation,
\begin{equation}
\Delta_X (t)=\int _{-\infty }^{+\infty }\frac{d\omega }{2\pi } \, e^{-i\omega t} \frac{1}{\omega ^{2} -m^{2} } ,
\label{10}
\end{equation}
it is necessary to bypass the poles on the real axis that arise for $\omega =\pm m.$ There are four possibilities, which correspond to four Green's functions:
\begin{eqnarray}
\Delta_{F}(t'-t) &=& \int_{-\infty}^{+\infty}\frac{d\omega }{2\pi}
e^{-i\omega (t'-t)} \frac{1}{ \omega^{2} - m^{2} +i0}
\nonumber \\
&=&-i\left( f^{(+)}(t') f^{(+)*}(t)\theta(t'-t) + f^{(-)}(t')f^{(-)*}(t)\theta(-t'+t)\right),
\label{11} \\
\Delta _{F} ^{c} (t'-t)&=&\int _{-\infty }^{+\infty }\frac{d\omega }{2\pi } \, e^{-i\omega (t'-t)} \frac{1}{\omega ^{2} -m^{2} -i0} \nonumber \\
&=& i\left(f^{(-)} (t')f^{(-)*} (t)\theta (t'-t)+f^{(+)} (t')f^{(+)*} (t)\theta (-t'+t)\right),
\label{12} \\
\Delta _{\mathrm{ret}} (t'-t)&=&\int _{-\infty }^{+\infty }\frac{d\omega }{2\pi } e^{-i\omega (t'-t)} \frac{1}{\omega ^{2} -m^{2} +i0\mathrm{sgn}(\omega )} \nonumber \\
&=&-i\left( f^{(+)} (t')f^{(+)*} (t)-f^{(-)} (t')f^{(-)*} (t)\right)\theta (t'-t),
\label{13} \\
\Delta _{\mathrm{adv}} (t'-t)&=&\int _{-\infty }^{+\infty }\frac{d\omega }{2\pi } e^{-i\omega (t'-t)} \frac{1}{\omega ^{2} -m^{2} -i0\mathrm{sgn}(\omega )} \nonumber \\
&=& i\left(f^{(+)} (t')f^{(+)*} (t)-f^{(-)} (t')f^{(-)*} (t)\right)\theta (-t'+t).
\label{14}
\end{eqnarray}
Each of these functions satisfies Eq.~(\ref{8}). The difference between any two Green's functions is a solution of the free equation (\ref{2}).
It is instructive to verify by the direct calculation that the representation (\ref{11}) satisfies Eq.~(\ref{8}). With the help of equation
\begin{equation*}
f(x)\delta ^{\prime }(x)= f(0)\delta ^{\prime }(x)-f^{\prime }(0)\delta (x),
\end{equation*
one finds
\begin{eqnarray}
\left( \frac{d^{2}}{dt^{\prime 2}} + m^{2} \right) i\Delta_F (t^{\prime }-t) &=&
2 \left( \dot{f}^{(+)}(t^{\prime })f^{(+)\ast }(t) - \dot{f}^{(-)}(t^{\prime })f^{(-)\ast }(t)\right) \delta (t^{\prime }-t) \nonumber \\
&+&\left( f^{(+)}(t^{\prime })f^{(+)\ast}(t)
-f^{(-)}(t^{\prime })f^{(-)\ast}(t)\right) \delta^{\prime }(t^{\prime }-t) \nonumber \\
&=&
\left( \dot{f}^{(+)}(t^{\prime })f^{(+)\ast }(t)
-\dot{f}^{(-)}(t^{\prime })f^{(-)\ast }(t) \right) \delta (t^{\prime }-t) \nonumber \\
&+&
\left( f^{(+)}(t)f^{(+)\ast }(t)
-f^{(-)}(t)f^{(-)\ast }(t)\right) \delta ^{\prime }(t^{\prime }-t).
\label{delta1}
\end{eqnarray
Using Eqs.~(\ref{Fund1}) and (\ref{Fund3}), we arrive at Eq.~(\ref{8}).
In terms of quantized variables, the Feynman propagator is defined by
\begin{equation}
i\Delta _{F}(t^{\prime }-t)=\langle 0|T\hat{\phi}_{0}(t^{\prime })\hat{\phi}_{0}(t)|0\rangle.
\label{A7}
\end{equation}
The $T$ product entering this expression
occurs naturally in solutions of the evolution equation
$i\partial_{t} \Psi(t) = \hat{H}(t) \Psi(t)$
of systems with a time-dependent Hamiltonian.
If, at various times, $\hat{H}$ does not commute with itself, namely $[\hat{H}(t'),\hat{H}(t)] \neq 0$, then
the solution $\Psi(t) = U(t,0) \Psi(0)$ is expressed in terms of the time-ordered exponential
$U(t,0) = T \exp (-i\int^{t}_{0} \hat{H}(t^{\prime})dt^{\prime})$.
In perturbation theory $\Delta _{F}(t^{\prime }-t)$ then arises by Wick's theorem,
which explains why $\Delta _{F}(t^{\prime }-t)$ plays a special role in quantum theory.
The definition (\ref{A7}) is consistent with the definition (\ref{11}).
\subsubsection{Superposition principle from Kirchhoff's integral theorem}
Let us compute the Wronskian of the Feynman propagator $\Delta _{F}^{} (t^{\prime} -t)$ and a solution $\phi _{0} (t)$ of Eq.~(\ref{2}). By taking the derivative with respect to $t$ of $W[\Delta _{F} (t^{\prime} -t),\phi _{0} (t)]$ and integrating the result over the interval $(t_{1} ,t_{2} )$, the following equation is obtained for $t_{1} < t^{\prime} < t_{2} $:
\begin{equation}
\phi _{0} (t^{\prime} )=W[\Delta _{F}^{} (t^{\prime} -t_{2} ),\phi _{0} (t_{2} )]-W[\Delta _{F}^{} (t^{\prime} -t_{1} ),\phi _{0} (t_{1} )].
\label{15}
\end{equation}
This relation is the harmonic oscillator analog of Kirchhoff's integral theorem.
Despite the drastic simplification, the fundamental meaning is maintained and is amenable to interpretation. According to Eq.~(\ref{15}), the coordinate $\phi_{0}(t)$ is determined by both the past and the future.
From the past, the Wronskian selects the positive-frequency component of $\phi _{0} (t_{1} )$ and propagates it into the future up to the moment $t=t^{\prime} >t_{1} $. From the future, the Wronskian selects the negative-frequency component of $\phi _{0} (t_{2} )$ and propagates it into the past up to the moment $t=t^{\prime} <t_{2} $. The result is a superposition of the two
\textit{waves}.
Equation (\ref{2}) is commonly regarded as the equation of motion of a particle (oscillator) in the 1-dimensional space. A less obvious interpretation of this equation as an evolution equation of a wave in the 0-dimensional space is also possible. Equation (\ref{15}) underlines the second interpretation.
The analogy with quantum field theory is apparent: particles are identified with positive-frequency solutions of wave equations, and antiparticles are identified with negative-frequency solutions. Particles move \textit{forward in time}, whereas antiparticles move \textit{backward in time}. In accordance with the Huygens-Fresnel superposition principle adapted here
for the Feynman asymptotic conditions,
the wave $\phi _{0} (t^{\prime} )$ is equal to the sum of the negative-frequency component of $\phi _{0} (t_{2} )$, propagating backward in time, and the positive-frequency component of $\phi _{0} (t_{1} )$, propagating forward in time. Equation (\ref{15}) can thus be interpreted both in the spirit of the Huygens-Fresnel superposition principle and in the spirit of the GF method, thereby establishing the close relationship between them.
According to Eq.~(\ref{15}), the coordinate $\phi _{0} (t^{\prime} )$ is determined by its value and its first derivative at the other two time points. Arguing reversely, this suggests that the evolution equation contains time derivatives of no higher than second order.
If $t^{\prime} \notin (t_{1} ,t_{2} )$, then there is a zero on the left-hand side of Eq.~(\ref{15}):
\begin{equation}
0=W[\Delta _{F}(t^{\prime} -t_{2} ),\phi _{0} (t_{2} )]
- W[\Delta _{F}(t^{\prime} -t_{1} ),\phi _{0} (t_{1} )].
\label{16}
\end{equation}
Equations (\ref{15}) and (\ref{16}) remain valid after the replacement $\Delta _{F}$ with any other propagator. For the retarded Green's function, the analog of Eqs.~(\ref{15}) and (\ref{16}) for $t_2 \to +\infty$ reads
\begin{equation}
\phi _{0}^{} (t^{\prime} )\theta (t^{\prime} - t_{1})=-W[\Delta _{\mathrm{ret}}^{} (t^{\prime} -t_{1}),\phi _{0}^{} (t_{1})].
\label{17}
\end{equation}
Here, the positive- and negative-frequency components propagate forward in time, corresponding to the usual formulation of the Huygens-Fresnel superposition principle, so that $\phi_{0}(t)$ is determined by the past only.
\subsubsection{Superposition principle from the completeness condition}
Here, we present a different formulation of the
superposition principle. To begin, let us find the Wronskian $W$ of $\Delta _{F} (t^{\prime} -t)$ and $\phi _{0} (t)$. The expression (\ref{11}), when substituted into $W$, yields
\begin{eqnarray}
W[\Delta _{F}(t^{\prime} -t),\phi _{0} (t)] &=&
- if^{(+)}(t^{\prime} )W[f^{(+)*} (t)\theta(t^{\prime} -t),\phi _{0} (t)]
-if^{(-)}(t^{\prime} )W[f^{(-)*} (t)\theta(t-t^{\prime} ),\phi _{0}] \nonumber \\
&=&-if^{(+)}(t^{\prime} )\theta(t^{\prime} -t)W[f^{(+)*} ,\phi _{0} ]-if^{(-)} (t^{\prime} )\theta (t-t^{\prime} )W[f^{(-) *} ,\phi _{0} ] \nonumber \\
&& + \Delta(t^{\prime} - t) \phi _{0} (t)\delta (t^{\prime} -t),
\label{18}
\end{eqnarray}
where
\begin{equation}
i\Delta(t^{\prime} - t) = f^{(+)}(t^{\prime} )f^{(+) *} (t)-f^{(-)} (t^{\prime} )f^{(-) *} (t).
\label{comm}
\end{equation}
By virtue of Eq.~(\ref{bracket1})
\begin{equation*}
\Delta(t^{\prime } - t) = \{\phi _{0}(t^{\prime }),\phi _{0}(t)\}.
\end{equation*}
In the transition to the last lines of Eq.~(\ref{18}), the properties of the Wronskian and the definitions of the basis functions (\ref{3}) are used.
According to Eq.~(\ref{Fund1}), the term $\sim \Delta(t^{\prime } - t) \delta(t^{\prime } - t)$ vanishes, yielding
\begin{equation}
\phi _{0}^{(+)} (t^{\prime} )\theta (t^{\prime} -t)-\phi _{0}^{(-)} (t^{\prime} )\theta (t-t^{\prime} )=-W[\Delta _{F}^{} (t^{\prime} -t),\phi _{0}^{} (t)].
\label{19}
\end{equation}
Equation (\ref{19}) can be regarded as an equation for
$\Delta _{F}(t^{\prime} -t)$. By taking the time ($t$) derivative of both sides, we obtain Eq.~(\ref{8}). The superposition principle, formalized as in (\ref{19}), thus determines the Green's function up to a solution of the free equation. To obtain a unique Green's function, the asymptotic behavior must be fixed. By taking the differences between both sides of Eq.~(\ref{19}) for $t=t_{2} $ and $t=t_{1}< t_{2}$, we obtain Eq.~(\ref{15}), provided that $t^{\prime} \in (t_{1} ,t_{2} )$. If the inverse condition, $t^{\prime} \notin (t_{1} ,t_{2} )$, holds, then we obtain Eq.~(\ref{16}). Finally, by taking the time ($t^{\prime} $) derivative, we obtain the superposition principle for the canonical momentum $\pi _{0}^{} (t)=\dot{\phi }_{0}^{} (t)$:
\begin{equation}
\pi _{0}^{(+)} (t^{\prime} )\theta (t^{\prime} -t)-\pi _{0}^{(-)} (t^{\prime} )\theta (t-t^{\prime} )=-W[\Delta _{F}^{} (t^{\prime} -t),\pi _{0}^{} (t)].
\label{19momentum}
\end{equation}
The proof of Eq.~(\ref{19}) is not based on Kirchhoff's theorem nor its obvious modification. For the retarded Green's function, the completeness condition does not lead to a new equation (compared with (\ref{17})).
In quantum field theory, the diagram technique is based on the Feynman propagator; thus, what is of interest to us here is the superposition principle formalized as in (\ref{15}), (\ref{16}) and (\ref{19}).
\subsubsection{Path integral}
Kirchhoff's integral theorem can also be used as a starting point for developing path integral method.
To show this, we note a useful relation
\begin{eqnarray}
iW[\Delta _{F}(t_{3}-t_{2}),\Delta _{F}(t_{2}-t_{1})]&=&-\theta
(t_{3}-t_{2})\theta (t_{2}-t_{1})f^{(+)}(t_{3})f^{(+)\ast }(t_{1}) \nonumber \\
&&+\theta(t_{1}-t_{2})\theta (t_{2}-t_{3})f^{(-)}(t_{3})f^{(-)\ast }(t_{1}).
\label{twopro}
\end{eqnarray
This relation tells that a wave propagating toward the future continues to propagate forward in time.
A similar property holds for waves propagating backward in time.
We choose a sequence of the intervals
$(t_{1},t_{2})\subset (t_{3},t_{4})\subset \ldots \subset
(t_{2n-1},t_{2n})$ and consider $t^{\prime} \in
(t_{1},t_{2})$. Eq.~(\ref{15}) being iterated $n$ times gives
\begin{eqnarray}
\phi _{0}(t^{\prime})&=& W[\Delta _{F}(t^{\prime }-t_{2}),W[\Delta
_{F}(t_{2}-t_{4}),W[\ldots ,W[\Delta _{F}(t_{2n}-t_{2n + 2}),\phi_{0}(t_{2n + 2})]\ldots ]]] \label{pi} \\
&+&(-)^{n + 1}W[\Delta _{F}(t^{\prime }-t_{1}),W[\Delta_{F}(t_{1}-t_{3}),W[\ldots ,W[\Delta _{F}(t_{2n-1}-t_{2n+1}),\phi_{0}(t_{2n+1})]\ldots ]]]. \nonumber
\end{eqnarray}
According to this equation, $\phi _{0}(t_{2n+2})$ generates a secondary wave that propagates into the past.
In the neighboring instant of time $t=t_{2n}< t_{2n+2}$ it generates new secondary wave, etc.
The same interpretation is valid for the wave propagating forward in time.
Equation~(\ref{15}) is reproduced with $n=0$ for $t_{-1} = t_{0} = t^{\prime}$.
The mixed terms containing forward and backward propagation do not arise as a consequence of (\ref{twopro}).
In the limit of $n \rightarrow \infty $, $t_2 - t_1 \to 0$ and $(t_{l+3} - t_{l+2}) \to (t_{l+1} - t_{l})$, we arrive at the continuous product \textit{over history}.
Equation (\ref{pi}) can be regarded as a path-integral representation in the space $\mathbb{R}^{1,0}$.
Path integral in the space $\mathbb{R}^{1,3}$ is discussed in Sect. III.E.
\subsection{Harmonic oscillator with a time-dependent frequency }
A field theoretical version of the evolution problem with a time-dependent oscillator frequency, in light of the superposition principle, is discussed in Sect. IV, where proofs are presented. Here, we restrict ourselves to statements of the main assertions.
We consider the equation
\begin{equation}
\left(\frac{d^{2} }{dt^{2} } +m^{2} +\Delta m^{2} (t)\right)\phi (t)=0,
\label{20}
\end{equation}
where $\Delta m^{2} (\pm \infty )=0.$ The perturbation $\Delta m^{2} (t)$ is switched on and off adiabatically. Let $\Delta _{F} (t',t)$ be the Feynman propagator for Eq.~(\ref{20}). The following superposition schemes hold: As a consequence of Kirchhoff's integral theorem,
\begin{eqnarray*}
\phi (t^{\prime} )&=& W[\Delta_{F}(t^{\prime} ,t_{2}),\phi (t_{2} )]
- W[\Delta_{F}(t^{\prime} ,t_{1}),\phi (t_{1} )]
\;\;\;\;\textrm{for} \;\;\;\; t^{\prime} \in (t_{1} ,t_{2} ), \\
0&=& W[\Delta_{F}(t^{\prime} ,t_{2}),\phi (t_{2} )]
-W[\Delta_{F}(t^{\prime} ,t_{1}),\phi (t_{1} )] \;\;\;\; \textrm{for} \;\;\;\; t^{\prime} \notin (t_{1} ,t_{2} ),
\end{eqnarray*}
and, as a consequence of the completeness condition,
\begin{equation*}
\phi _{}^{(+)} (t^{\prime} )\theta (t^{\prime} -t)-\phi _{}^{(-)} (t^{\prime} )\theta (t-t^{\prime} )=-W[\Delta _{F}^{} (t^{\prime} ,t),\phi (t)],
\end{equation*}
where $\phi^{(\pm )} (t) \sim f^{(\pm )} (t)$ at $t\to \pm \infty $. The expansion of $\phi (t)$ into positive- and negative-frequency components $ \phi^{(\pm)} (t)$ has an objective meaning because the evolution equation is linear.
\subsection{Anharmonic oscillator}
Let us consider a more general case. We add to the oscillator potential an arbitrary potential $V(\phi )$. The equation of motion takes the form
\begin{equation}
\left(\frac{d^{2} }{dt^{2} } +m^{2} \right)\phi (t)=-V'(\phi (t)).
\label{21}
\end{equation}
\subsubsection{Superposition principle from Kirchhoff's integral theorem}
Equation (\ref{15}) is modified as follows:
\begin{equation}
\phi (t^{\prime} )=W[\Delta _{F}^{} (t^{\prime} -t_{2} ),\phi (t_{2} )]-W[\Delta _{F}^{} (t^{\prime} -t_{1} ),\phi (t_{1} )]+\int _{t_{1} }^{t_{2} }dt\Delta _{F}^{} (t^{\prime} -t)V'(\phi (t)) .
\label{22}
\end{equation}
The propagator $\Delta _{F}^{} (t)$ is determined from Eq.~(\ref{8}). On the interval $(t_{1} ,t_{2} )$ the sum of the first two terms satisfies the evolution equation of the harmonic oscillator. We denote this sum as
\begin{equation}
\phi _{0} (t^{\prime} )\equiv W[\Delta _{F}^{} (t^{\prime} -t_{2} ),\phi (t_{2} )]-W[\Delta _{F}^{} (t^{\prime} -t_{1} ),\phi (t_{1} )].
\label{23}
\end{equation}
The solution takes the form
\begin{equation}
\phi (t^{\prime} )=\phi _{0} (t^{\prime} )+\int _{t_{1} }^{t_{2} }dt\Delta _{F}^{} (t^{\prime} -t)V'(\phi (t)) .
\label{24}
\end{equation}
Given that the Green's function properties of the harmonic oscillator are known, the solution can be written immediately. If $t^{\prime} \notin (t_{1} ,t_{2} )$, then we obtain
\begin{equation}
0=\phi _{0} (t^{\prime} )+\int _{t_{1} }^{t_{2} }dt\Delta _{F}^{} (t^{\prime} -t)V'(\phi (t)) .
\label{25}
\end{equation}
The last two equations constitute a version of Kirchhoff's integral theorem for the one-dimensional anharmonic oscillator.
Equation (\ref{24}) cannot be interpreted canonically. Although the first term has the standard meaning under the Fresnel superposition scheme, the second term indicates that a component arises among the secondary waves that is generated continuously in time.
According to the Huygens-Fresnel superposition principle, to describe the propagation of a wave, it is sufficient to know its phase and amplitude at a fixed time. However, this is true only in linear theories. In nonlinear theories, the propagation of a wave
is determined by its entire history (for retarded solutions, its prehistory), even if the original wave equation is local. The dependence of the wave observables on the entire history of the wave indicates, in general, the nonlocal nature of its evolution. Only a narrow family of representations that contain an integral over time correspond to local but nonlinear theories.
The derivative of the potential is an additional source of secondary waves (corrections to the coordinate), and the potential depends on the exact coordinate. This means that Eq. (\ref{24}) is self-consistent and that its solution is obvious only in the context of perturbation theory.
In quantum field theory, an equation similar to Eq.~(\ref{24}) serves as the starting point for the development of the diagram technique (see, e.g., \cite{Bjor64}). The equations obtained by replacing the Feynman propagator in Eq.~(\ref{24}) with the retarded and advanced propagators are used to develop the axiomatic scattering theory
(see, e.g., \cite{Bjor65}).
\subsubsection{Positive- and negative-frequency solutions}
In the theory of interacting fields, the decomposition of solutions into positive- and negative-frequency components makes sense only asymptotically for outgoing and incoming states. We assume that the nonlinear interaction is adiabatically switched on at $t \to - \infty$ and adiabatically switched off at $t \to + \infty$. If positive- and negative-frequency components $\phi^{(\pm )} (t)$ are somehow defined, then the subsequent modification of Eq.~(\ref{19}) is obvious:
\begin{equation}
\phi^{(+)}(t^{\prime} )\theta (t^{\prime} -t)-\phi _{}^{(-)} (t^{\prime} )\theta (t-t^{\prime} )=-W[\Delta _{F}^{} (t^{\prime} -t),\phi (t)]
+\int _{t}^{t^{\prime} }d\tau\Delta _{F} (t^{\prime} - \tau)V'(\phi (\tau)) .
\label{26}
\end{equation}
By taking the time ($t$) derivative, after some simple transformations, we obtain $\phi (t)=\phi^{(+)} (t)+\phi^{(-)} (t)$ and Eq.~(\ref{8}). The difference in this equation at two unequal time points leads to Eqs.~(\ref{22}) - (\ref{25}). It might seem, therefore, that Eq. (\ref{26}) is no less general than Eqs.~(\ref{22}) - (\ref{25}). However, we do not have an independent definition of the decomposition into positive- and negative-frequency components. We are forced, therefore, to regard Eq.~(\ref{26}) as a definition of $\phi^{(\pm)} (t)$. According to this equation, $\phi _{}^{(\pm )} (t)\sim f^{(\pm )} (t)$ at $t\to \pm \infty $.
The calculation of the first derivative of Eq.~(\ref{26}) in $t^{\prime} $ leads to the superposition principle for the canonical momentum
\begin{equation}
\pi^{(+)} (t^{\prime} )\theta (t^{\prime} -t)-\pi _{}^{(-)} (t^{\prime} )\theta (t-t^{\prime} )=-W[\Delta _{F}(t^{\prime} -t),\pi (t)]
+\int _{t}^{t^{\prime} }d\tau\Delta _{F}^{} (t^{\prime} -\tau)V''(\phi (\tau))\pi (\tau) .
\end{equation}
This equation is consistent with the evolution equation for $\pi^{(\pm )} (t)=\dot{\phi }^{(\pm )}(t)$.
Obviously, in nonlinear theories, a full generalization of (\ref{19}) does not exist.
A field theoretical version of the anharmonic oscillator problem is discussed in Sect. V.
\subsubsection{Numerical example}
We use a numerical example to demonstrate the application of the superposition scheme (\ref{24}) for the description of radial motion in the Keplerian problem.
After separation of the angular variables, the evolution problem reduces to solving a problem of one-dimensional motion in an effective potential
\[
U=-\frac{\alpha }{r}+\frac{L^{2}}{2\mu r^{2}},
\
where $\alpha =GM_{\odot }\mu $, $M_{\odot }$ is the solar mass, $\mu $ is
the mass of a celestial body, and $L$ is the angular momentum.
We add and subtract
from the potential $U$ an oscillator potential
\[
U_{osc}=\frac{1}{2}\mu m^{2}(r-a)^{2}
\
and treat $U_{osc}$ as the undisturbed potential.
The perturbation potential is thus $V=U-U_{osc}.$
In order to improve convergence and eliminate the need to determine optimized $U_{osc}$,
the frequency parameter $m$ is chosen in agreement with the exact solution (see, e.g., \cite{Arno89}):
$m=2\pi/T$, where $T=2\pi \mu ab/L$ is the orbital period, $a = (r_{\min } + r_{\max })/2$ and $b = \sqrt{pa}$ are the major and minor semi-axes of the ellipse
and $L = \sqrt{p \alpha \mu}$; the variable $r$ lies in the interval $(r_{\min },r_{\max})$, where $r_{\min } = p/(1 + e)$, $r_{\max } = p/(1 - e)$,
$p$ is the semi-latus rectum, and $e$ is the
eccentricity.
As a zeroth-order approximation for $\phi(t) \equiv r(t) - a$, we choose a free solution
\begin{equation}
\phi^{[0]}(t) = C_{+}^{[0]}f^{(+)}(t)+C_{-}^{[0]}f^{(-)}(t)
\end{equation}
with unknown coefficients $C_{\pm}^{[0]}$ and $f^{(\pm)}(t)$ defined by Eq.~(\ref{3}).
The motion begins at perihelion $\phi^{[0]}(0)=r_{\min } - a$,
with the vanishing velocity $\dot{\phi}^{[0]}(0)=0$. These conditions allow $C_{\pm}^{[0]}$ to be fixed.
Given the $l$th-order approximation, $r^{[l]} (t) = a + \phi^{[l]}(t)$ can be substituted in place of the argument of $V^{\prime }$
in Eq.~(\ref{24}) to produce the next-order iteration
\begin{equation}
\phi^{[l+1]}(t) = C_{+}^{[l+1]}f^{(+)}(t)+C_{-}^{[l+1]}f^{(-)}(t)+\int_{t_{1}}^{t_{2}}
\tau \Delta _{F}(t-\tau )V^{\prime }(a + \phi^{[l]}(\tau )), \label{25 bis}
\end{equation
where $\Delta _{F}(t)$ is defined by (\ref{11}).
The interval $(t_{1},t_{2})$ covers an interval within which we seek
the solution. $C_{\pm}^{[l+1]}$ are fixed
by the conditions $\phi^{[l+1]}(0)=r_{\min } - a$ and $\dot{\phi}^{[l+1]}(0)=0$.
\begin{table}[h]
\caption{Expansion coefficients of free solutions in the unperturbed potential
for the first two iterations and for the exact solution ($l=\infty$).}
\label{tab:2}
\centering
\vspace{2pt}
\begin{tabular}{|c|l|l|}
\hline\hline
$l$ & $~~~~~~~~~~~~~C^{[l]}_{+}$ & $~~~~~~~~~~~~~C^{[l]}_{-}$ \\
\hline\hline
$0$ & $-0.142872$ & $-0.142872$ \\
$1$ & $-0.155969 - i0.040544$ & $-0.155322 - i0.068246$ \\
$\infty$ & $-0.151619 - i0.033743$ & $-0.151875 - i0.033990$ \\
\hline\hline
\end{tabular
\end{table}
The numerical convergence of the recursion is a subtle issue that should be studied separately.
Assuming the convergence of the approximate sequence, we should obtain an identity when using $r(t)$ to evaluate the integral in Eq.~(\ref{24}):
\begin{equation}
\phi^{[\infty]}(t) = C_{+}^{[\infty]}f^{(+)}(t)+C_{-}^{[\infty]}f^{(-)}(t)+\int_{t_{1}}^{t_{2}}
\tau \Delta _{F}(t-\tau )V^{\prime }(a + \phi(\tau )).
\label{N2}
\end{equation
The exact solutions are parameterized in terms of the eccentric anomaly $E$:
$r=a(1-e\cos E )$ and $t=\sqrt{ma^{3}/\alpha }(E -e\sin E )$, where $t$ is time.
For our numerical estimates, we choose $\alpha = \mu = p = 1$ and $e=0.2$.
The values $t_{1}$ and $t_{2}$ are taken arbitrarily; they correspond to $E_{1} = -1$ and $E_{2} = 7.2$.
The coefficients $C_{\pm}^{[l]}$ for $l=0,1,\infty$ found as described above are presented in Table \ref{tab:2}.
Table \ref{tab:1} shows $r^{[0]}$, $r^{[1]}$ and $r^{[\infty]}$ for seven values of $E \in [ 0,2\pi ].$
The inclusion of the secondary waves generated by the nonlinear source $V^{\prime}$ reduces the standard deviation
$\chi^2 = \sum (r^{[l]} - r)^2$ from $0.0038$ to $0.0015$, whereas $r^{[\infty]}$ coincides with $r$.
Equation~(\ref{24}) can also be derived directly,
under the assumption of $t \in (t_1,t_2)$,
by using the GF method,
whereas Eqs.~(\ref{23}) and (\ref{25}) are specific consequences of Kirchhoff's integral theorem.
We verified that the free term in Eq.~(\ref{N2}) fulfills, numerically, Eq.~(\ref{23})
and checked Eq.~(\ref{25}) for a sample set of time points $t \notin (t_1,t_2)$ as well.
\vspace{4pt}
Summarizing, the idea of Kirchhoffβs integral theorem was explained in this section with a one-dimensional toy model (a harmonic oscillator).
Such a pedagogical approach illustrates formalism while the attempt to draw a physical analogy with well-known phenomena
leads to the seemingly paradoxical observation:
no waves in the $\mathbb{R}^{1,0}$ space, but the superposition principle is there,
and even the problem of celestial mechanics was solved using Kirchhoffβs integral theorem in a technically consistent manner.
A parallelism between classical mechanics and geometrical optics
was regarded as purely formal until the advent of quantum mechanics.
The possibility of solving the problems of classical mechanics using the methods of wave optics seems to be a surprising circumstance.
\begin{table}[t]
\caption{First two iterations $r^{[l]}$ for the approximate solution of the radial equation of motion
as compared to the exact solution $r=r^{[\infty]}$ for seven values of $E \in [0,2\pi]$.
}
\label{tab:1}
\centering
\vspace{2pt}
\begin{tabular}{|c|c|c|c|}
\hline\hline
$E$ & $r^{[0]}$ & $r^{[1]}$ & $r^{[\infty]}$ \\
\hline\hline
$0$ & 0.8333 & 0.8333 & 0.8333 \\
$\pi/3$ & 0.9079 & 0.9410 & 0.9375 \\
$2\pi/3$ & 1.1131 & 1.1619 & 1.1458 \\
$\pi$ & 1.2500 & 1.2500 & 1.2500 \\
$4\pi/3$ & 1.1132 & 1.1232 & 1.1458 \\
$5\pi/3$ & 0.9079 & 0.9094 & 0.9375 \\
$2\pi$ & 0.8333 & 0.8333 & 0.8333 \\
\hline\hline
\end{tabular
\end{table}
\section{ Kirchhoff's integral theorem for a free scalar field}
\renewcommand{\theequation}{III.\arabic{equation}}
\setcounter{equation}{0}
\subsection{Complete orthonormal basis functions}
A complete set of solutions to the Klein-Gordon equation is formed by the functions
\begin{equation*}
f_{\mathbf k}^{(+)}(x) =
\frac{e^{-ikx}} {\sqrt{2\omega_{\mathbf k}}} \;\;\;\;
\mathrm{and} \;\;\;\;
f_{\mathbf k}^{(-)}(x) =
\frac{e^{ ikx}} {\sqrt{2\omega_{\mathbf k}}},
\end{equation*}
where $k=(\omega_{\mathbf k},{\mathbf k})$, $\omega _{\mathbf k} = \sqrt{{\mathbf k}^{2} + m^{2} }$, $x=(t,{\mathbf x}) \in \mathbb{R}^{1,3}$,
and $kx = \omega_{\mathbf k}t - {\mathbf k}{\mathbf x}$. These functions correspond to the positive- and negative-frequency solutions in the oscillator problem. The orthonormality conditions are
\begin{eqnarray}
i\int d{\mathbf x} W[f_{{\mathbf k'}}^{(\pm )*} (x),f_{{\mathbf k}}^{(\pm )} (x)]&=&\pm (2\pi )^{3} \delta ({\mathbf k'}-{\mathbf k}), \nonumber \\
\int d{\mathbf x} W[f_{{\mathbf k'}}^{(\mp )*} (x),f_{{\mathbf k}}^{(\pm )} (x)]&=&0.
\label{27}
\end{eqnarray}
For any function $\phi _{0} (x)$ that is a solution of the Klein-Gordon equation,
\begin{equation}
\phi _{0} (x)=\int \frac{d{\mathbf k}}{(2\pi )^{3} } \left(f_{{\mathbf k}}^{(+)} (x)i\int d{\mathbf y} W[f_{{\mathbf k}}^{(+)*} (y),\phi _{0} (y)]
-f_{{\mathbf k}}^{(-)} (x)i\int d{\mathbf y} W[f_{{\mathbf k}}^{(-)*} (y),\phi _{0} (y)]\right).
\label{28}
\end{equation}
After the second quantization, the time-independent quantities
\begin{equation*}
a({\mathbf k}) = i\int d{\mathbf y}
W[f_{\mathbf k}^{(+)*}(y),\phi_{0}(y)]\;\;\;\; \mathrm{and}
\;\;\;\; a^{*}({\mathbf k})=-i\int d{\mathbf y} W[f_{\mathbf k}^{(-)*}(y),\phi_{0}(y)]
\end{equation*}
become annihilation and creation operators.
The first and the second terms in Eq.~(\ref{28}) are identified with the positive- and negative-frequency components of $\phi_{0} (x)$. According to the completeness condition (\ref{28}), the solutions of the free equation thereby split into the sum
\begin{equation*}
\phi_{0} (x)=\phi _{0} ^{(+)} (x)+\phi _{0} ^{(-)} (x).
\end{equation*}
This decomposition is analogous to the decomposition of Eq.~(\ref{7}). The orthonormality conditions (\ref{27}) and the completeness condition (\ref{28}) are the generalized equivalents to Eqs. (\ref{4}) and (\ref{6}), respectively, for the oscillator problem.
Using the analogy with equations (\ref{PB1}) - (\ref{Fund3}) and the Poisson bracket relations
\begin{eqnarray}
\{\phi _{0}(x),\phi _{0}(y)\}|_{x^0 = y^0} &=&0, \label{PB3} \\
\{\phi _{0}(x),\pi _{0}(y)\}|_{x^0 = y^0} &=&
\delta(\mathbf{x} - \mathbf{y}), \label{PB4}
\end{eqnarray}
one can prove that
\begin{eqnarray}
\int \frac{d\mathbf{k}}{(2\pi)^3} \left(
f^{(+)}_{\mathbf{k}}(x)f^{(+)\ast}_{\mathbf{k}}(y) - f^{(-)}_{\mathbf{k}}(x)f^{(-)\ast}_{\mathbf{k}}(y) \right)|_{x^0 = y^0}
&=&0, \label{Fund4} \\
\int \frac{d\mathbf{k}}{(2\pi)^3} \left(
f^{(+)}_{\mathbf{k}}(x)\dot{f}^{(+)\ast }_{\mathbf{k}}(y) - f^{(-)}_{\mathbf{k}}(x)\dot{f}^{(-)\ast }_{\mathbf{k}}(y)
\right)|_{x^0 = y^0} &=& i\delta(\mathbf{x} - \mathbf{y}), \label{Fund5} \\
\int \frac{d\mathbf{k}}{(2\pi)^3} \left(
\dot{f}^{(+)}_{\mathbf{k}}(x)f^{(+)\ast }_{\mathbf{k}}(y) - \dot{f}^{(-)}_{\mathbf{k}}(x)f^{(-)\ast }_{\mathbf{k}}(y)
\right)|_{x^0 = y^0} &=& -i\delta(\mathbf{x} - \mathbf{y}). \label{Fund6}
\end{eqnarray}
Equations (\ref{Fund4}) and (\ref{Fund5}) can be used to show that the completeness condition (\ref{28}) holds for arbitrary functions at $x^0 = y^0$.
\subsection{Feynman propagator}
The equation for the Feynman propagator is
\begin{equation}
(\Box +m^{2} )\Delta_F (x)=-\delta ^{4} (x).
\label{29}
\end{equation}
It is easiest to find the solution in four-momentum space and then apply the Fourier transform to convert it into coordinate space. Here, as in the oscillator problem, we must shift the contour of the integral over $k^0$ from the real axis in the vicinity of
$k^0 = \pm \omega_{\mathbf k}$. The four possible ways to do so correspond to four Green's functions.
The Feynman propagator can be written as follows:
\begin{eqnarray}
\Delta _{F}(x-y)&=&\int \frac{d^{4} k}{(2\pi )^{4} } \frac{e^{-ik(x-y)} }{k^{2} -m^{2} +i0} \nonumber \\
&=&-i\int \frac{d{\mathbf k}}{(2\pi )^{3} } \left( f_{{\mathbf k}}^{(+)} (x)f_{{\mathbf k}}^{(+)*} (y)\theta (x^{0} -y^{0} )+f_{{\mathbf k}}^{(-)} (x)f_{{\mathbf k}}^{(-)*} (y)\theta (- x^{0} + y^{0} ) \right).
\label{Fpro}
\end{eqnarray}
In comparison with Eq.~(\ref{11}), the phase space integral is added here. After the replacement $f^{(\pm)}(t) \to f_{\mathbf k}^{(\pm)}(x)$ and the integration over the phase space in Eqs.~(\ref{12}), (\ref{13}), and (\ref{14}), the form of the other propagators is restored.
Using the analogy with Eq.~(\ref{delta1}) and Eqs.~(\ref{Fund4}) and (\ref{Fund6}), one can verify that the propagator (\ref{Fpro}) satisfies Eq.~(\ref{29}).
\subsection{Superposition principle from Kirchhoff's integral theorem}
\subsubsection{General form of the superposition principle}
We start from the identity
\begin{equation}
\phi _{0} (\xi )\delta ^{4} (\xi -x)=\Delta _{F} (x-\xi )\left((\Box _{\xi } +m^{2} )\phi _{0} (\xi )\right)-\left((\Box _{\xi } +m^{2} )\Delta _{F} (x-\xi )\right)\phi _{0} (\xi ).
\label{30}
\end{equation}
The right-hand side can be written in divergence form as follows:
\begin{equation}
\phi _{0} (\xi )\delta ^{4} (\xi -x)=\frac{\partial }{\partial \xi _{\mu } } \left(\Delta _{F} (x-\xi )\frac{\stackrel{\leftrightarrow}{\partial }}{\partial \xi ^{\mu } } \phi _{0} (\xi )\right).
\label{31}
\end{equation}
By taking the integral over a four-dimensional region $\Omega $ and transforming the right-hand side into a surface integral, the equation
\begin{equation}
\phi _{0} (x)\theta (x\in \Omega )=\int _{\partial \Omega }^{}dS_{\xi }^{\mu } \left(\Delta _{F} (x-\xi )\frac{\stackrel{\leftrightarrow}{\partial }}{\partial \xi ^{\mu } } \phi _{0} (\xi )\right),
\label{32}
\end{equation}
is obtained, where $\theta (x\in \Omega )$ is the indicator function of $\Omega $:
\begin{equation*}
\theta (x\in \Omega )=\left\{\begin{array}{cc} {1,} & {x\in \Omega \, ,} \\ {0,} & {x\notin \Omega \, .} \end{array}\right.
\end{equation*}
By choosing for the surface $\partial \Omega $ a hyperplane $\xi ^{0} =y^{0} $ in the past, i.e., three-dimensional space at a time $\xi ^{0} = y^{0} <x^{0} $, and a three-dimensional space $\xi ^{0} =z^{0} $ at a time
$\xi ^{0} = z^{0} > x^{0} $ in the future, and then combining these spaces at infinity, where the integral vanishes, we arrive at
\begin{equation}
\phi _{0} (x) =\int d{\mathbf z}W[\Delta _{F} (x-z),\phi _{0} (z)] -\int d{\mathbf y}W[\Delta _{F} (x-y),\phi _{0} (y)] .
\label{33}
\end{equation}
If $x\notin \Omega $, we obtain
\begin{equation}
0 =\int d{\mathbf z}W[\Delta _{F} (x-z),\phi _{0} (z)] -\int d{\mathbf y}W[\Delta _{F} (x-y),\phi _{0} (y)] .
\label{34}
\end{equation}
Equation (\ref{33}) states that $\phi _{0} (x)$ is determined by its past and future. Equation (\ref{34}) suggests that the interference of secondary waves outside the interval $(y^{0} ,z^{0} )$ is strictly destructive.
Equation (\ref{32}) and its consequences (\ref{33}) and (\ref{34}) constitute a version of Kirchhoff's theorem in the most general form;
these equations hold for any choice of propagator.
\subsubsection{Monochromatic field }
The Fourier transform simplifies the superposition scheme of secondary waves. We restrict ourselves to the case of monochromatic, spatially inhomogeneous waves. Consider the following Fourier transforms in time of the scalar field and the Green's function:
\begin{equation}
\phi _{0} (\omega ,{\mathbf x})=\int _{-\infty }^{+\infty }dte^{i\omega t} \phi _{0} (t,{\mathbf x}),\, \, \, \, \Delta _{F} (\omega ,{\mathbf x})=\int _{-\infty }^{+\infty }dte^{i\omega t} \Delta _{F} (t,{\mathbf x}).
\label{36}
\end{equation}
They satisfy the equations
\begin{equation*}
(\Delta +{\mathbf k}^{2} )\phi _{0} (\omega ,{\mathbf x})=0,\, \, \, \, (\Delta +{\mathbf k}^{2} )\Delta _{F} (\omega ,{\mathbf x})=\delta ({\mathbf x}),
\end{equation*}
where ${\mathbf k}^{2} =\omega ^{2} -m^{2} $. The right-hand side of the identity
\begin{eqnarray}
\phi _{0} (\omega ,{\boldsymbol\xi })\delta ({\boldsymbol \xi}-{\mathbf x}) &=&
-\Delta _{F} (\omega ,{\mathbf x}-{\boldsymbol\xi})\left((\Delta _{{\xi}} +{\mathbf k}^{2} )\phi _{0} (\omega ,{\boldsymbol\xi})\right) \nonumber \\
&& +\left((\Delta _{{\xi}} +{\mathbf k}^{2} )\Delta _{F} (\omega ,{\mathbf x}-{\boldsymbol\xi})\right)\phi _{0} (\omega ,{\boldsymbol\xi}).
\label{G2I}
\end{eqnarray}
can be represented as the divergence
\begin{equation*}
\phi _{0} (\omega ,{\boldsymbol\xi })\delta ({\boldsymbol\xi }-{\mathbf x})=-\frac{\partial }{\partial \xi ^{\alpha } } \left(\Delta _{F} (\omega ,{\mathbf x}-{\boldsymbol\xi })\frac{\stackrel{\leftrightarrow}{\partial }}{\partial \xi ^{\alpha } } \phi _{0} (\omega ,{\boldsymbol\xi })\right).
\end{equation*}
Integrating over the region $\Omega _{3} $, we obtain von Helmholtz's theorem for the monochromatic field: \cite{Helm1860}
\begin{equation}
\phi _{0} (\omega ,{\mathbf x})\theta ({\mathbf x}\in \Omega _{3} )=-\int _{\partial \Omega _{3} }dS^{\alpha }_{\xi } \Delta _{F} (\omega ,{\mathbf x}-{\boldsymbol\xi })\frac{\stackrel{\leftrightarrow}{\partial }}{\partial \xi ^{\alpha } } \phi _{0} (\omega ,{\boldsymbol\xi }),
\label{37}
\end{equation}
which is a particular case of the third Green's identity \cite{GGreen1828} and a precursor of Kirchhoff's integral theorem.
The integration is performed over the surface $\partial \Omega _{3} $, which is the boundary of $\Omega _{3} $.
The equation shows that the field at the point ${\mathbf x}$ is determined by its values on any surrounding surface. This surface is not required to be the wave surface. If the point ${\mathbf x}$ lies outside the closed surface, then the integral vanishes. Regardless of the specific form of $\Delta _{F} (\omega ,{\mathbf x})$, we can conclude from the form of the equation alone that if the field $\phi _{0} (\omega ,{\mathbf x})$ satisfies a differential equation, then this equation contains derivatives over the spatial coordinates that are no higher than second order.
Equation (\ref{37}) is used to describe the diffraction phenomena of light \cite{Land71,Born99}.
In the monochromatic, spatially inhomogeneous case, the integration is over the surface rather than over the volume, as in Eq.~(\ref{33}). However, because we are discussing the calculation of the Fourier transform in time, an implicit time integration enters the problem.
\subsubsection{Massless field }
For massless particles the interference scheme for secondary waves simplifies. Let us apply the inverse Fourier transform in Eq.~(\ref{37}):
\begin{equation}
\phi _{0} (t,{\mathbf x})\theta ({\mathbf x}\in \Omega _{3} )=-\int _{\partial \Omega _{3} }dS_{{\xi }}^{\alpha } \int _{-\infty }^{+\infty }dt' \Delta _{F} (t-t',{\mathbf x}-{\boldsymbol\xi })\frac{\stackrel{\leftrightarrow}{\partial }}{\partial \xi ^{\alpha } } \phi _{0} (t',{\boldsymbol\xi }).
\label{38}
\end{equation}
This equation follows from Eq.~(\ref{32}) if we select for $\Omega $ an infinite cylinder whose spatial section $\Omega _{3} $ is covered by the surface of integration $\partial \Omega _{3} $ and the axis is parallel to the time axis.
As is well known, the propagator $\Delta _{F} (t,{\mathbf x})$ does not vanish outside of the light cone $t^{2} -{\mathbf x}^{2} <0$. This property does not generally violate causality, as $\Delta _{F} (t,{\mathbf x})$ also describes the propagation of the wave surfaces at which the phase remains constant. In the relativistic theory,
the phase velocity $v_{p} \equiv \omega _{{\mathbf k}} /\left|{\mathbf k}\right|\ge 1$ is greater than the speed of light; however, it is the group velocity $v_{g} \equiv \partial \omega _{{\mathbf k}} /\partial \left|{\mathbf k}\right| $$=\left|{\mathbf k}\right|/\omega _{{\mathbf k}} \le 1$ with which the propagation of signals is associated.
In the limit of zero mass, the propagator $\Delta _{F} (t,{\mathbf x})$ takes the following form (see \cite{Bjor65}, Appendix B or Eq.~(29) in Ref. \cite{Zavi79} in the massless limit):
\begin{equation}
\Delta _{F} (t,{\mathbf x})=\frac{i}{4\pi ^{2} } \frac{1}{t^{2} -|{\mathbf x}|^{2} -i0}.
\label{39}
\end{equation}
Substituting (\ref{39}) into (\ref{38}) and taking into account that
\begin{equation*}
\int _{-\infty }^{+\infty }dt' \Delta _{F} (t-t',{\mathbf x})e^{\mp i\omega _{{\mathbf k}} t'} = \frac{1}{4\pi |{\mathbf x}| }
e^{\mp i\omega_{{\mathbf k}} (t \mp |{\mathbf x}|) },
\end{equation*}
we obtain
\begin{eqnarray}
\phi _{0} (t,{\mathbf x})\theta ({\mathbf x}\in \Omega _{3} ) =
\frac{1}{4\pi } \int _{\partial \Omega _{3} }dS_{{\boldsymbol\xi }}^{\alpha } \left[
- \frac{1}{\rho } \frac{\partial }{\partial \xi ^{\alpha } }
\right.
\left(\phi _{0}^{(+)} (t-\rho /c,{\boldsymbol\xi })+\phi _{0}^{(-)} (t+\rho/c ,{\boldsymbol\xi }) \right)&& \nonumber \\
+\left(\frac{\partial }{\partial \xi ^{\alpha } } \frac{1}{\rho } \right)\left(\phi _{0}^{(+)} (t-\rho /c ,{\boldsymbol\xi })+\phi _{0}^{(-)} (t+\rho /c,{\boldsymbol\xi }) \right)&& \nonumber \\
- \left. \frac{1}{\rho } \frac{\partial \rho }{\partial \xi ^{\alpha } } \frac{\partial }{c \partial t} \left(\phi _{0}^{(+)} (t-\rho /c,{\boldsymbol\xi })-\phi _{0}^{(-)} (t+\rho /c,{\boldsymbol\xi }) \right)
\right],&&
\label{40}
\end{eqnarray}
where $\rho =\left|{\boldsymbol\xi }-{\mathbf x}\right|$ and in the first term the differentiation with respect to $\xi ^{\alpha } $ does not apply to $\rho $.
The dependence on the speed of light $c$ is here made explicit.
Equation (\ref{40}) represents a general form of Kirchhoff's integral theorem for the Feynman asymptotic conditions. The function is determined by its values on the selected arbitrary closed surface, taking into account the delay of the positive-frequency component and the advancement of the negative-frequency component. This representation is possible because massless particles travel at the speed of light, regardless of their momentum.
\footnote{In Euclidean space of dimension $n \geq 3$ Green's function has the form $\Delta(x) \sim 1/(x^2)^{(n-2)/2}$. Performing a Wick rotation, we find that the Green's function as an analytic function of the variable $ t=x^0$ has two isolated poles in the spaces of even dimension and two root branching points in the spaces of odd dimension. This means that in the massless case the Green's function is effectively localized on the light cone in the spaces of even dimension only. Here an analogue of the representation (\ref{40}) holds. In the spaces of odd dimension, the superposition scheme involves the integration over all spatial coordinates. This property of the Green's function suggests that the requirement of equal phase and group velocities and the speed of light is a necessary but not sufficient condition for the representation of superposition scheme in the form of a surface integral.}
By contrast, the speed of a massive particle depends on its momentum; therefore, the more general representation (\ref{38}) includes the integral over time delay and advance. Kirchhoff's theorem is a precise mathematical formulation of the Huygens-Fresnel superposition principle. A special feature of the Feynman asymptotic conditions is that the negative-frequency components are determined by the future.
An analogue of Eq.~(\ref{40}) for the retarded solutions is the original version of Kirchhoff's integral theorem. It is
briefly outlined in Appendix and discussed in detail in Ref.~\cite{Born99}
\subsection{Superposition principle from the completeness condition}
As a formalization of the superposition principle for the Feynman asymptotic conditions, by analogy with Eq.~(\ref{19}), we can consider
\begin{equation}
\phi _{0} ^{(+)} (x)\theta ( x^{0} - y^{0} )
-\phi _{0} ^{(-)} (x)\theta (-x^{0} + y^{0} )
=-\int d{\mathbf y}W[\Delta _{F} (x-y),\phi _{0} (y)] .
\label{35}
\end{equation}
The physical content of this equation is quite traditional: At the moment $y^{0} $, the wave is a source of secondary waves, and the propagation from point $y$ to point $x$ is described by $\Delta _{F} (x-y)$. To construct the positive-frequency waves, the past $y^{0} <x^{0} $ must be known, and to construct the negative-frequency waves, the future $x^{0} <y^{0} $ must be known. This property is reflected in the presence of the theta functions on the left-hand side of the equation.
The proof of Eq.~(\ref{35}) is similar to the proof of Eq.~(\ref{19}). It is not based on Kirchhoff's theorem but instead relies on the completeness condition (\ref{28}) and the expansion of the Feynman propagator into plane waves. Given that Eq.~(\ref{35}) is postulated, the Green's function is uniquely determined. Indeed, let us take the derivative over $y^{0} $ on both sides of the equation. After the transformation of the integrand, we obtain Eq.~(\ref{29}); it must then be supplemented by asymptotic conditions.
We take the difference (\ref{35}) for two instants of time, $z^{0}$ and $y^{0} $, such that $y^{0} < x^{0} < z^{0} $. The result is
Eq.~(\ref{33}). If $x^{0} \notin (y^{0} ,z^{0} )$, then we obtain (\ref{34}).
According to equations (\ref{35}) and (\ref{33}), the field is determined by its values and first derivatives at two time points. This property indicates the local nature of the evolution equation. Arguing backward, since the initial conditions required to determine the field are the field values and the first derivatives, the evolution equation may contain time derivatives of no higher than second order.
Additionally, in Eq.~(\ref{32}), a hypersurface $\partial \Omega $ in the form of an infinite cylinder with its axis parallel to the time axis, can be chosen. In such a case, Kirchhoff's theorem would assert that the wave is determined by its values and gradients on a two-dimensional surface at all times. This version of the theorem indicates the local nature of the evolution equation in the spatial coordinates. The corresponding differential equation may contain derivatives of the spatial coordinates of no higher than second order.
\subsection{Path integral}
The path integral representation is a consequence of Eq.~(\ref{32}).
We choose a set of four-dimensional regions $\Omega _{1}\subset \Omega _{2}\subset \ldots \subset \Omega _{n} \subset \mathbb{R}^{1,3}$.
By iterating Eq.~(\ref{32}), we obtain
\begin{eqnarray}
\phi _{0}(x)\theta (x \in \Omega _{1}) &=& \int_{\partial \Omega _{1}}dS_{\xi
_{1}}^{\mu _{1}}\int_{\partial \Omega _{1}}dS_{\xi _{2}}^{\mu _{2}}\ldots
\int_{\partial \Omega _{n}}dS_{\xi _{n}}^{\mu _{n}} \label{pir13} \\
&\times& \Delta _{F}(x-\xi _{1})\frac{\overset{\leftrightarrow }{\partial }}{\partial \xi _{1}^{\mu _{1}}}
\Delta _{F}(\xi _{1}-\xi
_{2})\frac{\overset{\leftrightarrow }{\partial }}{\partial \xi _{2}^{\mu
_{2}}} \ldots \Delta _{F}(\xi _{n-1}-\xi _{n})\frac{\overset
\leftrightarrow }{\partial }}{\partial \xi _{n}^{\mu _{n}}}\phi _{0}(\xi
_{n}). \nonumber
\end{eqnarray}
There exists considerable freedom in choosing $\Omega_{i}$. A similar freedom exists in the factorization of
unitary evolution operator $U(t_2,t_1)$ in quantum mechanics,
where the equation $U(t_2,t_1) = U (t_2,t) U (t,t_1)$ holds for any instant of time $t \in (t_1,t_2)$.
While the evolution operator is factorized in time, the integration in the path integral goes over the coordinates in three-dimensional space.
Such a representation easily follows from Eq.~(\ref{pir13}).
Indeed, choosing $\Omega_{i}$ to be cylinders with infinite radii and axes parallel to the time axis, we arrive at a representation of this kind.
The broken lines connecting the points $x$ and $\xi_{n} \in \partial \Omega_{n}$ through $\xi_{i} \in \partial \Omega_{i}$ ($i=1,\ldots,n-1$)
form in the continuum limit the class of paths over which the continual integral is defined.
The comparison of Eqs.~(\ref{32}) and (\ref{pir13}) also yields, in the limit of $n \to \infty$, an integral representation
for the Green's function in the form of a continual integral.
\section{Charged scalar field in an external electromagnetic field}
\renewcommand{\theequation}{IV.\arabic{equation}}
\setcounter{equation}{0}
Equations (\ref{32}) and (\ref{35}) and their particular cases were obtained for a free field. The following question arises: which relations can be generalized in the presence of an external field? We restrict ourselves to scalar electrodynamics.
\subsection{Complete orthonormal basis functions}
Substituting the normal derivatives with respect to the space-time coordinates in the Klein-Gordon equation with gauge covariant derivatives,
\begin{equation}
\partial _{\mu } \to D_{\mu } =\partial _{\mu } +ieA_{\mu } \label{41}
\end{equation}
yields the evolution equation for a complex scalar field in an external electromagnetic field,
\begin{equation}
(D_{\mu } D^{\mu } +m^{2} )\phi (x)=0.
\label{42}
\end{equation}
The external field is adiabatically switched on at $t\to -\infty $ and off at $t\to +\infty $. The set of positive- and negative-frequency asymptotic solutions $f_{{\mathbf k}}^{(\pm )} (x)$ is complete and orthonormal. The second-order Eq.~(\ref{43}) has a set of independent solutions $F_{{\mathbf k}}^{(\pm )} (x)$. The asymptotic conditions can be taken as
\begin{equation*}
F_{{\mathbf k}}^{(\pm )} (x)\to f_{{\mathbf k}}^{(\pm )} (x)\equiv \frac{e^{\mp ikx} }{\sqrt{2\omega_{ \mathbf{k} } } } \, \, \, \, {\mathrm {for}}\, \, \, \, t \to -\infty .
\end{equation*}
All other solutions of Eq.~(\ref{42}) are expressed as linear superpositions of the basis functions $F_{{\mathbf k}}^{(\pm )} (x)$.
It would be natural to use the prescription (\ref{41}) for extending the Huygens-Fresnel superposition principle. It can be assumed that in an external electromagnetic field, the suitable generalization of the Wronskian is given by
\begin{equation*}
W_{A} [\varphi ^{*} ,\chi ]\equiv \varphi ^{*} (\stackrel{\leftrightarrow}{\partial }_{t} +2ieA_{0} )\chi =\varphi ^{*} (D_{t} \chi )-(D_{t} \varphi )^{*} \chi .
\end{equation*}
We note a useful property:
\begin{eqnarray}
\partial _{t} W_{A} [\varphi ^{*} ,\chi ]
&=&
\partial _{t} (\varphi ^{*} (D_{t} \chi )-(D_{t}^{} \varphi )^{*}
\chi ) \nonumber \\
&=& \varphi^{*} (D_{t} D_{t} \chi )-(D_{t}^{} D_{t}^{} \varphi )^{*} \chi .
\label{43}
\end{eqnarray}
It is not difficult to show that if $\varphi $ and $\chi $ are two solutions of Eq.~(\ref{42}), then the following condition holds:
\begin{equation*}
\partial_{t} \int d \mathbf{x} W_{A} [\varphi ^{*} ,\chi ]=0.
\end{equation*}
This condition allows us to calculate the normalization integral by sending the time variable to negative infinity, where solutions are represented as plane waves. The orthonormality conditions thus take the form
\begin{eqnarray*}
i\int d {\mathbf x} W_{A} [F_{{\mathbf k}'}^{(\pm )*} (x),F_{{\mathbf k}}^{(\pm )} (x)]&=&\pm (2\pi )^{3} \delta ({\mathbf k}'-{\mathbf k}), \\
\int d {\mathbf x} W_{A} [F_{{\mathbf k}'}^{(\pm )*} (x),F_{{\mathbf k}}^{(\mp )} (x)]&=&0.
\end{eqnarray*}
The completeness condition is obvious:
\begin{equation}
\phi (x)=\int \frac{d\mathbf{k}}{(2\pi )^{3} } \left(F_{{\mathbf k}}^{(+)} (x)i\int d{\mathbf y} W_A [F_{\mathbf k}^{(+)*} (y),\phi (y)]-F_{{\mathbf k}}^{(-)} (x)i\int d{\mathbf y} W_A [F_{{\mathbf k}}^{(-)*} (y),\phi (y)]\right).
\label{extf}
\end{equation}
In the theory of a charged scalar field, the canonical momenta are defined by equations $\pi^{\ast} (x) = D_t \phi (x)$ and $\pi (x) = (D_t \phi (x))^{\ast}$. The canonically conjugate variables satisfy
\begin{equation}
\{\phi(x),\pi(y)\}|_{x^0 = y^0} = \{\phi(x)^{\ast},\pi^{\ast}(y)\}|_{x^0 = y^0} = \delta(\mathbf{x} - \mathbf{y}), \label{PB6}
\end{equation}
while other pairs have the vanishing Poisson bracket.
The generalization of the corresponding relations of a free scalar field can be written as follows
\begin{eqnarray}
\int \frac{d\mathbf{k}}{(2\pi)^3} \left(
F^{(+)}_{\mathbf{k}}(x)F^{(+)\ast}_{\mathbf{k}}(y) - F^{(-)}_{\mathbf{k}}(x)F^{(-)\ast}_{\mathbf{k}}(y) \right)|_{x^0 = y^0}
&=&0, \label{Fund7} \\
\int \frac{d\mathbf{k}}{(2\pi)^3} \left(
F^{(+)}_{\mathbf{k}}(x)D_t^{\ast } F^{(+)\ast }_{\mathbf{k}}(y) - F^{(-)}_{\mathbf{k}}(x)D_t^{\ast } F^{(-)\ast }_{\mathbf{k}}(y)
\right)|_{x^0 = y^0} &=& i \delta(\mathbf{x} - \mathbf{y}), \label{Fund8} \\
\int \frac{d\mathbf{k}}{(2\pi)^3} \left(
D_t F^{(+)}_{\mathbf{k}}(x)F^{(+)\ast }_{\mathbf{k}}(y) -
D_t F^{(-)}_{\mathbf{k}}(x) F^{(-)\ast }_{\mathbf{k}}(y)
\right)|_{x^0 = y^0} &=& -i \delta(\mathbf{x} - \mathbf{y}). \label{Fund9}
\end{eqnarray}
Equations (\ref{Fund7}) and (\ref{Fund8}) show that the completeness condition (\ref{extf}) holds for arbitrary functions evaluated at $x^0 = y^0$.
In conclusion we note that the zeroth component of vector potential can be removed by a gauge transformation, in which case $W_A = W$ and other relations and their proofs take the form more similar to the free case.
\subsection{Feynman propagator}
The decomposition of the Feynman propagator over the basis functions has the form
\begin{equation}
\Delta _{F} (x,y)=-i\int \frac{d\mathbf{k}}{(2\pi )^{3} } \left(F_{{\mathbf k}}^{(+)} (x)F_{{\mathbf k}}^{(+)*} (y)\theta (x^{0} -y^{0} )+F_{{\mathbf k}}^{(-)} (x)F_{{\mathbf k}}^{(-)*} (y)\theta (-x^{0} +y^{0} )\right).
\label{proext}
\end{equation}
The use of Eqs.~(\ref{Fund7}) and (\ref{Fund9}) allows to verify by the direct calculation that
\begin{equation}
(D_{\mu } D^{\mu } +m^{2} )\Delta _{F} (x,\xi )=-\delta ^{4} (x-\xi ).
\label{44}
\end{equation}
\subsection{Superposition principle from Kirchhoff's integral theorem }
To derive Eq.~(\ref{31}), the identity (\ref{30}) was used.
After recapitulating the arguments used in
the proof of Eq.~(\ref{43}), we rewrite the divergence of
\begin{equation*}
\varphi \stackrel{\leftrightarrow}{D}_{\mu } \chi \equiv \varphi (D_{\mu } \chi )-(D_{\mu }^{*} \varphi )\chi ,
\end{equation*}
where $\varphi $ and $\chi $ are arbitrary functions, in the form
\begin{equation*}
\partial _{\mu } (\varphi \stackrel{\leftrightarrow}{D^{\mu }} \chi )=\varphi (D_{\mu } D^{\mu } \chi )-(D^{*} _{\mu } D^{*\mu } \varphi )\chi .
\end{equation*}
Substituting $\Delta _{F} (x,\xi )$ and $\phi (\xi )$ in place of $\varphi $ and $\chi $, respectively, we obtain
\begin{equation}
\phi (\xi )\delta ^{4} (x-\xi )=\frac{\partial }{\partial \xi _{\mu } } \left( \Delta _{F} (x,\xi )(D_{\mu } \, \phi (\xi ))-(D_{\mu }^{*} \Delta _{F} (x,\xi ))\phi (\xi )\right).
\label{45}
\end{equation}
By choosing as the integration region a four-dimensional space with the variable $\xi ^{0} $ running in the interval $(y^{0} ,z^{0} )$, we find for $x^{0} \in (y^{0} ,z^{0} )$
\begin{eqnarray}
\phi (x) = \int d{\mathbf z}W_{A} [\Delta _{F} (x,z),\phi (z)] -\int d{\mathbf y}W_{A} [\Delta _{F} (x,y),\phi (y)].
\end{eqnarray}
In the opposite case of $x^{0} \notin (y^{0} ,z^{0} )$ the left-hand side vanishes.
\subsection{Superposition principle from the completeness condition}
The linearity of the evolution equation allows for the generalization of the superposition principle (\ref{35}) in the presence of an external electromagnetic field. The completeness condition leads to the following scheme:
\begin{equation}
\phi ^{(+)} (x)\theta (x^{0} -y^{0} )-\phi ^{(-)} (x)\theta (-x^{0} +y^{0} )=-\int d{\mathbf y}W_{A} [\Delta _{F} (x,y),\phi (y)] .
\label{46}
\end{equation}
Under the integral sign, the derivative entering $W_A$ also generates the term
\begin{equation*}
\Delta(x,y) = -i \int \frac{d\mathbf{k}}{(2\pi )^{3} } \left(
F_{\mathbf k}^{(+)} (x) F_{\mathbf k}^{(+)*} (y)
- F_{\mathbf k}^{(-)} (x) F_{\mathbf k}^{(-)*} (y) \right)
\end{equation*}
multiplied by $\phi (y) \delta (x^{0} -y^{0})$.
In view of the relationship $x^{0} = y^{0}$ and Eq.~(\ref{Fund7}), this term vanishes. By calculating the derivative of Eq.~(\ref{46}) with respect to $y^{0} $, one can prove that the propagator obeys equation
\begin{equation}
(D_{\mu }^{*} D_{}^{\mu *} +m^{2} )\Delta _F (x,\xi )=-\delta (x-\xi ),
\label{47}
\end{equation}
where the differentiation is over $\xi $. This equation is equivalent to Eq.~(\ref{44}), where $D^{\mu } $ acts on $x$.
The superposition scheme for the retarded propagator is as follows
\begin{equation}
\phi (x)\theta (x^{0} -y^{0} )=-\int d{\mathbf y}
W_{A} [\Delta _{\mathrm{ret}} (x,y),\phi (y)] .
\label{48}
\end{equation}
This equation is the analog of Eq.~(\ref{17}). It can also be derived from Eq.~(\ref{45}).
To conclude, the superposition schemes for a free scalar field are fundamentally valid for a scalar complex field in an external electromagnetic field.
\section{Nonlinear field theory}
\renewcommand{\theequation}{V.\arabic{equation}}
\setcounter{equation}{0}
The superposition principle for secondary waves, which is the consequence of the GF method, should be distinguished from the superposition principle
as a manifestation of the linearity of the problem.
In linear theory, the wave is a source of secondary waves.
In nonlinear theory, two sources of secondary waves exist: the wave itself plus a function $V^{\prime}(\phi)$. In both cases, secondary waves satisfy free linear wave equations, so the superposition principle applies to secondary waves universally.
\subsection{Superposition principle from Kirchhoff's integral theorem}
For a Lagrangian ${\rm {\mathcal L}}={\rm {\mathcal L}}_{{\rm free}} - V$, that contains a term $V=V(\phi )$ of a general form, the identity (\ref{30}) is modified as follows:
\begin{eqnarray}
\phi(\xi )\delta^{4} (\xi - x) &=& \Delta_{F} (x-\xi )
((\Box_{\xi } + m^{2} )
\phi(\xi )+ V'(\phi (\xi )))
- ((\Box_{\xi } + m^{2} )
\Delta_{F} (x-\xi ))\phi (\xi ) \nonumber \\
&=& \frac{\partial }{\partial \xi _{\mu}}
\left(
\Delta _{F} (x - \xi )
\frac{\stackrel{\leftrightarrow}{\partial }}{\partial \xi ^{\mu } } \phi (\xi )
\right)
+\Delta _{F} (x-\xi )V'(\phi (\xi )).
\end{eqnarray}
For $x^{0} \in (y^{0} ,z^{0} )$, this equation gives
\begin{equation}
\phi (x)=\phi _{0} (x)-\int d^{4} \xi \Delta _{F} (x-\xi )V'(\phi (\xi )),
\label{49}
\end{equation}
where the integration over $\xi ^{0} $ runs over $\xi ^{0} \in (y^{0},z^{0})$ and the integral in $\boldsymbol\xi$ extends over all space. The field $\phi _{0} (x)$ is defined by the relation
\begin{equation}
\phi _{0} (x)=\int d{\mathbf z}W[\Delta _{F} (x-z),\phi (z)] -\int d{\mathbf y}W[\Delta _{F} (x-y),\phi (y)] .
\label{50}
\end{equation}
For $x^{0} \in (y^{0} ,z^{0} )$, $\phi _{0} (x)$ satisfies the free Klein-Gordon equation. If $x^{0} \notin (y^{0} ,z^{0} )$, then
\begin{equation}
0=\phi_{0}(x) - \int d^{4} \xi \Delta _{F} (x-\xi ) V'(\phi (\xi )).
\label{51}
\end{equation}
In quantum field theory, Eq.~(\ref{49}) in the infinite limits $(y^{0} ,z^{0} )=(-\infty ,+\infty )$ is used in the development of perturbation theory. Unlike in the canonical formulation of the Fresnel superposition scheme, the integrand contains the nonlinear term $V'(\phi (\xi ))$ as an additional source of secondary waves and the integration spans the entire four-dimensional space.
Equations (\ref{49}) - (\ref{51}) in nonlinear scalar field theory are analogous to Eqs.~(\ref{22}) - (\ref{24}) in the anharmonic oscillator problem.
The mass term of ${\rm {\mathcal L}}$ can be attributed either to
${\rm {\mathcal L}}_{{\rm free}}$ or to the potential $V$.
In the last case, ${\rm {\mathcal L}}_{{\rm free}}$ describes massless particles.
This might seem disadvantageous, because asymptotic states of ${\rm {\mathcal L}}$ are massive in general.
The positive feature is that the retarded Green's function of massless particles,
being localized on the light cone (see Eq.~(\ref{A1})),
ensures reduction of four-dimensional integrals in Eqs.~(\ref{49}) and (\ref{51}) to three-dimensional integrals
and transformation of integrals in Eq.~(\ref{50}) to surface integrals.
\subsection{Positive- and negative-frequency solutions}
Interacting fields can be decomposed into a sum of positive- and negative-frequency solutions only asymptotically.
In Sect. II.C.2, we demonstrated that
the straightforward generalization of the Fresnel superposition scheme to nonlinear dynamical systems
is possible and consistent; however,
its value is limited to only providing the definitions of positive- and negative-frequency solutions for arbitrary $t$. For the sake of completeness, we present here a field theoretical version of the nonlinear superposition scheme (\ref{26}):
\begin{eqnarray}
\phi^{(+)}(x)\theta ( x^{0} - y^{0})
- \phi^{(-)}(x)\theta (- x^{0} + y^{0})
= &-&\int d\mathbf{y}W[\Delta _{F}(x - y),\phi (y)] \nonumber \\
&+&\int d^4 \xi \Delta _{F}(x - \xi)V'(\phi(\xi)),
\label{261}
\end{eqnarray}
where the integral over $ \xi^0 $ runs from $y^0$ to $x^0$.
The derivative over $y^{0}$ leads to the relation $\phi (t)=\phi^{(+)} (t)+\phi^{(-)} (t)$ and Eq.~(\ref{29}). The difference in Eq.~(\ref{261}) at two different time points leads to Eqs.~(\ref{49}) - (\ref{51}). Equation (\ref{261}) ensures that
$\phi^{(\pm )} (x)$ is a linear superposition in $\mathbf{k}$ of the basis functions $f^{(\pm )}_{\mathbf k} (x)$ at $t\to \pm \infty $.
The calculation of the first derivative of Eq.~(\ref{261}) in $x^{0} $ yields a superposition scheme for the canonical momentum:
\begin{eqnarray}
\pi^{(+)}(x)\theta ( x^{0} - y^{0})
- \pi^{(-)}(x)\theta (- x^{0} + y^{0})
=& -&\int d\mathbf{y}W[\Delta _{F}(x - y),\pi (y)] \nonumber \\
&+&\int d^4\xi \Delta _{F}(x - \xi)V''(\phi(\xi))\pi(\xi),
\label{262}
\end{eqnarray}
where the integral over $\xi^0$ runs from $y^0$ to $x^0$.
\section{Conclusions}
\renewcommand{\theequation}{6.\arabic{equation}}
\setcounter{equation}{0}
The evolution of the ideas underlying the Huygens-Fresnel superposition principle from geometrical and wave optics to the theory of interacting fields is highly instructive:
In geometrical optics, a \textit{wave front} refers to the two-dimensional surface that defines the farthest extent to which the wave has arrived after a certain period of time. Huygens' principle (1678), based on the Fermat principle, allows for the determination of how the wave front is propagating.
In wave optics, the term wave front has no strict definition. Instead, the term \textit{wave surface} is used. The wave surface is the two-dimensional surface on which the phase of the wave is constant.
A.-J. Fresnel proposed the principle of superposition (1816), which details the wave process. A wave is a
result of interference of secondary waves
emitted at an earlier time. At any fixed point, it is determined by the phase and amplitude at a wave surface corresponding to a preceding instant of time. The wave surface in the past can be chosen arbitrarily. The superposition principle anticipates informal content of the GF method (1828).
Kirchhoff's integral theorem (1883) is a dynamic, four-dimensional extension of Green's third identity of the static potential theory. More than half a century separates this theorem from Green's major work \cite{GGreen1828}, which introduced
the basic concepts of the GF method.
\footnote{In 1839 year G. Green came closely to the notion of the four-dimensional Green's function. The value of the GF method in quantum field theory is highly appreciated. \cite{Schw93}}
Kirchhoff's integral theorem provides a mathematical proof of the superposition principle, clarifying and quantifying it.
First, the theorem demonstrates that the amplitude of the secondary waves is determined by the Wronskian of the Green's function and the field at a previous time.
Second, the wave surfaces are not highlighted; this is perfectly consistent with the fact that they are not necessarily observable (in the massive theory, e.g., the speed of a wave surface of a plane wave is always greater than the speed of light). The surface must be closed and contain the point at which the wave is calculated; otherwise, it can be arbitrarily chosen. Outside the closed surface, the interference of the secondary waves is strictly destructive: for any exterior point, the calculation of the surface integral yields zero.
The reasoning used in the proof
can be regarded as a standard piece of the GF method;
it is of high generality, goes beyond the problem of propagation of electromagnetic waves and allows for an understanding of how the superposition principle should be modified in the theory of interacting fields. Note the most significant modifications:
i)~~According to the Huygens-Fresnel superposition principle, a wave at a given point is expressed as a superposition of secondary waves emitted from centers located on a two-dimensional surface. This property arises only in massless theories, including the theory of electromagnetic fields, where the group and phase velocities coincide with the speed of light, which is the necessary condition for the integral over time delay and advance be not available in Eq.~(\ref{40}). Kirchhoff's integral theorem for massive particles, Eq.~(\ref{38}), states that a wave is determined by its values on a closed surface at all times. The physical interpretation of this fact is quite transparent. The Fourier expansion of a massive field contains components of various momenta corresponding to various group and phase velocities, which leads to a spread in time lags. As a result, the two-dimensional integral over the sources of secondary waves is transformed into a three-dimensional integral.
ii)~~In the nonlinear theory,
there is a need for a more extensive modification of the superposition scheme. In addition to the wave itself,
a nonlinear function of the field $V'(\phi (\xi ))$ becomes the source of secondary waves. The summation runs over distributed sources: from a two-dimensional surface in theories with massless particles to a two-dimensional surface and the time axis in theories with massive particles and the entirety of four-dimensional space. This type of representation holds for both local nonlinear and nonlocal theories.
We see that after each modification, the effectiveness of the superposition principle weakens. In the most general nonlinear case, the modified principle certainly does not promise fast
results. To determine the field, it is necessary to calculate a four-dimensional integral in a self-consistent manner. In linear theories, the superposition principle solves the evolution problem, but in nonlinear theories, it only offers a different formulation of the problem. Nevertheless, relations of this type are still useful when searching for solutions within the framework of perturbation theory, when the non-linearity is small. In other cases, the solutions found using other techniques can be checked. The four-dimensional representation given by Eq.~(\ref{49}) is a consequence of Kirchhoff's integral theorem, but in quantum field theory, it is typically derived directly from the properties of Green's functions.
In the context of a field theory,
the original form of
the superposition principle only has
heuristic value.
The superposition schemes for the secondary waves that are used to solve specific problems are unified by Kirchhoff's integral theorem, which exploits the properties of the Wronskian of the Green's functions and solutions of the wave equation under consideration. The spectrum of such problems is quite comprehensive: from the harmonic oscillator to scalar electrodynamics and nonlinear field theories.
In addition to the use of the GF method, which has found a variety of applications in quantum field theory, Kirchhoff's theorem has a wider range of corollaries. Equation (\ref{38}), which represents one version of Kirchhoff's theorem, does not arise in quantum field theory because of the boundary conditions, which are atypical of a scattering problem. However, the superposition scheme represented by Eq. (\ref{35}), which is not based on Kirchhoff's theorem, is not sufficiently general because it does not extend to theories with interaction.
The statement of Kirchhoff's integral theorem depends on the asymptotic conditions imposed on the Green's function.
In the main part of the paper, because we were interested in the place of this remarkable principle and well-known theorem in quantum field theory, we applied the Feynman asymptotic conditions almost universally.
\begin{acknowledgments}
This work was supported in part by RFBR Grant No.~16-02-01104 and Grant~No.~HLP-2015-18~of Heisenberg-Landau~Program.
\end{acknowledgments}
| 4c72394f76db8b28831545e7d50af5786642e6c2 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:introduction}}
\section{Introduction}
\label{sec:introduction}
\IEEEPARstart{S}{oftware} testing \cite{ammann2008introduction} is an important activity used for verification and validation by observing the software, executed using a set of test inputs. In practice, engineers are creating these inputs based on different test goals and test design techniques (e.g., specification-based, random, combinatorial, code coverage-based). These techniques have so far been performed manually or semi-automatically with respect to distinct software development activities (i.e, unit and integration testing). With the emerging use of large complex software products, the traditional way of testing software has changed; engineers need to deliver high-quality software while devoting less time for properly testing the software.
In practice, test suites are still created manually by handcrafting them using specific test design techniques and domain-specific experience. Although the automatic or semi-automatic creation of test suites has been the focus of a great deal of research, manual testing is still widely used \cite{andersson2002verification,beer2008role} in the software development industry. However, over the past few decades, several test design techniques \cite{ammann2008introduction} have been proposed for the creation of test suites with less effort. Combinatorial testing \cite{cohen1996combinatorial} is a technique that creates test inputs based on combinations among the input values. Pairwise testing is an approach to combinatorial testing that generates a test suite which covers each combination of value pairs at least once. There is some evidence\cite{borazjany2012combinatorial, hagar2015introducing} suggesting that pairwise testing is efficient and effective at detecting software faults. However, even if pairwise techniques have been found useful and applicable in industrial applications, the experimental evidence regarding its effectiveness in practice is still limited.
In this paper we compare pairwise testing and manual testing performed by industrial engineers on industrial software created using the {{\sc IEC 61131-3}} programming language \cite{international3iec} that runs on Programmable Logic Controllers (PLCs). The paper makes the following contributions:
\begin{itemize}
\item Empirical evidence showing that pairwise testing achieves marginally lower levels of code coverage while in the same time using more tests cases on average than manual testing performed by industrial engineers.
\item Results showing that manual testing is not significantly better at finding faults than pairwise testing. Our paper suggests that pairwise testing is just as good in fault detection as manual testing for 64\% of the programs considered.
\item A discussion of the implications of these results for test engineers and researchers.
\end{itemize}
\section{Background}
\label{sec:background}
This paper describes a case study evaluating pairwise testing when used on PLC industrial programs implemented in the IEC 61131-3 FBD language. In this section, we provide a background on PLC industrial software and pairwise testing. According to Ammann and Offutt \cite{ammann2008introduction}, a test case is a set of inputs, expected outputs and actual outputs executed on the specified program. A test suite is a set of ordered test cases. Throughout the paper, we will use the terms test case and test suite in this way.
\subsection{Programmable Logic Controllers}
\label{subsec:plc}
A Programmable Logic Controller (PLC) is a computer system containing a processor, a memory, and a communication bus. PLCs \cite{lewis1998programming} have a programmable memory for storing the software used for expressing logical behaviour, timing and input/output control, networking and data processing. Safety-critical industrial systems implemented using PLCs are used in many applications \cite{parnas1990evaluation} such as transportation, robotics, nuclear and pharmaceutical. Software running on a PLC execute in a loop called scan cycle, in which the iteration follows the βread-execute-writeβ semantics. The PLC reads the input signals, computes the logical behaviour without interruption and updates its output signals \cite{donandt1989improving}.
\begin{figure}
\includegraphics[width=\columnwidth]{figures/new_fbd.pdf}
\caption{A PLC program with seven inputs and three outputs written
using the FBD IEC 61131-3 programming language.}
\label{figure:fbd}
\end{figure}
Programming a PLC differs from general-purpose computers; the PLC software follows a standardized programming paradigm: the IEC (International Electronical Commission) 61131-3 standard \cite{international3iec}. IEC 61131-3 is a popular programming language standard for PLCs used in industrial practice. As shown in Figure \ref{figure:fbd}, computational blocks in an IEC 61131-3 program can be represented in a Function Block Diagram (FBD). This program contains
predefined logical and/or stateful blocks (i.e., OR, RS, TOF, GE, AND and TON in Figure \ref{figure:fbd}) and signals (i.e., connections) between blocks representing the whole behavior of an FBD program. PLC software contains a particular type of blocks named timers that are used to activate or deactivate an output signal after a specific time interval ~\cite{lewis1998programming}. A timer block (e.g., TON and TOF in Figure \ref{figure:fbd}) keeps track of the number of times its input is either true or false and outputs different signals. The IEC 61131-3 standard contains four other programming languages: Instruction List (IL), Structured Text (ST), Ladder Diagram (LD) and Sequential Function Chart (SFC)~\cite{bolton2015programmable}. For more details on PLC programming and FBDs we refer the reader to the work of John et al. \cite{john2010iec}.
\subsection{Pairwise Testing}
\label{subsec:auto}
The process of test generation is that of finding suitable test inputs
using a certain test goal that guides the search in an algorithmic way \cite{ammann2008introduction}. Many algorithms and techniques \cite{orso2014software} for test generation have been proposed. One such technique is combinatorial testing which is used to reveal faults
caused by interactions between input parameters inside a software program. Such techniques design test cases by combining different input parameters based on a combinatorial strategy. Grindal et al. \cite{grindal2006evaluation} surveyed several strategies used for combinatorial testing (e.g., each-used, pair-wise, t-wise, base choice). One of the most commonly used strategy is pairwise (also known as two-way) testing in which each combination for all possible pairs of input parameters are covered by at least one test case. Several empirical studies \cite{borazjany2012combinatorial, hagar2015introducing,li2016applying,grindal2006evaluation} on the use of combinatorial testing for industrial software have been reported and showed that pairwise testing is a very effective technique. In this paper we seek to investigate the use of pairwise testing for industrial control software and compare this technique with manual testing performed by industrial engineers.
\section{Related Work}
Most studies concerning pairwise testing and related to the work included in this paper have focused on how to generate tests as quickly as possible, measure the code coverage score and/or compare with other combinatorial criteria \cite{grindal2006evaluation} or with random tests \cite{ghandehari2014empirical,schroeder2004comparing}. For example, Cohen et al.~\cite{cohen1996combinatorial} found that pairwise generated test suites can achieve 90\% block code coverage. These test suites where generated by the AETG tool. The same tool was used by Burr and Young~\cite{burr1998combinatorial} in a different study. In this paper, pairwise testing achieved 93\% block coverage on average. In addition, Vilkomir and Anderson~\cite{vilkomir2015relationship} showed that pairwise test suites could achieve 77\% MC/DC code coverage.
Other studies \cite{cohen1994automatic,dalal1998model,sampath2012improving} have reported the use of pairwise testing on real systems and how it can help in the detection of additional bugs when compared to standard test techniques. On the other hand, a few other studies compared manual with pairwise testing \cite{ellims2008effectiveness,ellims2007aetg} and the results suggest that pairwise testing is not able to detect more faults than manually created tests. These results encouraged our interest in investigating on a larger case study how manual testing performed by industrial engineers compares to pairwise testing for industrial software systems. Is there any compelling evidence on how pairwise test suites compare with rigorously handcrafted test suites in terms of test effectiveness?
\begin{figure*}[!ht]
\vspace{-30mm}
\centering
\includegraphics[width=0.65\linewidth]{figures/new_design.pdf}
\caption[Methodology overview]{Overview of the experimental method used to perform the case study. For each program in the Train Control Management System (TCMS), test suites are created manually by industrial engineers, generated by {\ctt} for pairwise testing, and executed on both the original and the mutated programs in order to collect scores for code coverage, mutation and number of tests.}
\label{figure:methodology}
\end{figure*}
\section{Method}
\label{sec:method}
The goal of this paper is to study the comparison between manual test suites created by industrial engineers and automatically generated test suites using a generation tool for pairwise testing in terms of efficiency and effectiveness of testing. To achieve this goal, we designed a case study (mirrored in Figure \ref{figure:methodology}) using industrial software programs from an already developed train control management system to answer the following research questions:
\begin{itemize}
\item {\it RQ1: Are pairwise generated test suites able to cover more code than test suites manually created by industrial engineers?}
\item {\it RQ2: Are pairwise generated test suites able to detect more faults than test suites manually created by industrial engineers?}
\item {\it RQ3: Is the size of pairwise generated test suites smaller than those manually created by industrial engineers?}
\end{itemize}
For each selected program, we executed the test suites produced by both manual testing and pairwise testing and collected the following measures: branch coverage in terms of achieved code coverage, the number of generated test cases and the mutation score as a proxy for fault detection. In order to calculate the mutation score, each test suite was executed on the mutated versions of the original program to determine whether it detects the injected fault.
This section describes the design of our case study, including the subject programs, the evaluation metrics and the test generation and selection.
\subsection{Subject Programs}
\label{subsec:sut}
Our case study uses an industrial safety-critical system developed by Bombardier Transportation Sweden AB, a large-scale company developing and manufacturing railway equipment. The system is a train control management software ({\sc TCMS}) that has been in development for several years and is tested according to safety standards and regulations. TCMS is a control system containing both software and hardware components in charge of the safety-related functionality of the train and is used by Bombardier Transportation Sweden AB for the control and communication functions in high speed trains. These functions are developed as software programs for PLCs using the Function Block Diagram (FBD) {{\sc IEC 61131-3}} graphical programming language \cite{international3iec}. Programs in TCMS are developed in a graphical development environment, compiled into PLC code and saved in standardized PLCOpen XML \footnote{http://www.plcopen.org/} containing structural and behavioural declarations.
We selected the subject programs for our case study by investigating the TCMS programs provided by Bombardier Transportation Sweden AB. We identified 53 programs and excluded eight programs due to the following reasons: one program contained only one input parameter, for another program the test generation got a memory exception, while for the six remaining programs the test execution failed due to wrong parameter ranges that resulted in an execution exception. Our final set of subjects contains 45 programs. These programs contain nine input parameters and 1076 Lines of XML Source Code (LOC) on average per program. The studied programs were already thoroughly manually tested and are currently used in an operational train.
\subsection{Test Case Creation}
\label{subsec:testcase}
We used manual test suites created by industrial engineers working at Bombardier Transportation Sweden AB. These manual test suites were obtained by using a post-mortem analysis of the data provided. In testing these programs, engineers perform testing according to specific safety standards and regulations. Specification-based testing is used by engineers to manually create test suites as this is mandated by the EN50128 standard \cite{en200150128}. The test suites collected in this study were based on functional requirement specifications written in a natural language.
In addition, we generate pairwise test suites using {\ctt} \cite{peter_2017_439253}. {\ctt} is the only available combinatorial test suite generation tool for {{\sc IEC 61131-3}} control software. {\ctt} is open source
software and is available at https://github.com/CharByte/SEAFOX \footnote{For more details on the {\ctt} tool we refer the reader to the work of Charbachi and Eklund \cite{thesischa}.}.
{\ctt} supports the generation of test suites using pairwise, base choice and random strategies. For pairwise generation, {\ctt} uses the IPOG algorithm as well as a first pick tie breaker\cite{lei2008ipog}. {\ctt} was used in this study as it supports as input a standard PLCOpen XML implementation of the programs.
A developer using {\ctt} can automatically generate test suites needed for a given {{\sc IEC 61131-3}} program after manually providing the input parameter range information based on the defined behaviour written in the specification.
In order to collect realistic data, we asked one test engineer from Bombardier Transportation, responsible for testing {{\sc IEC 61131-3}} software used in this study, to identify the range values for each input parameters and constraints. We used these predetermined input parameter ranges for each program variable for generating pairwise test suites using {\ctt} in order to maintain the same input model as the one used to create manual test suites.
\subsection{Evaluation Measurements}
\label{subsec:measurements}
In this section, we present how the case study is conducted with respect to each research question. We first discuss the evaluation measurements used for efficiency and effectiveness of testing.
\subsubsection*{Code Coverage} We use code coverage criteria to assess the test suites thoroughness \cite{ammann2008introduction} and answer RQ1. These coverage criteria are used to evaluate the extent to which the program has been exercised by a certain test suite. In this study, code coverage is directly measured using the branch coverage criterion. For the programs selected in this study the EN50128 safety standard \cite{en200150128} involves achieving high branch coverage. A test suite achieves 100\% branch coverage if executing the program causes each branch in the {{\sc IEC 61131-3}} program to have the value \emph{true} and \emph{false} at least once. A branch coverage score was obtained for each generated test suite using our own tool implementation based on the PLC execution framework provided by Bombardier Transportation Sweden AB.
\subsubsection*{Fault Detection} Ideally, in order to measure fault detection, real faulty versions of the programs are required. In our case, the data provided did not contain any information about what faults occurred during the testing of these programs. To overcome this issue and answer RQ2, we used mutation analysis by generating faulty versions of the original programs. Mutation analysis is a method of automatically creating artificial faulty versions of a program in order to examine the fault detection ability of a test suite \cite{ammann2008introduction}. A \textit{mutant} is a different version of the original program containing a small syntactical change. For example, in an {{\sc IEC 61131-3}} program, a mutant is created by replacing a constant value with another one, negating a signal or changing the type of a computational block. If the execution of the resulting mutant on a test is producing a different output as the execution of the original program, the test suite \textit{detects} the mutant. The mutation score is computed using an output-only verdict (i.e., using the expected values for all of the program outputs) against the set of mutants. The fault detection capability of each test suite was calculated as the ratio of mutants detected to the total number of mutants. Just et al. \cite{just2014mutants} provided compelling experimental evidence that the mutation score is a proxy for real fault detection.
In the creation of mutants we used common type of faults in {{\sc IEC 61131-3}} software \cite{oh2005software} as a basis for establishing the following mutation operators:
\begin{itemize}
\item {\it Logic Block Replacement}. Replacing a logical block with another block from the same category (e.g., an OR block is replaced by an AND block).
\item {\it Comparison Block Replacement}. Replacing a comparison block with another block from the same category (e.g., a Greater-Or-Equal (GE) block is replaced by a Greater-Than (GT) block).
\item {\it Arithmetic Block Replacement}. Replacing an arithmetic block with another block from the same category (e.g., replacing a maximum calculation block (MAX) with a minimum calculation block (MIN)).
\item {\it Negation Insertion}. Negating an input or output connection (e.g., an output boolean connection is negated).
\item {\it Value Replacement}. Replacing a value of a constant variable connected to a block (e.g., a constant variable is replaced by its boundary values).
\item {\it Timer Block Replacement}. Replacing a timer block with another block from the same function category (e.g., a Timer-On (TON) block is replaced by a Timer-Off (TOF) block).
\end{itemize}
Each of the mutation operators were applied to each program element. In total, for all of the selected programs, 1597 mutants were generated (i.e., 35 mutants on average per program). A mutant was considered detected by a test suite if the output from the mutated program differed from that of the original program. A mutation score was obtained for each generated test using our own tool implementation.
\begin{figure*}[tbp]
\centering\resizebox{\textwidth}{!}{
\includegraphics[width=0.3\textwidth]{figures/code_coverage_noRand}
\includegraphics[width=0.3\textwidth]{figures/mutation_score_noRand}
\includegraphics[width=0.3\textwidth]{figures/size_testsuite_log}
}
\vspace*{-1cm}
\caption[Box plots]{Code Coverage, mutation score and number of test cases comparison between manually handcrafted test suites (Manual) and test suites generated using pairwise testing (Pairwise); boxes span from 1st to 3rd quartile with black middle lines marking the median and the whiskers extending up to 1.5x the inter-quartile range; the circle symbols represent the outliers.}
\label{figure:boxplots}
\end{figure*}
\begin{figure*}[tbp]
\centering
\resizebox{\textwidth}{!}{
\includegraphics[height=4cm]{figures/pie_code}
\hspace{0.7cm}
\includegraphics[height=4cm]{figures/pie_mutation}
\hspace{0.8cm}
\includegraphics[height=4cm]{figures/pie_tests}
}
\caption[Manual and Pairwise test suite comparison]{Charts of number of tests, code coverage and mutation scores depicting how each technique (M stands for manual testing and P is short for pairwise testing) compared to the other for each measure. M$>$P indicates that manual testing achieved a higher measurement score, M$<$P shows how many times manual testing achieved a lower measurement score and M$=$P shows when both techniques achieve the same measurement score.
\label{figure:caseEval}
\end{figure*}
\subsubsection*{Number of Tests} In an ideal situation, the cost of testing is measured by taking into account direct and indirect type of cost, by measuring directly the test suite creation, the test suite execution and the checking of test suite results. However, since this is a case study using programs for which the development was performed a few years back, this kind of cost data was not available. To answer RQ3, we used the number of created test cases as a proxy for efficiency as we assume that all human costs are depended on the number of tests. The higher the number of tests, the higher are the respective testing costs. For example, a complex program will require more effort for test creation, execution and checking of the results.
\section{Experimental Results}
\label{sec:results}
In this section, we quantitatively answer the three research questions posed in Section \ref{sec:method}. We collected the data to answer these research questions by generating test suites for pairwise testing using {\ctt}; collecting manual test suites created by experienced industrial engineers; measuring their code coverage; and measuring their effectiveness in terms of mutation score. The overall results of this study are summarized in the form of boxplots in Figure \ref{figure:boxplots}. In Tables \ref{table:codeCoverage}, \ref{table:mutationScore} and \ref{table:testsuiteSize} we present the code coverage scores, mutation scores and the number of tests in each test suite by listing the standard deviation, mean, median, minimum and maximum values. In addition, statistical analysis was performed using the R statistical software \footnote{https://www.r-project.org/}. We assume that the collected data is drawn from an unknown distribution. In order to evaluate if there is any statistical difference between manual and pairwise testing we use a Wilcoxon-Man-Whitney U-test\cite{howell2012statistical}, a non-parametric hypothesis test used for checking if two data samples are randomly obtained from identical populations. We also use the Vargha-Delaney test (also known as the standardized effect size) to calculate the statistical significance. The Vargha-Delaney $\hat{A}$-measure is also "\textit{a measure of stochastic superiority}"\cite{vargha2000critique} and is used to measure the difference between two populations. The test result is denoted as $\hat{A}$, and simply specifies the amount of times population A is expected to be better than population B\cite{neumann2015transformed}. Its significance is determined when the effect size is above 0.7 or below 0.2.
\subsection{Code Coverage}
\label{subsec:code}
RQ1 asked if pairwise testing achieves better code coverage scores than manual testing. The coverage scores achieved by pairwise testing are ranging between 50\% and 100\% while for manual testing these are varying between 63\% and 100\%. As shown in Table \ref{table:codeCoverage}, the use of manual testing achieves on average 97\% branch coverage (3\% on average higher than pairwise testing). Results for all programs (in Table \ref{table:statisticalAnalysis}) show that differences in code coverage between manual and pairwise testing are statistically significant with a p-value of 0.04 but their effect is not strong (i.e., an effect size of 0.6).
\input{tables/codeCoverage}
As seen in Figure \ref{figure:caseEval}, for 62\% of the programs considered, pairwise performs equally good as manual testing; for 29\% of the programs manual testing performed better in terms of achieved code coverage while for 9\% of the programs pairwise testing covers more code than manual testing.
The results for all programs were surprising: test suites created using pairwise testing achieved relatively high code coverage (94\% on average). This shows that, for the programs studied in this experiment, pairwise testing achieves high branch coverage. This is likely due to the complexity of the studied programs. It is possible that more complex software would yield a greater code coverage difference between manual and pairwise test suites.
Overall, as shown in Figure \ref{figure:caseEval}, we confirm that pairwise test suites achieve just as good or better code coverage scores as manual testing for 71\% of programs considered in this study. This can be explained by the fact that pairwise testing if properly used is quite good at covering the logical behaviour of the code.
\begin{mdframed}[style=style1]
{\it Answer RQ1: Code coverage scores achieved by pairwise test suites are slightly lower than the ones created manually by industrial engineers.}
\end{mdframed}
\input{tables/mutationScore}
\subsection{Fault Detection}
\label{subsec:fault}
To answer RQ2, we first computed the mutation score of each manual and pairwise test suites. Figure \ref{figure:boxplots} shows box plots of our results for fault detection in terms of mutation score. Table \ref{table:mutationScore} summarizes statistics for these test suites. For all programs the fault detection scores obtained by manually written test suites are higher on average with 7\% than those achieved by pairwise testing. However, there is no statistically significant difference at 0.05; as the p-value is 0.67 and the effect size is 0.53 in Table \ref{table:statisticalAnalysis}. A larger sample size would be needed to obtain more confidence in our results. Interestingly, as show in Figure \ref{figure:caseEval}, our results suggest that fault detection scores achieved by manual testing are not significantly better at finding faults than pairwise testing. It seems that test suites generated using pairwise testing are just as good in terms of fault detection as manual test suites for 64\% of the cases considered in this study. For 42\% of the programs, pairwise testing performs as well as manual testing while for 36\% of the programs manual testing performed better in terms of fault detection.
The difference in effectiveness between manual and pairwise could be due to other factors such as the number of test cases and the test design techniques used to manually create test suites (e.g., testing the timed behavior of the PLC software).
\begin{mdframed}[style=style1]
{\it Answer RQ2: Pairwise testing is able to produce comparable results to manual testing in terms of fault detection. However, manual testing produced better mutation scores on average.}
\end{mdframed}
\subsection{Number of Tests}
\label{subsec:size}
\input{tables/testsuiteSize}
This section aims to answer RQ3 regarding the relative cost of performing manual testing versus pairwise testing. Analysing the cost in this study is directly related to the number of test cases giving a picture of the effort needed per created test suite. Based on the results highlighted in Figure \ref{figure:boxplots}, the use of pairwise testing results in very inconsistent number of tests created, compared to manual testing which seems to create tests with more diverse number of steps than pairwise testing. Examining Table \ref{table:testsuiteSize}, we see a different pattern: less number of tests on average are manually created by industrial engineers (13 test cases on average in a test suite) than when using pairwise testing (21 test cases on average in a test suite). As seen in Figure \ref{figure:caseEval}, for 44\% of the programs considered, pairwise test suites produced equally or larger test suites than the manual ones, leading to 55\% where pairwise produced fewer. Table \ref{table:statisticalAnalysis} shows an interesting pattern in the statistical analysis: the standardized effect size being 0.53, with p-value being higher than the traditional statistical significance limit of 0.05. Results are not strong in terms of effect size and we did not obtain any statistical difference for the number of tests.
The results for all programs matched our expectations: manual tests are handcrafted by experienced industrial engineers that can create very diverse tests. It is possible that more complex software would yield greater number of tests differences between tests written manually and pairwise testing.
\begin{mdframed}[style=style1]
{\it Answer RQ3: The use of pairwise testing results in more number of tests created on average than the use of manual testing. Even so, pairwise testing produced more cases with less tests created compared to manual testing, resulting in very inconsistent results.}
\end{mdframed}
\input{tables/statisticalAnalysis}
\section{Discussion}
\label{sec:discussion}
The goal of this work was to compare pairwise testing with manual testing performed by industrial engineers in terms of code coverage, fault detection and the number of created test cases. We found out that pairwise testing achieves high code coverage, but slightly lower scores than manual testing. In addition, we found the fault detection scores for pairwise testing to be lower on average than the ones written manually by industrial engineers. Interestingly, pairwise testing achieves equally good or better fault detection scores than manual testing for 64\% of the programs considered, which might indicate that for more than half of the programs the fault detection scores for pairwise testing are a good predictor of test effectiveness. This is reinforced by the mutation score box plot in Figure \ref{figure:boxplots}: the median mutation score for pairwise testing is 95\%. This in combination with the achieved high code coverage, suggests that pairwise testing can cover as much code as manual testing performed by industrial engineers, but this is not entirely reflected in its fault detection capability. The fault detection rate between manual and pairwise testing was found, in some of the published studies \cite{ellims2008effectiveness,ellims2007aetg} to be similar to our results. Interestingly enough, our results indicate that pairwise test suites might be even better in some cases in terms of fault detection than manual test suites. However, a larger pool of programs and tests is needed to statistically confirm this hypothesis.
The mean value for fault detection of 81\% for pairwise testing is right in line with the proportion of 2-way faults seen in other domains \cite{ kuhn20132combinatorial}. It is interesting to note here that the fault distribution for PLC industrial control software is similar to other types of software.
As part of our study, we used the number of tests to estimate the test efficiency in terms of creation, execution and result checking. While the cost of creating and executing a test for pairwise testing can be low compared to manual testing, the cost of checking the result is usually human intensive. Practically, the higher the number of test cases, the higher the cost of checking the test result. Our study suggests that pairwise test suites, while inconsistent, are longer on average in terms of created test steps (number of tests) than manual test suites. By considering generating optimized or shorter test suites, one could improve the cost of performing pairwise testing.
The idea of using pairwise testing in practice stands as a significant progress in the development of automatic test generation approaches. This progress implies, to some extent, that pairwise testing should be at least as effective and more efficient than manual testing for it to be considered ready to be used as a replacement to the manual effort of creating tests. The overall result, from this case study, is that pairwise testing alone is not better than manual testing. However, pairwise testing or stronger combinatorial criteria are capable of at least aiding an engineer in testing of industrial software. Our observations showed that experienced engineers are very effective at generating the right choice of values and considering the timing of the input parameters. Industrial PLC software typically have a complex and time-dependent behaviour. This behaviour require inputs to retain and change a sequence of inputs for some time in order to trigger a certain logical event. Several manual test suites collected in this case study contained that kind of test cases. The engineers creating these test suites had years of experience in developing and testing this type of software, including good knowledge of what combinatorial interactions are needed to cover the code and detect faults. As it turns out, pairwise testing is not particularly useful for some of the programs considered in this study. By considering generating more complex (stronger t-wise) and time-depended tests, one could improve both the achieved code coverage as well as the fault detection capability.
\section{Threats to Validity}
\label{sec:validity}
There are many tools (e.g., ACTS \cite{yu2013acts}, AETG\cite{cohen1997aetg}, TCG\cite{tung2000automating}) for automatically generating tests using pairwise testing and these may produce different results. The use of these tools in this study is complicated by the modelling of the input space for a PLC program. Hence, we choose a tool specifically developed for testing PLC industrial programs. For more details on the comparison between {\ctt} and ACTS, we refer the reader to the work of Charbachi and Eklund \cite{thesischa}. {\ctt} is using the IPOG algorithm. This algorithm continually expands the test suite to fit the input parameters with the use of vertical and horizontal extension. {\ctt} currently uses a first element tie breaker which was chosen when a clean-cut best choice for tie breaking was absent \cite{huang2014tie}. This choice of a tie breaker might affect the effectiveness of the generated tests.
Another threat to the validity of this study is also related to the use of the {\ctt} tool. Vertical extension \cite{lei2007ipog} is a step which adds new tests, if needed, to the test set produced by horizontal growth when there are no modifiable test cases to cover a specific pair. The IPOG algorithm, in this case, creates a new test case, thus expanding the size of the test suite by making the test case to cover the specific pair and keeping all the other parameters modifiable. This is done to reduce the need of further vertical extensions and can be used to avoid creating a new test case. However, when trying to execute a test suite, these values are syntactically illegal and need to be changed. {\ctt} currently handles this by randomising the modifiable values to a default option for each parameter. A more accurate option for each parameter would be needed to obtain more confidence in the test suite effectiveness.
The data collected is based on a study in one company using one industrial system containing 45 programs and manual test suites created by industrial engineers. We argue that even if the number of programs can be considered relatively small, reporting a case study using industrial artefacts can be representative.
Since this case study was performed post-mortem, the cost information was not available. We used the number of test cases as a valid proxy measure and a more detailed cost model should be used to obtain more accurate results.
\section{Conclusion}
\label{sec:conclusions}
In this paper, we studied the comparison between pairwise testing and manual testing in terms of branch coverage, mutation score and number of created test cases. From the 45 PLC industrial programs we studied, we drew the following conclusions:
\begin{itemize}
\item The use of pairwise testing results in high branch coverage and mutation scores for the majority of the programs considered.
\item The results of this paper support the claim that pairwise testing is not quite as effective (i.e., achieved branch coverage and fault detection) and efficient (in terms of number of tests created) as manual testing.
\item The use of pairwise testing results in similar or better fault detection scores than the use of manual testing for 64\% of the programs.
\end{itemize}
The results imply that pairwise testing can achieve high branch coverage, but slightly lower scores than manual testing.
In summary, our results suggests that pairwise testing can perform in some cases comparably with manual testing performed by industrial engineers. This is a significant experimental evidence on the progress of pairwise testing that needs to be further studied; we need to consider the cost of using pairwise testing in practice. In addition, pairwise testing is only one type of combination strategy and we would need to evaluate the use of stronger criteria such as t-wise (with $t>2$) testing.
\ifCLASSOPTIONcompsoc
\section*{Acknowledgments}
\else
\section*{Acknowledgment}
\fi
This research was supported by the Swedish Research Council (VR) through the βAdequacy-based testing of extra-functional properties of embedded systemsβ project and by VINNOVA and ECSEL (EU's Horizon 2020) under grant agreement No 737494. The authors would like to thank Bombardier Transportation AB for the valuable assistance in the planning and execution of this work.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
\balance
\bibliographystyle{IEEEtran}
\section{Introduction}
This demo file is intended to serve as a ``starter file''
for IEEE Computer Society conference papers produced under \LaTeX\ using
IEEEtran.cls version 1.8b and later.
I wish you the best of success.
\hfill mds
\hfill August 26, 2015
\subsection{Subsection Heading Here}
Subsection text here.
\subsubsection{Subsubsection Heading Here}
Subsubsection text here.
\section{Conclusion}
The conclusion goes here.
\ifCLASSOPTIONcompsoc
\section*{Acknowledgments}
\else
\section*{Acknowledgment}
\fi
The authors would like to thank...
\section{Introduction}\label{sec:introduction}}
\IEEEPARstart{T}{his} demo file is intended to serve as a ``starter file''
for IEEE Computer Society journal papers produced under \LaTeX\ using
IEEEtran.cls version 1.8b and later.
I wish you the best of success.
\hfill mds
\hfill August 26, 2015
\subsection{Subsection Heading Here}
Subsection text here.
\subsubsection{Subsubsection Heading Here}
Subsubsection text here.
\section{Conclusion}
The conclusion goes here.
\appendices
\section{Proof of the First Zonklar Equation}
Appendix one text goes here.
\section{}
Appendix two text goes here.
\ifCLASSOPTIONcompsoc
\section*{Acknowledgments}
\else
\section*{Acknowledgment}
\fi
The authors would like to thank...
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
| 32085c31855d106b2a82b0f657039e43b31b4f1f | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Method}
\subsection{Three-dimensional model of origami structures}
A mathematical model of each experimentally generated origami structure is formulated for numerical simulation and 3D reconstruction. The model is composed of the vertex coordinates and the constraints imposed by the crease length between neighboring vertices. Given a 2D projection of all the vertex positions ($x_i$, $y_i$), the $z_i$ coordinates in the third dimension are obtained by minimizing a penalty function \cite{Tachi:2013im} $V(z)=\frac{k}{2} \sum_{i,j} ' \left[\sqrt{(x_i-x_j)^2+(y_i-y_j)^2+(z_i-z_j)^2} - l_{ij}\right] ^2$, where $l_{ij}$ is the length of the crease between two connected vertices of indices $i$ and $j$ on the triangular crease lattice. The $z$ coordinates can thus be obtained by solving the associated ordinary differential equation (ODE) array as $\dot{z_i}=-k \sum_j ' \left[\sqrt{(x_i-x_j)^2+(y_i-y_j)^2+(z_i-z_j)^2} - l_{ij}\right] \hat{r}_{ij} \cdot \hat{z}$, where the unit vector $\hat{r}_{ij}=((x_i-x_j)\hat{x}+(y_i-y_j)\hat{y}+(z_i-z_j)\hat{z})/\sqrt{(x_i-x_j)^2+(y_i-y_j)^2+(z_i-z_j)^2}$, and $k$ is the virtual spring constant for perturbing the crease lengths. The ambiguity due to mountain-valley alignments is avoided by shifting the vertices up or down with a small amount in the $z$ direction as the initial condition.
\subsection{Configuration space analysis}
The configuration space of each vertex is described by the accessible folding angles of the connected creases. In the case of a degree-6 vertex, the configuration space is three dimensional (since DOF=3). Any potentially foldable structure is realized by rotating the creases through those 3 independent folding angles. The foldability is validated by the criterion that the distance between any pair of transformed vertices cannot be greater than that in the flat unfolded state for any inextensible sheet. Self-interceptions of the structures are also identified and excluded from the configuration space (see SI).
\subsection{Experiments}
The origami structures were made from a flat paper sheet (Stardream Metallics 81lb) with measured thickness $\tau=0.16$ mm and flexural rigidity $D=4.0 \times 10^{-4}$ N$\cdot$m \cite{Silverberg:2015gb}. The crease lines were perforated by a laser cutter and weakened manually by flat-folding along the perforation lines in both directions. Its folding kinematics were recorded by a top-view camera, which was calibrated for extracting the vertex locations. The $x-y$ coordinates of the vertices were fed into the aforementioned mathematical model of the origami to extract the $z$ depths of the vertices and the 3D configurations of the folded origami sheets were reconstructed using Matlab.
\section{Acknowledgments}
\begin{acknowledgments}
The authors thank Andy Ruina, Tim Healy, James Jenkins, Uyen Nguyen, Lea Freni, and the Cohen lab for useful discussions.
We also thank F. Parish for assistance with the laser cutter. This work was supported by the National Science Foundation Grant No.~EFRI ODISSEI-1240441. I.C. received continuing support from DMREF-1435829. C.D.S. thanks the KITP, Santa Barbara, for hospitality.
\end{acknowledgments}
| 142309cd9e2904b41b3740dd3e0ce155ffd8a234 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Purpose}
\label{sec:intro}
Volume calculation of lung lesions from Computed Tomography (CT) images plays a vital role in the evaluation of the level of lung disease. An efficient and accuracy method for volume measurement can provide reliable information for clinical diagnosis, which helps doctors to evaluate the state of patients and make appropriate therapies. Furthermore, doctors can estimate the timeliness and treatment effect of drugs with the help of the volume of lung lesions~\cite{Wang2014Direct}.
However, many previous studies about the volume calculation for lesions tissue or organ from CT or MRI images are based on the diameters of lung lesions which roughly expresses the size, or based on 3D model reconstruction and surface triangulation which are not efficient and sometimes fail to build a perfect 3D model due to the high complexity of medical images~\cite{Punithakumar2010Detection}.
Monte Carlo method (MC) is very useful tool for physical and mathematical problems, especially in optimization, numerical integration~\cite{Afshin2014Regional}. Quasi-Monte Carlo methods (qMC) (also called low-discrepancy sequences), which were proposed to generate random points with better uniformity and lower discrepancy, are often considered as a replacement of MC. Some qMC methods have already been used for computing the volume~\cite{Zhen2014Direct}.
In this paper, to accurately and directly calculate the volume of lung lesions, we introduce the low-discrepancy sequences to compute volume. In our study, the Halton low-discrepancy sequences are generated to calculate the area of each CT slice and the volume are obtained by accumulating these slices' area with a smooth formula. Experiments are conducted to evaluate our method compared with other methods.
\section{Method}
MC method is a simple and efficient numerical method based on random sampling. However, in \cite{davies1998low}, Davies \emph{et al.} pointed out that MC method is subject to errors and in particular all parts of space can not guarantee the equal sampling. Therefore, the sequence of Monte Carlo is not entirely uniform and it may lead to inaccuracy results. Fig \ref{fig:compare}(a) illustrates the distribution of pseudorandom number tends to be lumped.
To generate more uniform sequences, low-discrepancy sequences are proposed. Compared with MC method, quasi-Monte Carlo method solves numerical problems using low-discrepancy sequences. The distribution of the Halton sequence is more uniformly as shown in Fig \ref{fig:compare}(b). Low-discrepancy sequences also have lower statistical errors. Low-discrepancy sequences have theoretical error bound O($N^{-1}$$(logN)^{D}$) compared with pseudorandom sequences O($N^{-1/2}$) \cite{davies1998low}, where $N$ is the number of samples and $D$ denotes the dimensions. In this paper, we employ Halton sequences to generate quasi-random numbers. The Halton sequences are constructed according to a deterministic method that uses a prime number as its base. The sequence is implemented easily and has a very uniform distribution on low dimension.
\begin{figure}[H]
\centering
\subfigure[pseudorandom number sequence]{\includegraphics[height=1.7in,width=2.2in]{mcscatter.eps}}\hspace{0.7cm}
\subfigure[Halton sequence]{\includegraphics[height=1.7in,width=2.2in]{haltonscatter.eps}}
\caption{2-D plots of 1000 points generated respectively by (a) the pseudo-random number sequence, and (b) the Halton sequence.}\label{fig:compare}
\end{figure}
\textbf{Lesion Segmentation and Volume Calculation}. Our previous segmentation work \cite{wang20143d} is introduced to obtain the segmented results. The procedure of volume calculation has five steps: (1) building reference coordinate system in every slice of CT scanning image; (2) generating a certain number of random points in each slice using Halton low-discrepancy sequences; (3) finding out the number of random points who are inside the lung lesion region; (4) calculating the area ($S_i$) of the lung lesion region for each slice by the proportion of random points to all random points, the area can be calculated by ${S_i} \approx \frac{m}{N}{S_N}$ where $m$ means the number of points inside the lung lesion, $N$ represents the number of all points generated, and $S_N$ is the area of each slice; and (5) calculating the volume of lung lesion using the formula of the frustum model as:
${\rm{V}} = \sum\limits_{i = 0}^{z - 1} {\frac{{({S_i} + {S_{i + 1}} + \sqrt {{S_i}{S_{i + 1}}} ) \times h}}{3}}$, where $S_i$ and $S_{i+1}$ define the areas of lesions in the $i$th slice and the $(i+1)$th slice, and $h$ denotes the slice thickness. Fig \ref{fig:volumepoints} illustrates the random points generated by Halton sequence in 3D in the CT lung scans.
\begin{figure}[H]
\centering
\includegraphics[width=2in]{volume_point.eps}\\
\caption{The distribution of Halton sequence in the 3D CT lung scans.}\label{fig:volumepoints}
\end{figure}
\section{Experimental Results}
\label{sec:sections}
In this section, experiments are conducted on two kinds of CT data sets. The data was acquired at the 174 Hospital's CT room in Xiamen. The first data set is regular model, including cube, cuboid, and cylinder. See \ref{fig:RegularModels} for an illustration of the parameters of regular models, and we can calculate volumes of each model easily. Next, the CT scan data of each model was obtained using CT scanner in the CT room. And then the data can be used on our experiment. Table \ref{fig:regular} shows volumes of three regular models, which obviously demonstrates the result of three methods compared with ground truth volume. It has also shown that our method provides more accurate and precise results.
\begin{figure}[H]
\centering
\includegraphics[width=2in]{RegularModels.eps}\\
\caption{Illustration of regular models' parameters.}\label{fig:RegularModels}
\end{figure}
The second data set is manual tumor delineation of different size of lung lesions. Therefore, the data sets are divided into three groups: small, medium and large lung lesions. Each group has three data sets. The ground truth of each data set is provided by the experienced doctors who manually tag the CT lesion in each slice. The binary image of each slice is got using two value ($-4.0$ standing for background and $4.0$ indicating the lung lesion). This can guarantee the accurate segmentation of the lung lesion at the most extent. In order to evaluate our method, our results are compared with the Monte Carlo method and the method named LSTK using a discretized version of the divergence theorem described by Alyassin et al.\cite{alyassin1994evaluation}. Our experiments were run in Visual Studio 2010 using MicroSoft Windows 8.1 platform on a CPU of Intel 3.50GHz core i3-4150 with 4GB of RAM. In our proposed method, the dimension was set as $5$ to generate $5$ columns of low-discrepancy value and we chose the first and the fifth column (dimension) of the sequences.
We define the relative error $E$ as the evaluation criteria by the following formula:
\[E = \frac{{\left| {ComputedVolume - TrueVolume} \right|}}{{TrueVolume}}\]
\begin{table*}[ht]
\centering
\begin{tabular}{l|ccccccccc}
\hline
\hline
Dataset &True Volume &LSTK (-0.5) &MC (10000) &Ours (10000)\\ \hline
Cube &512.000 &516.356(0.85\%) &510.732(0.25\%) &512.790(\textbf{0.15\%})\\
Cuboid &231.000 &232.898(0.82\%) &230.177(0.36\%) &230.788(\textbf{0.09\%})\\
Cylinder &424.115 &427.033(0.69\%) &421.189(0.69\%) &425.106(\textbf{0.23\%})\\
\hline
\end{tabular}
\caption{The results of three regular models' volume calculation of three methods compared to ground truth volume. The value of each volume is in $cm^3$. MC and our method use 10000 random points. LSTK uses the iso-surface threshold $-0.5$ at its best. The relative error of each volume calculation is shown in brackets after the volume value.}
\label{fig:regular}
\end{table*}
Table \ref{fig:table} shows the results of volume calculation and its corresponding relative errors for all methods. It is demonstrated that the results of LSTK have a lower accuracy and MC method's results is unstable. However, our method achieve more accurate results greatly closed to the true volume at most cases and more stable than MC method.
We also analyze the accuracy of volume calculation when increase the random points for the MC method and our proposed method.
Fig.~\ref{fig:RegularPoints} illustrates the results of obtained of three regular models, the relative error versus the number of points used for volume measurement using the logs to base 10. For each graph, the theoretical error is plotted, which is compared to the actual relative error using our method. In the Fig.~\ref{fig:RE}, we compare the two methods among the three typical cases.
As is shown in the Fig.~\ref{fig:RegularPoints} and Fig.~\ref{fig:RE}. It is illustrated that our proposed method has lower relative errors in all cases compared with MC method. Because for our method, fewer number of points is supposed to achieve a given accuracy.
\begin{figure*}[ht]
\centering
\subfigure[cube]{\includegraphics[height=1in,width=2in]{cube_points.eps}}
\subfigure[cuboid]{\includegraphics[height=1in,width=2in]{cuboid_points.eps}}
\subfigure[cylinder]{\includegraphics[height=1in,width=2in]{cylinder_points.eps}}
\caption{The relative error analysis of calculating the volume of three regular objects for MC and our proposed method.}\label{fig:RegularPoints}
\end{figure*}
\begin{figure*}[ht]
\centering
\subfigure[small case (case 3)]{\includegraphics[height=1in,width=2in]{case3.eps}}
\subfigure[medium cases (case 6)]{\includegraphics[height=1in,width=2in]{case6.eps}}
\subfigure[large cases (case 9)]{\includegraphics[height=1in,width=2in]{case9.eps}}
\caption{The relative error analysis of calculating the lung lesion volume in three cases for MC and our proposed method.}\label{fig:RE}
\end{figure*}
\begin{table*}[ht]
\centering
\begin{tabular}{l|ccccccccc}
\hline
\hline
Dataset &True Volume &LSTK (-0.5) &MC (10000) &Ours (10000)\\ \hline
Small(case1) &425.182 &432.334(1.68\%) &417.224(1.87\%) &419.995(\textbf{1.22\%})\\
Small(case2) &1417.62 &1433.16(1.10\%) &1436.13(1.31\%) &1415.85(\textbf{0.13\%})\\
Small(case3) &1533.52 &1551.76(1.19\%) &1505.83(1.81\%) &1537.21(\textbf{0.24\%})\\
Medium(case4) &3444.63 &3489.74(1.31\%) &3518.6(2.15\%) &3415.86(\textbf{0.84\%})\\
Medium(case5) &5234.59 &5305.19(1.35\%) &5165.49(1.32\%) &5186.13(\textbf{0.93\%})\\
Medium(case6) &6820.19 &6898.35(1.15\%) &6836.36(0.24\%) &6782.16(\textbf{0.56\%})\\
Large(case7) &9277.13 &9352.53(0.81\%) &9296.87(0.21\%) &9260.66(\textbf{0.18\%})\\
Large(case8) &9415.15 &9533.36(1.26\%) &9179.48(2.50\%) &9338.65(\textbf{0.81\%})\\
Large(case9) &13634.1 &13758.1(0.91\%) &13635.1(\textbf{0.01\%}) &13577.4(0.42\%)\\
\hline
\end{tabular}
\caption{The results of volume calculation of three methods compared to ground truth volume. The value of each volume is in $mm^3$. MC and our method use 10000 random points. LSTK uses the iso-surface threshold $-0.5$ at its best. The relative error of each volume calculation is shown in brackets after the volume value.}
\label{fig:table}
\end{table*}
Comparing the MC method and ours, the graph above give a distinct analysis. Taking a different look, Table \ref{fig:table} shows the relative efficiency three regular models and nine cases for the two method. To achieve 1\% accuracy, the number of points approximately needed is shown in the table, and the ratio of MC method and our method is computed.
\begin{table*}[ht]
\centering
\begin{tabular}{l|ccccccccc}
\hline
\hline
Dataset &MC &Ours &Ratio \\ \hline
Cylinder &350 &100 &3.2 \\
Cuboid &720 &100 &7.2 \\
Cube &590 &100 &5.9 \\
Small(case1) &1040 &120 &8.7 \\
Small(case2) &6680 &160 &41.8 \\
Small(case3) &420 &260 &1.6 \\
Medium(case4) &1100 &330 &3.3 \\
Medium(case5) &370 &170 &2.2 \\
Medium(case6) &1350 &110 &12.3 \\
Large(case7) &8610 &110 &78.3 \\
Large(case8) &9110 &10 &911 \\
Large(case9) &230 &110 &2.1 \\
\hline
\end{tabular}
\caption{The number of points approximately needed to achieve an accuracy of 1\%.}
\label{fig:table}
\end{table*}
\section{Discussion and Conclusion}
In this study, the Halton low-discrepancy sequence is applied to calculate the volume for CT lung lesions. Compared with LSTK, the proposed method can calculate volume without reconstructing 3D model of lung lesions and surface triangulating. The proposed method can also generate more uniform random points compared with MC method. The experimental results are demonstrated that our proposed method can achieve more accurate results compared with LSTK and MC method and more stable than MC method.
In our near future study, we will improve the random points selection strategy for volume calculation since the more number of points selected, the more time costs in calculation.
\section*{Acknowledgement}
This work was supported by National Natural Science Foundation of China (Grant No. 61671399, 61327001), Research Fund for the Doctoral Program of Higher Education (20130121120045) and by the Fundamental Research Funds for the Central Universities (20720150110).
| 9fd06e6f119227a6b314178ad8f3616ad7f7d155 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Earth's atmosphere poses a persistent challenge to astronomers making ground-based observations at near infrared (NIR) wavelengths, where telluric water vapor absorption lines are numerous. Furthermore, the strengths of these absorption lines are variable in time and depend on the amount of precipitable water vapor (PWV) present in the atmospheric column above the observatory.
Mitigating the effects of water vapor absorption in astronomical data typically involves differential measurement techniques. While differential photometry removes first order effects of atmospheric variability, second order effects can be large depending on the spectral energy distributions of the target and reference stars. Because red stars incur more absorption than blue ones, a mismatch in the temperatures of the target star and the reference stars can still leave second-order photometric errors of over 1\% \citep{Ivezic07,Li16,Stubbs07,blake08,Blake11}. Ground-based follow-up studies of exoplanet detections around M-dwarf host stars and future surveys aiming for sub 1\% photometric precision in the NIR are among those observations most affected. For example, the MEarth survey, which is performing a ground-based search for exoplanets around M-dwarfs, rarely have comparison stars in their field-of-view (FOV) as red as their targets. As a result, the common-mode of their light curves undergo fluctuations from anywhere between 1 and 10 mmag between nights that correlate with ground-level humidity measurements \citep{Berta12}.
Spectroscopic detection of exoplanets via the radial velocity (RV) method is also affected by telluric lines. While the large patches of telluric absorption features imprinted on the data can be masked out, \cite{Cunha14} found that micro-telluric lines can still limit RV measurements at the level of 1 m/s depending on the RV of the target star and the PWV during the observation. These microtelluric lines therefore cannot be ignored by future exoplanet RV surveys aiming to reach 10-20 cm/s precision with instruments such as ESPRESSO \citep{espresso} and NEID \citep{neid,neidhalverson}.
Radiative transfer models of Earth's atmosphere scaled according to a single optical depth parameter do an excellent job of describing the wavelength-dependent transmission of the atmosphere (e.g. \citet{Blake11}). However, a real-time measurement of the optical depth, $\tau$, is still needed to produce an appropriate atmospheric model corresponding to a specific time. While this optical depth parameter can be extracted from the high resolution science spectral data itself \citep{Artigau14, Brogi14, Osip07}, telluric lines can be blended with stellar lines. Additionally, spectroscopic data solely for the purpose of atmospheric characterization is difficult to obtain, especially if it needs to be coincident with photometric observations.
Monitoring the PWV in real-time using a dedicated instrument would enable the generation of accurate atmospheric transmission models and alleviate the need to sacrifice valuable observing time for calibration observations. While GPS monitoring systems are commonly used for measuring PWV, these instruments require maintenance and are less accurate in dry conditions \citep{buehler12,hagemann2003}, which are typical of observing sites. An alternative to the GPS monitoring system is a multiband photometer \citep{Stubbs07}. Recently, a team of scientists from the Dark Energy Survey (DES) built a multiband photometer using narrowband filters called the aTmCam to monitor telluric absorption by constantly observing a bright star that serves to backlight the atmosphere \citep{Li13,Li14}.
Here, we present results from a similar instrument, called the Camera for the Automatic Monitoring of Atmospheric Lines (CAMAL) that we built for use by the MINiature Exoplanet Radial Velocity Array (MINERVA) \citep{minerva15}, MINERVA-Red \citep{minervared}, and MEarth \citep{mearth} surveys located at the Fred Lawrence Whipple Observatory (FLWO) on Mount Hopkins in Arizona. CAMAL focuses on measuring PWV, which is the most important variable atmospheric absorption component for the exoplanet surveys CAMAL is aiding. In \S \ref{sec:require} we discuss the system requirements for CAMAL. We present the design of CAMAL in \S \ref{sec:Instrum} then discuss the observations taken in \S \ref{sec:data}. The methods for extracting PWV from the data are described in \S \ref{sec:extract} and in \S \ref{sec:uncertainty} we address the absolute and relative uncertainties affecting CAMAL PWV measurements. In \S \ref{sec:results} we present the results compared to nearby GPS monitors and PWVs extracted from the Tillinghast Reflector Echelle Spectrometer (TRES), which is also located on Mount Hopkins. In \S \ref{sec:conc} we summarize the performance of CAMAL and discuss future improvements to the system.
\section{System Requirements}\label{sec:require}
To determine how accurate our PWV monitoring instrument must be in order to aid an exoplanet survey designed to detect transits of millimagnitude (hereafter mmag) depth, we must first determine the magnitude of the error induced by nightly PWV variations after differential photometry is performed \citep{blake08}. To model this, we assume stellar and planet parameters matching the TRAPPIST-1 system \citep{trappist} and assume a reference star of 4600 K, which is the average temperature of stars in the Tycho catalog \citep{tycho}. We use PHOENIX stellar models \citep{phoenix} for each star's spectral energy distribution and calculate these stars' fluxes for a range of PWV conditions by multiplying each stellar spectrum by a TAPAS water vapor transmission spectrum \citep{Bertaux14} appropriate for Mount Hopkins and integrating over the effective passband of the MEarth survey \citep{mearthpassband}. In Figure \ref{fig:TRAPPIST}a we plot the magnitude difference between TRAPPIST-1a (2600K) and the 4600K reference star as a function of PWV. For relatively dry conditions (PWV \textless 5 mm), the photometric error induced by PWV absorption is 2 mmag/mm, as indicated by the black line in Figure \ref{fig:TRAPPIST}a, which is a linear fit for PWV $<$ 5 mm. In situations where there only exists reference stars hotter than 4600 K, the second order atmospheric effects will be larger than 2 mmag per millimeter change in PWV. The TRAPPIST-1 planet transit depths range from 3 mmag to 8 mmag, meaning millimeter changes in PWV could produce spurious transit-like signals or potentially obscure a true transit signal.
How PWV actually evolves over the transit duration will ultimately determine whether the transit event will be detectable or not. For an hour long transit, which is the typical transit duration for the TRAPPIST planets, 1-2 millimeter changes in PWV can occur even during drier nights (PWV \textless 5 mm). We show an example of this in Figure \ref{fig:TRAPPIST}b using Kitt Peak PWV values from the SuomiNet network \citep{Suominet}. In this example the PWV above Kitt Peak dropped from 3 mm to 1 mm then increased back to 3.5 mm over the course of approximately four hours. An event like this could create a 4 mmag transit-like photometric dip for a star like TRAPPIST-1a, showing the importance of monitoring PWV to avoid confusing transit events for PWV variability. Though trends in PWV such as the one shown in Figure \ref{fig:TRAPPIST}b will not occur every night, transit events are intrinsically infrequent and thus monitoring PWV will help minimize the chance of either false or non-detections due to changes in PWV.
Based on these simulations, we conclude that correcting for PWV variations is crucial for precise NIR differential photometry of cool stars. Specifically, contemporaneous PWV measurements with a relative precision of $\pm 0.5$mm should provide a means to directly correct photometric measurements for second order extinction effects and achieve mmag precision. Other methods for correcting PWV variations such as using the common-mode trend of a multitude of photometry \citep{Berta12} may also reduce the impact of these second order extinction effects. While a discussion of the optimal method for making this correction is beyond the scope of this paper, we have designed CAMAL to achieve PWV measurement precision of $\pm 0.5$mm.
\begin{figure*}
\gridline{\fig{figs/figure1a.eps}{0.48\textwidth}{(a)}
\fig{figs/figure1b.eps}{0.48\textwidth}{(b)}
}
\caption{Motivation for monitoring PWV in the case of the TRAPPIST-1 system. On the left we show magnitude differences as a function of PWV for a 2600 K star like TRAPPIST 1a and a 4600 K reference star. For drier conditions (PWV \textless 5 mm) the bias in differential photometry for this hypothetical observation would be 2 mmag/mm, as indicated by the slope of the linear fit. On the right we show an example time sequence of PWVs changing at Kitt Peak on April 18th, 2016. \label{fig:TRAPPIST}}
\end{figure*}
\vspace{5mm}
\section{CAMAL Instrument Design}\label{sec:Instrum}
CAMAL utilizes a Celestron six-inch aperture Schmidt-Cassegrain optical tube mounted on a robotic Software Bisque German Equatorial MyT mount. For imaging we use an SBIG STF-8300 monochrome CCD and a filter wheel that contains three narrowband filters, one of which overlaps a dense set of atmospheric water absorption lines. The other two filters are chosen to overlap regions of very low atmospheric absorption. We use the Software Bisque telescope driver along with Python scripts to automate the instrument's pointing and tracking through Software Bisque's TheSkyX control software. Figure \ref{fig:camal} shows an annotated picture of the instrument in its permanent home on Mount Hopkins in one of the MINERVA domes, in addition to an internal view of the filter wheel with the three CAMAL filters in place.
We chose filters 25 mm in diameter with central wavelengths of 780 nm, 823 nm, and 860 nm. The 780 nm filter was an in-stock laser cleanup filter purchased from Semrock\footnote{https://www.semrock.com}, while the other two filters were production overrun filters from Omega Optical\footnote{http://www.omegafilters.com}. We list the filter properties in Table \ref{tab:filters} and plot the filter profiles over a model atmospheric transmission spectrum in Figure \ref{fig:filters}.
\begin{figure}
\centering
\includegraphics[width=0.95\textwidth]{figs/figure2.jpg}
\caption{CAMAL is shown inside a dome housing two MINERVA telescopes. The key CAMAL components are labeled. On the right, the three CAMAL filters are shown in their custom filter holders in the SBIG filter wheel.}
\label{fig:camal}
\end{figure}
We selected these filters to have central wavelengths within 100 nm of each other with the central filter overlapping water vapor absorption lines while the other two filters avoid overlapping any significant atmospheric features. Given a stellar source that serves as a backlight to the atmosphere, changes in the central 823 nm filter flux will correspond to changes in PWV, while the other filters pick up any achromatic changes in flux due to cloud coverage, changes in airmass, or intrinsic stellar variation. The three filters also were chosen to have similar physical sizes and similar narrow bandwidths on the order of a few nanometers. These requirements ensure that the exposure times for imaging our target stars are similar in each of the three filters. The narrow FWHM of these filters also ensures that the 823 nm filter is localized to a strong set of water absorption features. The 860 nm filter does not overlap with the 866.2 nm ionized calcium triplet line, which is often found in the spectra of F, G, and K type stars and can vary in time \citep{Calcium2000}. While CAMAL's targets will tend to be blue stars (T$_{\rm{Eff}}>9000$K) to avoid strong stellar features, in order to keep exposure times per filter below 10s, our limiting $i$-band magnitude is about $i$ \textless 2.5, making it necessary to occasionally observe cooler stars to remain above an airmass limit or to avoid the Moon, for example.
Currently, CAMAL is set up to continuously observe a bright star through one filter at a time while cycling through the filter wheel. Originally, the instrument was designed to have a simultaneous imaging setup, where the filters were located at the front aperture of the telescope with wedge prisms to deviate the locations of the stellar images in the different filters on the focal plane. This was similar to the design envisaged by \cite{Stubbs07} and tested in \cite{Li13}. \cite{Li14} ultimately abandoned this design due to different optimal exposure times for each filter and filter irregularities. While our filters were chosen to ensure similar exposure times, we found that the two filters from Omega Optical were not optically flat, causing a very large change in the focus position between filters when the filters were placed at the telescope pupil. This focus difference and a lack of other filter options led us to modify our design to incorporate a filter wheel that would quickly cycle through each filter. While we may upgrade in the future to multiple telescopes and CCDs to enable truly simultaneous imaging in each filter, the filter wheel transitions are fast enough such that stars brighter than $i$-band magnitude of 2.5 can be observed with a complete filter cycle in less than 30s. \cite{Li14} analyzed the temporal variations of PWV over several nights and found that in three minutes, the PWV changes were mostly attributed to instrumental uncertainty. Although the \cite{Li14} results are from a different observing site, we feel it is safe to assume that the 30s time-scale for a filter cycle is short compared to the expected timescale of atmospheric PWV variations above Mt. Hopkins.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{figs/figure3.eps}
\end{center}
\caption{SDSS and CAMAL filter profiles with Earth's atmospheric transmission spectrum. SDSS i and z filter profiles overlap strong water absorption features at 820 nm and 940 nm.}
\label{fig:filters}
\end{figure}
Though PWV is not expected to change significantly during the timescale of one complete filter cycle, the filter wheel setup is susceptible to changes in atmospheric opacity due to nonuniform, patchy clouds moving and changing on minute timescales. In our test observations, only one night was considered to be photometric throughout, but overall we found non-photometric conditions (i.e. thin clouds) did not significantly impact our ability to extract PWV estimates from the CAMAL fluxes. The impact of clouds on CAMAL's PWV measurements is addressed in detail in \S \ref{sec:clouds}.
\begin{table}
\centering
\begin{tabular}{cccc}
\hline
Centers (nm) & FWHM (nm) & Peak Transmission (\%) & Vendors\\
\hline
779.5 & 2.8 & 88.2 & Semrock\footnote{Due to a blue leak in the 780 nm filter, it was combined with a high-pass filter from Omega Optical. The peak transmission for the 780 nm filter therefore includes the 96.5\% decrease in maximum throughput due to the addition of the high-pass filter.}\\
823.1 & 3.2 & 91.4 & Omega\\
860.3 & 1.9 & 85.4 & Omega\\
\hline
\end{tabular}
\caption{Center wavelength, full width at half maximum, peak transmission, and vendor for each of the three CAMAL filters.}
\label{tab:filters}
\end{table}
\section{Observations \& Data Reduction}\label{sec:data}
\subsection{CAMAL}
During a commissioning phase, we collected data with CAMAL over four nights in the summer of 2016 and seven nights in the late fall of 2016. In the summer the main target star was Polaris while in the fall the target list included Vega, Elnath, and Regulus. CAMAL continuously imaged each of its targets according to a preset schedule, cycling through the filters in the order of the 780 nm filter first, then the 823 nm filter, and ending with the 860 nm filter. Each night began by taking calibration images, including biases and twilight flats. Biases from each night are median combined to produce a nightly master bias frame. A stock set of three bias-subtracted, median combined, then normalized flat-fields for each filter are created. Flat-field corrections can become important because our small filters cause significant vignetting. In practice, the flat-field corrections typically have a negligible effect because the target star remains on the same few pixels in the center of the field of view (FOV) throughout the night. Science frames are bias-subtracted then flat-field corrected, then fluxes for the target star are extracted and the sky background subtracted using the \texttt{photutils} Python package \citep{photutils}. We refer to the flux values measured through each filter as $ f_{780}$, $f_{823}$, and $f_{860}$.
We show example data from a CAMAL observing sequence in Figure \ref{fig:fluxexample}. On the left we show the relative fluxes for targets Elnath and Regulus observed on December 12th, 2016. The flux values have all been divided by a scaling factor that places $f_{780}$ around unity. The corresponding colors $c_1 = f_{823}/f_{780}$ and $c_2 = f_{823}/f_{860}$ are shown on the right in Figure \ref{fig:fluxexample}. The change in targets (and visible shift in fluxes) occurs after 5.5 hours past the start time. The increasing trend in the colors for each target star is due to the fact that the targets on the night shown here were each tracked from higher airmass (above 2) to airmass 1. Though the fluxes change by 10-20$\%$ because of clouds, this high level variability divides out when the flux ratios are taken to make the CAMAL colors.
\begin{figure}
\plottwo{figs/figure4a.eps}{figs/figure4b.eps}
\caption{Example CAMAL flux time-series for the night of December 12th, 2016 (left) and CAMAL colors made from these fluxes (right). Targets switch from Elnath to Regulus at approximately 5.5 hours after the start of the run. The constant variation in the fluxes indicate that this was a very cloudy night. \label{fig:fluxexample}}
\end{figure}
To assess the CAMAL photometry, we evaluate the expected photon error and compare this to the measured photometric error for the CAMAL color, $c_1$. To do this we use five nights of observations of Elnath and compute the standard deviation of $c_1$ over five minute intervals (10-11 filter cycles), which we plot as a function of airmass in Figure \ref{fig:photunc} in addition to the expected photometric uncertainty, which is shown as the red dashed line. The expected uncertainty is calculated by adding in quadrature the expected photon noise, scintillation noise, and read noise using the nominal $f_{780}$ and $f_{823}$ Elnath photon counts and then propagating this to the uncertainty in $c_1$. The estimate for scintillation noise is based on Young's approximation \citep{young67}. On average, the measured photometric error is twice the theoretical expectation based on scintillation, photon, and read noise. This underestimation may be due to an underestimation of the scintillation noise, which has been addressed by \citet{scint1} who found that Young's approximation underestimates the median scintillation noise by a factor of 1.5.
\begin{figure}
\centering
\includegraphics[width=0.45\linewidth]{figs/figure5.eps}
\caption{The RMS of five minute intervals of Elnath $c_1$ measurements plotted as a function of airmass. Five nights of data are shown here. The expected error is plotted as the red dashed line and takes into account photon noise, scintillation noise, and CCD read noise.}
\label{fig:photunc}
\end{figure}
\subsection{TRES}\label{sec:tresdata}
Additional observations from the Tillinghast Reflector Echelle Spectrograph (TRES) on the 1.5-meter Tillinghast telescope at FLWO on Mount Hopkins are used to validate the CAMAL PWV measurements and calibrate the absolute CAMAL PWV scale. TRES is located approximately 100 meters from CAMAL and produces high-resolution (R $\sim$ 30,000) stellar spectra over wavelength ranges covering the 823 nm water absorption features. The TRES spectra were reduced and blaze corrected according to the standard pipeline for TRES spectra \citep{tres}. We use these spectra to extract PWV values directly from water line depths and compare to nearly contemporaneous CAMAL PWV measurements. The TRES instrument is on-sky in cycles of approximately two weeks as it alternates with another instrument installed on the Tillinghast telescope. Due to this, there is not always overlap between TRES and CAMAL observations.
The TRES spectra that were selected were chosen to be the bluest, brightest star observed each night in order to reduce the influence of stellar lines on the fitting of the telluric water lines and additionally to ensure the spectra have high signal to noise ratios. Vega was a common target, but because Vega was typically observed during twilight several hours before CAMAL started observing, we also utilized TRES spectra of other targets that were observed closer in time to the CAMAL observations. In total, we utilize eight TRES spectra that overlap with seven CAMAL observing nights. Not all TRES observations are exactly contemporaneous with CAMAL observations because TRES begins observations of the brightest stars well before astronomical twilight. The airmasses of the TRES observations are also different than the CAMAL observations. We discuss this more when comparing CAMAL and TRES PWV measurements in \S \ref{sec:results}.
To extract PWV from the TRES data, we fit a 15\AA $\:$portion of the spectrum around 828 nm. We use a model atmospheric water vapor transmission spectrum appropriate for Mt. Hopkins from TAPAS \citep{Bertaux14} to fit the data. We scale the water vapor transmission spectrum to a particular PWV value following Equation \ref{eq:PWV}. We then convolve the model spectrum with a Gaussian of width $\sigma$ to model the line spread function of the spectrograph, then multiply by a linear function to account for a slope in the continuum. Two parameters are used to describe the wavelength solution of the data, then the model spectrum and data are compared on the same wavelength grid. We use the \texttt{MC3} Python package to find the best fit parameters and parameter uncertainties. Since we do not account for airmass in the fit, we must divide the final best fit PWV value by the airmass of the observation to scale PWV to its value at an airmass of 1. This allows for a direct comparison to CAMAL PWVs, which also determines PWV scaled to an airmass of 1 (see Equation \ref{eq:flux}). An example fit of a TRES spectrum is shown in Figure \ref{fig:tresfit}. For stars that contain more spectral absorption features, we fit multiple 15\AA $\:$ regions and average the results for PWV.
\begin{figure}
\begin{center}
\includegraphics[width=0.6\textwidth]{figs/figure6.eps}
\end{center}
\caption{Example fit of a TRES spectrum of Vega on November 23rd, 2016. The top panel shows the spectral data in blue with the best fit plotted as the black dashed line. The bottom panel shows the residuals of the fit. }
\label{fig:tresfit}
\end{figure}
\section{Extracting PWV from the CAMAL Colors}\label{sec:extract}
To derive absolute PWV measurements from the CAMAL photometry, we construct a forward model that produces model CAMAL fluxes by integrating over a template stellar spectrum multiplied by the throughput of the entire optical system in each filter band. The forward model is pre-evaluated on a three dimensional grid of PWVs, airmasses, and stellar temperatures. Then, for a specific airmass and target-star temperature, the PWV can be quickly evaluated by matching the observed CAMAL fluxes to those generated by the forward model. This quick calculation is useful when real-time PWV measurements are required. Alternatively, the time sequence of PWV over the night can be represented as a smoothly varying function and the whole night's data can be fit using an MCMC approach. We use this as a second method to extract PWVs from the data that allows us to characterize the statistical uncertainty in our PWV measurements.
\subsection{Generating Model Color Grids}\label{sec:pwvquick}
We make use of the instrument response curves for the optical system, CCD, and filters provided by the vendors to predict the overall wavelength dependent efficiency for CAMAL. For the model spectrum of the atmospheric absorption, we use the TAPAS \citep{Bertaux14} web form\footnote{http://www.pole-ether.fr/tapas/} to download one model spectrum containing only water vapor absorption features and another spectrum that only includes the effect of Rayleigh scattering. We scale the TAPAS telluric absorption model using $\textrm{v}$, the precipitable water in a column of unit area in the atmosphere, according to
\begin{equation}\label{eq:PWV}
\textrm{W}(\textrm{v}^\prime) = \textrm{W}(\textrm{v})^{\textrm{v}^\prime/\textrm{v}} ,
\end{equation}
\begin{figure}
\begin{center}
\plottwo{figs/figure7a.eps}{figs/figure7b.eps}
\end{center}
\caption{The flux ratios for $f_{823}$/$f_{780}$ and $f_{823}$/$f_{860}$ plotted versus airmass for a 10,000 K star at different PWV values (in millimeters) as determined by the forward model. The location on this plot of the measured CAMAL color for a given observation indicates the PWV at the time of the observation.}
\label{fig:fluxairmass}
\end{figure}
\noindent where $\textrm{W}$ is a telluric water vapor absorption spectrum generated using a known value of $\textrm{v}$. We define the parameter $\tau$, which is related to the water optical depth, as $\tau = \textrm{v}^\prime/\textrm{v}$ .
For each of the stars that are observed, we use PHOENIX \citep{phoenix} synthetic spectra of the corresponding temperature, surface gravity, and metallicity to predict the flux per unit telescope area, $f$, through each filter, $i$, which is derived using the following equation:
\begin{equation}\label{eq:flux}
f_i = \int d\lambda \: \textrm{W}^{\tau \mathit{X}}(\lambda) R^\mathit{X}(\lambda) B_i(\lambda) S(\lambda, T) E(\lambda) .
\end{equation}
\noindent We use $R$ as the Rayleigh scattering function slope at a given airmass, $X$, $B$ for the filter bandpass, $S$ as the stellar spectrum, and $E$ as the combined efficiency of the telescope and CCD. We always take PWV to be at an airmass of 1 and then scale by the airmass of an observation. In this equation, the only time-dependent variables are airmass and stellar temperature, which are both known, and $\tau$, which we wish to measure. We pre-evaluate Equation \ref{eq:flux} for all possible observations by integrating over wavelength assuming a grid of airmass values from 1.0 to 3.5 in steps of 0.01, stellar temperatures and properties matching those of our targets, and PWV values from 0 mm to 25 mm in steps of 0.1 mm. We use these model flux grids to generate the model colors $m_1 = \frac{f_{823}}{f_{780}}$ and $m_2 = \frac{f_{823}}{f_{860}}$ to compare to the observed CAMAL colors, $c_1$ and $c_2$. For each observation of the CAMAL color of a target star, the PWV during that measurement can be determined by where the observed color is located on this grid of model colors. In Figure \ref{fig:fluxairmass} we show example model colors, $m_1$ and $m_2$, both plotted versus airmass for a 10,000 K star and various values of PWV. We find that both colors are very sensitive to PWV. At an airmass of 1.3 and PWV value of 7 mm, there is a 1\% change in $m_1$ per millimeter change in PWV.
\subsection{Modeling PWV as a Smooth Function}\label{sec:fitting}
In the grid-based approach to extracting PWV, each CAMAL color must be matched to its respective color grid, which produces two different mappings to PWV. Although these two PWV mappings agree with each other under normal conditions (i.e. no external chromatic effects) and are just limited by photon noise to within our PWV grid spacing of 0.1 mm, we would prefer a method that fits both colors at the same time to find the best fit given all of the photometry and derives a single PWV time series for the night that is smoothly varying. To do this, we model the PWV as a smoothly varying function in time under the assumption that the typical timescale for PWV variability is 30 min \citep{Li14}. Modeling PWV in this way allows us to fit the entire night of data at once using an MCMC sampler that allows us to characterize our uncertainties.
The fit is performed by using a spline function to define PWV as a function of time. We map this PWV time sequence to the time variation of CAMAL colors using the model color grids described in \S \ref{sec:pwvquick}. The optimal PWV function parameters are determined by minimizing the difference between the model and observed colors following value:
\begin{equation}\label{eq:min}
\sum_i \Big[ \big(c_1(t_i) - m_1(\textrm{PWV}(t_i))\big)^2 + \big(c_2(t_i) - m_2(\textrm{PWV}(t_i))\big)^2\Big] \: .
\end{equation}
\noindent Here, $c_1$ and $c_2$ represent the CAMAL observed colors, and $t_i$ is the time of exposure sequence $i$. To define PWV($t_i$) we use a 3rd order B-spline, which we define using the Python function \texttt{splev} from the \texttt{scipy.interpolate} package. More information regarding the definition of the spline can be found in the references for \texttt{splev} \citep{bspline1,bspline2,bspline3}. We force the spline knot points to be stationary and equally spaced such that there is at least one knot point in each hour interval, including the end points. The hourly spaced knot points allow for variations on 30-minute time scales without overfitting the data. If $N_{in}$ is the number of internal knots, then there are $N_{in} + 4$ coefficients whose initial values are set to a starting value between 1 and 20 mm based on the model grid estimates. We use the \texttt{MC3} Python package \citep{mc3} to find best fit coefficients that describe $\textrm{PWV}(t)$. The various realizations of $\textrm{PWV}(t)$ using the parameters in the converged MCMC chains are used to determine the uncertainty in the PWV measurements while the median of them is taken to be the best fit.
\section{Assessing PWV Uncertainty}\label{sec:uncertainty}
Depending on the application, either relative or absolute PWV estimates may be necessary for correcting NIR spectroscopic or photometric measurements. We assess CAMAL's PWV accuracy in terms of how well CAMAL can measure absolute PWV values and we discuss precision in the context of CAMAL's ability to measure relative PWV values over time. In addition to this we discuss and characterize how clouds can contribute to both types of these uncertainties.
\subsection{Accuracy of Absolute PWV Measurement}\label{sec:accuracy}
The absolute calibration of CAMAL PWV estimates based on our forward modeling of the measured stellar flux depends on the total system throughput as a function of wavelength and the accuracy of our stellar models. However, the instrument efficiency profiles provided by the manufacturers are not direct measurements of our instrument. Additionally, while the 780 nm and 823 nm bands do not fall on any broad stellar spectral features in our hot target stars, the 860 nm band does overlap a Paschen series line that can be mismatched to our model spectrum depending on the true broadening of that line. While this could be alleviated by obtaining high-resolution spectra for each of our target stars over CAMAL's wavelength region, we can instead rely on the fact that the CAMAL flux through the 780 nm and 860 nm bands should match our model since flux through these two bands is independent of PWV. To calibrate the 823 nm model throughput, which is degenerate with PWV, we use independent PWV values extracted from TRES spectra.
Because the 860 nm band overlaps a broad stellar spectral feature, we find that each star requires its own adjustment to its efficiency at 860 nm. The CAMAL filters are very narrow and so we can take the correction to both the instrument efficiency and stellar spectrum to be a scalar, $A_{1}$, that is specific for each star. We indicate this by making it a function of temperature. $A_1(T)$ is just the ratio of the true stellar flux and telescope efficiency (i.e. mirror reflectivity, CCD QE) at 860 nm divided by what was originally assumed:
\begin{equation}
A_1(T) = \frac{S_{\mathrm{true}}(\lambda, T) E_{\mathrm{true}}(\lambda)}{S(\lambda, T) E(\lambda) } \biggr|_{\lambda=860\rm{nm}}
\end{equation}
\noindent This factor $A_1(T)$ is determined such that $f_{780}/f_{860}$ from our model agrees with our average flux measurement at airmass 1.0 for each star. We find that $A_1(T)$ ranges between 0.88 and 0.94 for the range of targets used in this work. To assess how reasonable these offsets are, we use a true spectrum of Vega from the CALSPEC Calibration Database\footnote{http://www.stsci.edu/hst/observatory/crds/calspec.html} to calculate $f_{780}/f_{860}$ and compare this to what we estimated using a PHOENIX synthetic spectrum. We find that $f_{860}/f_{780} $ increases by 4\% when we use the real spectrum of Vega compared to the PHOENIX model. Given the uncertainties in the reported efficiencies of the system components, such as the CCD, this is consistent with $A_1$ = 93.4\%, which is what was found for Vega.
For calibrating $f_{823}$, we find no significant difference in $f_{823}/f_{780}$ when calculated with a real and synthetic spectrum for Vega. We therefore only determine one unique factor to account for uncertainty in CAMAL's assumed instrument efficiency at 823 nm:
\begin{equation}
A_2 = \frac{E_{\mathrm{true}}(\lambda) }{ E(\lambda) } \biggr|_{\lambda = 823 \rm{nm}}
\end{equation}
\noindent To determine $A_2$, we compare CAMAL PWV estimates to the TRES measurements of PWV. Because TRES PWVs are extracted from high resolution spectra using an atmospheric model, these PWVs are unbiased, independent measurements of PWV. We compute $A_2$ by minimizing the root mean square difference between the TRES direct measurements of PWV and the corresponding PWV estimates from the CAMAL colors. We find that $A_2$ is 97.9\%, which implies that we were overestimating CAMAL's throughput by 2.1\%, which is reasonable due to the uncertainties in the telescope reflectivity and CCD QE. Using the two scalars $A_1(T)$ and $A_2$ to recalculate Equation \ref{eq:flux} and remake our model grids. We use these calibrated grids to recompute PWV for each night. A plot of the CAMAL-TRES comparison before $A_2$ is applied and after each of these two calibrations are performed is presented in Figure \ref{fig:trescamal}.
\subsection{PWV Measurement Precision}\label{sec:precision}
Assuming there are no other time varying chromatic effects in the CAMAL system other than PWV changing, the precision of CAMAL's measurements relies only on its ability to measure precise colors of its target star. We can therefore estimate its measurement precision by calculating the variation in our color measurements and then relating this to PWV measurement precision by propagating these color uncertainties through our PWV extraction method. In reality, clouds can potentially act as a time varying absorption that affects each filter slightly differently because the observations are not exactly contemporaneous. We address this in \S \ref{sec:clouds}.
Using five nights of CAMAL photometry on the target star Elnath, we estimate the photometric precision of our observations by taking the standard deviation over ten minute intervals of binned CAMAL colors $c_1$ and $c_2$. We plot the standard deviation of each of these ten minute intervals for $c_1$ as a function of airmass in Figure \ref{fig:photerr}. This differs from Figure \ref{fig:photunc} because here we have binned $c_1$ and $c_2$ on 2 minute timescales (4-5 filter cycles) to average down the photometric noise in order to see how precise we measure $c_1$ on longer timescales. For future CAMAL operations, we will limit observations to airmass less than 1.5 to reduce the effects of scintillation noise on CAMAL photometry.
\begin{figure}
\centering
\begin{minipage}{.48\textwidth}
\centering
\includegraphics[width=0.99\linewidth]{figs/figure8.eps}
\caption{Empirical photometric precision of CAMAL colors as a function of airmass for five nights of observations of Elnath. }
\label{fig:photerr}
\end{minipage}
\hfill
\begin{minipage}{.48\textwidth}
\centering
\includegraphics[width=0.99\linewidth]{figs/figure9.eps}
\caption{The error in PWV as a function of error in $c_1$ for various values of PWV assuming a 10,000K star and an airmass of 1.3. The resulting PWV error is step-like due to the 0.1 mm resolution of the model color grids. }
\label{fig:pwverr}
\end{minipage}
\end{figure}
Using the empirical range in photometric precision (0.002 \textless $\sigma_{c_1}$ \textless 0.01) we found for $c_1$, we propagate this photometric uncertainty into PWV uncertainty by using our model color grids to map color to PWV. We determine $\Delta$PWV, the difference in true PWV and PWV determined after adding $\sigma_{c_1}$ to $c_1$. We plot $\Delta$PWV versus $\sigma_{c_1}$ for various PWV conditions in Figure \ref{fig:pwverr}. For this calculation we assume an airmass of 1.3 and target star temperature of 10,000K. At an airmass of 1.3, $\sigma_{c_1}$ clusters around 0.006 in Figure \ref{fig:photerr}. This propagates to a precision of better than 0.5 mm in drier conditions, when PWV \textless 4 mm. These estimates of CAMAL's PWV precision agree with the statistical uncertainties in PWV determined from the fit to the photometry using a spline functional form for PWV as described in Section 5.2.
\subsection{The Impact of Clouds}\label{sec:clouds}
Due to the filter wheel design of CAMAL, the instrument does not make perfectly simultaneous measurements in all three bands. Though filter cycles are short compared to the expected timescale of variations in the atmosphere, clouds move very quickly across the field of view and therefore can effectively introduce a chromatic effect as the cloud extinction varies over a complete CAMAL filter cycle. If the magnitude of the cloud extinction changes randomly, clouds will increase the scatter in the photometry, worsening the photometric precision. If the cloud extinction is inherently chromatic, this could also systematically bias our measurements of PWV. To understand the magnitude of this effect, we use a night of data (Dec. 6th, 2016) that was clear at the beginning of the night, was affected by thin clouds for for a period of approximately an hour, then finished with a very sharp drop in flux due to thickening clouds.
To understand if clouds cause a systematic bias in the estimated PWVs, we look for systematics in the extracted best fit PWV that correlate with cloud absorption. If clouds are not acting as gray absorbers, we would expect a systematic trend in PWV correlating with changes in the overall flux attenuation due to clouds. In the top panel of Figure \ref{fig:clouderr}, we plot $f_{780}$ for Elnath along with CAMAL PWVs in gray and PWVs from the nearby site at Kitt Peak National Observatory plotted in red. Where flux is decreasing due to clouds (time \textgreater 2.8 hours) we see no systematic trend in the CAMAL PWVs. While the PWV between Mt. Hopkins and Kitt Peak are not perfectly correlated (see in \S \ref{sec:GPS}) the Kitt Peak PWVs are consistent with constant for this night, implying that the overall PWV level at Mt. Hopkins was likely stable over this night as well, consistent with the CAMAL measurements shown in Figure 10. Based on this, we conclude from this that chromatic effects in clouds do not significantly bias the CAMAL photometric PWV estimates.
To estimate how clouds affect PWV precision, we take the `real time' PWVs extracted from the model grids. We would expect increased photometric scatter due to clouds to be evident in the scatter of the grid-based PWV estimates. We show the standard deviation of ten minute intervals of the grid-based PWVs binned by 5 filter cycles (2 minutes) in Figure \ref{fig:clouderr}. The scatter in PWV decreases with time because of decreasing airmass, but increases again at times coincident with cloud absorption. While the effects of clouds are not debilitating, in future runs we will choose to move targets if cloud absorption reduces the nominal target flux by over 20\%, which will aid CAMAL in maintaining its goal of 0.5 mm precision in PWV.
\begin{figure}
\begin{center}
\includegraphics[width=0.5\textwidth]{figs/figure10.eps}
\end{center}
\caption{In the top panel, the $f_{780}$ flux of Elnath on December 6th, 2016 in addition to CAMAL PWVs for the night and Kitt Peak PWVs in red. The gray points are from PWV measurements mapped from each individual CAMAL color and the best fit PWV time sequence is given as the gray line. In the bottom panel we plot PWV error estimated from the individually mapped PWVs. High airmass observations (X \textgreater 1.5, time \textless 1 hour) and regions where flux is changing due to clouds (time \textgreater 2.8 hours) show greater error.}
\label{fig:clouderr}
\end{figure}
\section{Results \& Discussion}\label{sec:results}
\begin{figure*}
\gridline{\fig{figs/figure11a.eps}{0.48\textwidth}{(a)}
\fig{figs/figure11b.eps}{0.48\textwidth}{(b)}
}
\gridline{\fig{figs/figure11c.eps}{0.48\textwidth}{(c)}
\fig{figs/figure11d.eps}{0.48\textwidth}{(d)}
}
\gridline{\fig{figs/figure11e.eps}{0.48\textwidth}{(e)}
\fig{figs/figure11f.eps}{0.48\textwidth}{(f)}
}
\caption{Data from six nights of CAMAL data taken in December, 2016. CAMAL PWVs are plotted as gray points with the best-fit function of PWV overplotted as a gray shaded region. Red and blue error bars show PWV measurements from the nearby GPS PWV monitors at Kitt Peak and Amado, respectively. Regions with high scatter in the CAMAL PWVs typically correspond to changing to a high airmass target or because of clouds passing in the FOV.\label{fig:camaldata}}
\end{figure*}
\begin{figure}
\centering
\begin{minipage}{.95\textwidth}
\centering
\includegraphics[width=0.57\linewidth]{figs/figure12.eps}
\caption{CAMAL PWVs plotted versus independent PWV measurements using TRES. The gray points are using CAMAL measurements before calibrating the instrument throughput. The gray dashed line is a linear fit to these data, which were used to determine the calibration factor, $A_2$. The blue points compare TRES with the CAMAL data after calibrating with $A_2$. The solid black line is a one-to-one line. The standard deviation of residuals of TRES compared to the calibrated CAMAL PWVs is 0.7mm.}
\label{fig:trescamal}
\end{minipage}
\hfill
\begin{minipage}{.48\textwidth}
\centering
\includegraphics[width=0.9\linewidth]{figs/figure13.eps}
\caption{CAMAL PWVs plotted versus Amado PWVs. The black solid line shows the line of one-to-one correspondence, while the red dashed line is a fit to the data. The standard deviation of residuals from the fit is 1.6 mm. }
\label{fig:azamcamal}
\end{minipage}%
\hfill
\begin{minipage}{.48\textwidth}
\centering
\includegraphics[width=0.9\linewidth]{figs/figure14.eps}
\caption{CAMAL PWVs compared to Kitt Peak PWVs. The black solid line shows the line of one-to-one correspondence, while the red dashed line is a fit to the data. The standard deviation of residuals from the fit is 0.8 mm.}
\label{fig:kittcamal}
\end{minipage}
\end{figure}
Here we present results using data from the commissioning of CAMAL over four nights in the summer of 2016 and seven full nights in the late fall of 2016. We show in Figure \ref{fig:camaldata} the PWVs extracted following both methods described in \S \ref{sec:extract} for the six nights of data taken in December 2016. In each panel, the gray data points are `real time' PWV measurements from CAMAL colors $c_1$ and $c_2$ using the model grids described in \S \ref{sec:pwvquick}. The gray shaded region is the 3$\sigma$ region from the fit to all the data using a smooth function to describe the time evolution of PWV over the night, as described in \S \ref{sec:fitting}. PWV data from two nearby GPS PWV monitoring systems on the SuomiNet system \citep{Suominet} are also plotted. In red are PWV data from Kitt Peak Observatory and in blue are PWVs from the Amado site.
For the nights shown in Figure \ref{fig:camaldata}, the PWVs found by CAMAL do not vary more than 4 mm in one night. Each night's temporal variation is unique, with some nights showing an increasing trend in PWV, while others show a decrease in PWV over the course of the night. The two different methods for extracting PWV agree well, with no systematic offsets. Times in the night where target stars were changed are noticeable in some nights of data because we typically slewed to a target at high airmass and tracked it upwards towards zenith. The scatter in the photometric data increases after this transition due to the increase in scintillation noise at higher airmass.
\subsection{TRES Comparison}\label{sec:trescompare}
Seven nights of TRES PWV values are compared in Figure \ref{fig:trescamal} to CAMAL values. The gray points are the CAMAL values that were used to determine $A_2$ and the blue points are after $A_2$ was used to calibrate the CAMAL instrument throughput. Because all but two of the TRES and CAMAL measurements are not perfectly contemporaneous (some TRES spectra were taken outside of CAMAL's observing window by as much as a few hours), we plot the CAMAL PWV value nearest in time to the corresponding TRES spectrum taken on the same night. For one night we plot two points because two TRES spectra were taken that night at times separated by a few hours. The values and uncertainties of the CAMAL PWV values are taken from the results of the fit of PWV defined as a spline function, as described in \S \ref{sec:fitting}. These data were used to calibrate the CAMAL instrument throughput at 823 nm. The calibrated model of the instrument throughput results in the excellent correspondence shown here between TRES PWVs and CAMAL PWVs, with a correlation coefficient of 0.98.
The residuals between CAMAL (post $A_2$ calibration) and TRES values have a root mean square scatter of 0.7 mm. This remaining scatter in the relationship between the CAMAL and TRES data may be due to a combination of spatial and/or temporal variations in PWV between the observations since the corresponding measurements are not all perfectly contemporaneous, and the target stars had a range in angular separation on the sky. The residuals of the CAMAL data relative to the TRES PWV measurements data do not correlate with the temporal or spatial offsets between the two measurements. Our understanding of these residuals will continue to improve with more nights of contemporaneous observations between TRES and CAMAL. Since the TAPAS atmospheric models are directly tied to a physical model of the atmosphere and PWV in an absolute sense, from these comparisons between TRES and CAMAL we conclude that CAMAL produces PWVs with an absolute accuracy of approximately 0.7 mm.
\pagebreak
\subsection{Comparison to Nearby GPS Monitors}\label{sec:GPS}
We investigate how correlated CAMAL PWVs are with the Kitt Peak and Amado PWV monitors by comparing the various sites' PWVs over the nights CAMAL observed. The GPS station at Kitt Peak Observatory is at an altitude of 6900 ft and is approximately 90 km away while the GPS PWV monitoring system in the nearby town of Amado is approximately 30 km away from Mount Hopkins at an altitude of 3000 ft.
In Figures \ref{fig:azamcamal} and \ref{fig:kittcamal} we compare CAMAL PWVs to Amado and Kitt Peak PWVs, respectively. Error bars are omitted from these plots for clarity, however the GPS water vapor monitors have a reported measurement uncertainty of $\sim$1 mm. In total there were ten nights that had both Kitt Peak and Amado PWV measurements. Since the GPS monitor takes measurements every half hour, we plot all the GPS measurements in a night versus the coincident CAMAL measurements taken as a median over a ten-minute bin of CAMAL PWVs derived from the model grids and centered at the time of the GPS measurement. However, data from two nights in June were excluded in Figure \ref{fig:kittcamal} when the Kitt Peak PWV monitor was reporting extremely high values (\textgreater 25 mm), which are likely due to active precipitation.
The PWV comparisons between sites show highly correlated conditions between Kitt Peak and Mt. Hopkins with a Pearson correlation coefficient of 0.96 and an RMS difference of 0.8 mm from a linear fit. Amado and CAMAL are slightly less correlated, with a Pearson correlation coefficient of 0.92 and an RMS difference of 1.6 mm from a linear fit. Both Kitt Peak and Amado have wetter conditions when compared to CAMAL values. Due to the elevation differences between the CAMAL site on Mount Hopkins and the GPS sites of Kitt Peak (900 ft lower) and Amado (4800 ft lower), the drier conditions at Mt. Hopkins as shown by CAMAL are expected. The expected average relationship between Kitt Peak and Mt. Hopkins PWVs based on TAPAS atmospheric models, which only differ by site altitude, is that Kitt Peak PWVs are 1.12 times higher than PWVs at Mount Hopkins. This is similar to slope of 1.16 found in the linear fit between CAMAL and Kitt Peak PWVs.
Correlating PWV between two different sites has been done in previous studies. For example, in \cite{Blake11} the authors were able to correlate two GPS monitors that were 50 km away from each other with a 4000 ft elevation difference. \cite{Blake11} found that the two were highly correlated with deviations from linear at PWV \textless 1 mm. The comparison presented here between Amado and CAMAL also shows deviations from linearity at low PWV values, but this is not the case when comparing Kitt Peak to Mt. Hopkins. The PWV comparisons presented in \cite{Blake11} and this work shows the utility of using one site's PWV measurements to estimate conditions at a nearby site. Selecting PWV data from another site at a similar altitude may be preferable to simply using PWV data from the nearest location.
\section{Conclusions}\label{sec:conc}
We have successfully developed and installed an automatic PWV monitoring system, CAMAL, at the Fred Lawrence Whipple Observatory on Mount Hopkins that will aid the MINERVA, MINERVA-Red, and MEarth exoplanet surveys. CAMAL uses three narrowband filters, one of which is centered on water vapor telluric lines at 823 nm. By observing bright stars throughout a night, changes in flux through the 823 nm on-band filter correspond to changes in telluric water vapor content, or PWV, above Mount Hopkins, while the other two off-band filters serve as a reference to track any achromatic changes in flux not due to water vapor. By constructing a forward model and using independent measurements of PWV derived from high-resolution optical spectra to calibrate CAMAL's throughput, we are able to map the observed CAMAL flux ratios to absolute PWV measurements.
The commissioning data presented in this work spread over 11 nights in the summer and late fall of 2016. With these measurements we show that our photometric precision allows CAMAL to achieve better than 0.5 mm PWV precision for PWV conditions of \textless 4 mm. This level of precision will allow for mmag photometry for transiting exoplanet surveys targeting cool stars in the NIR. Additionally, comparisons between CAMAL PWVs and those from a GPS water vapor monitor at Kitt Peak National Observatory show that the PWVs at these sites are highly correlated with a 0.8 mm RMS difference from a linear fit. This demonstrates the utility of using the PWV conditions from a nearby site at a similar altitude as a first estimate of local PWV.
In future work we plan to upgrade CAMAL to a new design that allows for truly simultaneous measurements of a star through each of CAMAL's three filters. This will involve a total of three optical tubes and CCDs each equipped with one of the three filters that together will simultaneously image the same star. Although the current filter wheel design has allowed CAMAL to achieve its goal of measuring PWVs with $\pm$0.5 mm precision, a simultaneous setup would eliminate CAMAL's current vulnerability to highly variable clouds effectively reducing photometric precision. Additionally, we must currently limit CAMAL to stars brighter than $i=2.5$ mag in order to keep filter cycles periods under 30s to minimize the impact of atmospheric variability due to clouds. With a simultaneous setup, we would not have to limit exposure times and thus could observe fainter stars and increase the number of CAMAL targets. Having more target options could become useful if it is necessary for CAMAL targets to be nearby on the sky with respect to the science targets that will be aided by CAMAL's PWV measurements. Though \cite{Li14} did not find measurable variation on the sky in an 8 day campaign at CTIO, we will perform a similar study to assess the spatial variation of PWV on the sky at Mount Hopkins using CAMAL.
\acknowledgments
We would like to thank the anonymous referee for his or her helpful comments that improved this manuscript. We also thank the Fred Whipple Observatory support staff for their help machining the base attachment for CAMAL and for their quick on-site support when it was needed. We thank the MINERVA Collaboration for letting us install CAMAL in one of their domes and Jason Eastman for his coding help. We thank David Latham and Allyson Bieryla for picking out and sending us the TRES spectra used in this work. We acknowledge the use of \texttt{astropy} in this work and thank the Astropy Collaboration \citep{Astropy13} for their hours spent working on this open source code. This material is based upon work supported by the National Science Foundation Graduate Research Fellowship under Grant No. DGE-1321851 to A.D.B.
| 594d4280cfb9f6117ad6f47a90cafbbf1bb20469 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Despite experimental realizations of cold-atom Bose--Einstein condensation (BEC) in 1995, it remains difficult to measure the critical temperature change caused by interactions \cite{Smith}. This is especially true for homogeneous gases, for which, in spite of experiments in trapped gases \cite{Smith2} and novel experimental techniques \cite{GauntSmith}, no measurements were ever made.
To theorists, homogeneous gases also pose challenges: a \textit{non-interacting}, or \textit{free}, 3D Bose gas with density $\rho$ forms a BEC below
\begin{equation}
\label{Tfc}
T_{\rm fc}=4\pi\zeta(3/2)^{-2/3}\rho^{2/3}
\end{equation}
(with $\hbar=2m=k_B=1$), but how does an interaction change this \textit{free critical temperature}? Feynman \cite{Feynman-53} used path integrals to qualitatively answer this question for liquid helium. He predicted that the potential increases the effective mass, lowering the critical temperature---something that had already been measured.
To make quantitative predictions, various simplifications were considered. For a hard-core Bose gas with core radius (or scattering length) $a$, one can assume that the gas is \textit{dilute} and apply perturbation theory. The relevant parameter is $\rho^{1/3}a\ll1$, implying that the particles tend to be far apart on the scale of the interaction.
Studying this set-up, Lee and Yang \cite{LeeYan-58} replaced the hard-core potential with a pseudopotential \cite{HuaYan-57,LeeHuaYan-57} and simplified the resulting Hamiltonian with the Bogoliubov approximation \cite{Bogoliubov-47b}. Through its energy spectrum, they found that the critical temperature shifts by an amount proportional to $\rho^{1/3}a$, that is,
\begin{equation}
\label{exprTc}
T_{\rm{c}}=T_{\rm{fc}}(1+1.79(\rho^{1/3}a)+o(\rho^{1/3}a)),
\end{equation}
see (A2) in \cite{LeeYan-58}. This result was not taken seriously as the approach wrongly predicts a first-order phase transition.
What followed was an intense debate about the size of the critical temperature shift \cite{Andersen-04,Baymetal-01}. Motivated by the discrepancy between the observed decrease in $T_{\rm c}$ for helium and \eqref{exprTc}, some early theoretical results such as \cite{FetWal} disagreed with the predicted temperature increase. Other papers disputed the linear dependence on $\rho^{1/3}a$ (\cite{GKW,Hua1,Hua2} predict exponents of 1/2, 3/2 and 1/2, respectively). Ultimately field-theoretic methods showed that expression \eqref{exprTc} was accurate \cite{Stoof,BijSto-96,Baymetal-99,Yukalov}, and Monte Carlo simulations predicted that the constant's value 1.79 should be closer to 1.3 \cite{Arnold,Kash,NhoLan-04}. The fact that Bogoliubov's excitation spectrum manifestly predicts the correct expression seems to be little known. In our view, it is a worthwhile point to emphasize because the techniques involved are standard and accessible compared to more accurate, renormalization-based methods.
In a recent work \cite{NapReuSol2-15}, we approach the Bogoliubov approximation from a variational angle, rediscovering a model first introduced in \cite{CriSol-76} but scarcely used since. This allows us to include some interaction terms not present in the standard Bogoliubov Hamiltonian. Although the approach is still Bogoliubov-based, and so wrongly predicts a first-order phase transition, it sharpens \eqref{exprTc} to
\begin{equation}
T_{\rm{c}}=T_{\rm{fc}}(1+1.49(\rho^{1/3}a)+o(\rho^{1/3}a)).
\end{equation}
This analytical result is the closest to numerical predictions to date, which means it does better than more advanced methods, but that can of course be a coincidence.
In this paper, we repeat our calculation for a dilute 2D Bose gas to give the first analytical calculation of the Kosterlitz-Thouless transition temperature that includes the constant in the logarithm.
That may require some explanation. Our 3D calculation \cite{NapReuSol2-15} closely resembles the approach presented here (in fact, the paper can be read as an outline of that case). This could seem surprising because the phase transition in 3D involves BEC, which does not occur in homogeneous 2D Bose gases according to the Mermin--Wagner--Hohenberg theorem \cite{Hohenberg}. In reality, the 2D phase transition we study here is the formation of a quasi-condensate \cite{Popov,Kagan}, and it has been argued that Bogoliubov theory can accurately describe these \cite{MoraCastin} (see also \cite{Review2D} for a review). Quasi-condensates have been observed experimentally \cite{2Dexp}, and do not necessarily indicate the superfluid phase associated with Kosterlitz--Thouless physics \cite{KosterlitzThouless}, but as we explain before Theorem \ref{Crit}, (quasi-)condensation and superfluid pairing always occur together in this model and so our result is indeed this model's prediction of the KT transition temperature.
As in the 3D case, we expect that renormalization group approaches describe the full physics more accurately, but nevertheless the Bogoliubov Hamiltonian predicts the anticipated critical temperature. Historically, it was found that the dilute gas in 2D ($\rho^{1/2}a\ll1$) has a critical temperature \cite{Popov} in terms of $b=1/|\ln(\rho a^2)|\ll1$,
\begin{equation}
T_{\rm{c}}\approx-\frac{4\pi\rho}{\ln(b)},
\end{equation}
and this was confirmed with Bogoliubov theory by Fisher and Hohenberg \cite{FisherHohenberg} and later proved as an exact upper bound in \cite{SeiUel-09}. We should point out that both \cite{Popov} and \cite{FisherHohenberg} assume $\ln(1/b)\gg1$ instead of the appropriate and weaker $b\ll1$ \cite{Review2D, Schick}, and that \cite{FisherHohenberg} still requires renormalization group techniques. Here, we only use $b\ll1$ and the variational model to find leading behaviour
\begin{equation}
\label{this expression}
T_{\rm{c}}=4\pi\rho\left(\frac{1}{\ln(\xi/4\pi b)}+o(1/\ln^2{b})\right)
\end{equation}
with $\xi=14.4$. Besides the numerical prediction of $\xi=380$ \cite{Prokof'ev}, no calculations of $\xi$, in particular no analytical ones, were ever done.
\section{Set-up}
We start from the Hamiltonian for a gas of $N$ bosons with a repulsive pair interaction $V$ in a $n$-dimensional box $\left[-l/2,l/2\right]^n$ and periodic boundary conditions, where $n=2,3$. In units $\hbar=2m=k_B=1$,
\begin{equation}
\label{HN1}
H_N=\sum_{1\leq i\leq N}-\Delta_i+\sum_{1\leq i<j\leq N}V_{ij},
\end{equation}
with second-quantized form in momentum space
\begin{equation}
\label{HN}
H=\sum_p p^2 a^\dagger_p a_p+\frac{1}{2l^n}\sum_{p,q,k} \widehat{V}(k) a_{p+k}^\dagger a_{q-k}^\dagger a_q a_p.
\end{equation}
The \textit{canonical} Gibbs state at temperature $T$ and particle density $\rho=N/l^n$ can be found by minimizing
\begin{equation}
\label{tomin2}
\inf_{\omega}\big[\langle H_N\rangle_\omega-TS(\omega)\big],
\end{equation}
where $\omega$ is an $N$-boson state and $S$ is the von Neumann entropy.
The \textit{grand canonical} Gibbs state at temperature $T$ and chemical potential $\mu$ is the minimizer of
\begin{equation}
\label{tomin}
\inf_{\omega}\big[\langle H-\mu\mathcal{N}\rangle_\omega-TS(\omega)\big],
\end{equation}
where $\omega$ is now a state on the bosonic Fock space, $\mathcal{N}$ is the particle number operator and the infimum itself is the \textit{free energy}.
Throughout this paper we are working in the \textit{thermodynamic limit} $l\to\infty$. The two quantities \eqref{tomin2} and \eqref{tomin} are then related by a Legendre transform.
We say that a system displays \textit{BEC} if the 1-particle reduced density matrix of the minimizing $\omega$ of \eqref{tomin} has an eigenvalue of order 1 \cite{PenOns}. Therefore, \textit{one needs to find minimizers of \eqref{tomin} to determine $T_{\rm c}$}.
This cannot be done exactly: the exact free energy \eqref{tomin} has only been analysed in \cite{Sei-08, Yin-10}; all other results, such as \cite{ZagBru-01}, concern approximations. We will study one such model \cite{CriSol-76}. It restricts the minimization problem \eqref{tomin} to \textit{quasi-free states}, resulting in a variational upper bound to the free energy.
There are good arguments why this upper bound is accurate. The first is that Bogoliubov's approach renders the Hamiltonian quadratic in creation and annihilation operators. Ground and Gibbs states of such Hamiltonians are quasi-free states; exactly the states considered in our minimization problem. Also, quasi-free states are good trial states for the ground state energy of Bose gases \cite{Solovej-06,ErdSchYau-08,GiuSei-09} and may therefore also be for the free energy.
As we shall soon see, expressing $\langle H-\mu\mathcal{N}\rangle_\omega-TS(\omega)$ for a general quasi-free state leads to a non-linear functional \eqref{funct}. Linearizing the functional by removing the terms quartic in creation and annihilation operators, the authors of \cite{CriSol-76} conclude that the Gibbs state coincides with that of Bogoliubov's (approximated) Hamiltonian.
Motivated by the discovery in \cite{ErdSchYau-08} that the correct first-order energy is only found when the terms quartic in creation and annihilation operators are included, we consider the functional \textit{without} the linearization---hence including interacting terms that were ignored in the original Bogoliubov approximation---and use it to give a variational calculation of $T_{\rm c}$.
\section{Model}
We expect the particles to form a condensate at momentum $p=0$. We therefore mimic Bogoliubov's c-number substitution (justified in \cite{LieSeiYng-05}) by using a Bogoliubov transformation to include a \textit{condensate density} $\rho_0\geq0$ in the Hamiltonian \eqref{HN}, effectively replacing $a_0\rightarrow a_0+\sqrt{l^n\rho_0}$. A minimizer with $\rho_0>0$ \textit{indicates (quasi-)condensation}, whereas $\rho_0=0$ signifies its absence.
We evaluate the expectation value of the resulting Hamiltonian for quasi-free states only, so that we can use Wick's rule to split $\langle a_{p+k}^\dagger a_{q-k}^\dagger a_q a_p\rangle$ as
\begin{equation}
\langle a^\dagger_{p+k}a^\dagger_{q-k}\rangle\langle a_q a_p\rangle+\langle a^\dagger_{p+k}a_q\rangle\langle a^\dagger_{q-k}a_p\rangle+\langle a^\dagger_{p+k}a_p\rangle\langle a^\dagger_{q-k}a_q\rangle.
\end{equation}
Assuming translation invariance and $\langle a_p a_{-p}\rangle=\langle a^\dagger_{-p} a^\dagger_p\rangle$, the two (real-valued) functions \mbox{$\gamma(p):=\langle a^\dagger_p a_p\rangle\geq0$} and \mbox{$\alpha(p):=\langle a_p a_{-p}\rangle$}, together with the number $\rho_0$, now fully determine the expectation value in \eqref{tomin}.
Here, $\gamma(p)$ is the \textit{density of particles with momentum $p$}, and $\alpha$ describes \textit{pairing in the system}. It is well-known that \mbox{$\alpha^2\leq\gamma(\gamma+1)$}.
Taking the thermodynamic limit $l\to\infty$, we have now evaluated the expectation in \eqref{tomin} for \textit{Bogoliubov trial states}---quasi-free states with an added condensate---resulting in the \textit{Bogoliubov free energy functional}
\begin{equation}
\label{funct}
\begin{aligned}
\mathcal{F}&_{\mu,T} (\gamma,\alpha,\rho_0)=(2\pi)^{-n}\int p^2\gamma(p)dp-\mu\rho-TS(\gamma,\alpha)\\
&+\rho_0(2\pi)^{-n}\int\widehat{V}(p)(\gamma(p)+\alpha(p))dp+\frac{1}{2}\widehat{V}(0)\rho^2\\
&+\frac{1}{2}(2\pi)^{-2n}\int\gamma(p)(\widehat{V}\ast\gamma)(p)+\alpha(p)(\widehat{V}\ast\alpha)(p)dp,
\end{aligned}
\end{equation}
with chemical potential $\mu\in\mathbb{R}$ and density \mbox{$\rho=\rho_0+\rho_\gamma$} (i.e.\ the sum of the condensate density $\rho_0$ and the density of particles with positive momentum \mbox{$\rho_\gamma=(2\pi)^{-n}\int\gamma$}).
The entropy, defined in terms of $\beta(p)=\sqrt{(\gamma(p)+\frac{1}{2})^2-\alpha(p)^2}$, is
\begin{equation}
\begin{aligned}
S(\gamma,\alpha)&=(2\pi)^{-n}\int\left(\beta(p)+\frac{1}{2}\right)\ln\left(\beta(p)+\frac{1}{2}\right)\\
&\hspace{2cm}-\left(\beta(p)-\frac{1}{2}\right)\ln\left(\beta(p)-\frac{1}{2}\right)dp.
\end{aligned}
\end{equation}
A precise derivation can be found in the appendix of \cite{NapReuSol1-15}.
For a canonical formulation with fixed \textit{average} density $\rho$ and temperature $T$, we consider
\begin{equation}
\label{canproblem}
\mathcal{F}^{\rm can}_{\rho,T}(\gamma,\alpha,\rho_0)=\mathcal{F}_{\mu,T} (\gamma,\alpha,\rho_0)+\mu\rho
\end{equation}
for states $(\gamma,\alpha,\rho_0)$ with $\rho_0+\rho_\gamma=\rho$. This amounts to evaluating the expectation in \eqref{tomin2} for Bogoliubov trial states.
In what follows, we drop the subscripts of $\mathcal{F}^{\rm can}_{\rho,T}$ and $\mathcal{F}_{\mu,T}$. Note that in contrast to \eqref{tomin2} and \eqref{tomin}, the infima of these functionals are not automatically related by a Legendre transform because of the restricted minimization. For example, the canonical infimum is not convex in 3D \cite{NapReuSol2-15}.
\section{Results}
We now restrict to 2D and assume that the two-body interaction potential is repulsive, integrable and bounded. Its Fourier transform $\widehat{V}$ is assumed to be positive and it has its maximum at zero since $V\geq0$.
The gas is dilute, so $\rho^{1/2}a\ll1$ in 2D, which implies $b=1/|\ln(\rho a^2)|\ll1$ for the expansion parameter. To find $T_{\rm c}$, we consider temperatures that satisfy $\sqrt{T}a\ll1$; otherwise the first line in \eqref{funct} dominates and it is easy to show that $\rho_0=0$.
We also use
\begin{equation}
\label{expand}
\widehat{V}(p)=\widehat{V}(0)+Ca^2p^2+o(a^2p^2),
\end{equation}
where the first derivative is absent since $\widehat{V}$ has its maximum at zero, and the second derivative is assumed to be of order $a^2$ (in accordance with its units).\footnote{In 3D, $\rho^{1/3}a\ll1$ and $\widehat{V}(0)=O(a)$, modifying \eqref{expand} accordingly.}
In paper \cite{NapReuSol1-15}, we prove that there exist minimizers for both the grand canonical \eqref{tomin} and canonical \eqref{tomin2} minimization problems when the minimization is restricted to Bogoliubov trial states (resulting in a minimization of the functionals \eqref{funct} and \eqref{canproblem}, respectively): the grand canonical phase diagram is shown in fig.\ \ref{phasediagram}. In particular, there is a phase transition at positive $T$ for all fixed $\mu>0$. Also note that $\alpha\neq0$ iff $\rho_0>0$, so that (quasi-)condensation and pairing always occur together. Canonically, there is a phase transition at positive $T$ for all fixed $\rho>0$.
To calculate the critical temperature, the definition of diluteness $\rho^{1/2}a\ll1$ suggests that we study the canonical functional \eqref{canproblem}.
\begin{figure}
\includegraphics[scale=0.28]{grand_can_diagram2.png}
\caption{The grand canonical phase diagram of the model. No diluteness is assumed. At $\mu\leq0$ and $T=0$, all quantities are zero, and there is no (quasi-)condensation. Increasing $T$ does not lead to a phase transition, although $\gamma$ becomes non-zero. For $\mu>0$ fixed and $T=0$, there is (quasi-)condensation. This remains the case when $T$ increases (darkest region), eventually leading to a phase transition somewhere in the lighter region before we enter the white region where $\rho_0=0$. We can only locate the phase transition exactly for $\mu\to0$; this corresponds to the dilute limit studied in Theorem \ref{Crit}.}
\label{phasediagram}
\end{figure}
\begin{theorem}
\label{Crit}
Consider the canonical problem \eqref{tomin2} in 2D restricted to Bogoliubov trial states with $\rho=\rho_0+\rho_\gamma$ fixed, resulting in the canonical functional \eqref{canproblem} with $n=2$. The critical temperature, defined by the properties $\rho_0> 0$ if $T>T_{\rm c}$, $\rho_0= 0$ if $0\leq T<T_{\rm c}$, is
\begin{equation}
\label{critT}
T_{\rm{c}}=4\pi\rho\left(\frac{1}{\ln(\xi/4\pi b)}+o(1/\ln^2{b})\right),
\end{equation}
with $\xi=14.4$ for $\widehat{V}(0)\approx8\pi b$.
\end{theorem}
The proof relies on a careful expansion of the free energy.
\section{Proof of Theorem \ref{Crit}}
We fix $T$ and aim to find the critical density $\rho_{\rm c}$, which can easily be inverted to \eqref{critT}.
\textit{Outline.}\hspace{0.5cm} We could try to minimize the functional by solving the Euler--Lagrange equations of \eqref{funct}. However, the terms with convolutions give non-local contributions $\widehat{V}\ast\gamma$ and $\widehat{V}\ast\alpha$. Even with a Fourier transform, this cannot be solved.
We therefore approximate these terms so that we obtain a functional $\mathcal{F}^{\rm sim}$ that \textit{can} be minimized explicitly in $\gamma$ and $\alpha$. We expand the resulting energy integrals, and finally minimize in $\rho_0$ to determine whether it is zero or not.
To sketch this once more,
\begin{equation}
\label{appr}
\inf_{\substack{\text{$(\gamma,\alpha,\rho_0)$}\\\text{$\rho_0+\rho_\gamma=\rho$}}}
\mathcal{F}^{\rm can}\approx\inf_{0\leq\rho_0\leq\rho}\inf_{\substack{\text{$(\gamma,\alpha)$}\\\text{$\rho_\gamma=\rho-\rho_0$}}}\mathcal{F}^{\rm sim},
\end{equation}
where the infimum over $\gamma$ and $\alpha$ is calculated explicitly, then expanded in $b\ll1$, and finally minimized in $\rho_0$.\\
\textit{Step 1a.}\hspace{0.2cm} To approximate the convolution term involving $\gamma$, we use a comparison with the free Bose gas. Its energy is given exactly by
\begin{equation}
\mathcal{F}_{0}(\gamma)=(2\pi)^{-2}\int p^2\gamma(p)dp-TS(\gamma,0),
\end{equation}
whose minimizer for fixed $\rho$ is
\begin{equation}
\label{sommin}
\gamma_{\mu(\rho)}(p)=\frac{1}{e^{(p^2-\mu(\rho))/T}-1},
\end{equation}
where $\mu(\rho)\leq0$ is such that $(2\pi)^{-2}\int\gamma_{\mu(\rho)}=\rho$. Since the integral diverges as $\mu(\rho)\to0$, this definition works for all $\rho\geq0$.
We would like to show that the minimizing $\gamma$ for the interacting problem lives on the same scale as $\gamma_{\mu(\rho)}$, that is, most particles have momentum $|p|\leq O(\sqrt{T})$. Indeed, a careful comparison shows that the minimizer has to satisfy
\begin{equation}
\label{step1}
\mathcal{F}_0(\gamma_{\mu(\rho_\gamma)})\leq \mathcal{F}_0(\gamma)\leq \mathcal{F}_0(\gamma_{\mu(\rho)})+\rho^2\widehat{V}(0),
\end{equation}
which says that the energy does not deviate much from the minimal energy in the free case. Because of the $\int p^2\gamma$ term, this means that $\gamma$ cannot be very large for \mbox{$|p|\gg\sqrt{T}$}.
For $|p|\leq O(\sqrt{T})$, \eqref{expand} implies that \mbox{$|\widehat{V}(p)-\widehat{V}(0)|=O(Ta^2)\ll 1$}.
Since $\gamma$ is only large on $|p|\leq O(\sqrt{T})$, we can approximate $(2\pi)^{-2}\widehat{V}\ast\gamma\approx\widehat{V}(0)\rho_\gamma$.
To be more precise: a careful analysis shows
\begin{equation}
\left|(2\pi)^{-2}\rho_0\int\widehat{V}(p)\gamma(p)dp
- \rho_0\widehat{V}(0)\rho_\gamma\right|= O(\rho^2(\rho^{1/2}a)),
\end{equation}
and similar estimates for the term involving the convolution. This error is negligible compared to contributions of order $T^2b$---the order to which we shall expand the energy---for example if $\rho\leq O(T/\sqrt{b})$. The latter we can assume without loss of generality by showing that $\mathcal{F}^{\rm can}(\gamma,0,0)$ grows rapidly compared to $\mathcal{F}^{\rm can}(0,0,\rho_\gamma)$, demonstrating that $\rho_0>0$ if the density is large enough.\\
\textit{Step 1b.}\hspace{0.2cm} The strategy for the convolution term with $\alpha$ is different.
Adapting ideas in \cite{ErdSchYau-08} to the 2D case, we expect $\alpha$ to be related to the function $w:=2bw_0$, where $w_0$ is the \textit{scattering solution} satisfying
\begin{equation}
-\Delta w_0+\frac12 Vw_0=0,
\end{equation}
with $w_0(x)\sim \ln(|x|/a)$ as $|x|\to\infty$.
To work towards a good approximation for $\alpha$, we define
\begin{equation}
\label{alpha0}
\alpha_0:=(\rho_0+t_0)\widehat{w}-(2\pi)^2 \rho_0\delta_0,
\end{equation}
where $\delta_0$ is a delta function and $-\rho_0\leq t_0\leq0$ is an additional parameter that will be tuned to achieve a self-consistency equation \mbox{$\int(\alpha-\alpha_0)=0$}.
In 2D, $w$ has logarithmic asymptotic behaviour, and its Fourier transform is more complicated than in the 3D case. Computing it, we find
\begin{equation}
\label{alpha02}
\alpha_0=(2\pi)^2 t_0\delta_0-(\rho_0+t_0)\left(\frac{\widehat{Vw}(p)}{2p^2}\chi_{|p|> p_0}-\langle h,.\rangle\right),
\end{equation}
where $\chi$ indicates an indicator function and $p_0$ is a momentum scale $2\rho^{1/2}e^{-\Gamma}$ involving the Euler--Mascheroni constant $\Gamma$.
The third contribution in \eqref{alpha02} is a distribution that acts on test functions $f$ as
\begin{equation}
\langle h,f\rangle=\int_{|p|\leq p_0}\frac{\widehat{Vw}(p)f(p)-\widehat{Vw}(0)f(0)}{2p^2}.
\end{equation}
The motivation for defining $\alpha_0$ in this way is as follows: at momentum scales bigger than $\sqrt{Tb}$, we expect $\alpha$ to by related to the scattering solution. Its structure on smaller scales is more complicated, but the exact shape is irrelevant. We approximate this part by a $\delta$-function and eventually optimize our approximation in $t_0$.
So how does the guess \eqref{alpha0} help? We add and subtract terms to replace the convolution term with $\alpha$ by
\begin{equation}
\int(\alpha-\alpha_0)(p)(\widehat{V}\ast(\alpha-\alpha_0))(p)dp,
\end{equation}
which we later show to be small for the minimizing $\alpha$.
By doing this we have of course introduced terms involving $\widehat{V}\ast\alpha_0$, but
\begin{equation}
(2\pi)^{-2}\widehat{V}*\alpha_0(p)=(\rho_0+t_0)\widehat{Vw}(p)-\rho_0\widehat{V}(p),
\end{equation}
so that no convolution terms remain in our functional. This has the added effect that $\widehat{V}$ gets replaced by $\widehat{Vw}$ in the term linear in $\alpha$, but this Fourier transform is well-defined and satisfies $\widehat{Vw}(0)=8\pi b$. To simplify the resulting functional, we make sure to obtain a similar replacement for the $\rho_0\int\widehat{V}\gamma$-term.\\
\textit{Step 1c.}\hspace{0.2cm} To specify \eqref{appr}, we define
\begin{equation}
\label{errors}
\begin{aligned}
E_1&:=\frac12(2\pi)^{-4}\int(\alpha-\alpha_0)(p)(\widehat{V}\ast(\alpha-\alpha_0))(p)dp\\
E_2&:=\rho_0\left((2\pi)^{-2}\int\widehat{V}(p)\gamma(p)dp-\widehat{V}(0)\rho_\gamma\right)\\
E_3&:=-(\rho_0+t_0)\left((2\pi)^{-2}\int\widehat{Vw}(p)\gamma(p)dp-\widehat{Vw}(0)\rho_\gamma\right)\\
E_4&:=\frac12(2\pi)^{-4}\int\gamma(p)(\widehat{V}\ast\gamma)(p)dp-\frac12\widehat{V}(0)\rho_\gamma^2,
\end{aligned}
\end{equation}
and following Steps 1a and 1b, we now derive a simplified functional $\mathcal{F}^{\rm sim}$ satisfying\
\begin{equation}
\label{bounds}
\mathcal{F}^{\rm{can}}(\gamma,\alpha,\rho_{0})-\mathcal{F}^{\rm{sim}}(\gamma,\alpha,\rho_0)=E_1+E_2+E_3+E_4.
\end{equation}
\textit{Step 2a.}\hspace{0.2cm} Now that we have $\mathcal{F}^{\rm sim}$, we calculate and expand its minimum as a function of $\rho$ and $\rho_0$. It turns out that we need to expand to order $T^2b$ to derive \eqref{critT}.
The part of $\mathcal{F}^{\rm sim}$ that depends on $\gamma$ and $\alpha$ is
\begin{equation}
\label{simplestf}
\begin{aligned}
(2\pi)^{-2}\left[\right.\int p^2\gamma(p)dp+(\rho_0+t_0)\int\widehat{Vw}(p)(\gamma(p)+\alpha(p))dp\\
+(\rho_0+t_0)^2\int\frac{\widehat{Vw}(p)^2-\chi_{|p|\leq p_0}\widehat{Vw}(0)^2}{4p^2}dp\left.\right]-TS(\gamma,\alpha).
\end{aligned}
\end{equation}
Given $\rho$ and $\rho_0$ we find minimizers $\gamma^{\rho_0,\rho}$ and $\alpha^{\rho_0,\rho}$ by adding a Lagrange multiplier term $\delta\rho_\gamma$ (with $\delta\geq0$) to this expression.
The dominant contribution of \eqref{simplestf} to the energy $\mathcal{F}^{\rm sim}(\gamma^{\rho_0,\rho},\alpha^{\rho_0,\rho},\rho_0)$ is
\begin{equation}
\label{enexample}
(2\pi)^{-2}T\int\ln(1-e^{-T^{-1}\sqrt{(p^2+\delta)^2+2(p^2+\delta)(\rho_0+t_0)\widehat{Vw}(p)}})dp,
\end{equation}
which resembles the energy of the free gas. \\
\textit{Step 2b.}\hspace{0.2cm} We now expand \eqref{enexample}. To do this, we judiciously define
\begin{equation}
\label{replacements}
\rho_0=\frac{\sigma}{8\pi}T, \hspace{0.3cm} \rho=\frac{\ln(k)-\ln(b)}{4\pi}T, \hspace{0.3cm} \delta=dbT, \hspace{0.3cm} t_0=\frac{\tau}{\sigma}\rho_0,
\end{equation}
where $\sigma,k,d,\tau$ are parameters of order 1.
Changing variables $p\to \sqrt{T}p$ and using \eqref{expand} for $\widehat{Vw}$, we expand \eqref{enexample} for $b\ll1$, resulting in
\begin{equation}
\begin{aligned}
&\frac{T^2}{4\pi}\left[\right.-\frac16\pi^2-b\ln(b)(d+\sigma+\tau)+b\left(\right.d+\sigma+\tau\\
&-\frac12(d+2\sigma+2\tau)\ln(d+2\sigma+2\tau)-\frac12d\ln(d)\left.\right)+o(b)\left.\right].
\end{aligned}
\end{equation}
Of course, there is a relation between $d$, $k$ and $\sigma$, since $d$ was a Lagrange multiplier.
We eliminate $d$ by calculating, expanding and solving \mbox{$\int\gamma^{\rho_0,\rho}=\rho-\rho_0$}, finding
\begin{equation}
\label{d}
d=-(\sigma+\tau)+\sqrt{(\sigma+\tau)^2+e^\sigma/k^2}.
\end{equation}
\textit{Step 2c.}\hspace{0.2cm} We now use \eqref{bounds} to show that we can accurately approximate the energy. We claim the minimizers have to satisfy
\begin{equation}
|\mathcal{F}^{\rm can}(\gamma,\alpha,\rho_0)-\mathcal{F}^{\rm sim}(\gamma^{\rho_0,\rho},\alpha^{\rho_0,\rho},\rho_0)|=o(T^2b),
\end{equation}
There are two bounds to show.
The lower bound follows from \eqref{bounds} and the a priori results from Step 1a. Note that no a priori information on $\alpha$ is needed since $E_1\geq0$.
For the upper bound, we simply verify that all errors in \eqref{errors} are $o(T^2b)$ for $\gamma^{\rho_0,\rho}$ and $\alpha^{\rho_0,\rho}$.
This is only non-trivial for $E_1$, for which it suffices to choose $\tau$ such that \mbox{$\int(\alpha^{\rho_0,\rho}-\alpha_0)=0$}. We conclude
\begin{equation}
\label{tau}
\tau=\ln\left(\frac{d}{d+2\sigma+2\tau}\right)+o(1),
\end{equation}
which can be used to eliminate $\tau$.\\
\textit{Step 3.}\hspace{0.2cm} We have now reduced \eqref{appr} to
\begin{equation}
\inf_{\substack{\text{$(\gamma,\alpha,\rho_0)$}\\\text{$\rho_0+\rho_\gamma=\rho$}}}\mathcal{F}^{\rm can}=\frac{T^2}{4\pi}\cdot\inf_{\sigma\geq0}f(k,\sigma)+o(T^2b),
\end{equation}
with, in the specific case $\widehat{V}(0)\to8\pi b$,
\begin{equation}
\label{f}
\begin{aligned}
f&(k,\sigma)=-\frac16\pi^2+2b\ln^2b-4b\ln(b)\ln(k)\\
&+b\left[\right.d+\sigma+\tau-(\sigma+\tau)\ln(d+2\sigma+2\tau)\\
&+2\ln^2k+\frac14(\sigma+\tau)^2-\frac12\tau^2-(\sigma+\tau)\ln(k)\left.\right] ,
\end{aligned}
\end{equation}
and $d$ and $\tau$ as in \eqref{d} and \eqref{tau}. We can determine $\rho_{\rm c}$ by fixing $k$ and checking when the minimizing $\sigma$ changes from zero to non-zero, as illustrated in fig.\ \ref{123}. We find $\rho_{\rm c}=\ln(1.145/b)T/4\pi$, which can be rewritten as the desired $T_{\rm c}$ \eqref{critT}. Note that this strategy works for any value of $\widehat{V}(0)$, not just $8\pi b$, but the expression \eqref{f} simplifies in this case.
\begin{figure}
\includegraphics[scale=0.42]{Plot3.pdf}
\caption{Plots of the part of the free energy $f(k,\sigma)$ between the square brackets in \eqref{f} for three values of $k$, where $\rho=\ln(k/b)T/4\pi$ and $\rho_0=\sigma T/8\pi$. For $k=1.1$, $\sigma=\rho_0=0$ gives the lowest energy: no quasi-condensation. For $k=1.2$, the minimum occurs at some $\rho_0>0$: quasi-condensation. The critical value is $k_c=1.145$.}
\label{123}
\end{figure}
\section{Conclusion and discussion}
Adapting our earlier calculation in 3D, we analytically calculate the Kosterlitz-Thouless transition temperature of a dilute, translation-invariant Bose gas in 2D, finding \eqref{this expression} with $\xi=14.4$.
We use a variational model that can be seen as a reformulation of Bogoliubov's approximation, although, in fact, it is slightly more accurate. The relevant trial states are quasi-free states with an added (quasi-)condensate. As any approach that relies on Bogoliubov theory, this produces unphysical results like an incorrect first-order phase transition to BEC in 3D, but, perhaps surprisingly, Bogoliubov's approximation does make accurate predictions for both the 2D and 3D critical temperatures.
To improve the model one would have to extend the class of trial states, which is both very interesting and very challenging. Another compelling direction for future research is to ask what role superfluidity plays in this model, and what we can learn from it.
\acknowledgments
We thank Robert Seiringer and Daniel Ueltschi for bringing the issue of the change in critical temperature to our attention. We also thank the Erwin Schr\"odinger Institute (all authors) and the Department of Mathematics, University of Copenhagen (MN) for the hospitality during the period this work was carried out. We gratefully acknowledge the financial support by the European Unionβs Seventh Framework Programme under the ERC Grant Agreement Nos. 321029 (JPS and RR) and 337603 (RR) as well as support by the VILLUM FONDEN via the QMATH Centre of Excellence (Grant No. 10059) (JPS and RR), by the National Science Center (NCN) under grant No. 2016/21/D/ST1/02430 and the Austrian Science Fund (FWF) through project Nr. P 27533-N27 (MN).
| e226f3ee2645d64d70d7aa1736c616b085547257 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In the study of arithmetic geometry over a Noetherian local ring $R$, we are often faced with the problem to recover some arithmetic invariant attached to a given module $M$ over $R$ from those attached to modules $M/IM$ over $R/I$, where $I$ runs through a certain set of principal ideals of $R$. We call it a \textit{specialization method} here. A typical work of this type appears in \textit{Iwaswa main conjecture for two-variable Hida deformations} of the first named author's articles \cite{Oc1} and \cite{Oc2}. In these papers, we considered the characteristic ideal of the Selmer group attached to the multi-variable Galois deformation with a continuous action of the Galois group $G_{\mathbb{Q}}=\operatorname{Gal}(\overline{\mathbb{Q}}/\mathbb{Q})$ over a certain Galois deformation ring $R$, under the assumption that $R \simeq \mathcal{O}[[x_1,\ldots,x_d]]$, where $\mathcal{O}$ is a complete discrete valuation ring (note that $R$ is of Krull dimension $d+1$). It was quite essential to assume the regularity of $R$ in \cite{Oc1}
In our previous work \cite{OcSh}, we proved the \textit{local Bertini theorem} for the normality on a Noetherian complete local ring $(R,\frak{m})$ with an extra assumption that $\operatorname{depth} R \ge 3$, which forces $\dim R \ge 3$. In fact, it was already visible in the work \cite{Oc1} that for $\mathcal{O}[[x_1,\ldots,x_d]]$, the case $d \ge 2$ is quite different from the case $d=1$. The present article as well as \cite{OcSh} attempt to extend the results of \cite{Oc1} to the case that $R$ is a normal domain which is torsion free and finite over $\mathbb{Z}_p[[x_1,\ldots,x_d]]$. In the present article, we complete the missing case $\dim R=2$. With these results at hand, we establish Euler system bound over Cohen-Macaulay normal domains and one of the divisibilities in the Iwasawa main conjecture for two-variable Hida deformation without assuming that the branch of the Hecke algebra is isomorphic to the power series ring $\mathcal{O}[[X,Y]]$. The method of Euler system is quite powerful in giving an upper bound of a size of a Selmer group. In this article, it will be important to consider the $\Lambda$-adic (Hida-theoretic) version of Euler system whose specialization at an arithmetic point coincides with Beilinson-Kato Euler system constructed from the elements of the $K_2$-group of modular curves.
The structure of this paper is twofold. The first part is concerned with commutative algebra and the second part is concerned with applications of results from commutative algebra to Iwasawa theory.
\subsection{Main results in commutative algebra}
To establish the specialization method in Krull dimension two, we need the local Bertini theorem in this setting. However, if $R$ is a two-dimensional local normal domain and $I$ is a nonzero principal ideal, then $R/I$ is normal if and only if $R/I$ is regular. For this reason, it is too much to expect that $R/I$ is normal. Thus, we attempt to find a large set of specializations ideals $\{I_{\lambda}\}$ for which $R/I_{\lambda}$ are reduced rings of mixed characteristic. The local Bertini theorem is stated as follows (see Theorem \ref{MainBertini}, Corollary \ref{LocalBertini} and Corollary \ref{LocalBertini2}), which is of independent interest in commutative algebra (see Remark \ref{Bertinifail} for some relevant remarks)
\begin{maintheorema}
\label{mainresult1}
Assume that $B \hookrightarrow C$ is a module-finite extension such that $B$ is a Noetherian unique factorization domain, $C$ is a normal domain and $\operatorname{Frac}(B) \to \operatorname{Frac}(C)$ is a separable field extension. Let us define a subset $\mathcal{S}_{C/B}$ of $\operatorname{Spec} B$ as follows:
$$
\mathcal{S}_{C/B}=\{\frak{p} \in \operatorname{Spec} B~|~\frak{p}~\mbox{is a height-one prime in}~\operatorname{Ass}_B(\Omega_{C/B})\}.
$$
Then $\mathcal{S}_{C/B}$ is a finite set and $C/\frak{p} C$ is a reduced ring for any height-one prime $\frak{p} \in \operatorname{Spec} B$ such that $\frak{p} \notin \mathcal{S}_{C/B}$.\\
\end{maintheorema}
We need the theorem it when $B$ is a complete regular local domain in \S~\ref{controltheorem}. Assume that $(R,\frak{m},\mathbb{F})$ is a two-dimensional complete local normal domain with mixed characteristic and finite residue field. We define certain large sets $\mathscr{L}_{R,W(\mathbb{F})}(z,\{a_n\}_{n \in \mathbb{N}})$ and $\mathscr{E}_{R,W(\mathbb{F})}(z,r,\{a_n\}_{n \in \mathbb{N}})$ (see (\ref{Fittideal}) and (\ref{Fittideal1}) in \S~\ref{controltheorem}, respectively), which are comprised of principal ideals generated by \textit{specialization elements} in $R$ attached to finitely generated torsion $R$-modules and these elements play a role in the control of the behavior of characteristic ideals under specialization. Let $\operatorname{char}_R(M)$ denote the \textit{characteristic ideal} of a finitely generated torsion module $M$ over a Noetherian normal domain $R$ (see Definition \ref{charideal}). Characteristic ideals appear in the formulations of the Iwasawa main conjectures in various forms. By combining Main Theorem \ref{mainresult1}, we establish the following theorem (see Theorem \ref{prop1} and Theorem \ref{prop2} for precise statements).
\begin{maintheorema}
\label{mainresult2}
Let $(R,\frak{m},\mathbb{F})$ be a two-dimensional Noetherian complete local normal domain with mixed characteristic $p>0$ and finite residue field $\mathbb{F}$. Suppose that $M$ and $N$ are finitely generated torsion $R$-modules.
\begin{enumerate}
\item
The following statements are equivalent:
\begin{enumerate}
\item
Let $\frak{q}$ be any height-one prime of $R$ which does not lie over $p$. Then we have
$$
\operatorname{char}_R(N)_{\frak{q}} \subseteq \operatorname{char}_R(M)_{\frak{q}}.
$$
\item
There exists a constant $c \in \mathbb{N}$, depending on $M$ and $N$, such that
$$
c \cdot \frac{|N/\mathbf{x}N |}{|M/\mathbf{x}M |} \in \mathbb{N}
$$
for all but finitely many principal ideals
$$
(\mathbf{x}) \in \bigcup_{1\le i \le k} \mathscr{L}_{R,W(\mathbb{F})}(z_i,\{a_n\}_{n \in \mathbb{N}}).
$$
\end{enumerate}
\item
The following statements are equivalent:
\begin{enumerate}
\item
Let $\frak{p}$ be any height-one prime of $R$ which lies over $p$. Then we have
$$
\operatorname{char}_R(N)_{\frak{p}} \subseteq \operatorname{char}_R(M)_{\frak{p}}.
$$
\item
There exists a constant $c \in \mathbb{N}$, depending on $M$ and $N$, such that
$$
c \cdot \frac{|N/\mathbf{x}N |}{|M/\mathbf{x}M |} \in \mathbb{N}
$$
for all but finitely many principal ideals
$$
(\mathbf{x}) \in \bigcup_{1\le i \le h} \mathscr{E}_{R,W(\mathbb{F})}(z,r_i,\{a_n\}_{n \in \mathbb{N}}).
$$\\
\end{enumerate}
\end{enumerate}
\end{maintheorema}
For arithmetic applications, we will need to combine Main Theorem \ref{mainresult2} with a modified version of \cite[Theorem 8.8]{OcSh} in a practical form (see Theorem \ref{theorem:previous}).
\subsection{Arithmetic applications}
We study \textit{Euler system bound} associated with a $p$-adic Galois representation over a complete Noetherian local ring (see Definition \ref{definitionEuler} and the book \cite{Rub} for the classical Euler system theory over a discrete valuation ring). As remarked in the beginning of the introduction, the method of Euler system is useful for finding an upper bound of the size of a Selmer group, hence for proving one of the predicted divisibilities of the Iwasawa main conjecture. First, we state Euler system bound for a $p$-adic Galois representation with coefficients in a complete Noetherian reduced local ring that is torsion free and finite over the ring of $p$-adic integers (see Theorem \ref{Euler}).
\begin{maintheorema}
\label{mainresult3}
Let $(R,\frak{m},\mathbb{F})$ be a complete Noetherian reduced local ring which is finite and flat over $\mathbb{Z}_p$ for a prime number $p>2$. Suppose that
$$
\big\{\mathbf{z}_n\in H^1(G_{\Sigma,n},T^*(1))\big\}_{n \in \mathfrak{N}}
$$
is an Euler system for $(T,R,\Sigma)$, $T$ is a free $R$-module of rank two and suppose that the following conditions hold:
\begin{enumerate}
\item[\rm{(i)}]
$T \otimes_R R/\frak{m}$ is absolutely irreducible as a representation of $G_{\mathbb{Q}}$.
\item[\rm{(ii)}]
The quotient $H^1(G_{\Sigma},T^*(1))/R \mathbf{z}_1$ is a finite group.
\item[\rm{(iii)}]
$H^2(G_{\mathbb{Q}_{\ell}},T^*(1))=0$ for every $\ell \in \Sigma\setminus\{\infty\}$ and $H^2(G_{\Sigma},D)$ is a finite group.
\item[\rm{(iv)}]
The determinant character $\wedge^2 \rho:G_{\mathbb{Q}} \to R^{\times}$ $($resp. $\wedge^2 \rho^*(1):G_{\mathbb{Q}} \to R^{\times}$$)$ associated with the $G_{\mathbb{Q}}$-representation $T$ (resp. $T^*(1)$) has an element of infinite order.
\item[\rm{(v)}]
The $R$-module $T$ splits into eigenspaces: $T=T^+ \oplus T^-$ with respect to the complex conjugation in $G_{\mathbb{Q}}$, and $T^+_{\frak{p}}$ (resp. $T^-_{\frak{p}}$) is of $R_{\frak{p}}$-rank one for each minimal prime $\frak{p}$ of $R$.
\item[\rm{(vi)}]
There exist $\sigma_1 \in G_{\mathbb{Q}(\mu_{p^{\infty}})}$ and $\sigma_2 \in G_{\mathbb{Q}}$ such that $\rho(\sigma_1) \simeq \begin{pmatrix} 1 & \epsilon \\ 0 & 1 \end{pmatrix} \in GL_2(R)$ for a nonzero divisor $\epsilon \in R$ and $\sigma_2$ acts on $T$ as multiplication by $-1$.
\end{enumerate}
Then Euler system bound holds for $(T,R,\Sigma)$:
\begin{enumerate}
\item
$\textcyr{Sh}^2_{\Sigma}(T^*(1))$ is a finite group.
\item
We have
$$
c \cdot |H^1(G_{\Sigma},T^*(1))/R \mathbf{z}_1|~\mbox{is divisible by}~| \textcyr{Sh}^2_{\Sigma}(T^*(1))|,
$$
where $c:=|R/(\epsilon^k)|$ is an error term for Euler system bound and $k$ is the number of minimal generators of the $R$-module $\textcyr{Sh}^2_{\Sigma}(T^*(1))$.\\
\end{enumerate}
\end{maintheorema}
Main Theorem \ref{mainresult3} has been known to hold when $R$ is a complete discrete valuation ring. We reduce the proof of the theorem to the case where $R$ is a product of complete discrete valuation rings after normalizing the reduced ring $R$ in its total ring of fractions. Euler system bound is less powerful, if the image of the Galois representation attached to an ordinary modular form is small. This phenomena occurs when the modular form under consideration has \textit{complex multiplication}. The paper \cite{OcPr} discusses the case of elliptic cusp forms with complex multiplication and \cite{HaOc15} generalized it to the case of Hilbert modular cusp forms with complex multiplication.
Our primary concern, as in the hypothesis $\rm(iv)$ of Main Theorem \ref{mainresult3}, is when the Galois representation has big image and $R$ is a certain (Hecke) deformation ring. As we will see soon, the Main Theorem \ref{mainresult3} is not a mere generalization of the classical Euler system bound over a discrete valuation ring. The theorem is used to deduce Euler system bound for a $p$-adic Galois representation with coefficients in a complete Noetherian local domain of Krull dimension at least two. Indeed, we obtain, from Theorem \ref{theorem:previous} (to descend from Krull dimension at least three to two) and Main Theorem \ref{mainresult2}, the following Euler system bound over a Cohen-Macaulay normal domain (see Theorem \ref{veryfinal}):
\begin{maintheorema}
\label{mainresult4}
Let $(R,\frak{m},\mathbb{F})$ be a Noetherian complete local Cohen-Macaulay normal domain
of Krull dimension $d\geq 2$ with mixed characteristic $p>2$ and finite residue field $\mathbb{F}$. Suppose that
$$
\big\{\mathbf{z}_n\in H^1(G_{\Sigma,n},T^*(1))\big\}_{n \in \mathfrak{N}}
$$
is an Euler system for $(T,R,\Sigma)$, $T$ is a free $R$-module of rank two with
continuous $G_{\mathbb{Q}}$-action and suppose that the following conditions hold:
\begin{enumerate}
\item[\rm{(i)}]
$T \otimes_R R/\frak{m}$ is absolutely irreducible as a representation of $G_{\mathbb{Q}}$.
\item[\rm{(ii)}]
The quotient $H^1(G_{\Sigma},T^*(1))/R \mathbf{z}_1$ is an $R$-torsion module.
\item[\rm{(iii)}]
$H^2(G_{\mathbb{Q}_{\ell}},T^*(1))=0$ for every $\ell \in \Sigma\setminus\{\infty\}$ and $H^2(G_{\Sigma},D)$ is a finite group.
\item[\rm{(iv)}]
The determinant character $\wedge^2 \rho:G_{\mathbb{Q}} \to R^{\times}$ (resp. $\wedge^2 \rho^*(1):G_{\mathbb{Q}} \to R^{\times}$) associated with the $G_{\mathbb{Q}}$-representation $T$ (resp. $T^*(1)$) has an element of infinite order.
\item[\rm{(v)}]
The $R$-module $T$ splits into eigenspaces: $T=T^+ \oplus T^-$ with respect to the complex conjugation in $G_{\mathbb{Q}}$, and $T^+$ (resp. $T^-$) is of $R$-rank one.
\item[\rm{(vi)}]
There exist $\sigma_1 \in G_{\mathbb{Q}(\mu_{p^{\infty}})}$ and $\sigma_2 \in G_{\mathbb{Q}}$ such that $\rho(\sigma_1) \simeq \begin{pmatrix} 1 & P \\ 0 & 1 \end{pmatrix} \in GL_2(R)$ for a nonzero element $P \in R$ and $\sigma_2$ acts on $T$ as multiplication by $-1$.
\end{enumerate}
Then Euler system bound holds for $(T,R,\Sigma)$:
\begin{enumerate}
\item
$\textcyr{Sh}^2_{\Sigma}(T^*(1))$ is a finitely generated torsion $R$-module.
\item
We have an inclusion of reflexive ideals:
$$
(P^k ) \operatorname{char}_R\big(H^1 (G_{\Sigma},T^*(1))/R \mathbf{z}_1\big) \subseteq \operatorname{char}_R\big(\textcyr{Sh}^2_{\Sigma} (T^*(1))\big),
$$
where $k$ is the number of minimal generators of $\textcyr{Sh}^2_{\Sigma}(T^*(1))$ as an $R$-module.\\
\end{enumerate}
\end{maintheorema}
The proof of Main Theorem \ref{mainresult4} is reduced to Main Theorem \ref{mainresult3} by an inductive argument with respect to the Krull dimension of the Cohen-Macaulay normal domain $R$. The most essential part of hypotheses in Main Theorem \ref{mainresult4} is:
\begin{enumerate}
\item[-]
$H^1(G_{\Sigma},T^*(1))/R \mathbf{z}_1$ is a torsion $R$-module.
\end{enumerate}
This is deeply related to the non-triviality of Euler system, which is conjectured to yield the existence of a $p$-adic $L$-function. We will apply the theorem in the case that $T=\mathcal{T}^{\operatorname{n.ord}}$ is a \textit{nearly ordinary Hida deformation space} attached to a Hida family $\mathbf{f}$. The hypotheses in Main Theorem \ref{mainresult4} are not so restrictive, except for the vanishing of $H^2(G_{\mathbb{Q}_{\ell}},T^*(1))$. At the moment, this seems necessary for the proof to work and it is related to the local automorphic representation of $GL_2(\mathbb{Q}_{\ell})$ spanned by an arithmetic specialization $\mathbf{f}_{\kappa}$ of $\mathbf{f}$. However, we believe that it suffices to assume that $H^2(G_{\mathbb{Q}_{\ell}},T^*(1))$ is an $R$-torsion module, although it is not clear at the moment. See \cite{Oc2} for the detailed study of the local monodromy of $\mathcal{T}^{\operatorname{n.ord}}$. Now we state the Iwasawa main conjecture for two-variable Hida deformations (see \S~\ref{Hidafamily} for notation and the conditions $(\bf{NOR})$, $(\bf{IRR})$ and $(\bf{FIL})$).
\begin{conjecture}
\label{MainConj}
Let $\mathbf{f}$ be a Hida family of $p$-ordinary $p$-stabilized cusp newforms of tame level $N$, and assume that $\mathcal{T}$ is the nearly ordinary Hida deformation associated with $\mathbf{f}$. Fix an $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$-basis $\mathbf{B}^{\pm}_{\mathbf{f}}$ of the modules of $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$-adic modular symbols $\mathbf{MS}^{\pm}_{\mathbf{f}}$. If the conditions $(\bf{NOR})$, $(\bf{IRR})$ and $(\bf{FIL})$ hold, then the Pontryagin dual of the Selmer group $\operatorname{Sel}_{\mathbb{Q}}(\mathcal{T})$ is a finitely generated, torsion $\mathbf{I}^{\operatorname{n.ord}}_\mathbf{f}$-module and we have
$$
\operatorname{char}_{\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}}\big(\operatorname{Sel}_{\mathbb{Q}}(\mathcal{T})^{\operatorname{PD}}\big)=\big(L_p(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})\big),
$$
where $L_p(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})$ is a two-variable $p$-adic $L$-function attached to $\mathbf{B}^{\pm}_{\mathbf{f}}$.\\
\end{conjecture}
The two-variable $p$-adic $L$-function $L_p(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})$ was constructed by Mazur and Kitagawa and we review its construction in this paper. In \cite{Hid0} and \cite{Hid1}, Hida showed that $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$ is a three-dimensional Noetherian complete local domain. This ring is called a \textit{branch} of the nearly ordinary Hecke algebra attached to $\mathbf{f}$ and we review its properties in \S~\ref{Hidafamily}. A partial answer to Conjecture \ref{MainConj} is given in \cite{Oc1} under the assumption that $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}} \simeq \mathcal{O}[[X,Y]]$. The Euler system bound as stated in Main Theorem \ref{mainresult4} is used to prove one of the divisibilities of Conjecture $\ref{MainConj}$ under the assumption that $R=\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$ is a normal domain. If $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$ is not a regular ring, the quotient $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}/\frak{a}$ can only be a reduced ring for an ideal $\frak{a}=(x,y)$ of height-two, as explained in the beginning of the introduction. This is the main reason why we need some technical results from commutative algebra, including authors' previous work \cite{OcSh}. As already said, the method of Euler system provides the following inclusion relation:
$$
(p\mbox{-adic}~L\mbox{-function}) \subseteq (\mbox{characteristic ideal of the Pontryagin dual of Selmer group})
$$
To achieve this, we will make an essential use of Beilinson-Kato Euler system constructed from elements of the $K_2$-group of modular curves. The following result is obtained as a partial answer to Conjecture \ref{MainConj} (see Theorem \ref{MainTh}).
\begin{corollary}
Assume that $(\bf{NOR})$, $(\bf{IRR})$ and $(\bf{FIL})$ hold for the nearly ordinary Hida deformation $\mathcal{T}$ attached to a Hida family $\mathbf{f}$. Fix an $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$-basis $\mathbf{B}^{\pm}_{\mathbf{f}}$ of the modules of $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$-adic modular symbols $\mathbf{B}^{\pm}_{\mathbf{f}}$. Assume further that
\begin{enumerate}
\item[\rm{(i)}]
There exists an element $\sigma_1 \in G_{\mathbb{Q}(\mu_{p^{\infty}})}$ such that $\rho_{\mathbf{f}}^{\operatorname{n.ord}}(\sigma_1) \simeq \begin{pmatrix} 1 & \epsilon \\ 0 & 1 \end{pmatrix} \in GL_2(\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}})$ for a nonzero element $\epsilon \in \mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$.
\item[\rm{(ii)}]
There exists an element $\sigma_2 \in G_{\mathbb{Q}}$ such that $\sigma_2$ acts on $\mathcal{T}$ as multiplication by $-1$.
\item[\rm{(iii)}]
If $\ell$ is any prime dividing $Np$, the maximal Galois invariant quotient vanishes; $(\mathcal{T}^*)_{G_{\mathbb{Q}_{\ell}}}=0$.
\end{enumerate}
Let $k$ be the number of minimal $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$-generators of $\textcyr{Sh}^2_{\Sigma}(\mathcal{T}^*(1))$. Then we have
$$
\big(\epsilon^k L_p(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})\big) \subseteq \operatorname{char}_{\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}}\big(\operatorname{Sel}_{\mathbb{Q}}(\mathcal{T})^{\operatorname{PD}}\big).
$$\\
\end{corollary}
\subsection{Outline of the paper}
In \S~\ref{LocalBertiniTheorem}, we prove some preliminary lemmas to prove the local Bertini theorem for reduced quotients. Then we prove its corollary which is fundamental in the specialization methods.
In \S~\ref{controltheorem}, we establish specialization methods on characteristic ideals over complete local normal domains with mixed characteristic in dimension two. Together with the main results in \cite{OcSh}, we obtain all algebraic tools for the specialization methods.
In \S~\ref{GaloisCoho}, after reviewing global (Poitou-Tate) and local (Tate) dualities of Galois cohomology groups over number fields, we prove some preliminary results of Galois cohomology with a coefficient ring that is a module-finite extension over either $\mathbb{F}_p[[x_1,\ldots,x_d]]$ or $\mathbb{Z}_p[[x_1,\ldots,x_d]]$ based on the techniques developed in Greenberg's paper \cite{Gr2}.
In \S~\ref{EulerSystem}, we prove Euler system bound with coefficient in a reduced local ring that is finite torsion free over $\mathbb{Z}_p$ as Theorem \ref{Euler}. Then we establish Euler system bound with coefficient in a Cohen-Macaulay normal domain as Theorem \ref{veryfinal}, whose proof is based on Theorem \ref{Euler} and the specialization methods.
In \S~\ref{Hidafamily}, we review the theory of $p$-adic modular forms and Hida theory to an extent we need, including Galois representations attached to a $\Lambda$-adic family of ordinary cusp forms (called a \textit{Hida family}) with basic properties. Then we introduce a two-variable $p$-adic $L$-function defined over a branch of the nearly ordinary Hecke algebra. To this aim, we define $p$-optimal complex period and $p$-adic period after introducing the module of modular symbols. We use these periods to formulate the interpolation formula of the two-variable $p$-adic $L$-function.
In \S~\ref{GaloisDeformation}, we begin by recalling Greenberg's Selmer group associated to the nearly ordinary Hida deformation. After giving a summary on Beilinson-Kato Euler system, we explain its $\Lambda$-adic version and a construction of $p$-adic $L$-function via Coleman's map. The rest of this section is devoted to proving finiteness results on local Galois cohomology groups.
In \S~\ref{proof2}, we apply our main result to Beilinson-Kato Euler systems over the nearly ordinary Hida deformation and we give a partial result on Conjecture \ref{MainConj} (see Theorem \ref{MainTh} and Corollary \ref{MainCo}).\\
\section{Local Bertini theorem for reduced quotients}
\label{LocalBertiniTheorem}
\subsection{Preliminary lemmas}
In this section, we prove that a Noetherian complete local normal domain $R$ of Krull dimension $\ge 2$ admits infinitely many nonzero principal ideals $\{(\mathbf{x}_n)\}_{n \in \mathbb{N}}$ such that $R/(\mathbf{x}_n)$ are reduced rings. This is a fundamental tool for the specialization method. In \cite{OcSh}, the authors proved the local Bertini theorem with its application to characteristic ideals for Noetherian complete local normal domains in dimension $\ge 3$. Note that the quotient of a two-dimensional local normal domain by a nonzero element is not normal, if the ring itself is not regular.
The nature of this section is pure commutative algebra, and we refer the reader to \cite{BrHer} and \cite{Mat} for relevant commutative algebra and to \cite{Kunz} for modules of K\"ahler differentials. We begin with some preliminary lemmas.
\begin{lemma}
\label{Koszul}
Let $f,g$ be a regular sequence in an integral domain $B$. Then, we have
$$
B=B[\tfrac{1}{f}] \cap B[\tfrac{1}{g}].
$$
\end{lemma}
\begin{proof}
The inclusion $B \subseteq B[\frac{1}{f}] \cap B[\frac{1}{g}]$ is clear. We will prove the other inclusion
$B[\frac{1}{f}] \cap B[\frac{1}{g}] \subseteq B$.
Let us take $x \in B[\frac{1}{f}] \cap B[\frac{1}{g}]$. Let $m$ be the smallest non-negative integer
such that $f^m x \in B$ and we put $a=f^m x$. We thus have $x =\frac{a}{f^m}$. We take another presentation
$x =\frac{b}{g^n}$ with $b\in B$ and $n$ a positive integer (note that we do not necessarily assume the minimality for $n$).
It suffices to prove that $m=0$. Assume $m>0$ to prove the lemma by contradiction. We have $bf^m=ag^n$.
Since $f,g$ is a regular sequence and $m,n$ are positive integers, $f^m ,g^n$ is also a regular sequence.
Hence, there exists an element $a' \in B$ such that $a=a'f$. We have
$x=\frac{a'}{f^{m-1}}$, which contradicts to the minimality of $m$. Hence $m=0$, showing that $x \in B$.
This completes the proof of the inclusion $B[\frac{1}{f}] \cap B[\frac{1}{g}] \subseteq B$.
\end{proof}
\begin{lemma}
\label{Koszul2}
Let $B \hookrightarrow C$ be a ring extension of normal domains. Let $(y) \subseteq B$ be a principal prime ideal, let $B_{(y)}$ denote the localization of $B$ at the prime $yB$ and let $C_{(y)}:= C \otimes_B B_{(y)}$. Assume the following conditions:
\begin{enumerate}
\item[\rm{(i)}]
$C$ is integral over $B$.
\item[\rm{(ii)}]
$C_{(y)}/(y)$ is a reduced ring.
\end{enumerate}
Then $C/(y)$ is a reduced ring.
\end{lemma}
\begin{proof}
Assume that $C/(y)$ is not reduced and deduce a contradiction. Then there exists $t \in C$ such that $t^N \in (y) \subseteq C$ and $t \notin (y) \subseteq C$ for some $N>0$. Since $C_{(y)}/(y)$ is reduced by assumption, we have $t \in (y) \subseteq C_{(y)}$ and we can write
$$
t=\tfrac{by}{x}~\mbox{for some}~x \in B \setminus (y)~\mbox{and}~b \in C.
$$
Since $t \in C$, we have $\frac{b}{x}=\frac{t}{y} \in C[\frac{1}{y}]$ and $\frac{b}{x} \in C[\frac{1}{x}]$. Since $B/(y)$ is an integral domain, $x$ is a regular element on $B/(y)$. Hence, $y,x$ is a regular sequence on $B$ and we have $B=B[\frac{1}{y}] \cap B[\frac{1}{x}]$ by Lemma \ref{Koszul}. On the other hand, we claim that
$$
B=B[\tfrac{1}{y}] \cap B[\tfrac{1}{x}] \to C[\tfrac{1}{y}] \cap C[\tfrac{1}{x}]
$$
is an integral extension. For this, let $\operatorname{Frac}(B)[X]$ denote the polynomial algebra over $\operatorname{Frac}(B)$ and let $F(X) \in \operatorname{Frac}(B)[X]$ be the monic minimal polynomial of an element $\alpha \in C[\frac{1}{y}] \cap C[\frac{1}{x}]$. Then $\alpha$ is integral over $B[\frac{1}{y}]$ and $B[\frac{1}{x}]$ which are both normal domains by normality of $B$. It follows from \cite[Theorem 2.1.17]{SwHu} that
$$
F(X) \in B[\tfrac{1}{y}][X]~\mbox{and}~F(X) \in B[\tfrac{1}{x}][X].
$$
Thus, it follows that $F(X) \in B[X]$ and $B \to C[\frac{1}{y}] \cap C[\frac{1}{x}]$ is an integral extension of integral domains. Since $C[\frac{1}{y}]$ and $C[\frac{1}{x}]$ are normal domains, $C[\frac{1}{y}] \cap C[\frac{1}{x}]$ is also normal. This implies that $C=C[\frac{1}{y}] \cap C[\frac{1}{x}]$ by normality of $C$. Finally,
$$
\tfrac{b}{x}\in C[\tfrac{1}{y}] \cap C[\tfrac{1}{x}]=C,
$$
which contradicts to our hypothesis. Hence we must have $t \in (y) \subseteq C$.\\
\end{proof}
\subsection{Local Bertini theorem}
Let $B \hookrightarrow C$ be a module-finite extension of Noetherian domains. Then it is known that the support of the $C$-module $\Omega_{C/B}$ is the largest closed subset of $\operatorname{Spec} C$ on which $\operatorname{Spec} C \to \operatorname{Spec} B$ is ramified \cite[Corollary 6.10]{Kunz}. We prove the following theorem as a main corollary of the preliminary lemmas. A ring map $B \to C$ is \textit{etale}, if it is flat, unramified and of finite presentation. Let $\Omega_{C/B}$ be the module of K\"ahler differentials and let $\operatorname{Ass}_B(\Omega_{C/B})$ be the set of associated primes of $\Omega_{C/B}$ as a $B$-module.
\begin{theorem}[Local Bertini Theorem]
\label{MainBertini}
Assume that $B \hookrightarrow C$ is a module-finite extension such that $B$ is a Noetherian unique factorization domain, $C$ is a normal domain and $\operatorname{Frac}(B) \to \operatorname{Frac}(C)$ is a separable field extension. Let us define a subset $\mathcal{S}_{C/B}$ of $\operatorname{Spec} B$ as follows:
\begin{equation}
\label{equation:scb}
\mathcal{S}_{C/B}=\{\frak{p} \in \operatorname{Spec} B~|~\frak{p}~\mbox{is a height-one prime in}~\operatorname{Ass}_B(\Omega_{C/B})\}.
\end{equation}
Then $\mathcal{S}_{C/B}$ is a finite set and $C/\frak{p} C$ is a reduced ring for any height-one prime $\frak{p} \in \operatorname{Spec} B$ such that $\frak{p} \notin \mathcal{S}_{C/B}$.
\end{theorem}
\begin{proof}
Since $\operatorname{Frac}(B) \to \operatorname{Frac}(C)$ is separable by assumption, we have
$$
\Omega_{C/B} \otimes_C \operatorname{Frac}(C) \simeq \Omega_{\operatorname{Frac}(C)/\operatorname{Frac}(B)}=0,
$$
which implies that $\Omega_{C/B}$ is a finitely generated torsion $C$-module. Since $C$ is a finitely generated $B$-module, $\Omega_{C/B}$ is a finitely generated torsion $B$-module. Let $\{ \frak{p}_1,\ldots,\frak{p}_k \}$ be the finite set of all associated primes of the finitely generated torsion $B$-module $\Omega_{C/B}$ which are of height-one. Then, for a height-one prime $\frak{p} \in \operatorname{Spec} B$, we have $(\Omega_{C/B})_{\frak{p}}=\Omega_{C_{\frak{p}}/B_{\frak{p}}} \ne 0$
if and only if $\frak{p} \in \{\frak{p}_1,\ldots,\frak{p}_k\}$, where $C_{\frak{p}}=C \otimes_B B_{\frak{p}}$. In other words, if $\frak{p} \notin \{\frak{p}_1,\ldots,\frak{p}_k\}$, the localization $B_{\frak{p}} \to C_{\frak{p}}$
is an etale extension of semi-local Dedekind domains.
Note that every height-one prime of $B$ is principal because $B$ is a unique factorization domain. Now take an arbitrary height-one prime $\frak{p} \notin \{\frak{p}_1,\ldots,\frak{p}_k\}$ of $B$ and
we will apply Lemma \ref{Koszul2} with $(y) =\mathfrak{p}$.
Since $B \hookrightarrow C$ is a module-finite extension, $C$ is integral over $B$, which is nothing but the condition (i) of Lemma \ref{Koszul2}. The condition (ii) of Lemma \ref{Koszul2} was already verified above. We thus conclude that $C/\frak{p} C$ is a reduced ring by Lemma \ref{Koszul2}.\\
\end{proof}
\subsection{Cohen structure theorem}
We discuss applications of Theorem \ref{MainBertini} to Betini-type theorems on complete local normal domains, which we will apply to study the characteristic ideals. Now let $(R,\frak{m},\mathbf{k})$ be a Noetherian complete local domain of mixed characteristic, where $\frak{m}$ is its unique maximal ideal and $\mathbf{k}$ is its residue field of characteristic $p>0$. Let $(A,pA,\mathbf{k})$ be a complete discrete valuation ring of mixed characteristic $p>0$. We fix a \textit{coefficient ring} of $R$:
$$
(A,pA,\mathbf{k}) \xrightarrow{\pi} R.
$$
That is, $\pi$ is a flat local map from $A$ to $R$ inducing an isomorphism on residue fields. Let $I \subseteq \frak{m}$ be an ideal of a Noetherian local ring $(R,\frak{m},\mathbf{k})$. We say that $I$ is a \textit{parameter ideal}, if it is generated by a system of parameters of $R$. Let us recall that Cohen structure theorem is essential in the construction of a module-finite extension of a complete local ring over a complete regular local ring via a certain choice of a system of parameters \cite{Cohen}.
\begin{theorem}[Cohen structure theorem]
\label{Cohen}
A Noetherian complete local domain $(R,\frak{m},\mathbf{k})$ with mixed characteristic $p>0$ and dimension $d$ has a coefficient ring $(A,pA,\mathbf{k})$ and there exists a regular local subring $A[[x_2,\ldots,x_d]]$ of $R$ such that $A[[x_2,\ldots,x_d]] \hookrightarrow R$ is module-finite.
\end{theorem}
The difficult part of Cohen structure theorem is to prove that a coefficient ring exists. Once a coefficient ring is determined, we use the following lemma to find a module-finite extension from a complete regular local ring.
\begin{lemma}
\label{CohenStructure}
Let the notation be as in Theorem \ref{Cohen}, let $p,z_2,\ldots,z_d$ be any system of parameters of $R$ and let $A[[z_2,\ldots,z_d]]$ be the completion of the subring $A[z_2,\ldots,z_d]$ of $R$ with respect to the $(p,z_2,\ldots,z_d)$-adic topology. Then there is a module-finite extension $A[[z_2,\ldots,z_d]] \to R$. Moreover, $A[[z_2,\ldots,z_d]]$ is a $d$-dimensional regular local ring.
\end{lemma}
\begin{proof}
Taking the $(p,z_2,\ldots,z_d)$-adic completion, $A[z_2,\ldots,z_d] \hookrightarrow R$ extends to a ring map:
$$
A[[z_2,\ldots,z_d]] \to R.
$$
Let us prove that $R$ is a finitely generated $A[[z_2,\ldots,z_d]]$-module. Since $A[[z_2,\ldots,z_d]]$ and $R$ have the same residue field, $R/(p,z_2,\ldots,z_d)$ is an $A[[z_2,\ldots,z_d]]$-module of finite length. Applying the topological Nakayama's theorem \cite[Theorem 8.4]{Mat}, it follows that $A[[z_2,\ldots,z_d]] \to R$ is module-finite. Since the Krull dimension of $A[[z_2,\ldots,z_d]]$ is $d$ and its maximal ideal is generated by $d$ elements, it is a regular local ring.\\
\end{proof}
\subsection{An application of the local Bertini theorem}
We will need the following corollary of Theorem \ref{MainBertini} in \S~\ref{controltheorem}. Let us fix a prime number $p>0$.
\begin{corollary}
\label{LocalBertini}
Let $(R,\frak{m},\mathbf{k})$ be a two-dimensional Noetherian complete local normal domain with mixed characteristic $p>0$, where $A$ is a coefficient ring of $R$. Then the following statements hold.
\begin{enumerate}
\item
Let us choose an element $z \in \frak{m}$ such that $(p,z)$ is a parameter ideal of $R$. Then, for each $n>0$, $R/(z+a_np^n)$ is a reduced ring with mixed characteristic for all but finitely many $a_n \in A^{\times}$.
\item
Let us choose an element $z, r \in \frak{m}$ such that $(p,z^n+r)$ is a parameter ideal of $R$ for all $n>0$.
Then, for each $n>0$, $R/(z^n+r+a_n p)$ is a reduced ring with mixed characteristic for all but finitely many $a_n \in A^{\times}$.
\end{enumerate}
\end{corollary}
By a theorem of Auslander and Buchsbaum, it is known that a regular local ring is a unique factorization domain. Now we apply Theorem \ref{MainBertini} by setting $B=A[[z]]$ or $B=A[[z^n+r]]$ to prove the corollary.
\begin{proof}
The proofs of the assertions $\rm (1)$ and $\rm (2)$ proceed almost in the same way. The situation
of $\rm (2)$ is slightly more complicated since the ring $B=A[[z^n+r]]$ is different for varying $n$. Thus, we will prove only the assertion $\rm (2)$ omitting the proof of the assertion $\rm (1)$.
Let us start the proof of the assertion $\rm (2)$. Since $(p,z^n+r)$ is a parameter ideal, $A[[z^n+r]] \hookrightarrow R$ is module-finite. Note that this map depends on the choice of $n>0$. For each $n>0$, we denote by $\mathcal{S}_n$ the finite subset $\mathcal{S}_{R/ A[[z^n+r]]}$ of $\operatorname{Spec} A[[z^n+r]]$
defined in \eqref{equation:scb}.
Let $\mathcal{T}_n \subseteq \operatorname{Spec} R$ be a finite set defined to be the inverse image of $\mathcal{S}_n$ via the map
$\operatorname{Spec} R \rightarrow \operatorname{Spec} A[[z^n+r]]$.
When we have two distinct elements $a_n , a'_n \in A^{\times}$,
$(z^n+r+a_np)$ and $(z^n+r+a'_np)$ are pairwise distinct primes of $A[[z^n+r]]$.
Hence we have $
(z^n+r+a_np) \notin \mathcal{T}_n $ for all but finitely many $a_n \in A^{\times}$.
If we have $(z^n+r+a_np) \notin \mathcal{T}_n $, then $R/(z^n+r+a_np)$ is a reduced ring of mixed characteristic. Thus, we complete the proof of $\rm(2)$.\\
\end{proof}
\begin{remark}
\label{heightoneprime}
An inspection of the proof of Lemma \ref{LocalBertini} reveals that a fixed height-one prime ideal of $R$ cannot contain infinitely many distinct principal ideals from $\{(z+a_np^n)\}_{n \in \mathbb{N}}$ (resp. $\{(z^n+r+a_np)\}_{n \in \mathbb{N}}$), as long as $z \in \frak{m}$ is chosen to be nonzero. This fact will be used in the proof of Theorem \ref{veryfinal}.
\end{remark}
\subsection{An application of Cohen-Gabber theorem}
Let us recall that Cohen-Gabber theorem is a refinement of Cohen structure theorem in the equal characteristic $p>0$ case. We prove the following corollary as a straightforward consequence of Theorem \ref{MainBertini} for independent interest.
\begin{corollary}
\label{LocalBertini2}
Suppose that $(R,\frak{m},\mathbf{k})$ is a Noetherian complete local normal domain of dimension $d \ge 2$ and of equal characteristic. Let $\frak{q}_1,\ldots,\frak{q}_r$ be pairwise distinct height-one primes. Then there exists an infinite set of nonzero principal ideals $\{(\mathbf{x}_n)\}_{n \in \mathbb{N}}$ such that $R/(\mathbf{x}_n)$ are reduced rings and every $\mathbf{x}_n$ is not contained in $\frak{q}_1 \cup \cdots \cup \frak{q}_r$.
\end{corollary}
\begin{proof}
By assumption, $R$ contains the field $\mathbf{k}$.
When $\mathbf{k}$ is of characteristic $p>0$, we need to use Cohen-Gabber theorem. By \cite[Th\'eor\`em 2.1.1]{Book} (see \cite{KS} for an elementary proof) , there exist a regular local ring $\mathbf{k}[[t_1,\ldots,t_d]]$ and a module-finite extension: $\mathbf{k}[[t_1,\ldots,t_d]] \to R$ such that
$$
\operatorname{Frac}(\mathbf{k}[[t_1,\ldots,t_d]]) \to \operatorname{Frac}(R)
$$
is a separable extension, which is automatic in the characteristic zero case. Let us put
$$
\frak{p}_i:=\mathbf{k}[[t_1,\ldots,t_d]] \cap \frak{q}_i.
$$
Then $\frak{p}_i$ is a height-one prime ideal. We denote by $\mathcal{S}$ the set of height-one primes of $\mathbf{k}[[t_1,\ldots,t_d]]$ attached by Theorem \ref{MainBertini} applied to the map $B:=\mathbf{k}[[t_1,\ldots,t_d]] \to C:=R$. Now we can find pairwise distinct elements $\mathbf{x}_1,\mathbf{x}_2,\ldots \in \mathbf{k}[[t_1,\ldots,t_d]]$ inductively so that $(\mathbf{x}_n) \notin \mathcal{S} \cup \{\frak{p}_1,\ldots,\frak{p}_r\}$ for $n>0$. Hence $R/(\mathbf{x}_n)$ is a reduced ring.
\end{proof}
\begin{remark}
\label{Bertinifail}
\begin{enumerate}
\item
Theorem \ref{MainBertini} fails without the assumption that $\operatorname{Frac}(B) \to \operatorname{Frac}(C)$ is separable. For example, let $C=\mathbf{k}[x,y]$ for a perfect field $\mathbf{k}$ of characteristic $p>0$. Let $B=\mathbf{k}[x^p,y^p]$. Then $B \to C$ is a purely inseparable extension. Set $\mathbf{x}=ax^p+by^p$ with $a,b \in \mathbf{k}^{\times}$. Then $C/(\mathbf{x})$ is not reduced for any choice $a,b \in \mathbf{k}^{\times}$.
\item
Corollary \ref{LocalBertini} can also be proved by using the basic elements and the derivation of completed K\"ahler differential modules, which are the main tools for the proof of the Bertini theorem for normality on local rings in the article \cite{OcSh}.
\item
In \S~\ref{controltheorem}, we will make a special choice of $r \in R$ in Corollary \ref{LocalBertini}, which is necessary to take care of height-one primes lying above $p$ in a normal domain.
\end{enumerate}
\end{remark}
Finally let us prove a lemma on when an ind-etale extension of a coefficient ring $A$ of a complete local ring $R$ preserves many good properties of $R$.
\begin{lemma}
\label{localring}
Let $(R,\frak{m},\mathbf{k})$ be a Noetherian complete local ring with mixed characteristic and residue field $\mathbf{k}$ and let $(A,pA,\mathbf{k}) \to R$ be a coefficient ring map. Assume that $(A,pA,\mathbf{k}) \to (B,pB,\mathbf{k}')$ is an integral flat extension of discrete valuation rings. We put $R_B:=R \widehat{\otimes}_A B$, where $\widehat{\otimes}$ is the complete tensor product. Then we have the following statements.
\begin{enumerate}
\item
$R_B$ is a Noetherian complete local ring and $R \otimes_A B$ is a quasilocal ring.
\item
If we assume further that $R$ is a normal domain and $A \to B$ is an ind-etale extension, then $R_B$ is a normal domain.
\end{enumerate}
\end{lemma}
\begin{proof}
Let us prove the assertion $\rm (1)$. There is a surjetion $A[[t_1,\ldots,t_n]] \twoheadrightarrow R$ by Cohen structure theorem. Then we have a surjection
$B[[t_1,\ldots,t_n]] \twoheadrightarrow R_B$, which shows that $R_B$ is a Noetherian complete local ring. We prove that $R \otimes_A B$ is local. Note that $p$ is in the Jacobson radical of $R \otimes_A B$ and so it suffices to see that $(R \otimes_A B)/(p)=R/(p) \otimes_{\mathbf{k}} \mathbf{k}'$ is a local ring. Since $R/(p)$ is a local ring with the residue field $\mathbf{k}$ and $\mathbf{k} \to \mathbf{k}'$ is an algebraic extension, it is not hard to see
that $R/(p) \otimes_{\mathbf{k}} \mathbf{k}'$ is a local ring (see \cite[Theorem]{Swe} for example). This completes the proof of $\rm (1)$.
For the proof of the assertion $\rm (2)$, notice that $R \to R \otimes_A B$ is an ind-etale local extension and $R \otimes_A B$ is quasilocal by the assumption that $A \to B$ is ind-etale and $\rm(1)$. In particular, we have $R \to R \otimes_A B \to R^{\rm{sh}}$, where $R^{\rm{sh}}$ is the strict henselization of $R$. Since it is known that $R^{\rm{sh}}$ is Noetherian and $R \otimes_A B \to R^{\rm{sh}}$ is faithfully flat, $R \otimes_A B$ is also Noetherian. Since $R \to R \otimes_A B$ is ind-etale and $R$ is a complete local normal domain, we find that $R\otimes_A B$ is normal and excellent by \cite[Theorem 5.3 (iv)]{Gre}. Then the $\frak{m}$-adic completion of $R \otimes_A B$, which is $R_B$, is a normal domain.\\
\end{proof}
\section{Characteristic ideals and specialization methods}
\label{controltheorem}
In this section, we develop the specialization method for characteristic ideals over Noetherian local normal domains in dimension two. In classical Iwasawa theory, one considers the characteristic ideal attached to the torsion module arising as the projective limit of the $p$-part of the ideal class groups of a certain tower of number fields, which gives rise to the completed group algebra $\mathbb{Z}_p[[1+p\mathbb{Z}_p]]$. In Hida theory, one considers the projective limit of the space of $p$-adic ordinary modular forms and the $p$-adic ordinary Hecke algebras. Thus, the resulting projective limit would be more complicated than $\mathbb{Z}_p[[1+p\mathbb{Z}_p]]$. This motivates us to consider the structure of characteristic ideals attached to finitely generated torsion modules over normal domains. We first make the definition of characteristic ideals. Let $R$ be a Noetherian normal domain and let
$$
N^{\operatorname{rc}}:=\operatorname{Hom}_R(\operatorname{Hom}_R(N,R),R)
$$
denote the \textit{reflexive closure} of an $R$-module $N$ and let $\ell_R(M)$ denote the length of an Artinian $R$-module $M$.
\begin{definition}
\label{charideal}
Let $R$ be a Noetherian normal domain.
\begin{enumerate}
\item
Let $M$ be a finitely generated torsion module over $R$. The \textit{characteristic ideal} associated to $M$ is defined as the reflexive ideal of $R$ as follows:
$$
\operatorname{char}_R(M):=\big(\prod_{\frak{p}} \frak{p}^{\ell_{R_{\frak{p}}}(M_{\frak{p}})}\big)^{\operatorname{rc}},
$$
where $\frak{p}$ ranges over all height-one prime ideals of $R$. If $M$ is not a torsion $R$-module, we set $\operatorname{char}_R(M)=0$.
\item
A finitely generated $R$-module $M$ is said to be \textit{pseudo-null}, if $M_{\frak{p}}=0$ for all height-one primes $\frak{p}$ of $R$.
\item
Let $M$ and $N$ be finitely generated $R$-modules. Then an $R$-module map $f:M \to N$ is said to be a \textit{pseudo-isomorphism} if both kernel and cokernel of $f$ are pseudo-null $R$-modules.
\end{enumerate}
\end{definition}
More details on characteristic ideals are found in \cite[\S~7]{OcSh}. If $R$ is a unique factorization domain (UFD), the characteristic ideal is principal and it is not necessary to take the reflexive closure in Definition \ref{charideal}. The following proposition is a collection of the basic properties \cite[Proposition 9.6]{OcSh}.
\begin{proposition}
The following statements hold.
\begin{enumerate}
\item
Let $R$ be a finitely generated torsion $R$-module. Then we have
$$
\operatorname{char}_R(M)=\big(\prod_{\operatorname{ht} \frak{p}=1} \frak{p}^{\ell_{R_{\frak{p}}}(M_{\frak{p}})}\big)^{\operatorname{rc}}=\operatorname{Fitt}_R(M)^{\operatorname{rc}},
$$
where $\operatorname{Fitt}_R(M)$ is the Fitting ideal of $M$. In particular, $\operatorname{Fitt}_R(M) \subseteq \operatorname{char}_R(M)$. If $R$ is a UFD, then
$$
\operatorname{Fitt}_R(M) \subseteq \prod_{\operatorname{ht} \frak{p}=1} \frak{p}^{\ell_{R_{\frak{p}}}(M_{\frak{p}})}=\operatorname{char}_R(M).
$$
\item
Let $0 \to L \to M \to N \to 0$ be a short exact sequence of finitely generated $R$-modules. Then
$$
\operatorname{char}_R(M)=\big(\operatorname{char}_R(L) \operatorname{char}_R(N)\big)^{\operatorname{rc}}.
$$
\item
Let $M$ and $N$ be finitely generated torsion $R$-modules. Then $\operatorname{char}_R(M) \subseteq \operatorname{char}_R(N)$ if and only if $\operatorname{char}_R(M)_{\frak{p}} \subseteq \operatorname{char}_R(N)_{\frak{p}}$ for all height-one primes $\frak{p} \in \operatorname{Spec} R$.
\end{enumerate}
\end{proposition}
In what follows, we set
$$
M[x]:=\{m \in M~|~xm=0\}
$$
for a module $M$ over a ring $R$ and $x \in R$. This is an $R$-submodule of $M$.
\begin{lemma}
\label{lemma2}
Let $0 \to L \to M \to N \to 0$ be a short exact sequence of modules over a ring $R$. Then for $x \in R$, there is a short
exact sequence:
$$
0 \to L[x] \to M[x] \to N[x] \to L/xL \to M/xM \to N/xN \to 0.
$$
\end{lemma}
\begin{proof}
The lemma follows immediately by applying the snake lemma to the following diagram:
$$
\begin{CD}
0 @>>> L @>>> M @>>> N @>>> 0\\
@. @V\times xVV @V\times xVV @V\times xVV \\
0 @>>> L @>>> M @>>> N @>>> 0.\\
\end{CD}
$$
\end{proof}
For a finitely generated module $M$ over a domain $R$, the \textit{rank} of $M$ is defined as the $\operatorname{Frac}(R)$-dimension of the vector space $M \otimes_R \operatorname{Frac}(R)$. We have the following lemma:
\begin{lemma}
\label{lemma3}
Assume that $(R,\frak{m})$ is a one-dimensional local Noetherian domain and $M$ is a finitely generated $R$-module. Then, for a nonzero element $x \in \frak{m}$, we have
$$
\ell_R(M/xM)-\ell_R(M[x])=\operatorname{rank}_R(M) \cdot \ell_R(R/(x)).
$$
\end{lemma}
\begin{proof}
Since the only prime ideals of $R$ are $0$ and $\frak{m}$, there exists a filtration $0=M_0 \subseteq M_1 \subseteq \cdots \subseteq M_n=M$ such that each quotient $M_i/M_{i+1}$ is isomorphic to $R$ or $R/\frak{m}$ by \cite[Theorem 6.4]{Mat}. By Lemma \ref{lemma2}, the left-hand side of the displayed equation is additive with respect to short exact sequences.
Since $\ell_R(R/(x))$ is a constant which has nothing to do with $M$,
the right-hand side of the displayed equation is additive with respect to short exact sequences. Hence, it suffices to consider the case that $M=R$ or $M=R/\frak{m}$. If $M=R$, then both sides are equal to $\ell_R(R/(x))$. If $M=R/\frak{m}$, then both sides are obviously zero, as desired.
\end{proof}
\begin{lemma}
\label{zerodivisor}
If $x,y$ are nonzero divisors of a commutative ring $R$, then there is a short exact sequence:
$$
0 \to R/(x) \to R/(xy) \to R/(y) \to 0.
$$
\end{lemma}
\begin{proof}
This is an easy exercise.
\end{proof}
\begin{definition}
Let $R[[X]]$ be a power series ring over a Noetherian local ring $(R,\frak{m},\mathbf{k})$. Then we say that $f \in R[[X]]$ is a \textit{distinguished polynomial}, if it is of the form:
$$
f =X^n+a_{n-1}X^{n-1}+\cdots+a_1X+a_0
$$
for some integer $n \ge 0$ and $a_0,\ldots,a_{n-1} \in \frak{m}$.
\end{definition}
The following theorem is a classical result of much importance (see \cite[Chap. 7, \S 3.8]{Bourbaki} or \cite[Theorem 5.3.4]{NSW} and see also \cite{Ger} for a short proof)
\begin{theorem}[Weierstrass Preparation Theorem]
\label{Weierstrass}
Let $(R,\frak{m},\mathbf{k})$ be a complete Noetherian local ring and let
$f=\sum_{i=0}^{\infty}a_i X^i \in R[[X]]$ be a nonzero element. Assume that there exists $n \in \mathbb{N}$ such that $a_i \in \frak{m}$ for all $i < n$ and $a_n \notin \frak{m}$. Then there exist a unit $u$ in $R[[X]]$ and a distinguished polynomial $f_0$ of degree $n$ such that $f=uf_0$. Furthermore, $u$ and $f_0$ are uniquely determined by $f$.\\
\end{theorem}
\subsection{Specialization method in dimension two for height-one primes outside $p$}
Let $(R,\frak{m},\mathbb{F})$ be a two-dimensional Noetherian complete local normal domain with mixed characteristic $p>0$ and finite residue field $\mathbb{F}$. Let $W(\mathbb{F})$ be the ring of Witt vectors of a finite field $\mathbb{F}$. Since $\mathbb{F}$ is a perfect field, there is a unique ring map $W(\mathbb{F}) \hookrightarrow R$ which induces an identity map on the residue field $\mathbb{F}$. Note that for any local ring $A$ with finite residue field $\mathbb{F}$ and a finitely generated Artinian $A$-module $M$, we have
\begin{equation}
\label{cardinality}
\big|M\big|=\big|\mathbb{F}\big|^{\ell_A(M)},
\end{equation}
as verified easily by the filtration argument. Let us choose a parameter ideal $(p,z)$ of $R$ and define the set of principal ideals:
\begin{equation}
\label{Fittideal}
\mathscr{L}_{R,W(\mathbb{F})}(z,\{a_n\}_{n \in \mathbb{N}}):=\{(\mathbf{x}_n)\}_{n \in \mathbb{N}},
\end{equation}
where $\mathbf{x}_n=z+a_n p^n$ is attached by Corollary \ref{LocalBertini} $\rm (1)$ and $\{a_n\}_{n \in \mathbb{N}}$ is any sequence of elements in $W(\mathbb{F})^{\times}$. Suppose that $M$ and $N$ are finitely generated torsion $R$-modules. Let $\{\frak{q}_i\}_{1 \le i \le k}$ be the set of all height-one primes outside $(p)$ and that appear as components in the characteristic ideal of either $M$ or $N$. Consider an additional condition:
\begin{equation}
\label{linearideal}
z_t \in \frak{q}_t~\mbox{and}~z_t \notin \bigcup_{i \ne t} \frak{q}_i~\mbox{for each}~1 \le t \le k.~\mbox{If}~k=1,~\mbox{then choose}~z \in \frak{q}_1\setminus \{0\}.
\end{equation}
We notice that the existence of $z_t \in R$ satisfying $(\ref{linearideal})$ is ensured by Prime Avoidance Lemma \cite[Theorem A.1.1 at page 392]{SwHu}.
\begin{theorem}[Specialization Method outside $p$]\label{prop1}
Let $(R,\frak{m},\mathbb{F})$ be a two-dimensional Noetherian complete local normal domain with mixed characteristic $p>0$ and finite residue field $\mathbb{F}$. Suppose that $M$ and $N$ are finitely generated torsion $R$-modules and fix a finite set of elements $\{z_i\}_{1 \le i \le k}$ satisfying $(\ref{linearideal})$. Then the following statements are equivalent:
\begin{enumerate}
\item[(a)]
Let $\frak{q}$ be any height-one prime of $R$ which does not lie over $p$. Then we have
$$
\operatorname{char}_R(N)_{\frak{q}} \subseteq \operatorname{char}_R(M)_{\frak{q}}.
$$
\item[(b)]
There exists a constant $c \in \mathbb{N}$, depending on $M$ and $N$, such that
$$
c \cdot \frac{|N/\mathbf{x}N |}{|M/\mathbf{x}M |} \in \mathbb{N}
$$
for all but finitely many principal ideals
$$
(\mathbf{x}) \in \bigcup_{1\le i \le k} \mathscr{L}_{R,W(\mathbb{F})}(z_i,\{a_n\}_{n \in \mathbb{N}}).
$$
\end{enumerate}
\end{theorem}
\begin{proof}
Take the fundamental pseudo-isomorphisms
\begin{equation}
\label{equation : fundamental_isomorphism}
f_M:M \to \bigoplus_i R/\frak{p}_i^{e_i} \oplus \bigoplus_r R/\frak{q}_r^{f_r}~\mbox{and}~f_N:N \to \bigoplus_j R/\frak{p}_j^{e'_j}\oplus \bigoplus_s R/\frak{q}_s^{f'_s},
\end{equation}
where $\{\frak{p}_i\}_{1 \le i \le h}$ is the set of all height-one primes of $R$ containing $p$, and $\{\frak{q}_i\}_{1 \le i \le k}$ is the set of height-one primes not containing $p$. We call the target of \eqref{equation : fundamental_isomorphism} a \textit{fundamental module} associated to $M$ (resp. $N$) and denote it by $F(M)$ (resp. $F(N)$). Then kernel and cokernel of $f_M$ (resp. $f_N$) are finite, since $R$ is of Krull dimension two and $f_M$ (resp. $f_N$) is a pseudo-isomorphism.
By the same argument as \cite[Proposition 3.11]{Oc1}, the proof of $\rm (a) \iff (b)$ for a pair of modules $(M,N)$ is reduced to the proof of $\rm(a) \iff (b)$ for a pair of modules $(F(M),F(N))$. We are thus reduced to the case where $M$ and $N$ are fundamental modules. That is, we may assume that
$$
M=R/J_M \oplus R/I_M ~\mbox{and}~N=R/J_N \oplus R/I_N,
$$
where $J_M=\prod_i \frak{p}_i^{e_i}$, $I_M=\prod_r \frak{q}_r^{f_r}$ and $J_N=\prod_j \frak{p}_j^{e'_j}$, $I_N=\prod_s \frak{q}_s^{f'_s}$. Note that $\operatorname{Ass}_R(M)$ and $\operatorname{Ass}_R(N)$ are the subsets of $\{\frak{p}_1,\ldots,\frak{p}_h,\frak{q}_1,\ldots,\frak{q}_k\}$.
First we prove $\rm (a) \Rightarrow (b)$. Then we find that
$$
|M/\mathbf{x}M|=|R/J_M+(\mathbf{x})| \cdot |R/I_M+(\mathbf{x})|
~\mbox{and}~|N/\mathbf{x}N|=|R/J_N+(\mathbf{x})| \cdot |R/I_N+(\mathbf{x})|.
$$
For those elements $\mathbf{x} \in R$ of specified type, it follows that $|R/J_M+(\mathbf{x})|$ is finite and bounded. So the implication $\rm(a) \Rightarrow (b)$ follows by putting
$$
c=\max_{\mathbf{x}}\{|R/J_M+(\mathbf{x})|\}.
$$
Next we prove $\rm (b) \Rightarrow (a)$. Let us prove the equations:
\begin{equation}
\label{linear1}
\ell_R(M/\mathbf{x}M)=\sum_{i=1}^h \ell_{R_{\frak{p}_i}}(M_{\frak{p}_i}) \cdot \ell_R(R/\frak{p}_i+(\mathbf{x}))+\sum_{i=1}^k \ell_{R_{\frak{q}_i}}(M_{\frak{q}_i}) \cdot \ell_R(R/\frak{q}_i+(\mathbf{x}))
\end{equation}
and
\begin{equation}
\label{linear2}
\ell_R(N/\mathbf{x}N)=\sum_{i=1}^h \ell_{R_{\frak{p}_i}}(N_{\frak{p}_i}) \cdot \ell_R(R/\frak{p}_i+(\mathbf{x}))+\sum_{i=1}^k \ell_{R_{\frak{q}_i}}(N_{\frak{q}_i}) \cdot \ell_R(R/\frak{q}_i+(\mathbf{x})),
\end{equation}
under the assumption $\mathbf{x} \notin \bigcup_{i=1}^h \frak{p}_i \cup \bigcup_{i=1}^k \frak{q}_i$.
As the proofs of $(\ref{linear1})$ and $(\ref{linear2})$ are exactly the same, we will prove only $(\ref{linear1})$. As we will show below, the equation $(\ref{linear1})$ is reduced to the following equation.
\begin{multline}
\label{equation:specialization11}
\ell_R(M/\mathbf{x}M)-\ell_R(M[\mathbf{x}])
\\
=\sum_{i=1}^h \ell_{R_{\frak{p}_i}}(M_{\frak{p}_i}) \cdot \ell_R(R/\frak{p}_i+(\mathbf{x}))+\sum_{i=1}^k \ell_{R_{\frak{q}_i}}(M_{\frak{q}_i}) \cdot \ell_R(R/\frak{q}_i+(\mathbf{x})).
\end{multline}
In fact, since $M[\mathbf{x}]=0$ by the assumption that $M$ is a fundamental module, $(\ref{linear1})$ follows immediately from \eqref{equation:specialization11}. To prove \eqref{equation:specialization11}, its left-hand side is additive with respect to short exact sequences of $R$-modules by Lemma \ref{lemma2}. Its right-hand side is also additive. Thus, we reduce the proof to the case where $M=R/P$ for $P \in \{\frak{p}_1,\ldots,\frak{p}_h,\frak{q}_1,\ldots,\frak{q}_k\}$
by taking a prime filtration of $M$. In this case, the left-hand side of the equation is $\ell_R(R/P+(\mathbf{x}))$, while the right-hand side is $\ell_{R_P}(R_P/PR_P) \cdot \ell_R(R/P+(\mathbf{x}))$. They are equal to each other. To prove $\rm (b) \Rightarrow (a)$, assume that
\begin{equation}
\label{linear5}
\ell_{R_{\frak{q}_j}}(M_{\frak{q}_j})>\ell_{R_{\frak{q}_j}}(N_{\frak{q}_j})
\end{equation}
for some $1 \le j \le k$. Thus, we may put $j=1$ for simplicity. Fix an element $z:=z_1 \in R$ such that
\begin{equation}
\label{linear55}
z \in \frak{q}_1~\mbox{and}~z \notin \bigcup_{i \ne 1}\frak{q}_i.
\end{equation}
By definition, $(p, z)$ is a parameter ideal of $R$ and $\Lambda_{W(\mathbb{F})}:=W(\mathbb{F})[[z]] \to R$ is module-finite. Since $\Lambda_{W(\mathbb{F})}$ is a UFD, we have $(g_i(z))=\Lambda_{W(\mathbb{F})} \cap \frak{q}_i$, where $g_i(z)$ is an irreducible monic polynomial by Theorem \ref{Weierstrass}. From $(\ref{linear55})$, we find that
\begin{equation}
\label{linear56}
(z)=(g_1(z))~\mbox{and}~(z) \ne (g_i(z))~\mbox{for any}~i \ne 1.
\end{equation}
Applying Lemma \ref{lemma3} to the extension: $\Lambda_{W(\mathbb{F})}/(g_i(z)) \to R/\frak{q}_i$, we obtain
\begin{equation}
\label{linear3}
\ell_R(R/\frak{q}_i+(\mathbf{x}))=\operatorname{rank}_{\Lambda_{W(\mathbb{F})}/(g_i(z))}(R/\frak{q}_i) \cdot
\ell_{\Lambda_{W(\mathbb{F})}}(\Lambda_{W(\mathbb{F})}/(g_i(z),\mathbf{x})),
\end{equation}
under the assumption $\mathbf{x} \notin (g_i(z))$ in $\Lambda_{W(\mathbb{F})}$. Let us put
$$
g_M(z):=\prod_{i=1}^kg_i(z)^{m_i},~\mbox{where}~m_i:={\operatorname{rank}_{\Lambda_{W(\mathbb{F})}/(g_i(z))}(R/\frak{q}_i) \cdot \ell_{R_{\frak{q}_i}}(M_{\frak{q}_i})}
$$
and
$$
g_N(z):=\prod_{i=1}^kg_i(z)^{n_i},~\mbox{where}~n_i:={\operatorname{rank}_{\Lambda_{W(\mathbb{F})}/
(g_i(z))}(R/\frak{q}_i) \cdot \ell_{R_{\frak{q}_i}}(N_{\frak{q}_i})}).
$$
It follows that $g_N(z)$ is not divisible by $g_M(z)$ in view of $(\ref{linear5})$ and $(\ref{linear56})$. To be more precise, we have the following:
\begin{equation}
\label{linear557}
g_1(z)^{n_1}~\mbox{is not divisible by}~g_1(z)^{m_1}.
\end{equation}
Set $\mathbf{x}_n:=z+a_np^n \in \Lambda_{W(\mathbb{F})}$ for $n>0$. Then by $(\ref{linear56})$ and $(\ref{linear557})$, a calculation over $\Lambda_{W(\mathbb{F})}$ (see \cite[Proposition 3.11 (i)]{Oc1} for the same calculation) yields that
\begin{equation}
\label{inequality1}
\text{$\frac{\vert \Lambda_{W(\mathbb{F})}/(g_N(z),\mathbf{x}_n)\vert}
{\vert \Lambda_{W(\mathbb{F})}/(g_M(z),\mathbf{x}_n) \vert}$ tends to $0$ as $n$ goes to
$\infty$.}
\end{equation}
On the other hand, since $p \in \frak{p}_i$, we have the following:
\begin{equation}
\label{linear57}
\limsup_{n \to \infty} \ell_R(R/\frak{p}_i+(\mathbf{x}_n)) < \infty~\mbox{for}~i=1,\ldots,h.
\end{equation}
By $(\ref{cardinality})$, $(\ref{linear1})$, $(\ref{linear2})$, $(\ref{linear3})$ and using Lemma \ref{zerodivisor} repeatedly, we get
\begin{align}
\label{absolutevalue1}
|M/\mathbf{x}_n M|&=|\mathbb{F}|^{\ell_R (M/\mathbf{x}_n M)}
\\
&=
|\mathbb{F}|^{\sum_{i=1}^h \ell_{R_{\frak{p}_i}}(M_{\frak{p}_i}) \cdot \ell_R(R/\frak{p}_i+(\mathbf{x}_n))} \cdot |\mathbb{F}|^{\sum_{i=1}^k \ell_{R_{\frak{q}_i}}(M_{\frak{q}_i}) \cdot \ell_R(R/\frak{q}_i+(\mathbf{x}_n))} \nonumber
\\
&=|\mathbb{F}|^{\sum_{i=1}^h \ell_{R_{\frak{p}_i}}(M_{\frak{p}_i}) \cdot \ell_R(R/\frak{p}_i+(\mathbf{x}_n))} \cdot |\Lambda_{W(\mathbb{F})}/(g_M(z),\mathbf{x}_n)| \nonumber
\end{align}
and similarly,
\begin{equation}
\label{absolutevalue2}
|N/\mathbf{x}_n N|=|\mathbb{F}|^{\sum_{i=1}^h \ell_{R_{\frak{p}_i}}(M_{\frak{p}_i}) \cdot \ell_R(R/\frak{p}_i+(\mathbf{x}_n))} \cdot |\Lambda_{W(\mathbb{F})}/(g_N(z),\mathbf{x}_n)|.
\end{equation}
Now putting together $(\ref{inequality1})$, $(\ref{linear57})$,
$(\ref{absolutevalue1})$ and $(\ref{absolutevalue2})$, we conclude that
$$
\dfrac{|N/\mathbf{x}_n N|}{|M/\mathbf{x}_n M|}~\mbox{tends to}~ 0~\mbox{as}~n~\mbox{goes to}~\infty.
$$
However, this is a contradiction to the hypothesis. That is, $(\ref{linear5})$ is false and we have thus proved $\rm (b) \Rightarrow (a)$. We complete the proof.\\
\end{proof}
\subsection{Specialization method in dimension two for height-one primes over $p$}
Let $(R,\frak{m},\mathbb{F})$ be a two-dimensional Noetherian complete local normal domain with mixed characteristic $p>0$ and finite residue field $\mathbb{F}$. Let us choose a parameter ideal $(p,z)$ of $R$ and define the set of principal ideals:
\begin{equation}
\label{Fittideal1}
\mathscr{E}_{R,W(\mathbb{F})}(z,r,\{a_n\}_{n \in \mathbb{N}}):=\{(\mathbf{x}_n)\}_{n \in \mathbb{N}},
\end{equation}
where $\mathbf{x}_n=z^n+r+a_n p$ is attached by Corollary \ref{LocalBertini} $\rm (2)$ and $\{a_n\}_{n \in \mathbb{N}}$ is any sequence of elements in $W(\mathbb{F})^{\times}$. It is necessary to make a special choice of $r \in R$. Let $\{\frak{p}_i\}_{1 \le i \le h}$ be the set of all height-one primes of $R$ over $p$. Consider an additional condition:
\begin{equation}
\label{linearideal2}
r_t \in \frak{p}_t~\mbox{and}~r_t \notin \bigcup_{i \ne t} \frak{p}_i~\mbox{for each}~1 \le t \le h.~\mbox{If}~h=1,~\mbox{then choose}~r=0.
\end{equation}
We notice that the existence of $r_t \in R$ satisfying $(\ref{linearideal2})$ is ensured by Prime Avoidance Lemma. By the choice made as above, $(p,z^n+r_i)$ is a parameter ideal of $R$ for $n \in \mathbb{N}$ and $1 \le i \le h$.
\begin{theorem}[Specialization Method over $p$]
\label{prop2}
Let $(R,\frak{m},\mathbb{F})$ be a two-dimensional Noetherian complete local normal domain with mixed characteristic $p>0$ and finite residue field $\mathbb{F}$.
Suppose that $M$ and $N$ are finitely generated torsion $R$-modules and fix a finite set of elements $\{r_i\}_{1 \le i \le h}$ satisfying $(\ref{linearideal2})$. Then the following statements are equivalent:
\begin{enumerate}
\item[(a)]
Let $\frak{p}$ be any height-one prime of $R$ which lies over $p$. Then we have
$$
\operatorname{char}_R(N)_{\frak{p}} \subseteq \operatorname{char}_R(M)_{\frak{p}}.
$$
\item[(b)]
There exists a constant $c \in \mathbb{N}$, depending on $M$ and $N$, such that
$$
c \cdot \frac{|N/\mathbf{x}N |}{|M/\mathbf{x}M |} \in \mathbb{N}
$$
for all but finitely many principal ideals
$$
(\mathbf{x}) \in \bigcup_{1\le i \le h} \mathscr{E}_{R,W(\mathbb{F})}(z,r_i,\{a_n\}_{n \in \mathbb{N}}).
$$
\end{enumerate}
\end{theorem}
\begin{proof}
Take the fundamental pseudo-isomorphisms
$$
f_M:M \to \bigoplus_i R/\frak{p}_i^{e_i} \oplus \bigoplus_r R/\frak{q}_r^{f_r}~\mbox{and}~f_N:N \to \bigoplus_j R/\frak{p}_j^{e'_j}\oplus \bigoplus_s R/\frak{q}_s^{f'_s},
$$
where $\{\frak{p}_i\}_{1 \le i \le h}$ is the set of all height-one primes of $R$ containing $p$, and $\{\frak{q}_i\}_{1 \le i \le k}$ is the set of height-one primes not containing $p$. Then as in the proof of Theorem \ref{prop1}, it is sufficient to prove the theorem for modules of the following types:
$$
M=R/J_M \oplus R/I_M~\mbox{and}~N=R/J_N \oplus R/I_N,
$$
where $J_M=\prod_i \frak{p}_i^{e_i}$, $I_M=\prod_r \frak{q}_r^{f_r}$ and $J_N=\prod_j \frak{p}_j^{e'_j}$, $I_N=\prod_s \frak{q}_s^{f'_s}$.
First we prove $\rm (a) \Rightarrow (b)$. Then we find that
$$
|M/\mathbf{x}M|=|R/J_M+(\mathbf{x})| \cdot |R/I_M+(\mathbf{x})|~\mbox{and}~|N/\mathbf{x}N|=|R/J_N+(\mathbf{x})| \cdot |R/I_N+(\mathbf{x})|.
$$
For those elements $\mathbf{x} \in R$ of specified type, it follows that $|R/I_M+(\mathbf{x})|$ is finite and bounded. So the implication $\rm (a) \Rightarrow (b)$ follows by putting
$$
c=\max_{\mathbf{x}}\{|R/I_M+(\mathbf{x})|\}.
$$
Next we prove $\rm (b) \Rightarrow (a)$. To prove it by contradiction, let us assume that
\begin{equation}
\label{linear6}
\ell_{R_{\frak{p}_j}}(M_{\frak{p}_j})>\ell_{R_{\frak{p}_j}}(N_{\frak{p}_j})
\end{equation}
for some $1 \le j \le h$. Thus, we may put $j=1$ for simplicity. Set $\mathbf{x}_n:=z^n+r_1+a_np \in R$ for $n>0$, where $r_1 \in \frak{p}_1$ and $r_1 \notin \bigcup_{i \ne 1} \frak{p}_i$. Recall the following formula:
\begin{equation}
\label{linear7}
\ell_R(M/\mathbf{x}M)=\sum_{i=1}^h \ell_{R_{\frak{p}_i}}(M_{\frak{p}_i}) \cdot \ell_R(R/\frak{p}_i+(\mathbf{x}))+\sum_{i=1}^k \ell_{R_{\frak{q}_i}}(M_{\frak{q}_i}) \cdot \ell_R(R/\frak{q}_i+(\mathbf{x}))
\end{equation}
and
\begin{equation}
\label{linear8}
\ell_R(N/\mathbf{x}N)=\sum_{i=1}^h \ell_{R_{\frak{p}_i}}(N_{\frak{p}_i}) \cdot \ell_R(R/\frak{p}_i+(\mathbf{x}))+\sum_{i=1}^k \ell_{R_{\frak{q}_i}}(M_{\frak{q}_i}) \cdot \ell_R(R/\frak{q}_i+(\mathbf{x}))
\end{equation}
under the assumption $ \mathbf{x} \notin \bigcup_{i=1}^h \frak{p}_i \cup \bigcup_{i=1}^k \frak{q}_i$. Let us denote by $\widetilde{R}_i$ the integral closure of $R_i:=R/\frak{p}_i$ in $\operatorname{Frac}(R_i)$, where $\frak{p}_i \in \operatorname{Spec} R$ is as above. As $R_i$ is a complete local domain of Krull dimension one, $R_i \to \widetilde{R}_i$ is module-finite and $\widetilde{R}_i$ is a discrete valuation ring. Moreover, $\widetilde{R}_i/R_i$ is a finitely generated torsion $R$-module. Hence we have $\ell_R(\widetilde{R}_i/R_i)<\infty$. Consider the following commutative diagram:
$$
\begin{CD}
0 @>>> R_i @>>> \widetilde{R}_i @>>> \widetilde{R}_i/R_i @>>> 0 \\
@. @V\times \mathbf{x}_nVV @V\times \mathbf{x}_nVV @V\times \mathbf{x}_nVV \\
0 @>>> R_i @>>> \widetilde{R}_i @>>> \widetilde{R}_i/R_i @>>> 0 \\
\end{CD}
$$
The left vertical and the middle vertical maps are injective, as the maps are the multiplication by a nonzero element in a domain $R_i$. Thus by applying the snake lemma to the above commutative diagram, we have the following exact sequence:
$$
0 \to \ker \big(\widetilde{R}_i/R_i \overset{\times\mathbf{x}_n}{\longrightarrow } \widetilde{R}_i/R_i\big)
\to R_i /(\mathbf{x}_n) \to \widetilde{R}_i/(\mathbf{x}_n)
\to \operatorname{Coker} \big(\widetilde{R}_i/R_i \overset{\times\mathbf{x}_n}{\longrightarrow} \widetilde{R}_i/R_i\big)
\to 0.
$$
Since we know $\ell_R(\widetilde{R}_i/R_i)<\infty$, it follows that the $R$-lengths of the modules in the first and the last terms of the sequence are equal to each other. Hence, the equality $\ell_R(R_i/(\mathbf{x}_n))=\ell_R(\widetilde{R}_i/(\mathbf{x}_n))$ follows. Let $v$ be the valuation of $\widetilde{R}_i$. Then we have $\ell_R(\widetilde{R}_i/(\mathbf{x}_n))=[\widetilde{\mathbb{F}}_i:\mathbb{F}] \cdot v(\mathbf{x}_n)$, where $\widetilde{\mathbb{F}}_i$ is the residue field of $\widetilde{R}_i$. We find that
\begin{equation}
\label{linear9}
\limsup_{n \to \infty}\ell_R(R/\frak{p}_i+(\mathbf{x}_n)) < \infty~\mbox{for any}~i \ne 1,
\end{equation}
and
\begin{equation}
\label{linear10}
\lim_{n \to \infty}\ell_R(R/\frak{p}_1+(\mathbf{x}_n))=\infty.
\end{equation}
On the other hand, since $a_np \notin \frak{q}_i$, we find that
\begin{equation}
\label{linear11}
\limsup_{n \to \infty}\ell_R(R/\frak{q}_i+(\mathbf{x}_n)) < \infty~\mbox{for any}~i.
\end{equation}
Now putting together $(\ref{cardinality})$, $(\ref{linear7})$, $(\ref{linear8})$, $(\ref{linear9})$, $(\ref{linear10})$ and $(\ref{linear11})$, we conclude that
$$
\dfrac{|N/\mathbf{x}_nN |}{|M/\mathbf{x}_nM |}~\mbox{tends to}~0~\mbox{as}~n~\mbox{goes to}~\infty.
$$
However, this is a contradiction to the hypothesis. That is, $(\ref{linear6})$ is false and we have thus proved $\rm (b) \Rightarrow (a)$. This completes the proof.\\
\end{proof}
\subsection{Specialization method in dimension at least three}
Let us briefly recall the specialization method in the case of dimension at least three, together with notation from \cite{OcSh}. The philosophy behind the method is that, if the inclusion $\operatorname{char}_{R/(\mathbf{x})}(M/\mathbf{x}M) \subseteq \operatorname{char}_{R/(\mathbf{x})}(N/\mathbf{x}N)$ is fulfilled for sufficiently many $\mathbf{x} \in R$, then we can recover the inclusion $\operatorname{char}_R(M) \subseteq \operatorname{char}_R(N)$.
Let $(R,\frak{m},\mathbb{F})$ be a Noetherian complete local domain with mixed characteristic and finite residue field $\mathbb{F}$, let $\overline{\mathbb{F}}$ be a fixed algebraic closure of $\mathbb{F}$ and let $W(\overline{\mathbb{F}})$ denote the ring of Witt vectors. Let $\mathbb{P}^n(W(\overline{\mathbb{F}}))$ denote the $n$-dimensional projective space with coordinates in $W(\overline{\mathbb{F}})$ such that every point of $\mathbb{P}^n(W(\overline{\mathbb{F}}))$ is \textit{normalized}; let us choose a point $a=(a_0:\cdots:a_d) \in \mathbb{P}^n(W(\overline{\mathbb{F}}))$. Then we require $a_i \in W(\overline{\mathbb{F}})^{\times}$ for some $0 \le i \le d$. Let us set
$$
R_{W(\overline{\mathbb{F}})}:=R \widehat{\otimes}_{W(\mathbb{F})} W(\overline{\mathbb{F}})~\mbox{and}~R_{W(\mathbb{F}')}:=R \otimes_{W(\mathbb{F})} W(\mathbb{F}'),
$$
where $\mathbb{F} \to \mathbb{F}'$ is a finite field extension. We shall need the following facts.
\begin{enumerate}
\item
$R_{W(\overline{\mathbb{F}})}$ and $R_{W(\mathbb{F}')}$ are complete local rings in view of Lemma \ref{localring}. Moreover, they are normal domains if $R$ is normal.
\item
The natural extension $R \to R_{W(\mathbb{F}')}$ is module-finite. The extension $R_{W(\mathbb{F}')} \to R_{W(\overline{\mathbb{F}})}$ is not integral, because $W(\mathbb{F}') \to W(\overline{\mathbb{F}})$ is not integral.
\item
Fix a set of minimal generators $x_0,\ldots,x_n$ of $\frak{m}$. Recall that we constructed a set-theoretic injective map \cite[(4.7)]{OcSh}:
$$
\theta_{W(\overline{\mathbb{F}})}: \mathbb{P}^n(\overline{\mathbb{F}}) \hookrightarrow \mathbb{P}^n(W(\overline{\mathbb{F}})).
$$
Then the composite map $\operatorname{Sp}_{W(\overline{\mathbb{F}})} \circ \theta_{W(\overline{\mathbb{F}})}$ is the identity map, where
$$
\operatorname{Sp}_{W(\overline{\mathbb{F}})}:\mathbb{P}^n(W(\overline{\mathbb{F}})) \twoheadrightarrow \mathbb{P}^n(\overline{\mathbb{F}})
$$
is the \textit{specialization map}. This map plays an essential role in the formulation of the local Bertini theorems in mixed characteristic.
\end{enumerate}
Let us fix a point $a=(a_0:\cdots:a_n) \in \theta_{W(\overline{\mathbb{F}})}(\mathbb{P}^n(\overline{\mathbb{F}}))$ and let $\widetilde{a}=(\widetilde{a}_0,\ldots,\widetilde{a}_n) \in \mathbb{A}^{n+1}(W(\overline{\mathbb{F}}))$ be its \textit{representative}. This we explain below. The point $\widetilde{a}=(\widetilde{a}_0,\ldots,\widetilde{a}_n) \in \mathbb{A}^{n+1}(W(\overline{\mathbb{F}}))$ is chosen in such a way that $\widetilde{a}_i \in W(\overline{\mathbb{F}})^{\times}$ for some $0 \le i \le n$. Let us form a linear combination
\begin{equation}
\label{linearcomb}
\mathbf{x}_{\widetilde{a}}:=\sum_{i=0}^n \widetilde{a}_ix_i,
\end{equation}
where $\mathbf{x}_{\widetilde{a}}$ itself depends on the choice of $\widetilde{a}$, but the principal ideal generated by it does not. We recall the following Avoidance Lemma (see \cite[Lemma 4.2]{OcSh}).
\begin{lemma}
\label{Case1}
Let us fix a set of minimal generators $x_0,\ldots,x_n$ of the maximal ideal of $R_{W(\overline{\mathbb{F}})}$, together with a non-maximal prime ideal $\frak{p}$ of $R_{W(\overline{\mathbb{F}})}$. Then there exists a non-empty Zariski open subset $U \subseteq \mathbb{P}^n(\overline{\mathbb{F}})$ such that
$$
\mathbf{x}_{\widetilde{a}}=\sum_{i=0}^n \widetilde{a}_ix_i \notin \frak{p}
$$
for every $a=(a_0:\cdots:a_n) \in \operatorname{Sp}_{W(\overline{\mathbb{F}})}^{-1}(U)$. If moreover $\widetilde{a} \in \theta_{W(\overline{\mathbb{F}})}(U)$, then there exists a finite field extension $\mathbb{F} \to \mathbb{F}'$ such that $\mathbf{x}_{\widetilde{a}}$ is an element of $R_{W(\mathbb{F}')}$.
\end{lemma}
Lemma \ref{Case1} ensures that one can find infinitely many specialization elements with certain properties. Let $M$ be a finitely generated torsion $R$-module. Then, in \cite[Definition 8.1]{OcSh}, we defined a certain subset:
$$
\mathcal{L}_{W(\overline{\mathbb{F}})}(M_{W(\overline{\mathbb{F}})}) \subseteq \mathbb{P}^n(W(\overline{\mathbb{F}}))
$$
consisting of specialization elements with good properties for $M$.
\begin{enumerate}
\item[-]
By \cite[Lemma 8.6]{OcSh}, the set $\mathcal{L}_{W(\overline{\mathbb{F}})}(M_{W(\overline{\mathbb{F}})})$ is identified with a Zariski open subset of $\mathbb{P}^n(\overline{\mathbb{F}})$ through the map $\theta_{W(\overline{\mathbb{F}})}: \mathbb{P}^n(\overline{\mathbb{F}}) \to \mathbb{P}^n(W(\overline{\mathbb{F}}))$.
\end{enumerate}
In other words, one can consider $\mathcal{L}_{W(\overline{\mathbb{F}})}(M_{W(\overline{\mathbb{F}})})$ as a classifying space of specialization elements \cite[Proposition 5.1]{OcSh}. Let $I$ be an ideal of a ring $A$. Let us denote by $\operatorname{Min}_A(I)$ the set of all prime ideals of $A$ that are minimal over $I$. Let us formulate \cite[Theorem 8.8]{OcSh} in the form we need. Let $M$ and $N$ be finitely generated torsion $R$-modules. Let $U_{M,N} \subseteq \mathbb{P}^n(\overline{\mathbb{F}})$ be the Zariski open subset such that $U_{M,N}$ corresponds to the intersection:
$$
\mathcal{L}_{W(\overline{\mathbb{F}})}(M_{W(\overline{\mathbb{F}})}) \cap \mathcal{L}_{W(\overline{\mathbb{F}})}(N_{W(\overline{\mathbb{F}})})
$$
via the map $\theta_{W(\overline{\mathbb{F}})}: \mathbb{P}^n(\overline{\mathbb{F}}) \to \mathbb{P}^n(W(\overline{\mathbb{F}}))$ defined in the above.
\begin{theorem}[Local Bertini theorem]
\label{theorem:previous}
Let $(R,\frak{m},\mathbb{F})$ be a Noetherian complete local normal domain with mixed characteristic $p>0$, finite residue field and $\operatorname{depth} R \ge 3$. Let us choose a non-empty Zariski open subset $U \subseteq U_{M,N}$ and let $(\mathbf{x}_{\widetilde{a}}) \subseteq R_{W(\overline{\mathbb{F}})}$ be the height-one prime ideal as in $(\ref{linearcomb})$ associated to a fixed point $a \in \theta_{W(\overline{\mathbb{F}})}(U)$. Then we can find a finite field extension $\mathbb{F} \to \mathbb{F}'$ such that
$\mathbf{x}_{\widetilde{a}}$ is an element of $R_{W(\mathbb{F}')}$ and $R_{W(\mathbb{F}')}/(\mathbf{x}_{\widetilde{a}})$ is a complete local normal domain with mixed characteristic. Moreover, the following statements are equivalent:
\begin{enumerate}
\item[\rm{(1)}]
$\operatorname{char}_R(M) \subseteq \operatorname{char}_R(N)$.
\item[\rm{(2)}]
For all but finitely many height-one primes:
$$
(\mathbf{x}_{\widetilde{a}}) \subseteq R_{W(\overline{\mathbb{F}})}~\mbox{with}~a \in \theta_{W(\overline{\mathbb{F}})}(U),
$$
we have
$$
\operatorname{char}_{R_{W(\mathbb{F}')}/(\mathbf{x}_{\widetilde{a}})}(M_{W(\mathbb{F}')}/\mathbf{x}_{\widetilde{a}}M_{W(\mathbb{F}')}) \subseteq \operatorname{char}_{R_{W(\mathbb{F}')}/(\mathbf{x}_{\widetilde{a}})}(N_{W(\mathbb{F}')}/\mathbf{x}_{\widetilde{a}}N_{W(\mathbb{F}')}),
$$
where $\mathbb{F}'$ is any finite field extension of $\mathbb{F}$ such that $\mathbf{x}_{\widetilde{a}}$ is an element of $R_{W(\mathbb{F}')}$.
\item[\rm{(3)}]
For all but finitely many height-one primes:
$$
(\mathbf{x}_{\widetilde{a}}) \subseteq R_{W(\overline{\mathbb{F}})}~\mbox{with}~a \in \theta_{W(\overline{\mathbb{F}})}(U),
$$
we have
$$
\operatorname{char}_{R_{W(\overline{\mathbb{F}})}/(\mathbf{x}_{\widetilde{a}})}(M_{W(\overline{\mathbb{F}})}/\mathbf{x}_{\widetilde{a}}M_{W(\overline{\mathbb{F}})}) \subseteq \operatorname{char}_{R_{W(\overline{\mathbb{F}})}/(\mathbf{x}_{\widetilde{a}})}(N_{W(\overline{\mathbb{F}})}/\mathbf{x}_{\widetilde{a}}N_{W(\overline{\mathbb{F}})}).
$$
\end{enumerate}
\end{theorem}
\begin{proof}
For the proof of the statement that $R_{W(\mathbb{F}')}/(\mathbf{x}_{\widetilde{a}})$ is a complete local normal domain with mixed characteristic, see \cite[Corollary 4.6]{OcSh}.
For the equivalence of the above statements, we refer the reader to \cite[Theorem 8.8]{OcSh} for complete details. Here, we indicate the requisite modifications. It suffices to prove the implication $(3) \Rightarrow (1)$ and we may assume that
$$
M=\bigoplus_i R_{W(\overline{\mathbb{F}})}/P_i^{e_i}~\mbox{and}~
N=\bigoplus_i R_{W(\overline{\mathbb{F}})}/Q_i^{f_i},
$$
where $\{P_i\}$ and $\{Q_i\}$ are certain finite sets of height-one primes. Fix a prime ideal $Q_k$ associated to $N$. As demonstrated in {\bf Step 1} of the proof of \cite[Theorem 8.8]{OcSh}, it is sufficient to find an infinite set of principal ideals $\{(\mathbf{x}_{\widetilde{a}_i})\}_{i \in \mathbb{N}}$ corresponding to the points of $\theta_{W(\overline{\mathbb{F}})}(U)$ such that
\begin{equation}
\label{MinMax}
\bigcup_{i \in \mathbb{N}} \operatorname{Min}_{R_{W(\overline{\mathbb{F}})}}\big(Q_k+(\mathbf{x}_{\widetilde{a}_i})\big)~\mbox{is an infinite set}.
\end{equation}
Since every prime ideal which belongs to $(\ref{MinMax})$ is of height-two in view of \cite[Definition 8.1]{OcSh} and the Krull dimension of $R$ is greater than or equal to three, these primes are properly contained in the maximal ideal of $R$. Then as in the proof of \cite[Lemma 8.6]{OcSh}, a choice of the sequence $\{\mathbf{x}_{\widetilde{a}_i}\}_{i \in \mathbb{N}}$ satisfying $(\ref{MinMax})$ is made possible by applying repeatedly Lemma \ref{Case1} to each finite set
$\operatorname{Min}_{R_{W(\overline{\mathbb{F}})}}(Q_k+(\mathbf{x}_{\widetilde{a}_i}))$. Now we complete the proof of the theorem.\\
\end{proof}
\section{Galois cohomology over number fields}
\label{GaloisCoho}
\subsection{Galois cohomology}
We fix a prime number $p>2$. Let $K$ be a number field and let $G_K=\operatorname{Gal}(\overline{K}/K)$ be the absolute Galois group. For a finite set $\Sigma$ of primes of $K$ containing all infinite primes, denote by $K_{\Sigma}$ the maximal extension of $K$ which is unramified outside $\Sigma$ and $G_{\Sigma}=\operatorname{Gal}(K_{\Sigma}/K)$. Most part of the discussions in this section goes over a complete Noetherian local ring that is torsion free and finite over $\Lambda$, where $\Lambda$ is either $\mathbb{Z}_p[[x_1,\ldots,x_d]]$ or $\mathbb{F}_p[[x_1,\ldots,x_d]]$. The symbol $(R,\frak{m},\mathbb{F})$ will denote a Noetherian complete local ring, where $\frak{m}$ is its unique maximal ideal and $\mathbb{F}$ is its finite residue field. Let $M$ be a compact or discrete $R$-module with a continuous $G_K$-action.
We put
$$
M^{\operatorname{PD}}:=\operatorname{Hom}(M,\mathbb{Q}_p/\mathbb{Z}_p),~M^*:=\operatorname{Hom}_R(M,R)~\mbox{and}~ M^*(1):=\operatorname{Hom}_R(M,R) \otimes_{\mathbb{Z}_p}\mathbb{Z}_p(1),
$$
where $\mathbb{Z}_p(1)$ is the \textit{Tate twist}. The $R$-module structure of $M^{\operatorname{PD}}$ is obvious. The $G_K$-action of $M^{\operatorname{PD}}$ is as follows. For $f:M \to \mathbb{Q}_p/\mathbb{Z}_p \in M^{\operatorname{PD}}$, we define $g \cdot f$ to be $(g \cdot f)(m):=f(g^{-1}m)$ for $g \in G_K$ and $m \in M$. For an $R$-module $M$ and an ideal $I \subseteq R$, we put
$$
M[I]:=\{m \in M~|~rm=0~\mbox{for all}~r \in I\}.
$$
For Galois cohomology groups for finite modules, we refer the reader to \cite{NSW}. In what follows, we will consider Galois cohomology for modules over $(R,\frak{m},\mathbb{F})$. If $M$ is a compact $R$-module with a continuous $G$-action for a profinite group $G$, its Galois cohomology is computed as
$$
H^i(G,M)=\varprojlim_n H^i(G,M/\frak{m}^nM).
$$
In the case that $M$ is a discrete $G$-module, we get
$$
H^i(G,M)=\varinjlim_n H^i(G,M[\frak{m}^n]).
$$
These cohomology groups coincide with the cohomology groups defined by continuous cochains, if $G$ is either $G_{\Sigma}$ or $G_{K_v}$ (of cohomological dimension $\ge 2$ if $p>2$). Here $G_{K_v}$ denotes the local absolute Galois group. It is known that if $M^{\operatorname{PD}}$ is a finitely generated $R$-module, then $H^i(G_{\Sigma},M)^{\operatorname{PD}}$ is finitely generated for all $i \ge 0$ \cite[Proposition 3.2]{Gr2}. We set
\begin{equation}
\label{TateShafarevich}
\textcyr{Sh}^i_{\Sigma}(M):=\ker\big(H^i(G_{\Sigma},M) \to \bigoplus_{v \in \Sigma} H^i(G_{\mathbb{Q}_v},M)\big).
\end{equation}
This is called the \textit{Tate-Shafarevich group}.
\begin{theorem}[Poitou-Tate duality]\label{theorem:P_T}
Let $M$ be either a compact or discrete $G_{\Sigma}$-module over a complete Noetherian local ring $(R,\frak{m},\mathbb{F})$. Then we have a perfect pairing:
$$
\textcyr{Sh}^1_{\Sigma}(M^{\operatorname{PD}}(1)) \times \textcyr{Sh}^2_{\Sigma}(M) \to \mathbb{Q}_p/\mathbb{Z}_p
$$
and there is the following exact sequence:
$$
0 \to H^0(G_{\Sigma},M) \to \bigoplus_{v \in \Sigma} H^0(G_{\mathbb{Q}_v},M) \to H^2(G_{\Sigma},M^{\operatorname{PD}}(1))^{\operatorname{PD}}
$$
$$
\to H^1(G_{\Sigma},M) \to \bigoplus_{v \in \Sigma} H^1(G_{\mathbb{Q}_v},M) \to H^1(G_{\Sigma},M^{\operatorname{PD}}(1))^{\operatorname{PD}}
$$
$$
\to H^2(G_{\Sigma},M) \to \bigoplus_{v \in \Sigma} H^2(G_{\mathbb{Q}_v},M) \to H^0(G_{\Sigma},M^{\operatorname{PD}}(1))^{\operatorname{PD}} \to 0.
$$
\end{theorem}
\begin{proof}
Assume that $M$ is a compact module. Since $R$ is the inverse limit of finite $p$-groups, say $R=\varprojlim_n R/\frak{m}^n$, $M=\varprojlim_n M/\frak{m}^n M$, and $M^{\operatorname{PD}}=\varinjlim_n M^{\operatorname{PD}}[\frak{m}^n]$, the theorem follows by taking inverse and inductive limits for the usual Poitou-Tate duality theorem \cite[Theorem 8.6.7 and 8.6.10]{NSW} for finite discrete $G_{\Sigma}$-modules. Note that the inverse limit is an exact functor in this case. The case that $M$ is discrete can be treated similarly.
\end{proof}
Let us recall (local and global) Euler-Poincar\'e characteristic formula.
\begin{theorem}
\label{Tate}
Let $K$ be a number field with $r_1$ real and $r_2$ complex places, respectively, and let $M$ be a finite discrete $p$-torsion $G_K$-module. Let $v$ be any finite place of $K$ with $v|p$. Then
\begin{equation}
\label{EulerPo1}
\frac{|H^0(G_{K_v},M)| \cdot |H^2(G_{K_v},M)|}{|H^1(G_{K_v},M)|}=p^{-[K_v:\mathbb{Q}_p] \cdot v_p(|M|)},
\end{equation}
where $v_p(-)$ is the $p$-adic valuation. Fix a finite set $\Sigma$ of places containing $p$ and all infinite places of $K$. Assume the $G_K$-module $M$ is unramified outside $\Sigma$. Then
\begin{equation}
\label{EulerPo2}
\frac{|H^0(G_{\Sigma},M)| \cdot |H^2(G_{\Sigma},M)|}{|H^1(G_{\Sigma},M)|}=\prod_{v:\ \mathrm{infinite}}|H^0(G_v,M)| \cdot p^{-(r_1+r_2) \cdot v_p(|M|)}.
\end{equation}
If $v$ does not divide $p$, then the quantity in the local Euler-Poincar\'e characteristic formula $(\ref{EulerPo1})$ is equal to one.
\end{theorem}
\begin{proof}
For $(\ref{EulerPo1})$ and $(\ref{EulerPo2})$, we refer the reader to \cite[Theorem 7.3.1 and Theorem 8.7.4]{NSW} and \cite[Theorem 3.2]{TK} for the proof and generalizations.\\
\end{proof}
We often use the following (well-known) lemma.
\begin{lemma}
\label{Pont}
Let $(R,\frak{m},\mathbb{F})$ be a Noetherian complete local ring with finite residue field, let $\frak{a} \subseteq R$ be an ideal and let $M$ be either a compact or discrete $R$-module. Then
$$
(M/\frak{a} M)^{\operatorname{PD}} \simeq M^{\operatorname{PD}}[\frak{a}].
$$
\end{lemma}
\begin{proof}
The claimed isomorphism follows easily from the short exact sequence $0 \to \frak{a} M \to M \to M/\frak{a} M \to 0$.
\end{proof}
\subsection{Control of Galois cohomology groups}
Let $(R,\frak{m},\mathbb{F})$ be a reduced local, finite and torsion free ring over $\Lambda$ with residual characteristic $p>2$, where as usual, $\Lambda$ is either $\mathbb{Z}_p[[x_1,\ldots,x_d]]$ or $\mathbb{F}_p[[x_1,\ldots,x_d]]$. Let $T$ be a continuous $R[G_{\mathbb{Q}}]$-module free of finite rank over $R$ (the base field $\mathbb{Q}$ may be replaced with a number field $K$ in the following discussion).
In this section, we prepare some technique of the specialization method inspired by some papers of Greenberg (see \cite{Gr1}, \cite{Gr2} for more discussions).
\begin{lemma}
\label{last}
Let $(R,\frak{m},\mathbb{F})$ be a Noetherian complete local domain of dimension at least two with finite residue field $\mathbb{F}$ of characteristic $p>2$.
Let $T$ be a continuous $R[G_{\mathbb{Q}}]$-module free of rank at least two over $R$ which is ramified at a finite set of primes $\Sigma$ which
contains the prime $p$. Let us assume the following conditions:
\begin{enumerate}
\item[\rm{(i)}]
The $G_{\mathbb{Q}}$-representation $T \otimes_R R/\frak{m}$ is irreducible.
\item[\rm{(ii)}]
$H^2 (G_{\mathbb{Q}_\ell},T^\ast (1))=0$ for every $\ell \in \Sigma \setminus \{\infty\}$.
\end{enumerate}
Then, the natural $R/\frak{a}$-module map:
$$
\textcyr{Sh}^2_{\Sigma}(T^\ast (1))/ \frak{a} \textcyr{Sh}^2_{\Sigma}(T^\ast (1))
\to
\textcyr{Sh}^2_{\Sigma}((T^*/\frak{a} T^*)(1)),
$$
which is induced by a surjection of $R[G_{\mathbb{Q}}]$-modules: $T^*(1) \twoheadrightarrow
(T^*/\frak{a} T^*)(1)$, is an isomorphism for any ideal $\frak{a} \subseteq R$.
\end{lemma}
Before proving the lemma, we prepare the notation.
Let $D:=T \otimes_R R^{\operatorname{PD}}$ on which $G_{\Sigma}=\operatorname{Gal}(\mathbb{Q}_{\Sigma}/\mathbb{Q})$ acts by hypothesis and let us note that
\begin{equation}
\label{niceexact}
(T^*)^{\operatorname{PD}} \simeq \operatorname{Hom}_{\mathbb{Z}_p}(T^* \otimes_R R,\mathbb{Q}_p/\mathbb{Z}_p) \simeq \operatorname{Hom}_R(T^*,R^{\operatorname{PD}}) \simeq D.
\end{equation}
\begin{proof}
By $(\ref{niceexact})$ together with the Poitou-Tate duality (Theorem \ref{theorem:P_T}) and Lemma \ref{Pont},
the natural $R/\frak{a}$-module map:
$$
\textcyr{Sh}^2_{\Sigma}(T^\ast (1))/ \frak{a} \textcyr{Sh}^2_{\Sigma} (T^\ast (1))
\to
\textcyr{Sh}^2_{\Sigma}((T^*/\frak{a} T^*)(1))
$$
is an isomorphism if and only if the natural map of $R/\frak{a}$-modules:
\begin{equation}\label{equation:sha^1_isom}
\textcyr{Sh}^1_{\Sigma}(D[\frak{a}]) \to \textcyr{Sh}^1_{\Sigma}(D)[\frak{a}]
\end{equation}
is an isomorphism. In order to prove this, we consider the following commutative diagram:
\begin{equation}\label{equation:commutative_diagram_d[a]}
\begin{CD}
0 @>>> \textcyr{Sh}^1_{\Sigma}(D[\frak{a}]) @>>> H^1 ( G_{\Sigma} , D[\frak{a} ] ) @>>> \displaystyle{\prod_{\ell \in \Sigma \setminus \{\infty\}}}
H^1(G_{\mathbb{Q}_\ell},D[\frak{a} ]) \\
@. @VVV @VVV @VVV \\
0 @>>> \textcyr{Sh}^1_{\Sigma}(D)[\frak{a}] @>>> H^1 ( G_{\Sigma} , D )[\frak{a} ] @>>> \displaystyle{\prod_{\ell \in \Sigma \setminus \{\infty\}}}
H^1(G_{\mathbb{Q}_\ell},D) [\frak{a} ]. \\
\end{CD}
\end{equation}
In order to show that \eqref{equation:sha^1_isom} is an isomorphism, it suffices to show that the map
$\prod_{\ell \in \Sigma} H^1(G_{\mathbb{Q}_\ell},D[\frak{a}]) \to \prod_{\ell \in \Sigma} H^1(G_{\mathbb{Q}_\ell},D)[\frak{a}]$ of
\eqref{equation:commutative_diagram_d[a]} is injective and the map $H^1(G_{\Sigma},D[\frak{a} ]) \to H^1(G_{\Sigma} ,D)[\frak{a}]$
of \eqref{equation:commutative_diagram_d[a]} is an isomorphism by the snake lemma.
First, we prove that $\prod_{\ell \in \Sigma} H^1(G_{\mathbb{Q}_\ell},D[\frak{a}]) \to \prod_{\ell \in \Sigma} H^1(G_{\mathbb{Q}_\ell},D)[\frak{a}]$ is injective. Let $\frak{a}=(x_1,\ldots,x_n)$. Then there is a short exact sequence:
\begin{equation}\label{equation:shortD}
0 \to D[x_1,\ldots,x_i] \to D[x_1,\ldots,x_{i-1}] \xrightarrow{\times x_i} x_i D[x_1,\ldots,x_{i-1}] \to 0.
\end{equation}
The condition $\rm(ii)$ together with the local Tate duality shows that $H^0(G_{\mathbb{Q}_{\ell}},D)=0$. Then since $x_i D[x_1,\ldots,x_{i-1}]$ is a submodule of $D$, the induced map:
$$
\alpha_i:H^1(G_{\mathbb{Q}_\ell},D[x_1,\ldots,x_i]) \to H^1(G_{\mathbb{Q}_\ell},D[x_1,\ldots,x_{i-1}])[x_i ]
$$
is injective. Composing the maps $\alpha_i$ for $i=1,\ldots,n$, we see that the map $H^1(G_{\mathbb{Q}_\ell},D[\frak{a}]) \to
H^1(G_{\mathbb{Q}_\ell},D)[\frak{a} ]$ is injective. We have proved that the map \eqref{equation:sha^1_isom} is injective and thus, the map $\prod_{\ell \in \Sigma} H^1(G_{\mathbb{Q}_\ell},D[\frak{a}]) \to \prod_{\ell \in \Sigma} H^1(G_{\mathbb{Q}_\ell},D)[\frak{a}]$ is also injective.
Next, we prove that $H^1(G_{\Sigma},D[\frak{a} ]) \to H^1(G_{\Sigma} ,D)[\frak{a}]$ is an isomorphism. Since the rank of $T$ is at least two by assumption, the condition $\rm(i)$ together with Nakayama's lemma implies that $D$ admits no nonzero subquotient on which $G_{\Sigma}$ acts trivially. In particular, we have $H^0(G_{\Sigma},D)=0$. Then by the same argument as above, we can show that
$H^1(G_{\Sigma},D[\frak{a}]) \to H^1(G_{\Sigma},D)[\frak{a}]$ is injective. In order to prove the surjectivity of $H^1(G_{\Sigma},D[\frak{a}]) \to H^1(G_{\Sigma},D)[\frak{a}]$,
we consider the short exact sequence:
$$
0 \to x_i D[x_1,\ldots,x_{i-1}] \to D[x_1,\ldots,x_{i-1}] \to D[x_1,\ldots,x_{i-1}]/x_i D[x_1,\ldots,x_{i-1}] \to 0.
$$
Since $D[x_1,\ldots,x_{i-1}]/x_i D[x_1,\ldots,x_{i-1}]$ is a subquotient of $D$, we have
$$
H^0(G_{\Sigma},D[x_1,\ldots,x_{i-1}]/x_i D[x_1,\ldots,x_{i-1}])=0
$$
and hence the following injection is induced:
\begin{equation}\label{eaution:injection}
H^1(G_{\Sigma},x_i D[x_1,\ldots,x_{i-1}]) \hookrightarrow H^1(G_{\Sigma},D[x_1,\ldots,x_{i-1}]).
\end{equation}
We also note that $D[x_1,\ldots,x_{i-1}] \xrightarrow{\times x_i} D[x_1,\ldots,x_{i-1}]$ factors as
\begin{equation}\label{equation:composition}
D[x_1,\ldots,x_{i-1}] \to x_iD[x_1,\ldots,x_{i-1}] \to D[x_1,\ldots,x_{i-1}].
\end{equation}
The injection \eqref{eaution:injection} combined with the map of cohomology induced from \eqref{equation:composition} yields
\begin{multline*}
\ker\big[H^1(G_{\Sigma},D[x_1,\ldots,x_{i-1}]) \to H^1(G_{\Sigma},x_iD[x_1,\ldots,x_{i-1}])\big]
\\
=\ker\big[ H^1(G_{\Sigma},D[x_1,\ldots,x_{i-1}]) \xrightarrow{\times x_i} H^1(G_{\Sigma},D[x_1,\ldots,x_{i-1}])\big] .
\end{multline*}
This equality combined with the long exact sequence of cohomology induced from \eqref{equation:shortD}
shows the surjectivity of the natural map $\beta_i:H^1(G_{\Sigma},D[x_1,\ldots,x_i]) \to H^1(G_{\Sigma},D[x_1,\ldots,x_{i-1}])[x_i]$. Composing the maps $\beta_i$ for $i=1,\ldots,n$, we see that $H^1(G_{\Sigma},D[\frak{a} ]) \to H^1(G_{\Sigma} ,D)[\frak{a}]$ is surjective. This completes the proof of the lemma.
\end{proof}
\begin{lemma}
\label{last2}
Let $(R,\frak{m},\mathbb{F})$ be a Noetherian complete local normal domain of dimension at least two with finite residue field $\mathbb{F}$ of characteristic $p>0$ and assume that the following conditions hold:
\begin{enumerate}
\item[\rm{(i)}]
$H^2(G_{\Sigma},D)$ is a finite group (resp. a trivial group).
\item[\rm{(ii)}]
$T^*(1)/(T^*(1))^{G_{\mathbb{Q}_\ell}}$ is $R$-reflexive for all $\ell \in \Sigma$.
\item[\rm{(iii)}]
$(T^*(1))^{G_{\mathbb{Q}_{\ell_0}}}=0$ for at least one $\ell_0 \in \Sigma \setminus \{\infty\}$.
\end{enumerate}
Then there exist finitely many height-one primes $\frak{p}_1,\ldots,\frak{p}_m$ with the property that, if $\frak{a} \subseteq R$ is a principal ideal with $\frak{a} \nsubseteq \bigcup_{i=1}^m \frak{p}_i$, then
$$
H^2(G_{\Sigma},D[\frak{a}])
$$
is a finite group (resp. a trivial group).
\end{lemma}
\begin{proof}
Recall that a Noetherian domain is \textit{reflexive}, if we have $R=\cap_{\frak{p}} R_{\frak{p}}$, where $\frak{p}$ ranges over all height-one primes of $R$. Since $R$ is a normal domain, it is a reflexive domain and a result of Greenberg \cite[Theorem 1]{Gr2} applies to our case in view of $\rm(ii)$ and $\rm(iii)$. The Pontryagin dual of $\textcyr{Sh}^2_{\Sigma}(D)$ is $R$-reflexive. However, since $H^2(G_{\Sigma},D)$ is finite by $\rm(i)$, it follows that $\textcyr{Sh}^2_{\Sigma}(D)$ is finite and hence it is trivial. We can apply \cite[Proposition 6.10]{Gr2} to our situation. The proof of \cite[Proposition 6.10]{Gr2} will work over a module-finite extension of $\Lambda$. Write $\frak{a}=(f)$ with $f \in R$. Then
we have $H^2(G_{\Sigma},D[\frak{a}])=H^2(G_{\Sigma},D[f])$. Consider the long exact sequence:
$$
H^1(G_{\Sigma},D) \xrightarrow{\times f} H^1(G_{\Sigma},D) \to H^2(G_{\Sigma},D[f]) \to H^2(G_{\Sigma},D) \xrightarrow{\times f} H^2(G_{\Sigma},D) \to 0,
$$
which is induced by $0 \to D[f] \to D \xrightarrow{\times f} D \to 0$. Then the map $H^1(G_{\Sigma},D) \xrightarrow{\times f} H^1(G_{\Sigma},D)$ is surjective, if $\frak{a} \nsubseteq \bigcup_{i=1}^m \frak{p}_i$ is satisfied for a certain set of height-one primes $\frak{p}_1,\ldots,\frak{p}_m$ of $R$ by \cite[Proposition 6.10]{Gr2}. Hence the map $H^2(G_{\Sigma},D[f]) \to H^2(G_{\Sigma},D)$ is injective, which proves that $H^2(G_{\Sigma},D[f])$ is finite by $\rm(i)$. The case where $H^2(G_{\Sigma},D)$ is trivial can be treated similarly.\\
\end{proof}
\section{Euler system theory over a Noetherian complete local ring}
\label{EulerSystem}
\subsection{Definition of Euler system}
An Euler system for a $p$-adic Galois representation is defined as a norm-compatible system of elements in the tower of the first Galois cohomology groups. We fix a prime number $p>0$ and let $\Sigma$ be a finite set of primes of $\mathbb{Q}$ containing $\{p,\infty \}$. Let $n>0$ be a square-free integer that is not divisible by primes in $\Sigma$ and let $\mathbb{Q}(\mu_n)$ be the $n$-th cyclotomic extension by adjoining a primitive $n$-th root of unity $\mu_n$. Let $G_{\mathbb{Q}(\mu_n)}:=\operatorname{Gal}(\overline{\mathbb{Q}}/\mathbb{Q}(\mu_n))$ and let $\mathbb{Q}(\mu_n)_{\Sigma}/\mathbb{Q}(\mu_n)$ be the maximal Galois extension that is unramified outside primes lying above those in $\Sigma$ and put $G_{\Sigma,n}:=\operatorname{Gal}(\mathbb{Q}(\mu_n)_{\Sigma}/\mathbb{Q}(\mu_n))$. We will write $G_{\Sigma}:=G_{\Sigma,1}$ for simplicity.
\begin{definition}[Euler system over $\mathbb{Q}$]
\label{definitionEuler}
Let $T$ be a finite free module over a Noetherian complete local ring $(R,\frak{m},\mathbb{F})$ with finite residue field, where $T$ is equipped with a continuous $G_{\mathbb{Q}}$-action and $T$ is unramified outside a finite set of primes $\Sigma$ containing $\{p, \infty\}$. Let $\mathfrak{N}$ be the set of square-free integers which are all relatively prime to $\Sigma$. An \textit{Euler system} for $(T,R,\Sigma)$ is a collection of cohomology classes
$$
\big\{\mathbf{z}_n \in H^1(G_{\mathbb{Q}(\mu_n)},T^*(1))\big\}_{n \in \mathfrak{N}}
$$
which satisfies the following properties:
\begin{enumerate}
\item[\rm{(i)}]
$\mathbf{z}_n$ is unramified at all primes of $\mathbb{Q}(\mu_n)$ which are not over primes dividing $np$.
\item[\rm{(ii)}]
For a prime $\ell$ with $(n,\ell)=1$ and for $n\ell \in \mathfrak{N}$, we have
$$
\operatorname{Cor}_{\mathbb{Q}(\mu_{n\ell})/\mathbb{Q}(\mu_{n})}(\mathbf{z}_{n\ell})=P(\mathrm{Frob}_{\ell};T)\mathbf{z}_n,
$$
where $P(t;T)=\det(1-\operatorname{Frob}_{\ell}t:T \to T)$ and $\operatorname{Frob}_{\ell}$ is the geometric Frobenius in the Galois group $\operatorname{Gal}(\mathbb{Q}(\mu_n)/\mathbb{Q})$.
\end{enumerate}
\end{definition}
For Euler system over a general number field, we refer the reader to \cite{Rub}. By Definition \ref{definitionEuler} (i), we see that $\mathbf{z}_n$ descends to an element in a finitely generated $R$-module $H^1(G_{\Sigma,n},T^*(1))$. There are a few examples of Euler systems discovered and its finding is a deep arithmetic problem. We will need the following condition:
\begin{enumerate}
\item[($\bf{RED}$):] $(R,\frak{m},\mathbb{F})$ is a reduced local ring and finite flat over $\mathbb{Z}_p$.
\end{enumerate}
A commutative ring satisfying the condition $(\bf{RED})$ is a one-dimensional local ring, but it is not necessarily a \textit{principal ideal ring}. Thus, it is necessary to develop techniques which allow us to calculate Galois cohomology groups.
\begin{lemma}
\label{dis1}
Assume that $(R,\frak{m},\mathbb{F})$ is a Noetherian complete reduced local ring with finite residue field. Let $T$ be a continuous $R[G_{\Sigma}]$-module free of rank at least two over $R$. Assume that the $G_{\Sigma}$-representation $T \otimes_R R/\frak{m}$ is irreducible. Then $H^1(G_{\Sigma},T^*(1))$ is an $R$-torsion free module.
\end{lemma}
\begin{proof}
The assumption implies that $T^*(1) \otimes_R R/\frak{m}$ is irreducible, since $T^*(1)$ is a twist of $T^*$ by a cyclotomic character and we have $(T^*(1) \otimes_R R/\frak{m})^{G_{\Sigma}}=0$ by the assumption that $\operatorname{rank}_R T \ge 2$. It follows that for any nonzero divisor $x \in \frak{m}$, the quotient $T^*(1)/x T^*(1)$ is irreducible as a representation of $G_{\Sigma}$ by Nakayama's lemma. Then this implies that
$$
(T^*(1)/x T^*(1))^{G_{\Sigma}}=0,
$$
and the short exact sequence $0 \to T^*(1) \xrightarrow{\times x} T^*(1) \to T^*(1)/x T^*(1) \to 0$ yields that the map $H^1(G_{\Sigma},T^*(1)) \xrightarrow{\times x} H^1(G_{\Sigma},T^*(1))$ is injective. In particular, $H^1(G_{\Sigma},T^*(1))$ contains no nonzero pseudo-null submodules.
\end{proof}
\begin{definition}
\label{definition:isogeneous}
Let $T_1$ and $T_2$ be finitely generated torsion-free $R$-modules over a Noetherian complete local ring $(R,\frak{m},\mathbb{F})$ with finite residue field, equipped with continuous $G_K$-actions. Then $T_1$ and $T_2$ are \textit{isogeneous}, if there exists an injective $R[G_K]$-homomorphism $\phi:T_1 \to T_2$ whose cokernel is annihilated by a nonzero divisor of $R$. In this case, $\phi$ is called an \textit{isogeny}.
\end{definition}
\begin{lemma}
The relation of being isogeneous defined in Definition \ref{definition:isogeneous} is an equivalence relation.
\end{lemma}
\begin{proof}
Let us take an injective $R[G_K]$-homomorphism $\phi:T_1 \hookrightarrow T_2$ as in Definition \ref{definition:isogeneous}, whose cokernel is annihilated by a nonzero divisor $x$ of $R$. Identify $T_1$ with its image $\phi(T_1)$. Since $x$ annihilates $T_2/T_1$, the $R[G_K]$-module $xT_2$ is identified with a submodule of $T_1$. So we have that $T_1/xT_2$ is a submodule of $T_2/xT_2$ and $x(T_1/xT_2)=0$. Since $x$ is a nonzero divisor of $R$, the multiplication $R \xrightarrow{\times x} R$ is injective. Since $T_2$ is a torsion-free $R$-module, we have an isomorphism $T_2 \simeq xT_2$. So we obtained a $G_K$-equivariant injection $T_2 \simeq xT_2 \hookrightarrow T_1$ whose cokernel is annihilated by $x$.
Suppose that we are given another injective $R[G_K]$-homomorphism $\phi':T_2 \to T_3$ as in Definition \ref{definition:isogeneous}, whose cokernel is annihilated by a nonzero divisor $y$ of $R$. Then the cokernel of the injective $R[G_K]$-homomorphism $\phi' \circ \phi :T_1 \to T_3$ is annihilated by $xy$.
Thus we proved that the relation of being isogeneous defined in Definition \ref{definition:isogeneous}
is an equivalence relation.
\end{proof}
We construct an Euler system in an isogeneous representation.
\begin{lemma}
Assume that the ring $(R,\frak{m},\mathbb{F})$ satisfies the condition $(\bf{RED})$. Let $T$ be a continuous $R[G_{\Sigma}]$-module which is finite free over $R$ and let $\widetilde{T}:=T \otimes_R \widetilde{R}$, where $\widetilde{R}$ is the normalization of $R$ in its total ring of fractions. Let $\big\{\mathbf{z}_n \in H^1(G_{\mathbb{Q}(\mu_n)},T^*(1))\big\}_{n \in \mathfrak{N}}$ be an Euler system for $(T,R,\Sigma)$. Then there is a natural $R$-module map
$$
H^1(G_{\mathbb{Q}(\mu_n)},T^*(1)) \to H^1(G_{\mathbb{Q}(\mu_n)},\widetilde{T}^*(1))
$$
and one can obtain an Euler system for $(\widetilde{T},\widetilde{R},\Sigma)$:
$$
\big\{\widetilde{\mathbf{z}}_n \in H^1(G_{\mathbb{Q}(\mu_n)},\widetilde{T}^*(1))\big\}_{n \in \mathfrak{N}}
$$
which is given as the image of $\big\{\mathbf{z}_n \in H^1(G_{\mathbb{Q}(\mu_n)},T^*(1))\big\}_{n \in \mathfrak{N}}$.
Finally, assume that $|H^1(G_{\Sigma},T^*(1))/R \mathbf{z}_1|<\infty$. Then we have $|H^1(G_{\Sigma},\widetilde{T}^*(1))/\widetilde{R} \widetilde{\mathbf{z}}_1|<\infty$.
\end{lemma}
\begin{proof}
Since $\widetilde{R}/R$ is finite, the cokernel of the injective composed map
$$
\operatorname{Hom}_R(T,R) \to \operatorname{Hom}_R(T,R)\otimes_R \widetilde{R} \to \operatorname{Hom}_{\widetilde{R}}(T \otimes_R \widetilde{R},\widetilde{R})
$$
is finite. Let us put $\widetilde{T}^*(1):=\operatorname{Hom}_{\widetilde{R}}(T \otimes_R \widetilde{R},\widetilde{R})(1)$. This induces a natural map of $R$-modules with finite kernel and cokernel:
\begin{equation}
\label{induced1}
H^1(G_{\mathbb{Q}(\mu_n)},T^*(1)) \to H^1(G_{\mathbb{Q}(\mu_n)},\widetilde{T}^*(1)).
\end{equation}
Then via $(\ref{induced1})$, one can define a family of elements:
$$
\big\{\widetilde{\mathbf{z}}_n \in H^1(G_{\mathbb{Q}(\mu_n)},\widetilde{T}^*(1))\big\}_{n \in \mathfrak{N}}
$$
as the image of $\big\{\mathbf{z}_n \in H^1(G_{\mathbb{Q}(\mu_n)},T^*(1))\big\}_{n \in \mathfrak{N}}$. This gives an Euler system for $(\widetilde{T},\widetilde{R},\Sigma)$. Indeed, it is evident from the construction that $\widetilde{\mathbf{z}}_n$ is unramified outside $\Sigma$, which descends to an element in $H^1(G_{\Sigma,n},\widetilde{T}^*(1))$ via the surjection $G_{\mathbb{Q}(\mu_n)} \twoheadrightarrow G_{\Sigma,n}$.
Finally, assume that $|H^1(G_{\Sigma},T^*(1))/R \mathbf{z}_1|< \infty$. Since both kernel and cokernel of $(\ref{induced1})$ are finite, it follows that $|H^1(G_{\Sigma},\widetilde{T}^*(1))/\widetilde{R} \widetilde{\mathbf{z}}_1|< \infty$ by snake lemma.
\end{proof}
\begin{lemma}
\label{dis2}
Let $(R,\frak{m},\mathbb{F})$ be a Noetherian complete local ring with finite residue field and let $M$ be a finitely generated $R$-module with a continuous $G$-action, where $G$ is a profinite group.
\begin{enumerate}
\item[(i)]
Assume that $R$ is reduced and $R \to \widetilde{R}$ is the normalization map. Then the induced map
$$
H^i(G,M) \otimes_R \widetilde{R} \to H^i(G,M \otimes_R \widetilde{R})
$$
has $R$-torsion kernel and cokernel for all $i \ge 0$.
\item[(ii)]
Let $(R,\frak{m},\mathbb{F}) \to (S,\frak{n},\mathbb{F}')$ be a flat map of complete Noetherian local rings with finite residue fields. Then the natual map:
$$
H^i(G,M) \otimes_R S \to H^i(G,M \otimes_R S)
$$
is an isomorphism for all $i \ge 0$.
\end{enumerate}
\end{lemma}
\begin{proof}
We prove the assertion $\rm(i)$. Let $P_{\bullet}$ denote the projective resolution of the trivial $G$-module $\mathbb{Z}$. Then the natural map
$$
\operatorname{Hom}(P_{\bullet},M) \otimes_R \widetilde{R} \to \operatorname{Hom}(P_{\bullet},M \otimes_R \widetilde{R})
$$
is an isomorphism, since every $P_i$ is a finitely generated projective module. Since $R$ is a complete local ring and $\operatorname{Frac}(R)=\operatorname{Frac}(\widetilde{R})$, it follows that $R \to \widetilde{R}$ is module-finite and $\widetilde{R}/R$ is an $R$-torsion module. Thus, both kernel and cokernel of the natural map $H^i(G,M) \otimes_R \widetilde{R} \to H^i(\operatorname{Hom}(P_{\bullet},M) \otimes_R \widetilde{R})$ are $R$-torsion. Composing this map with $H^i(\operatorname{Hom}(P_{\bullet},M) \otimes_R \widetilde{R}) \simeq H^i(G,M \otimes_R \widetilde{R})$, both kernel and cokernel of the cohomology map
$$
H^i(G,M) \otimes_R \widetilde{R} \to H^i(G,M \otimes_R \widetilde{R})
$$
are $R$-torsion modules.
We prove the assertion $\rm(ii)$. Since $R \to S$ is flat, we have $(M \otimes_R S)^H=M^H \otimes_R S$ for any subgroup $H \subseteq G$. The proof may be completed by keeping track of $\rm(i)$ by replacing $R \to \widetilde{R}$ with $R \to S$.
\end{proof}
\begin{lemma}
\label{cond}
Assume that $(R,\frak{m},\mathbb{F})$ satisfies the condition $(\bf{RED})$ and $T$ is a finite free $R$-module with a continuous $G_{K_v}$-action, where $K$ is a number field and $v$ is a non Archimedean prime of $K$. Let $\widetilde{T}:=T \otimes_R \widetilde{R}$ and assume that $H^2(G_{K_v},T^*(1))=0$. Then $H^2(G_{K_v},\widetilde{T}^*(1))=0$.
\end{lemma}
\begin{proof}
Since $H^2(G_{K_v},\widetilde{T}^*(1))^{\operatorname{PD}} \simeq H^0(G_{K_v},(\widetilde{T}^*)^{\operatorname{PD}})=((\widetilde{T}^*)_{G_{K_v}})^{\operatorname{PD}}$ by local Tate duality theorem \cite[Theorem 1.4.1]{Rub}, we have $H^2(G_{K_v},\widetilde{T}^*(1))=0$ if and only if $(\widetilde{T}^*)_{G_{K_v}}=0$. Hence, it suffices to prove that $(\widetilde{T}^*)_{G_{K_v}}=0$.
We remark that $H^2(G_{K_v},{T}^*(1))=0$ if and only if $({T}^*)_{G_{K_v}}=0$ by the same argument using local Tate duality.
Consider the short exact sequence of $R$-modules $0 \to R \to \widetilde{R} \to \widetilde{R}/R \to 0$.
Since $R$ satisfies $(\bf{RED})$, the quotient $\widetilde{R}/R$ is an
Artinian $R$-module.
By applying the tensor product $(-)\otimes_R T^*$ with free $R$-module $T^*$
to the above short exact sequence,
we have a short exact sequence $0 \to T^* \to \widetilde{T}^* \to T^* \otimes_R \widetilde{R}/R \to 0$.
Taking the $G_{K_v}$-coinvariant quotient, we have an exact sequence:
$$
(T^*)_{G_{K_v}} \to (\widetilde{T}^*)_{G_{K_v}} \to (T^* \otimes_R \widetilde{R}/R)_{G_{K_v}} \to 0.
$$
By the above remark together with the assumption $H^2(G_{K_v},T^*(1))=0$,
the first term $(T^*)_{G_{K_v}}$ of this sequence is already known to be zero.
From this, it suffices to prove that the last term
$(T^* \otimes_R \widetilde{R}/R)_{G_{K_v}}$ is zero in order to prove that the middle
term is zero. Since $R$ is local, any simple
$R$-subquotient of $\widetilde{R}/R$ is isomorphic to $R/\frak{m} R$.
Hence, the $R$-module $T^* \otimes_R \widetilde{R}/R$ has finite length and is obtained by taking successive extensions of $T^*/\frak{m} T^*$. Recall that we have $(T^*/\frak{m} T^*)_{G_{K_v}}=0$ since
$(T^*)_{G_{K_v}}=0$.
If we have an exact sequence $0 \to M' \to M \to M'' \to 0$ of Artinian $R$-modules,
we have an exact sequence as follows
$$
(T^* \otimes_R M' )_{G_{K_v}} \to (T^* \otimes_R M)_{G_{K_v}} \to (T^* \otimes_R M'' )_{G_{K_v}} \to 0 .
$$
by the argument similar to the above. By the induction argument with respect to the $R$-length of $M$,
we prove $(T^* \otimes_R M)_{G_{K_v}}=0$ for any Artinian $A$-module $M$ and hence $(T^* \otimes_R \widetilde{R}/R)_{G_{K_v}}=0$.\\
\end{proof}
\subsection{Euler system bound over a reduced local ring of dimension one}
Let $T$ be as in the setting of Euler system over $\mathbb{Q}$ and recall the notation $D:=T \otimes_R R^{\operatorname{PD}}$, the discrete Galois module on which the Galois group $G_{\mathbb{Q}}$ acts through the first factor.
\begin{definition}
Let $(R,\frak{m},\mathbb{F})$ be a complete reduced local ring satisfying $(\bf{RED})$. Then \textit{Euler system bound} holds for $(T,R,\Sigma)$, if both of the following conditions are satisfied:
\begin{enumerate}
\item
$\textcyr{Sh}^2_{\Sigma}(T^*(1))$ is a finite group.
\item
$c \cdot |H^1(G_{\Sigma},T^*(1))/R \mathbf{z}_1|$ is divisible by $| \textcyr{Sh}^2_{\Sigma}(T^*(1))|$ for some $c \in \mathbb{N}$. Let us call $c$ an \textit{error term}, which is explicitly specified by $(T,R,\Sigma)$.
\end{enumerate}
\end{definition}
It is a deep arithmetic problem to determine under which conditions $H^1(G_{\Sigma},T^*(1))/R \mathbf{z}_1$ (or $\textcyr{Sh}^2_{\Sigma}(T^*(1))$) is a finite group. Our aim in \S~\ref{proof2} is to prove that Euler system bound holds over a module-finite extension $\Lambda=\mathbb{Z}_p[[x_1,\ldots,x_d]] \to R$, where $R$ is a Cohen-Macaulay normal domain. To this aim, we prove the following theorem as a direct generalization of Euler system bound over a discrete valuation ring, which was previously established by Kato, Perrin-Riou and Rubin independently (see \cite{KatoKodai}, \cite{Pe} and \cite{Rub} for their work), to a local ring satisfying $(\bf{RED})$. See also \cite[Theorem 4.7]{Oc1}. Let $\mu_A(M)$ denote the number of minimal generators of a finitely generated module $M$ over a local ring $(A,\frak{m})$. This is known to be equal to $\dim_{A/\frak{m}}(M/\frak{m} M)$ by Nakayama's lemma.
\begin{theorem}\label{Euler}
Let $(R,\frak{m},\mathbb{F})$ be a complete Noetherian reduced local ring which is finite and flat over $\mathbb{Z}_p$ for a prime number $p>2$. Suppose that
$$
\big\{\mathbf{z}_n \in H^1(G_{\Sigma,n},T^*(1))\big\}_{n \in \mathfrak{N}}
$$
is an Euler system for $(T,R,\Sigma)$, $T$ is a free $R$-module of rank two and suppose that the following conditions hold:
\begin{enumerate}
\item[\rm{(i)}]
$T \otimes_R R/\frak{m}$ is absolutely irreducible as a representation of $G_{\mathbb{Q}}$.
\item[\rm{(ii)}]
The quotient $H^1(G_{\Sigma},T^*(1))/R \mathbf{z}_1$ is a finite group.
\item[\rm{(iii)}]
$H^2(G_{\mathbb{Q}_{\ell}},T^*(1))=0$ for every $\ell \in \Sigma \setminus \{\infty\}$ and $H^2(G_{\Sigma},D)$ is a finite group.
\item[\rm{(iv)}]
The determinant character $\wedge^2 \rho:G_{\mathbb{Q}} \to R^{\times}$ (resp. $\wedge^2 \rho^*(1):G_{\mathbb{Q}} \to R^{\times}$) associated with the $G_{\mathbb{Q}}$-representation $T$ (resp. $T^*(1)$) has an element of infinite order.
\item[\rm{(v)}]
The $R$-module $T$ splits into eigenspaces: $T=T^+ \oplus T^-$ with respect to the complex conjugation in $G_{\mathbb{Q}}$, and $T^+_{\frak{p}}$ (resp. $T^-_{\frak{p}}$) is of $R_{\frak{p}}$-rank one for each minimal prime $\frak{p}$ of $R$.
\item[\rm{(vi)}]
There exist $\sigma_1 \in G_{\mathbb{Q}(\mu_{p^{\infty}})}$ and $\sigma_2 \in G_{\mathbb{Q}}$ such that $\rho(\sigma_1) \simeq \begin{pmatrix} 1 & \epsilon \\ 0 & 1 \end{pmatrix} \in GL_2(R)$ for a nonzero divisor $\epsilon \in R$ and $\sigma_2$ acts on $T$ as multiplication by $-1$.
\end{enumerate}
Then Euler system bound holds for $(T,R,\Sigma)$:
\begin{enumerate}
\item[\rm{(1)}]
$\textcyr{Sh}^2_{\Sigma}(T^*(1))$ is a finite group.
\item[\rm{(2)}]
We have
$$
c \cdot |H^1(G_{\Sigma},T^*(1))/R \mathbf{z}_1|~\mbox{is divisible by}~| \textcyr{Sh}^2_{\Sigma}(T^*(1))|,
$$
where $c:=|R/(\epsilon^k)|$ is an error term for Euler system bound and $k$ is the number of minimal generators of the $R$-module $\textcyr{Sh}^2_{\Sigma}(T^*(1))$.
\end{enumerate}
\end{theorem}
Before starting the proof, we fix notation. We denote by $\widetilde{R}$ the integral closure of the reduced local ring $R$ in its total ring of fractions. Let $\frak{p}_1,\ldots,\frak{p}_n$ be the set of all minimal primes of $R$. Set $R_i:=R/\frak{p}_i$, $\widetilde{T}:=T \otimes_R \widetilde{R}$ and $\widetilde{T}_i:=T \otimes_R \widetilde{R}_i$.
\begin{proof}
First, we prove the assertion $\rm(1)$, which is not so hard. Then we establish the assertion $\rm(2)$, which requires some computations using Poitou-Tate and Pontryagin dualities.
Let us first show that the hypotheses of the theorem remain true after replacing $(T,R,\Sigma)$ with $(\widetilde{T},\widetilde{R},\Sigma)$. By construction, we have the decomposition $\widetilde{R} =\bigoplus_i \widetilde{R}_i$, where the product is the finite dirct product of rings and each $\widetilde{R}_i$ is a $p$-torsion free complete discrete valuation ring with finite residue field. Moreover, there is the decomposition $\widetilde{T}=\bigoplus_i \widetilde{T}_i$ and each $\widetilde{T}_i$ is a free $\widetilde{R}_i$-module of rank two. The map $R \to \widetilde{R}_i$ factors as $R \twoheadrightarrow R_i \hookrightarrow \widetilde{R}_i$ and the cokernel of $R_i \hookrightarrow \widetilde{R}_i$ is finite. Putting this together with Lemma \ref{dis2} and Lemma \ref{cond}, it is straightforward to see that the hypotheses of the theorem are satisfied for the $G_{\mathbb{Q}}$-representation $\widetilde{T}_i$, and the assumption on the Galois elements $\sigma_1, \sigma_2$ in the theorem remains true for $\widetilde{T}_i$. It follows from the snake lemma that
\begin{equation}
\label{EuSys00}
c=\big|R/(\epsilon^k)\big|=\big|\widetilde{R}/(\epsilon^k)\big|.
\end{equation}
Let us prove $\rm(1)$. Consider a natural injection:
$$
T^*:=\operatorname{Hom}_R(T,R) \to \operatorname{Hom}_R(T,R)\otimes_R \widetilde{R} \to \widetilde{T}^*:=\operatorname{Hom}_{\widetilde{R}}(T \otimes_R \widetilde{R},\widetilde{R}).
$$
This induces a map of $R$-modules $\textcyr{Sh}^2_{\Sigma}(T^*(1)) \to \textcyr{Sh}^2_{\Sigma}(\widetilde{T}^*(1))$. Since the injection $R \hookrightarrow \widetilde{R}$ has finite cokernel, the above map has finite kernel and cokernel. By the result of Euler system bound over a discrete valuation ring $\widetilde{R}_i$ (see \cite[Theorem 4.7]{Oc1}), the $\widetilde{R}_i$-module $\textcyr{Sh}^2_{\Sigma}(\widetilde{T}^*_i(1))$ is a finite group. The same is true for $\textcyr{Sh}^2_{\Sigma}(\widetilde{T}^*(1))$. Then we see that $\textcyr{Sh}^2_{\Sigma}(T^*(1))$ is also finite. Thus, we have proved $\rm(1)$.
Let us prove $\rm(2)$. According to Euler system bound over a discrete valuation ring applied to $(\widetilde{T}_i,\widetilde{R}_i,\Sigma)$, we claim the following statement:
\begin{equation}
\label{EuSys1}
c_i \cdot |H^1(G_{\Sigma},\widetilde{T}^*_i(1))/\widetilde{R}_i \mathbf{z}_1|~\mbox{is divisible by}~|\textcyr{Sh}^2_{\Sigma}(\widetilde{T}^*_i(1))|,
\end{equation}
for each $i$, where $c_i:=|\widetilde{R}_i/(\epsilon^k)|$ and $k$ is as in the theorem. Then $(\ref{EuSys1})$ implies the following statement:
\begin{equation}
\label{EuSys2}
c \cdot |H^1(G_{\Sigma},\widetilde{T}^*(1))/\widetilde{R} \mathbf{z}_1|~\mbox{is divisible by}~|\textcyr{Sh}^2_{\Sigma}(\widetilde{T}^*(1))|,
\end{equation}
where $c=|R/(\epsilon^k)|=|\widetilde{R}/(\epsilon^k)|$ by $(\ref{EuSys00})$. So let us prove $(\ref{EuSys1})$. Indeed, in view of \cite[Theorem 4.7]{Oc1}, it suffices to prove that
\begin{equation}
\label{EuSys3}
\mu_{\widetilde{R}_i}\big(\textcyr{Sh}^2_{\Sigma}(\widetilde{T}_i^*(1))\big)=k.
\end{equation}
Since $H^2(G_{\mathbb{Q}_{\ell}},\widetilde{T}_i^*(1))=0$ for $\ell \in \Sigma\setminus\{\infty\}$, we have the following isomorphism by Lemma \ref{last}:
$$
\textcyr{Sh}^2_{\Sigma}(\widetilde{T}_i^*(1))/\varpi_i \textcyr{Sh}^2_{\Sigma}(\widetilde{T}_i^*(1)) \simeq \textcyr{Sh}^2_{\Sigma}((\widetilde{T}_i^*/\varpi_i \widetilde{T}_i^*)(1)),
$$
where $\varpi_i$ is a generator of the maximal ideal of $\widetilde{R}_i$. By Nakayama's lemma, we have
$$
\mu_{\widetilde{R}_i}\big(\textcyr{Sh}^2_{\Sigma}(\widetilde{T}_i^*(1))\big)=\dim_{\widetilde{R}_i/(\varpi_i)}\big(\textcyr{Sh}^2_{\Sigma}((\widetilde{T}^*_i/\varpi_i \widetilde{T}^*_i)(1))\big).
$$
Applying the field extension $R/\frak{m} \to \widetilde{R}_i/(\varpi_i)$ to Lemma \ref{dis2}, it follows from $(\ref{TateShafarevich})$ that
$$
\dim_{\widetilde{R}_i/(\varpi_i)}\big(\textcyr{Sh}^2_{\Sigma}((\widetilde{T}^*_i/\varpi_i \widetilde{T}^*_i)(1))\big)=\dim_{R/\frak{m}}\big(\textcyr{Sh}^2_{\Sigma}((T^*/\frak{m} T^*)(1))\big).
$$
By the hypothesis $\rm(iii)$ and Lemma \ref{last}, we get
$$
\textcyr{Sh}^2_{\Sigma}(T^*(1))/\frak{m} \textcyr{Sh}^2_{\Sigma}(T^*(1)) \simeq \textcyr{Sh}^2_{\Sigma}((T^*/\frak{m} T^*)(1)).
$$
Putting all above equations together, $(\ref{EuSys3})$ (and thus $(\ref{EuSys1})$) follows, as required.
The rest of the proof will be devoted to proving the equality:
\begin{equation}
\label{EuSys4}
|\textcyr{Sh}^2_{\Sigma}(\widetilde{T}^*(1))|=r \cdot |\textcyr{Sh}^2_{\Sigma}(T^*(1))|,
\end{equation}
where
\begin{equation}
\label{number1}
r:=|\operatorname{Coker}[H^1(G_{\Sigma},T^*(1)) \to H^1(G_{\Sigma},\widetilde{T}^*(1))]|.
\end{equation}
Applying the functor $(-)^{\operatorname{PD}}(1)$ to the exact sequence $0 \to T^*(1) \to \widetilde{T}^*(1) \to \mathcal{C} \to 0$ (the cokernel $\mathcal{C}$ is a finite $p$-group), we get an exact sequence of discrete modules: $0 \to \mathcal{C}^{\operatorname{PD}}(1) \to \widetilde{D} \to D \to 0$. We claim that this induces the commutative diagram with exact rows:
$$
\footnotesize{
\begin{CD}
0 @>>> H^1(G_{\Sigma},\mathcal{C}^{\operatorname{PD}}(1)) @>\phi_1>> H^1(G_{\Sigma},\widetilde{D}) @>>> H^1(G_{\Sigma},D) @>\psi_1>> H^2(G_{\Sigma},\mathcal{C}^{\operatorname{PD}}(1)) \\
@. @V\operatorname{loc}^1_{\mathcal{C}^{\operatorname{PD}}(1)}VV @V\operatorname{loc}^1_{\widetilde{D}} VV @V\operatorname{loc}^1_{D}VV @V\operatorname{loc}^2_{\mathcal{C}^{\operatorname{PD}}(1)}VV \\
0 @>>> \displaystyle{\bigoplus_{\ell \in \Sigma}} H^1(G_{\mathbb{Q}_\ell},\mathcal{C}^{\operatorname{PD}}(1)) @>\phi_2>> \displaystyle{\bigoplus_{\ell \in \Sigma}} H^1(G_{\mathbb{Q}_\ell},\widetilde{D}) @>>> \displaystyle{\bigoplus_{\ell \in \Sigma}} H^1(G_{\mathbb{Q}_\ell},D) @>\psi_2>> \displaystyle{\bigoplus_{\ell \in \Sigma}} H^2(G_{\mathbb{Q}_\ell},\mathcal{C}^{\operatorname{PD}}(1)) \\
\end{CD}}
$$
Henceforth, we denote by $\mathcal{C}_1 \xrightarrow{\phi} \mathcal{C}_2$ the map $\operatorname{Coker}(\phi_1) \to \operatorname{Coker}(\phi_2)$ (resp. by $\mathcal{C}'_1 \xrightarrow{\phi'} \mathcal{C}'_2$ the map $\operatorname{Im}(\psi_1) \to \operatorname{Im}(\psi_2)$). Let us check the exactness of the above commutative diagram. We can verify that $\phi_1$ is injective as follows. Since $(T^*(1)/\frak{m} T^*(1))_{G_{\Sigma}}=0$ in view of the hypothesis $\rm(i)$ and
$\operatorname{rank}_R T =2$, Nakayama's lemma implies that $(T^*(1))_{G_{\Sigma}}=0$, and thus $\phi_1$ is injective. Also, since $p>2$ and $\mathcal{C}^{\operatorname{PD}}(1)$ is a finite $p$-group, $H^1(G_{\mathbb{R}},\mathcal{C}^{\operatorname{PD}}(1))=0$. Thus, $\phi_2$ is injective by the hypothesis (iii). For later use, we prove the following:
\begin{equation}
\label{surj1}
\operatorname{loc}^2_{\mathcal{C}^{\operatorname{PD}}(1)}~\mbox{is surjective}.
\end{equation}
Indeed, the cokernel of $\operatorname{loc}^2_{\mathcal{C}^{\operatorname{PD}}(1)}$ is equal to $H^0(G_{\Sigma},\mathcal{C})^{\operatorname{PD}}$ by Poitou-Tate duality. To prove that this group vanishes, it suffices to prove that $H^0(G_{\Sigma},\mathcal{C})=0$. By the exact sequence,
$$
\cdots \to H^0(G_{\Sigma},\widetilde{T}^*(1)) \to H^0(G_{\Sigma},\mathcal{C}) \to H^1(G_{\Sigma},T^*(1)) \to \cdots
$$
we know that $H^0(G_{\Sigma},\widetilde{T}^*(1))$ vanishes by the hypothesis (i) and $H^1(G_{\Sigma},T^*(1))$ is $R$-torsion free by Lemma \ref{dis1}. Since $H^0(G_{\Sigma},\mathcal{C})$ is finite, it vanishes. Thus, $\operatorname{loc}^2_{\mathcal{C}^{\operatorname{PD}}(1)}$ is surjective.
Next, we prove the following:
\begin{equation}
\label{surj2}
\textcyr{Sh}^2_{\Sigma}(\widetilde{D})=0.
\end{equation}
Let us recall that $\ker(\operatorname{loc}^2_{\widetilde{D}})=\textcyr{Sh}^2_{\Sigma}(\widetilde{D})$, where $\operatorname{loc}^2_{\widetilde{D}}:H^2(G_{\Sigma},\widetilde{D}) \to \bigoplus_{\ell \in \Sigma} H^2(G_{\mathbb{Q}_{\ell}},\widetilde{D})$. The Pontryagin dual of $\textcyr{Sh}^2_{\Sigma}(\widetilde{D})$, which is $\textcyr{Sh}^1_{\Sigma}(\widetilde{T}^*(1))$, injects into $H^1(G_{\Sigma},\widetilde{T}^*(1))$ by definition. Then $H^1(G_{\Sigma},\widetilde{T}^*(1))$ is torsion free by Lemma \ref{dis1} and the hypothesis (i). Then, $\textcyr{Sh}^1_{\Sigma}(\widetilde{T}^*(1))$ is also torsion free, while the hypothesis (iii) implies that $\textcyr{Sh}^2_{\Sigma}(\widetilde{D})$ (say $\textcyr{Sh}^1_{\Sigma}(\widetilde{T}^*(1))$) is finite. Hence $\textcyr{Sh}^2_{\Sigma}(\widetilde{D})=0$.
Applying the snake lemma to the commutative diagram whose rows are exact:
$$
\begin{CD}
0 @>>> H^1(G_{\Sigma},\mathcal{C}^{\operatorname{PD}}(1)) @>\phi_1>> H^1(G_{\Sigma},\widetilde{D}) @>>> \mathcal{C}_1@>>> 0\\
@. @V\operatorname{loc}^1_{\mathcal{C}^{\operatorname{PD}}(1)}VV @V\operatorname{loc}^1_{\widetilde{D}}VV @V\phi VV \\
0 @>>> \displaystyle{\bigoplus_{\ell \in \Sigma}} H^1(G_{\mathbb{Q}_\ell},\mathcal{C}^{\operatorname{PD}}(1)) @>\phi_2>> \displaystyle{\bigoplus_{\ell \in \Sigma}} H^1(G_{\mathbb{Q}_\ell},\widetilde{D}) @>>> \mathcal{C}_2 @>>> 0\\
\end{CD}
$$
we obtain
\begin{equation}
\label{EuSys5}
\frac{|\textcyr{Sh}^1_{\Sigma}(\widetilde{D})|}{r_1}=\frac{|H^1(G_{\Sigma},\mathcal{C}^{\operatorname{PD}}(1))|}{|\bigoplus_{\ell \in \Sigma} H^1(G_{\mathbb{Q}_{\ell}},\mathcal{C}^{\operatorname{PD}}(1))|} \cdot |\ker(\phi)|,
\end{equation}
where
$$
r_1:=|\ker[\operatorname{Coker}(\operatorname{loc}^1_{\widetilde{D}}) \to \operatorname{Coker}(\phi)]|.
$$
Next, consider the following commutative diagram whose rows are exact:
$$
\begin{CD}
0 @>>> \mathcal{C}'_1 @>>> H^2(G_{\Sigma},\mathcal{C}^{\operatorname{PD}}(1)) @>>> H^2(G_{\Sigma},\widetilde{D}) \\
@. @V\phi' VV @V\operatorname{loc}^2_{\mathcal{C}^{\operatorname{PD}}(1)}VV @V\operatorname{loc}^2_{\widetilde{D}}VV \\
0 @>>> \mathcal{C}'_2@>>> \displaystyle{\bigoplus_{\ell \in \Sigma}} H^2(G_{\mathbb{Q}_\ell},\mathcal{C}^{\operatorname{PD}}(1)) @>>> \displaystyle{\bigoplus_{\ell \in \Sigma}} H^2(G_{\mathbb{Q}_\ell},\widetilde{D}) \\
\end{CD}
$$
Since $\operatorname{loc}^2_{\widetilde{D}}$ is injective by $(\ref{surj2})$, we have
$\ker(\phi')=\ker(\operatorname{loc}^2_{\mathcal{C}^{\operatorname{PD}}(1)})$. On the other hand, it follows from the injectivity of $\operatorname{loc}^2_{\widetilde{D}}$ and the surjectivity of $\operatorname{loc}^2_{\mathcal{C}^{\operatorname{PD}}(1)}$ by $(\ref{surj1})$ that
\begin{equation}
\label{EuSys666}
\phi'~\mbox{is surjective}.
\end{equation}
Moreover, we have
\begin{equation}
\label{EuSys6666}
\frac{|\mathcal{C}'_2|}{|\mathcal{C}'_1|}=\frac{|\bigoplus_{\ell \in \Sigma} H^2(G_{\mathbb{Q}_\ell},\mathcal{C}^{\operatorname{PD}}(1))|}{|H^2(G_{\Sigma},\mathcal{C}^{\operatorname{PD}}(1))|}.
\end{equation}
Then applying the snake lemma together with $(\ref{EuSys666})$ to the commutative diagram with exact rows:
$$
\begin{CD}
0 @>>> \mathcal{C}_1 @>>> H^1(G_{\Sigma},D) @>\psi_1>> \mathcal{C}'_1 @>>> 0 \\
@. @V\phi VV @V\operatorname{loc}^1_{D}VV @V\phi'VV \\
0 @>>> \mathcal{C}_2@>\phi_2>> \displaystyle{\bigoplus_{\ell \in \Sigma}} H^1(G_{\mathbb{Q}_\ell},D) @>\psi_2>> \mathcal{C}'_2 @>>> 0\\
\end{CD}
$$
we obtain
\begin{equation}
\label{EuSys66}
\frac{1}{|\textcyr{Sh}^1_{\Sigma}(D)|}=\frac{|\mathcal{C}'_2|} {|\mathcal{C}'_1|} \cdot \frac{1}{|\ker(\phi)|} \cdot \frac{|\operatorname{Coker}(\phi)|}{|\operatorname{Coker}(\operatorname{loc}^1_D)|}= \frac{|\mathcal{C}'_2|} {|\mathcal{C}'_1|} \cdot \frac{r_2}{|\ker(\phi)|},
\end{equation}
where
$$
r_2:=|\ker[\operatorname{Coker}(\phi) \to \operatorname{Coker}(\operatorname{loc}^1_{D})]|.
$$
Now putting together $(\ref{EuSys6666})$ and $(\ref{EuSys66})$, we get
\begin{equation}
\label{EuSys6}
\frac{1}{|\textcyr{Sh}^1_{\Sigma}(D)|}=\frac{|\bigoplus_{\ell \in \Sigma} H^2(G_{\mathbb{Q}_\ell},\mathcal{C}^{\operatorname{PD}}(1))|} {|H^2(G_{\Sigma},\mathcal{C}^{\operatorname{PD}}(1))|} \cdot \frac{r_2}{|\ker(\phi)|},
\end{equation}
Putting together $(\ref{EuSys5})$ and ($\ref{EuSys6})$ and applying the Poitou-Tate duality theorem, we get the following equality:
$$
\frac{|\textcyr{Sh}^2_{\Sigma}(\widetilde{T}^*(1))|}{|\textcyr{Sh}^2_{\Sigma}(T^*(1))|}=\frac{|H^1(G_{\Sigma},\mathcal{C}^{\operatorname{PD}}(1))|}{| H^2(G_{\Sigma},\mathcal{C}^{\operatorname{PD}}(1))|} \cdot \frac{|\bigoplus_{\ell \in \Sigma} H^2(G_{\mathbb{Q}_\ell},\mathcal{C}^{\operatorname{PD}}(1))|}{|\bigoplus_{\ell \in \Sigma} H^1(G_{\mathbb{Q}_\ell},\mathcal{C}^{\operatorname{PD}}(1))|} \cdot r_1 \cdot r_2.
$$
Note that $H^0(G_{\Sigma},\mathcal{C}^{\operatorname{PD}}(1))=0$ by $H^0(G_{\Sigma},\widetilde{D})=0$. Then by the Euler-Poincar\'e characteristic formula $(\ref{EulerPo1})$ and $(\ref{EulerPo2})$ in Theorem \ref{Tate}, treating cases according as $\ell=p$, $\ell \in \Sigma \setminus \{p,\infty\}$, or $\ell=\infty$, respectively, we obtain the following equality:
$$
|\textcyr{Sh}^2_{\Sigma}(\widetilde{T}^*(1))|=|\textcyr{Sh}^2_{\Sigma}(T^*(1))| \cdot r_1 \cdot r_2.
$$
It remains to show that $r_1 \cdot r_2$ is equal to $r$ which is given in $(\ref{number1})$. For its proof, consider the commutative diagram induced by Poitou-Tate duality, whose rows are exact:
$$
\begin{CD}
0 @>>> \operatorname{Coker}(\operatorname{loc}^1_{\widetilde{D}}) @>>> H^1(G_{\Sigma},\widetilde{T}^*(1))^{\operatorname{PD}} @>>> \textcyr{Sh}^2_{\Sigma}(\widetilde{D}) @>>> 0\\
@. @VV\tau_1V @VV\tau_2V @VVV \\
0 @>>> \operatorname{Coker}(\operatorname{loc}^1_{D}) @>>> H^1(G_{\Sigma},T^*(1))^{\operatorname{PD}} @>>> \textcyr{Sh}^2_{\Sigma}(D) @>>> 0\\
\end{CD}
$$
We claim that $r=|\ker(\tau_2)|$. Since $r_1 \cdot r_2=|\ker(\tau_1)|$, this is equivalent to showing that the induced injective map $\ker(\tau_1) \to \ker(\tau_2)$ is surjective. However, since $\textcyr{Sh}^2_{\Sigma}(\widetilde{D})=0$ by $(\ref{surj2})$, the equality $r=r_1 \cdot r_2$ follows. So we have proved ($\ref{EuSys4}$).
Finally in view of the hypothesis (ii), we claim that
\begin{equation}
\label{EuSys7}
|H^1(G_{\Sigma},\widetilde{T}^*(1))/\widetilde{R} \mathbf{z}_1|=\frac{r}{|\widetilde{R}/R|} \cdot |H^1(G_{\Sigma},T^*(1))/R \mathbf{z}_1|,
\end{equation}
and all groups are finite. To see this, consider the commutative diagram of $R$-modules with exact rows:
$$
\begin{CD}
0 @>>> R @>>> \widetilde{R} @>>> \widetilde{R}/R @>>> 0 \\
@. @V\times \mathbf{z}_1VV @V\times \mathbf{z}_1VV @V\times 0VV \\
0@>>> H^1(G_{\Sigma},T^*(1)) @>>> H^1(G_{\Sigma},\widetilde{T}^*(1)) @>>> \mathcal{M} @>>> 0 \\
\end{CD}
$$
We have $r=|\mathcal{M}|$ by definition and $(\ref{EuSys7})$ follows from the snake lemma. Now, incorporating both $(\ref{EuSys4})$ and $(\ref{EuSys7})$ into $(\ref{EuSys2})$, the required Euler system bound for $(T,R,\Sigma)$ is obtained. This finishes the proof of the theorem.\\
\end{proof}
\subsection{Euler system bound over a Cohen-Macaulay normal domain}
In this subsection, we prove an Euler system bound over Cohen-Macaulay normal complete local domains with Krull dimension at least two using Theorem \ref{Euler}. Later in \S \ref{proof2}, we shall show one of divisibilities of the Iwasawa main conjecture for two-variable Hida deformations by applying this Euler system bound. The main tool in the proof of the Euler system bound is the specialization methods that we have developed so far. We mention several facts on the behavior of Galois cohomology and characteristic ideals under etale extensions.
Let $(R,\frak{m},\mathbb{F})$ be a Noetherian complete local normal domain, whose coefficient ring is $W(\mathbb{F})$.
For a finite field extension $\mathbb{F} \to \mathbb{F}'$, we define $R_{W(\mathbb{F}')}:=R \otimes_{W(\mathbb{F})} W(\mathbb{F}')$.
Then we have the following.
\begin{enumerate}
\item
Let $M$ be a finitely generated $R$-module with continuous $G$-action. Then we have
\begin{equation}
\label{controlGal}
H^i(G,M) \otimes_R R_{W(\mathbb{F}')} \simeq H^i(G,M \otimes_R R_{W(\mathbb{F}')}),
\end{equation}
\begin{equation}
\label{controlSha}
\textcyr{Sh}^2_{\Sigma}(M) \otimes_R R_{W(\mathbb{F}')} \simeq \textcyr{Sh}^2_{\Sigma}(M \otimes_R R_{W(\mathbb{F}')}),
\end{equation}
which follow from Lemma \ref{dis2} and the definition of Tate-Shafarevich group.
\item
The following equality between reflexive ideals holds:
\begin{equation}
\label{controlchar}
\operatorname{char}_R(M) R_{W(\mathbb{F}')}=\operatorname{char}_{R_{W(\mathbb{F}')}}(M \otimes_R R_{W(\mathbb{F}')}),
\end{equation}
which is easy to check from the definition.
\end{enumerate}
\begin{lemma}
\label{topological}
Let $(R,\frak{m},\mathbb{F})$ be a two-dimensional Noetherian complete local normal domain with mixed characteristic $p>0$ and finite residue field. Let $u \in R^{\times}$ be an element of infinite order. Then the following statements hold.
\begin{enumerate}
\item[\rm{(1)}]
Let us choose $z \in \frak{m}$ such that $(p,z)$ is a parameter ideal of $R$.
Then, for each $n>0$, $R/(z+a_np^n)$ is a reduced ring of mixed characteristic, and
the image of $u \in R^{\times}$ under the reduction map
$$
R^{\times} \to \big(R/(z+a_np^n)\big)^{\times}
$$
is of infinite order for infinitely many different choices $a_n \in W(\mathbb{F})^{\times}$.
\item[\rm{(2)}]
Let us choose $z, r \in \frak{m}$ such that $(p,z^n+r)$ is a parameter ideal of $R$ for all $n>0$. Then, for each $n>0$, $R/(z^n+r+a_n p)$ is a reduced ring of mixed characteristic, and the image of $u \in R^{\times}$ under the reduction map
$$
R^{\times} \to \big(R/(z^n+r+a_np^n)\big)^{\times}
$$
is of infinite order for infinitely many different choices $a_n \in W(\mathbb{F})^{\times}$.
\end{enumerate}
\end{lemma}
\begin{proof}
Since the proofs of the assertions $\rm(1)$ and $\rm(2)$ proceed almost in the same manner, we only prove $\rm(2)$. Let us fix an integer $n>0$ and consider a module-finite extension $W(\mathbb{F})[[z^n+r]] \to R$. By Corollary \ref{LocalBertini} $\rm(2)$, there is a finite subset $S_n \subseteq W(\mathbb{F})^{\times}$ such that $R/(z+r+a_n p^n)$ is reduced of mixed characteristic for every $a_n \in W(\mathbb{F})^{\times} \setminus S_n$. Let us put
$$
\mathcal{U}_n:=\{P \in \operatorname{Spec} R~|~(u^k-1) \subseteq P~\mbox{for some}~k~\mbox{and}~\operatorname{ht}(P)=1\}.
$$
Let $\mathcal{V}_n$ denote the image of $\mathcal{U}_n$ under the map $\operatorname{Spec} R \to \operatorname{Spec} W(\mathbb{F}) [[z^n+r]]$. Then we can check that $\mathcal{V}_n$ is a countable set. Let us put
$$
T_n:=\{a_n \in W(\mathbb{F})^{\times}~|~(z^n+r+a_np) \in \mathcal{V}_n\}.
$$
Notice that if we choose $a_n \ne a'_n$, then
$$
(z^n+r+a_n p)~\mbox{and}~(z^n+r+a'_n p)
$$
are distinct prime ideals of $W(\mathbb{F})[[z^n+r]]$. Moreover, $W(\mathbb{F})^{\times}$ is an \textit{uncountable} set (for this, one uses the fact that elements of $W(\mathbb{F})$ are written as $p$-series) and it follows that $W(\mathbb{F})^{\times} \setminus S_n \cup T_n$ is uncountable. In particular, it is infinite. Now we conclude that the ring $R/(z^n+r+a_np)$ is reduced of mixed characteristic and the image of $u$ under $R^{\times} \to \big(R/(z^n+r+a_np)\big)^{\times}$ is of infinite order for every $a_n \in W(\mathbb{F})^{\times} \setminus S_n \cup T_n$. This completes the proof of the lemma.
\end{proof}
We shall prove the following theorem via reduction to Theorem \ref{Euler}.
\begin{theorem}\label{veryfinal}
Let $(R,\frak{m},\mathbb{F})$ be a Noetherian complete local Cohen-Macaulay normal domain
of Krull dimension $d\geq 2$ with mixed characteristic $p>2$ and finite residue field $\mathbb{F}$. Suppose that
$$
\big\{\mathbf{z}_n \in H^1(G_{\Sigma,n},T^*(1))\big\}_{n \in \mathfrak{N}}
$$
is an Euler system for $(T,R,\Sigma)$, $T$ is a free $R$-module of rank two with
continuous $G_{\mathbb{Q}}$-action and suppose that the following conditions hold:
\begin{enumerate}
\item[\rm{(i)}]
$T \otimes_R R/\frak{m}$ is absolutely irreducible as a representation of $G_{\mathbb{Q}}$.
\item[\rm{(ii)}]
The quotient $H^1(G_{\Sigma},T^*(1))/R \mathbf{z}_1$ is an $R$-torsion module.
\item[\rm{(iii)}]
$H^2(G_{\mathbb{Q}_{\ell}},T^*(1))=0$ for every $\ell \in \Sigma\setminus\{\infty\}$ and $H^2(G_{\Sigma},D)$ is a finite group.
\item[\rm{(iv)}]
The determinant character $\wedge^2 \rho:G_{\mathbb{Q}} \to R^{\times}$ (resp. $\wedge^2 \rho^*(1):G_{\mathbb{Q}} \to R^{\times}$) associated with the $G_{\mathbb{Q}}$-representation $T$ (resp. $T^*(1)$) has an element of infinite order.
\item[\rm{(v)}]
The $R$-module $T$ splits into eigenspaces: $T=T^+ \oplus T^-$ with respect to the complex conjugation in $G_{\mathbb{Q}}$, and $T^+$ (resp. $T^-$) is of $R$-rank one.
\item[\rm{(vi)}]
There exist $\sigma_1 \in G_{\mathbb{Q}(\mu_{p^{\infty}})}$ and $\sigma_2 \in G_{\mathbb{Q}}$ such that $\rho(\sigma_1) \simeq \begin{pmatrix} 1 & P \\ 0 & 1 \end{pmatrix} \in GL_2(R)$ for a nonzero element $P \in R$ and $\sigma_2$ acts on $T$ as multiplication by $-1$.
\end{enumerate}
Then Euler system bound holds for $(T,R,\Sigma)$:
\begin{enumerate}
\item[\rm{(1)}]
$\textcyr{Sh}^2_{\Sigma}(T^*(1))$ is a finitely generated torsion $R$-module.
\item[\rm{(2)}]
We have an inclusion of reflexive ideals:
\begin{equation}
\label{inclusionchar}
(P^k) \operatorname{char}_R\big(H^1 (G_{\Sigma},T^*(1))/R \mathbf{z}_1\big) \subseteq \operatorname{char}_R\big(\textcyr{Sh}^2_{\Sigma} (T^*(1))\big),
\end{equation}
where $k$ is the number of minimal generators of $\textcyr{Sh}^2_{\Sigma}(T^*(1))$ as an $R$-module.
\end{enumerate}
\end{theorem}
\begin{proof}
Let us put $d:=\dim R$. Let $\mathcal{S}$ be a set of ideals of $R$ (resp. $R_{W(\overline{\mathbb{F}})}$) which consists of certain nonzero principal ideals
(resp. certain nonzero principal \textit{prime} ideals) when $d=2$ (resp. $d\geq 3$). Later, we make more precise the definition of $\mathcal{S}$. Then we want to define the \textit{subset} $\mathcal{S}_{\rm{(i)}}$ (resp. $\mathcal{S}_{\rm{(ii)}}$, $\mathcal{S}_{\rm{(iii)}}$,
$\mathcal{S}_{\rm{(iv)}}$, $\mathcal{S}_{\rm{(v)}}$, $\mathcal{S}_{\rm{(vi)}}$) of $\mathcal{S}$
to be the set consisting of height-one ideals $\mathfrak{a} \in \mathcal{S}$ for which the condition $\rm(i)$ (resp. $\rm(ii)$, $\rm(iii)$, $\rm(iv)$, $\rm(v)$, $\rm(vi)$) does not hold, after replacing the quantities: $T$, $T^* (1)$, $R$, $P$ and $\mathbf{z}_1$ with their respective quotients
$T/\mathfrak{a}T$, $(T^*/\mathfrak{a}T^*) (1)$, $R/\mathfrak{a}$, $\overline{P}:=P \pmod{\frak{a}} \in R/\mathfrak{a}$ and $\overline{\mathbf{z}}_1$ which is the image of $\mathbf{z}_1$ in $H^1(G_{\Sigma},(T^*/\mathfrak{a}T^*)(1))$. We put
$$
\mathcal{S}_{\rm{bad}}:=\mathcal{S}_{\rm{(i)}} \cup \mathcal{S}_{\rm{(ii)}} \cup \mathcal{S}_{\rm{(iii)}} \cup \mathcal{S}_{\rm{(iv)}} \cup \mathcal{S}_{\rm{(v)}} \cup \mathcal{S}_{\rm{(vi)}}.
$$
We prove the assertion $\rm(1)$ and the formula $(\ref{inclusionchar})$ in the assertion $\rm(2)$ simultaneously. Let us point out that $\rm(1)$ is required to prove $\rm(2)$ theoretically. However, the proof of both assertions requires that $\mathcal{S} \setminus \mathcal{S}_{\rm{bad}}$ is large enough. In fact, to prove the assertion $\rm{(1)}$, we need to check that $\mathcal{S} \setminus \mathcal{S}_{\rm{bad}}$ is not empty. To prove the assertion $\rm{(2)}$, we need to check that $\mathcal{S} \setminus \mathcal{S}_{\rm{bad}}$ is infinite and large enough to apply Theorem \ref{prop1} and Theorem \ref{prop2} in the case $d=2$ and Theorem \ref{theorem:previous} in the case $d \ge 3$, respectively. Let us proceed by induction with respect to the Krull dimension $d \geq 2$ of $(R,\frak{m},\mathbb{F})$. As the initial step for induction of the proofs of assertions $\rm(1)$ and $\rm(2)$, we start from the proof of the case $d=2$.
\\
First, we prove the assertion $\rm(1)$ when $d=2$, assuming that $\mathcal{S} \setminus \mathcal{S}_{\rm{bad}}$ is not empty. Let $\frak{a} \in \mathcal{S} \setminus \mathcal{S}_{\rm{bad}}$. Then $\frak{a}$ is a nonzero principal ideal and
the $(R/\frak{a})[G_{\mathbb{Q}}]$-representation $T/\frak{a} T$ satisfies the hypotheses \rm{(i)} through \rm{(vi)} of Theorem \ref{Euler}, which in particular implies that $\textcyr{Sh}^2_{\Sigma}((T^*/\frak{a} T^*)(1))$ is finite. By Lemma \ref{last} applied to the map $\textcyr{Sh}^2_{\Sigma}(T^*(1))/\frak{a} \textcyr{Sh}^2_{\Sigma}(T^*(1)) \to \textcyr{Sh}^2_{\Sigma}((T^*/\frak{a} T^*)(1))$, and by Nakayama's lemma, it follows by the finiteness of $\textcyr{Sh}^2_{\Sigma}((T^*/\frak{a} T^*)(1))$ due to
Theorem \ref{Euler}, that $\textcyr{Sh}^2_{\Sigma}(T^*(1))$ is a finitely generated torsion $R$-module.
Next, we prove the assertion $(\ref{inclusionchar})$ when $d=2$. We define
$\mathcal{S}$ to be the set of height-one ideals $\{(\mathbf{x}_n) \subseteq R\}_{n \in \mathbb{N}}$, where we take $\mathbf{x}_n=z+a_np^n$ as in Lemma \ref{topological} $\rm(1)$ and $(\ref{linearideal})$ for the choice $z \in R$, and where we take $\mathbf{x}_n=z^n+r+a_n p$ as in Lemma \ref{topological} $\rm(2)$ and $(\ref{linearideal2})$ for the choice $z,r \in R$.
We set
\begin{align*}
& M:=\textcyr{Sh}^2_{\Sigma} (T^*(1)), \\
& N:=R/(P^k) \oplus H^1 (G_{\Sigma},T^*(1))/R \mathbf{z}_1.
\end{align*}
The inclusion \eqref{inclusionchar} holds if and only if
\begin{equation}\label{inclusionchar_p}
\operatorname{char}_R (N)_\frak{p} \subseteq \operatorname{char}_R (M)_\frak{p}~\mbox{holds for every height one prime}~\frak{p}~\mbox{of}~R.
\end{equation}
The proof of \eqref{inclusionchar_p} is divided into the case when $\frak{p}$ is not lying over $p$ and the case when $\frak{p}$ is lying over $p$.
Theorem \ref{prop1} plays a role in the former case and Theorem \ref{prop2} plays a role in the latter case. Since the proof goes in the same way in both cases, we shall concentrate on the former case.
Since we are assuming $d=2$, the maximal pseudo-null submodule of $H^2(G_{\Sigma},T^*(1))$ is finite. Let the \textit{error term} $c$ equal the cardinality of the maximal pseudo-null submodule of $H^2(G_{\Sigma},T^*(1))$. By applying the Euler system bound of Theorem \ref{Euler} for the $(R/\mathfrak{a})[G_{\mathbb{Q}}]$-module $T/\mathfrak{a}T$, we find that
\begin{equation}
\label{reduce1}
c \cdot \frac{|N/\mathfrak{a}N|}{|M/\mathfrak{a}M|} \in \mathbb{N}
\end{equation}
for all $\mathfrak{a} \in \mathcal{S} \setminus \mathcal{S}_{\rm{bad}}$. Once we prove that the set $\mathcal{S}_{\rm{bad}}$ is finite, Theorem \ref{prop1} applies to deduce \eqref{inclusionchar_p} for every height-one prime $\frak{p}$ of $R$ which is not lying over $p$. Similarly, the finiteness of $\mathcal{S}_{\rm{bad}}$ and Theorem \ref{prop2} allow us to deduce \eqref{inclusionchar_p} for every height-one prime $\frak{p}$ of $R$ which is lying over $p$.
Hence the proof of \eqref{inclusionchar} will be completed when $d=2$. Thus, in order to prove the assertion $\rm(2)$ for the case $d=2$, it remains to show that $\mathcal{S}_{\rm{bad}}$ is a finite set.
First, we note that if the condition $\rm(i)$ (resp. $\rm(v)$) holds for $R[G_{\mathbb{Q}}]$-module $T$,
the condition $\rm(i)$ (resp. $\rm(v)$) holds for the $(R/\mathfrak{a})[G_{\mathbb{Q}}]$-module $T/\mathfrak{a}T$ and for all $\mathfrak{a} \in \mathcal{S}$. So we get $\mathcal{S}_{\rm{(i)}}=\mathcal{S}_{\rm{(v)}}=\emptyset$. It remains to check that $\mathcal{S}_{\rm{(ii)}}$, $\mathcal{S}_{\rm{(iii)}}$, $\mathcal{S}_{\rm{(iv)}}$, $\mathcal{S}_{\rm{(vi)}}$ are finite.
$\mathbf{Finiteness~of}$ $\mathcal{S}_{\rm{(ii)}}$: We need to find an ideal $\mathfrak{a} \in \mathcal{S}$ such that
\begin{equation}
\label{finitegroup}
\big|H^1(G_{\Sigma},(T^*/\frak{a} T^*)(1))/(R/\mathfrak{a}) \mathbf{z}_1\big|< \infty.
\end{equation}
Let us put $(f)=\mathfrak{a}$ and write $R \xrightarrow{\times f} R$ for the multiplication map. Then the short exact sequence $0 \to R \xrightarrow{\times f} R \to R/\mathfrak{a} \to 0$ yields the long exact sequence:
$$
\cdots \to H^1(G_{\Sigma},T^*(1)) \xrightarrow{\times f} H^1(G_{\Sigma},T^*(1)) \to H^1(G_{\Sigma},(T^*/\frak{a} T^*)(1)) \to H^2(G_{\Sigma},T^*(1)) \xrightarrow{\times f} \cdots
$$
For simplicity, we put
$$
N_i:=H^i(G_{\Sigma},T^*(1))~\mbox{and}~J_i:=\operatorname{Ann}_R(N_i/R \mathbf{z}_1),
$$
which is the annihilator of the $R$-module $N_i/R \mathbf{z}_1$.
Let us consider $N_1$. By hypothesis $\rm(ii)$, $\operatorname{Supp}_R(N_1/R \mathbf{z}_1)=\operatorname{Spec} (R/J_1)$ is a proper closed subset of $\operatorname{Spec} R$. First, we suppose that $\operatorname{Spec}(R/J_1) \subseteq \{\frak{m}\}$. Then we choose any nonzero element $f \in \frak{m}$. Second, we suppose that $\operatorname{Spec}(R/J_1)=\{\frak{p}_1,\ldots,\frak{p}_n,\frak{m}\}$. Then we choose $f \in \frak{m}$ such that $f \notin \bigcup_{i=1}^n \frak{p}_i$.
Let us consider $N_2$, in which the image of $\mathbf{z}_1$ is zero and we have $\operatorname{Supp}_R(N_2)=\operatorname{Spec}(R/J_2)$. It suffices to choose $f \in \frak{m}$ such that $\ker\big[ N_2 \xrightarrow{\times f} N_2\big] $ is finite. First, we suppose that $\operatorname{Spec}(R/J_2) \subseteq \{\frak{m}\}$. Then we choose any nonzero element $f \in \frak{m}$. Second, we suppose that $\operatorname{Spec}(R/J_2)=\{\frak{q}_1,\ldots,\frak{q}_{n'},\frak{m}\}$. Then we choose $f \in \frak{m}$ such that $f \notin \bigcup_{i=1}^{n'} \frak{q}_i$. Finally, suppose that $\operatorname{Spec}(R/J_2)=\operatorname{Spec} R$. Since $R$ is assumed to be an integral domain, this implies that $J_2=(0)$. That is, $N_2$ is a torsion-free $R$-module. Then we choose any nonzero element $f \in \frak{m}$. In all cases considered above, $\ker\big[ N_2 \xrightarrow{\times f} N_2\big] $ is finite. Let $\frak{a} \in \mathcal{S}$. Then
$$
\frak{a} \in \mathcal{S}_{\rm{(ii)}} \iff \frak{a} \subseteq \bigcup_{i=1}^n \frak{p}_i \cup \bigcup_{i=1}^{n'} \frak{q}_i,
$$
which shows that $\mathcal{S}_{\rm{(ii)}}$ is finite (see Remark \ref{heightoneprime}). So if $\frak{a} \in \mathcal{S} \setminus \mathcal{S}_{\rm{(ii)}}$,
then $(\ref{finitegroup})$ holds, as required.
$\mathbf{Finiteness~of}$ $\mathcal{S}_{\rm{(iii)}}$: We prove that the finiteness of $H^2(G_{\Sigma},D[\mathfrak{a}])$ follows from Lemma \ref{last2}. Indeed, by Poitou-Tate duality and the hypothesis $|H^2(G_{\Sigma},D)| < \infty$, the group $H^2(G_{\mathbb{Q}_{\ell}},D)$ is finite for every $\ell \in \Sigma$. The local Tate duality shows that $T^*(1)^{G_{\mathbb{Q}_{\ell}}}$ is finite. Since $T^*(1)^{G_{\mathbb{Q}_{\ell}}}$ is $R$-torsion free, it is finite only when $T^*(1)^{G_{\mathbb{Q}_{\ell}}}=0$. Thus, the assumptions of Lemma \ref{last2} are satisfied and we find that $H^2(G_{\Sigma},D[\mathfrak{a}])$ is finite. Next, we prove that $H^2(G_{\mathbb{Q}_{\ell}},(T^*/\frak{a} T^*)(1))=0$ for $\ell \in \Sigma\setminus\{\infty\}$. Notice that the second local Galois cohomology group is exactly controlled. That is, we have $H^2(G_{\mathbb{Q}_{\ell}},T^*(1))/\mathfrak{a} H^2(G_{\mathbb{Q}_{\ell}},T^*(1)) \simeq H^2(G_{\mathbb{Q}_{\ell}},(T^*/\frak{a} T^*)(1))$. So the required vanishing follows from the assumption $H^2(G_{\mathbb{Q}_{\ell}},T^*(1))=0$. Hence we have $\mathcal{S}_{\rm{(iii)}}=\emptyset$.
$\mathbf{Finiteness~of}$ $\mathcal{S}_{\rm{(iv)}}$: Since $\wedge^2 \rho:G_{\mathbb{Q}} \to R^{\times}$ (resp. $\wedge^2 \rho^*(1):G_{\mathbb{Q}} \to R^{\times}$) is of infinite order, the image contains an element $u \in R^\times$ (resp. $u' \in R^\times$) of infinite order. Then by Lemma \ref{topological} $\rm(1)$, the image of $u$ (resp. $u'$) under $R^{\times} \to \big(R/\mathfrak{a} \big)^{\times}$ is of infinite order for every $n \in \mathbb{N}$ and $\mathfrak{a}=(z+a_n p^n) \in \mathcal{S}$. Hence we have $\mathcal{S}_{\rm{(iv)}}=\emptyset$.
$\mathbf{Finiteness~of}$ $\mathcal{S}_{\rm{(vi)}}$: It suffices to choose $(f)=\mathfrak{a} \subseteq R$ such that $\mathfrak{a}$ is not contained in $\frak{p}$, where $\frak{p} \in \operatorname{Min}_R(R/(P))$ and $P \in R$ is given as in $\rm(vi)$. Let
$\frak{a} \in \mathcal{S}$. Then
$$
\mathfrak{a} \in \mathcal{S}_{\rm{(vi)}} \iff \mathfrak{a} \subseteq \frak{p}~\mbox{for some}~\frak{p} \in \operatorname{Min}_R(R/(P)),
$$
which shows that $\mathcal{S}_{\rm{(vi)}}$ is finite (see Remark \ref{heightoneprime}).
Now $\mathcal{S}_{\rm{bad}}=\mathcal{S}_{\rm{(i)}} \cup \mathcal{S}_{\rm{(ii)}} \cup \mathcal{S}_{\rm{(iii)}} \cup \mathcal{S}_{\rm{(iv)}} \cup \mathcal{S}_{\rm{(v)}} \cup \mathcal{S}_{\rm{(vi)}}$ forms a finite set of height-one ideals of $R$.
As explained at the beginning of the proof, we complete the proofs of assertions $\rm(1)$ and $\rm(2)$ in the case $d=2$.
\\
Suppose that $d \ge 3$ and that the theorem has been proved in the case that $\dim R=d-1$. Then we establish the theorem in the case $\dim R=d$ by using an induction hypothesis that $(\ref{inclusionchar})$ holds in the case that $\dim R=d-1$, in which the local Bertini theorem (Theorem \ref{theorem:previous}) plays a crucial role. First of all, we set
\begin{align*}
& M:=\textcyr{Sh}^2_{\Sigma} (T^*(1)), \\
& N:=R/(P^k) \oplus H^1 (G_{\Sigma},T^*(1))/R \mathbf{z}_1.
\end{align*}
Then we want to prove that $\operatorname{char}_R(N) \subseteq \operatorname{char}_R (M)$. To this aim, we consider scalar extensions $M_{W(\overline{\mathbb{F}})}$ and $M_{W(\overline{\mathbb{F}})}$ as
$R_{W(\overline{\mathbb{F}})}:=R \widehat{\otimes}_{W(\mathbb{F})} W(\overline{\mathbb{F}})$-modules. Notice that $R_{W(\overline{\mathbb{F}})}$ is a complete local Cohen-Macaulay normal domain with residue field $\overline{\mathbb{F}}$ in view of Lemma \ref{localring}. With the notation as in Theorem \ref{theorem:previous}, we set
$$
\widetilde{\mathcal{S}}:=\{(\mathbf{x}_{\widetilde{a}}) \subseteq R_{W(\overline{\mathbb{F}})}~|~a \in \theta_{W(\overline{\mathbb{F}})}(U_{M,N})\}.
$$
Let us take an element $\frak{p}:=(\mathbf{x}_{\widetilde{a}}) \in \widetilde{\mathcal{S}}$. Then there exists a finite extension of fields $\mathbb{F} \to \mathbb{F}'$ such that $\frak{p} \subseteq R_{W(\mathbb{F}')}$. By Theorem \ref{theorem:previous}, we want to prove that
\begin{equation}
\label{charcheck}
\operatorname{char}_{R_{W(\mathbb{F}')}/\frak{p}}(N_{W(\mathbb{F}')}/\frak{p} N_{W(\mathbb{F}')}) \subseteq \operatorname{char}_{R_{W(\mathbb{F}')}/\frak{p}}(M_{W(\mathbb{F}')}/\frak{p} M_{W(\mathbb{F}')}).
\end{equation}
It is easy to check that the conditions $\rm(i),\rm(ii),\rm(iii),\rm(iv),\rm(v)$ and $\rm(vi)$ are not effected by base change $T \to T_{W(\mathbb{F}')}$. After checking that the conditions $\rm(i),\rm(ii),\rm(iii),\rm(iv),\rm(v)$ and $\rm(vi)$ hold for the $R_{W(\mathbb{F}')}/\frak{p}[G_{\mathbb{Q}}]$-module $T_{W(\mathbb{F}')}/\frak{p} T_{W(\mathbb{F}')}$, we obtain the following by induction hypothesis:
\begin{equation}
\label{charcheck2}
(\overline{P}^k) \operatorname{char}_{R_{W(\mathbb{F}')}/\frak{p}}\big(H^1 (G_{\Sigma},(T^*_{W(\mathbb{F}')}/\frak{p} T^*_{W(\mathbb{F}')})(1))/(R_{W(\mathbb{F}')}/\frak{p}) \overline{\mathbf{z}}_1\big)
\end{equation}
$$
\subseteq \operatorname{char}_{R_{W(\mathbb{F}')}/\frak{p}}\big(\textcyr{Sh}^2_{\Sigma}((T^*_{W(\mathbb{F}')}/\frak{p} T^*_{W(\mathbb{F}')})(1))\big).
$$
By Lemma \ref{last} together with $(\ref{controlSha})$, we find that
\begin{equation}
\label{charcheck3}
\operatorname{char}_{R_{W(\mathbb{F}')}/\frak{p}}(M_{W(\mathbb{F}')}/\frak{p} M_{W(\mathbb{F}')})=\operatorname{char}_{R_{W(\mathbb{F}')}/\frak{p}}\big(\textcyr{Sh}^2_{\Sigma}((T^*_{W(\mathbb{F}')}/\frak{p} T^*_{W(\mathbb{F}')})(1))\big).
\end{equation}
Next, let us consider the long exact sequence:
\begin{multline*}
\cdots \to H^1 (G_{\Sigma},T^*_{W(\mathbb{F}')}(1)) \xrightarrow{\times \mathbf{x}_{\widetilde{a}}} H^1 (G_{\Sigma},T^*_{W(\mathbb{F}')}(1)) \to H^1 (G_{\Sigma},(T^*_{W(\mathbb{F}')}/\frak{p} T^*_{W(\mathbb{F}')})(1))
\\
\to H^2(G_{\Sigma},T^*_{W(\mathbb{F}')}(1)) \xrightarrow{\times \mathbf{x}_{\widetilde{a}}} \cdots
\end{multline*}
which is induced by the exact sequence $0 \to T^*_{W(\mathbb{F}')} \xrightarrow{\times \mathbf{x}_{\widetilde{a}}} T^*_{W(\mathbb{F}')} \to T^*_{W(\mathbb{F}')}/\frak{p} T^*_{W(\mathbb{F}')} \to 0$. The above long exact sequence yields the following short exact sequence:
\begin{multline}\label{equation:shortsequence1}
0 \to H^1 (G_{\Sigma},T^*_{W(\mathbb{F}')}(1))/\frak{p} H^1 (G_{\Sigma},T^*_{W(\mathbb{F}')}(1)) \to H^1 (G_{\Sigma},(T^*_{W(\mathbb{F}')}/\frak{p} T^*_{W(\mathbb{F}')})(1))
\\
\to H^2(G_{\Sigma},T^*_{W(\mathbb{F}')}(1))[\mathbf{x}_{\widetilde{a}}] \to 0.
\end{multline}
Let us choose $\mathbf{x}_{\widetilde{a}}$ such that
$$
\mathbf{x}_{\widetilde{a}} \notin \bigcup \frak{q},
$$
where $\frak{q}$ ranges over all associated prime ideals of the $R_{W(\mathbb{F}')}$-module $H^2(G_{\Sigma},T^*_{W(\mathbb{F}')}(1))$ which are of height
at most two. Notice that the set of such $\mathbf{x}_{\widetilde{a}}$ is non-empty by the assumption $d \ge 3$.
For such $\mathbf{x}_{\widetilde{a}}$, we have the following equality by combining \eqref{controlGal} and \eqref{equation:shortsequence1}:
\begin{multline}
\label{charcheck4}
\operatorname{char}_{R_{W(\mathbb{F}')}/\frak{p}}(N_{W(\mathbb{F}')}/\frak{p} N_{W(\mathbb{F}')})
\\
=(\overline{P}^k) \operatorname{char}_{R_{W(\mathbb{F}')}/\frak{p}}\big(H^1 (G_{\Sigma},(T^*_{W(\mathbb{F}')}/\frak{p} T^*_{W(\mathbb{F}')})(1))/(R_{W(\mathbb{F}')}/\frak{p}) \overline{\mathbf{z}}_1\big).
\end{multline}
Thus, in view of $(\ref{charcheck2})$, $(\ref{charcheck3})$ and $(\ref{charcheck4})$, it suffices to define the subset $\mathcal{S}$ of $\widetilde{\mathcal{S}}$ as follows:
$$
\mathcal{S}:=\big\{(\mathbf{x}_{\widetilde{a}})~\big|~\mathbf{x}_{\widetilde{a}} \notin \frak{q}~\mbox{for all}~\frak{q} \in \operatorname{Ass}_{R_{W(\overline{\mathbb{F}})}}\big(H^2(G_{\Sigma},T^*_{W(\overline{\mathbb{F}})}(1))\big) \setminus\{\frak{m}_{R_{W(\overline{\mathbb{F}})}}\}\big\} \cap \widetilde{\mathcal{S}}
$$
to prove $(\ref{charcheck})$. Let us define $U_{\frak{q}} \subseteq \mathbb{P}^n(\overline{\mathbb{F}})$ to be the Zariski-dense open subset attached to each $\frak{q} \in \operatorname{Ass}_{R_{W(\overline{\mathbb{F}})}}\big(H^2(G_{\Sigma},T^*_{W(\overline{\mathbb{F}})}(1))\big) \setminus\{\frak{m}_{R_{W(\overline{\mathbb{F}})}}\}$ via Lemma \ref{Case1}. Now we apply Theorem \ref{theorem:previous} to the Zariski open subset:
$$
U:=\big(\bigcap U_{\frak{q}}\big) \cap U_{M,N} \subseteq \mathbb{P}^n(\overline{\mathbb{F}})
$$
for $\frak{q} \in \operatorname{Ass}_{R_{W(\overline{\mathbb{F}})}}\big(H^2(G_{\Sigma},T^*_{W(\overline{\mathbb{F}})}(1))\big) \setminus\{\frak{m}_{R_{W(\overline{\mathbb{F}})}}\}$.
As the proof of the assertion $\rm(1)$ is similar to the case $d=2$, we skip the details. So let us prove the assertion $\rm(2)$. Write $R$ for $R_{W(\mathbb{F}')}$ to simplify the notation for a finite field extension $\mathbb{F} \to \mathbb{F}'$. As we did in the case $d=2$, we need to determine the subsets $\mathcal{S}_{\rm{(i)}}, \mathcal{S}_{\rm{(ii)}}, \mathcal{S}_{\rm{(iii)}}, \mathcal{S}_{\rm{(iv)}}, \mathcal{S}_{\rm{(v)}}$ and $\mathcal{S}_{\rm{(vi)}}$ of $\mathcal{S}$, respectively. Notice that the conditions $\rm(i)$ and $\rm(v)$ in the theorem hold for the $R/\frak{p}[G_{\mathbb{Q}}]$-module $T/\frak{p} T$ for any $\frak{p} \in \mathcal{S}$ and we have $\mathcal{S}_{\rm{(i)}}=\mathcal{S}_{\rm{(v)}}=\emptyset$. Moreover, the conditions $\rm(iii)$ and $\rm(vi)$ are treated in a similar manner to the case $d=2$ and we find that $\mathcal{S}_{\rm{(iii)}}=\emptyset$
and $\mathcal{S}_{\rm{(vi)}}$ is finite. So it remains to discuss the conditions $\rm(ii)$ and $\rm(iv)$ for the $R/\frak{p}[G_{\mathbb{Q}}]$-module $T/\frak{p} T$.
$\mathbf{Finiteness~of}$ $\mathcal{S}_{\rm{(ii)}}$: Write $R \xrightarrow{\times \mathbf{x}_{\widetilde{a}}} R$ for the multiplication map with $(\mathbf{x}_{\widetilde{a}})=\frak{p}$. Then the short exact sequence $0 \to R \xrightarrow{\times \mathbf{x}_{\widetilde{a}}} R \to R/\frak{p} \to 0$ yields the long exact sequence:
$$
\cdots \to H^1(G_{\Sigma},T^*(1)) \xrightarrow{\times \mathbf{x}_{\widetilde{a}}} H^1(G_{\Sigma},T^*(1)) \to H^1(G_{\Sigma},(T^*/\frak{p} T^*)(1)) \to H^2(G_{\Sigma},T^*(1)) \xrightarrow{\times \mathbf{x}_{\widetilde{a}}} \cdots .
$$
Now we have to find $\frak{p}=(\mathbf{x}_{\widetilde{a}}) \subseteq R$ such that
\begin{equation}
\label{torsionmodule}
H^1(G_{\Sigma},(T^*/\frak{p} T^*)(1))/(R/\frak{p}) \mathbf{z}_1~\mbox{is a torsion}~R/\frak{p} \mbox{-module}.
\end{equation}
As the proof of $(\ref{torsionmodule})$ is done similarly to the case $d=2$, we skip the details.
$\mathbf{Finiteness~of}$ $\mathcal{S}_{\rm{(iv)}}$: By hypothesis, there exists an element $u \in R^{\times}$ of infinite order, which is in the image of the determinant character $\wedge^2 \rho: G_{\mathbb{Q}} \to R^{\times}$. Since the ring of Witt vectors $W(\mathbb{F})$ is a coefficient ring of $R$, we have $R=W(\mathbb{F})+\frak{m}$ and $u=a+t$ for some $a \in W(\mathbb{F})^{\times}$ and $t \in \frak{m}$. We shall make a choice of a principal prime ideal $\frak{p}$ such that the following condition holds:
\begin{equation}
\label{character}
\frak{p}~\mbox{has the property that}~u^n-1 \notin \frak{p}~\mbox{for all}~n \ge 1.
\end{equation}
If $a \ne 1$, then there is nothing to prove due to $u^n-1\in R^{\times}$. So we assume that $a=1$ in what follows. Then
$$
u^n-1=(1+t)^n-1=\sum_{k=1}^n{n \choose k}t^k=t \big(n+\frac{n(n-1)}{2}t+\cdots+t^{n-1}\big).
$$
Writing $n=mp^e$ with $e \ge 0$ and $p \nmid m$, we see that $\dfrac{u^m-1}{t} \in R^{\times}$ since $t \in \frak{m}$.
Thus, there are only finitely many height-one primes in $R$ containing $u^m-1$ for varying $m$. Consider an ideal of $R$:
$$
I_{m,e}:=(p,u^{mp^e}-1).
$$
Since $u^{mp^e}-1 \equiv (u^m-1)^{p^e} \pmod{pR}$ and $\dfrac{u^m-1}{t} \in R^{\times}$, it follows that
\begin{equation}
\label{character2}
\bigcup_{m, e \ge 1}\operatorname{Min}_R(I_{m,e})~\mbox{is a finite set}.
\end{equation}
Every prime ideal $\frak{q}$ belonging to $\bigcup_{m, e \ge 1}\operatorname{Min}_R(I_{m,e})$ is of height at most 2. Moreover, since $d=\dim R \ge 3$, such a prime ideal has the property that $\frak{q} \ne \frak{m}$. Then Lemma \ref{Case1} allows us to find sufficiently many prime ideals $\frak{p} \in \operatorname{Spec} R$ so as to satisfy $(\ref{character})$. In other words, the reduced character $\wedge^2 \overline{\rho}:G_{\mathbb{Q}} \to (R/\frak{p})^{\times}$ is of infinite order.
Now we get the desired $(\ref{charcheck})$ for all height-one prime ideals $\frak{p} \in \mathcal{S} \setminus \mathcal{S}_{\rm{bad}}$. Hence, $\operatorname{char}_R(N) \subseteq \operatorname{char}_R(M)$, that is, $(\ref{inclusionchar})$ has been established in the case $d=\dim R$. This completes the proof of the theorem.
\end{proof}
\section{Modular forms and Hida theory}
\label{Hidafamily}
In this section, we give a brief review on modular form, Hecke algebra and their extension in the context of Hida theory.
\subsection{$p$-optimal complex period via modular symbols}
In this section, we construct a $p$-adic $L$-function associated to a normalized $p$-ordinary Hecke eigencusp form which interpolates the special values of the complex $L$-function. In the next section, we will give the construction of two-variable $p$-adic $L$-function attached to a branch of the nearly ordinary Hecke algebra. This was first constructed by Mazur and Kitagawa \cite{Kit}. Another constructions were given by Greenberg and Stevens \cite{GS93} and Emerton, Pollack and Weston \cite{EPW}.
Let us introduce \textit{p-optimal complex period} and \textit{p-adic period} in order to have the algebraicity of complex $L$-functions. As the complex period of Deligne type \cite{De79} is determined only up to multiple of $\overline{\mathbb{Q}}^{\times}$, it is necessary to introduce the $p$-optimal complex period to formulate the interpolation formula for the Hida deformation of a $p$-adic $L$-function without ambiguities. Fix a prime number $p>0$ and embeddings
$i_{\infty}:\overline{\mathbb{Q}} \hookrightarrow \mathbb{C}$ and $i_p:\overline{\mathbb{Q}} \hookrightarrow \overline{\mathbb{Q}}_p$.
\begin{definition}
Let $k \ge 1$ be an integer and let $p>0$ be a prime number. Fix an embedding $i_p:\overline{\mathbb{Q}} \hookrightarrow \overline{\mathbb{Q}}_p$.
\begin{enumerate}
\item
Let $M>0$ be an integer such that $p \mid M$. Then we say that a cusp form $f \in S_k(\Gamma_1(M);\overline{\mathbb{Q}})$ is \textit{p-stablized}, if it is a normalized Hecke eigencusp form.
\item
Let $N>0$ is an integer such that $p \nmid N$. Let $f \in S_k(\Gamma_1(N);\overline{\mathbb{Q}})$ be a normalized Hecke eigencusp form and let $\alpha,\beta \in \overline{\mathbb{Q}}_p$ be the roots of the equation $X^2-a_p(f)X+p^{k-1}\psi_f(p)=0$, where $\psi_f$ is the neben character of $f$. Let
$$
f_{\alpha}:=f(q)-\beta f(q^p),~f_{\beta}:=f(q)-\alpha f(q^p).
$$
which are shown to be cusp forms in the space $S_k(\Gamma_1(Np),\overline{\mathbb{Q}})$. We call $f_{\alpha}$ and $f_{\beta}$ the \textit{p-stabilization} of $f$.
\item
We say that $ f \in S_k(\Gamma_1(M),\overline{\mathbb{Q}})$ is \textit{p-ordinary}, if $i_p(a_p(f)) \in \overline{\mathbb{Q}}_p$ is a $p$-adic unit via $i_p:\overline{\mathbb{Q}} \hookrightarrow \overline{\mathbb{Q}}_p$. A normalized $p$-ordinary Hecke eigencusp form $f \in S_k(\Gamma_1(M),\overline{\mathbb{Q}})$ is called a \textit{p-stabilized newform}, if it comes from the $p$-stabilization of a certain primitive form.
\end{enumerate}
\end{definition}
\begin{remark}
\begin{enumerate}
\item
Let $f \in S_k(\Gamma_1(M),\overline{\mathbb{Q}})$ be a normalized $p$-stabilized Hecke eigencusp form Then it is proved that $f \in S_k(\Gamma_1(Mp),\overline{\mathbb{Q}})$ is $p$-stabilized.
\item
It is necessary to raise the power of $p$ in the ``level'' for constructing Hida deformations and it is essential to consider $p$-ordinary $p$-stabilized newforms.
\end{enumerate}
\end{remark}
Let $f \in S_k(\Gamma_1(M),\overline{\mathbb{Q}})$ be a $p$-ordinary $p$-stabilized Hecke eigencusp form of weight $k \ge 2$; in particular, we have $p|M$. Denote by $\mathbb{Q}_f:=\mathbb{Q}(\{a_n(f)\})$ the \textit{Hecke field} associated to $f$. Set $\mathbb{Z}_{f,(p)}:=i^{-1}_p(\overline{\mathbb{Z}}_p)$. This is a discrete valuation ring with the field of fractions $\mathbb{Q}_f$. For an integer $k \ge 2$ and a ring $A$, we define an $A$-module
$$
L_k(A):=\bigoplus_{0 \le i \le k-2} A \cdot X^i Y^{k-2-i},
$$
where $A \cdot X^iY^{k-2-i}$ denotes the $A$-submodule of the polynomial ring $A[X,Y]$ generated by $X^iY^{k-2-i}$. We let the matrix
$$
g=
\begin{pmatrix}
a & b \\ c & d
\end{pmatrix}
\in M_2(\mathbb{Z}) \cap GL_2(\mathbb{Q})
$$
act on $L_k(A)$ by the formula:
$$
g \cdot p(X,Y)=p\Big((X,Y)\begin{pmatrix} a & c \\ b & d \end{pmatrix}\Big).
$$
for $p(X,Y) \in L_k(A)$.
Let $Y_1(M)_{\mathbb{C}}:=\Gamma_1(M) \setminus \mathfrak{H}$ be the complex affine modular curve. We let $\Gamma_1(M)$ act on the product $\mathfrak{H} \times L_k(A)$ diagonally. We define $\mathcal{L}_k(M)$ as a sheaf of continuous sections of the topological covering map:
$$
\Gamma_1(M) \setminus \mathfrak{H} \times L_k(A) \twoheadrightarrow \Gamma_1(M) \setminus \mathfrak{H}=Y_1(M)_{\mathbb{C}}.
$$
Let $H^1_{\mathrm{Betti}}(Y_1(M)_{\mathbb{C}},\mathcal{L}_k(A))$ be the sheaf cohomology using complex topology on $Y_1(M)_{\mathbb{C}}$.
Let $\mathcal{O}$ be the ring of $p$-adic integers with $\mathcal{K}$ its field of fractions. Then one can define the etale sheaf $\mathcal{L}_k(A)_{\operatorname{et}}$ on $Y_1(M)_{\overline{\mathbb{Q}}}$ when $A=\mathcal{O}$ or $\mathcal{K}$ and the etale cohomology $H^1_{\operatorname{et}}(Y_1(M)_{\overline{\mathbb{Q}}},\mathcal{L}_k(A)_{\operatorname{et}})$. We remark that both Betti and etale cohomologies groups are equipped with Hecke actions via Hecke correspondences. We take $A=\mathcal{O}$ or $\mathcal{K}$. By the comparison theorem between etale and Betti cohomologies, one obtains an isomorphism:
\begin{equation}
\label{Betti}
H^1_{\operatorname{Betti}}(Y_1(M)_{\mathbb{C}},\mathcal{L}_k(A)) \simeq H^1_{\operatorname{et}}(Y_1(M)_{\overline{\mathbb{Q}}},\mathcal{L}_k(A)_{\operatorname{et}})~\mbox{via}~i_{\infty}:\overline{\mathbb{Q}} \hookrightarrow \mathbb{C}.
\end{equation}
Moreover, the isomorphism $(\ref{Betti})$ preserves Hecke actions on both sides. Deligne constructed the Galois representation attached to a cusp form by using etale cohomology, which will be discussed later.
\begin{definition}
We define \textit{parabolic cohomology} by
$$
H^1_{\Bettip}(Y_1(M)_{\mathbb{C}},\mathcal{L}_k(A)):=\operatorname{Im} \Big(H^1_{\operatorname{Betti},c}(Y_1(M)_{\mathbb{C}},\mathcal{L}_k(A)) \to H^1_{\operatorname{Betti}}(Y_1(M)_{\mathbb{C}},\mathcal{L}_k(A))\Big),
$$
where $H^1_{\operatorname{Betti},c}(Y_1(M)_{\mathbb{C}},\mathcal{L}_k(A))$ denotes the compactly supported cohomology.
\end{definition}
The importance of parabolic cohomology is expressed by the following theorem (see \cite[Theorem 1 in \S~6.2]{Hid2} or \cite[Theorem 8.4]{Sh71} for the proof).
\begin{theorem}[Eichler-Shimura isomorphism]
\label{ES}
Fix integers $k \ge 2$ and $M\ge 1$. Then there is a natural isomorphism of $\mathbb{C}$-vector spaces:
$$
\ES:S_k(\Gamma_1(M)) \oplus \overline{S_k(\Gamma_1(M))} \to H^1_{\Bettip}(Y_1(M)_{\mathbb{C}},\mathcal{L}_k(\mathbb{C})).
$$
Moreover, the map $\ES$ is compatible with the action of Hecke operators. Here, $\overline{S_k(\Gamma_1(M))}$ is the $\mathbb{C}$-vector space spanned by the complex conjugates $\overline{f(z)}$ for all $f(z) \in S_k(\Gamma_1(M))$.
\end{theorem}
Let $\mathcal{D}$ be the free abelian group generated by the set of cusps $\mathbb{P}^1(\mathbb{Q})$. Let $\mathcal{D}_0$ denote the subgroup of $D$ generated by the elements of the form $\{\alpha\}-\{\beta\}$ for $\{\alpha\}, \{\beta\} \in \mathbb{P}^1(\mathbb{Q})$. The group $\Gamma_1(M)$ acts on $\mathfrak{H} \cup \mathbb{P}^1(\mathbb{Q})$ via linear fractional transformations. Via this action, we may regard $\mathcal{D}$ and $\mathcal{D}_0$ as being equipped with $\Gamma_1(M)$-action.
\begin{proposition}
\label{modsy}
Let $A$ be any commutative ring. Then there is an isomorphism of $A$-modules:
$$
\operatorname{Hom}_{\Gamma_1(M)}(\mathcal{D}_0;L_k(A)) \simeq H^1_{\operatorname{Betti},c}(Y_1(M)_{\mathbb{C}},\mathcal{L}(A)).
$$
This isomorphism is compatible with the Hecke action on both sides.
\end{proposition}
Let $A$ be a ring such that $\frac{1}{2} \in A$. As the $2 \times 2$-matrix $\begin{pmatrix} 1 & 0 \\ 0 & -1 \end{pmatrix} \in \Gamma_1(M)$ induces an involution on any $A[\Gamma_1(M)]$-module, we have the decomposition:
$$
\operatorname{Hom}_{\Gamma_1(M)}(\mathcal{D}_0;L_k(A))=\operatorname{Hom}_{\Gamma_1(M)}(\mathcal{D}_0;L_k(A))^+ \oplus \operatorname{Hom}_{\Gamma_1(M)}(\mathcal{D}_0;L_k(A))^-
$$
\begin{definition}[Modular symbols]
Let $A$ be a commutative ring. Then the $A$-module
$$
\operatorname{Hom}_{\Gamma_1(M)}(\mathcal{D}_0;L_k(A))
$$
is called the \textit{space of modular symbols of weight k} with values in $A$. Let $f=\sum_{i=1}^{\infty} a_n(f) q^n \in S_k(\Gamma_1(M),\overline{\mathbb{Q}})$ be a normalized Hecke eigencusp form of weight $k \ge 2$ and assume that $A$ is a $\mathbb{Z}_{f,(p)}[\frac{1}{2}]$-algebra. We define
$$
\operatorname{Hom}_{\Gamma_1(M)}(\mathcal{D}_0;L_k(A))[f]^{\pm}:=\Big\{h \in \operatorname{Hom}_{\Gamma_1(M)}(\mathcal{D}_0,L_k(A))^{\pm}~\Big|~h|T(n)=a_n(f)h~\mbox{for all}~n \ge 0\Big\},
$$
where $T(n)$ is the Hecke operator. Then
$$
MS_f(A):=\operatorname{Hom}_{\Gamma_1(M)}(\mathcal{D}_0,L_k(A))[f]
$$
is called the \textit{space of modular symbols attached to f}. One can take the $\pm$-decomposition $MS_f^{\pm}(A).$
\end{definition}
\begin{example}
\label{specialcycle}
We give a special cocycle in the space of modular symbols which is associated to a cusp form as follows. Let $f \in S_k(\Gamma_1(M),\overline{\mathbb{Q}})$ be a cusp form of weight $k \ge 2$. We define a map $\eta_f:\mathcal{D}_0 \to L_k(\mathbb{C})$ by
$$
\eta_f(\{\alpha\}-\{\beta\})=\int_{\beta}^{\alpha} f(z)(zX+Y)^{k-2}dz.
$$
This is called the \textit{modular symbol associated to f}. It is easy to verify that $\eta_f \in MS_f(\mathbb{C})$. Indeed, Eichler-Shimura map is related by the formula $\ES(f,0)=\eta_f$.
\end{example}
Let $f \in S_k(\Gamma_1(M),\overline{\mathbb{Q}})$ be a $p$-ordinary $p$-stabilized newform of weight $k \ge 2$. Let $\mathbb{Z}^{\wedge}_{f,(p)}$ be the completion of $\mathbb{Z}_{f,(p)}$ with respect to the $p$-adic valuation induced by $i_p:\overline{\mathbb{Q}} \hookrightarrow \overline{\mathbb{Q}}_p$. Denote by $\mathbb{Q}^{\wedge}_{f,(p)}$ the field of fractions of $\mathbb{Z}^{\wedge}_{f,(p)}$. By Deligne's theorem, we can attach the $G_{\mathbb{Q}}$-representation $V_f:=H^1_{\operatorname{et},c}(Y_1(M)_{\overline{\mathbb{Q}}},\mathcal{L}_k(\mathbb{Q}^{\wedge}_{f,(p)})_{\operatorname{et}})[f]$. As is well-known, $V_f$ is a $\mathbb{Q}^{\wedge}_{f,(p)}$-vector space of dimension two. Then one can check that $MS_f^+(\mathbb{Z}_{f,(p)})$ and
$MS_f^-(\mathbb{Z}_{f,(p)})$ are free $\mathbb{Z}_{f,(p)}$-modules of rank one, as each spans the sub $\mathbb{Q}_{f,(p)}^{\wedge}$-vector space of dimension one inside $V_f$ under the composition:
$$
MS_f(\mathbb{Z}_{f,(p)}) \to
H^1_{\operatorname{Betti},c}(Y_1(M)_{\mathbb{C}},\mathcal{L}_k(\mathbb{Q}_{f,(p)}^{\wedge}))[f] \simeq V_f=H^1_{\operatorname{et},c}(Y_1(M)_{\overline{\mathbb{Q}}},\mathcal{L}_k(\mathbb{Q}_{f,(p)}^{\wedge}))[f],
$$
where the first map is due to Proposition \ref{modsy} and the second isomorphism is due to the compactly supported version of $(\ref{Betti})$.
\begin{definition}[$p$-optimal complex period]
Fix a $\mathbb{Z}_{f,(p)}$-basis $b^+_{T_f}$ of the module $MS_f^+(\mathbb{Z}_{f,(p)})$ (resp. $\mathbb{Z}_{f,(p)}$-basis $b^-_{T_f}$ of the module $MS_f^-(\mathbb{Z}_{f,(p)})$). Via an isomorphism $MS_f^{\pm}(\mathbb{Z}_{f,(p)}) \otimes_{\mathbb{Z}_{f,(p)}} \mathbb{C} \simeq MS_f^{\pm}(\mathbb{C})$, one may regard $\{b^{\pm}_{T_f}\}$ as a $\mathbb{C}$-basis of $MS_f^{\pm}(\mathbb{C})$. Let $\eta_f=\eta^+_f+\eta^-_f$ be the sum in the decomposition
$MS_f(\mathbb{C})=MS_f^+(\mathbb{C}) \oplus MS_f^-(\mathbb{C})$. We define a \textit{p-optimal complex period} $\Omega_{\infty}(f,b^{\pm}_{T_f}) \in \mathbb{C}^{\times}$ to be a constant satisfying the equality
$$
\eta^{\pm}_f=\Omega_{\infty}(f,b^{\pm}_{T_f}) \cdot b^{\pm}_{T_f},
$$
where $\eta^{\pm}_f$ is the modular symbol associated to $f$ as in Example \ref{specialcycle}.
\\
\end{definition}
\subsection{Nearly ordinary Hida deformation and Galois representation}
We fix notation used in Hida theory (see \cite{Hid0} and \cite{Hid1} for details). For a fixed prime number $p>2$ (we will assume this for the moment), let $\mathbb{Q}_{\infty}/\mathbb{Q}$ be the unique cyclotomic $\mathbb{Z}_p$-extension and let $C_{\infty}$ be the Galois group of $\mathbb{Q}_{\infty}/\mathbb{Q}$. Then there is a canonical character $\chi_{\operatorname{cyc}}:C_{\infty} \simeq 1+p\mathbb{Z}_p \subseteq \mathbb{Z}_p^{\times}$ (the $p$-adic cyclotomic character). Let $N>0$ be an integer such that $p \nmid N$. Let $Y_1(Np^t)_{\mathbb{C}}:=\Gamma_1(Np^t) \setminus \mathfrak{H}$. Then, one can consider an algebraic curve $Y_1(Np^t)$ as defined over $\mathbb{Q}$ associated to $Y_1(Np^t)_{\mathbb{C}}$. Then for
$d \in (\mathbb{Z}/Np^t\mathbb{Z})^{\times}$ with $d \equiv 1 \pmod p$, the diamond operator $\langle d \rangle$ maps $[E,P]$ to $[E,dP]$. The group of diamond operators forms an inverse system compatible with the inverse system $\{Y_1(Np^t)\}_{t \ge 1}$. We denote the inverse limit by $D_{\infty}$. Then there is a canonical character $\eta:D_{\infty} \simeq 1+p\mathbb{Z}_p \subseteq \mathbb{Z}_p^{\times}$.
Fix an integer $k \ge 2$ and let $\mathcal{O}$ be the ring of integers of a finite extension of $\mathbb{Q}_p$. Let $h_k(\Gamma_{1}(Np^r);\mathcal{O})$ be the Hecke algebra acting on $S_k(\Gamma_1(Np^r);\mathcal{O})$. We define \textit{ordinary Hecke algebra} $h^{\operatorname{ord}}_k(\Gamma_{1}(Np^r);\mathcal{O})$ as the maximal algebra direct summand of $h_k(\Gamma_{1}(Np^r);\mathcal{O})$ such that $h^{\operatorname{ord}}_k(\Gamma_{1}(Np^r);\mathcal{O})$ acts on the space of ordinary forms: $S_k^{\operatorname{ord}}(\Gamma_1(Np^r);\mathcal{O}):=e\big(S^{\operatorname{ord}}_k(\Gamma_1(Np^r);\mathcal{O})\big) \subseteq S_k(\Gamma_1(Np^r);\mathcal{O})$,
where $e$ is Hida's idempotent operator.
We define the \textit{universal ordinary (Hida's) Hecke algebra} with tame level $N$:
$$
\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}}:=\varprojlim_r h^{\operatorname{ord}}_k(\Gamma_{1}(Np^r);\mathcal{O})
$$
which acts on the $\Lambda$-module:
$$
S_k^{\operatorname{ord}}(\Gamma_1(Np^{\infty});\mathcal{O}):=\bigcup_{r \ge 1} S_k^{\operatorname{ord}}(\Gamma_1(Np^r);\mathcal{O}).
$$
Then we can check that $\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}}$ is regarded as an algebra over $\Lambda:=\mathbb{Z}_p[[D_{\infty}]]$ via the action of $D_{\infty}$. More explicitly, $\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}}$ is a sub-algebra of $\End_{\Lambda}\big(S_k^{\operatorname{ord}}(\Gamma_1(Np^{\infty});\mathcal{O})\big)$ generated by Hecke operators $T_{\ell}$; $\ell \nmid Np$ and $U_{\ell}$; $\ell | Np$. It is a fundamental theorem of Hida theory that the algebra $\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}}$ does not depend on the choice of $k \ge 2$ (see \cite[Theorem 1.1]{Hid1}). Moreover, $\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}}$ is commutative and a semi-local ring which is finite flat over $\Lambda$ and is characterized by the algebra isomorphism:
$$
\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}} \otimes_{\Lambda} \Lambda/(P^r_k) \simeq h^{\operatorname{ord}}_k(\Gamma_{1}(Np^r);\mathcal{O}_{r,k})
$$
for $\mathcal{O}_{r,k}:=\Lambda/(P^r_k)$ and $P^r_k:=u^{p^{r-1}}-\eta(u)^{(k-2)p^{r-1}}$, where $u \in D_{\infty}$ is the fixed topological generator. The above isomorphism is often called the \textit{control theorem for Hida deformations} and it is proved in \cite[Theorem 1.2]{Hid1}.
We say that $\kappa \in \operatorname{Hom}_{\mathbb{Z}_p}(\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}},\overline{\mathbb{Q}}_p)$ is an \textit{arithmetic character of weight $w(\kappa)$}, if there exists an open subgroup $\Gamma \subseteq D_{\infty}$ for which the restriction of $\kappa$ to the subring $\mathbb{Z}_p[[\Gamma]] \subseteq \mathbb{Z}_p[D_{\infty}]$ coincides with the ring map induced by the character $\eta^{w(\kappa)-2}$ for an integer $w(\kappa) \ge 2$. One may also define the notion of arithmetic character for any $\mathbb{Z}_p[[D_{\infty}]]$-algebra. The \textit{nearly ordinary Hecke algebra} is defined as the completed group algebra:
$$
\mathbf{H}^{\operatorname{n.ord}}_{Np^{\infty}}:=\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}} \widehat{\otimes}_{\mathbb{Z}_p} \mathbb{Z}_p[[C_{\infty}]],
$$
which is a finite flat extension of $\mathbb{Z}_p[[C_{\infty}\times D_{\infty}]]$. The completed group algebra $\mathbb{Z}_p[[C_{\infty}\times D_{\infty}]]$ is non-canonically isomorphic to the power series ring $\mathbb{Z}_p[[X,Y]]$. In fact, an isomorphism is given by a fixed pair of topological generators of $C_{\infty}$ and $D_{\infty}$, respectively. We say that the localization $(\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}})_{\frak{m}}$ for a maximal ideal $\frak{m} \subseteq \mathbf{H}^{\operatorname{ord}}_{Np^{\infty}}$ is a \textit{local component} of $\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}}$.
We let $\mathbf{I}$ denote the integral closure of $\mathbb{Z}_p[[D_{\infty}]]$ in a finite field extension of $\operatorname{Frac}(\mathbb{Z}_p[[D_{\infty}]])$. Let $\mathbf{f}=\sum_{n=1}^{\infty}a_n(\mathbf{f})q^n \in \mathbf{I}[[q]]$ be a $p$-ordinary $p$-stabilized $\mathbf{I}$-adic newform with level $Np^{\infty}$. This means that the specialization $\mathbf{f}_{\kappa}$ is a $p$-ordinary $p$-stabilized cusp newform for all but finitely many arithmetic characters $\kappa \in \operatorname{Hom}_{\mathbb{Z}_p}(\mathbf{I},\overline{\mathbb{Q}}_p)$. We can define the ring map $\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}} \to \mathbf{I}$ by sending $T_{\ell}$ to $a_{\ell}(\mathbf{f})$. Since $\mathbf{I}$ is a domain, this map factors through the quotient $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}:=\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}}/\frak{a}_{\mathbf{f}}$ for a minimal prime $\frak{a}_{\mathbf{f}} \subseteq \mathbf{H}^{\operatorname{ord}}_{Np^{\infty}}$. We call $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$ a \textit{branch} of the Hecke algebra $\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}}$. Then $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$ is a local domain which is a finite extension of $\mathbb{Z}_p[[D_{\infty}]]$ and there is a continuous Galois representation attached by Hida:
$$
\rho_{\mathbf{f}}^{\operatorname{ord}}:G_{\mathbb{Q}} \to \operatorname{Aut}_{\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}}(\mathcal{T}^{\operatorname{ord}}),
$$
unramified outside a finite set of primes $\Sigma$, which consists of $\{\infty\}$ and all prime factors of $Np$. The $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$-module $\mathcal{T}^{\operatorname{ord}}$ is defined as a reflexive module
$$
\mathcal{T}^{\operatorname{ord}}:=\big(\varprojlim_n H^1_{\operatorname{et},c}(Y_1(Np^n)_{\overline{\mathbb{Q}}};\mathbb{Z}^{\wedge}_{\mathbf{f}_{\kappa},(p)})^{\operatorname{ord}}[\mathbf{f}_{\kappa}]\big)^{\operatorname{rc}},
$$
where $\mathbf{f}_{\kappa}$ is any weight two specialization of $\mathbf{f}$ and $\varprojlim_n H^1_{\operatorname{et},c}(Y_1(Np^n)_{\overline{\mathbb{Q}}};\mathbb{Z}^{\wedge}_{\mathbf{f}_{\kappa},(p)})^{\operatorname{ord}}$ is the inverse limit of the ordinary part of compactly supported etale cohomology groups of affine modular curves:
$$
\cdots \to Y_1(Np^{n+1})_{\overline{\mathbb{Q}}} \to Y_1(Np^n)_{\overline{\mathbb{Q}}} \to \cdots \to Y_1(Np)_{\overline{\mathbb{Q}}},
$$
where the transition map of the cohomology is given by the trace map. The representation $\rho_{\mathbf{f}}^{\operatorname{ord}}$ is called \textit{Hida deformation}.
Under the above setting, there is a well-defined notion of the \textit{residual representation} of $\rho_{\mathbf{f}}^{\operatorname{ord}}$. The original account of residual representations is \cite{MazWi}. A more detailed account is found in \cite{BookIwasawa} and Carayol's article in \cite{Kit}. Let us recall the following facts about $\mathcal{T}^{\operatorname{ord}}$, whose proof was originally given in \cite[Theorem 2.1]{Hid1}. Refer also to \cite[Theorem 2.1.1, Theorem 2.1.3 and Theorem 2.2.1]{EPW} for the review of these facts:
\begin{enumerate}
\item[\rm{(i)}]
$\rho_{\mathbf{f}}^{\operatorname{ord}}$ is unramified outside a finite set $\Sigma$ as above.
\item[\rm{(ii)}]
For a geometric Frobenius $\operatorname{Frob}_{\ell} \in G_{\mathbb{Q}}$ with $\ell \notin \Sigma$, we have $\operatorname{Tr}\rho_{\mathbf{f}}^{\operatorname{ord}}(\operatorname{Frob}_{\ell}))=T_{\ell}$, where $T_{\ell}$ is the Hecke operator in $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$.
\item[\rm{(iii)}]
Let us define $\mathcal{V}^{\operatorname{ord}} :=\mathcal{T}^{\operatorname{ord}} \otimes_{\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}}\operatorname{Frac}(\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}})$ with $\operatorname{Frac}(\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}})$ being the field of fractions of
$\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$.
Then, the restriction $\rho_{\mathbf{f}}^{\operatorname{ord}}|_{G_{\mathbb{Q}_p}}$ admits the following $G_{\mathbb{Q}_p}$-invariant filtration:
$$
0 \to F^+\mathcal{V}^{\operatorname{ord}} \to \mathcal{V}^{\operatorname{ord}} \to \mathcal{V}^{\operatorname{ord}}/F^+\mathcal{V}^{\operatorname{ord}} \to 0
$$
in which both $F^+\mathcal{V}^{\operatorname{ord}} $ and $\mathcal{V}^{\operatorname{ord}} /F^+\mathcal{V}^{\operatorname{ord}}$ are $\operatorname{Frac}(\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}})$-modules of rank one and $G_{\mathbb{Q}_p}$ acts on $\mathcal{V}^{\operatorname{ord}}/F^+\mathcal{V}^{\operatorname{ord}}$ via $\alpha$, where $\alpha:G_{\mathbb{Q}_p} \to (\mathbf{I}^{\mathrm{ord}}_{\mathbf{f}})^{\times}$ is an unramified character such that $\alpha(\operatorname{Frob}_p)=U_p$.
\end{enumerate}
Let $\widetilde{\chi}_{\operatorname{cyc}}$ be the universal cyclotomic character $G_{\mathbb{Q}} \twoheadrightarrow C_\infty \hookrightarrow \mathbb{Z}_p [[C_\infty ]]$ induced by $\chi_{\operatorname{cyc}}$. We denote by $\rho_{\mathbf{f}}^{\operatorname{n.ord}}$ the continuous $G_{\mathbb{Q}}$-representation $\rho_{\mathbf{f}}^{\operatorname{ord}} \widehat{\otimes}\widetilde{\chi}_{\operatorname{cyc}}$ and denote the representation space of $\rho_{\mathbf{f}}^{\operatorname{ord}} \widehat{\otimes}\widetilde{\chi}_{\operatorname{cyc}}$ by $\mathcal{T}^{\operatorname{n.ord}}$. This is a finitely generated module over $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}=\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}} \widehat{\otimes}_{\mathbb{Z}_p} \mathbb{Z}_p[[C_{\infty}]]$. The representation
$$
\rho_{\mathbf{f}}^{\operatorname{n.ord}}:G_{\mathbb{Q}} \to \operatorname{Aut}_{\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}}(\mathcal{T}^{\operatorname{n.ord}})
$$
is called \textit{nearly ordinary deformation} attached to the Hida family $\mathbf{f}$. We also call it \textit{(nearly ordinary) Hida deformation}.
We will consider the following conditions, which are not a major obstacle in practice:
\begin{enumerate}
\item[($\bf{NOR}$):]
$\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$ is a normal domain.
\item[($\bf{IRR}$):]
The residual representation associated with $\rho_{\mathbf{f}}^{\operatorname{ord}}$ is absolutely irreducible.
\item[($\bf{FIL}$):]
The restriction $\rho_{\mathbf{f}}^{\operatorname{ord}}|_{G_{\mathbb{Q}_p}}$ admits a $G_{\mathbb{Q}_p}$-invariant filtration:
$$
0 \to F^+\mathcal{T}^{\operatorname{ord}} \to \mathcal{T}^{\operatorname{ord}} \to \mathcal{T}^{\operatorname{ord}}/F^+\mathcal{T}^{\operatorname{ord}} \to 0,
$$
where $F^+\mathcal{T}^{\operatorname{ord}} $ and $\mathcal{T}^{\operatorname{ord}} /F^+\mathcal{T}^{\operatorname{ord}}$ are direct summands as $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$-modules, and the above sequence gives rise to the $G_{\mathbb{Q}_p}$-invariant filtration:
$$
0 \to F^+\mathcal{V}^{\operatorname{ord}} \to \mathcal{V}^{\operatorname{ord}} \to \mathcal{V}^{\operatorname{ord}} /F^+\mathcal{V}^{\operatorname{ord}} \to 0,
$$
after taking the base extension to $\mathcal{V}^{\operatorname{ord}} $.
\end{enumerate}
We note the following fact.
\begin{lemma}
Assume that the condition $(\bf{NOR})$ holds. Then $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$ is a three-dimensional local Cohen-Macaulay normal domain.
\end{lemma}
\begin{proof}
Since $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$ is a two-dimensional normal local domain, it is Cohen-Macaulay by Serre's normality criterion. Now we observe that $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}=\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}[[C_{\infty}]]$ is a three-dimensional local Cohen-Macaulay normal domain.
\end{proof}
Let us assume the condition $(\bf{IRR})$. Then the $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$-lattice $\mathcal{T}^{\operatorname{ord}}$ is uniquely determined and it is a free module of rank two. Thus, it defines a representation $\rho_{\mathbf{f}}^{\operatorname{ord}}:G_{\mathbb{Q}} \to GL_2(\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}})$. Moreover, since $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$ is a local domain, there is a unique maximal ideal $\frak{m}_{\mathbf{f}}$ of $\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}}$ such that $\frak{a}_{\mathbf{f}} \subseteq \frak{m}_{\mathbf{f}}$. Then the condition $(\bf{IRR})$ implies that the local component $(\mathbf{H}^{\operatorname{ord}}_{Np^{\infty}})_{\frak{m}_{\mathbf{f}}}$ is Gorenstein. These facts together with the proofs are found in \cite[Proposition 2 at page 254]{MazWi}. There is a sufficient condition that assures the local filtration to consist of rank one free modules when $\mathcal{T}^{\operatorname{ord}}$ is free of rank two.
\begin{enumerate}
\item[($\bf{DIST}$):]
The semi-simplification of the $G_{\mathbb{Q}_p}$-module $\mathcal{T}^{\operatorname{ord}}/\frak{m}_{\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}}\mathcal{T}^{\operatorname{ord}}$ is the sum of two distinct characters of $G_{\mathbb{Q}_p}$
with values in $(\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}/\frak{m}_{\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}})^\times$, where $\frak{m}_{\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}}$ is the maximal ideal of
$\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$.
\end{enumerate}
It is known that, if the condition $(\bf{DIST})$ holds and $\mathcal{T}^{\operatorname{ord}}$ is free of rank two, then both $F^+\mathcal{T}^{\operatorname{ord}}$ and $F^-\mathcal{T}^{\operatorname{ord}}$ are free $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$-modules of rank one. Especially, we have the implication: $(\bf{IRR})+(\bf{DIST}) \Rightarrow (\bf{FIL})$. In the rest of the present article, let us denote $\mathcal{T}^{\operatorname{n.ord}}$ by $\mathcal{T}$ for simplicity.\\
\subsection{Two-variable $p$-adic $L$-function over Hida deformations}
Recall that Kitagawa constructed $\Lambda$-adic modular symbols in \cite{Kit} in order to construct a two-variable $p$-adic $L$-function.
\begin{theorem}[Kitagawa]
\label{MazurKitagawa}
Assume that the condition $(\bf{IRR})$ holds. Then there exist free $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$-modules $\mathbf{MS}^+_{\mathbf{f}}$ and $\mathbf{MS}^-_{\mathbf{f}}$ of rank one such that for any arithmetic character $\kappa$ of $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$ with $w(\kappa) \ge 2$, the specialization $\mathbf{MS}^{\pm}_{\mathbf{f}} \otimes_{\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}} \kappa(\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}})$ is isomorphic as $\mathbb{Z}^{\wedge}_{\mathbf{f}_{\kappa},(p)}$-module, to the intersection:
$$
i \big(MS_{\mathbf{f}_{\kappa}}(\mathbb{Z}^{\wedge}_{\mathbf{f}_{\kappa},(p)})\big) \cap H^1_{\operatorname{Betti},c}(Y_1(M_{\kappa})_{\mathbb{C}},\mathcal{L}_{w(\kappa)}(\mathbb{Q}^{\wedge}_{\mathbf{f}_{\kappa},(p)}))[\mathbf{f}_{\kappa}]^{\pm},
$$
where $M_{\kappa}$ denotes the level of the cusp form $\mathbf{f}_{\kappa}$ and the composition map $i:MS_{\mathbf{f}_{\kappa}}(\mathbb{Z}^{\wedge}_{\mathbf{f}_{\kappa},(p)}) \simeq H^1_{\operatorname{Betti},c}(Y_1(M_{\kappa})_{\mathbb{C}},\mathcal{L}_{w(\kappa)}(\mathbb{Z}^{\wedge}_{\mathbf{f}_{\kappa},(p)})) \to H^1_{\operatorname{Betti},c}(Y_1(M_{\kappa})_{\mathbb{C}},\mathcal{L}_{w(\kappa)}(\mathbb{Q}^{\wedge}_{\mathbf{f}_{\kappa},(p)}))$ is the natural map induced by $\mathbb{Z}^{\wedge}_{\mathbf{f}_{\kappa},(p)} \to \mathbb{Q}^{\wedge}_{\mathbf{f}_{\kappa},(p)}$.
\end{theorem}
The modules $\mathbf{MS}^{\pm}_{\mathbf{f}}$ are called the \textit{spaces of $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$-adic modular symbols}. As we noted earlier, $(\bf{IRR})$ ensures that the local ring $(\mathbf{H}_{Np^{\infty}}^{\operatorname{ord}})_{\frak{m}_{\mathbf{f}}}$ is Gorenstein, where $\frak{m}_{\mathbf{f}}$ is the unique maximal ideal of $\mathbf{H}_{Np^{\infty}}^{\operatorname{ord}}$ such that $\frak{a}_{\mathbf{f}} \subseteq \frak{m}_{\mathbf{f}}$. The proof of Theorem \ref{MazurKitagawa} is found in \cite[\S~5.5]{Kit}.
\begin{definition}[$p$-adic period]
Assume that the condition $(\bf{IRR})$ holds. Fix a basis $\mathbf{B}^{+}_{\mathbf{f}}$ of the $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$-module $\mathbf{MS}^{+}_{\mathbf{f}}$ (resp.
$\mathbf{B}^{-}_{\mathbf{f}}$ of the $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$-module $\mathbf{MS}^-_{\mathbf{f}}$). Fix a basis $b_{\mathcal{T}^{\operatorname{ord}}_{\kappa}}^+$ of the $\mathbb{Z}_{\mathbf{f}_{\kappa},(p)}$-module $MS^+_{\mathbf{f}_{\kappa}}(\mathbb{Z}_{\mathbf{f}_{\kappa},(p)})$ (resp. $b_{\mathcal{T}^{\operatorname{ord}}_{\kappa}}^-$ of the $\mathbb{Z}_{\mathbf{f}_{\kappa},(p)}$-module $MS^-_{\mathbf{f}_{\kappa}}(\mathbb{Z}_{\mathbf{f}_{\kappa},(p)})$). We define the \textit{p-adic period} $\Omega_p(\mathbf{B}_{\mathbf{f}}^{\pm},
b^{\pm}_{\mathcal{T}^{\operatorname{ord}}_{\kappa}}) \in \overline{\mathbb{Z}}_p$ to be a constant satisfying the equality:
$$
\kappa(\mathbf{B}^{\pm}_{\mathbf{f}})=
\Omega_p(\mathbf{B}_{\mathbf{f}}^{\pm},
b^{\pm}_{\mathcal{T}^{\operatorname{ord}}_{\kappa}}) \cdot b^{\pm}_{\mathcal{T}^{\operatorname{ord}}_{\kappa}}.
$$
\end{definition}
We have the following theorem (see \cite[Theorem 1.1]{Kit}).
\begin{theorem}[Kitagawa]
Assume that the conditions $(\bf{NOR})$ and $(\bf{IRR})$ hold and let $\xi=(\chi^j_{\operatorname{cyc}}\phi,\kappa)$ be an arithmetic character such that $\phi:C_{\infty} \to \overline{\mathbb{Q}}^{\times}$ is a finite order Dirichlet character of $p$-power conductor and $1 \le j \le w(\kappa)+1$ (criticality condition). Then there exists an element $L_p(\{\mathbf{B}^{\pm}_{\mathbf{f}}\}) \in \mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$ which satisfies the following interpolation property over an $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$-adic family of cusp forms:
\begin{equation}
\label{interpolation}
\frac{\xi(L_p(\{\mathbf{B}^{\pm}_{\mathbf{f}}\}))}{\Omega_p(\mathbf{B}_{\mathbf{f}}^{\pm},
b_{\mathcal{T}^{\operatorname{ord}}_{\kappa}}^{(-1)^{j-1}\phi(-1)})}=(-1)^j (j-1)! \times
\operatorname{Eul}(\mathbf{f}_{\kappa,}j,\phi) \times
G(\phi) \times \frac{L(\mathbf{f}_{\kappa},\phi^{-1},j)}{(2\pi\sqrt{-1})^j \Omega_{\infty}(\mathbf{f}_{\kappa},b_{\mathcal{T}^{\operatorname{ord}}_{\kappa}}^{(-1)^{j-1}\phi(-1)})},
\end{equation}
where
$$
\operatorname{Eul}(\mathbf{f}_{\kappa},j,\phi)=\begin{cases}
\Big(1-\frac{p^{j-1}}{a_p(\mathbf{f}_{\kappa})}\Big), & \phi=\mathbf{I} \\
\Big(\frac{p^{j-1}}{a_p(\mathbf{f}_{\kappa})}\Big)^{\operatorname{ord}_p(\phi)}, & \phi \ne \mathbf{I}
\end{cases}
$$
and $\Omega_p(\mathbf{B}_{\mathbf{f}}^{\pm},
b_{\mathcal{T}^{\operatorname{ord}}_{\kappa}}^{(-1)^{j-1}\phi(-1)}) \in \overline{\mathbb{Z}}_p$ is a $p$-adic period and $\Omega_{\infty}(\mathbf{f}_{\kappa},b_{\mathcal{T}^{\operatorname{ord}}_{\kappa}}^{(-1)^{j-1}\phi(-1)}) \in \mathbb{C}$ is a complex period attached to $\mathbf{f}_{\kappa}$ and $b_{\mathcal{T}^{\operatorname{ord}}_{\kappa}}^{\pm}$, determined by the Eichler-Shimura map. $G(\phi)$ is the Gauss sum of the character $\phi$. Finally, the symbol $\xi(L_p(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})) \in \overline{\mathbb{Q}}_p$ denotes the specialization of $L_p(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})$ at the arithmetic point $\xi=(\chi_{\operatorname{cyc}}^j\phi,\kappa)$.
\end{theorem}
\begin{remark}
By Drinfeld-Manin theorem,
it is known that we have
$$
\frac{L(\mathbf{f}_{\kappa},\phi^{-1},j)}{(2\pi\sqrt{-1})^j \Omega_{\infty}(\mathbf{f}_{\kappa},b_{\mathcal{T}^{\operatorname{ord}}_{\kappa}}^{(-1)^{j-1}\phi(-1)})}
\in \overline{\mathbb{Q}}.
$$
Hence, the right-hand side of $(\ref{interpolation})$ is regarded as a $p$-adic number through the fixed embedding $i_p:\overline{\mathbb{Q}} \hookrightarrow \overline{\mathbb{Q}}_p$ (see also \cite{De79}).\\
\end{remark}
\section{Algebraic invariants of Hida deformations}
\label{GaloisDeformation}
In the present section, we first review Greenberg's Selmer group attached to the nearly ordinary (Hida) deformation based on the article \cite{Gr1} and prove some necessary results. After that, we recall the definition of Beilinson-Kato Euler system based on \cite{Kato} and introduce a $\Lambda$-adic family of these elements.\\
\subsection{Greenberg Selmer group}
Let $\mathcal{T}$ be the nearly ordinary deformation space attached to an $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$-adic newform $\mathbf{f}$ as introduced in \S~\ref{Hidafamily}. Take the discrete module $\mathcal{D}^{\operatorname{n.ord}}:=\mathcal{T} \otimes_{\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}}(\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}})^{\operatorname{PD}}$ viewed as a Galois module via $g \in G_{\mathbb{Q}} \mapsto g(m \otimes \sigma)=g(m) \otimes \sigma$. Then the filtration ($\bf{FIL}$) in \S~\ref{Hidafamily} induces a filtration: $0 \to F^+\mathcal{D}^{\operatorname{n.ord}} \to \mathcal{D}^{\operatorname{n.ord}} \to F^-\mathcal{D}^{\operatorname{n.ord}} \to 0$.
The \textit{Greenberg Selmer group} is defined as
$$
\operatorname{Sel}_{\mathbb{Q}}(\mathcal{T})=\ker\big(H^1(G_{\Sigma},\mathcal{D}^{\operatorname{n.ord}}) \to
H^1(I_{\mathbb{Q}_p},\mathcal{D}^{\operatorname{n.ord}}/ F^+ \mathcal{D}^{\operatorname{n.ord}}) \times
\prod_{\ell \nmid p } H^1(I_{\mathbb{Q}_{\ell}},\mathcal{D}^{\operatorname{n.ord}})\big),
$$
where $H^1(G_{\Sigma},\mathcal{D}^{\operatorname{n.ord}}) \to H^1(G_{\mathbb{Q}_{\ell}},\mathcal{D}^{\operatorname{n.ord}})$ is the restriction map.
We define the \textit{localization map for the nearly ordinary deformations} as the natural composite maps:
$$
\mathbf{Loc}_{/f}: H^1(G_{\Sigma},\mathcal{T}^*(1)) \to H^1(G_{\mathbb{Q}_p},\mathcal{T}^*(1)) \to H^1_{/f}(G_{\mathbb{Q}_p},\mathcal{T}^*(1)),
$$
where $H^1_{/f}(G_{\mathbb{Q}_p},\mathcal{T}^*(1))$ is the \textit{singular part} of the local Galois cohomology as defined in \cite[Lemma 3.10]{Oc}. Specializing to $\mathbf{f}_{\kappa}$ for an arithmetic character $\kappa \in \operatorname{Hom}_{\mathbb{Z}_p}(\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}},\overline{\mathbb{Q}}_p)$ with $w(\kappa) \ge 2$, let $T_{\mathbf{f}_{\kappa}}$ be a $\mathbb{Z}_{\mathbf{f}_{\kappa},(p)}^{\wedge}$-lattice of the $G_{\mathbb{Q}}$-representation attached to $\mathbf{f}_{\kappa}$. Let $\phi:C_{\infty} \to \overline{\mathbb{Q}}^{\times}$ be a finite order Dirichlet character of $p$-power conductor and let $1 \le j \le w(\kappa)+1$ (criticality condition). Then we define the \textit{localization map} as the
composite maps:
$$
\operatorname{loc}_{/f}:H^1(G_{\Sigma},T^*_{\mathbf{f}_{\kappa}}(1-j) \otimes \phi^{-1}) \to H^1(G_{\mathbb{Q}_p},T^*_{\mathbf{f}_{\kappa}}(1-j) \otimes \phi^{-1}) \to H^1_{/f}(G_{\mathbb{Q}_p},T^*_{\mathbf{f}_{\kappa}}(1-j) \otimes \phi^{-1}).
$$
Finally, the \textit{Bloch-Kato's dual exponential map} is defined by
$$
\exp^*:H^1_{/f}(G_{\mathbb{Q}_p},V^*_{\mathbf{f}_{\kappa}}(1-j) \otimes \phi^{-1}) \to \Fil^0 D_{\dR}(V^*_{\mathbf{f}_{\kappa}}(1-j) \otimes \phi^{-1}),
$$
where $V_{\mathbf{f}_{\kappa}}:=T_{\mathbf{f}_{\kappa}} \otimes_{\mathbb{Z}_p} \mathbb{Q}_p$.
\subsection{$p$-adic $L$-function and Beilinson-Kato Euler system}
In \cite{Oc} and \cite{Oc2}, we gave another construction of a two-variable $p$-adic $L$-function attached to $\mathcal{T}$ based on Mazur-Kitagawa $L$-function \cite{Kit}. Recall that Beilinson-Kato Euler system, which is constructed and studied extensively in \cite{Kato}, is adapted into the Hida deformations and it thus defines a $\Lambda$-adic version of Euler system.
We are ready to formulate the main content of Beilinson-Kato elements for nearly ordinary Hida deformations.
\begin{theorem}
\label{Coleman}
Assume that the conditions $(\bf{NOR})$ and $(\bf{IRR})$ hold for the nearly ordinary deformation space $\mathcal{T}$ attached to a Hida family $\mathbf{f}$. Fix an $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$-basis $\{\mathbf{B}^{\pm}_{\mathbf{f}}\}$ of the $\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}}$-modules $\mathbf{MS}^{\pm}_{\mathbf{f}}$. Let us put
$$
\mathfrak{R}:=\{r \in \mathbb{Z}_{\ge 1}~|~(r,2p)=1~\mbox{and}~r~\mbox{is a square-free integer}\}.
$$
Let $G_{\Sigma,r}$ be the Galois group of $\mathbb{Q}(\mu_r)_{\Sigma}/\mathbb{Q}(\mu_r)$ with $\Sigma=\{p,\infty\}$. Then there exists a collection of elements:
$$
\big\{\mathbf{z}^{\operatorname{BK}}_r(\{\mathbf{B}^{\pm}_{\mathbf{f}}\}) \in H^1(G_{\Sigma,r},\mathcal{T}^*(1))\big\}_{r \in \mathfrak{R}}
$$
which defines Euler system for $(\mathcal{T},\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}},p)$ such that the following statements hold:
\begin{enumerate}
\item[\rm(i)]
For $ r \ell\in \mathfrak{R}$ and the inertial group $I_{\ell} \subseteq G_{\mathbb{Q}}$, we have the following equality
$$
\operatorname{Cor}_{\mathbb{Q}(\mu_{r\ell})/\mathbb{Q}(\mu_{r})}(\mathbf{z}^{\operatorname{BK}}_{r \ell}(\{\mathbf{B}^{\pm}_{\mathbf{f}}\}))=P(\operatorname{Frob}_{\ell};\mathcal{T})\mathbf{z}^{\operatorname{BK}}_{r}(\{\mathbf{B}^{\pm}_{\mathbf{f}}\}),
$$
where $P(\operatorname{Frob}_{\ell};\mathcal{T})=\det(1-\operatorname{Frob}_{\ell}X;\mathcal{T}^*(1)^{I_{\ell}})$ is the characteristic polynomial for $\operatorname{Frob}_{\ell}$.
\item[\rm (ii)]
Let us put
$$
\mathbf{z}^{\operatorname{BK}}(\{\mathbf{B}^{\pm}_{\mathbf{f}}\}):=
\mathbf{z}^{\operatorname{BK}}_1(\{\mathbf{B}^{\pm}_{\mathbf{f}}\}) \in H^1(G_{\Sigma,1},\mathcal{T}^*(1)).
$$
Then for any arithmetic character $\xi:=(\chi_{\operatorname{cyc}}^j \phi, \kappa) \in \operatorname{Hom}_{\mathbb{Z}_p}(\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}},\overline{\mathbb{Q}}_p)$ such that $1 \le j \le w(\kappa)+1$ and $\phi$ is a finite order Dirichlet character of $p$-power conductor, the element
$$
\operatorname{loc}_{/f} \circ \xi(\mathbf{z}^{\operatorname{BK}}_{r \ell}(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})) \in H^1_{/f}(G_{\mathbb{Q}_p}, V^*_{\mathbf{f}_{\kappa}}(1-j) \otimes \phi^{-1})
$$
satisfies the following interpolation formula
$$
\exp^*\big(\operatorname{loc}_{/f} \circ \xi(\mathbf{z}^{\operatorname{BK}}(\{\mathbf{B}^{\pm}_{\mathbf{f}}\}))\big)=
\Omega_p(\mathbf{B}_{\mathbf{f}}^{\sgn(j,\phi)},
b_{\mathcal{T}^{\operatorname{ord}}_{\kappa}}^{\sgn(j,\phi)}) \times
\frac{L(\mathbf{f}_{\kappa},\phi^{-1},j)}{(2 \pi \sqrt{-1})^j \Omega_{\infty}(\mathbf{f}_{\kappa},b_{\mathcal{T}^{\operatorname{ord}}_{\kappa}}^{\sgn(j,\kappa)})} \cdot \overline{\mathbf{f}_{\kappa}}
$$
which holds in $\Fil^0 D_{\dR}(V^*_{\mathbf{f}_{\kappa}}(1-j) \otimes \phi^{-1})$.
\item[\rm(iii)]
There is an injective $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$-module map (Coleman map):
$$
\Xi:H^1_{/f}(G_{\mathbb{Q}_p},\mathcal{T}^*(1)) \to \mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}
$$
with pseudo-null cokernel and such that $L_p(\{\mathbf{B}^{\pm}_{\mathbf{f}}\}):=\Xi(\mathbf{Loc}_{/f}(\mathbf{z}^{\operatorname{BK}}(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})))$ satisfies the following properties:
\begin{enumerate}
\item[\rm{(a)}]
$\operatorname{char}_{\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}}\big(H^1_{/f}(G_{\mathbb{Q}_p},\mathcal{T}^*(1))/\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}} \cdot \mathbf{Loc}_{/f}(\mathbf{z}^{\operatorname{BK}}(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})\big)
=\big(L_p(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})\big)$.
\item[\rm{(b)}]
$L_p(\{\mathbf{B}^{\pm}_{\mathbf{f}}\}) \in \mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$ is a two-variable $p$-adic $L$-function which satisfies the interpolation property as described in $(\ref{interpolation})$.
\end{enumerate}
\end{enumerate}
\end{theorem}
We refer the reader to \cite{Oc}, \cite{Oc1} and \cite{Oc2} for Beilinson-Kato elements and the construction of Coleman map for Hida deformations. Theorem \ref{Coleman} is the main content of \cite{Oc}, where the $p$-adic $L$-function is constructed via the dual exponential map. Various $p$-adic $L$-functions have been constructed by several people through different methods. We refer the reader to \cite{Oc2} for comparison results on these functions with their connection to the main conjecture.
\begin{lemma}
\label{deRham}
Under the notation as in Theorem \ref{Coleman}, we have the following assertions.
\begin{enumerate}
\item[\rm{(1)}]
There is an exact sequence of finitely generated $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$-modules $:$
$$
0 \to H^1(G_{\Sigma},\mathcal{T}^*(1))/\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}} \cdot \mathbf{z}^{\operatorname{BK}}(\{\mathbf{B}^{\pm}_{\mathbf{f}}\}) \to H^1_{/f}(G_{\mathbb{Q}_p},\mathcal{T}^*(1))/\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}} \cdot \mathbf{Loc}_{/f}(\mathbf{z}^{\operatorname{BK}}(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})
$$
$$
\to \operatorname{Sel}_{\mathbb{Q}}(\mathcal{T})^{\operatorname{PD}} \to \textcyr{Sh}^2_{\Sigma}(\mathcal{T}^*(1)) \to 0.
$$
\item[\rm{(2)}]
$\operatorname{Sel}_{\mathbb{Q}}(\mathcal{T})^{\operatorname{PD}}$ is finitely generated and torsion over $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$.
\end{enumerate}
\end{lemma}
\begin{proof}
For the assertion $\rm(1)$, we have $(\mathcal{D}^{\operatorname{n.ord}})^{\operatorname{PD}} \simeq \mathcal{T}^*$ as $G_{\mathbb{Q}}$-modules. Since we know $\textcyr{Sh}^1_{\Sigma}(\mathcal{D}^{\operatorname{n.ord}}) \subseteq \operatorname{Sel}_{\mathbb{Q}}(\mathcal{T})$ and $\mathcal{D}^{\operatorname{n.ord}} \simeq M^{\operatorname{PD}}(1)$ for $M:=\mathcal{T}^*(1)$, the surjection on the most right side of the above sequence follows. Hence, the existence of the above exact sequence from the Poitou-Tate duality.
The assertion $\rm(2)$ is found in \cite[Proposition 4.9]{Oc2}.
\end{proof}
In order to prove one of the expected divisibilities in the Iwasawa main conjecture for the Hida deformations, we will need a deep result of Kato concerning the finiteness of Selmer groups \cite[Section 14]{Kato}. Using this result, it follows that
$$
H^2(G_{\Sigma},\mathcal{D}^{\operatorname{n.ord}})=0~(\mathrm{weak~Leopoldt~conjecture}).
$$
See \cite[Lemma 8.3]{Oc2} for the proof of this fact. For our purpose, it will be sufficient to have a weaker condition that $H^2(G_{\Sigma},\mathcal{D}^{\operatorname{n.ord}})$ is a finite group.\\
\subsection{Finiteness of local Galois cohomology groups}
As we have seen in Theorem \ref{Euler} and Theorem \ref{veryfinal}, the vanishing of $H^2(G_{\mathbb{Q}_{\ell}},T^*(1))$ plays a role in Euler system bound. Let us discuss its finiteness in terms of the Fourier expansion of a modular form. Let $f$ be a normalized eigen newform over $\overline{\mathbb{Q}}_p$, of weight $\ge 2$ and tame level $N$. Take $\mathcal{O}$ to be the ring of $p$-integers such that
$a_n(f) \in \mathcal{O}$ for all $n \ge 1$ with $f=\sum_{n=1}^{\infty}a_n(f)q^n$. Denote by $T$ an $\mathcal{O}$-lattice of the $p$-adic Galois representation attached to $f$. Now let $V:=T \otimes_{\mathcal{O}} K$ for the quotient field $K$ of $\mathcal{O}$. Denote by $T^*$ the $\mathcal{O}$-dual representation of $T$. We have the following result.
\begin{lemma}
\label{finite1}
Assume $\ell$ is a prime dividing $N$ and the notation is as above. Then $a_{\ell}(f) \ne 1$ if and only if $H^2(G_{\mathbb{Q}_{\ell}},T^*(1))$ is a finite group.
\end{lemma}
\begin{proof}
$H^2(G_{\mathbb{Q}_{\ell}},T^*(1))$ is a finitely generated $\mathcal{O}$-module and $H^2(G_{\mathbb{Q}_{\ell}},V^*(1)) \simeq H^2(G_{\mathbb{Q}_{\ell}},T^*(1)) \otimes_{\mathbb{Z}_p} \mathbb{Q}_p$. Therefore, the finiteness of $H^2(G_{\mathbb{Q}_{\ell}},T^*(1))$ is equivalent to the vanishing: $H^2(G_{\mathbb{Q}_{\ell}},V^*(1))=0$. Let $I_\ell$ be the inertia group at $\ell$. Then by local and Pontryagin dualities, it suffices to show that $(V^*)_{G_{\mathbb{Q}_{\ell}}}=0$. Note that $(V^*)_{G_{\mathbb{Q}_{\ell}}}$ is the quotient of $(V^*)_{I_{\ell}}$. As $f$ is a newform of tame level $N$, $V$ is ramified at all primes $\ell$ dividing $N$. Hence we have $\dim_K (V^*)_{I_{\ell}} \le 1$.
If $(V^*)_{I_{\ell}}=0$, there is nothing to prove. So let us assume that
$(V^*)_{I_{\ell}}$ is one-dimensional. Then it defines a character $\varphi:G_{\mathbb{Q}_{\ell}} /I_{\ell} \to K^{\times}$ and thus, the eigenvalue of $\operatorname{Frob}_{\ell}$, acting on this space invertibly, is $a_{\ell}(f)$. Thus,
$$
(1-a_{\ell}(f))(V^*)_{I_{\ell}}=(V^*)_{I_{\ell}} \iff a_{\ell}(f) \ne 1,
$$
which yields that $(V^*)_{G_{\mathbb{Q}_{\ell}}}=0$. The lemma now follows.
\end{proof}
\section{Applications to the Iwasawa Main Conjecture}
\label{proof2}
The aim of this section is to prove the following theorem as an application of Theorem \ref{veryfinal}
\begin{theorem}
\label{MainTh}
Assume that $(\bf{NOR})$, $(\bf{IRR})$ and $(\bf{FIL})$ hold for the nearly ordinary Hida deformation $\mathcal{T}$ attached to a Hida family $\mathbf{f}$. Fix an $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$-basis $\mathbf{B}^{\pm}_{\mathbf{f}}$ of the modules of $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$-adic modular symbols $\mathbf{MS}^{\pm}_{\mathbf{f}}$. Assume further that
\begin{enumerate}
\item[\rm{(i)}]
There exists an element $\sigma_1 \in G_{\mathbb{Q}(\mu_{p^{\infty}})}$ such that $\rho_{\mathbf{f}}^{\operatorname{n.ord}}(\sigma_1) \simeq \begin{pmatrix} 1 & \epsilon \\ 0 & 1 \end{pmatrix} \in GL_2(\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}})$ for a nonzero element $\epsilon \in \mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$.
\item[\rm{(ii)}]
There exists an element $\sigma_2 \in G_{\mathbb{Q}}$ such that $\sigma_2$ acts on $\mathcal{T}$ as multiplication by $-1$.
\item[\rm{(iii)}]
If $\ell$ is any prime dividing $Np$, the maximal Galois invariant quotient vanishes; $(\mathcal{T}^*)_{G_{\mathbb{Q}_{\ell}}}=0$.
\end{enumerate}
Let $k$ be the number of minimal $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$-generators of $\textcyr{Sh}^2_{\Sigma}(\mathcal{T}^*(1))$. Then we have
$$
\big(\epsilon^k L_p(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})\big) \subseteq \operatorname{char}_{\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}}\big(\operatorname{Sel}_{\mathbb{Q}}(\mathcal{T})^{\operatorname{PD}}\big).
$$
\end{theorem}
\begin{proof}
Let $\mathcal{D}^{\operatorname{n.ord}}$ denote the discrete module associated to $\mathcal{T}$. Then $\mathcal{D}^{\operatorname{n.ord}} \simeq (\mathcal{T}^*)^{\operatorname{PD}}$. By assumption, $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$ is a three-dimensional Cohen-Macaulay normal domain. Recall that the characteristic ideal is a reflexive ideal and is additive with respect to short exact sequences. Putting together Theorem \ref{Coleman} and Lemma \ref{deRham}, it follows that
$$
\big(\epsilon^k L_p(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})\big) \subseteq \operatorname{char}_{\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}}\big(\operatorname{Sel}_{\mathbb{Q}}^{\Sigma}(\mathcal{T})^{\operatorname{PD}}\big)
$$
if and only if
$$
\operatorname{char}_{\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}}\big(\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}/(\epsilon^k) \oplus H^1(G_{\Sigma},\mathcal{T}^*(1))/\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}} \mathbf{z}^{\operatorname{BK}}_1(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})\big) \subseteq \operatorname{char}_{\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}}\big(\textcyr{Sh}^2_{\Sigma}(\mathcal{T}^*(1))\big).
$$
Or equivalently,
$$
(\epsilon^k) \operatorname{char}_{\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}}\big(H^1(G_{\Sigma},\mathcal{T}^*(1))/\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}} \mathbf{z}^{\operatorname{BK}}_1(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})\big) \subseteq \operatorname{char}_{\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}}\big(\textcyr{Sh}^2_{\Sigma}(\mathcal{T}^*(1))\big).
$$
Note that all modules appearing inside the characteristic ideals are finitely generated torsion $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$-modules. Then the conclusion of the theorem follows from Theorem \ref{veryfinal}, as all the hypotheses are satisfied.
\end{proof}
We obtain the following corollary of the main theorem. In \cite{MazWi}, some examples of modular $p$-adic Galois representations containing $SL_2(\mathbb{Z}_p)$ are constructed.
\begin{corollary}
\label{MainCo}
Assume that $(\bf{NOR})$ and $(\bf{FIL})$ hold for the nearly ordinary Hida deformation $\mathcal{T}$ attached to a Hida family $\mathbf{f}$ and assume that the maximal Galois invariant quotient vanishes; $(\mathcal{T}^*)_{G_{\mathbb{Q}_{\ell}}}=0$ for every prime $\ell$ dividing $Np$. Fix an $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$-basis $\mathbf{B}^{\pm}_{\mathbf{f}}$ of the modules of $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$-adic modular symbols $\mathbf{MS}^{\pm}_{\mathbf{f}}$. If the image of the restriction map
$$
\rho_{\mathbf{f}}^{\operatorname{ord}}:G_{\mathbb{Q}(\mu_{p^{\infty}})} \to GL_2(\mathbf{I}^{\operatorname{ord}}_{\mathbf{f}})
$$
contains a conjugate of $SL_2(\mathbb{Z}_p[[D_{\infty}]])$, then we have
$$
\big(L_p(\{\mathbf{B}^{\pm}_{\mathbf{f}}\})\big) \subseteq \operatorname{char}_{\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}}\big(\operatorname{Sel}_{\mathbb{Q}}(\mathcal{T})^{\operatorname{PD}}\big).
$$
\end{corollary}
\begin{proof}
It suffices to check that the condition $(\bf{IRR})$ holds. Since the image of $\rho_{\mathbf{f}}^{\operatorname{ord}}$ contains a conjugate of $SL_2(\mathbb{Z}_p[[D_{\infty}]])$ by assumption, the residual representation associated to $\rho_{\mathbf{f}}^{\operatorname{ord}}$ is absolutely irreducible.
\end{proof}
\begin{remark}
\begin{enumerate}
\item
If the isomorphism holds: $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}} \simeq \mathcal{O}[[X,Y]]$, then it is not necessary to assume that $(\mathcal{T}^*)_{G_{\mathbb{Q}_{\ell}}}=0$ for every prime $\ell$ dividing $Np$ in Corollary \ref{MainCo}, as it was already considered in \cite{Oc1}.
\item
It is an important problem to find an example of $\mathbf{I}^{\operatorname{n.ord}}_{\mathbf{f}}$ which is normal, but not regular. The authors do not know if there is such an example. It will be interesting to ask the same problem in the case of the nearly ordinary Hecke algebra attached to a $\Lambda$-adic family of Hilbert modular cusp forms as constructed in \cite{Hid11}. A formulation of the $\Lambda$-adic version of the main conjecture in the totally real case requires an elaborate analysis as given in \cite{Oc3}.
\end{enumerate}
\end{remark}
| 03b4464ef20addd737fec2a91d411a7104f6df38 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Multi-Task Learning (MTL) aims to enhance the overall generalization performance by learning multiple related tasks simultaneously.
It has been extensively studied from various points of view~\cite{Caruana97,Evgeniou05,WATR2010,qi2010semi}.
As an example, the common tastes of users (i.e., tasks) with respect to movies (i.e., instances) can be harnessed into a movie recommender system using MTL~\cite{pan2010transfer}.
Most MTL methods run under the offline learning setting where the training data for each task is available beforehand.
However, offline learning methods are generally inefficient, since they suffer from a high training cost and poor scalability.
This is especially true when it comes to the large-scale streaming data.
As a remedy, MTL has been studied under the online setting, in which the model runs over a sequence of data by processing them one by one~\cite{anderson2008theory}.
After updating the model in each round, the current input will be discarded.
As a result, online learning algorithms are efficient and scalable, and have been successfully applied to a number of MTL applications~\cite{Saha,lugosi2009online,ruvolo2014online,attenberg2009collaborative,yang2016learning}.
In this paper, we investigate MTL under the online setting.
Existing online MTL methods assume that all tasks are related with each other and simply constrain their relationships via a presumed structure~\cite{Saha,CavallantiCG10}.
However, such a constraint may be too restrictive and rarely hold in the real-life applications, as the personalized tasks with individual traits often exist~\cite{gong2012robust}.
We attempt to address this drawback through a creative formulation of online MTL that consists of two components:
the first component captures a low-rank \emph{correlative} structure over the related tasks,
while the second one represents the \emph{personalized} patterns of individual tasks.
Specifically, our algorithm learns a weight matrix which is decomposed into two components as aforementioned.
A nuclear norm regularization is imposed on the first component to induce a low-rank \emph{correlative} structure of the related tasks.
A group lasso penalty is applied onto the second component of all individual tasks to identify the outliers.
Next, we apply an online projected gradient scheme to solve this non-smooth problem with a closed-form solution for the \emph{correlative} and \emph{personalized} components.
This gives our algorithm two advantages: 1) it is efficient to make predictions and update models in a real-time manner; 2) it can achieve a good trade-off between the common and personalized structures.
We provide a theoretical evaluation for our algorithm by giving a proof that our algorithm can achieve a sub-linear regret compared to the best linear model in hindsight.
Although our algorithm achieves good performance, it may not accurately approximate a low-rank matrix: the nuclear norm is essentially the $l_1$ norm of singular values, known for being biased in estimation since large singular values are detrimental to the approximation. To address this issue, we use a log-determinant function to approximate the matrix rank, that is able to reduce the contributions of large singular values while keeping those of small singular values close to zero. To solve this non-convex optimization problem, a proximal gradient algorithm is derived to adaptively learn such a low-rank structure with a closed-form solution.
In addition, we prove that there is a unique root of the refined objective under a proper parameter setting.
Finally, we conduct comparative experiments against a variety of state-of-the-art techniques on three real-world datasets.
Empirically, the refined algorithm with the log-determinant function achieves better performance than that with the nuclear norm due to a better low-rank approximation.
The rest of this paper is organized as follows.
Section 2 introduces related work.
The problem setting and the proposed algorithm with analysis are presented in Section 3 and Section 4, respectively.
Section 5 provides experimental results.
Section 6 concludes this paper.
\section{Related Work}
In this section, we briefly introduce works related to MTL in the offline and online settings, followed by the low-rank matrix approximation.
\subsection*{Multi-Task Learning}
Conventional offline or batch MTL algorithms can be broadly classified into the following two categories: \emph{explicit} parameter sharing and \emph{implicit} parameter sharing.
In the first category, all tasks can be made to share some common parameters explicitly.
Such common parameters include hidden units in neural networks~\cite{baxter2000model}, prior in hierarchical Bayesian models~\cite{bakker2003task,yu2005learning}, feature mapping matrix~\cite{AndoZ05} and classification weight~\cite{Evgeniou04}.
On the other side, the shared structure can be estimated in an implicit way by imposing a low rank subspace~\cite{pong2010trace,negahban2011estimation}, e.g. Trace-norm Regularized Multi-task Learning (TRML)~\cite{zhou2012mutal} captured a common low-dimensional subspace of task relationship with a trace-norm regularization; or a common set of features~\cite{argyriou2008convex,yang2009heterogeneous}, e.g. Multi-Task Feature Learning (MTFL)~\cite{ArgyriouEP06} learned a common feature across the tasks in an unsupervised manner. Besides, \cite{AbernethyBR07} and \cite{Agarwal} proposed a few experts to learn the task relatedness on the entire task set. These MTL techniques have been successfully used in the real-world applications, e.g. multi-view action recognition~\cite{yan2014multitask}, spam detection~\cite{haideceptive}, head pose estimation~\cite{yan2016multi}, etc.
Compared to the offline learning, online learning techniques are more efficient and suitable to handle massive and sequential data~\cite{yang2015min,conf/icml/ZhaoHJY11,yang2016efficient}. An early work ~\cite{Dekel,DekelLS06}, Online Multi-Task Learning (OMTL), studied online learning of multiple tasks in parallel. It exploited the task structure by using a global loss function. Another work \cite{LiCHLJ11} proposed a collaborative online framework, Confidence-weighted Collaborative Online Multi-task Learning (CW-COL), which learned the take relativeness via combining the individual and global variations of online Passive-Aggressive (PA) algorithms~\cite{Crammer}.
Instead of fixing the task relationship via a presumed structure~\cite{CavallantiCG10}, a recent Online Multi-Task Learning approach introduced an adaptive interaction matrix which quantified the task relevance with LogDet Divergence (OMTLLOG) and von-Neumann Divergence (OMTLVON)~\cite{Saha}, respectively. Most Recently, \cite{crammer2012learning} proposed an algorithm, Shared Hypothesis model (SHAMO), which used a K-means-like procedure to cluster different tasks in order to learn the shared hypotheses.
Similar to SHAMO, \cite{murugesan2016adaptive} proposed an Online Smoothed Multi-Task Learning with Exponential updates (OSMTL-e). It jointly learned both the per-task model parameters and the inter-task relationships in an online MTL setting.
The algorithm presented in this paper differs from existing ones in that it can learn both a common structure among the correlative tasks and the individual structure of outlier tasks.
\subsection*{Low-Rank Matrix Approximation}
In many areas (e.g. machine learning, signal and image processing), high-dimensional data are commonly used. Apart from being uniformly distributed, high-dimensional data often lie on the low-dimensional structures. Recovering the low-dimensional subspace can well preserve and reveal the latent structure of the data. For example, face images of an individual under different lighting conditions span a low-dimensional subspace from an ambient high-dimensional space \cite{basri2003lambertian}. To learn low-dimensional subspaces, recently proposed methods, such as Low-Rank Representation (LRR) \cite{liu2013robust} and Low-Rank Subspace and Clustering (LRSC) \cite{favaro2011closed}, usually depended on the nuclear norm as a convex rank approximation function to seek low-rank subspaces. Unlike the rank function that treats them equally, the nuclear norm simply adds all nonzero singular value together, where the large values may contribute exclusively to the approximation, rendering it much deviated from the true rank. To resolve this problem, we propose a log-determinant function to approximate the rank function, which is able to reduce the contributions of large singular values while keeping those of small singular values close to zero. To the best of our knowledge, this is the first work that exploits a log-determinant function to learn a low-rank structure of task relationship in the online MTL problem.
\section{Problem Setting}
In this section, we first describe our notations, followed by the problem setting of online MTL.
\subsection{Notations}
Lowercase letters are used as scalars, lowercase bold letters as vectors, uppercase letters as elements of a matrix, and bold-face uppercase letters as matrices.
$\x_i$ and $X_{ij}$ denote the $i$-th column and the $(i,j)$-th element of a matrix $\mathbf{X}$, respectively.
Euclidean and Frobenius norms are denoted by $\|\cdot\|$ and $\|\cdot\|_F$, respectively.
In particular, for every $q,p \geq 1$, we define the $(q,p)$-norm of $\mathbf{A} \in \mathbb{R}^{d\times m}$ as $\|\mathbf{A}\|_{q,p} = (\sum_{i=1}^m\|\a_i\|_q^p)^{\frac{1}{p}}$.
When the function $f(\w)$ is differentiable, we denote its gradient by $\nabla f(\w)$.
\subsection{Problem Setting}
According to the online MTL setting, we are faced with $m$ different but related classification problems, also known as tasks.
Each task has a sequential instance-label pairs, i.e., $\{(\x_t^i,y_t^i)\}_{1\leq t \leq T}^{1\leq i\leq m}$, where $\x_t^i \in \mathbb{R}^d$ is a feature vector drawn from a single feature space shared by all tasks, and $y_t^i \in \{\pm 1\}$.
The algorithm maintains $m$ separate models in parallel, one for each of the $m$ tasks.
At the round $t$, $m$ instances $\{\x_t^1,\ldots,\x_t^m\}$ are presented at one time.
Given the $i$-th task instance $\x_t^i$, the algorithm predicts its label using a linear model $\z_t^i$, i.e., $\hat{y}_t^i = \textrm{sign}(\hat{p}_t^i)$, where $\hat{p}_t^i = \z_t^{i\top}\x_t^i$ and $\z_t^i$ is the weight parameter of the round $t$.
The true label $y_t^i$ is not revealed until then.
A hinge-loss function is applied to evaluate the prediction,
\bqs
f_t^i(\z_t^i) = [1 - y^i_t \hat{p}_t^i]_+= [1 - y^i_t\z_t^i\cdot\x^i_t]_{+},
\eqs
where $[a]_{+} = \max\{0,a\}$.
The cumulative loss over all $m$ tasks at the round $t$ is defined as
\bqs
F_t(Z_t) = \sum_{i=1}^m f_t^i(\z_t^i),
\eqs
where $Z_t = [\z_t^1,\ldots,\z_t^m] \in \mathbb{R}^{d\times m}$ is the weight matrix for all tasks.
Inspired by the Regularized Loss Minimization (RLM) in which one minimizes an empirical loss plus a regularization term jointly~\cite{shalev2011stochastic}, we formulate our online MTL to minimize the regret compared to the best linear model in hindsight,
\bqs\label{RegretStatic}\notag
R_\phi \triangleq \sum_{t=1}^T [F_t(Z_t) + g(Z_t)] - \inf_{Z\in\Omega}\sum_{t=1}^T [F_t(Z) + g(Z)],
\eqs
where $\Omega\subset\mathbb{R}^d$ is a closed convex subset and the regularizer $g:\Omega\rightarrow\mathbb{R}$ is a convex regularization function that constraints $\Omega$ into simple sets, e.g. hyperplanes, balls, bound constraints, etc.
For instance, $g(Z) = \|Z\|_1$ constrains $Z$ into a sparse matrix.
\vspace{-0.01in}
\begin{figure
\centering
\caption{Learning Personalized and Low-rank Structures from Multiple Tasks}\label{Flow-Chart}
\subfigure {\includegraphics[width=0.52\textwidth,height=5.1cm]{figures_flow_2.eps}}
\end{figure}
\section{Algorithm}
We propose to solve the regret $R_\phi$ by two steps: 1) to learn the correlative and personalized patterns over multiple tasks; 2) to achieve an optimal solution for the regret $R_\phi$.
\subsection{Correlative and Personalized Structures}
We propose a novel formulation for online MTL that incorporates two components, as illustrated in Fig. \ref{Flow-Chart}. The first component captures a low-rank \emph{common} structure $U$ over the similar tasks, where one model (or pattern) can be shared cross the related tasks. As outlier tasks often exist in real-world scenarios, the second one, $V$, identifies the \emph{personalized} patterns specific to individual tasks. Thus, incorporation of two structures $U$ and $V$ could make the final model $W$ more robust and reliable.
To learn both correlative and personalized structures from multiple tasks, we decompose the weight matrix $Z$ into two components: \emph{correlative} matrix $U$ and \emph{personalized} matrix $V$, and define a new weight matrix,
\bqs
\label{decompositionW}
\Omega = \{W | W = \begin{bmatrix} U \\ V \end{bmatrix}, U\in\mathbb{R}^{d\times m}, V\in\mathbb{R}^{d\times m}\},
\eqs
where $\w^{i} = \begin{bmatrix} \u^{i} \\ \v^{i} \end{bmatrix}\in\mathbb{R}^{2d}$ is the $i$-th column of the weight matrix $W = [\w^{1},\ldots,\w^{m}]\in\mathbb{R}^{2d\times m}$.
Denoted by matrix $Z$ the summation of $U$ and $V$, we obtain
\bqs
\label{decomposeZ}
Z = U + V = \begin{bmatrix} I_d, I_d \end{bmatrix}W,
\eqs
where $I_d\in\mathbb{R}^{d\times d}$ is an identity matrix.
Given an instance $(\x^{i}_{t},y^{i}_{t})$ , the algorithm makes prediction based on both the correlative and personalized parameters,
\bqs
\hat{p}_t^i & = \z_t^{i}\cdot\x_t^i \overset{(\ref{decomposeZ})} {=} (\begin{bmatrix} I_d , I_d \end{bmatrix} \w^{i}_t) \cdot \x_t^i \\
& = (\u_t^i+\v_t^i)^{\top}\x_t^i,
\eqs
with the corresponding loss function,
\bqs
\notag
f_t^i(\z^i_t) = f_t^i(\begin{bmatrix} I_d, I_d \end{bmatrix}\w^{i}_t) = [1 - y^{i}_t(\u^{i}_t+\v^{i}_t)^{\top}\x^{i}_t]_{+}.
\eqs
We thus can reformat the cumulative loss function with respect to $W$,
\bqs
\label{LossFunctionW}
L_t(W_t) = F_t(Z_t) \overset{(\ref{decomposeZ})} {=} F_t(\begin{bmatrix} I_d, I_d \end{bmatrix}W_t).
\eqs
We impose a regularizer on $U$ and $V$, respectively,
\bqs
\label{RegularizerW}
& r(W) = g(\begin{bmatrix} I_d, I_d \end{bmatrix}W) \triangleq \lambda_1 r(U) + \lambda_2 r(V),
\eqs
where $\lambda_1$ and $\lambda_2$ are non-negative trade-off parameters.
Substituting Eq. (\ref{LossFunctionW}) and (\ref{RegularizerW}) into the regret $R_\phi$, it can be formatted as
\bqs
\label{RegretStaticW}
R_\phi \triangleq \sum_{t=1}^T [L_t(W_t) + r(W_t)] - \inf_{W\in\Omega}\sum_{t=1}^T [L_t(W) + r(W)],
\eqs
where $\phi_t(W) = L_t(W) + r(W)$ is a non-smooth convex function.
We next show how to achieve an optimal solution to the reformatted regret (\ref{RegretStaticW}).
\subsection{Online Task Relationship Learning}
Inspired by~\cite{bertsekas1999nonlinear}, we can solve the regret (\ref{RegretStaticW}) by a subgradient projection,
\bqs
\label{subgradient-projection}
\displaystyle{\mathop{\mathrm{argmin}}_{W\in\Omega}} \; \|W - (W_t - \eta\nabla\phi_t(W_t))\|_F,
\eqs
where $\eta > 0$ is the learning rate.
In the following lemma, we show that the problem (\ref{subgradient-projection}) can be turned into a linearized version of the proximal algorithm~\cite{rockafellar1976monotone}.
To do so, we first introduce a Bregman-like distance function~\cite{bregman1967relaxation},
\bqs
\notag
B_\psi(W,W_t) = \psi(W) - \psi(W_t) - \langle W-W_t,\nabla \psi(W_t)\rangle,
\eqs
where $\psi$ is a differentiable and convex function.
\begin{lemma}
\label{Lemma_Linearization}
Assume $\psi(\cdot)=\frac{1}{2}\|\cdot\|_F^2$, then using first-order Taylor expansion of $\phi_t(W_t)$, the algorithm (\ref{subgradient-projection}) is equivalent to a linearized form with a step-size parameter $\eta > 0$,
\bqs
\notag
\label{linearization_projection}
W_{t+1} = \displaystyle{\mathop{\mathrm{argmin}}_{W\in\Omega}} \; \langle \nabla\phi_t(W_t),W - W_t\rangle + \frac{1}{\eta}B_\psi(W,W_t).
\eqs
\end{lemma}
\noindent
Instead of balancing this trade-off individually for each of the multiple tasks, we balance it for all the tasks jointly.
However, the subgradient of a composite function, i.e. $\nabla\phi_t(W_t) = \nabla L_t(W_t) + \nabla r(W_t)$ cannot lead to a desirable effect, since we should constrain the projected gradient (i.e. $W_t - \eta\nabla\phi_t(W_t)$) into a restricted set.
To address this issue, we refine the optimization function by adding a regularizer on $W$,
\bqs
\label{optimal_solution}
W_{t+1} \triangleq \displaystyle{\mathop{\mathrm{argmin}}_{W\in\Omega}} \;
& \langle \nabla L_t(W_t),W - W_t\rangle \\
& + \frac{1}{\eta}B_\psi(W,W_t) + r(W).
\eqs
Note that the formulation (\ref{optimal_solution}) is different from the Mirror Descent (MD) algorithm~\cite{beck2003mirror}, since we do not \emph{linearize} the regularizer $r(W)$.
Given that $W = \begin{bmatrix} U^{\top}, V^{\top} \end{bmatrix}^{\top}$, we show that the problem (\ref{optimal_solution}) can be presented with $U$ and $V$ in the lemma below.
\begin{lemma}
Assume that $\psi(W) = \frac{1}{2}\|W\|^2_F$ and $W = \begin{bmatrix} U \\ V \end{bmatrix}$, the problem (\ref{optimal_solution}) turns into an equivalent form in terms of $U$ and $V$,
\bqs
\label{OMTL-LRO_solution}
& (U_{t+1},V_{t+1}) \triangleq \displaystyle{\mathop{\mathrm{argmin}}_{U,V\in\Omega}} \; \lambda_1r(U) + \lambda_2r(V) \\
& + \frac{1}{2\eta_1}\|U - U_t\|^2_F + \frac{1}{2\eta_2}\|V - V_t\|^2_F \\
& + \langle\nabla_{U}L_t(U_t), U - U_t\rangle + \langle\nabla_{V}L_t(V_t), V - V_t\rangle,
\eqs
where the parameters $\eta_1$ and $\eta_2$ control previous learned knowledge retained by $U$ and $V$.
\end{lemma}
\begin{proof}
Assume that $\psi(W) = \frac{1}{2}\|W\|^2_F$, we obtain:
\bqs\notag
\label{B_psi_UV}
B_\psi(W,W_t) & = \frac{1}{2}\|W\|^2_F - \frac{1}{2}\|W_t\|^2_F - \langle W - W_t, W_t\rangle \\
& = \frac{1}{2}\|W - W_t\|^2_F \\
& \overset{(\ref{decompositionW})} {=} \frac{1}{2}\|U - U_t\|^2_F + \frac{1}{2}\|V - V_t\|^2_F.
\eqs
The linearized gradient form can be rewritten as:
\bqs\notag
\label{Linear_UV}
& \langle \nabla L_t(W_t),W - W_t \rangle
\overset{(\ref{decompositionW})} {=} \langle \begin{bmatrix} \nabla_{U}L_t(U_t) \\ \nabla_{V}L_t(V_t) \end{bmatrix} , \begin{bmatrix} U - U_t \\ V - V_t \end{bmatrix} \rangle \\
& = \langle\nabla_{U}L_t(U_t), U - U_t\rangle + \langle\nabla_{V}L_t(V_t), V - V_t\rangle.
\eqs
Substituting above two inferences and (\ref{RegularizerW}) into problem (\ref{optimal_solution}), we complete this proof.
\end{proof}
We next introduce the regularization $r(U)$ and $r(V)$, and then present how to optimize this non-smooth convex problem with a closed-form solution.
\subsection{Regularization}
As mentioned above, restricting task relatedness to a presumed structure via a single weight matrix~\cite{Saha} is too strict and not always plausible in practical applications.
To overcome this problem, we thus impose a regularizer on $U$ and $V$ as follows,
\bqs\label{regularization-UV}
r(U) = \|U\|_{*} & & r(V) = \|V\|_{2,1}.
\eqs
A nuclear norm~\cite{pong2010trace} is imposed on $U$ (i.e., $\|U\|_{*}$) to represent multiple tasks ($\u_i,i\in[1,m]$) by a small number (i.e. $n$) of the basis ($n\ll m$).
Intuitively, a model performing well on one task is likely to perform well on the similar tasks.
Thus, we expect that the best model can be shared across several related tasks.
However, the assumption that all tasks are correlated may not hold in real applications.
Thus, we impose the $l_{(2,1)}$-norm~\cite{kim2010tree} on $V$ (i.e., $\|V\|_{2,1}$), which favors a few non-zero columns in the matrix $V$ to capture the personalized tasks.
Note that our algorithm with the regularization terms above is able to detect personalized patterns, unlike the algorithms~\cite{AbernethyBR07,Agarwal,chen2012learning}.
Although prior work~\cite{gong2012robust} considers detecting the personalized task, it was designed for the offline setting, which is different from our algorithm since we learn the personalized pattern adaptively with online techniques.
\subsubsection{Optimization}
Although the composite problem (\ref{OMTL-LRO_solution}) can be solved by~\cite{vandenberghe1996semidefinite}, the composite function with linear constraints has not been investigated to solve the MTL problem.
We employ a projected gradient scheme~\cite{boyd2004convex,boyd2011distributed} to optimize this problem with both smooth and non-smooth terms.
Specifically, by substituting (\ref{regularization-UV}) into (\ref{OMTL-LRO_solution}) and omitting the terms unrelated to $U$ and $V$, the problem can be rewritten as a projected gradient schema,
\bqs
\notag
(U_{t+1},V_{t+1}) = \; & \displaystyle{\mathop{\mathrm{argmin}}_{U,V\in\Omega}} \;
\frac{1}{2\eta_1}\|U - \hat{U}_t\|_F^2 + \lambda_1\|U\|_* \\
& + \frac{1}{2\eta_2}\|V - \hat{V}_t\|_F^2 + \lambda_2\|V\|_{2,1}.
\eqs
where
\bqs
\notag
\hat{U}_t = U_t - \eta_1\nabla_U L_t(U_t), & & \hat{V}_t = V_t - \eta_2\nabla_V L_t(V_t).
\eqs
Due to the decomposability of the objective function above, the solution for $U$ and $V$ can be optimized separately,
\bqs
\label{OptimalU}
U_{t+1} = \displaystyle{\mathop{\mathrm{argmin}}_{U\in\Omega}} \; \frac{1}{2\eta_1}\|U - \hat{U}_t\|_F^2 + \lambda_1\|U\|_{*}.
\eqs
\bqs
\label{OptimalV}
V_{t+1} = \displaystyle{\mathop{\mathrm{argmin}}_{V\in\Omega}} \; \frac{1}{2\eta_2}\|V - \hat{V}_t\|_F^2 + \lambda_2\|V\|_{2,1}.
\eqs
This has two advantages: 1) there is a closed-form solution for each update; 2) the update for the $U$ and $V$ can be performed in parallel.
\noindent
\textbf{Computation of U:}
\noindent
Inspired by~\cite{boyd2004convex}, we show that the optimal solution to (\ref{OptimalU}) can be obtained via solving a simple convex optimization problem in the following theorem.
\begin{thm}\label{them_U}
Denote by the eigendecomposition of $\hat{U}_t = P\hat{\Sigma} Q^{\top}\in\mathbb{R}^{d\times m}$ where $r = $\textrm{rank}$(\hat{U}_t)$, $P\in\mathbb{R}^{d\times r}$, $Q\in\mathbb{R}^{m\times r}$, and $\hat{\Sigma}=$\textrm{diag}$(\hat{\sigma}_1,\ldots,\hat{\sigma}_r)\in\mathbb{R}^{r\times r}$.
Let $\{\sigma_i\}_{i=1}^r, \sigma_i \geq 0$ be the solution of the following problem,
\bqs
\label{OptimalSigma}
\min_{\{\sigma_i\}_{i=1}^r} & \frac{1}{2\eta_1}\sum_{i=1}^r(\sigma_i - \hat{\sigma}_i)^2 + \lambda_1 \sum_{i=1}^r \sigma_i.
\eqs
It is easy to obtain the optimal solution for (\ref{OptimalSigma}): $\sigma_i^* = [\hat{\sigma}_i - \eta_1\lambda_1]_{+}$ for $i\in[1,r]$.
Assume that $\Sigma^{*} =$diag$(\sigma_1^*,\ldots,\sigma_r^*)\in\mathbb{R}^{r\times r}$, the optimal solution to Eq. (\ref{OptimalU}) is given by,
\bqs
\label{OptimalSolutionU}
U^* = P\Sigma^{*} Q^{\top},
\eqs
\end{thm}
\noindent
\textbf{Computation of V:}
\noindent
We rewrite (\ref{OptimalV}) by solving an optimization problem for each column,
\bqs
\label{optimizationV1}
\min_{\v_i\in\mathbb{R}^d} \sum_{i=1}^m \frac{1}{2\eta_2}\|\v_i - \hat{\v}_i\|^2 + \lambda_2\sum_{i=1}^m\|\v_i\|_2.
\eqs
where $\hat{\v}_i\in\mathbb{R}^{d}$ denotes the $i$-th column of $\hat{V}_t= V_t - \eta_2\nabla_V L_t(V_t) = [\hat{\v}_1,\ldots,\hat{\v}_m]$.
The optimal operator problem (\ref{optimizationV1}) above admits a closed-form solution with time complexity of $O(dm)$~\cite{tibshirani1996regression},
\bqs
\label{OptimalSolutionV}
\v^{i}_{t+1} = \max(0,1 - \frac{\eta_2\lambda_2}{\|\hat{\v}^{i}_t\|_2})\hat{\v}^{i}_t, \quad \forall i\in[1,m].
\eqs
We observe that $\v^i$ would be retained if $\|\hat{\v}^{i}\|_2 > \eta_2\lambda_2$, otherwise, it decays to $\textbf{0}$.
Hence, we infer that only the personalized patterns among the tasks, which differs from the low-rank common structures and thus cannot be captured by $U$, would be retained in $V$.
The two quantities $V_t$ and $U_t$ can be updated according to a closed-form solution on each round $t$. A mistake-driven strategy is used to update the model.
Finally, this algorithm, which we call Robust Online Multi-tasks learning under Correlative and persOnalized structures with NuClear norm term (ROMCO-NuCl), is presented in Alg. \ref{ROMCO-NuCl}.
\begin{algorithm}[t]
\caption{ROMCO-NuCl} \label{ROMCO-NuCl}
\begin{algorithmic}[1]
\STATE {\bf Input}: a sequence of instances $(\x^{i}_t,y^{i}_t,), \forall t\in[1,T]$, and the parameter $\eta_1$, $\eta_2$, $\lambda_1$ and $\lambda_2$.
\STATE {\bf Initialize}: $\u_{0}^i = \mathbf{0}, \v_{0}^i = \mathbf{0}$ for $\forall i\in[1,m]$;
\FOR{$t=1,\ldots, T$}
\FOR{$i=1,\ldots, m$}
\STATE Receive instance pair ($\x^{i}_t$ $y^{i}_t$);
\STATE Predict $\hat{y}_t^{i} = \sign[(\u^{i}_t + \v^{i}_t)\cdot\x_t^{i}]$;
\STATE Compute the loss function $f_t^i(\w_t^{i})$;
\ENDFOR
\IF {$\exists i\in[1,m], f_t^i(\w_t^i) > 0$}
\STATE Update $U_{t+1}$ with Eq.~(\ref{OptimalSolutionU});
\STATE Update $V_{t+1}$ with Eq.~(\ref{OptimalSolutionV});
\ELSE
\STATE $U_{t+1} = U_{t}$ and $V_{t+1} = V_{t}$;
\ENDIF
\ENDFOR
\STATE {\bf Output}: $\w^{i}_T$ for $i\in[1,m]$
\end{algorithmic}
\end{algorithm}
\subsection{Log-Determinant Function}
\vspace{-0.01in}
\begin{figure
\centering
\caption{The rank, nuclear norm and log-determinant objectives in the scalar case}\label{Rank-Value-Comparison}
\subfigure {\includegraphics[width=0.52\textwidth,height=5.3cm]{Rank-Value-Comparison.eps}}
\end{figure}
\vspace{-0.0in}
While the nuclear norm has been theoretically proven to be the tightest convex approximation to the rank function, it is usually difficult to theoretically prove whether the nuclear norm near-optimally approximates the rank function, e.g., the incoherence property \cite{candes2009exact}\cite{candes2010power}.
In addition, the nuclear norm may not accurately approximate the rank function in practice, since the matrix "rank" regards all nonzero singular values to have equal contributions, i.e., regarding all positive singular values as ''1", as shown by the red line in Fig. \ref{Rank-Value-Comparison}; while the nuclear norm, as shown by the yellow star line in Fig. \ref{Rank-Value-Comparison}, treats the nonzero singular values differently, i.e., it simply adds all nonzero values together, thus the larger singular values make more contribution to the approximation.
To solve this issue, we introduce a log-determinant function as follows,
\begin{definition}
Let $\sigma_i$ ($i=1,2,\ldots,n$) be the singular values of the matrix $U$,
\bqs\label{log_determinant}\notag
r^{ld}(U) = \log\det\left(I + U^{\top}U\right) = \sum_{i=1}^n \log\left(1 + \sigma_i^2\right).
\eqs
\begin{itemize}
\item When $\sigma_i = 0$, the term $\log(1+\sigma_i^2) = 0$, which is the same as the true rank function;
\item When $0 < \sigma_i < 1$, $\log(1 + \sigma_i^2) < \sigma_i$, implying that those small singular values can be reduced further;
\item For those large singular values $\sigma_i > 1$, $\log(1 + \sigma_i^2) \ll \sigma_i$, which is a significant reduce over large singular values.
\end{itemize}
\end{definition}
In this case, $r^{ld}(U)$ approximates the rank function better than the nuclear norm by significantly reducing the weights of large singular values, meanwhile regarding those very small singular values as noise, as presented with the blue circle line in Fig. \ref{Rank-Value-Comparison}.
\subsubsection{Optimal Solution}
Replacing the nuclear norm $\|\cdot\|_*$ with the log-determinant function $r^{ld}(U)$, the minimization of $U$ is reduced to the following problem:
\bqs\label{log_determinant_Objective}
U_{t+1} = \arg\min_{U} \frac{1}{2\eta_1}\|U - \hat{U}_t\|_F^2 + \lambda_1r^{ld}(U).
\eqs
To solve the objective function above,
we show that the optimal solution could be obtained by solving the roots of a cubic equation in the following theorem.
\begin{thm}
Let $\hat{U}_t = P\hat{\Sigma} Q^{\top}\in\mathbb{R}^{d\times m}$ where $\hat{\Sigma}=$\textrm{diag}$(\hat{\sigma}_1,\ldots,\hat{\sigma}_r)\in\mathbb{R}^{r\times r}$ and $r = $\textrm{rank}$(\hat{U}_t)$. Let $\{\sigma_i\}_{i=1}^r, \sigma_i \geq 0$ be the solution of the following problem,
\bqs
\label{OptimalSigma_log}
\min_{\{\sigma_i\}_{i=1}^r} \sum_{i=1}^r\left[\frac{1}{2\eta_1}(\sigma_i - \hat{\sigma}_i)^2 + \lambda_1\log(1+\sigma_i^2)\right].
\eqs
Then the optimal solution to Eq. (\ref{log_determinant_Objective}), similar to Thm. \ref{them_U}, is given by
$U^* = P\Sigma^{*} Q^{\top}$, where $\Sigma^{*} =$diag$(\sigma_1^*,\ldots,\sigma_r^*)\in\mathbb{R}^{r\times r}$ and $\sigma_i^*$ is the optimal solution of (\ref{OptimalSigma_log}) .
To obtain the solution, the problem is reduced to solving the derivative of Eq. (\ref{OptimalSigma_log}) for each $\{\sigma_i\}_{i=1}^r \geq 0$ with $\rho = \eta_1\lambda_1$,
\bqs\label{derivate_objective}
\frac{1}{\rho}\sigma_i^3 - \frac{1}{\rho}\hat{\sigma}_i\sigma_i^2 + (\frac{1}{\rho} + 2)\sigma_i - \frac{1}{\rho}\hat{\sigma}_i = 0.
\eqs
\end{thm}
\noindent
\begin{algorithm}[t]
\caption{ROMCO-LogD} \label{LogD_ROMCO}
\begin{algorithmic}[1]
\STATE {\bf Input}: a sequence of instances $(\x^{i}_t,y^{i}_t,), \forall t\in[1,T]$, and the parameter $\eta_1$, $\eta_2$, $\lambda_1$ and $\lambda_2$.
\STATE {\bf Initialize}: $\u_{0}^i = \mathbf{0}, \v_{0}^i = \mathbf{0}$ for $\forall i\in[1,m]$;
\FOR{$t=1,\ldots, T$}
\FOR{$i=1,\ldots, m$}
\STATE Receive instance pair ($\x^{i}_t$ $y^{i}_t$);
\STATE Predict $\hat{y}_t^{i} = \sign[(\u^{i}_t + \v^{i}_t)\cdot\x_t^{i}]$;
\STATE Compute the loss function $f_t^i(\w_t^{i})$;
\ENDFOR
\IF {$\exists i\in[1,m], f_t^i(\w_t^i) > 0$}
\STATE Update $U_{t+1}$ by solving Eq. (\ref{derivate_objective});
\STATE Update $V_{t+1}$ with Eq. (\ref{OptimalSolutionV});
\ELSE
\STATE $U_{t+1} = U_{t}$ and $V_{t+1} = V_{t}$;
\ENDIF
\ENDFOR
\STATE {\bf Output}: $\w^{i}_T$ for $i\in[1,m]$
\end{algorithmic}
\end{algorithm}
In general, the equation (\ref{derivate_objective}) has three roots. The details of root computation is given in Appendix.
In addition, in the following proposition, we prove that a unique positive root for (\ref{derivate_objective}) can be obtained in a certain parameter setting.
\begin{proposition}\label{Proposition1}
Assume that $\Theta(\sigma_i) = \frac{1}{2\rho}(\sigma_i - \hat{\sigma}_i)^2 + \log(1+\sigma_i^2)$ with $\rho = \eta_1\lambda_1$, when $\hat{\sigma}_i = 0$, $ \sigma_i = 0$, under the condition that $\hat{\sigma}_i > 0$ and $\frac{1}{\rho} > 1/4$, $\sigma_i$ located in $(0, \hat{\sigma}_i)$ is the unique positive root in cubic Eq. (\ref{derivate_objective}).
\end{proposition}
\begin{proof}
We need to minimize Eq. (\ref{OptimalSigma_log}) under the constraint of $\sigma_i \geq 0$. The derivative of $\Theta(\sigma_i)$ is
\bqs\notag
\Theta^{\prime}(\sigma_i) = \frac{2\sigma_i}{1 + \sigma_i^2} + \frac{1}{\rho}(\sigma_i - \hat{\sigma}_i),
\eqs
and the second derivative is
\bqs\notag
\Theta^{\prime\prime}(\sigma_i) = \frac{\frac{1}{\rho}\sigma_i^4 + (\frac{2}{\rho} - 2)\sigma_i^2 + (2 + \frac{1}{\rho})}{(1+\sigma_i^2)^2}.
\eqs
\emph{Case 1}: If $\hat{\sigma}_i = 0$, because $\sigma_i \geq 0$, we have $\Theta^{\prime}(\sigma_i) \geq 0$. That is, $\Theta(\sigma_i)$ is nondecreasing for any $\sigma_i \geq 0$ and strictly increasing with $\sigma_i > 0$. Thus, the minimizer of $\Theta(\sigma_i)$ is $\sigma_i^* = 0$.\\
\emph{Case 2}: If $\hat{\sigma}_i > 0$, then the roots exist only in a region of $(0,\hat{\sigma}_i)$ to let $\Theta^{\prime}(\sigma_i) = 0$, since $\Theta^{\prime}(\sigma_i)$ is monotonic with $\sigma_i$, and $\Theta^{\prime}(0) = -\frac{1}{\rho}\hat{\sigma}_i < 0$ while $\Theta^{\prime}(\hat{\sigma}_i) > 0$.
\begin{itemize}
\item If $\frac{1}{\rho} > 1/4$, then $\Theta^{\prime\prime}(\sigma_i) > 0$ since $\frac{1}{\rho}\sigma_i^4 + (\frac{2}{\rho} - 2)\sigma_i^2 + (2 + \frac{1}{\rho}) > 0$. In this case, $\Theta(\sigma_i)$ is a strictly convex function with a unique root in $(0,\hat{\sigma}_i)$. Thus, the proposition is proven.
\item If $0 < \frac{1}{\rho} \leq 1/4$, we determine the minimizer in the following way: Denote the set of positive roots of Eq. (\ref{derivate_objective}) by $\Omega_+$. By the first-order necessary optimality condition, the minimizer needs to be chosen from $\{0\}\cup\Omega_+$, that is, $\sigma^*_i = \arg\min_{\sigma_i\in\{0\}\cup\Omega_+}\Theta(\sigma_i)$.
\end{itemize}
\noindent
In our experiments, we initialize $\frac{1}{\rho} = 1$ and increase its value in each iteration. Therefore, when $\hat{\sigma}_i > 0$, the minimizer $\sigma^*_i\in(0, \hat{\sigma}_i)$ is the unique positive root of (\ref{derivate_objective}); when $\hat{\sigma}_i = 0$, $\sigma^*_i = 0$.
\end{proof}
We are ready to present the algorithm: RMOCO with Log-Determinant function for rank approximation, namely ROMCO-LogD, which also exploits a mistake-driven update rule. We summarize ROMCO-LogD in Alg. \ref{LogD_ROMCO}. To the best of our knowledge, this is the first work that proposes a log-determinant function to learn a low-rank structure of task relationship in the online MTL problem. In the next section, we will theoretically analyze the performance of the proposed online MTL algorithms ROMCO-NuCl/LogD.
\section{Theoretical Analysis}
We next evaluate the performance of our online algorithm ROMCO-NuCl/LogD in terms of the regret bound.
We first show the regret bound of the algorithm (\ref{optimal_solution}) and its equivalent form in the following lemma, which is essentially the same as Theorem 2 in the paper~\cite{duchi2010composite}:
\begin{lemma}
\label{General-Regret}
Let $\{W_t\}$ be updated according to (\ref{optimal_solution}).
Assume that $B_\psi(\cdot,\cdot)$ is $\alpha$-strongly convex w.r.t. a norm $\|\cdot\|_p$ and its convex conjugate $\|\cdot\|_q$ with $\frac{1}{p} + \frac{1}{q} = 1$, then for any $W^{*}\in\Omega$,
\bqs
\label{general_regret}\notag
R_\phi \leq \frac{1}{\eta}B_\psi(W^{*},W_1) + r(W_1) + \frac{\eta}{2\alpha}\sum_{t=1}^{T}\|\nabla L_t(W_t)\|_{q}^2.
\eqs
\end{lemma}
\begin{Remark}
We show that the above regret is $O(\sqrt{T})$ with respect to the best linear model in hindsight.
Suppose that the functions $F_t$ are Lipschitz continuous, then $\exists G_{q}$ such that $\max_t\|\nabla L_t(W_t)\|_{q} \leq G_{q}$.
Then we obtain:
\bqs
\notag
R_\phi \leq \frac{1}{\eta}B_\psi(W^{*}, W_1) + r(W_1) + \frac{T\eta}{2\alpha}G_{q}^2.
\eqs
We also assume that $r(W_1) = 0$.
Then by setting $\eta = \sqrt{2\alpha B_\psi(W^{*},W_1)}/(\sqrt{T}G_{q})$, we have $R_\phi \leq \sqrt{2TB_\psi(W^{*},W_1)}G_{q}/\sqrt{\alpha}$.
Given that $G_{q}$ is constant, setting $\eta \propto 1/\sqrt{T}$, we have $R_\phi = O(\sqrt{T})$.
\end{Remark}
\begin{lemma}
\label{ProjectionEqua}
The general optimization problem (\ref{optimal_solution}) is equivalent to the two step process of setting:
\bqs
\notag
& \tilde{W}_t = \displaystyle{\mathop{\mathrm{argmin}}_{W\in\Omega}} \; \frac{1}{\eta}B_\psi(W,W_t) + \langle\nabla L_t(W_t), W \rangle, \\
& W_{t+1} = \displaystyle{\mathop{\mathrm{argmin}}_{W\in\Omega}} \; \frac{1}{\eta}\{B_\psi(W,\tilde{W}_t) + r(W) \}.
\eqs
\end{lemma}
\begin{proof}
The optimal solution to the first step satisfies
$\nabla\psi(\tilde{W}_t) - \nabla\psi(W_t) + \eta\nabla L_t(W_t) = 0$, so that
\bqs
\label{FirstStep}
\nabla\psi(\tilde{W}_t) = \nabla\psi(W_t) - \eta\nabla L_t(W_t).
\eqs
Then look at the optimal solution for the second step.
For some $r'(W_{t+1})\in \partial r(W_{t+1})$, we have
\bqs
\label{SecondStep}
\nabla\psi(W_{t+1}) - \nabla\psi(\tilde{W}_t) + \eta r'(W_{t+1}) = 0.
\eqs
Substituting Eq.~(\ref{FirstStep}) into Eq.~(\ref{SecondStep}), we obtain
\bqs
\notag
\frac{1}{\eta}(\nabla\psi(W_{t+1}) - \nabla\psi(W_t)) + \nabla L_t(W_t) + r'(W_{t+1}) = 0,
\eqs
which satisfies the optimal solution to the one-step update of (\ref{optimal_solution}).
\end{proof}
We next show that ROMCO can achieve a sub-linear regret in the following theory.
\begin{thm}
\label{Strong-Regret}
The algorithm ROMCO (Alg.~\ref{ROMCO-NuCl} and Alg. \ref{LogD_ROMCO}) runs over a sequence of instances for each of the $m$ tasks.
Assume that $r(0) = 0$, i.e., $W_1 = 0$ and $\max_t \|\nabla L_t(W_t)\| \leq G_2$, $U, V \in \mathbb{R}^{d\times m}$, then the following inequality holds for all $W^{*}\in\Omega$,
\bqs
\label{boundAlgorithm}\notag
R_\phi \leq \frac{1}{2\eta}\|W^{*}\|^2_F + T\eta G_{2}^2 = O(G_2\|W^{*}\|\sqrt{T}).
\eqs
\end{thm}
\begin{proof}
Let $\psi(\cdot)=\frac{1}{2}\|\cdot\|^2_F$, according to Lemma \ref{ProjectionEqua}, the solutions in subgradient projection (\ref{OptimalU}) and~\eqref{OptimalV} are equivalent to the one in form of the general optimization (\ref{optimal_solution}).
Based on Lemma \ref{General-Regret}, for any $U^{*}\in\Omega$,
\bqs
\notag
R_\phi \leq \frac{1}{\eta}B_\psi(W^{*}, W_1) + r(W_1) + \frac{\eta}{2\alpha}\sum_{t=1}^{T}\|\nabla L_t(W_t)\|^2_q.
\eqs
Because that $\psi(\cdot)=\frac{1}{2}\|\cdot\|^2_F$ (i.e., $p=q=2$), $\|\nabla L_t(W_t)\| \leq G_2$.
Assuming $W_1 = 0$, we obtain $r(W_1) = 0$ and $B_\psi(W^{*}, W_1)=\frac{1}{2}\|W^{*}\|_F^2$.
Thus,
\bqs
\notag
R_\phi \leq \frac{1}{2\eta}\|W^{*}\|^2_F + \frac{T\eta}{2}G_{2}^2.
\eqs
By setting $\eta = \frac{\|W^*\|}{\sqrt{T}G_2}$, we have $R_\phi = O(G_2\|W^{*}\|\sqrt{T})$.
\end{proof}
\section{Experimental Results}
We evaluate the performance of our algorithm on three real-world datasets.
We start by introducing the experimental data and benchmark setup, followed by discussions on the results of three practical applications.
\subsection{Data and Benchmark Setup}
\begin{table}[t]
\centering
\caption{Statistics of three datasets}
\label{statistic_data}
\scriptsize
\begin{tabular}[2.1\textwidth]{|c|c|c|c|} \hline
& Spam Email & MHC-I & EachMovie \\ \hline
\#Tasks & 4 & 12 & 30 \\ \hline
\#Sample & 7068 & 18664 & 6000 \\ \hline
\#Dimesion & 1458 & 400 & 1783 \\ \hline
\#MaxSample & 4129 & 3793 & 200 \\ \hline
\#MinSample & 710 & 415 & 200 \\ \hline
\end{tabular}
\end{table}
\subsubsection{Experimental Datasets and Baseline}
We used three real-world datasets to evaluate our algorithm: \emph{Spam Email}\footnote{http://labs-repos.iit.demokritos.gr/skel/i-config/},
\emph{Human MHC-I}\footnote{http://web.cs.iastate.edu/~honavar/ailab/} and \emph{EachMovie} \footnote{http://goldberg.berkeley.edu/jester-data/}.
Table~\ref{statistic_data} summarizes the statistics of three datasets.
Each of the datasets can be converted to a list of binary-labeled instances, on which binary classifiers could be built for the applications of the three real-world scenarios: Personalized Spam Email Filtering, MHC-I Binding Prediction, and Movie Recommender System.
We compared two versions of the ROMCO algorithms with two batch learning methods: multi-task feature learning (\emph{MTFL})~\cite{ArgyriouEP06} and trace-norm regularized multi-task learning (\emph{TRML}) \cite{zhou2012mutal}, as well as six online learning algorithms: online multi-task learning (\emph{OMTL}) \cite{DekelLS06}, online passive-aggressive algorithm (\emph{PA}) \cite{Crammer}, confidence-weighted online collaborative multi-task learning (\emph{CW-COL}) \cite{LiCHLJ11} and three recently proposed online multi-task learning algorithms: \emph{OMTLVON}, \emph{OMTLLOG} \cite{Saha} and \emph{OSMTL-e} \cite{murugesan2016adaptive}. Due to the expensive computational cost in the batch models, we modified the setting of MTFL and TRML to handle online data by periodically retraining them after observing 100 samples. All parameters for MTFL and TRML were set by default values. To further examine the effectiveness of the PA algorithm, we deployed two variations of this algorithm as described below:
\emph{PA-Global} learns a single classification model from data of all tasks;
\emph{PA-Unique} trains a personalized classifier for each task using its own data.
The parameter C was set to 1 for all related baselines and the ROMCO algorithms. Other parameters for CW-COL, OMTLVON(OMTLLOG) and OSMTL-e were tuned with a grid search on a held-out random shuffle.
The four parameters, $\eta_1$, $\eta_2$, $\lambda_1$ and $\lambda_2$ for ROMCO-NuCl/LogD, were tuned by the grid search $\{10^{-6}$,\ldots,$10^{0}\}$ on a held-out random shuffle.
\begin{table*}[t]
\centering
\caption{Cumulative error rate (\%) and F1-measure (\%) with their standard deviation in the parenthesis on the Spam Email Dataset}
\label{email_comparison_result}
\scriptsize
\begin{tabular}[2.1\textwidth] {|c|c|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{Algorithm} & \multicolumn{2}{|c|}{\emph{User1}} & \multicolumn{2}{|c|}{\emph{User2}} & \multicolumn{2}{|c|}{\emph{User3}} & \multicolumn{2}{|c|}{\emph{User4}} \\
\cline{2-9}
& Error Rate & \tabincell{c}{Legit F1} & Error Rate & \tabincell{c}{Legit F1} & Error Rate & \tabincell{c}{Legit F1} & Error Rate & \tabincell{c}{Legit F1} \\
\hline\hline
MTFL & 13.16(1.21) & 88.61(1.06) & 8.72(0.58) & 94.67(0.38) & 14.84(0.91)& 86.70(0.88) & 16.87(0.78) & 83.59(0.77) \\ \hline
TRML & 17.71(0.99) & 84.61(0.88) & 12.45(0.94) & 92.24(0.65) & 13.89(0.73)& 87.57(0.66) & 20.78(1.08) & 79.67(1.09) \\ \hline
PA-Global & 6.15(0.45) & 94.54(0.41) & 8.59(0.82) & 94.72(0.51) & 4.12(0.30) & 96.33(0.27) & 9.75(0.61) & 90.20(0.64) \\ \hline
PA-Unqiue & 5.05(0.49) & 95.51(0.44) & 8.28(0.85) & 94.91(0.52) & 3.67(0.36) & 96.73(0.32) & 8.43(0.86) & 91.52(0.89) \\ \hline
CW-COL & 5.10(0.61) & 95.45(0.55) & \textbf{6.58(0.60)} & \textbf{95.91(0.38)} & 4.14(0.12) & 96.29(0.10) & 7.95(0.73) & 92.08(0.73) \\ \hline
OMTL & 5.00(0.48) & 95.55(0.43) & 8.01(0.77) & 95.07(0.48) & 3.55(0.29) & 96.84(0.26) & 8.24(0.73) & 91.71(0.75) \\ \hline
OSMTL-e & 7.34(0.69) & 93.46(0.59) & 10.25(0.89) & 93.61(0.57) & 5.86(0.69) & 94.74(0.61) & 10.42(1.18) & 89.68(1.07) \\ \hline
OMTLVON & 18.88(5.70) & 85.55(3.70) & 19.76(0.10) & 89.04(0.06) & 3.54(0.30) & 96.85(0.29) & 10.54(2.47) & 89.52(2.97) \\ \hline
OMTLLOG & 4.81(0.36) & 95.73(0.32) & 7.58(0.65) & 95.36(0.39) & 2.91(0.18) & 97.41(0.16) & 7.16(0.53) & 92.87(0.51) \\ \hline \hline
\textbf{ROMCO-NuCl} & 4.12(0.50) & 96.34(0.44) & 7.06(0.49) & 95.68(0.30) & 2.87(0.42) & 97.43(0.38) & 6.85(0.68) & 93.23(0.66) \\ \hline
\textbf{ROMCO-LogD} & \textbf{4.00(0.42)} & \textbf{96.45(0.37)} & 7.31(0.45) & 95.55(0.27) & \textbf{2.74(0.16)} & \textbf{97.56(0.14)} & \textbf{6.68(0.44)} & \textbf{93.40(0.43)} \\ \hline
\end{tabular}
\end{table*}
\vspace{-0.05in}
\subsubsection{Evaluation Metric}
We evaluated the performance of the aforementioned algorithms by two metrics:
1) Cumulative error rate: the ratio of predicted errors over a sequence of instances.
It reflects the prediction accuracy of online learners.
2) F1-measure: the harmonic mean of precision and recall.
It is suitable for evaluating the learner's performance on class-imbalanced datasets.
We followed the method of \cite{yang2015aggressive} by randomly shuffling the ordering of samples for each dataset and repeating the experiment 10 times with new shuffles. The average results and its standard deviation are reported below.
\begin{table}[t]
\centering
\caption{Run-time (in seconds) for each algorithm}
\label{running_time}
\scriptsize
\begin{tabular}[2.1\textwidth]{|c|c|c|c|c|} \hline
Algorithm & Spam Email & MHC-I & EachMovie \\ \hline
TRML & 73.553 & 361.42 & 391.218 \\ \hline
MTFL & 78.012 & 198.90 & 302.170 \\ \hline
PA-Global & 0.423 & 1.79 & 23.337 \\ \hline
PA-Unique & 0.340 & 1.53 & 26.681 \\ \hline
CW-COL & 0.86 & 4.35 & 31.002 \\ \hline
OMTL & 26.428 & 40.314 & 85.105 \\ \hline
OSMTL-e & 0.360 & 2.327 & 11.43 \\ \hline
OMTLVON & 1.230 & 1.785 & 21.586 \\ \hline
OMTLLOG & 1.145 & 1.371 & 20.232 \\ \hline \hline
\textbf{ROMCO-NuCl} & 11.49 & 4.88 & 33.235 \\ \hline
\textbf{ROMCO-LogD} & 10.59 & 5.352 & 28.716 \\ \hline
\end{tabular}
\label{tab3}
\end{table}
\vspace{-0.01in}
\subsection{Spam Email Filtering}
\vspace{-0.05in}
\begin{figure*}[t]
\centering
\caption{Cumulative error rate on the Email Spam dataset along the entire online learning process}
\label{EmailSpam-figure}
\subfigure {\includegraphics[width=0.2435\textwidth,height=4.4cm]{Email-Spam-Error-logD-T1-new-1.eps}}
\subfigure {\includegraphics[width=0.2435\textwidth,height=4.4cm]{Email-Spam-Error-logD-T2-new-1.eps}}
\hfil
\subfigure {\includegraphics[width=0.2435\textwidth,height=4.4cm]{Email-Spam-Error-logD-T3-new-1.eps}}
\subfigure {\includegraphics[width=0.2435\textwidth,height=4.4cm]{Email-Spam-Error-logD-T4-new-1.eps}}
\end{figure*}
We applied online multi-task learning to build effective personalized spam filters.
The task is to classify each new incoming email massage into two categories: \emph{legitimate} or \emph{spam}.
We used a dataset hosted by the Internet Content Filtering Group.
The dataset contains 7068 emails collected from mailboxes of four users (denoted by \emph{user1}, \emph{user2}, \emph{user3}, \emph{user4}). Basically, the set of all emails received by a user was not specifically generated for that user.
However, the characteristic of each user's email could be said to match user's interest.
Each mail entry was converted to a word document vector using the TF-IDF (term frequency-inverse document frequency) representation.
Since the email dataset had no time-stamp, each email list was shuffled into a random sequence. The cumulative error rate and
F1-measure results of 10 shuffles are listed in Table \ref{email_comparison_result}. In addition, the cumulative error rate for the four specific users along the learning process is presented in Fig. \ref{EmailSpam-figure}. We also report each algorithm's run-time, i.e., the time consumed by both training and test phase during the complete online learning process in Table \ref{running_time}. From these results, we can make several observations.
First, the proposed ROMCO-NuCl/LogD outperform other online learners in terms of the error rate and F1-measure.
In particular, in accordance to the results from the four specific users, learning tasks collaboratively with both the common and personalized structures consistently beats both the global model and the personalized model.
Second, the performance of the proposed online multi-task learning methods are better than that of the two batch learning algorithms (MTFL and TRML). It should be noted that compared to online learners which update models based only on the current
samples, batch learning methods have the advantage of keeping a substantial amount of recent training samples, at the cost
of storage space and higher complexity. In fact, the proposed ROMCO-NuCL/LogD are more efficient than the batch incremental methods,
e.g., it could be more than 100 times faster than batch MTFL in large-sized dataset (28.72 secs versus 302.17 secs in EachMovie as shown in Table \ref{running_time}). ROMCO-NuCL/LogD do not store recent training samples. They only use the current training sample and a simple rule to update the model. In contrast, batch learning algorithms need to keep a certain number of recent training samples in memory,
learning to extra burden on storage and complexity. In addition, both MTFL and TRML need to solve an optimization problem in
an iterative manner. For practical applications involving hundreds of millions of users and features, the batch learning algorithms are no longer feasible, while online learners remain highly efficient and scalable.
We also observed that ROMCO-NuCL/LogD are slightly slower than CW-COL, OMTLVON/LOG and OSMTL-e. This is expected as ROMCO-NuCL/LogD have to update two component weight matrices. However, the extra computational cost is worth considering the significant improvement over the two measurements achieved by using the two components.
\vspace{-0.05in}
\subsection{MHC-I Binding Prediction}
\begin{table}[t]
\centering
\caption{Cumulative error rate (\%) and F1-measure (\%) and its Standard Deviation in the parenthesis on the MHC-I Dataset Results over 12 Tasks}
\label{Bio-table}
\scriptsize
\begin{tabular}[2.1\textwidth] {|c|c|c|c|}
\hline
\multirow{2}{*}{Algorithm} & \multirow{2}{*}{Error Rate} & Positive Class & Negative Class \\
& & F1-measure & F1-measure \\
\hline\hline
MTFL & 43.84(6.05) & 51.04(10.12) & 59.12(7.35) \\ \hline
TRML & 44.26(5.98) & 50.50(9.97) & 58.80(7.40) \\ \hline
PA-Global & 44.70(2.68) & 45.44(9.86) & 61.28(3.17) \\ \hline
PA-Unqiue & 41.62(3.95) & 51.08(10.23) & 63.02(2.89) \\ \hline
CW-COL & 41.32(4.46) & 50.89(10.70) & 63.59(3.08) \\ \hline
OMTL & 41.56(3.97) & 51.13(10.25) & 63.08(2.89) \\ \hline
OSMTL-e & 42.78(0.65) & 50.48(0.43) & 61.59(0.96) \\ \hline
OMTLVON & 38.13(5.03) & 54.73(10.44) & 66.39(4.02) \\ \hline
OMTLLOG & 38.08(5.16) & 54.83(10.56) & 66.40(4.13) \\ \hline \hline
\textbf{ROMCO-NuCl} & 38.09(5.32) & 55.03(10.53) & 66.34(4.09) \\ \hline
\textbf{ROMCO-LogD} & \textbf{37.91(4.97)} & \textbf{55.09(10.10)} & \textbf{66.55(4.14)} \\ \hline
\end{tabular}
\end{table}
\begin{figure*}[t]
\centering
\caption{Cumulative error rate on the 12 tasks of MHC-I dataset along the entire online learning process}
\label{Bio-figure}
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.4cm]{MHC-I-1-logD-new-1.eps}}
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.4cm]{MHC-I-2-logD-new-1.eps}}
\hfil
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.4cm]{MHC-I-3-logD-new-1.eps}}
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.4cm]{MHC-I-4-logD-new-1.eps}}
\hfil
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.4cm]{MHC-I-5-logD-new-1.eps}}
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.4cm]{MHC-I-6-logD-new-1.eps}}
\hfil
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.4cm]{MHC-I-7-logD-new-1.eps}}
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.4cm]{MHC-I-8-logD-new-1.eps}}
\hfil
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.4cm]{MHC-I-9-logD-new-1.eps}}
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.4cm]{MHC-I-10-logD-new-1.eps}}
\hfil
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.4cm]{MHC-I-11-logD-new-1.eps}}
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.4cm]{MHC-I-12-logD-new-1.eps}}
\end{figure*}
Computational methods have been widely used in bioinformatics to build models to infer properties from biological data~\cite{yang2014ensemble,yang2014ldsplit}. In this experiment, we evaluated several methods to predict peptide binding to human MHC (major histocom-patibility complex) class I molecules. It is known that peptide binding to human MHC-I molecules plays a crucial role in the immune system. The prediction of such binding has valuable application in vaccine designs, the diagnosis and treatment of cancer, etc. Recent
work has demonstrated that there exists common information between related molecules (alleles) and such information can be
leveraged to improve the peptide MHC-I binding prediction.
We used a binary-labeled MHC-I dataset.
The data consists of 18664 peptide sequences for 12 human MHC-I molecules.
Each peptide sequence was converted to a 400 dimensional feature vector following~\cite{LiCHLJ11}.
The goal is to determine whether a peptide sequence (instance) binds to a MHC-I molecule (task) or not, i.e., \emph{binder} or \emph{non-binder}.
We reported the average cumulative error rate and F1-measure of 12 tasks in Table \ref{Bio-table}. To make a clear comparison
between the proposed ROMCO-NuCL/LogD and baselines, we showed the variation of their cumulative error rate along the entire online learning process averaged over the 10 runs in Fig. \ref{Bio-figure}.
From these results, we first observed that the permutations of the dataset have little influence on the performance of each method,
as indicated by the small standard deviation values in Table \ref{Bio-table}. Note that the majority of the dataset belongs to the negative class, thus predicting more
examples as the majority class decreases the overall error rate, but also degrades the accuracy of the minority positive class.
The consistently good performance achieved by the proposed ROMCO-NuCL/LogD in terms of the error rate and F1-measures of both classes further demonstrates effectiveness of our algorithms over imbalanced datasets. Moreover, among the six online models, learning related tasks jointly still achieves better performance than learning the tasks individually, as shown by the improvement of ROMCO and OMTL models over the PA-Unique model.
\begin{figure}[t]
\centering
\caption{Average error rate and F1-measure on the EachMovie dataset over 30 tasks along the entire online learning process}\label{EachMovie_figure}
\subfigure {\includegraphics[width=0.24\textwidth,height=4.5cm]{Movie-Error-logD-new-1.eps}}
\subfigure {\includegraphics[width=0.24\textwidth,height=4.5cm]{Movie-F-measure-logD-new-1.eps}}
\end{figure}
\begin{table}[t]
\centering
\caption{Average Cumulative Error rate (\%) and F1-measure(\%) with their standard deviation in the parenthesis over 30 tasks of EachMovie Dataset Results}
\label{EachMovie-table}
\scriptsize
\begin{tabular}[2.1\textwidth] {|c|c|c|c|}
\hline
\multirow{2}{*}{Algorithm} & \multirow{2}{*}{Error Rate} & Positive Class & Negative Class \\
& & F1-measure & F1-measure \\
\hline\hline
MTFL & 27.51(12.25) & 79.18(12.87) & 36.06(14.85) \\ \hline
TRML & 26.58(11.82) & 79.89(12.49) & 37.64(15.05) \\ \hline
PA-Global & 31.80(5.87) & 74.43(8.61) & 47.96(14.47) \\ \hline
PA-Unqiue & 19.68(7.39) & 82.97(9.35) & 57.80(21.05) \\ \hline
CW-COL & 25.45(6.96) & 78.89(9.30) & 53.95(16.71) \\ \hline
OMTL & 19.44(7.28) & 83.18(9.29) & 57.77(21.39) \\ \hline
OSMTL-e & 20.73(7.15) & 82.31(9.04) & \textbf{58.76(18.71)} \\ \hline
OMTLVON & 18.61(7.29) & 84.45(8.64) & 55.92(23.94) \\ \hline
OMTLLOG & 18.61(7.29) & 84.45(8.64) & 55.92(23.94) \\ \hline \hline
\textbf{ROMCO-NuCl} & 19.14(7.20) & 83.46(9.25) & 58.27(21.17) \\ \hline
\textbf{ROMCO-LogD} & \textbf{18.21(6.71)} & \textbf{84.63(8.41)} & 55.53(25.02) \\ \hline
\end{tabular}
\end{table}
\vspace{-0.05in}
\subsection{Movie Recommender System}
\begin{figure*}
\centering
\caption{Sensitivity analysis on the effect of the parameter $\lambda_1$ and $\lambda_2$ in terms of the error rate}\label{Sensitivity_Figure}
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.6cm]{sensitivity_EmailSpam_lemda1_LogD_1.eps}}
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.6cm]{sensitivity_EmailSpam_lemda2_LogD_1.eps}}
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.6cm]{sensitivity_MHC-I_lemda1_LogD_1.eps}}
\subfigure {\includegraphics[width=0.2445\textwidth,height=4.6cm]{sensitivity_MHC-I_lemda2_LogD_1.eps}}
\end{figure*}
In recent years, recommender systems have achieved great success in many real-world applications.
The goal is to predict users' preferences on targeted products, i.e., given partially observed user-movie rating entries
in the underlying (ratings) matrix, we would like to infer their preference for unrated movies.
We used a dataset hosted by the DEC System Research Center that collected the EachMovie recommendation data for 18 months.
During that time, 72916 users entered a total of 2811983 numeric ratings for 1628 different movies.
We randomly selected 30 users (tasks) who viewed exactly 200 movies with their rating as the target classes.
Given each of the 30 users, we then randomly selected 1783 users who viewed the same 200 movies and used their ratings as the features of the movies.
The six possible ratings (i.e., $\{1,\ldots,6\}$) were converted into binary classes (i.e., \emph{like} or \emph{dislike}) based on the rating order.
Table~\ref{EachMovie-table} shows the comparison results in terms of the average cumulative error rate and F1-measure.
Fig.~\ref{EachMovie_figure} depicts the detailed the cumulative error rate along the entire online learning process over the averaged 30 tasks of the EachMovie dataset.
From these results, we can make several conclusions.
First, it can be seen that the proposed ROMCO-NuCl/LogD outperform other baselines: ROMCO-NuCl/LogD always provide smaller error rates and higher F1-measures compared to other baselines.
It shows that our algorithms can maintain a high quality of prediction accuracy.
We believe that the promising result is generally due to two reasons: First, the \emph{personalized} and \emph{correlative} patterns are effective to discover the personalized tasks and task relativeness, and these patterns are successfully captured in three real-world datasets.
Second, once an error occurs in at least one task, ROMCO-NuCl/LogD would update the entire task matrix.
That would benefit other related tasks with few learning instances since the shared subspaces would be updated accordingly.
Next, we observed that ROMCO-LogD is better than ROMCO-NuCl in Fig. ~\ref{EachMovie_figure} in terms of the error rate and F1-measure. This is expected because compared to the nuclear norm, ROMCO-LogD is able to achieve better rank approximation with the log-determinant function, i.e., it reduces the contribution of large singular values while approximating the small singular values into zeros.
\subsection{Effect of the Regularization Parameters}
\vspace{-0.05in}
We used Spam Email and Human MHC-I datasets as the cases for parameter sensitivity analysis.
In the Spam Email dataset, by fixing $\lambda_2=0.0001$ as well as varying the value of $\lambda_1$ in the tuning set, i.e., $[10^{-6},\ldots,10^0]$, we studied how the parameter $\lambda_1$ affects the classification performance of ; by fixing $\lambda_1=0.0001$ as well as varying the value of $\lambda_2$ in tuning set of $[10^{-7},\ldots,10]$, we study how the parameter affects the performance of ROMCO-NuCl/LogD.
Similarly, in the Human MHC-I dataset, we studied the pair of $(\lambda_1,\lambda_2)$ by fixing $\lambda_2=0.0001$ with the tuning set of $\lambda_1$ $[10^{-5},\ldots,10^0]$ and by fixing $\lambda_1=1$ with tuning set of $\lambda_2$ $[10^{-6},\ldots,10^0]$.
In Fig.~\ref{Sensitivity_Figure}, we show the classification performance of ROMCO in terms of the error rate for each pair of $(\lambda_1,\lambda_2)$.
From Fig.~\ref{Sensitivity_Figure}, we observed that the performance is worse with an increment of either $\lambda_1$ or $\lambda_2$ over the Spam Email dataset.
It indicates a weak relativeness among the tasks and many personalized tasks existing in the Email dataset.
In Human MHC-I, the poor performance is triggered by a small value of $\lambda_1$ or a large value of $\lambda_2$.
Compared with the Email data, MHC-I contains fewer personalized tasks, meanwhile most tasks are closely related and well represented by a low-dimensional subspace.
\vspace{-0.1in}
\section{Conclusion}
We proposed an online MTL method that can identify sparse personalized patterns for outlier tasks, meanwhile capture a shared low-rank subspace for correlative tasks. As an online technique, the algorithm can achieve a low prediction error rate via leveraging previously learned knowledge. As a multi-task approach, it can balance this trade-off by learning all the tasks jointly. In addition, we proposed a log-determinant function to approximate the rank of the matrix, which, in turn, achieves better performance than the one with the nuclear norm. We show that it is able to achieve a sub-linear regret bound with respect to the best linear model in hindsight, which can be regarded as a theoretical support for the proposed algorithm. Meanwhile, the empirical results demonstrate that our algorithms outperform other state-of-the-art techniques on three real-world applications. In future work, online active learning could be applied in the MTL scenario in order to save the labelling cost.
\section*{Appendix}
\subsection*{root computation of the Log-Determinant function}
To solve the log-determinant function, we set the derivative of Eq. (\ref{OptimalSigma_log}) for each $\{\sigma_i\}_{i=1}^r \geq 0$ to zero with $\rho = \eta_1\lambda_1$,
\bqs\notag
\frac{1}{\rho}\sigma_i^3 - \frac{1}{\rho}\hat{\sigma}_i\sigma_i^2 + (\frac{1}{\rho} + 2)\sigma_i - \frac{1}{\rho}\hat{\sigma}_i = 0.
\eqs
Assume that $a = \frac{1}{\rho}$, $b = - \frac{1}{\rho}\hat{\sigma}_i$, $c = \frac{1}{\rho} + 2$ and $d = - \frac{1}{\rho}\hat{\sigma}_i$, we define that
$\Delta = \alpha^2 + \beta^3$, where
\bqs\notag
\alpha = \frac{bc}{6a^2} - \frac{b^3}{27a^3} - \frac{d}{2a}, & & \beta = \frac{c}{3a} - \frac{b^2}{9a^2}.
\eqs
Then the three possible roots of the above cubic equation include one real root and two complex roots,
\bqs\notag
\sigma_i^{(1)} = & -\frac{b}{3a} + \sqrt[3]{\alpha + \sqrt{\Delta}} + \sqrt[3]{\alpha - \sqrt{\Delta}}, \\
\sigma_i^{(2)} = & -\frac{b}{3a} + \frac{\sqrt{3}i - 1}{2}\sqrt[3]{\alpha + \sqrt{\Delta}} - \frac{\sqrt{3}i + 1}{2}\sqrt[3]{\alpha - \sqrt{\Delta}}, \\
\sigma_i^{(3)} = & -\frac{b}{3a} - \frac{\sqrt{3}i + 1}{2}\sqrt[3]{\alpha + \sqrt{\Delta}} + \frac{\sqrt{3}i - 1}{2}\sqrt[3]{\alpha - \sqrt{\Delta}}.
\eqs
According to the $\Delta$, three different scenarios of roots are as follows,
\begin{itemize}
\item if $\Delta > 0$, the equation has a real root and a conjugate pair of complex roots
\item if $\Delta = 0$, the equation has three real roots:
\begin{enumerate}
\item when $\alpha^2 = \beta^3 = 0$, there are three equal real roots;
\item when $\alpha^2 = - \beta^3 \neq 0$, there are two equal real roots.
\end{enumerate}
\item if $\Delta < 0$, there are three unequal real roots.
\end{itemize}
{
\bibliographystyle{IEEEtran}
| b492b3b1e586d1433a0d4a80ca8bf14dd95b5466 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The gaseous circumgalactic medium (CGM) around galaxies serves as a
massive reservoir of baryons and metals \citep{peeples2014, werk2014,
keeney2017} and an import driver of galactic evolution. For example,
gas and metals accreted from the CGM may provide the fuel for as much
as half of a massive galaxy's stellar mass by $z\sim0$
\citep[e.g.,][]{oppenheimer2010, ford2014, angles-alcazar2016}. The
CGM includes contributions from a broad range of sources. It includes
material ejected from the central galaxy in the form of winds
\citep[e.g.,][]{weiner2009, tripp2011, shen2013, rubin2014, cafg2015,
muratov2017}, which may ultimately return to the galaxy. The CGM
includes matter removed from satellites as gas stripped via tidal
forces or ram pressure \citep{grcevich2009, spekkens2014, emerick2016}
or gas ejected via the winds of those same satellites
\citep[e.g.,][]{angles-alcazar2016}.\footnote{These are not
necessarily mutually exclusive, as ram pressure may play a role in
removing wind-driven matter from the satellite's halo.} It likely
also includes matter that has accreted from the IGM, either as cold
gas or gas that is heated via its interaction with the existing CGM
\citep{keres2005, fumagalli2011, van-de-voort2011,cafg2011}. Most of
these processes produce (at least initially) relatively dense, cool
concentrations of gas. This is the case even for galactic winds, for
which a significant fraction of the mass ejected may come from cool,
entrained or condensed material \citep[e.g.,][]{heckman1990,
schwartz2004, rupke2005, rubin2014, kacprzak2014}. These cool
streams of matter may ultimately fall onto the central galaxy itself,
perhaps fueling future star formation \citep{keres2005,
oppenheimer2010, ford2014} or dissipate within the hotter, more
diffuse coronal matter in the halo \citep[e.g.,][]{joung2012,
voit2015}. Thus, the observable CGM about galaxies is expected to be
multiphase, with the very diffuse warm and hot gas
\citep[$10^5 \la T \la 10^7$ K;][]{tumlinson2011, prochaska2011,
wakker2009, anderson2013} threaded with denser cool gas
\citep[$10^4 \la T \la 10^5$ K;][]{nielsen2013, kacprzak2013,
werk2014, keeney2017}, including the dense structures that could
represent new matter flowing into the CGM for the first time
\citep{lehner2013, wotta2016, van-de-voort2012}.
In the context of the CGM, such dense ($n_{\rm H} \ga 10^{-3}$ cm$^{-3}$)
and cool ($T < 10^5$ K) structures are ionized entities. They have
characteristic densities and size scales that produce optical depths
at the Lyman limit of order unity and above \citep{schaye2001,
fumagalli2011a, cafg2011, van-de-voort2012}. In QSO absorption line
studies of the CGM, these streams would thus be categorized as Lyman
limit systems \citep[LLSs; ][]{tytler1982,steidel1990}, with \ensuremath{\mbox{\ion{H}{1}}}\
column densities $N(\mbox{\ensuremath{\mbox{\ion{H}{1}}}}) \ge 1.6\times10^{17}$ cm$^{-2}$\
($\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \ge 17.2$), which corresponds to an optical depth at the H
ionization edge of $\tau_{\rm 1 Ryd} \ge 1$
\citep{spitzer1978}. Studies of low-redshift Lyman limit systems have
shown they exhibit a broad range of metallicities, from $<1\%$ solar
to super-solar metallicities \citep{lehner2013, wotta2016}, as
expected if they are tracing a range of origins, including accreting,
stripped, or expelled matter.
Simulations of galaxies at $z\approx2$ show that optically-thick \ensuremath{\mbox{\ion{H}{1}}}\
material may cover $\sim2\% - 10\%$ of the area within the virial
radius of typical galaxies \citep[e.g.,][]{fumagalli2011a, cafg2011},
with a signicantly higher covering fraction at the smallest impact
parameters. Only recently have simulations been able to regularly
study the high-density structures in the CGM to $z=0$ with full
feedback prescriptions. The covering factors of optically thick gas in
simulated galaxies at low redshift tend to be lower on average than
their higher-redshift counterparts, partly due to the general
dimunition expected in cold accretion as well as generally diminished
feedback. For example, \citealt{hafen2017} (hereafter
\citetalias{hafen2017}) find a median $f_c \sim 1\%$ for
$\log M_{\rm halo}/\ensuremath{M_\odot}\relax \approx 12$ galaxies at $z=0$. Still, the
covering factors for individual galaxies can be as high as
$f_c \sim 10\% - 15\%$, depending on the feedback recipes and the
histories of recent star formation, mergers, and accretion
\citep[][\citetalias{hafen2017}]{gutcke2016}. Measurements of
optically-thick \ensuremath{\mbox{\ion{H}{1}}}\ thus constrain the total mass in the densest
material associated flows of matter through the CGM while also
constraining prescriptions for feedback and other processes in galaxy
simulations.
The study of galactic halos at $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \sim 17-18$ has largely been
limited to absorption line measurements. However, the the best radio
observatories now reach sensitivities to \ensuremath{\mbox{\ion{H}{1}}}\ 21-cm emission that get
nearly to $\tau_{\rm 1 \, Ryd} \approx 1$ gas around low-redshift
galaxies \citep[e.g.,][]{braun2004, lockman2012, wolfe2013,
wolfe2016}, bridging the gap between emission and absorption
experiments. Radio searches for faint emission have the distinct
advantage of (potentially at least) providing more spatial information
than can be gleaned from the small numbers of skewers that are
available for absorption line spectroscopy, with its need for bright
background sources (especially in the ultraviolet).
Here we use the Robert C. Byrd Green Bank Telescope (GBT), part of the
Green Bank Observatory,\footnote{The Green Bank Observatory is a
facility of the National Science foundation operated under a
cooperative agreement by Associated Universities, Inc.} to search
for 21-cm emission from the halo of the nearby Andromeda galaxy. This
21-cm search is in support of our Project AMIGA (Absorption Maps In
the Gas of Andromeda)\footnote{Not to be confused with the AMIGA
survey (Analysis of the interstellar Medium in Isolated GAlaxies, PI
Lourdes Verdes-Montenegro); {\tt http://amiga.iaa.es}.}, a large
{\em Hubble Space Telescope} ({\em HST}) program to study the absorption
from the extended CGM about the Andromeda galaxy (M31) using the
Cosmic Origins Spectrograph (COS). Project AMIGA will measure UV
absorption lines along $\sim$25 AGN sight lines passing within
$\rho \approx \ensuremath{R_{\rm vir}}\relax \approx 300$ kpc of the center of M31. This builds
on our recent study demonstrating that M31 has a massive, diffuse CGM
stretching to approximately \ensuremath{R_{\rm vir}}\relax\ \citep[][hereafter
\citetalias{lehner2015}]{lehner2015}. Project AMIGA will characterize
the radial and azimuthal dependences of metal ion surface densities
(for \ion{C}{2}, \ion{C}{4}, \ion{Si}{2} \ion{Si}{3}, \ion{Si}{4}, and
others) in M31's CGM. However, our {\em HST}\ observations will provide
little information on the \ensuremath{\mbox{\ion{H}{1}}}\ distribution at the velocities where
\citetalias{lehner2015} find metal absorption from M31's CGM
($\ensuremath{v_{\rm LSR}\relax} \ga -500$ km~s$^{-1}$), as the strongly-saturated Ly$\alpha$\relax\ absorption
from the Milky Way completely blocks the light from the background
AGNs over these wavelengths. This is not a problem with \ensuremath{\mbox{\ion{H}{1}}}\ emission,
where the Milky Way emission is largely constrained to higher
velocities than M31's CGM.
In the present work we target 48\ directions through the
Andromeda galaxy's halo in order to measure the covering factor of
highly optically-thick gas ($\tau_{\rm 1\, Ryd} \approx 2.5$). Our
observations achieve sensitivities to \ensuremath{\mbox{\ion{H}{1}}}\ column densities that
overlap with absorption line measurements of CGM \ensuremath{\mbox{\ion{H}{1}}}\ about more
distant galaxies \citep[e.g.,][]{tumlinson2013, lehner2013,
prochaska2017} and UV absorption line studies of high-velocity
clouds in the CGM of our own Milky Way \citep[e.g.,][]{fox2006}. As
the 21-cm line is so easy to excite (the critical density for
excitation is $n_{\rm crit} \approx 3\times10^{-5}$ cm$^{-3}$ , well
below the density of gas we are considering, and Ly$\alpha$ pumping is
also an effective excitation mechanism), these GBT observations allow
us to trace gas even in conditions expected of the CGM.
There is good reason to suspect that M31 may have a significant CGM,
even one detectable in 21-cm emission. The relative dearth of neutral
gas associated with dwarf satellite galaxies has led several groups to
to argue for the existence of a low-density diffuse CGM about M31 to
impact parameters $\rho \la 270$ kpc \citep[][ see also
\citealt{spekkens2014} for similar work specific to the Milky
Way]{blitz2000, grcevich2009}. Similarly, although the giant southern
stream has a dynamical age of $<1$ Gyr \citep{fardal2008}, its star
formation was shut off well before it was tidally disrupted
\citep[$\ga 4$ Gyr ago;][]{brown2006}, suggesting its gas was stripped
by M31's CGM long ago. \citeauthor{grcevich2009} used ram pressure
stripping arguments to imply a density of a few $\times10^{-4}$
cm$^{-3}$\ for a diffuse (potentially hot) corona about M31.
\citet{rao2013} searched for metal line absorption from the halo of
M31, but their sensitivity was limited by the low spectral resolution
of their data. \citetalias{lehner2015} found significant metal ion
surface densities for sight lines projected within
$\rho \approx \ensuremath{R_{\rm vir}}\relax \approx 300$ kpc of M31, suggesting an extended
distribution of low-density gas. The total CGM gas mass implied by
the \citetalias{lehner2015} observations is significant at
$M_{\rm total} \ga 2\times10^8\, (Z/Z_\odot)^{-1}$ \ensuremath{M_\odot}\relax\ within
$0.2\ensuremath{R_{\rm vir}}\relax$, although the total mass could be a factor of 10 larger if
the covering factor is near unity to the virial radius.\footnote{The
results initially given in \citetalias{lehner2015} are too high by a
factor of $\sim10$ due to an unfortunate error.}
The \ensuremath{\mbox{\ion{H}{1}}}\ emission from the diffuse CGM implied by these studies may be
minimal, depending on the metallicity and ionization state of the
gas. In the Andromeda system, \citet{braun2004} reported the detection
of a diffuse ``bridge'' of low column density material stretching
between M31 and its companion M33 (projected $\sim200$ kpc from M31),
with an extension several degrees beyond M31 in the north
($1^\circ \approx 13$ kpc at M31). In total \citeauthor{braun2004}
found the filament stretched nearly 260 kpc. This work was done with
a coarse angular resolution of $49\arcmin$ and velocity resolution of
18 km~s$^{-1}$\ to achieve a sensitivity of $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \approx 17$ for emission
filling the beam, and much of the bridge emission appeared to have
beam-averaged column densities well below $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \sim
18$. Subsequent high-resolution maps of the vicinity of M33 have
called into questioned the presence of this bridge, given the lack of
a bridge component in the maps of M33
\citep{putman2009}. \citet{lockman2012} confirmed the existence of
faint 21-cm emission in this region between the two galaxies, but they
noted the gas must be very patchy, as they did not find it
consistently over several GBT pointings. \citet{wolfe2013} and
\citet{wolfe2016} showed that most of the emission identified by
\citet{braun2004} was associated with higher column density,
small-scale clumps of gas. The ``bridge'' appears to be constituted of
small clouds (each with masses $\sim10^5$ \ensuremath{M_\odot}\relax) that appeared as more
continuous, diffuse structure when diluted within the Westerbork beam
employed by the \citeauthor{braun2004} study. And while
\citet{wolfe2016} did detect some emission in the extended structure
found by \citet{braun2004} to the northwest of M31, the spatial
distribution of this gas appears to be much different than that
implied by the initial maps. Thus, M31's CGM appears to harbor at
least some small-scale concentrations of \ensuremath{\mbox{\ion{H}{1}}}\ detectable through their
21-cm emission.
In the inner regions of M31's halo, \citet{thilker2004} identified a
population of high-velocity clouds (HVCs) projected within
$\rho \sim 50$ kpc of the center of M31, which were subsequently
mapped at higher resolution by \citet{westmeier2005}. These clouds
individually have \ensuremath{\mbox{\ion{H}{1}}}\ masses $M_{\rm HI} \approx 10^5 - 10^6$ \ensuremath{M_\odot}\relax\
with size scales of $\sim1$ kpc and central column densities
$\ensuremath{N(\mbox{\ion{H}{1}})}\relax \sim10^{19}$ cm$^{-2}$\ \citep{westmeier2005}. They are discrete
structures, dense clouds ($n_{\rm HI} \approx 10^{-2}$ cm$^{-3}$) in the
CGM that may be similar to the clouds identified at larger distances
by \citet{wolfe2016}, though with higher masses.
In the present experiment (Project AMIGA GBT), we survey sight lines
toward UV-bright AGNs with impact parameters $\rho \la \ensuremath{R_{\rm vir}}\relax$ from M31
in order to characterize the distribution of dense, cool gas in the
CGM. In no cases have the observed directions been chosen with prior
knowledge of the local \ensuremath{\mbox{\ion{H}{1}}}\ content. Thus, we can use our observations
to directly determine the covering factor of optically-thick gas in
the CGM of M31. Our covering factor estimates are complementary to the
\citet{richter2012} assessment of the covering factor of HVCs about
M31 based on the \citet{thilker2004} maps (which have a $5\sigma$
column density sensitivity of $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \sim18.25$ per GBT beam for a
FWHM$ = 25$ km~s$^{-1}$, significantly higher than our
observations).\footnote{\citet{thilker2004} discuss several variations
of their datacube, each smoothed to different velocity and spatial
resolutions. Here we give the limits associated with their
higher-resolution data, with a $\approx13\farcm65$ beam and 18 km~s$^{-1}$\
velocity channels. Their limits imply a mass sensitivity of
$M_{\ensuremath{\mbox{\ion{H}{1}}}} \le 3.5\times 10^{4}$ \ensuremath{M_\odot}\relax\ per beam.}
Our paper is presented as follows. We summarize the GBT observations
and data reduction in \S \ref{sec:observations}. We derive covering
factors for cool gas about the Andromeda galaxy in \S
\ref{sec:coveringfactor}. In \S \ref{sec:comparisons} we consider the
covering factor results in the context of recent QSO absorption line
measurements and numerical simulations. Our discussion and summary
follow in \S \ref{sec:discussion} and \S \ref{sec:summary}. A
metallicity measurement for a sight line passing through the nearby
Magellanic Stream (MS) is included in an Appendix.
Throughout we assume a distance of 752\ kpc to M31
\citep{riess2012}. For comparison with other galaxies (simulated and
observed), we assume for M31 a stellar mass
$\log M_*/\ensuremath{M_\odot}\relax \approx 11.0$ and a halo mass
$\log M_{\rm h}/\ensuremath{M_\odot}\relax \approx 12.0$ \citep{tamm2012}.
\section{GBT 21-cm Observations and Analysis}
\label{sec:observations}
We used the 100-m diameter Robert C. Byrd Green Bank Telescope
\citep{prestage2009} to take pointed observations toward 48\
AGNs projected behind the CGM of M31 as part of NRAO programs
GBT15A-328 and GBT14B-436. A map showing the locations of the AGNs is
given in Figure \ref{fig:map}, and the details of the AGN sight lines
are summarized in Table \ref{tab:targets}. These sight lines represent
two samples that, together, should provide an unbiased sampling of the
\ensuremath{\mbox{\ion{H}{1}}}\ content of the Andromeda galaxy's gaseous halo. The first sample
consists of 25 objects in the Project AMIGA sample, shown with red
outlines in Figure \ref{fig:map}. These are selected to be UV-bright
AGNs projected within $\rho \approx 340$ kpc ($\sim 1.1 \ensuremath{R_{\rm vir}}\relax$) of the
center of M31. They are chosen to probe a broad range of impact
parameters, with sight lines focused about the major axis, minor axis,
and intermediate orientations. They do not sample the impact parameter
space randomly (i.e., with weighting $\propto \rho^2$). Furthermore,
they do not randomly sample the azimuthal distribution, both because
of the goals of Project AMIGA (to probe the azimuthal variations
systematically) and of a general dearth of identified UV-bright AGNs
behind the northern half of M31's CGM \citepalias[see][]{lehner2015}.
\begin{figure}
\epsscale{1.2}
\plotone{fig-GBTmap.pdf}
\caption{The locations of our GBT pointings relative to the M31-M33
system, where the axes show physical impact parameter from the
center of M31. In this orientation, north is up, east to the left.
The 25 Project AMIGA sight lines have red outlines; the supplemental
sight lines have black outlines and are smaller. Two probable MS
detections are filled in cyan. The remainder of the sight lines have
non-detections of \ensuremath{\mbox{\ion{H}{1}}}\ associated with M31. Dotted circles show
impact parameters $\rho = 100$, 200, 300 kpc, the last being roughly
\ensuremath{R_{\rm vir}}\relax. The sizes and orientations of the two galaxies are taken from
the RC3 \citep{de-vaucouleurs1991} and correspond to the optical
$R_{25}$ values. The dashed line shows the plane of the Magellanic
System ($b_{\rm MS} = 0^\circ$) as defined by
\cite{nidever2008}. The region within $b_{\rm MS} = \pm15^\circ$ is
shaded. \label{fig:map}}
\end{figure}
The remaining 23 pointings (shown with black outlines in Figure
\ref{fig:map}) sample sight lines toward UV-bright AGNs (and one blank
sky direction) that are not scheduled for {\em HST}\ observations. These
supplemental sight lines were chosen because they may eventually
represent quality target for UV follow-up, and because they provide an
expansion of our survey in a way that mimics the initial selection. In
both cases, the principal criterion for selecting a sight line is the
presence of a UV-bright background source. This selection is
independent of the \ensuremath{\mbox{\ion{H}{1}}}\ distribution, and we used no knowledge of the
\ensuremath{\mbox{\ion{H}{1}}}\ structure of M31 in selecting background sources. There is a bias
in the distribution of the sources to the southern half of Andromeda's
CGM. This is in part due to the increased extinction northward of M31
(at lower Galactic latitudes) and a lack of surveys at those lower
latitudes \citep[a result being rectified by QSO searches with, e.g.,
LAMOST; ][]{huo2013}. We note that if the goal of our program was simply
to derive information about the covering factor of low column density
\ensuremath{\mbox{\ion{H}{1}}}\ in the vicinity of M31, we would not need to search in directions
toward background AGNs. However, the present program was designed with
the UV observations (or potential future observations) in mind, so
that deep \ensuremath{\mbox{\ion{H}{1}}}\ observations might help constrain the metallicity of
the CGM viewed in UV absorption.
The pointed GBT observations were taken with the L-band receiver and
have a \ensuremath{9\farcm1}\ primary beam. Each AGN direction was initially
observed for a total of 50--60 minutes in individual 10 minute
scans. In a few cases difficulties with the receivers caused one of
the polarizations to be unusable. We received additional director's
discretionary time (program GBT16A-433) to reobserve 12 of the sight
lines. Our observations employed a 5.16 MHz bandwidth centered on the
\ensuremath{\mbox{\ion{H}{1}}}\ hyperfine transition and have a native channel spacing of 0.15
km~s$^{-1}$ . The typical system temperature of the GBT in this configuration
is 18 K at zenith in both linear polarizations. Our observations are
consistent with this expectation, with most values in the range
$17.3 \la T_{\rm sys} \la 19.0$ K during our observations. The
exceptions are those observations taken toward radio-bright AGNs 3C48
(45 K) and 3C59 (21.9 K) as well as the direction toward 3C66A (24.6
K). None of the other background AGNs is within a factor of 10 of the
brightness of 3C48 and 3C59 (which have 1.4 GHz fluxes of $\sim 16$
and 2 Jy). In-band frequency switching (with a throw of 3 MHz) was
used for background subtraction. This provided useable velocity
coverage of typically $-515 \la \ensuremath{v_{\rm LSR}\relax} \la +470$ km~s$^{-1}$.
We performed the data reduction within GBTIDL \citep{marganian2006},
following \citet{boothroyd2011} for the basic brightness temperature
calibration and stray radiation correction and \citet{lockman2012} for
the scan coaddition and baseline fitting. For each sight line we
individually examined each 10 minute scan, including both
polarizations. In some cases we interpolated over occasional localized
interference, i.e., for interference affecting a small number of
channels in a given scan. In a few cases, individual scans were
excluded from the coaddition if such interference occupied a large
number of channels. The quality of the spectral baselines is one of
the limiting factors in setting our column density sensitivity. For
some scans, one of the linear polarized receivers had much worse
baselines than typical. We excluded these from the subsequent data
processing. We fitted spectral baselines separately to each scan and
linear polarization over a very broad velocity range before
coaddition. We adopted third- to fifth-order polynomials, taking care
to exclude regions of potential emission from the fitting. The
individual baseline-corrected spectra for a given sight line were then
coadded with equal weights after applying an atmospheric extinction
correction to each. Finally we corrected for the GBT's main beam
efficiency at 21-cm ($\eta_{\rm mb} = 0.88$). The final data are
binned to $\sim0.6$ km~s$^{-1}$\ channel width. Several examples of our final
spectra are shown in Figure \ref{fig:spectra}.
We have performed an automated search for emission at $\ge 5\sigma$
significance over the velocity range $-515 \le \ensuremath{v_{\rm LSR}\relax} \le -170$ km~s$^{-1}$\
(M31 has a systemic velocity $v_{\rm sys} = -300$ km~s$^{-1}$ ). Our
observations have typical RMS brightness temperature fluctuations of
$\sigma_{\rm b} \approx 8$ mK over the search velocities, with a full
range between $\approx 7$ and 12 mK per 0.6 km~s$^{-1}$\ channel. The RMS
brightness temperatures for each sight line, derived empirically over
the full range of velocities searched for M31 emission, are given in
Table \ref{tab:targets}. Because we have calculated these empirically
they include both the random noise and the effects of imperfect
baseline subtractions or local baseline irregularities.
For each sight line we calculate detection limits for the \ensuremath{\mbox{\ion{H}{1}}}\ column
density assuming the optically thin approximation:
$\ensuremath{N(\mbox{\ion{H}{1}})}\relax = [1.8 \times 10^{18} \ \mbox{cm$^{-2}$\ (K km~s$^{-1}$)}^{-1}] \ \int T_b
\, dv$. Any \ensuremath{\mbox{\ion{H}{1}}}\ in the CGM of M31 will be optically thin in the 21-cm
line. The opacity of the 21-cm line is
$\tau = (5.2 \times 10^{-19} \ \mbox{K km~s$^{-1}$\ cm$^2$}) \, \ensuremath{N(\mbox{\ion{H}{1}})}\relax \, {\rm
FWHM}^{-1} \, T_{ex}^{-1}$ \citep{dickey1990}, where the excitation
temperature is generally close to the kinetic temperature of the \ensuremath{\mbox{\ion{H}{1}}}.
Thus for the expected \ensuremath{\mbox{\ion{H}{1}}}\ column densities $\approx 10^{18}$ cm$^{-2}$,
FWHM$\, = 25$ km~s$^{-1}$, and $T_{ex} > 1000$ K, the peak optical depth will
be $\tau < 2 \times 10^{-5}$. Although the 21-cm measurements were
made toward AGN, these almost always contribute a negligible amount of
radio continuum emission. Even for the exception, 3C48, which has a
continuum antenna temperature $\sim30$ K at the GBT, the continuum
temperature is much less than the 21-cm excitation temperature. It
would reduce the 21-cm emission line by $<1$ mK for the expected 21-cm
opacities, a signal that is well below the noise level.
Our detection limits are given in Table \ref{tab:targets} based on the
RMS brightness temperature fluctuations and an assumed line width of
25 km~s$^{-1}$\ (FWHM). This yields typical 5$\sigma$ detection limits of
$\ensuremath{N(\mbox{\ion{H}{1}})}\relax \le 4\times10^{17}$ cm$^{-2}$\ per beam or better. Our choice of
FWHM is based on the median FWHM in detections of M31 HVC and clump
emission from the studies of \cite{thilker2004},
\citet{westmeier2008}, \citet{lockman2012}, and \citet{wolfe2016}.
Our typical sensitivity is equivalent to an \ensuremath{\mbox{\ion{H}{1}}}\ mass of
$M_{\rm HI} \sim 10^4$ \ensuremath{M_\odot}\relax\ per beam ($\sim 800$
\ensuremath{M_\odot}\relax$/{\rm kpc}^2$), no matter the scale of the emission.
\begin{figure}
\epsscale{0.75}
\plotone{fig-spectra.pdf}
\caption{Four spectra from our GBT observations covering the velocity
range $-515 \le \ensuremath{v_{\rm LSR}\relax} \le -170$ km~s$^{-1}$\ (sight line identifications
are given in the upper left of each panel). The top two spectra are
characteristic examples of the vast majority of the observations,
showing no detectable \ensuremath{\mbox{\ion{H}{1}}}\ 21-cm emission from the CGM of M31. The
bottom two spectra show the \ensuremath{\mbox{\ion{H}{1}}}\ emission observed toward RBS~2055
and RBS~2061 (properties of the lines are given in Table
\ref{tab:MSdetections}). All spectra are shown Hanning-smooted to
$\Delta v\approx1.2$ km~s$^{-1}$. Limits are determined using the RMS
brightness temperature variations across the uncontaminated spectral
range of the observations and assuming FWHM$\, = 25$
km~s$^{-1}$. \label{fig:spectra}}
\end{figure}
Toward the direction of M31's extended halo, \ensuremath{\mbox{\ion{H}{1}}}\ emission from gas
associated with the Milky Way and its HVCs, the Magellanic Stream
(MS), and M31 can overlap in velocity, as discussed in detail by
several works \citepalias[\citealt{thilker2004},
][\citealt{kerp2016}]{lehner2015}. M31 itself has a systemic velocity
in the LSR frame $\ensuremath{v_{\rm LSR}\relax}({\rm M31}) = -300\pm4$ km~s$^{-1}$, and we expect
potential CGM emission within $\sim \pm225$ km~s$^{-1}$\ of the systemic
velocity, given that more than 90\% of the candidate dwarf galaxies in
the M31 system lie within this range \citep{mcconnachie2012}. The
upper velocity limit for our search is set by the likelihood of
contamination by emission from relatively local Galactic HVCs, which
are confined to $\ensuremath{v_{\rm LSR}\relax} \ga -170$ km~s$^{-1}$\
\citep[][\citetalias{lehner2015}]{lehner2011a,lehner2012}. The lower
velocity limit corresponds to the limits of the M31 HVC emission found
by \citet{thilker2004}. None of the emission detected in high-fidelity
GBT individual pointings or maps \citep{thilker2004, lockman2012,
wolfe2013, wolfe2016} have found emission at velocities more
negative than $\ensuremath{v_{\rm LSR}\relax} = -515$ km~s$^{-1}$; neither have absorption line
searches for gas associated with the Andromeda galaxy
\citepalias[\citealt{rao2013}]{lehner2015}. In addition, most of our
spectra do not give access to velocities more negative than
$\ensuremath{v_{\rm LSR}\relax} \la -515$ km~s$^{-1}$\ given the adopted frequency-switching
offset. At the same time, we can demonstrate that no Ly$\alpha$
absorption exists in the range $-1000 \le \ensuremath{v_{\rm LSR}\relax} \le -600$ km~s$^{-1}$\ toward
the 25 AGNs in the Project AMIGA sample. This is based on a careful
search of the blue edge of the Milky Way Ly$\alpha$ damping
wings. Given the strength of this line, we would detect any gas with
$\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \ga 13.5$. This search will be presented in a future paper.
In addition to potential contamination by Milky Way HVCs, the MS
crosses through this region of the sky at velocities that contaminate
searches for M31 emission, as detailed by \citetalias{lehner2015}
\citep[see also ][]{nidever2008, fox2014}. The emission from the MS is
most important at small Magellanic latitudes ($b_{\rm MS}$), although
metal ion absorption from the MS can be seen to at least
$b_{\rm MS}\approx\pm30^\circ$ \citep{fox2014} and at Magellanic
longitudes $l_{\rm MS} \ga -110\degr$. We show in Figure \ref{fig:map}
the position of $b_{\rm MS} = 0\degr$ as a dashed curve, and we shade
the region between $b_{\rm MS} = \pm15\degr$. The MS velocities can be
predicted following \citet{nidever2008}, as described in detail by
\citetalias{lehner2015} (see their Figure 3). Such contamination is
only important at $\ensuremath{v_{\rm LSR}\relax} \ga v_{\rm sys}({\rm M31}) \approx -300$
km~s$^{-1}$\ \citep{de-vaucouleurs1991}.
Significant \ensuremath{\mbox{\ion{H}{1}}}\ emission over $-515 \le \ensuremath{v_{\rm LSR}\relax} \le -170$ km~s$^{-1}$\ is
observed in only 2 of our 48 sightlines (see Figure \ref{fig:spectra})
at $5\sigma$ significance. The sight lines toward RBS~2055 and
RBS~2061 (the filled points in Figure \ref{fig:map}) show \ensuremath{\mbox{\ion{H}{1}}}\
emission at the level of $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \sim 17.8$. The \ensuremath{\mbox{\ion{H}{1}}}\ profiles for
these directions are shown in Figure \ref{fig:spectra}, and the line
properties are summarized in Table \ref{tab:MSdetections}. These
objects lie only $1\fdg2$ apart on the sky; both directions show
emission at $\ensuremath{v_{\rm LSR}\relax} \approx -336$ km~s$^{-1}$. RBS~2055 and RBS~2061 have
Magellanic coordinates $(l_{\rm MS}, b_{\rm MS}) = (-113.2, +11.0)$
and $(-112.1, +11.6)$, respectively. Thus, both sight lines lie close
to the great circle on which the Magellanic system resides, and the
detected emission has velocities close to those expected for the
extrapolation of the MS in these directions. Our Project AMIGA COS
observations of the sight line towards RBS~2055 also reveal
significant metal line absorption. In particular, and unique among the
Project AMIGA sight lines, this sight line shows absorption from \ensuremath{\mbox{\ion{O}{1}}}\
$\lambda1302$ at the same velocity as the \ensuremath{\mbox{\ion{H}{1}}}\ emission (as well as
other metal lines). As we show in the Appendix, the comparison of
$N(\mbox{\ensuremath{\mbox{\ion{O}{1}}}}) / \ensuremath{N(\mbox{\ion{H}{1}})}\relax$, which gives a good measure of O/H with very
small if any ionization corrections, yields a metallicity estimate
[O/H]$\, = \ensuremath{-0.95\pm0.12}$. This is consistent with the metallicities
of the main body of the MS derived by \citet{fox2010, fox2013}. This
agreement in metallicity with the general MS, and the location of
these directions relative to the MS, strongly suggests the detected
21-cm emission arises in the extension of the MS across this
region. We will assume these directions probe MS gas at
$\ensuremath{v_{\rm LSR}\relax} \approx -336\pm15$ km~s$^{-1}$\ moving forward. We note that no \ensuremath{\mbox{\ion{H}{1}}}\
emission is detected in these directions outside of those expected for
the MS, although we do find absorption from M31's CGM at velocities
distinct from the MS in the strongest lines covered by the Project
AMIGA COS data (see the Appendix).
\section{\ensuremath{\mbox{\ion{H}{1}}}\ Column Density and Covering Factor Limits}
\label{sec:coveringfactor}
We have in hand a sample of 48\ pointings through the halo of
the Andromeda galaxy at impact parameters $\rho < 1.2 \ensuremath{R_{\rm vir}}\relax$ with
sensitivity to \ensuremath{\mbox{\ion{H}{1}}}\ 21-cm emission of
$\ensuremath{N(\mbox{\ion{H}{1}})}\relax \approx 2 - 4\times10^{17}$ cm$^{-2}$\ per beam. The \ensuremath{N(\mbox{\ion{H}{1}})}\relax\ limits
for each sight line are shown as a function of impact parameter in
Figure \ref{fig:columnrho} and reported in Table \ref{tab:targets}. To
encompass the range of sensitivites, we adopt a survey search
sensitivity of $\ensuremath{N(\mbox{\ion{H}{1}})}\relax \le 4\times10^{17}$ cm$^{-2}$\
($\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \le 17.60$). The $5\sigma$ upper limits are shown for each
sight line, with the exception of the two MS detections discussed in
\S \ref{sec:observations}, for which the detected MS emission column
densities are shown. Outside of the velocity range
$-370 \la \ensuremath{v_{\rm LSR}\relax} \la -300$ km~s$^{-1}$, these two sight lines show no emission
to limits similar to those observed along the other sight lines (see
Table \ref{tab:targets}).
\begin{figure}[b]
\plotone{fig-NHIrho.pdf}
\caption{Observed column densities and limits along each of the
48\ GBT sight lines through the CGM of M31. Upper limits
(red) are given at the $5\sigma$ level assuming a ${\rm FWHM} = 25$
km~s$^{-1}$. They are determined over a velocity range
$-515 \la \ensuremath{v_{\rm LSR}\relax} \la -170$ km~s$^{-1}$. Two detections of apparent MS
emission are shown in cyan. Those two sight lines would have
$\log N({\rm H I}) \la 17.6$ outside of the velocities expected for
the MS ($-370 \la \ensuremath{v_{\rm LSR}\relax} \la -300$ km~s$^{-1}$). Also shown are profiles
from recent simulations of \ensuremath{\mbox{\ion{H}{1}}}\ around massive galaxies. The blue
line shows the median \ensuremath{\mbox{\ion{H}{1}}}\ profile from the $z=0$ simulation of
\citet{liang2016} that best fits the ensemble of low-redshift CGM
observations (their {\tt ALL\_Efb\_e001\_5ESN} run) along with the
95\% confidence range of their values (cyan shading). This
simulation follows a single $\log M_{\rm h}/M_\odot \sim 12$ halo to
$z=0$; while it does fairly well at matching CGM observations, its
stellar component is ``unrealistic'' \citep{liang2016}. The dashed
lines are examples of results from \citet{gutcke2016}, showing the
median values for their $\log M_*/M_\odot = 10.66$ and 10.89
galaxies. \label{fig:columnrho}}
\end{figure}
Clouds with \ensuremath{\mbox{\ion{H}{1}}}\ column densities $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \ge 17.6$ could be present
on scales smaller than the 2 kpc diameter GBT beam and remain
undetected, as our limits apply to the beam-averaged column
density. Indeed, high-resolution observations of HVCs in the inner
regions of the halo ($\rho \la 50$ kpc) by \citet{westmeier2005} have
demonstrated that at least some M31 HVCs have smaller sizes
($\langle D \rangle \sim 1$ kpc) and higher central column densities
($\langle \ensuremath{N(\mbox{\ion{H}{1}})}\relax \rangle \approx 4\times10^{19}$ cm$^{-2}$). Objects like
these specific clouds would be readily detected by our observations on
the basis of their masses ($\langle M_{\rm HI} \rangle \approx 10^5$
\ensuremath{M_\odot}\relax), but there may still be small-scale clouds that lie below the
detection limits of our data. In fact, the photoionization simulations
by \cite{lehner2013} suggested relatively small sizes for the $z\la1$
LLSs in that work (the vast majority of which have
$N(\mbox{\ensuremath{\mbox{\ion{H}{1}}}}) \la 10^{17}$ cm$^{-2}$), with only 11/25 of the systems
they modeled having sizes $>1$ kpc. We will discuss this issue further
below.
Also shown on Figure \ref{fig:columnrho} are predictions from the
recent ``zoom-in'' galaxy simulations. The blue curve shows the median
\ensuremath{\mbox{\ion{H}{1}}}\ profile from the {\tt ALL\_Efb\_e001\_5ESN} of \citet{liang2016},
which they note best fits recent COS observations of the CGM about
$z\la1$ galaxies while producing ``unrealistic'' results for the
stellar component of a typical low-redshift galaxy. The cyan shading
about the line represents the 95\% confidence range of the
measurements drawn from their simulation. Similar results from the
NIHAO simulations of \citet{gutcke2016} are shown in the dashed
lines. These represent the median profiles of their mock \ensuremath{\mbox{\ion{H}{1}}}\
distributions for individual galaxies with stellar masses
$\log M_*/\ensuremath{M_\odot}\relax = 10.66$ and 10.89, those closest to M31. (Here we do
not show the quantiles, but they are similar in magnitude to those
shown for the \citeauthor{liang2016} simulations.) The main point in
this comparison of both the \citeauthor{liang2016} and
\citeauthor{gutcke2016} simulations with our results is that \ensuremath{\mbox{\ion{H}{1}}}\
about a typical simulated galaxy falls off quite rapidly, save for the
filaments and smaller-scale structures tracing flows through the CGM.
The principal result of our work is a calculation of the cumulative
covering factor of \ensuremath{\mbox{\ion{H}{1}}}\ as a function of impact parameter from M31.
We calculate the covering factor from our data assuming a binomial
distribution. We follow \citet{cameron2011} in assessing the
likelihood function for values of the covering factor, \ensuremath{f_c(\le \rho)}\relax, given
the number of detections (successes) against the total sample (number
of sight lines within a given impact
parameter). \citeauthor{cameron2011} demonstrates that the normalized
likelihood function useful for calculating (Bayesian) confidence
intervals on a binomial distribution with a non-informative (uniform)
prior follows a beta distribution. We refer the reader to
\citet{cameron2011} for a detailed justification of this approach,
which is particularly useful for providing robust results in cases of
small samples (particularly important at small impact parameter in our
work).
\begin{figure}
\plotone{fig-coveringFactor.pdf}
\caption{The cumulative covering factors for impact parameters less
than $\rho$, \ensuremath{f_c(\le \rho)}\relax, from our GBT observations. These values are
appropriate for $\log N({\rm H I}) \ga 17.6$
($N(\mbox{\ensuremath{\mbox{\ion{H}{1}}}}) \ga 4\times10^{17}$ cm$^{-2}$). Upper limits are
shown at 90\% confidence. We assume a binomial distribution for
each bin and follow the approach of \citet{cameron2011} in
calculating the one-sided confidence limit using the incomplete
beta function estimator (for a number of targets given in the top
panel). The covering factor limit for impact parameters less than
the virial radius is $f_c(\le \ensuremath{R_{\rm vir}}\relax) \le 0.051$. Also
shown in cyan are the covering factors that would be derived if
one assumes the two sight lines with apparent MS emission are
instead tracing emission from the CGM of M31 (shifted to slightly
larger impact parameter for clarity). The central value is the
median of the posterior probability distribution from the beta
distribution generator, while the error bars denote the 80\%\
confidence interval (i.e., the 10\%\ and 90\%\ quantiles)
\citep{cameron2011}. The blue line shows the covering factor of
HVCs in the inner regions of the halo (limiting sensitivity of
$\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \sim 18.25$), adjusting the results of
\citet{richter2012} to reflect a cumulative covering
factor. \label{fig:coveringfactor}}
\end{figure}
The results of our covering factor calculations are given in Table
\ref{tab:coveringfactor} and shown in Figure
\ref{fig:coveringfactor}. We report {\em cumulative} covering factors:
each point represents the covering factor for all impact parameters
less than the one considered, \ensuremath{f_c(\le \rho)}\relax, rather than the differential
covering factors (i.e., we calculate the covering factors for
$\le \rho$ rather than for $\rho \pm \Delta \rho$). Given the lack of
detections, we report one-sided confidence limits (the 90\% quantile
of the distribution) in Table
\ref{tab:coveringfactor}.\footnote{Specifically, we've used
\citeauthor{cameron2011}'s Equation 3, determining the value of
\ensuremath{f_c(\le \rho)}\relax\ (equivalent to that equations's upper bound $p_u$) for which
the integration of the beta distribution between \ensuremath{f_c(\le \rho)}\relax\ and 1
yields a normalized probability of 0.1.} We emphasize that the
results for each impact parameter are not independent, as they include
the same measurements as those at smaller impact parameter. Thus the
results summarized in Figure \ref{fig:coveringfactor} do not give
information on the {\em distribution} of covering factors within M31's
halo, only on the 90\% confidence limits to the covering factor within
a given impact parameter. [For interested parties, the limits on the
{\em differential} covering factors are $f_c^\prime \le 0.162$, 0.188,
0.099 at 90\%\ confidence for 100 kpc-wide bins centered on
$\rho = 50$, 150, 250 kpc, respectively.] We also note that our
covering factors are only appropriate for regions outside of the \ensuremath{\mbox{\ion{H}{1}}}\
disk of M31 (i.e., in the CGM). We define positions associated with
the disk as those that would fall on a projected circle of radius
$r=30$ kpc at an inclination angle of $i\approx 78^\circ$
\citep{braun1991}.
We also show in Figure \ref{fig:coveringfactor} the results we would
obtain if we assume the two \ensuremath{\mbox{\ion{H}{1}}}\ detections are in fact associated
with the CGM of M31, although we stress that this seems highly
unlikely (see Appendix). The covering factors derived under this
assumption are summarized in Table \ref{tab:MScoveringfactor} at
impact parameters for which the results are different than those in
Table \ref{tab:coveringfactor}. We give three values in Table
\ref{tab:MScoveringfactor}: \ensuremath{f_c(\le \rho)}\relax $_{0.1}$, \ensuremath{f_c(\le \rho)}\relax $_{0.5}$, \ensuremath{f_c(\le \rho)}\relax
$_{0.9}$, which correspond to the 10\%, 50\%\ (median), and 90\%\
quantiles for the covering factor distributions \citep[again using the
incomplete beta generator discussed by ][]{cameron2011}.
We find the covering factor within \ensuremath{R_{\rm vir}}\relax\ is
$f_c(\le \ensuremath{R_{\rm vir}}\relax) \le 0.051$ (90\%\ confidence limit). Our
ability to provide meaningful constraints is dependent on the number
of sight lines probed, and thus the limits are not strong in the inner
regions where $<10$ sight lines were observed (see Figure
\ref{fig:coveringfactor}, top). Even though we do not detect \ensuremath{\mbox{\ion{H}{1}}}\
emission in our survey, one should not be under the impression that
there is {\em no} high column density gas associated with M31's CGM,
as \citet{wolfe2016, kerp2016, westmeier2005, westmeier2008} have
found clumps of high column density gas in the halo (in some cases to
$\rho\sim100$ kpc from the center of M31). However, any such gas must
have a very small covering factor at, e.g., $\rho \ga 50$
kpc. \citet{westmeier2008}, for example, find $\sim95\%$ of the HVCs
in their map of a region of the M31 CGM are confined to $\rho \la 50$
kpc (for $M_{\rm HVC} \la 1.3\times10^5$ \ensuremath{M_\odot}\relax\ at $5\sigma$). In our
data, gas at column densities $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \ga 17.6$ can be present on
scales smaller than the 2 kpc-diameter beam and remain undetected, so
long as the total mass is less than our typical sensitivity of
$M_{\rm HI} \approx 10^4$ \ensuremath{M_\odot}\relax.
Figure \ref{fig:coveringfactor} also shows an assessment of the
covering factor of HVCs about M31 by \cite{richter2012} \citep[based
on the maps of][]{thilker2004}. The blue curve in Figure
\ref{fig:coveringfactor} shows these results rescaled for our assumed
distance and transformed into the cumulative covering factors used
here based on \citeauthor{richter2012}'s fit. The results beyond
$\rho \approx 60$ kpc are an extrapolation of that fit. This fit
applies only to the HVC component as defined by \cite{thilker2004}. It
excludes the disk of M31. Our new limits are compatible with the HVC
covering factors, although the HVCs were observed over an area in
which we have few measurements and limited constraints. The HVCs (and
``bridge'' material) observed about M31 have typical column densities
in excess of $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \ga 18$ and masses $M_{\rm HI} \ga 10^5$ \ensuremath{M_\odot}\relax,
which would be readily detectable by our observations if such material
were common in our survey area. While we discuss our results in the
general context of the CGM, our definition of CGM is inclusive of
``high velocity clouds.'' Our limits show there is not an abundance of
HVCs beyond the mapping limits of those earlier works. A caveat is
that our covering factor determinations are limited to
$-515 \le \ensuremath{v_{\rm LSR}\relax} \le -170$ km~s$^{-1}$, the lower end set by
frequency-switching limits and the upper end set to avoid
contamination by Galactic HVCs.
\section{Comparison with Recent CGM Surveys and Simulations}
\label{sec:comparisons}
\subsection{Comparison with Recent Observations}
\label{sec:obscomparison}
The nearest comparison to our measurements from the QSO absorption
line literature are those of the COS-Halos\ survey
\citep{tumlinson2011, tumlinson2013, prochaska2017}. The original
survey targeted AGNs projected within $\rho \sim 150$ kpc of 44
$\sim$\ensuremath{L^*} galaxies at $z\sim0.2$, with detection of \ensuremath{\mbox{\ion{H}{1}}}\
Lyman-series absorption in nearly every case \citep[40/44;
see][]{tumlinson2013}. Such absorption line observations can be
sensitive to column densities as low as $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \approx 13$. Most of
the \ensuremath{\mbox{\ion{H}{1}}}\ detections in the original survey had \ensuremath{\mbox{\ion{H}{1}}}\ column densities
that were either orders of magnitude below our detection limits or
that were not able to be determined due to the strong saturation of
the available Lyman series lines. Follow-up measurements of the Lyman
break in some of these these galaxies have recently been completed,
providing stronger constraints on their \ensuremath{\mbox{\ion{H}{1}}}\ column densities
\citep{prochaska2017}. We compare our results with this most recent
update.
Figure \ref{fig:CHcomparison} compares the updated COS-Halos\
measurements with our measurements in the halo of the Andromeda
galaxy. The top panel shows \ensuremath{N(\mbox{\ion{H}{1}})}\relax\ versus $\rho$ from
\citet{prochaska2017}. All of the COS-Halos\ sight lines are
considered, though many of the \ensuremath{\mbox{\ion{H}{1}}}\ columns (19/44) fall below the
lower bounds of this plot. Several of the COS-Halos\ sight lines have
only broad constraints, with lower limits from saturated Lyman series
lines or a saturated Lyman break and upper limits from the lack of
damping wings on Ly$\alpha$\relax\ or Ly$\beta$\relax. \citeauthor{prochaska2017} argue these
systems should be treated as having a flat probability distribution in
\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax\ between their lower and upper bounds. We adopt this
recommendation, plotting these systems in Figure
\ref{fig:CHcomparison} as extended error brackets (representing the
95\% confidence interval) without central values.
\begin{figure}
\epsscale{0.75}
\plotone{fig-CHcomparison.pdf}
\caption{{\em Top:} The column density limits through the CGM of M31
compared with the results of the absorption-line COS-Halos\ survey
\cite[see also \citealt{thom2012,
tumlinson2013}]{prochaska2017}. The error bars without central
values represent sight lines for which \cite{prochaska2017}
bracket the column densities with upper and lower limit, arguing
for a flat probability distribution between those 95\% confidence
limits. Because the COS-Halos\ results probe galaxies over 2 dex
in stellar mass ($9.5 \la \log M_*/M_\odot \la 11.5$), we plot the
position of each galaxy according to its normalized impact
parameter, $\rho / \ensuremath{R_{\rm vir}}\relax$ (i.e., according to the impact parameter
scale on the top axis) for comparison with the M31 results. Only
about half of the COS-Halos\ systems are seen in this plot, the
rest having column densities $\log \ensuremath{N(\mbox{\ion{H}{1}})}\relax \le 16.5$. All of the
COS-Halos\ galaxies are at $\rho / \ensuremath{R_{\rm vir}}\relax \la 0.75$. {\em Bottom:}
Covering factor limits for M31 compared with the covering factor
of $\log \ensuremath{N(\mbox{\ion{H}{1}})}\relax \ge 17.6$ absorption in the ensemble of COS-Halos\
galaxies. Here we have calculated the cumulative covering factors
for the revised COS-Halos\ survey of \ensuremath{\mbox{\ion{H}{1}}}\ \citep[][see also
\citealt{thom2012,tumlinson2013}]{prochaska2017}. The central
value for the COS-Halos\ results is the median of the output
distribution from the Monte Carlo approach described in the text,
while the error bars denote the 10\%\ to 90\%\ quantiles. All of
the COS-Halos\ galaxies are at $\rho / \ensuremath{R_{\rm vir}}\relax \la 0.75$, so we only
plot covering factors within that limit. Note that the
distribution of sight lines for the COS-Halos\ sample (grey
histogram in the upper panel) is significantly different than that
of our Andromeda sample, with the former weighting smaller
normalized impact parameters more
heavily.\label{fig:CHcomparison}}
\end{figure}
There are several caveats to note when comparing our M31 measurements
to those of COS-Halos. The COS-Halos\ sample probes a broad range of
galaxy mass/luminosity centered roughly on ``\ensuremath{L^*}.'' The estimated
halo masses for the COS-Halos\ survey cover the range
$11.5 \la \log M_{\rm h} / \ensuremath{M_\odot}\relax \la 13.7$, with stellar masses
$9.6 \la \log M_* / \ensuremath{M_\odot}\relax \la 11.5$; these should be compared with the
\citet{tamm2012} estimates for M31 of
$\log M_{\rm h} / \ensuremath{M_\odot}\relax \approx 12.0$ and
$\log M_* / \ensuremath{M_\odot}\relax \approx 11.0$. Thus, we are comparing one galaxy to
an ensemble of galaxies with a broad range in masses. Furthermore, due
to this broad mass range, the relevant physical scales can vary
significantly. We have chosen to present all COS-Halos\ measurements
and results relative to their normalized impact parameter,
$\rho / \ensuremath{R_{\rm vir}}\relax$, which is shown across the top of the
figures.\footnote{The results of our comparison are not fundamentally
different if we consider the results with $\rho$ rather than the
normalized coordinate.} For this reason some of the COS-Halos\
measurements will appear at impact parameters larger than the survey
limit of $\rho \la 150$ kpc, as the more massive systems in COS-Halos\
have virial radii larger than that of
M31.
The bottom panel of Figure \ref{fig:CHcomparison} shows a comparison
of the covering factors derived for the \citet{prochaska2017} results
in comparison with ours for M31. The COS-Halos\ results are based on a
Monte Carlo sampling of the full \ensuremath{N(\mbox{\ion{H}{1}})}\relax\ distribution over a given
(normalized) impact parameter range. We do this to account for the
systems with lower and upper bounds on the column density that
straddle our detection limit (though we have verified that it gives
results indistinguishable from the \citealt{cameron2011} approach if
we consider only systems clearly above our detection limit). For each
impact parameter considered, we create an \ensuremath{N(\mbox{\ion{H}{1}})}\relax\ distribution that is
the sum of the probability distribution functions (PDFs) describing
each of the COS-Halos\ measurements. The direct measurements in
\citet{prochaska2017} have reasonably-symmetric errors, and we adopt a
normal distribution for each in \ensuremath{\log N(\mbox{\ion{H}{1}})}\relax. For the bounded values from
\citeauthor{prochaska2017}, we follow their recommendation and treat
each as having a flat PDF between the limiting columns. For upper
limits, we assume a flat PDF between $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax = 10.0$ and the quoted
upper limits. To calculate the covering factor within a given impact
parameter, we create 10,000 realizations of the observations, drawing
the same number of \ensuremath{\mbox{\ion{H}{1}}}\ column densities as sight lines within that
range from the full \ensuremath{N(\mbox{\ion{H}{1}})}\relax\ distribution. For each realization we
calculate the covering factor as the fraction of the mock sight lines
with $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \ge 17.6$. From the full sample of covering factors, we
derive the quantiles of the distribution. The bottom panel of Figure
\ref{fig:CHcomparison} shows the results, where the central values
give the median of the distribution with error bars representing the
[10\%, 90\%] quantiles (i.e., the 80\% confidence interval).
The implied covering factor of gas with $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \ge 17.6$ for
COS-Halos\ galaxies within $0.5 \ensuremath{R_{\rm vir}}\relax$ is $f_c(\le 0.5 \ensuremath{R_{\rm vir}}\relax) = 0.26$
[0.18, 0.37] (median and [10\%, 90\%] quantiles), and it is
$\approx0.36$ within $0.33 \ensuremath{R_{\rm vir}}\relax$. (Although they do not normalize to
the virial radius, these results are similar to those given by
\citet{prochaska2017}, who note the covering factor at $\rho \le 75$
kpc is $\ga 50\%$ for $\log \ensuremath{N(\mbox{\ion{H}{1}})}\relax \ga 17.2$.) The COS-Halos\ survey
thus shows a striking prevalence of high column density \ensuremath{\mbox{\ion{H}{1}}}\
absorption systems. A sizeable fraction of the COS-Halos\ absorption
line measurements (at least 8/44) have columns in excess of our
detection limit, $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax = 17.6$. The mean \ensuremath{N(\mbox{\ion{H}{1}})}\relax\ for gas with
$\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \ge 17.6$ from COS-Halos\ is
$\log \langle \ensuremath{N(\mbox{\ion{H}{1}})}\relax \rangle_{\ge17.6} = 19.15$ (with a geometric mean
$\langle \log \ensuremath{N(\mbox{\ion{H}{1}})}\relax \rangle_{\ge17.6} = 18.65$). This gas is detected
frequently enough that the mean \ensuremath{N(\mbox{\ion{H}{1}})}\relax\ drawn from our construction of
the full \ensuremath{N(\mbox{\ion{H}{1}})}\relax\ distribution of \citet{prochaska2017} is
$\log \langle \ensuremath{N(\mbox{\ion{H}{1}})}\relax \rangle = 18.51$.
In roughly 30\% of the COS-Halos\ galaxies, gas with
$\log \langle \ensuremath{N(\mbox{\ion{H}{1}})}\relax \rangle_{\ge17.6} = 19.15$ is observed within
$0.5 \ensuremath{R_{\rm vir}}\relax$. Any of our observations intercepting gas with this column
density would readily detect such gas if it covered 30\% of our 2 kpc
beams. Thus, if the mean COS-Halos\ statistics are a fair
representation of the structure of \ensuremath{N(\mbox{\ion{H}{1}})}\relax\ on small scales --
specifically as good reprentations of the fraction of a 2 kpc diameter
beam covered by any high column density gas clouds -- we should have
detected any clouds having high \ensuremath{\mbox{\ion{H}{1}}}\ columns like those seen in the
COS-Halos\ measurements. Indeed, gas with $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax = 19.15$ would need
to cover $\le 2.8\%$ of our beam to avoid detection, some $10\times$
lower than implied by the COS-Halos\ covering factors.
It is the case that the radial distributions of sight lines in
COS-Halos\ and our sample are weighted differently. The high column
density \ensuremath{\mbox{\ion{H}{1}}}\ absorbers in COS-Halos\ are at $\rho \la 0.33 \ensuremath{R_{\rm vir}}\relax$
(Figure \ref{fig:CHcomparison}). Even considering only this range in
impact parameter, however, our M31 covering factors
$f_c(\le 0.33 \ensuremath{R_{\rm vir}}\relax)$ aren't consistent with those in the COS-Halos\
galaxies. Thus, it is difficult to reconcile the ensemble results from
COS-Halos\ with our measurements of \ensuremath{\mbox{\ion{H}{1}}}\ in the CGM of the Andromeda
galaxy.
The COS-Halos\ galaxies that show gas at $\log \ensuremath{N(\mbox{\ion{H}{1}})}\relax \ge 17.6$ have a
mean stellar mass $\langle \log M_*/\ensuremath{M_\odot}\relax \rangle \approx 10.9$ (full
range 10.2 to 11.3), so this high column density gas resides in the
halos of galaxies similar to M31. It is perhaps noteworthy that the
COS-Halos\ survey attempted to select against pairs of \ensuremath{L^*}\
galaxies \citep{tumlinson2013}, such as the M31-Milky Way pair. Thus,
if M31 were at $z\sim0.2$, it may not have been been included in the
COS-Halos\ survey. If its Local Group membership plays a role in
determining the covering factor, the M31/COS-Halos\ comparison could
be inappropriate (although the halos in the Local Group-like
simulations of \citealt{nuza2014} show results similar to those of the
single-halo simulations discussed below). However, because the
initial selection was done using photometric redshifts, this selection
against galaxy pairs was not as clear cut as initially intended
\citep{werk2012}.
In general metal absorption lines show higher covering factors about
galaxies than our \ensuremath{\mbox{\ion{H}{1}}}\ 21-cm measurements about M31. Metal lines are
found with covering factors $>60\% - 75\%$ within $\sim0.5 \ensuremath{R_{\rm vir}}\relax$
\citep[e.g.,][]{chen2010, nielsen2013, stocke2013, werk2014,
borthakur2016}. Indeed, \citetalias{lehner2015} find covering
factors near unity for \ensuremath{\mbox{\ion{C}{2}}}\ and \ensuremath{\mbox{\ion{Si}{3}}}\ absorption within
$\rho \la 0.5 \ensuremath{R_{\rm vir}}\relax$, albeit with a very small sample. Preliminary
results from Project AMIGA still support a high covering factor in
these ions (N. Lehner et al. in prep). The metal lines show higher
covering factors in part due to the greater sensitivity of absorption
line techniques to low column density gas (e.g., Ly$\alpha$\relax\ absorption
lines are detectable to $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax < 12.5$), but it is also a reflection
of the general ionization level of gas in the CGM. A majority of the
cool and warm gas ($10^{4-5} \le T \le T_{\rm vir}$) in galaxy halos
is significantly ionized \citep[][]{werk2014,keeney2017}, including
the LLS-like regime that we are probing \citep{lehner2013,
fumagalli2016, lehner2016}.
The metal ion covering factors about M31 are not that different than
those found about COS-Halos\ and other galaxies. But, depending on the
degree of saturation within the detected metals, we should not
necessarily expect to see the same covering factor in high \ensuremath{\mbox{\ion{H}{1}}}\ column
density gas. We have compared the \citetalias{lehner2015} measurements
along the three sight lines with $\rho \la 50$ kpc relative to
M31 to the COS-Halos\ results. All of the COS-Halos\ with high
\ensuremath{\mbox{\ion{H}{1}}}\ column densities ($\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \ge 17.6$) also have significantly
higher \ion{Si}{2}, \ion{Si}{3}, and \ion{C}{2} columns than the three
inner-halo Andromeda sight lines. These COS-Halos\ systems have
$>2-3\times$ higher \ion{Si}{3} columns than M31 (all are lower limits
due to saturation in the COS-Halos\ sight lines) and $\sim10\times$
higher \ion{Si}{2} column densities. Thus, it appears the COS-Halos\
sight lines simply probe higher column densities of low-ionization
gas, including \ensuremath{\mbox{\ion{H}{1}}}. This is hidden by the comparison of metal ion
covering factors because the metal absorption lines are so sensitive
and quick to saturate. The metal / \ensuremath{\mbox{\ion{H}{1}}}\ ratios in M31 appear to be
consistent with those found by COS-Halos.
\subsection{Comparison with Recent Simulations}
The covering factors of \ensuremath{\mbox{\ion{H}{1}}}\ and metal ions are considered in a number
of simulation papers \citep{fumagalli2011, cafg2011, shen2012,
fernandez2012, fumagalli2014, cafg2015, suresh2015, cafg2016,
gutcke2016, liang2016}. The motivation for extracting this quantity
from simulations arises in part because this is an observable quantity
\citep{rudie2012,prochaska2013}, but also because the covering factor
can respond to changes in the accretion rate and feedback intensity in
a galaxy and the numerical approaches adopted. For example, the
covering factors of LLS absorption can change by factors of $>2$
depending on the wind models adopted or presence of AGN activity
\citep[e.g.,][]{suresh2015}. Most commonly the predictions are made
for $z\sim2$, not only because the simulations are less expensive to run
to these redshifts, but also because there are more measurements of
the \ensuremath{\mbox{\ion{H}{1}}}\ covering factor at high redshift than at $z\sim0$, which
typically requires space-based observations. At the same time, the
choice of the limiting \ensuremath{\mbox{\ion{H}{1}}}\ column for reporting the covering factors
varies widely, depending on the focus of the simulations.
With this in mind, we compare our observations of M31 with the recent
simulations reported in \citetalias{hafen2017}, who investigated the
relationship of $z<1$ LLSs to galaxies; these halos were simulated as
part of the FIRE project \citep{hopkins2014}. Figure
\ref{fig:coveringfactorModels} compares our results with an ensemble
of cosmological zoom simulations from \citetalias{hafen2017}, shown
with the green curve and shaded regions. The analysis of
\citetalias{hafen2017} produce simulated column density distributions
for their ensemble of galaxies, which they found to be consistent with
the cosmological incidence rates of low-redshift ($0 \le z \le 1.0$)
LLSs at $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \ge 17.5$ and $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax\ge 17.2$
\citep{ribaudo2011}. These simulations however have difficulties
matching the observed metallicity distribution of low-z LLSs
\citep{lehner2013, wotta2016}. The results summarized in Figure
\ref{fig:coveringfactorModels} are from the four halos studied in
\citetalias{hafen2017} whose final ($z = 0$) masses are
$11.8 \le \log M_{\rm h}/M_\odot \le 12.1$ (simulations m11.9a, m12i,
m12q, m12v of \citetalias{hafen2017}). The simulated halos have been
projected into a $512^3$ grid spanning $2.4\times R_{\rm vir}$ (with
cell sizes $\sim1.0$--1.2 kpc) for each simulation. We derived
cumulative covering factors for $\log N(\mbox{\ensuremath{\mbox{\ion{H}{1}}}}) \ge 17.6$ using
normalized impact parameters $\rho/R_{\rm vir}$ as we did in the
COS-Halos comparison. We consider each of the simulated galaxies from
three orthogonal perspectives and use 11 snapshots per simulation over
the redshift range $0 \le z \le 0.25$. Thus there are 132 total
models that go into calculating \ensuremath{f_c(\le \rho)}\relax\ and its quantiles for each
impact parameter bin (which are sampled in steps of
$0.1 \, \rho/\ensuremath{R_{\rm vir}}\relax$).
\begin{figure}
\plotone{fig-coveringFactorModels.pdf}
\caption{The cumulative covering factors for M31 compared with
simulation results. The green line represents the median cumulative
covering factor from the $11.8 \le \log M_{\rm h}/M_\odot \le 12.1$
simulations of \citeauthor{hafen2017} \citepalias{hafen2017}, while
the dark and light green shading shows the [25\%, 75\%]
(interquartile) and [5\%, 95\%] ranges from these models. The median
simulation result has a covering factor $f_c (\le \ensuremath{R_{\rm vir}}\relax) = 0.011$.
Our limits on the covering factor from M31 are consistent with
expectations from these
simulations. \label{fig:coveringfactorModels}}
\end{figure}
The thick curve in Figure \ref{fig:coveringfactorModels} shows the
median of the \citetalias{hafen2017} results, the darkest shading
shows the interquartile range (from the 25\%\ to 75\%\ quantiles), and
the lightest shading shows the 5\% to 95\% quantile range. The median
results show that the covering factor from these simulations of high
column density gas around galaxies in the mass range analyzed is
generally relatively small, with simulated median covering factors
$f_c (\le 0.5 \ensuremath{R_{\rm vir}}\relax ) = 0.037$ and $f_c(\le R_{\rm vir}) = 0.011$.
These compare with our upper limits of $<0.11$ and $<0.051$ at 90\%
confidence for the same two impact parameters; they are significantly
lower than the results from COS-Halos\ at these same impact
parameters. We point out that the simulations display significant
variations in the covering factors. The variations are a result of
both strong time-dependence related to wind and accretion activity (in
the case of the \citetalias{hafen2017}) as well as variations between
different galaxy halos \citep[][\citetalias{hafen2017}]{cafg2015,
muratov2015}.
The highest covering factors in the simulations arise in the inner
regions of the galaxies. We do not sample $\rho \la 50$ kpc well with
our 21-cm observations due to our focus on directions with known
UV-bright AGN (we have only 4 directions in this regime). In the inner
regions (within $\rho \approx 0.33 \ensuremath{R_{\rm vir}}\relax \approx 100$ kpc) the FIRE
simulations from \citetalias{hafen2017} seem to be in good
agreement with the covering factors derived for M31's HVC population
\citep{richter2012}.
Other simulations find results similar to those of
\citetalias{hafen2017}. For example, \citet{gutcke2016} find
$f_c\approx 0\% - 10\%$ for $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \ge 17.2$ in simulations of
M31-like mass galaxies at $z\sim0$, where the range represents the
variations seen in simulations of different
galaxies. \citet{fernandez2012} presented a high-resolution simulation
of a single M31-like halo, and they found \ensuremath{\mbox{\ion{H}{1}}}\ covering factors
reasonably consistent with the \citet{richter2012} distribution for
$\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \ge 17.7$. \citet{nuza2014} considered a Local Group-like
configuration of halos; they find very high covering factors in the
inner regions of those two halos, and results that are generally
consistent with our observations. Our measurements in M31 are
consistent with the median set of simulations in all of these cases,
showing a quite low covering factor of optically-thick \ensuremath{\mbox{\ion{H}{1}}}\ when
assessed within \ensuremath{R_{\rm vir}}\relax.
\section{Discussion}
\label{sec:discussion}
In this work, we have investigated the frequency of optically-thick
gas with $\log \ensuremath{N(\mbox{\ion{H}{1}})}\relax \ge 17.6$ about the local $\ensuremath{L^*}$ galaxy M31. On
the basis of GBT searches for 21-cm emission with a 2~kpc beam, we
find a minimal covering factor of such gas in the halo of the
Andromeda galaxy. We derive a covering factor within the virial radius
of $f_c (\le \ensuremath{R_{\rm vir}}\relax) < 0.051$. Our limits on the cumulative
covering factor are not particularly strong in the inner halo (esp.,
$\rho \la 0.25 \ensuremath{R_{\rm vir}}\relax$), where we have few pointings. Complementary
information exists from the previous HVC results within
$\rho \la 0.2 \ensuremath{R_{\rm vir}}\relax$ \citep{thilker2004, richter2012}.
There is some discrepancy in our assessment with previous work on the
\ensuremath{\mbox{\ion{H}{1}}}\ environment of the M31/M33 system. As noted earlier, the initial
maps by \citet{braun2004} seem to show more wide-spread emission than
is present \citep{lockman2012, wolfe2013, wolfe2016}.\footnote{These
works generally show a very low detection rate in their searches,
consistent with our low covering factor determinations. However,
because these works studied areas specifically to address previous
potential detections of \ensuremath{\mbox{\ion{H}{1}}}\ by \citet{braun2004}, their
observations are not suitable for general covering factor
calculations, as they may be biased in favor of \ensuremath{\mbox{\ion{H}{1}}}\
detections. Nonetheless, these efforts all support a small covering
factor on small scales, even in directions suspected originally to
have \ensuremath{\mbox{\ion{H}{1}}}\ emission.} More recently, \citet{kerp2016} have emphasized
the potential for blending of high-velocity M31 structures with the
Milky Way's 21-cm emission. They attempted to decompose the
complicated region around M31 with guidance from continuities in the
``difference second moment map,'' arguing that several very large \ensuremath{\mbox{\ion{H}{1}}}\
complexes within $\rho \sim 80$ kpc are associated with M31. These
complexes span tens of kpc at the distance of M31. We would not have
included them as detections in our survey because they lie outside of
our velocity range. The \citeauthor{kerp2016} structures are at
$-150 \la \ensuremath{v_{\rm LSR}\relax} \la -100$ km~s$^{-1}$\ (within $\approx200$ km~s$^{-1}$\ of M31's
systemic velocity), whereas our survey is restricted to
$-515 \le \ensuremath{v_{\rm LSR}\relax} \le -170$ km~s$^{-1}$\ in order to avoid contamination by
Galactic HVC emission. Furthermore, none of our pointings directly
intersect the clouds they identify, a by-product of our low sampling
density at small impact parameters. Assuming all of the clouds
\citeauthor{kerp2016} identify are indeed associated with M31, their
Figure 7 implies covering factors just consistent with our limits. For
example, their map implies $f_c(\le 0.33 \ensuremath{R_{\rm vir}}\relax) \approx 0.2$ (ignoring
emission within 25 kpc in order to exclude the disk), which compares
with our limit $f_c(\le 0.33 \ensuremath{R_{\rm vir}}\relax) \le 0.16$ at 90\% confidence. This
covering factor estimate is dominated by the two large structures that
they identify on opposite sides of the galaxy (their clouds {\em a}
and {\em b}). It is a bit disconcerting that the two structures,
separated by many tens of kpc, both lie at the extreme positive
velocities expected for circumgalactic material about M31 (both with
$\ensuremath{v_{\rm LSR}\relax} \approx -115$ km~s$^{-1}$). It is not clear why two structures
separated by more than the diameter of M31's disk should have the
exact same extreme velocity. The covering factors implied by the
\citet{kerp2016} are $\sim2\times$ higher than those of
\citet{richter2012} at the same impact parameters. This discrepancy is
wholly connected to the issues of blending with Milky Way HVC gas at
$\ensuremath{v_{\rm LSR}\relax} \ga -170$ km~s$^{-1}$. While \citeauthor{kerp2016} attempt to solve it
through a unique approach, it is difficult to verify their results
with other means.
In general our limits to the \ensuremath{\mbox{\ion{H}{1}}}\ covering factor are quite consistent
with the low values expected from cosmological zoom simulations of
individual galaxies at $z \approx 0$
\citep[][\citetalias{hafen2017}]{gutcke2016}. With no detections, we
do not have a valid characterization of the shape of the $f_c$
distribution about M31, although the work by \cite{richter2012}
provides a fit to the HVC covering factor within $\sim60$ kpc. Some of
the simulated galaxies show higher $f_c$ than is observed in M31, as
captured in the simulations ranges shown in Figure
\ref{fig:coveringfactorModels}. This is a result of the halo-to-halo
variations as well as the significant temporal variations within a
single halo \citep[][\citetalias{hafen2017}]{rahmati2015,
cafg2015}. Thus, while the covering fraction of optically-thick gas
about M31 is not in conflict with the simulations, it is a single
sample from a specific time. It may be drawn from a broad distribution
like those seen in the simulations.
As discussed in \S \ref{sec:obscomparison}, our results appear
discrepant from the COS-Halos\ galaxies \citep{tumlinson2013,
prochaska2017}. There are very few other works that put strong
constraints on the covering factor of optically-thick gas about
low-redshift galaxies. There could be a multitude of physical reasons
for the difference between M31 and COS-Halos : 1) M31 could have
experienced significant evolution in the last 2.5 Gyr, the time since
the typical redshift of the COS-Halos\ galaxies; 2) ``green valley''
galaxies with star formation rates like M31, which are not well
sampled by COS-Halos, may have distinct CGM properties; 3) the
presence of M31 in a group with another massive spiral, which would
have been (mildly) selected against in the COS-Halos\ sample, may
affect its CGM. However, it's unlikely we could distinguish these
effects from simple stochastic variations for a single galaxy.
The environment of M31 could plausibly play a role in shaping its
CGM. \cite{burchett2016} found a smaller detection rate (covering
factor) of \ion{C}{4} for galaxies in high density environments
(assessed over $\rho \approx 1.5$ Mpc). They do not find a
corresponding difference in \ensuremath{\mbox{\ion{H}{1}}}\ covering factor, but their dynamic
range in \ensuremath{N(\mbox{\ion{H}{1}})}\relax\ is small due to saturation of Ly$\alpha$\relax\ (they cannot probe
columns higher than $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \sim 14.5$). However, the Local Group
would not be included in their high density category, as only 2-3
galaxies would make their luminosity cut ($M_r < -19$ mag) for
identifying galaxies in their density counting scheme. We note that
\citet{nuza2014} have presented simulations of a Local Group-like pair
of halos, both of which show high covering factors. They find near
unity (differential) covering factors for gas at $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \ga 17.85$
to 30-50 kpc in their two halos. However, this is based on only one
pair of halos seen at a single time, so we cannot yet draw strong
conclusions based on these simulations.
In general higher covering fractions are found from metal absorption
line searches. Several studies of strong \ensuremath{\mbox{\ion{Mg}{2}}}\ absorption (typically
$W_r \ga 0.3$ m\AA ) have found quite high covering factors (e.g.,
$>0.6$) within $\sim\ensuremath{R_{\rm vir}}\relax$ \citep{chen2010, nielsen2013}, even for
relatively massive galaxies like M31. Some of this difference has to
do with the broad range of \ensuremath{\mbox{\ion{H}{1}}}\ column densities probed by \ensuremath{\mbox{\ion{Mg}{2}}}\
absorption, which can trace gas over nearly 5 orders of magnitude in
\ensuremath{N(\mbox{\ion{H}{1}})}\relax\ \citep[gas with $\log N(\mbox{\ensuremath{\mbox{\ion{Mg}{2}}}}) \sim 13$ --
$W_r(2796) \sim 0.3$ m\AA\ -- can probe solar metallicity gas with
$\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \sim 16$ or damped Ly$\alpha$ systems with $<1/300$ of the
solar metallicity; ][]{wotta2016}.
We note that \cite{rao2013} searched for the signature of \ensuremath{\mbox{\ion{Mg}{2}}}\ and
other low-ion absorption from the halo of M31 using low-resolution COS
spectra. They probed impact parameters $13 \la \rho \la 112$ kpc with
10 AGNs, finding absorption in four sight lines, all of which reside
at $\rho \la 40$ kpc and within the $\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \sim 18.3$ contours of
the 21-cm measurements. Their observations, sensitive to
$W_r(2796) \sim 0.3 - 0.5$ \AA, targeted AGNs projected along
M31's major axis. The lack of absorption in their four targets
at $40 \la \rho \la 112$ kpc is perhaps not surprising given the small
number of sight lines and the low covering factor of high column
density gas we find (their low-resolution observations have
sensitivity to only $W_r(2796) \sim 300$ m\AA, whereas Project AMIGA's
sensitivity is $\sim20$ m\AA).
Roughly half of the sight lines observed in this work have been
observed by {\em HST} /COS at intermediate resolution (G130M+G160M) as part
of Project AMIGA. These observations, a subset of which were published
by \citetalias{lehner2015}, are sensitive to relatively weak metal
line absorption, notably in transitions from the ions \ensuremath{\mbox{\ion{C}{2}}},
\ion{C}{4}, \ensuremath{\mbox{\ion{Si}{2}}}, \ensuremath{\mbox{\ion{Si}{3}}}, \ion{Si}{4}. The preliminary data from this
program suggest a quite high metal line covering factor. One of the
initial goals for Project AMIGA GBT was to provide an \ensuremath{\mbox{\ion{H}{1}}}\ reference
for potential metallicity estimates, since Ly$\alpha$\relax\ absorption from the
Andromeda galaxy is swamped by the Galactic absorption
trough. Unfortunately, with only upper limits on \ensuremath{N(\mbox{\ion{H}{1}})}\relax\ coupled with
potential beam-dilution effects, we cannot provide a hydrogen
reference for the metallicity determination. Without the hydrogen
reference, determining the metallicities along these sight lines will
not be possible \citep[on this point we disagree with the recent work
of ][]{koch2015}.
Better constraining the covering factor in the inner regions of M31's
CGM may be readily accomplished by simply observing more sight lines
(and more GBT observations are forthcoming). The \ensuremath{\mbox{\ion{H}{1}}}\ maps of
\citet{thilker2004} as analyzed by \citet{richter2012} provide some
guidance in this inner region, albeit at worse mass sensitivity than
the current observations and only at $\rho \la 50$ kpc. Pushing below
our current \ensuremath{\mbox{\ion{H}{1}}}\ column density limits, which are in the range of
$\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \sim 17.4$ to 17.6 at $5\sigma$, is in principle
possible. However, in many of the cases presented here the limiting
factor in the \ensuremath{\mbox{\ion{H}{1}}}\ sensitivity is the quality of the spectral
baselines. The GBT baselines are already among the best available, but
we are working near the limit for the current instrumentation.
\section{Summary}
\label{sec:summary}
We have used the Green Bank Observatory's 100-m Robert C. Byrd Green
Bank Telescope to search for 21-cm emission from the CGM of our
neighbor M31. We detect no \ensuremath{\mbox{\ion{H}{1}}}\ emission, with column density limits
that overlap studies of QSO absorption line experiments (notably the
LLSs). Our principal conclusions are as follows.
\begin{enumerate}
\item We constrain the covering factor of optically-thick \ensuremath{\mbox{\ion{H}{1}}}\ with
$\ensuremath{\log N(\mbox{\ion{H}{1}})}\relax \ga 17.6$ about M31 to be
$f_c (\le \ensuremath{R_{\rm vir}}\relax) < 0.051$ (90\% confidence).
\item Our covering factor limits for these high \ensuremath{\mbox{\ion{H}{1}}}\ column densities
are much lower than those found for metal lines about \ensuremath{L^*}\
galaxies, including M31 \citepalias{lehner2015}.
\item The covering factors derived here are also discrepant from
recent measurements by the COS-Halos\ project at the same \ensuremath{\mbox{\ion{H}{1}}}\
column densities \citep{prochaska2017}. The origin of this
difference is not clear. It may be related to characteristics of M31
that are not represented in the COS-Halos\ sample, although the
difference could also be consistent with stochastic variations in
M31's CGM with time.
\item The covering factors of optically-thick \ensuremath{\mbox{\ion{H}{1}}}\ about M31 are
consistent those found in recent cosmological zoom simulations.
\end{enumerate}
\acknowledgements
Part of this manuscript was written at the 2016 Arthur M. Wolfe
Symposium in Astrophysics hosted by IMPS of UC Santa Cruz Department
of Astronomy. We thank the Esalen Institute for its great setting and
wonderful hospitality during that retreat. Support for HST Program
number 14268 was provided by NASA through a grant from the Space
Telescope Science Institute, which is operated by the Association of
Universities for Research in Astronomy, Incorporated, under NASA
contract NAS5-26555. Some of this work was supported by NSF grants
AST-1212012 and AST-1517353 to Notre Dame and JHU. DJP recognizes
partial support from NSF CAREER grant AST-1149491. Contributions by
ZH and CAFG were additionally supported through NSF grants AST-1412836
and AST-1517491 and through NASA grant NNX15AB22G. This research made
use of Astropy, a community-developed core Python package for
Astronomy \citep[][]{robitaille2013}, and the matplotlib plotting
package \citep{hunter2007}.
\facility{GBT}
\software{Astropy \citep{robitaille2013},GBTIDL \citep{marganian2006},
Matplotlib \citep{hunter2007}}
| eea1b94ddabd6012f87dddd0b5d27a095a9488db | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{INTRODUCTION}
\nobreak
Present day halo globular clusters are
a tracer population in the dark matter of the Milky Way, providing important insights
into the assembly history of the galaxy and
some of the first observational evidence for galactic dark matter \citep{HS:78}.
\citet{SZ:78} proposed a basic conceptual picture in which star clusters formed
in ``transient protogalactic fragments" which accrete onto the developing galaxy
to form the stellar and dark matter
halo of the Milky Way galaxy. This picture is now elaborated within general models of hierarchical galaxy
formation within the Cold Dark Matter (CDM) cosmology.
Globular clusters evolve through the gravitational interactions of their constituent stars and the influence
of external gravitational fields.
Encounters between individual stars, binaries and more complex bound units, systematically increases the binding
energy of the central region causing the outer regions of the star cluster to expand.
The external gravitational field, which can often be approximated with a tidal tensor,
set a distance beyond which cluster stars are no longer
bound to the cluster. And, time varying tidal fields, from either other massive sub-halos or peri-galacticon passage,
pump energy into cluster stars.
The internal two-body relaxation time increases with increasing mass \citep{Spitzer:87} for the same
orbit within a galaxy.
Consequently, lower mass clusters lose mass more quickly, systematically
depleting the lower mass clusters so that over the
lifetime of the galaxy a wide range of initial mass functions come
to resemble the nearly universally observed log-normal mass distribution
\citep{FR:77,GO:97,FZ:01}.
The stars that become unbound from their parent cluster are pulled away in tidal
streams that are offset in angular momentum so lead and trail the cluster in its orbit.
In principle a stream provides a complete history of the globular
cluster and the properties of the potential along its orbit.
The density of stars along a tidal stream provides an
empirical estimate of the mass loss from its parent cluster, which
in principle could be used to reconstruct the cluster, including completely dissolved clusters.
The simulations here will shed light on some of the difficulties that will need to be overcome to allow even
a statistical reconstruction of the history of interest.
The thin components of streams
are sensitive indicators of the low mass sub-halos that
the LCDM power spectrum predicts should be orbiting in the halos of the galaxy
\citep{Klypin:99,Moore:99,YJH:11,Carlberg:12,Bovy:17}. Current methods of analysis have concentrated on
the relative simple streams that develop in a static potential, which may require modification to
take into account
the time variations of the gravitational fields in which the stream develops.
High resolution gas dynamical galaxy formation simulations
are beginning to explore formation of dense star clusters within a galaxy
\citep{KG:05,BDE:08,Kimm:16,RAG:17,Li:17,Kim:17}. The current spatial resolution of the
resulting clusters is not yet adequate to follow the dynamical evolution of individual
stars in the clusters. The study of star clusters remains a challenging numerical problem, largely
addressed through the \citet{Aarseth:99} program of gravitational n-body
codes without particle softening. The current version, Nbody6 and its specialized
variants, is a highly realistic cluster code, which allows external fields to be described with
an externally supplied time varying tidal tensor.
The numbers and masses of globular clusters at high redshift are of interest for the reionization of the
universe \citep{Bouwens:17,BK:17} via the the ionizing radiation from their massive stars present
at early times, and, the creation of binary black holes which inspiral to merge and emit gravitational waves
at low redshift.
The low metal abundance of globular cluster stars reduces the
stellar winds of the high mass stars which can lead to stellar remnant black holes in the $30{\rm\,M_\odot}$ range.
The black holes sink via gravitational interactions into the central regions where processes of
binary exchange and binary gravitational hardening over a Gyr or so can put together
massive binary black holes (BBHs) that will inspiral over a Hubble time to merge \citep{PZM:00}.
BBH merger rates predicted from the current
population of globular clusters generally come up somewhat low \citep{Rodriguez:15,Rodriguez:16,Chatterjee:17,Askar:17} relative
to the currently measured LIGO event rates \citep{LIGO2}. If
the high redshift population of globular clusters were significantly larger than
the low redshift numbers on which current predictions are based.
It is interesting that both BBHs and reionization may be connected through the high mass
stars in high redshift globular clusters.
To better understand the history of globular clusters and their tidal streams this paper follows
a population of globular clusters formed at high redshift
in the sub-galactic dark matter halos that merge over time to create a galactic halo.
The evolution of a pre-galactic dark matter distribution drawn from
a larger cosmological simulation to form a galaxy's dark matter halo
is a well understood n-body simulation problem. Here such
a simulation is augmented through the insertion of
dense star clusters in the central regions of the pre-galactic sub-halos. The added star clusters contribute so little
mass to the overall simulation, less than 0.01\% here, that the dark matter history is barely affected.
The simulation is conceptually started at a time after the massive stars in the clusters
have evolved and dispersed any surrounding gas so that hydrodynamical effects no longer
play a role. The star particles use a gravitational softening to speed up the calculation, requiring
that the expected level of two-body relaxation between stars in the clusters be added in
with a Monte Carlo velocity perturbation scheme.
Questions of interest here are to what degree the visible tidal streams be used to reconstruct the
initial dense stellar cluster distribution, and, what is the character of
dark matter sub-halo induced gaps in these streams.
Both of these questions can be answered without a high precision
cluster mass loss model, although the rate of mass loss must be realistic.
The next section lays out the details of the simulations with the starting conditions for the
dark matter, the construction of the individual star clusters and where they
are inserted into the dark matter distribution. The Monte Carlo scheme to
appropriately add two-body relaxation back into the softened star clusters
along with high accuracy calibration runs is described in Section 2.2.
Section 3 describes the evolution of the cluster masses and how the
stars pulled away from the clusters are distributed in streams.
\begin{figure}
\begin{center}
\includegraphics[angle=0,scale=0.8,trim=190 190 190 190,clip=true]{Figure_1.pdf}
\end{center}
\caption{The VL-2 halos at z=3, 2.078 Gyr (top) and z=8, 0.70 Gyr (bottom).
The red circles are the halos into which the star clusters will be inserted, with the circle radius
being equal to the $R_m$ of the halo. The remaining halos
are shown as dots at the halo center. No particles are shown.
}
\label{fig_xy0}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[angle=0,scale=0.8,trim=190 190 190 190,clip=true]{Figure_2.pdf
\end{center}
\caption{The reconstituted dark matter particle distribution
0.5~Gyr after starting at z=3 (top) and z=8 (bottom panel). The red circles are the locations
of the centers of the model star clusters. }
\label{fig_xy50}
\end{figure}
\section{Globular Clusters in a Cosmological Simulation}
The formation processes that lead to the low metallicity halo population of globular clusters
of the Milky Way and all other galaxies are not yet well understood.
At the high redshifts corresponding to their stellar ages
the most common host dark matter halos will have peak circular velocities in roughly the 10-30\ifmmode{\,\hbox{km\,s}^{-1}}\else {\rm\,km\,s$^{-1}$}\fi\ range,
sufficient to retain gas as reionization occurs, but the gas will largely be driven out as a result energy and momentum
input from the more massive stars that they form, leaving a largely dark matter dominated system
containing a low metallicity stellar population \citep{DS:86}.
Accordingly, it is a reasonable step to focus simulations on the dynamical processes of star clusters in dark matter halos,
temporarily setting aside the complications of gas and the stellar structure of the host halos.
The star clusters are conceptually placed in the dark halos
after their most massive stars have gone through their short and dramatic evolution in
the first few hundred million years of the star cluster's evolution, after which
stellar evolution produces smaller and slower changes in the mass of stars.
The initial conditions then become fairly straightforward, requiring a dark matter distribution at high redshift which evolves into
a good approximation of the Milky Way's dark halo. An appropriate subset of the numerous low mass dark halos present
at high redshift is then populated with dense star clusters. Subsequent
gravitational evolution creates a Milky Way like dark halo containing the remnant star clusters
and their tidal tails.
Most of the halo clusters orbit at sufficiently large radius that the galactic disk and bulge do not substantially disturb the clusters.
The simulations contain a mixture of star particles and dark matter particles. The star particles
begin in star clusters that are inserted within the top thirty or so dark matter halos identified
in the VL-2 simulation \citep{VL2}.
The resulting mixture of particles is followed using the gravity code Gadget2 \citep{Gadget2}, augmented
to provide the two-body relaxation between stars that drives cluster evolution.
\subsection{The Cosmological N-body Halos}
The initial conditions for the simulations use the snapshots of the Via Lactea II simulation (VL-2) \citep{VL2}.
A set of 20048 halos, with masses above about $2\times 10^5{\rm\,M_\odot}$,
were identified within VL-2 at redshift zero and then traced back in time
to give their size, mass, position and velocity.
The VL-2 data files for these progenitor sub-halos are the basis of
the initial conditions here. Two
snapshots, near redshifts 3 and 8, are selected as start times for the insertion of
dense star clusters. The lower redshift is within epoch of peak star formation for galaxies
and the higher redshift is within the epoch of reionization, which roughly spans
the likely range of times over which the old halo clusters are formed.
A continuous distribution of globular cluster formation
times could be recreated by summing over appropriately weighted starting times,
since the star clusters are sufficiently light
that they have essentially no effect on the dark matter halos.
Only those sub-halos above a minimum sub-halo mass, in the range of a few $10^9{\rm\,M_\odot}$,
are normally populated with star clusters, leaving the rest empty.
To create the distribution of
particle positions and velocities the individual dark matter halos
are approximated as \citet{Hernquist:90} spheres.
Each halo is characterized with a maximum circular
velocity, $V_m$, which occurs at radius $R_m$, together giving
a Hernquist mass $M_H = 4 G^{-1}V_m^2 R_m$.
At time 0.70 Gyr, redshift 7.8,
halos with $M_H \ge 0.1\times 10^{10}{\rm\,M_\odot}$ are selected, which identifies
31 halos, and at time 2.08 Gyr, redshift 3.24, $M_H\ge 0.5\times 10^{10}{\rm\,M_\odot}$ which selects
33 halos. The selected halos at the two times are shown in Figure~\ref{fig_xy0}.
\begin{figure}
\begin{center}
\includegraphics[angle=0,scale=0.8,trim=180 190 190 190,clip=true]{Figure_3.pdf
\end{center}
\caption{The dark matter at the end point for the z=3 (top) and z=8 (bottom) start systems.
The red circles are the centers of the globular clusters. The boxes are 60 kpc on an edge.
}
\label{fig_xydend}
\end{figure}
The total mass of the 20048 halos in the simulation is $2.54\times 10^{12}{\rm\,M_\odot}$.
The total mass in the halos at early times is
50\% and 5\% of the final total mass
at z=3 and z=8, respectively. To allow for the mass that is not in the Hernquist sphere halos,
the additional mass is added as an extended, low density, sphere of particles centered on the 20048 primary halos.
This starting setup is sufficiently similar to the VL-2 simulations that its gravitational evolution
to the dominant, sub-halo rich, Milky Way-like final halo.
The detailed procedure to create the particle distribution
first uses the \citet{Hernquist:90} mass distribution with radius to give
the particle positions. The analytic
distribution function as a function of energy
for the Hernquist mass distribution gives the velocity distribution at each radius from
which an equilibrium particle distribution for each halo is created.
The extra mass required for the total mass is distributed in proportion to the primary halo mass and placed in a
second Hernquist sphere given a scale radius 8 times the radius of
the central equilibrium distribution, with velocities reduced to
80\% of equilibrium in the low density region to ensure that they are accreted
onto the central halo over time. The secondary extended halo is low enough density
that it does not substantially disrupt the primary halo.
The reduced random
velocities allow an accretion flow onto the central halo to develop, but is
sufficient to suppress a coherent rapid collapse and local gravitational instability.
The dark matter distributions of the two simulations are shown 1 Gyr into the simulation in
Figure~\ref{fig_xy50}. The z=8 start still has strong cosmic web features, whereas
the z=3 start with its dominant halo is in the regime of accretion of satellites.
\subsection{Simulating Dense Star Clusters}
The internal evolution of star clusters sufficiently dense
to survive for a Hubble time
remains a vigorous research area. A primary goal of this study is to accurately follow the stars
that leave the cluster in the tidal streams.
A secondary goal is to study the mass evolution of clusters as a function of their cosmological history.
All the internal structural details of the star clusters are not needed, but the clusters need
to be sufficiently realistic to lose mass at approximately the correct rate. A simplified physical model
of cluster star-star encounters is developed with its parameters calibrated using a high accuracy
gravitational n-body model.
\subsubsection{Star Cluster Internal Evolution Model}
Star clusters evaporate as a result of gravitational encounters of stars
\citep{Henon:61,Spitzer:87}. Following the evolution of a single star cluster
in full detail remains a substantial computational challenge \citep{Aarseth:99}.
\citet{Carlberg:17} presented a model in which a
softened gravity cluster is augmented with a simplified single-zone heating model
to approximate the relaxation which drives evaporation.
The model calculates velocity from the two-body relaxation time, which is then randomly
added to the stars within some fraction of the tidal radius of the model star cluster.
The calculation of the required velocity increment is
done at intervals of $\delta t$.
The velocity increment for a cluster of characteristic mass $M_c$ is calculated using the RMS
velocity diffusion coefficient \citep{BT:08},
\begin{equation}
\delta v = {\sigma_c \over C_h} \left({\delta t\over 4.9\ {\rm Myr} } \right)^{1/2}
\left({M_c\over 10^{5}{\rm\,M_\odot}}\right) ^{1/4}
\left({r_v\over 10\ {\rm pc}}\right)^{-5/4},
\label{eq_dvheat}
\end{equation}
where $\delta v$ is in units of \ifmmode{\,\hbox{km\,s}^{-1}}\else {\rm\,km\,s$^{-1}$}\fi, $\sigma_c$ is the
cluster velocity dispersion, and $r_v$ is the virial radius, calculated with softening included, then
removed in quadrature from the result. The weak explicit dependence
on cluster mass, $M_c$, is useful to retain to ensure that the heating remains reasonable as the
clusters near complete evaporation.
The cluster
mass, $M_c$ and velocity dispersion are measured within some fiducial radius which is set to be some multiple of the
virial radius, usually set at $6r_v$. The virial radius is preferred over the half mass radius on the basis that it is simpler
to calculate and is a stable quantity with a weak dependence on the maximum radius. The clusters here
have $r_v$ values about 2/3 of $r_h$.
The kinematic properties are calculated for each cluster each time the velocity increments
need to be added to the star particles. The random velocities are added to all star
particles within a radial range of typically 0.5-1.5$r_v$, with no disturbance of more distance particles.
A dummy particle with nearly zero mass is used to mark the center, which speeds
up the calculation. At
each heating step the dummy's location is updated to the current center of mass and momentum.
The $\delta v$ calculated from Equation~\ref{eq_dvheat}
is generates a Gaussian distribution of velocity offsets in each velocity direction.
\subsubsection{N-body Cluster Calibration}
Nbody6 direct n-body simulations \citep{Aarseth:99} are used to calibrate
the model heating parameters.
The Nbody6 code is used to generate a Plummer model with 20,000 particles, all of equal mass of 0.7543267 ${\rm\,M_\odot}$.
The cluster of particles is placed in a logarithmic potential with a circular velocity of 240\ifmmode{\,\hbox{km\,s}^{-1}}\else {\rm\,km\,s$^{-1}$}\fi
at 30 kpc from the center and given an initial tangential velocity of 120\ifmmode{\,\hbox{km\,s}^{-1}}\else {\rm\,km\,s$^{-1}$}\fi.
The circular velocity is close to the value of the final dark matter halo of the simulation
and the orbit is fairly typical for both real world star clusters that produce streams and
many of those that the simulations create.
The mass as a function of time of the clusters and their
half mass radii are displayed in
Figures~\ref{fig_mtplum} and \ref{fig_rhplum}, respectively, for the Nbody6 run and precisely
the same cluster run with the heated cluster model used here.
It is clear that with the right choice
of parameters the heated cluster model can reproduce a full n-body result to the
required accuracy. An important caveat is that one demonstration holds
for one cluster mass on one orbit.
Some preliminary tests (Jeremy Webb private communication) have shown
a degree of mass dependence of the most suitable heating parameters. The maximum radius inside of which
heating is applied to the star particles is fixed at $1.5r_v$, but the inner heating radius rises from 0.5 to 1.0$r_v$
for higher mass clusters.
The inner radius is described with the interpolating function,
$r_{min}=0.5r_v(5.76\times 10^4{\rm\,M_\odot}/M_c)^{1/3}$ over the mass range
$M_c=5.76\times (10^3-10^4) {\rm\,M_\odot}$, with the value pegged at the upper and bottom limit outside the range.
The heating coefficient is similarly varied with cluster mass,
\begin{equation}
C_h= 36-9(5.76\times 10^4{\rm\,M_\odot}/M_c)^{1/3}.
\label{eq_chm}
\end{equation}
A caution is that the heating coefficients are based on only two Nbody6 runs.
An important future step is to extend the cluster mass range, the range of galactic orbits,
and the degree to which the reasonable range of stellar mass function, rotation, and binary stars effect the results.
\begin{figure}
\begin{center}
\includegraphics[angle=-90,scale=0.7,trim=20 40 20 20,clip=true]{Figure_4.pdf
\end{center}
\caption{An N=20,000 Plummer sphere of equal mass stars having a mass $1.51 \times 10^4{\rm\,M_\odot}$ created evolved with
Nbody6 (red line) and the heated cluster code used here
with the heating parameter set at 25 (the adopted value, blue) and 30 (green).
}
\label{fig_mtplum}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[angle=-90,scale=0.7,trim=20 40 20 20,clip=true]{Figure_5.pdf
\end{center}
\caption{A 20,000 Plummer sphere cluster with mass $1.4 \times 10^4{\rm\,M_\odot}$ created and evolved with
Nbody6 (red line) and the heated cluster code
with the heating parameter set at 25 (the adopted value, blue) and 30 (green).
}
\label{fig_rhplum}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[angle=0,scale=0.8,trim=120 100 120 100,clip=true]{Figure_6.pdf
\end{center}
\caption{The internal properties of the model clusters, density distribution (top panel), radial velocity dispersion (middle panel),
and velocity anisotropy (bottom panel) shown for all 198 clusters started as
$W_0=7$ King models.}
\label{fig_GCkin}
\end{figure}
\subsection{Cluster Sizes and Halo Orbits}
\citet{King:66} models of star clusters are inserted into the 33 halos more massive
than $0.5\times 10^{10}{\rm\,M_\odot}$ in the z=3 start and the 31 halos more massive than
$0.1\times 10^{10}{\rm\,M_\odot}$ in the z=8 start.
Each star cluster starts with 20,000 star particles, with
particle masses of 15${\rm\,M_\odot}$ for clusters of total mass $3\times 10^5 {\rm\,M_\odot}$. In
each halo, three clusters are placed on near circular orbits at $R_m/2$ and another three $R_m/4$.
The average $R_m$ is 8.4~kpc for the z=3 start and 2.6~kpc for the z=8 start, so the average cluster is at about 3 kpc
at z=3 and 1 kpc at z=8.
The clusters are started in a randomly oriented disk distribution,
at the local circular velocity with an added random velocity of 5\ifmmode{\,\hbox{km\,s}^{-1}}\else {\rm\,km\,s$^{-1}$}\fi\
in each cardinal direction to minimize cluster-cluster interactions.
The King model star particle
distribution is created with a potential parameter $W_0 = 7$, the ratio of the central
potential depth to the characteristic velocity dispersion of the cluster, resulting in a cluster with a ratio
of tidal to core radius of 22.2. A concentration
in this range is typical for the \citet{Harris:96} [Fe/H]$\le -1$ clusters beyond 10 kpc galactocentric
radius. The dark matter softening length is set at 2~kpc. The star particle softening length is
set at 2 pc, comparable to the size of a cluster core, which will not be resolved. However, the details of mass loss
here are not significantly dependent on accurately following core properties. Reducing the star particle
softening to 1 pc sharpens the density profile somewhat, but makes essentially no difference to
the mass loss rate for an orbit in the logarithmic potential, at the cost of reduced step sizes.
A key consideration for the rate of cluster mass loss
is the half-mass radius of the clusters. A King model has $r_h$ roughly one quarter of the outer radius of the cluster for
a wide range of central concentration.
If the model cluster of mass $M_c$ is scaled to the simple tidal radius, $r_t=r[M(<r)/M_c]^{1/3}$,
where $r$ is the orbital radius and $M(<r)$ the halo mass within that radius.
For the sub-galactic halos the tidal radii work out to around
around 100 pc, which leads to $r_h$ of 25 pc, which is much larger than most current halo clusters.
Evolving such large clusters in our n-body model leads to no surviving clusters
within 100 kpc of the center of the final Milky Way-like halo.
A somewhat artificial approach which avoids assumptions
is to scale the King model clusters to $r_t/10$ after which they naturally
expand to their locally preferred size over the first Gyr of the simulation.
The internal properties of all 198 clusters in the z=3 run are shown 1 Gyr
after the start in Figure~\ref{fig_GCkin}. The core has expanded from an unresolved 0.5 pc to about 5 pc.
The radial velocity dispersion shows the expected decline with radius then a rise into the unbound particles in the stream.
The cluster was started with an isotropic velocity dispersion, $\beta=0$, but with time $\beta$ has become a rising
positive function, indicating a radially oriented velocity ellipsoid.
Figure~\ref{fig_rhv} shows the evolution of the cluster half mass radius and the virial radius within the clusters of the
z=3 (top) and z=8 (bottom panel) simulations. The z=8 clusters that fall into the main halo are completely evaporated over
the course of the simulation. The \citet{Harris:96} catalog gives a half light radius of 8 pc for low metallicity clusters
beyond 10 kpc galactocentric radius and 15 pc for those beyond 30 kpc. The half light radius is expected to be somewhat
smaller than
the half mass radius as a consequence of mass segregation in clusters. The clusters that remain at the end of the
simulations here typically have half mass radii of 15-20 pc. Overall the half mass radii in the simulations overlap
the range seen in Milky Way halo clusters, but the simulated clusters are missing
the very dense clusters with half mass radii of a few parsecs.
\begin{figure}
\begin{center}
\includegraphics[angle=0,scale=1,trim=140 180 140 180,clip=true]{Figure_7.pdf
\end{center}
\caption{The half mass (pluses) and virial radii (crosses), with clusters inside 100~kpc from the
main halo center shown as red or blue.
The top panel is for the z=3 start and the z=8 start below.
}
\label{fig_rhv}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[angle=0,scale=1,trim=140 180 140 180,clip=true]{Figure_8.pdf
\end{center}
\caption{The gravitational at the location of sixteen star clusters, one per halo.
The top panel is for the z=3 start and the z=8 start below. The varying
potential depth is a combination of orbital motion within a sub-halo and the infall orbit
into the growing galactic halo.
}
\label{fig_pot}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[angle=0,scale=1,trim=140 180 140 180,clip=true]{Figure_9.pdf
\end{center}
\caption{The largest eigenvalue of the tidal tensor, in units of $(\ifmmode{\,\hbox{km\,s}^{-1}}\else {\rm\,km\,s$^{-1}$}\fi/{\rm kpc})^2$,
for one cluster per halo for sixteen halos.
The top panel is for the z=3 start and the z=8 start below.
}
\label{fig_tides}
\end{figure}
The gravitational potentials at the location of a representative
sixteen clusters in separate halos are
shown in Figure~\ref{fig_pot}. The z=3 simulation shows the pattern expected from accretion with little change
in radial period with time. The z=8 simulation shows a growing main halo that disrupts pre-galactic halos as internal orbital
motion ceases and as the radial oscillation period increases. The largest eigenvalue of the tidal tensors for the same
sixteen clusters is shown in Figure~\ref{fig_tides}. Tidal forces are dominated by the local density field which emphasizes
the variations in the tides due to local orbital motion in the pre-galactic halos, relative the motion in the main halo. The
tides are also somewhat noisy due to motion of nearby individual dark matter particles. The individual dark matter
particles here have a mass of $2\times 10^5{\rm\,M_\odot}$ with a softening of 0.2~kpc, so individual particles
will have a tidal field of about $100 (\ifmmode{\,\hbox{km\,s}^{-1}}\else {\rm\,km\,s$^{-1}$}\fi/{\rm kpc})^2$ or less, with an encounter time of about 1~Myr, which is not resolved
in these plots at intervals of 10~Myr. The clusters are generally in regions where the dark matter particle
density is of order $10^2$ particles per softening length.
\subsection{Numerical Considerations}
The dark matter particles are much heavier and faster moving than the star particles
and will slightly
heat the star particles over the course of the simulation. For encounters
at impact parameter $b$ and velocity $V$, over time $T$ the heating is the kinetic energy change per
encounter, times the rate of encounters and the duration \citep{BT:08}.
The heating rate of the cool, light, star particles, per dynamical time, $t_d=V/R_s$, for a system of size $R_s$ with $N$ heavy
dark matter particles moving at velocities $V$, is then,
\begin{equation}
{d(\Delta V)^2\over dt} = {6 V^2\over N t_d} \ln{\Lambda} .
\label{eq_heat}
\end{equation}
where $\ln{\Lambda}$ is the Coulomb logarithm, with $\Lambda=R_s/s$,
with $R_s$ being about 100 kpc here, and $s$ is the dark matter particle softening.
The VL-2 simulation had a force resolution of 0.162 kpc \citep{VL2}
of with particles of mass $1.8\times 10^5{\rm\,M_\odot}$.
The simulation here has very similar particle properties, with
dark matter particles of mass $2\times 10^5{\rm\,M_\odot}$ and
a softening set to 0.2 kpc, to preserve the halo structure of the original simulation.
The heating of the light, low velocity dispersion, star particles from heavy dark matter particles
over a Hubble time needs to be below the velocity dispersion of the clusters, which
is comparable to the spread of velocities in the tidal streams. The simulation
creates a halo with $\simeq 10^7$ halo particles with typical orbital velocities of 200\ifmmode{\,\hbox{km\,s}^{-1}}\else {\rm\,km\,s$^{-1}$}\fi,
for $\Lambda= 100/0.2$, $\ln{\Lambda}$ of 6, the heating over a Hubble time, about 20 dynamical times
in the inner regions of interest, is about 2\ifmmode{\,\hbox{km\,s}^{-1}}\else {\rm\,km\,s$^{-1}$}\fi. Lower mass cosmological halos have comparable dynamical
times, but fewer particles. Equation~\ref{eq_heat} indicates that heating varies as $V^{-1}$ for fixed $t_d$,
so the heating will be about 6\ifmmode{\,\hbox{km\,s}^{-1}}\else {\rm\,km\,s$^{-1}$}\fi in a 20\ifmmode{\,\hbox{km\,s}^{-1}}\else {\rm\,km\,s$^{-1}$}\fi\ dark matter halo.
The star clusters have characteristic velocity dispersions of 3 and 7\ifmmode{\,\hbox{km\,s}^{-1}}\else {\rm\,km\,s$^{-1}$}\fi, at masses of $10^5{\rm\,M_\odot}$ and $10^6{\rm\,M_\odot}$,
respectively. The heating is not negligible, but does not significantly compromise the results.
The purpose here is to examine the basic effects. Large N simulations
will be done later.
The star clusters inserted into the simulation have an initial mass of $3\times 10^5{\rm\,M_\odot}$, which is the
mass of cluster
at the peak of the current epoch globular cluster mass function. Each cluster in the primary simulations
begins with 20,000 particles, each of mass $15{\rm\,M_\odot}$.
The number of particles and the number
of clusters inserted, about 200, are selected to give reasonable resolution of the tidal tails and a reasonable balance
between dark matter and star particles. Since the entire dark matter distribution needs to be evolved it makes sense to
insert clusters in all the most massive pre-galactic sub-halos which comprise at least half of the dark matter mass.
Trial runs found
that a star particle softening of 0.002~kpc allowed a reasonable time step, the shortest being about $2\times 10^4$ yr.
The star particle softening suppresses relaxation giving the added heating of Equation~\ref{eq_dvheat}
complete control.
Clusters of a fixed initial mass realized with ten times more particles were compared to these cluster
both run in a logarithmic external potential, finding essentially no dependence on particle number.
The Gadget2 parallelized tree code \citep{Gadget2} is very efficient for cosmological simulations.
One complication in these simulations is the factor of $10^4$ scale difference between
the 100~kpc scale of the developing Milky Way-like halo and the star clusters with half mass
radii in the range of 10~pc. As the star clusters orbit through the dark matter background
imbalances in
memory and computational load between the processors develop
requiring relaxing various Gadget2 load balance parameters.
Runs end at 13.4~Gyr, or 13.7 of the calculation's time units which is close to
the Hubble age for Via Lactea.
There are typically a total of approximately 200,000 time steps in
a simulation.
\section{Cluster and Stream Evolution}
The distribution of the dark matter at the end of the
simulation is shown in Figure~\ref{fig_xydend} and the
stars at the same time are shown in Figure~\ref{fig_xysend} for the two runs.
The main dark matter halos in the two runs are comparable, with masses of $9.3\times 10^{11}{\rm\,M_\odot}$
and $7.5\times 10^{11}{\rm\,M_\odot}$, for the z=3 and z=8 start simulations, respectively.
The z=3 start simulation has a more spherical halo.
Not visible in the frame is that both runs have
a second large halo about half the mass of the dominant one, located
at a distance of 780 kpc for the z=8 start and 360 kpc for the z=3 start.
\begin{figure}
\begin{center}
\includegraphics[angle=0,scale=0.8,trim=180 180 180 180,clip=true]{Figure_10.pdf
\end{center}
\caption{The stars at the final moment of the z=3 (top) and z=8 (bottom) start simulations.
Only accreted systems are shown, with stars inserted in the dominant halo not plotted.
The boxes are 50 kpc center to edge or a 100 kpc on a side.
}
\label{fig_xysend}
\end{figure}
The two runs lead to very different cluster mass loss from the star clusters and spatial
distribution of the tidal streams. The clusters in the z=3 run,
top panel of Figure~\ref{fig_xysend}, have only a small
in a diffuse distribution and most of the star clusters remain intact. On
the other hand, the clusters in the z=8 run, bottom panel of Figure~\ref{fig_xysend}, have
a substantial diffuse component and the clusters accreted on the main halo are completely dissolved.
\subsection{Tidal Streams and Remnant Clusters}
The fractional mass remaining in the clusters as a function of time is shown in Figure~\ref{fig_mt}
for a random 10\% subset of the clusters in the two simulations.
Both simulations show a large range in cluster mass loss rates with time,
reflecting the wide range of cluster tidal environments. The z=8 start clusters
are much more likely to be entirely disrupted before the simulation
reaches an age of half a Hubble time, about 7 Gyr. In the z=8 start simulation,
larger halos build up as smaller halos encounter each other on nearly radial orbits,
so that their stellar contents pass through regions of strong tidal fields.
The z=3 start simulation has a large dominant halo into which significantly smaller halos fall
from relatively large distances and either miss or quickly pass through high tidal field regions,
causing far less star cluster disruption.
\begin{figure}
\begin{center}
\includegraphics[angle=-90,scale=0.85]{Figure_11.pdf
\end{center}
\caption{Fractional mass remaining for a randomly selected 10\% subset of
the 198 (red) z=3 start and 186 (green) z=8 start clusters. All clusters
were started with masses of $3\times 10^5{\rm\,M_\odot}$.
The heavy lines are the average fractional mass remaining of the clusters in the two simulations.
}
\label{fig_mt}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[angle=0,scale=0.8,trim=160 180 160 180,clip=true]{Figure_12.pdf
\end{center}
\caption{Radial distribution of the
tidally removed stars (blue) and dark matter (black) for the z=3 (top) and z=8 starts (bottom panel).
}
\label{fig_dens}
\end{figure}
The radial distribution of dark matter (black lines) and stars stripped from the globular clusters (blue lines)
is shown in Figure~\ref{fig_dens} for the z=3 start (top panel) and z=8 start (bottom panel).
The star clusters are in relatively denser regions in the z=8 start than in the z=3 start, therefore relative
to the approximately the same overall dark matter distribution of the Milky Way-like final halo
the stars removed from the z=8 stars remain more bound and are a more centrally concentrated
population than the z=3 remnants.
\begin{figure}
\begin{center}
\includegraphics[angle=0,scale=1,trim=170 240 170 240,clip=true]{Figure_13.pdf
\end{center}
\caption{The stars within 50 kpc (full weight with inverse distance squared weight on more distant star particles) of the galactic center
at the end point for the z=3 (top) and z=8 (bottom) runs
projected on the sky.
Only infall systems are shown, with clusters inserted in the dominant halo not plotted.
}
\label{fig_ah}
\end{figure}
The cluster stars are projected onto the sky in Figure~\ref{fig_ah} using an equal
area Hammer-Aitoff projection. The center is defined as
the center of mass of the clusters inserted into the main halo.
All stars to 50~kpc are plotted with equal weights, with more distant ones reduced with the inverse
square of the distance to allow a somewhat realistic decline in visibility. As
expected from the xy plots the z=3 start produces somewhat longer thin streams than the z=8 start.
Table~\ref{tbl-1} provides an overview of the basic statistics of the tidal streams in the two simulations.
Streams are identified using the current observational
technique, which is to identify relatively straight thin structures by eye.
The minimum angular length that qualifies is approximately 10 degrees.
These images have comparable numbers
of visible particles (a few thousand typically) as real streams have stars.
The simulation has the advantage that there are effectively no
field stars or dust that complicates finding Milky Way streams.
A few interesting conclusions can be taken from these two stream formation
simulations, but a full analysis of stream properties is deferred to a future paper.
In the z=3 simulation the infall clusters remain largely intact, with an average of 55\% of
their initial mass retained. In complete contrast, the z=8 start has no remaining clusters in the inner regions
with more than 0.1\% of
their initial mass.
The complete dissolution of the z=8 clusters may be at least partly due to
fluctuations in the tides and will need to be tested in larger N runs.
The table shows that every z=3 start cluster inside 50~kpc has a visible stream
produces a visible stream, and 67\% of the clusters inside 100~kpc has
a visible stream. The z=8 start leads to a more compact distribution,
with 78 cluster centers (no remnant clusters) inside 50 kpc,
but only 23 streams visible. Comparing the top and bottom panels of Fig~\ref{fig_ah}
shows that the z=8 start streams are generally shorter than the z=3 start ones,
likely a result of the z=3 infall being earlier and relatively smooth as indicated in
the potentials of Fig~\ref{fig_pot}.
These two simulations effectively bracket the redshift
range over which globular clusters accreted onto our halo are formed.
Because not all globular clusters have visible streams
the z=3 start is not a good explanation for all of the halo stream population,
and, because no z=8 clusters survive in the main halo it also fails
to explain the full population, but the combination of the two, and intermediate times,
would be a qualitatively acceptable mix.
The list of \citet{GC:16} finds that only 2 of the 12 streams have ``known or likely globular cluster progenitors", Pal~5 and NGC~5466.
Based on the result here that every accreted z=3 cluster should still
be present with its stream, one infers that
10/12 of the streams have their origin in globular clusters formed
at redshifts greater than 3. Similarly the two streams with remnant clusters are likely to have formed around z=3
and been accreted onto a dominant Milky Way-like halo fairly quickly.
These results depend on the density of the region in which the clusters formed in their pre-galactic halos.
Clusters here were inserted at radii of the 1/4 to 1/2 of the peak of the circular velocity.
If the clusters were inserted yet more deeply into the pre-galactic halos
they would be in stronger tidal fields which would boost
the rate of dissolution upwards to that experienced in higher
redshift start simulations.
The fact that only 2 of the 12 thin streams have progenitor clusters.
combined with the weak mass dependence of the rate of the evolution
of the cluster mass function seen here suggests that most globular clusters
that were accreted onto the halo have dissolved.
Poisson statistics for 2 expected events indicates that up to 4 events contains 95\% of the probability,
or, 2/3 of the cluster population that makes streams is dissolved. With the small sample and
large extrapolation in mind, these numbers indicate that at high redshift that with
fairly high confidence there should be about 3 times more globular clusters that formed than
currently observed.
More recent stream discoveries \citep{PS1,Grillmair:17}
find more streams and no more clearly associated globular clusters.
\begin{table}
\begin{center}
\begin{center}\caption{Cluster and Stream Counts.\label{tbl-1}}\end{center}
\begin{tabular}{lrrrrr}
\tableline\tableline
Start & $M_c$ &$R_g$ max & $N_0$ & $N_f$ & Streams \\
\tableline
z=3 & $3\times 10^5{\rm\,M_\odot}$ &50 kpc & 20 & 20 & 18 \\
& &100 kpc & 67 & 67 & 45 \\
z=8 & $3\times 10^5{\rm\,M_\odot}$&50 kpc & 78 & 0 & 28 \\
& & 100 kpc & 87 & 0 & 29 \\
z=3 & $N \propto M^{-1.5}$ & 35 kpc & 50 & 45 & 28\\
\tableline
\end{tabular}
\end{center}
\tablecomments{Col 1: Starting time;
Col 2: $M_c$, cluster masses; Col 3: $R_g$ max, the maximum distance for full weight of the star particles, more
distant particles are de-weighted with the inverse distance squared; Col 4: $N_0$, the number of cluster centers
that fall into this volume; Col 5: $N_f$, number of clusters with at least 0.1\% of their initial mass remaining,
Col 6: the number of streams visible}
\end{table}
\begin{figure}
\begin{center}
\includegraphics[angle=0,scale=1.,trim=140 185 140 190,clip=true]{Figure_14.pdf
\end{center}
\caption{The angular momentum calculated around the center of rotation
in the inner galaxy for stars within 50~kpc of the
center. The top panel is for the z=3 start, the bottom for the z=8.
The points are color coded by their cluster of origin, cycling over 7 colors.
Approximately horizontal stripes are star streams. Near vertical stripes are
stripped stars orbitting within a sub-galactic halo.
The Aitoff-Hammer $\lambda$ variable has a range of $\pm2\sqrt{2}$ at the equator
but declines to zero at the pole.
}
\label{fig_lt}
\end{figure}
Although a dynamical analysis of streams is left to a future paper, streams are often modeled with
their dynamical variables, such as angular momentum and energy. Figure~\ref{fig_lt} shows the
length of the angular momentum vector of individual star particles from a set of clusters that produce thin streams
in the two simulations. The horizontal variable is the same as in the Hammer-Aitoff plots of
Figure~\ref{fig_ah}, which wraps around the sky. The angular momentum is calculated with
respect to the center of rotation of the inner-most star clusters. The halos here are triaxial, so the
angular momentum will not be a conserved quantity, but it remains sufficiently close to being conserved
that it is useful guide.
Thin tidal streams emanating from star clusters that are orbiting freely in the main
halo are the thin, approximately linear, features in Figure~\ref{fig_lt}. Stripped stars that
are orbiting within a sub-galactic halo create a vertical feature. A dark matter halo
that passes through a stream pulls stream stars towards the centerline of its path, leading
to a sideways ``S'-like feature in the angular momentum of the stream, which with time
develops into a gap in the stream density. Both ``S'' and gap features appear to be present
in Figure~\ref{fig_lt}.
\subsection{Star Cluster Mass Function Evolution}
\begin{figure}
\begin{center}
\includegraphics[angle=0,scale=1,trim=140 185 140 190,clip=true]{Figure_15.pdf
\end{center}
\caption{Mass function of the globular cluster population as
a function of cosmic time for the z=3 (top) and z=8 starts (bottom panel).
In the top panel the times are 2.1 Gyr (black line), 7.9 Gyr (red), and 13.7 Gyr (green).
In the bottom panel the times are 0.7 Gyr (black) and 6.5 Gyr (red), with essentially no clusters
left at the current epoch.
The initial population is proportional to $M^{-1.5}$ distribution between $5\times 10^4$ and
$2\times 10^6 {\rm\,M_\odot}$.
}
\label{fig_nm}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[angle=-90,scale=0.7, trim=100 70 100 70, clip=true]{Figure_16.pdf
\end{center}
\caption{The stars within 35 kpc at the end point for a z=3 run with a
$M^{-1.5}$ mass spectrum of globular clusters over the
range of $0.05-2.0\times 10^6 {\rm\,M_\odot}$.
projected on the sky.
}
\label{fig_ahnm}
\end{figure}
The evolution of the mass function of the globular cluster population is of great interest for comparison to current
day globular cluster systems, and, for estimating the numbers of clusters that were present
at high redshifts, within a few Gyr of formation.
To provide an initial estimate of the mass function
evolution we populate the same dark matter simulations with
stellar clusters drawn from a $n(M)\propto M^{-1.5}$ mass distribution function, limited
to $0.05-2\times 10^6 {\rm\,M_\odot}$. At total of 1433 clusters were inserted, distributed over all halos
in proportion to the halo mass. The ratio of stellar mass to dark matter mass is $3\times 10^{-5}$ in the
final main halo.
The mass functions are shown with time in Figure~\ref{fig_nm}.
The mean number over the inserted mass range declines to about 50\% of the initial population to a time of 13.4 Gyr.
On the other hand, the z=8 population is reduced by half at z=1 and disappears by redshift zero, again with
the caveat that this strong evolution needs further numerical verification.
An interesting outcome is that the mass loss from the clusters has a relatively weak rise towards lower masses.
The significantly more rapid destruction of clusters in the z=8 start as compared to the z=3 start can
be attributed to the significantly stronger tidal forces that the clusters in the high redshift start experience. Although the
z=3 start clusters have a lower rate of mass loss, they also show little mass dependence, suggestive that
tidal heating continues to overwhelm the internal two-body heating.
An important caveat is that these results depend on both the internal heating model used,
and, the fluctuations in the tidal field that the current particle numbers allow.
The sky distribution of the variable mass clusters is shown in Figure~\ref{fig_ahnm}. Here the maximum distance
for full weight visibility is 35 kpc, to be comparable to the \citep{PS1} maps. It is important to note that the most
massive clusters make the most visible streams, because their lose stars at a higher absolute rate.
\citet{PS1} find 15 streams in $3\pi$ of sky to a depth of 35 kpc, although about another half dozen streams are known
from other data. The plane of the Milky Way reduces the effective sky area for stream finding significantly. If it is reduced
to $2\pi$, or half the total sky, then the 28 streams in the simulation,Table~\ref{tbl-1}, is
in reasonable agreement with the streams in the sky, although all of these have remnant clusters in them, which the
real sky does not.
Globular clusters are a site for the creation of some fraction of the
relatively heavy binary black holes that LIGO has detected \citep{LIGO1,LIGO2} through
the gravitational hardening and exchanges of binaries in the cores of globular clusters \citep{PZM:00,Oleary:16,
Mapelli:16,Hurley:16,Rodriguez:15,Rodriguez:16,Askar:17,Park:17}. The binary black holes
that are merging today formed at much earlier times.
The simulations here indicate that the clusters above $10^5{\rm\,M_\odot}$ that have survived are 1/2 or less of
the high redshift globular cluster population when the BBHs were produced.
The prediction of the cluster binary LIGO event rate
is therefore at least twice times the event rate that is predicted based on the numbers of globular clusters, with
the statistics indicating that in the mean a full factor of five more could have been present.
More recent stream discoveries \citep{PS1,Grillmair:17} pushes the number of streams up another 5-6 streams,
with one somewhat marginal case for a progenitor globular cluster \citep{Koposov:14,PS1}.
At present, these numbers rest on 12-18 streams, 2-3 globular clusters, and two simulations. Both
additional simulations and a more quantitative comparison to sky data are needed to harden these conclusions.
\section{Discussion and Conclusions}
Stellar dynamical, globular cluster-like, star clusters are placed in a
in a full cosmological simulation at approximately redshifts 3 and 8.
The simulations give a cosmologically realistic view of how the thin, and not-so thin,
stellar streams from globular clusters are formed and the overall spatial
distribution of the stars evaporated and tidally pulled from globular clusters.
The simulations are gravity-only, with no hydrodynamics or star formation mechanism.
The Gadget2 code is augmented to introduce an appropriate level of two-body driven heating
into the star clusters, with heating parameters calibrated with Nbody6 \citep{Aarseth:99} runs of a few clusters.
The current calibration runs are limited to low mass clusters and one or two orbits. The modified Gadget2 code
is able to reproduce the calibration runs to within about 10\% in mass loss and cluster size. These initial runs
typically have $10^7$ dark matter particles and a comparable but generally somewhat smaller number
of star particles. The particle numbers are sufficient, but just barely, to keep heating of star particles from
dark matter particles below the very low velocity dispersion present in the streams. The resulting star clusters
at the end of the simulation have half mass radii in the range of those seen in the halo, 15-20 pc, but
tend to be somewhat larger. These and other numerical issues will be examined further with additional simulations.
The basic stream morphologies are comparable to observed streams, with the simulations giving thin streams lengths
ranging from around 10 degrees up to about 100 degrees on the sky. Essentially all streams have an
extended diffuse component that is several kpc wide that is composed of stars tidally removed from the clusters
while they were in their pre-galactic halos which upon accretion into the growing Milky Way halo
are spread out over the width of their orbit in the dwarf. The longer thin streams generally exhibit
density variations along their length,
qualitatively as expected on the basis of dark matter sub-halos crossing the streams and creating gaps.
The survival rate of the z=3 start clusters with streams is 100\% inside 50kpc, whereas the z=8 start has no surviving
clusters in the inner halo. The observed streams have 2 or 3 visible progenitor clusters in 12-18 streams \citep{GC:16,PS1}.
Although the statistics have large uncertainties a tentative
conclusion is that for every massive current epoch globular cluster there were at least two, and possibly more, at high redshift.
The mass loss is, appropriately, sensitive to the details of the heating model for the star clusters. The model has been validated for only a few orbits
with low mass clusters. In particular the significant tidally driven evolution of
the massive clusters can be understood as tidally driven, but will be examined further in future n-body simulations.
\acknowledgements
Comments from an anonymous referee spurred substantial revisions to this paper.
This research was supported by NSERC of Canada.
| 83df24ad1f3c1b5b79e35d353e278e47acbee380 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Markov Chain Monte Carlo (MCMC) and Bayesian Statistics
are two independent disciplines, the former being a method to sample from a
distribution while the latter is a theory to interpret
observed data. When these two disciplines are combined
together, the effect is so dramatic and powerful that it has
revolutionized data analysis in almost all disciplines of
science, and astronomy is no exception.
This review explores the power of this combination.
What is so special about MCMC based Bayesian data analysis?
The usefulness of Bayesian methods in science and astronomy
is easy to understand.
In many situations, it is easy to predict the
outcome given a cause. But in science, most often, we are faced with the
opposite question. Given the outcome of an experiment what
are the causes, or what is the probability of a cause as
compared to some other cause? If we have some prior
information, how does that help us? This opposite problem
is more difficult to solve. The power of Bayesian theory lies
in the fact that it provides
a unified framework to quantitatively answer such
questions. Hence it has become indispensable for science.
As opposed to deductive logic, Bayesian theory provides a framework
for plausible reasoning, a concept which is more powerful and
general, an idea championed by \citet{jaynes2003probability}
in his book.
The question now is how does one solve a problem
that has been set up using Bayesian theory.
This mostly involves computing the
probability distribution function (pdf) of some parameters given the data
and is written as $p(\theta|D)$. Here, $\theta$ need not be a
single parameter; in general, it represents a set of parameters.
Usually here and elsewhere, such functions
do not have analytical solutions and so we need methods to
numerically evaluate the distribution. This is where
MCMC methods come to the rescue.
They provide an efficient and easy way to sample
points from any given distribution which is analogous
to evaluating the distribution.
Bayesian data analysis \citep{jeffreys193961} and
Markov Chain Monte Carlo \citep{metropolis1953equation}
techniques have existed for more than 50 years.
Their tremendous increase in popularity over the last decade
is due to an increase in computational power which has made
it affordable to do such computations.
The simplest and the most widely used MCMC algorithm is the ``random
walk'' Metropolis algorithm (\sec{mhalgo}). However, the efficiency of this
algorithm depends upon the ``proposal distribution'' which the user
has to supply. This means that there is some problem-specific
fine tuning to be done by the user. The problem to find
a suitable proposal distribution becomes worse as the
dimensionality of the space over which the sampling is done
increases. Correlations and degeneracies between the
coordinates further exacerbate the problem.
Many algorithms have been proposed to solve this
problem and it remains an active area of research.
Some algorithms work better for specific problems
and under special conditions but algorithms that work
well in general are in high demand.
Multimodal distributions pose additional problems for
MCMC algorithms. In such situations, an MCMC chain can easily
get stuck at a local density maximum. To overcome this,
algorithms like simulated tempering and parallel tempering
have been proposed (\sec{partemp}). Hence discussion of efficient
MCMC algorithms is one focus of this review.
Given its general applicability, the Bayesian framework can be used
in almost any field of astronomy. Hence, it is not possible
to discuss all its applications.
However, there are many examples where either alternatives
do not exist or are inferior. The aim of this review
is to specifically discuss such cases where
Bayesian-MCMC methods have enjoyed great success.
The Bayesian framework by itself is very simple. The difficult
part when attempting to solve a problem is to express the
problem within this framework and then to choose the appropriate
MCMC method to solve it. The best way to master this is by
studying a diverse set of applications, and we aim to
provide this in our review (\sec{casestudy}).
Finally, we also discuss a few advanced topics
like non-parametric models
and hierarchical Bayesian models (\sec{bhm}) which are not yet
main stream in astronomy but are very powerful and allow
one to solve complex problems.
To summarize, the review has three main aims. The first is to explain
the basics of Bayesian theory
using simple familiar problems, e.g., fitting
a straight line to a set of data points with errors in both
coordinates and in the presence of outliers. This is targeted
at readers who are new to the topic. The second goal is to
provide a concise overview of
recent developments. This will benefit people who are familiar
with Bayesian data analysis but are interested in learning
more. The final aim is to discuss emerging ideas that
hold great promise in future.
We also develop and distribute
downloadable software (available at
\verbatim{https://github.com/sanjibs/bmcmc/} or by running the
command \verb|pip install bmcmc|)
implementing
some of the algorithms and examples that we discuss.
\begin{figure}
\centering \includegraphics[width=0.95\textwidth]{metropolis_citations.pdf}
\caption{Citation history for the
\citet{metropolis1953equation} paper for three different
subject areas.
\label{fig:metropolis_citations}}
\end{figure}
\subsection{Rise of MCMC based Bayesian methods in astronomy and science}
The emergence of Bayesian statistics has a long and
interesting history dating back to 1763 when
Thomas Bayes laid down the basic ideas
of his new probability theory \citep[][published
posthumously by Richard Price]{mr1763essay}.
It was rediscovered
independently by Laplace \citep{de1774memoire}
and used in a wide variety of contexts, e.g., celestial
mechanics, population statistics, reliability, and
jurisprudence. However, after that it was largely ignored.
A few scientists like,
Bruno de Finetti and Harold Jeffreys kept the
Bayesian theory alive in the first half of the 20th
century. Harold Jeffreys published
the book Theory of Probability \citep{jeffreys193961},
which for a long time remained the
main reference for using the Bayes theorem.
The Bayes theorem was used in the Second World War at
Bletchley Park, United Kingdom, for cracking the German Enigma code,
but its use remained classified for many years afterwards.
From 1950 onwards, the tide turned towards
Bayesian methods. However, the lack of proper tools to do
Bayesian inference remained a challenge. The frequentist
methods in comparison were simpler to implement which made them more
popular. Recent statement by the American Statistical
Association, \citep{wasserstein2016asa} warning on the
misuse of P values is another example of the superiority of
the Bayesian methods of hypothesis testing.
Interestingly, efficient methods like MCMC
to sample distributions had been invented by 1954
in the context of solving problems in statistical
mechanics \citep{metropolis1953equation}. (The brand name Monte Carlo was coined by
\citet{metropolis1949monte} where they discussed
a stochastic method making use of random numbers to
solve a class of problems in mathematical physics
which are difficult to solve due to the large number
of dimensions.)
Such problems typically involve $N$ interacting particles.
A single configuration $\omega$ of such a system is fully specified by
giving the position and velocity of all the particles;
i.e., $\omega$ can be defined by a point in
$\mathcal{R}^{2N}$ space, also known as the
configuration space $\Omega$. The total energy is
a function of the configuration $E(\omega)$. For a system in
equilibrium, the probability of a configuration is given by
$\exp(-E(\omega)/kT)$, where $k$ is the Boltzmann constant and
$T$ is the temperature of the system.
Computing any thermodynamic property of the system, e.g.,
pressure or energy typically involves computing integrals
of the form
\begin{eqnarray}} \def\ee{\end{eqnarray}
\bar{F}=\frac{\int F(\omega) \exp(-E(\omega)/kT) {\rm d} \omega}{Z}
\ee
for which $Z=\int \exp(-E(\omega)/kT) d \omega$ is known
as the partition function. The integrals over $\omega$ are
in most cases analytically and computationally intractable.
The idea of Metropolis and colleagues was to start with an arbitrary
configuration of $N$ particles and then move each particle
by a random walk. If $\Delta E<0$, the move is always accepted,
otherwise, it is accepted stochastically with
probability $\exp(-\Delta E/kT)$, which is the
ratio of the probability of the new configuration
with respect to the old.
The method ends
up choosing a configuration $\omega$ sampled
from $\exp(-E(\omega)/kT)$.
The method immediately became popular in the
statistical physics community.
However, the fact
that the same method can be used for sampling
an arbitrary pdf $p(\omega)$
by simply replacing $E(\omega)/kT $
with $\ln(p(\omega))$ had to wait till the
important paper by \citet{hastings1970monte}.
He generalized the work of Metropolis and colleagues and derived the
essential condition for the acceptance ratio
that a Markov chain ought to satisfy in order
to sample the target distribution.
The generalized algorithm is
now known as the Metropolis-Hastings (MH) algorithm.
Later Hastings's student
Peskun showed that,
among the available choices, the one by Metropolis
and colleagues was the most efficient
\citep{peskun1973optimum}.
Despite its introduction
to the statistical community, the ideas remained
dormant till 1980.
Around 1980 things suddenly changed and a few influential algorithms
appeared.
Simulated annealing was presented
by \citet{kirkpatrick1983optimization}
to solve combinatorial optimization problems
using the MH algorithm in conjunction with
ideas of annealing from solid state physics.
It is especially useful for situations where we have
multiple maxima and applies to any setting
where we have to minimize an objective function $C(\omega)$.
This is done by sampling $\exp(-C(\omega)/T)$, with
progressively decreasing $T$ to allow annealing and
selection of a globally optimum solution.
A year later
\citet{geman1984stochastic} introduced what we currently
know as ``Gibbs sampling'' in the context of image
restoration.
This was the first proper use of MCMC techniques to solve a problem
set up in a Bayesian framework, in the sense that simulating from conditional
distributions is the same as simulating from the joint
distribution. However, there exists earlier
work related to Gibbs sampling;
the Hammersley-Clifford theorem which was developed in
the early 1970s and the work by \citet{besag1974spatial}.
At about this time, one of the most influential
methods of the 20th-century emerged $-$
the expectation maximization (EM)
algorithm by \citet{dempster1977maximum}.
This provided a way to deal with missing data
and hidden variables and vastly
increased the range of problems that can be
addressed by Bayesian methods.
The EM algorithm is deterministic and
has some sensitivity to the starting configuration.
To address this, stochastic versions were
developed \citep{celeux1985sem} quickly followed
by the data augmentation (DA)
algorithm \citep{tanner1987calculation}.
The watershed moment in the field of statistics is largely
credited to the paper by \citet{gelfand1990sampling} that
unified the ideas of Gibbs sampling, DA and
the EM algorithm \citep{tanner2010data,robert2011short}. It firmly
established that Gibbs sampling and Metropolis-Hastings based
MCMC algorithms can be used to solve a wide class of
problems that fall into the category of hierarchical
Bayesian models. The citation history of the famous
\citet{metropolis1953equation} paper shown in \fig{metropolis_citations}
corroborates the historical narrations on this topic.
In physics, the MH
algorithm was well known in the period 1970-1990, but
this was not so in statistics or astronomy.
In astronomy,
a watershed moment can be seen in 2002; this is visible
more clearly in \fig{arxiv_stats} where we track the usage of the words {\it MCMC} and {\it Bayesian}.
But prior to 2002, the Bayesian-MCMC technique was not unknown to
the astronomy community. We can see its use in
\citet{saha1994unfolding} who applied it to extract galaxy kinematics from
absorption line spectra.
Further seeds were planted down the line by
\citet{1998PhRvD..58h2001C} while studying gravitational
wave radiation, and then by \citet{2001CQGra..18.2677C}
and \citet{2001ApJ...563L..95K} in
the context of cosmological parameter estimation using
cosmic microwave background data. Inspired by these papers,
\citet{2002PhRvD..66j3511L}
more than any other paper
seems to have galvanized the astronomy community
in the use of Bayesian and MCMC techniques.
They laid out in detail the Bayesian-MCMC framework, applied it to one of the most
important data sets of the time (cosmic background
radiation) and used it to address a significant scientific
question $-$ the fundamental parameters of our universe. Additionally,
they made their MCMC code publicly available, which
was instrumental in lowering the barrier for new entrants
to the field.
\begin{figure}
\centering
\begin{minipage}{6cm}
\centering
\includegraphics[width=6.0cm]{arxiv_Bayesian.pdf}
\end{minipage}
\centering
\begin{minipage}{6cm}
\centering
\includegraphics[width=6.0cm]{arxiv_MCMC.pdf}
\end{minipage}
\caption{Percentage of articles in Arxiv astro-ph abstracts
containing the word Bayesian (left) and MCMC
(right). Computed using the code {\it arxiv.py}, courtesy Dustin Lang.}
\label{fig:arxiv_stats}
\end{figure}
\section{Bayesian Data Analysis}
In this section we briefly review the basics of the Bayesian
theory. We start with the Bayes theorem and then use it to
set up the problem of fitting a model to data. This is
followed by a discussion of the
role of priors in Bayesian analysis.
Next, the Bayesian solution of fitting a straight line is
discussed in detail to illustrate the ideas discussed.
Finally, we show how to perform model selection.
To further explore the topics discussed here,
many excellent resources are available.
A stimulating discussion on Bayesian theory can be
found in \citet{jaynes2003probability}.
\citet{sivia2006data} and \citet{gregory2005bayesian}
are excellent textbooks with a good emphasis on
applications in science. \citet{2010arXiv1008.4686H}
provides lucid tutorial on fitting models to data.
A fascinating discussion
on Bayesian versus frequentist approaches to solving
problems can be found in \citet{loredo1990laplace}.
A review with emphasis on cosmology is given by
\citet{2008ConPh..49...71T}.
\subsection{Bayes' Theorem}
\citet{cox1946probability} showed that the rules of Bayesian
probability theory can be derived from
just two basic rules:
\begin{eqnarray}} \def\ee{\end{eqnarray}
& p(H|I)+p(\bar{H}|I) = 1 &\textrm{ Sum Rule},\\
& p(H,D|I) = p(H|D,I)p(D|I)=p(D|H,I)p(H|I) &\textrm{ Product Rule}.
\ee
Here $H$ stands for some proposition being true
and $D$ stands for some other proposition being true,
and
$\bar{H}$ means the proposition $H$ is false.
So the sum rule just states that the probability of a proposition
being true plus the probability of it being false is unity.
The product rule expresses the joint probability of two
propositions being true in terms of conditional
probabilities, one being true given the other is true.
The vertical bar $|$ is a conditioning symbol and means
`given'. $I$ denotes relevant background information that is
used to construct the probabilities.
The product rule leads to the Bayes Theorem
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(H|D,I)=\frac{p(D|H,I)p(H|I)}{p(D|I)},\quad
{\rm Posterior=\frac{Likelihood \times Prior}{Evidence}},
\ee
where we identify $H$ with the hypothesis and $D$ with the
data.
The $p(D|H,I)$ is the probability of observing the data
$D$ if the hypothesis is true and is known as the {\it likelihood}.
The quantity $p(H|I)$ is
the {\it prior} and specifies our prior knowledge of $H$
being true. The
$p(H|D,I)$, known as {\it posterior}, expresses our updated
belief about the truth of the hypothesis in
light of the data $D$. The quantity $p(D|I)$ is a constant and
serves the purpose of normalizing $\int p(H|D,I)\:{\rm d}H$
to 1. It is known as the {\it evidence}.
Another important result that can be derived from the sum
rule and the product rule is the marginalization
equation,
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(X|I)=\int p(X,Y|I)dY=\sum_i p(X,Y_i|I).
\ee
First let us write the sum rule in an alternate form.
Instead of considering just $Y$ and $\bar{Y}$, we consider
a set of possibilities $\{Y_i\}$ that are mutually exclusive.
\begin{eqnarray}} \def\ee{\end{eqnarray}
\sum p(Y_i|I)=1, &\textrm{ Extended Sum Rule}.
\ee
Now, making use of the product rule and the sum rule we get
\begin{eqnarray}} \def\ee{\end{eqnarray}
\sum_i p(X,Y_i|I) &= \sum_i p(Y_i|X,I)p(X|I) & \textrm{ using product rule} \\
&= p(X|I) \sum_i p(Y_i|X,I) = p(X|I) & \textrm{ using sum rule}.
\ee
\subsection{Fitting a model to data}
Typically, we have some data and we want to use it for
scientific inference. One of the most effective approaches to
dealing with such problems is to develop a model
that describes how the data were created. Let $\theta$ be the
set of parameters of the model and $x^t$ a data
point generated by the model according to $f(x^t|\theta)$.
The observed data points $x$ can have some measurement
errors, described by a parameter $\sigma_x$. The probability
of the observed value is then given by $p(x|x^t,\sigma_x)$, which could be
$\mathcal{N}(x|x^t,\sigma_x^2)$ for Gaussian errors; hereafter,
$\mathcal{N}(.|\mu,\sigma^2)$ refers to a normal distribution
with mean $\mu$ and variance $\sigma^2$. The
probability of observed data point $x$ given a model
and an error is then
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(x|\theta,\sigma_x)=\int f(x^t|\theta)p(x|x^t,\sigma_x) dx^t.
\ee
We have integrated over true values $x^t$ which are
unknown.
If we have reason to believe that there are
outliers in the data, e.g., a fraction of points are not
described by the model, we can supplement a background model
$f_b(x^t|\theta_b)$ with probability $P_b$ and parameters
$\theta_b$ \citep{press1997understanding,2010arXiv1008.4686H}. The probability of the observed data points
can then be written as,
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(x|\theta,\theta_b,P_b,\sigma_x) & = & \int
\left[P_b f_b(x^t|\theta_b)+(1-P_b)f(x^t|\theta)\right]p(x|x^t,\sigma_x)
dx^t \\
& = & p(x|\theta_b)P_b +p(x|\theta)(1-P_b).
\ee
The total probability for a set of $N$ data points $X=\{x_1,...,x_N\}$ is then
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(X|\theta,\theta_b,P_b,\sigma_x) & = & \prod_{i=1}^{N} p(x_i|\theta,\theta_b,P_b,\sigma_{x,i}).
\ee
To infer the model parameters, one uses the Bayes theorem
and computes
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta,\theta_b,P_b|X,\sigma_x) & \propto & p(X|\theta,\theta_b,P_b,\sigma_x)p(\theta,\theta_b,P_b).
\ee
Here, $p(\theta,\theta_b,P_b)$ represents our prior
knowledge about the parameters. We discuss
this in detail in the next section.
\begin{figure}
\centering \includegraphics[width=0.95\textwidth]{stline_b1.pdf}
\caption{Fitting a straight line to data with outliers. The
outliers are shown as red points. The dashed line is the
best fit line when an outlier model is not used. The solid
line is the best fit line with an outlier model. The data
was generated with model parameters $m=2$ and $c=10$. 20\%
of the points were set as outliers and sampled from $\mathcal{N}(30,5^2)$.
\label{fig:stline_b1}}
\end{figure}
We consider the problem
of fitting a straight line
with equation $y=mx+c$ to some data points
$X=\{x_1,...,N\}$ and $Y=\{y_1,...,N\}$, with uncertainty
$\sigma_{y,i}$ on the $y$ ordinate. We generated 50 data
points with $m=2.0$ and $c=10.0$; 20\% of the data points
were set as outliers and were sampled from
$\mathcal{N}(30,5^2)$.
To simulate random uncertainty
the $y$ ordinate was scattered with a Gaussian function
having dispersion in range $0.25<\sigma_y<1.25$.
The data along with the results of our fitting exercise are
shown in \fig{stline_b1}.
The image shows the outliers and
data sampled from a straight line.
We first fitted a simple model without taking the
outliers into account (dashed line).
Here, $\theta=\{m, b\}$ and the generative model of the data
is
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(y_i|m,c,x_i,\sigma_{y,i})=\frac{1}{\sqrt{2\pi}\sigma_{y,i}}\exp\left(-\frac{(y_i-mx_i-b)^2}{2 \sigma_{y,i}^2}\right)
\ee
It can be seen that the ``best fit'' line is not a good
description for the data points that were sampled from
a straight line.
Next, we extended the model by adding a model for the outliers as
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(y_i|\mu_b,\sigma_b,x_i,\sigma_{y,i})=\frac{1}{\sqrt{2\pi(\sigma_{y,i}^2+\sigma_b^2)}}
\exp\left(-\frac{(y_i-\mu_b)^2}{2 (\sigma_{y,i}^2+\sigma_b^2)}\right)
\ee
The full model being
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(Y|m,c,P_b,\mu_b,\sigma_b,X,\sigma_y)=\prod_{i=1}^N
[p(y_i|\mu_b,\sigma_b,x_i,\sigma_{y,i})P_b+p(y_i|m,c,x_i,\sigma_{y,i})(1-P_b)] \label{equ:outliers}
\ee
The best-fit line resulting from this model obtained
by sampling the posterior distribution using a Markov Chain
Monte Carlo scheme
is shown in \fig{stline_b1}. The best-fit parameters of
the model resemble well the true parameters that were used
to create the synthetic data set (the example is implemented
in the software that we provide).
\subsection{Priors}
Priors are one of the most important ingredients of the
Bayesian framework.
Priors express our present state of knowledge about the
parameters of interest, which we wish to constrain
by analyzing new data.
In a multi-dimensional parameter space, it is quite common
to have degeneracies among the parameters. Here
priors can play a crucial rule in restricting the posterior
to a small region of the parameter space as compared to
the much larger region allowed by the likelihood function.
Priors can be broadly classified into two types,
uninformative and informative.
Uninformative priors express our state of ignorance
and have very little restricting power. They are also known as
ignorance prior. Typically
their distributions are diffuse. Informative
priors on the other hand
By contrast, informative priors
are typically very restricting.
They might come from the analysis of some previous data.
They are important when the data alone are
not very informative and without
external information the data cannot adequately constrain
the parameters being investigated.
Ignorance priors are used in cases where
we have very little knowledge about the parameters we want to
constrain, and we wish to express
our ignorance by using uninformative priors.
Certainly a prior with sudden jumps or oscillating features
is too detailed for expressing ignorance! So smoothness is
certainly an important criterion for an ideal uninformative
prior. In fact, if the data are informative,
almost any prior that is sufficiently smooth in the region
of high likelihood will lead to very similar conclusions.
Is there a formal and unique way to express our ignorance?
A number of techniques exist for constructing ignorance
priors. We here discuss a few simple and commonly used ones;
for a detailed review see \citet{kass1996selection}.
The simplest is Laplace's principle of insufficient
reason which assigns equal probability to all possible
values of the parameter. If the parameter space
consists of a finite set of points, then it is easy to
apply this principle. However, for a continuous parameter
space, the prior depends upon the chosen partitioning
scheme.
Ignorance priors can also be specified using the
invariance of the likelihood
function, $p(x'|\theta'){\rm d}x'=p(x|\theta){\rm d}x$, under the action of a
transformation group $(x',\theta')=h(x,\theta)$, e.g., translation, scaling or
rotation of coordinates. If the priors are really
uninformative, consistency demands that
we should make the same Bayesian inference,
which implies that the priors should also be invariant to the transformation
and satisfy $p(\theta'){\rm d}\theta'=p(\theta){\rm d}\theta$
\citep{jaynes2003probability}.
For two special types of parameters, this leads to unique choices
for expressing ignorance. These are the {\it location
parameters} and the {\it scale parameters}. An example is
the mean $\mu$ and dispersion $\sigma$
of a normal distribution $\mathcal{N}(x|\mu,\sigma^2)$ which
are the location and the scale parameters respectively.
The likelihood $\mathcal{N}(x|\mu,\sigma^2)$ is invariant
under transformation $(x',\mu')=(x+b,\mu+b)$, demanding
invariance for the prior leads to $p(\mu)={\rm constant}$.
Similarly, $\mathcal{N}(x|\mu,\sigma^2)$ is also invariant
under $(x'-\mu',\sigma')=(a(x-\mu),a\sigma)$, which leads
to $p(\sigma) \propto 1/\sigma$. In general, $\mu$ and
$\sigma$ are
location and scale parameters if likelihood is of the form
$f((x-\mu)/\sigma)/\sigma$.
Another commonly used technique to specify ignorance priors is the Jeffreys rule,
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta) \propto \det (\mathcal{I}(\theta))^{1/2}
\mathrm{,\ where\ }
[\mathcal{I}(\theta)]_{ij}=\int p(x|\theta) \frac{\partial^2
}{\partial
\theta_i \partial \theta_j} \ln p(x|\theta) {\rm d}x
\ee
is the Fisher information matrix and $\theta$ a vector of
parameters. It is based on the idea that the prior should
be invariant to reparameterization of $\theta$. Applying it
to the case where the likelihood is a normal
distribution $\mathcal{N}(x|\mu,\sigma^2)$, gives
$p(\mu)= {\rm constant}$ (for a fixed $\sigma$) and
$p(\sigma)= 1/\sigma$ (for a fixed $\mu$). However,
when applied to both $\mu$ and $\sigma$ together,
it gives $p(\mu,\sigma)=1/\sigma^2$. To avoid this
contradiction the rule was modified to
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\mu_1,..\mu_k,\theta) \propto \det
(\mathcal{I}(\theta))^{1/2},
\ee
where $\mu_i$ are location parameters
and $\mathcal{I}(\theta)$ is calculated keeping them fixed.
The principle of maximum entropy
\citep{Jaynes1957information} is also
helpful for selecting priors. Suppose we are interested
in knowing the pdf of a variable, e.g., the probability of a given
face of a six-faced die landing up.
Suppose we also have some macroscopic constraint
available to us, e.g., the mean value obtained when the die is rolled a
large number of times. Such a constraint cannot uniquely
identify a pdf but can be used to rule out a number of
pdfs. The principle says that out of all possible
pdfs satisfying the constraint,
the most likely one is the one having maximum entropy,
where the entropy is defined as $S=-\sum p_i\log p_i$.
We now use this principle to derive the most likely
distribution of a variable for two common cases.
\begin{itemize}
\item If for a variable $x$ we know the expectation
value $\bar{x}$ and the fact that it lies in the range $[0,\infty]$ then
the maximum entropy distribution of $x$ is
$p(x|\bar{x}) = \exp(-x/\bar{x})/\bar{x}. $
\item If $\bar{x}$ and variance
$\sigma^2=\langle (x-\bar{x})^2\rangle$
are known, then
$p(x|\bar{x},\sigma)=\frac{1}{\sigma\sqrt{2\pi}}\exp\left(-\frac{(x-\bar{x})^2}{2
\sigma^2}\right).$
\end{itemize}
\subsection{Fitting a straight line}
We now consider the Bayesian solution for fitting a straight
line in detail \citep[see also][]{jaynes1991straight,2010arXiv1008.4686H}.
We first discuss the solution for the general
case where we have uncertainties on both $x$ and $y$ coordinates
and then discuss the case where the uncertainties are
unknown.
Suppose we have a collection of points $(X=\{x_1,...,x_N\}$,
$Y=\{y_1,...,x_N\})$,
with uncertainties
$\mathbf{\Sigma}=\{\mathbf{\Sigma}_1,...,\mathbf{\Sigma}_N\}$.
Here $\mathbf{\Sigma}_i$ is the covariance matrix defined as
\begin{eqnarray}} \def\ee{\end{eqnarray}
\mathbf{\Sigma}_i= \left[ {\begin{array}{cc}
\sigma_{x,i}^2 & \sigma_{xy,i}^2\\
\sigma_{xy,i}^2 & \sigma_{y,i}^2\\
\end{array} } \right].
\ee
We want to fit a line $y=ax+b$ to these data.
For the time being, we assume $\mathbf{\Sigma}_i$ to be a
diagonal matrix with $\sigma_{xy,i}=0$.
Let $(x,y)$ be the true values
corresponding to the point $(x_i,y_i)$.
Then the probability of measuring the point $(x,y)$ at
$(x_i,y_i)$ is
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(x,y|x_i,y_i,\sigma_{x,i},\sigma_{y,i}) & = & \frac{1}{2 \pi
\sigma_{x,i}\sigma_{y,i} }\exp\left(-\frac{(x-x_i)^2}{2\sigma_{x,i}^2}-\frac{(y-y_i)^2}{2\sigma_{y,i}^2}\right).
\ee
Let us consider a generative model for the line.
We consider the pdf of
a line to be described by a Gaussian with width $\sigma_p$
along a direction perpendicular to the line and width $\sigma_h$
along the line. Here, $\sigma_p$ can be thought of as
an intrinsic scatter about the linear relation that
we wish to investigate.
In the limit $\sigma_h \to \infty$, the
probability of a point $(x,y)$ to be sampled from this
generative model is
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(x,y|a,b,\sigma_p) & = & \frac{1}{\sqrt{2 \pi}
\sigma_p}\exp\left(-\frac{(y-(ax+b))^2}{(1+a^2)2
\sigma_p^2}\right).
\ee
Hence, the probability of $(x_i,y_i)$ being sampled from the
generative model of the line is
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(x_i,y_i|a,b,\sigma_{x,i},\sigma_{y,i},\sigma_p) && =
\int \int p(x,y|a,b,\sigma_p)
p(x,y|x_i,y_i,\sigma_{x,i},\sigma_{y,i}) \ {\rm d}x{\rm d}y
\\
&& = \frac{1}{\sqrt{2 \pi
(\sigma_{\perp,i}^2+\sigma_p^2)}}\exp\left(-\frac{d_i^2}{2(\sigma_{\perp,i}^2+\sigma_p^2)}\right)
\ee
where
$\sigma_{\perp,i}=\sqrt{(\sigma_{y,i}^2+a^2\sigma_{x,i}^2)/(1+a^2)}$
is the component of the error vector perpendicular to the
line, and $d_i=(y_i-a x_i-b)/\sqrt{1+a^2}$ is the perpendicular
distance of the point from the line.
For a general matrix $\mathbf{\Sigma}_i$,
$\sigma_{\perp,i}=\hat{\mathbf{u}}^T\mathbf{\Sigma}_i
\hat{\mathbf{u}}$
for which $\hat{\mathbf{u}}=(-a/\sqrt{1+a^2},1/\sqrt{1+a^2})$ is a unit vector
perpendicular to the line.
For the full sample,
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(X,Y|\mathbf{\Sigma},a,b,\sigma_p) & = & \prod_i^N
\frac{1}{\sqrt{2 \pi(\sigma_{\perp,i}^2+\sigma_{p}^2)}}\exp\left(-\frac{d_i^2}{2(\sigma_{\perp,i}^2+\sigma_{p}^2)}\right).
\ee
\begin{figure}
\centering \includegraphics[width=0.75\textwidth]{stline_prior.pdf}
\caption{Prior for slope of a straight line. The left panel
represents a prior uniform in slope $a$, straight lines
with constant interval $\Delta a$. The right panel
represents a prior symmetric with respect to rotation,
straight lines with constant interval $\Delta \theta$ and
slope $a=\tan(\theta)$.
\label{fig:stline_prior}}
\end{figure}
If we desire to compute $a$ and $b$, then
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(a,b|X,Y,\mathbf{\Sigma}) & = &
A p(a,b) p(X,Y|\mathbf{\Sigma},a,b).
\ee
Henceforth, $A$ is a normalization constant which may be
different in different equations.
The $p(a,b)$ is the prior distribution of parameters of the
line. The two common
choices for the prior are the uniform (flat) and Jeffreys
prior. Neither is appropriate.
Given the rotational symmetry in the problem, a
sensible choice is to have priors that are
symmetric with respect to rotation. Let $\theta=\tan^{-1}a$
be the angle made by the
line with $x$ axis, and $b_{\perp}=b\cos(\theta)$ be the
distance of the line from the origin.
A uniform prior on $\theta$ and $b_{\perp}$ is symmetric
with respect to rotation. This leads to
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(a,b)\:{\rm d}a\:
{\rm d} b=\frac{{\rm d}\theta}{\pi}\frac{{\rm
d}b_{\perp}}{2B_{\perp}}=\frac{1}{(1+a^2)^{3/2}}\frac{{\rm d}a\:{\rm d}b}{2B_{\perp}\pi}
\ee
In \fig{stline_prior}, we graphically show how a prior
uniform in $a$ differs from a prior uniform
in $\theta=\tan^{-1}(a)$. In the left panel, we show straight lines
uniformly spaced in $a$. The
lines tend to crowd at high value of $a$, and this can
bias the estimate of the slope $a$. In the right panel,
the lines are uniformly spaced in $\theta$,
and there is no crowding effect.
The log-likelihood of the full solution after taking
the prior into account is
\begin{eqnarray}} \def\ee{\end{eqnarray}
\ln L & =& \ln p(a,b|\{x_i\},\{y_i\},\mathbf{\Sigma}_i,\sigma_p) \\
& = & K -\frac{3}{2}\ln (1+a^2) -\sum_{i=1}^{N}\frac{1}{2}\ln (\sigma_{\perp,i}^2+\sigma_{p}^2)-\sum_{i=1}^{N}\frac{d_i^2}{\sigma_{\perp,i}^2+\sigma_{p}^2}.
\ee
We now study the case where $\mathbf{\Sigma_i}$ is unknown and
$\sigma_p=0$.
For simplicity, we assume the uncertainty is the same for all data
points, i.e., $\sigma_{\perp,i}=\sigma_{\perp}$.
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(a,b,\sigma_\perp|\{x_i\},\{y_i\}) & = &
p(a,b,\sigma_\perp) p(\{x_i\},\{y_i\}|\sigma_{\perp},a,b) \\
& = & p(a,b)p(\sigma_\perp|a,b) \frac{1}{(2 \pi
\sigma_{\perp}^2)^{N/2}}\exp\left(-\frac{\sum_i d_i^2}{2\sigma_\perp^2}\right)
\ee
and integrate over
$\sigma_\perp$ using Jeffreys prior $p(\sigma_\perp|a,b)=1/\sigma_{\perp}$ to arrive at
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(a,b|\{x_i\},\{y_i\}) & = & A p(a,b)\left(\sum_i d_i^2\right)^{-N/2}
\ee
So, if we ignore the prior factor, the best fit line is
simply the line that minimizes the sum of the squared
perpendicular distances of points from the line.
\subsection{Model comparison}\label{sec:modelcomp}
When we have multiple models to explain data, we are faced
with the question of which model is better. There is no
unique definition of better and depending upon
what we mean by better we can come up with different
criteria to compare models. We have two main schools of
thought, a) to compare the probability of the model given
the data and b) to compare the expected predictive accuracy of the
model for the future data. The former is inherently
a Bayesian approach and is known as Bayesian model
comparison. The latter is inspired by frequentist ideas but
can also be argued from a Bayesian perspective
\citep{vehtari2012survey, gelman2014understanding_waic}.
\subsubsection{Bayesian model comparison}
In the Bayesian formulation, the usefulness of a model is
indicated by the probability of a model $M$ given the data $D$,
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(M|D)=\frac{p(D|M)p(M)}{p(D)}.
\ee
The prior model probability $p(M)$ is generally assumed to
be unity. Note in some cases it may not be so, and we might have
more reason to believe one model over the other.
The $p(D)$ is the same for all models, so it is irrelevant
when comparing models. Thus the main thing we need to compute
is the evidence $p(D|M)$ (also know as marginal likelihood).
Hence, for two models $M_1$ and $M_2$, the odds ratio in favor of
$M_2$ compared to $M_1$ is mainly determined by the
ratio of their evidences, $B_{21}$, also known as the
``Bayes factor'' \citep[for a review and a guide to
interpreting the Bayes factor, see][]{kass1995bayes}.
\begin{eqnarray}} \def\ee{\end{eqnarray}
\frac{p(M_2|D)}{p(M_1|D)} & = & \frac{p(D|M_2)p(M_2)}{p(D|M_1)p(M_1)}=\frac{p(D|M_2)}{p(D|M_1)}=B_{21}
\ee
For some given data $D$ and a model $M$ parameterized by $\theta$, we have
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta|D,M)=\frac{p(D|\theta,M)p(\theta|M)}{p(D|M)}.
\label{equ:bayesfac1}
\ee
The evidence $p(D|M)$ appears as the denominator on the right
hand side and can be obtained by integrating
both sides of \equ{bayesfac1} over all $\theta$. For
properly normalized quantities, the left
hand side integrates to unity, leading to
$p(D|M)=\int p(D|\theta,M)p(\theta|M) {\rm d}\theta$.
Note, the Bayes factor depends upon the adopted range of the
prior which leads to some
conceptual difficulties \citep[see the paradox in][]{lindley1957statistical}.
The range of prior
is not an issue for parameter estimation but it is for model
selection; we cannot use improper priors.
In most cases, we do have a reasonable sense of the range of
priors and they are unlikely to extend to infinity.
To better understand the role of priors, consider two models
$M_1$ and $M_2$, where $M_2$ has a free parameter $\theta$,
while $M_1$ has no free parameter (with $\theta$
being fixed to $\theta_0$).
Let $\Delta
\theta_{\rm likelihood}$ be the characteristic width of the likelihood
distribution and $\Delta \theta_{\rm prior}$ the range of a uniform
prior which encloses the likelihood peak.
The Bayes factor
in favor of model $M_2$ as compared to $M_1$ is then
\begin{eqnarray}} \def\ee{\end{eqnarray}
B_{21} &=& \frac{p(D|M_2)}{p(D|M_1)} = \frac{\int p(D|\theta)p(\theta) {\rm d}\theta}{L(\theta_{\rm
0})}
= \frac{\int p(D|\theta)
{\rm d}\theta}{L(\theta_{\rm
0}) \Delta \theta_{\rm prior}} = \frac{L(\theta_{\rm
max})}{L(\theta_{\rm 0})} \frac{\Delta \theta_{\rm likelihood}}{\Delta \theta_{\rm prior}}.
\ee
The first term on the right hand side will in general be
greater than one and will favor $M_2$, as the simpler model $M_1$ is a
special case of $M_2$. However, the second term penalizes $M_2$
if it has a large range in priors.
The conceptual difficulty associated with the dependence of
the Bayes factor on the adopted prior range is alleviated if one thinks
of hypothesis as a specification of a model as well as the prior on
its parameters. A model $M_2$ with a larger range in priors allows
for a larger number of possible data sets consistent with
the hypothesis as compared to a simpler model $M_1$ with narrow range
of prior. Hence $p(D|M_2)$, being a normalized probability
over possible data sets, will be lower as compared to
$p(D|M_1)$ \citep{mackay2003information}.
Also, $M_1$ is more precise as a hypothesis as compared with $M_2$.
If we have more free
parameters in a model, the penalty term in the Bayes factor
will be
higher, being of the form $\prod_{i=1}^{d} \Delta
\theta_{\rm likelihood,i}/\Delta \theta_{\rm prior,i}$.
In this sense, the Bayes factor has a
built-in safeguard to prevent overfitting
(a model with a large number of free parameters will
fit a given set of data better but will perform poorly when
presented with new data).
\begin{marginnote}
\entry{BIC}{Bayesian information criterion}
\entry{WBIC}{Widely applicable Bayesian information criterion}
\end{marginnote}
Computing the Bayes factor or the Bayesian evidence
is computationally challenging.
Generally, the likelihood is
peaked and confined to a narrow region in the
prior range, but has long tails whose contributions cannot be
neglected. Some commonly employed numerical techniques are
(1) simulated annealing, (2) nested sampling,
(3) Laplace's approximation, (4) Lebesgue integration
theory \citep{weinberg2012computing}, and (5) the
Savage-Dickey density ratio \citep{verdinelli1995computing}.
Two useful approximations of the Bayes free energy $\mathcal{F}=-\ln
p(D|M)$ are
\begin{eqnarray*}
{\rm BIC}/2 =& -\ln p(Y|\hat{\theta})+(d \ln n)/2 &\textrm{\citep{schwarz1978estimating}}\\
{\rm WBIC}/2 =& \mathbb{E}_{\theta}^{\beta}[- \ln p(Y|\theta)]
\mathrm{\ where\ }
\beta=\frac{1}{\ln n} &\textrm{\citep{watanabe2013widely}},
\end{eqnarray*}
Here, ${\rm E}_{\theta}^{\beta}$ denotes expectation taken over the
posterior distribution $p(\theta|Y) \propto p(Y|\theta)^{\beta}p(\theta)$
of $\theta$. The case of $\beta=1$
corresponds to the Bayesian estimation of the posterior. The
posterior can be sampled using an MCMC algorithm
Assuming weak priors and that the posterior is
asymptotically normal we have $\mathcal{F}=\mathrm{BIC} +O(1)$.
WBIC is an improved version of BIC, which is also
applicable for singular statistical models where BIC fails.
A model is singular if the Fisher information matrix is
not positive definite, which typically occurs when the
model contains hierarchical layers or has hidden variables.
\subsubsection{Predictive methods for Model comparison}
A statistical model $p(x|\theta)$ can be thought of as an approximation
of the true distribution $q(x)$ from which the observed data
$Y=\{y_1,y_2,...,y_n\}$ were generated.
$Y$ represents a set of independently
observed data points such that $p(Y|\theta)=\prod_{i=1}^{n}p(y_i|\theta)$.
The Bayesian predictive distribution can then be defined as
$p(x|Y)={\rm E}_{\theta}[p(x|\theta)]$, while the
maximum likelihood estimate is given by
$p(x|\hat{\theta}(Y))$.
Predictive methods judge models by their ability to
fit future data $X=\{x_1,x_2,...,x_n\}$, e.g., via
the log-likelihood function $-\ln p(X|Y)$.
Given that we do not have future data, the
idea is to measure out-of-sample-prediction error from the
sample at hand. Cross validation is a natural way to do
this, where we divide the current data set into training and
testing samples. But this is computationally costly.
Hence, alternate criteria have been developed.
We start by computing the training error
$T_e = -\frac{1}{n}\sum_{i=1}^{n} \ln p(y_i|Y)$.
However, this is a biased estimator of $\mathbb{E}_x[-\ln
p(x|Y)]$ as the data are used twice, once
to estimate the model and once more to compute the log likelihood
of the data. If we have more parameters in the model, it will
certainly fit the given data better but will also give rise
to larger variance in the estimator, and we need to penalize
the model for this. This variance,
which represents the effective degrees of freedom in the
model, can be calculated from the data and the model.
A list of some useful information criteria
based on the above idea are given below. They can be
easily computed using samples of $\theta$ obtained by an
MCMC simulation of the posterior $p(\theta|Y)$.
\begin{marginnote}
\entry{AIC} {Akaike information criterion}
\entry{DIC}{Deviance information criterion}
\entry{WAIC}{Widely applicable Bayesian information criterion}
\end{marginnote}
\begin{eqnarray*}
{\rm AIC}/2 =& -\ln p(Y|\hat{\theta})+d & \textrm{\citep{akaike1974new}}\\
{\rm DIC}_{1}/2 =& -\ln p(Y|{\rm E}_{\theta}^{1}[\theta])+2\left( \ln
p(Y| {\rm E}_{\theta}[\theta])-{\rm E}_{\theta}^{1}[\ln p(Y| \theta)]
\right) &\textrm{\citep{spiegelhalter2002bayesian}} \\
{\rm DIC}_{2}/2 =& -\ln p(Y|{\rm E}_{\theta}^{1}[\theta])+2 {\rm Var}_{\theta}^{1}[\ln
p(Y|\theta)] &\textrm{\citep{spiegelhalter2002bayesian}}\\
{\rm WAIC}_{1}/2 =& -\sum_i^{n} \ln {\rm
E}_{\theta}^{1}[p(y_i|\theta)] +2\sum_i^{n} \ln {\rm
E}_{\theta}^{1}[p(y_i| \theta)]-{\rm E}_{\theta}^{1}[\ln
p(y_i| \theta)] &\textrm{\citep{watanabe2010asymptotic}}\\
{\rm WAIC}_{2}/2 =& -\sum_i^{n} \ln {\rm
E}_{\theta}^{1}[p(y_i|\theta)] +\sum_i^{n} {\rm
Var}_{\theta}^{1}[\ln p(y_i|\theta)] &\textrm{\citep{watanabe2010asymptotic}}
\end{eqnarray*}
Here, ${\rm Var}_{\theta}^{1}$ denotes variance
taken over the posterior distribution
$p(Y|\theta)p(\theta)$ of $\theta$.
The first term is a measure of how well the model fits the
observed data while the second term is a penalty for
the degrees of freedom $d$ in the model.
In general, the predictive criteria have a well-defined
information-theoretic interpretation
\citep{burnham2002model, watanabe2010asymptotic}.
Specifically, the expected value of AIC and WAIC,
is equivalent to the expected Kullback-Leibler divergence
$\int q(x )\ln [(q(x)/
p(x|Y)] dx $ of the predictive distribution from the true
distribution, the expectation is taken over the random realizations
of the observed data set $Y$, which samples the true
distribution $q(x)$.
Also, in the asymptotic limit of large sample
size, both AIC and WAIC are equivalent to leave-one-out
cross-validation (LOOCV).
An extra parameter in a model need not necessarily
contribute to extra variance in the predictive density,
e.g., if we have informative priors on the parameter,
the likelihood has a very weak dependence on the parameter or if the
model is hierarchical then multiple parameters
might be restricted. The use of AIC can be problematic in
such cases. DIC and WAIC overcome this problem by estimating
the effective degrees of freedom
directly from the likelihood function of the data and samples
of $\theta$ obtained from the posterior $p(\theta|Y)$.
WAIC offers some additional advantages as compared to
AIC and DIC.
AIC and DIC use a point estimate for $\theta$ when
computing predictive density, whereas WAIC uses the Bayesian
predictive density.
If a model is singular, criteria such as AIC, DIC and BIC do not work well.
In contrast,
WAIC works for such cases, and in the asymptotic limit of large sample
size, WAIC is always equivalent to Bayesian LOOCV.
It is instructive to study the differences between BIC and
AIC, as they represent two very different approaches to the
problem of model selection
\citep[for a detailed discussion, see][]{burnham2002model}.
Due to the presence of the $\ln n$ term, for $n>7$ the
BIC penalizes free
parameters more heavily as compared to AIC.
So BIC is more parsimonious or cautious when it comes
to admitting new parameters in a model.
In situations where two models
can give rise to the same predictive distribution, BIC
will favor the model with fewer degrees of freedom while
AIC will treat them equally.
An example is nested models, where a simpler model can be
considered as a special case of a complex model but with
few of its parameters being fixed.
Interestingly, AIC can also
be argued to be using the approach of Bayes factors, but with a prior
whose variance decreases with sample size $n$, whereas
BIC would correspond to the choice of a weak prior with
fixed variance \citep{smith1980bayes}.
To conclude, the Bayesian and the predictive
methods both have their strengths and weaknesses.
If the choice of priors is well justified, then the
methods based on Bayes factor are best suited for model
selection. However, if our aim
is best predictive accuracy for future data, predictive
methods like WAIC are a better choice.
\section{Monte Carlo methods for Bayesian computations}\label{sec:mcmc}
Having discussed how to set up problems in the Bayesian
framework, we now discuss methods to perform the
inference, i.e., how to estimate the pdf
of parameters given the data. Except for some simple
cases, closed form analytical solutions are in general
not available. So one makes use of Monte Carlo based
methods to sample from the desired distribution.
The most popular method to do this today is the Markov Chain
Monte Carlo (MCMC) method. MCMC is a class of methods for
sampling a pdf using a Markov chain
whose equilibrium distribution is the desired distribution.
Once we have a sample
distributed according to some desired distribution, we can
compute expectation values and integrals of various
quantities in a process analogous to Monte Carlo
integration.
The word Monte Carlo in MCMC comes from the use of random numbers to
drive the Markov process and the close analogy to
Monte Carlo integration schemes. Note in conventional
Monte Carlo integration, the random samples are statistically
independent whereas in MCMC they are correlated.
We first broach the theory
behind Markov chains and then discuss specific MCMC
methods based on it.
\subsection{Markov Chain}\label{sec:markovchain}
A Markov chain is a sequence of random variables $X_1,
..., X_n$ such that, given the present state, the
future and past are independent. It is formally written as
\begin{eqnarray}} \def\ee{\end{eqnarray}
{\rm Prob}(X_{n+1}=x| X_1=x_1,X_2=x_2,...,X_n=x_n) & = &
{\rm Prob}(X_{n+1}=x|
X_n=x_n)
\ee
In other words, the conditional distribution of $X_{n+1}$ in future,
depends only upon the present state $X_n$.
If the probability of transition is independent of $n$, it is
a time-homogeneous chain. Such a chain is defined by
specifying the probabilities of transitioning
from one state to another.
To simplify mathematical notation, we sometimes
consider the state space to be continuous and sometimes
discrete.
But the presented results are equally valid for either type
of spaces.
For a continuous state space where a probability density can
be defined we can write the transition probability as
\begin{eqnarray}} \def\ee{\end{eqnarray}
K(x,y)={\rm Prob}(X_{n+1}=y|X_n=x)
\ee
For a discrete state space the transition probability
is a matrix and is written as $K_{xy}$.
On a given state space, a time-homogeneous Markov chain
has a stationary distribution (invariant measure) $\pi$ if
\begin{eqnarray}} \def\ee{\end{eqnarray}
\pi(y)=\int dx\ \pi(x) K(x,y)
\ee
A Markov chain is irreducible if it can go from any state
$x$ of a discrete state space to any
other state $y$ in a finite number of steps, i.e.,
there exists an integer $n$
such that $K^n_{xy}>0$.
If a chain having a stationary distribution is irreducible,
the stationary distribution is unique, and the chain is
positive recurrent.
For an aperiodic, positive recurrent chain with stationary
distribution $\pi$, the distribution is limiting
(equilibrium distribution).
This means if we start
with any initial distribution $\lambda$ (a row vector
specifying probability over states of a discrete state space)
and apply the transition
operator $K$ (a matrix) many times, the final distribution will
approach the stationary distribution $\pi$ (a row vector),
\begin{eqnarray}} \def\ee{\end{eqnarray}
\lim_{n \rightarrow \infty}\|\lambda K^n-\pi\|=0.
\ee
For an irreducible Markov chain with a unique
stationary distribution $\pi$, there is a law of large numbers
which says that the expectation value of a function $g(x)$
over $\pi$ approaches the average taken over the output of a Markov
chain,
\begin{eqnarray}} \def\ee{\end{eqnarray}
E_{\pi}[g(x)] = \int g(x) \pi(x) dx = \lim_{n \rightarrow \infty}\frac{1}{n}\sum_{i=1}^{n} g(x_i).
\ee
This property allows one to compute Monte Carlo estimates of specific
quantities of interest from a Markov chain. Techniques that
do this are known as Markov chain Monte Carlo or MCMC.
A chain having a stationary distribution is said to be
reversible if the chain starting from a stationary
distribution looks the same when run forward or
backward in time. In other words, if $X_n$ has distribution $\pi$
then the pair $(X_{n},X_{n+1})$ has the same joint
distribution as $(X_{n+1},X_{n})$.
\begin{eqnarray}} \def\ee{\end{eqnarray}
{\rm Prob}(X_n,X_{n+1})={\rm Prob}(X_{n+1},X_{n})
\ee
For the transition kernel $K$ this means
\begin{eqnarray}} \def\ee{\end{eqnarray}
\pi(X_n)K(X_{n},X_{n+1})=\pi(X_{n+1})K(X_{n+1},X_{n})
\ee
and is known as the {\it condition of detailed balance}.
For a Markov chain, it is not necessary to satisfy reversibility
in order to have a stationary distribution. However,
reversibility guarantees the existence of a stationary
distribution, and is thus a stronger condition. This
is the reason that most MCMC algorithms are designed to
satisfy detailed balance.
\subsection{Metropolis Hastings algorithm}\label{sec:mhalgo}
The most general MCMC algorithm is the Metropolis-Hastings (MH)
algorithm
\citep{metropolis1953equation,hastings1970monte}.
Suppose we are interested in sampling a
distribution $f(x)$ on a state space $E$, with $x \in E$.
To construct a transition kernel $K(x,y)$ to go from
$x$ to $y$, MH algorithm uses a two
step process:
\begin{itemize}
\item Specify a proposal distribution
$q(y|x)$.
\item Accept draws from
$q(y|x)$ with acceptance ratio $\alpha(x,y)=
{\rm min}\left[1,\frac{f(y)q(x|y)}{f(x)q(y|x)}\right]$.
\end{itemize}
So the transition kernel is given by $K(x,y)=q(y|x)\alpha(x,y)$.
The full algorithm is as follows
\begin{algorithm}
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{Starting point $x_1$, function $f(x)$,
transition kernel function $q(y|x)$}
\Output{An array of $N$ points ${x_1,x_2,...x_{N}}$}
\For{ $t=1$ \KwTo $N-1$}
{
Obtain a new sample $y$ from $q(y|x_{t})$ \;
Sample a uniform random variable U \;
\lIf{$U<\frac{f(y)q(x_{t}|y)}{f(x_{t})q(y|x_{t})}$}
{
$x_{t+1}=y$
}
\lElse
{
$x_{t+1}=x_t$
} \;
}
\caption{Metropolis Hastings Algorithm \label{MH1}}
\end{algorithm}
The transition kernel of the MH algorithm is reversible and satisfies
detailed balance, $f(x)K(x,y)=f(y)K(y,x)$.
Note the reversibility condition by
itself does not lead to a unique form for the acceptance ratio
$\alpha(x,y)$ and alternatives exist \citep{barker1965monte}.
However, it has been shown that the acceptance ratio of the
MH algorithm results in a chain with the fastest
mixing rate \citep{peskun1973optimum}.
There are multiple ways to construct the proposal
distribution $q$ each leading to a new version of the MH algorithm.
\begin{itemize}
\item {\bf Symmetric Metropolis:} $q(y|x)=q(x|y)$ which
simplifies the acceptance probability to ${\rm min}\left\{1,f(y)/f(x)\right\}$; this is the version that was
proposed by Metropolis and colleagues.
\item {\bf Random walk Metropolis-Hastings (RWMH):}
$q(y|x)=q(y-x)$; the direction and distance of the new
point from the current point is independent of the current
point. Common choices are $N(x,\sigma^2)$ and ${\rm Uniform}(x-\sigma,x+\sigma)$.
\item {\bf Independence sampler:} $q(y|x)=q(y)$; i.e., the new state is drawn independent of the current
state. The acceptance probability is given by ${\rm
min}\left\{1,\frac{f(y)q(x)}{f(x)q(y)}\right\}$, a generalization of the accept-reject algorithm. The quantity
$q(x)$ should resemble $f(x)$ but with longer
tails.
\item {\bf Langevin algorithm:} $q(y|x) \sim
N(x+\frac{\sigma^2}{2}\nabla \log f(x),\sigma^2)$; this is
useful when the gradient is available.
\end{itemize}
Except when $f(y)=f(x)$ (uniform target density), the mean
of the acceptance ratio $\alpha$ is always
less than unity. Decreasing $\sigma$ in the RWMH algorithm
increases $\alpha$ but
lowers the independence of the sampler. Increasing $\sigma$
improves the independence but lowers $\alpha$.
In the Langevin algorithm, one makes use of the information
in the gradient to allow faster mixing of the chain.
\subsection{Gibbs sampling}\label{sec:gibbs}
The Gibbs sampler introduced by \citet{geman1984stochastic}
is one of the most popular computational
methods for doing Bayesian computations.
Suppose we want to sample $f(x)$ where $x \in \chi \subseteq
\mathcal{R}^d$. In Gibbs sampling, the transition kernel $K(x,y)$ is split into
multiple steps. In each step, one coordinate is advanced
based on its conditional density with respect to other
coordinates. The algorithm is as follows:
\begin{algorithm}
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{Starting point $x^1$, Function $f(x)$}
\Output{An array of $N$ points ${x_1,...,x_{N}}$}
\For{ $t=1$ \KwTo $N-1$}
{
Sample $x^1_{t+1}$ from $f(x^1|x^2_{t},...x^d_{t})$ \;
Sample $x^2_{t+1}$ from $f(x^2|x^1_{t+1},x^3_{t}...x^d_{t})$ \;
Sample $x^d_{t+1}$ from $f(x^d|x^1_{t+1},...x^{d-1}_{t+1})$ \;
}
\caption{Gibbs Sampling Algorithm \label{GS1}}
\end{algorithm}
\noindent
The full transition kernel is written as,
\begin{equation}} \def\eeq{\end{equation}
\kappa_{1 \to d}(x_{t+1}|x_{t})=\prod_{i=1}^{d}
f(x^i_{t+1}|x^1_{t+1},...x^{i-1}_{t+1},x^{i+1}_{t},...,x^d_{t+1}).
\eeq
Similarly one can define a reverse move,
\begin{equation}} \def\eeq{\end{equation}
\kappa_{d \to 1}(x_t|x_{t+1})=\prod_{i=d}^{1}
f(x^i_{t}|x^1_{t+1},...x^{i-1}_{t+1},x^{i+1}_{t},...,x^d_{t+1}).
\eeq
It can be easily shown that
\begin{equation}} \def\eeq{\end{equation}
f(x_{t})\kappa_{1\to
d}(x_{t+1}|x_{t}) = f(x_{t+1})\kappa_{d\to
1}(x_{t}|x_{t+1}).
\eeq
Integrating both sides leads to
\begin{equation}} \def\eeq{\end{equation}
\int f(x_{t})\kappa_{1\to d}(x_{t+1}|x_{t}) dx= f(y).
\eeq
Thus, $f$ is the stationary distribution of the Markov
chain formed by the transition kernel $\kappa_{1\to
d}(x_{t+1}|x_{t})$. Note the Gibbs sampler as given above
({\it systematic} scan) is not reversible. However, the
reversible ones can easily
be produced, e.g., at each iteration picking a random
component to update (random-scan). The random-scan
Gibbs sampler can be viewed as a special case of
MH sampler with acceptance ratio ${\rm
min}(1,\frac{f(y)q(x|y)}{f(x)q(y|x)})=1$.
It follows that
\begin{eqnarray}} \def\ee{\end{eqnarray}
f(y)q(x|y) &= & f(y^i|y_{-i})f(y^{-i})f(x^i|y^{-i}) =
f(y^i|x^{-i})f(x^{-i})f(x^i|x^{-i}) \nonumber \\
& = & f(x^i|x^{-i})f(x^{-i})f(y^i|x^{-i}) = f(x)q(y|x).
\ee
Here,
$x^{-i}=\{x^1,...,x^{i-1},x^{i+1},...,x^{d}\}$ and
$y^{-i}=x^{-i}$, as only the $i$-th component is changed in
each step.
\subsection{Metropolis within Gibbs}\label{sec:mwg}
One problem with the Gibbs sampler is that it requires
one to sample from the conditional distributions
which can be difficult.
In such cases, one can replace the sampling of
conditional densities with the MH step. This then becomes
the Metropolis within Gibbs (MWG) scheme
\citep[see ][]{muller1991generic}, which is shown in Algorithm
\ref{MWG} (it is implemented in the code that we provide).
\begin{algorithm}
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{Starting point $x^1$, Function $f(x)$}
\Output{An array of $N$ points ${x^1,x^2,...x^{N}}$}
\For{ $t=1$ \KwTo $N-1$}
{
\For{ $i=1$ \KwTo $d$}
{
Generate $x^i_{*}$ from $q_i(x^i|x^1_{t+1},...,x^{i-1}_{t},x^{i+1}_{t}...x^d_{t})$ \;
Calculate
$r=\frac{f_i(x^i_{*}|x^1_{t+1},...,x^{i-1}_{t},x^{i+1}_{t}...x^d_{t})}{f_i(x^i_{t}|x^1_{t+1},...,x^{i-1}_{t},x^{i+1}_{t}...x^d_{t})}
\frac{q_i(x^i_{t}|x^1_{t+1},...,x^{i-1}_{t},x^{i+1}_{t}...x^d_{t})}{q_i(x^i_{*}|x^1_{t+1},...,x^{i-1}_{t},x^{i+1}_{t}...x^d_{t})}$\;
\lIf{$U<{\rm Min}(1,r)$}
{
$x^{i}_{t+1}=x^{i}_*$
} \lElse
{
$x^i_{t+1}=x_t$
}\;
}
}
\caption{Metropolis-Within-Gibbs Algorithm \label{MWG}}
\end{algorithm}
Rather than updating all the variables step by step,
one can also choose to update a subset of variables
together, leading to block updates.
The fact that the full sampling of a complicated
distribution can be broken up into a sequence
of smaller and easier samplings, is the main strength
of the Gibbs sampler and has resulted in its widespread use \citep[e.g.][]{2012MNRAS.427.2119S,2014ApJ...793...51S}.
\subsection{Adaptive Metropolis}\label{sec:adapmetro}
The efficiency of the MH algorithm depends crucially upon
the proposal distribution. By efficiency we typically mean
how independent are the samples. If the samples are not
independent then they have high correlation. For Markov chains,
the correlation falls off with distance between samples. If
the correlation is large, this means the mixing in the
chain is slow. If the width of the proposal
distribution is too small, the acceptance ratio is high
but the chain mixes very slowly. If the width of the proposal
distribution is too large, the acceptance ratio is too small
and the chain again mixes slowly (see \fig{adaptive1} for an
illustration of this effect).
\citet{gelman1996efficient} showed that optimal
covariance matrix $\Sigma$ for the RWMH algorithm using the multivariate
normal distribution is $(2.38^2/\mathcal{D})\Sigma_{\pi}$, where $\mathcal{D}$
is the dimensionality of the space and $\Sigma_{\pi}$ is the covariance
matrix of the target distribution $\pi$. The optimal acceptance
ratio $\alpha_{\rm opt}$ is 0.44 for dimension $\mathcal{D}=1$ and then
falls off with increasing number of dimensions reaching an
asymptotic value of $0.23$ for $\mathcal{D} \to \infty$.
The convergence
is quite fast
($\alpha=$[0.441, 0.352, 0.316, 0.279, 0.275, 0.266] for
$\mathcal{D}=$[1, 2, 3, 4, 5, 6]).
The efficiency as compared to independent samples is
$0.331/\mathcal{D}$.
\begin{figure}
\centering \includegraphics[width=0.75\textwidth]{adaptive1.pdf}
\caption{MCMC chains for different widths of the
proposal distribution. The variable $x$ is sampled from a
Gaussian distribution $N(0.0,1.0)$ using MCMC with
different proposal distributions. The proposal
distributions are also normal and are characterized by width
$\sigma$.
The ideal width is $\sigma=2.38$ and the chain for
this is shown in panel (b).
In panel (a), $\sigma$ is ten times larger while in panel
(c) it is one hundred times smaller. Panel (d) shows the
chain when an adaptive scheme is used to adjust $\sigma$.
The performance of this is same as for the ideal case shown
in panel (b).
\label{fig:adaptive1}}
\end{figure}
These results suggest a possible way to choose the optimal
proposal distribution.
Estimate the covariance matrix $\Sigma_{\pi}$ by a trial
run and then use it for the actual run. Even doing this
is cumbersome as it is unclear how long
the trial run should be. To circumvent this,
\citet{haario2001adaptive} proposed an adaptive scheme in
which $\Sigma$ is updated on the fly using past values.
Naively, any scheme that uses proposals that depend upon the
full past history violates the Markovian property, i.e.,
the future should only depend on the present and should be
independent of the past. The trick is to adapt the proposal
distribution in such a way that it converges to the optimal
one. The resulting chain then also converges to the target distribution.
\citet{andrieu2001controlled} showed that such a scheme
can be described as part of a more general adaptive
framework.
At the heart of most adaptive algorithms is the
\citet{robbins1951stochastic} recursion. They proposed an iterative
stochastic algorithm to find roots of functions that are
stochastic, i.e., their algorithm solves $M(x)=\alpha$, where instead of
$M(x)$ the function available is $N(x)$, which is
stochastic and is such that $\langle N(x) \rangle=M(x)$.
Starting with some initial value $x_0$
the algorithm to get the $n+1$th iterate is
\begin{eqnarray}} \def\ee{\end{eqnarray}
x_{n+1}=x_n+\gamma_n(\alpha-N(x_n))
\ee
Here $\gamma_1,\gamma_2,...$ is a sequence of positive
steps. The $x_n$ then converge to the true solution
provided the sequence $\gamma_n$ satisfies
\begin{eqnarray}} \def\ee{\end{eqnarray}
\sum_{n=0}^{\infty} \gamma_n=\infty \textrm{ and }
\sum_{n=0}^{\infty} \gamma_n^2 < \infty
\ee
The first condition makes sure that irrespective of where we
start, the solution can be reached in a finite number of
steps. The second condition makes sure that we do converge.
A possible choice of $\gamma_n$ is
$\gamma_n=\gamma/n^{\beta}$ where $0<\beta<1$.
A nice description of various adaptive
algorithms is given by \citet{Andrieu2008}.
Below we discuss Algorithm 4 from their paper
which is quite general and is implemented in the software that we provide.
\begin{algorithm}
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{Starting point $x_0$,$\mu_0$, $\Sigma_0$, $\alpha^{*}$, function $f(x)$}
\Output{An array of $N$ points ${x_0,x_2,...x_{N-1}}$}
\For{ $i=1$ \KwTo $N-1$}
{
Obtain a new sample $y$ from
$N(x_{i},\lambda_i\Sigma_{i})$ \;
Set $\alpha_{i}(x_i,y)=\frac{f(y)}{f(x_{i})}$ \;
Sample a uniform random variable U \;
\eIf{$U<\alpha_{i}(x_i,y)$}
{$x_{i+1}=y$ \; }
{$x_{i+1}=x_{i}$ \;}
$\log \lambda_{i+1}=\log
\lambda_{i}+\gamma_{i+1}(\alpha_{i}(x,y)-\alpha^*)$\;
$\mu_{i+1}=\mu_{i}+\gamma_{i+1}(x_{i+1}-\mu_{i})$ \;
$\Sigma_{i+1}=\Sigma_{i}+\gamma_{i+1}[(x_{i+1}-\mu_{i})(x_{i+1}-\mu_{i})^{\rm
T}-\Sigma_{i}]$ \;
}
\caption{Adaptive Symmetric Random Walk Metropolis Hastings Algorithm \label{AMH1}}
\end{algorithm}
If $\beta$ is too small, the convergence is too slow; if
$\beta$ is too large the convergence is too fast and
the simulation can quickly lean towards a wrong solution
and will take a long time to get out of it.
For adaptive MCMC, we find a choice of $\beta=0.6$
to be satisfactory for most test cases. \fig{adaptive1}d
shows an adaptive MCMC chain obtained using Algorithm
\ref{AMH1}.
The adaptive chain looks very similar to the ideal case
shown in \fig{adaptive1}b, and this demonstrates the
usefulness of the adaptive MCMC scheme.
\subsection{Affine invariant sampling}\label{sec:affine}
An elegant solution to the problem of tuning the proposal
density is to use the idea of ensemble samplers
\citep{gilks1994adaptive}. Here multiple chains (walkers)
are run in parallel but allowed to interact in such a way
that they can adapt their proposal densities.
\citet{Goodman2010} provide a general purpose algorithm
to do this, known as the affine invariant sampler \citep[see
also][]{christen2010general}. A python implementation
of this
(emcee: the MCMC hammer, \verbatim{http://dan.iel.fm/emcee/current/})
is provided by \citet{2013PASP..125..306F}
and is widely used in astronomy. We now
describe this algorithm.
We saw in the previous section that adapting the proposal density can
violate the Markovian property of a chain. The trick lies
in using the information available in the ensemble but
in a way that does not violate the Markovian property.
This is achieved by using the idea of partial resampling which is a
generalized version of the Gibbs sampling procedure.
Let us consider an ensemble of walkers
$X=(x_1,x_2,...,x_L)$ and a Markov chain that walks
on a product space with distribution
$\Pi(X)=\pi(x_1)\pi(x_2)...\pi(x_L)$.
Then if $x_i$ is updated conditional on other walkers
$x_{[-i]}=\{x_1,...,x_{i-1},x_{i+1},...,x_L\}$
(complementary set of walkers), but satisfying
detailed balance $p(y_i|x_i,x_{[-i]})=p(x_i|y_i,x_{[-i]})$,
then each walker samples from $\pi(x)$.
One way to do this is to choose a point $x_j$ from $x_{[-i]}$
and a scalar $r$ with density $g(r)$, and propose a new point $y$ as
\begin{eqnarray}} \def\ee{\end{eqnarray}
y=x_i+(r-1)(x_i-x_j)=x_j+r(x_i-x_j).
\ee
The inverse transformation is given by $x_i=x_j+(y-x_j)/r$.
Now if we want the proposal to be symmetric then
$q(y_i|x_i,x_{[-i]})=q(x_i|y_i,x_{[-i]})$, and this implies
$g(1/r)=rg(r)$.
A good choice of such a function is
\begin{eqnarray}} \def\ee{\end{eqnarray}
g(r)=\frac{1}{\sqrt{r}} \textrm{ for } r \in
\left[\frac{1}{a},a\right] \textrm{ and } a>1.
\ee
To satisfy detailed balance, the acceptance probability is
given by
$\textrm{min}\left[1,r^{n-1}\frac{\pi(Y)}{\pi(X_i)}\right]$.
The factor $r^{n-1}$ is because the proposal is restricted
along a line and not the full hypersphere over the actual
space. This means an appropriate Jacobian has to
be calculated; for details see \citet{gilks1994adaptive} and
\citet{roberts1994convergence} (the proof is much easier
when using the reversible jump MCMC formalism of \citet{green1995reversible}).
Moves other than the stretch move can also be constructed, e.g. a proposal $y=x_i+W$, where $W$ has
a covariance computed from a subset of walkers in the
complementary sample. It is also possible to construct
algorithms which use a combination of both the stretch and
the walk move. Although the \citet{Goodman2010} affine invariant algorithm
elegantly solves the problem of choosing a
suitable proposal distributions, it has one drawback.
The computational
cost of warm-up scales linearly with the number of walkers.
Note, like most other MCMC algorithms, multimodal distributions
(distributions with many well separated peaks)
also pose a problem for this algorithm.
\subsection{Convergence Diagnostics}\label{sec:convergence}
Having studied MCMC methods in order to sample from distributions, we now
discuss how to detect convergence; i.e., how long should
we run an MCMC chain. Several convergence diagnostics
have been proposed in the literature.
\citet{cowles1996markov} provide a good review of 13
convergence diagnostics. Other reviews include
\citet{brooks1998general} and \citet{robert2013monte}.
Unfortunately, because there is no method to detect convergence,
we can only detect failure to converge.
So convergence diagnostics are necessary conditions but not
sufficient. Below we present two schemes
to monitor convergence. The first scheme makes use of the
correlation length of the chain to compute the effective
number of independent samples in a chain. The second
scheme makes use of multiple chains to see if they
are converging.
\subsubsection{Effective sample size}
Let us begin by estimating how many independent
samples we need to get reliable estimates of mean
and variance of a quantity.
For a posterior of some variable $x$ with standard deviation
$\sigma_x$, the Monte Carlo standard error goes
as $\sigma_x/\sqrt{N}$ for sample of size $N$. So to measure
the mean of a quantity with about 3\% error as compared to
the overall uncertainty $\sigma_x$ we need $N=1000$.
\citet{raftery1992many} showed that to measure $0.025$
quantile to within $\pm0.005$ with probability 0.95 requires
about 4000 independent samples.
However, the MCMC is not an independent sampler. As we
have seen, the points in an MCMC chain are
correlated. Autocorrelation provides a measure of
this. Autocorrelation $\rho_{xx}(t)$ for a sequence is the correlation
between two points separated by a fixed distance $t$; i.e.
\begin{eqnarray}} \def\ee{\end{eqnarray}
\rho_{xx}(t)=\frac{\mathbb{E}[(x_i-\bar{x})(x_{i+t}-\bar{x})]}{\mathbb{E}[(x_i-\bar{x})^2]}
\ee
An automatic windowing procedure is discussed by
\cite{sokal1997} for the computation of integrated
autocorrelation \citep[see also][]{goodman1989multigrid,Goodman2010}.
Typically the autocorrelation falls off exponentially
as $\sim \exp^{-t/\tau_{x}}$ and $\tau_{x}$ is known as the
correlation time (or correlation length).
The integrated autocorrelation
is defined as $\tau_{\rm int,x}=(1/2)
\sum_{t=-\infty}^{\infty} \rho_{xx}(t)$. The variance of
the mean of $x$ for a sample of size $N$ can be shown to be
\begin{eqnarray}} \def\ee{\end{eqnarray}
{\rm Var}(\bar{x})=(2 \tau_{\rm int,x}) \frac{\mathbb{E}[(x_i-\bar{x})^2]}{N}
\ee
So for correlated samples the variance is $2
\tau_{\rm int,x}$ times larger than the variance of
independent samples. Using $\tau_{\rm int,x}$, one can
measure the number of effective independent samples in a correlated
chain $-$ also known as the {\it effective sample size} (ESS) $-$
as $N/(2\tau_{\rm int,x})$ and then use it to decide if we
have enough samples (e.g., $1000<\:$ESS$\:<4000$).
\subsubsection{Variance between chains}
The most widely used criterion for studying convergence was first
presented
by \citet{gelman1992inference}.
Let us suppose we have $M$ chains each consisting of $2N$
iterations out of which we use only the last $N$
iterations. For any given scalar parameter of interest
$\theta$, let
\begin{eqnarray}} \def\ee{\end{eqnarray}
\bar{\theta}_{j}=\frac{1}{n}\sum_{i=1}^{n}\theta_{i,j}
\textrm{\ \ \ \ and\ \ \ \ }
\bar{\theta}=\frac{1}{m}\sum_{j=1}^{m}\bar{\theta_j}
\ee
The index $i$ runs over points in a chain, and the index $j$
runs over the chains.
Then the between chain variance and the mean within chain variance
can be written as
\begin{eqnarray}} \def\ee{\end{eqnarray}
B=\frac{1}{m-1}\sum_{j=1}^m(\bar{\theta}_j-\bar{\theta})^2
\textrm{\ \ \ \ and\ \ \ \ }
W=\frac{1}{m}\sum_{j=1}^m \frac{1}{n-1}\sum_{i=1}^{n}(\theta_{i,j}-\bar{\theta_j})^2.
\ee
The total variance $\hat{\sigma}^2$ for
the estimator $\bar{\theta}$ can be written as a weighted
average of $W$ and $B$, $\hat{\sigma}^2=W(n-1)/n+B$.
If we account for the sampling variability of the estimator
$\bar{\theta}$, then this yields a pooled variance of
\begin{eqnarray}} \def\ee{\end{eqnarray}
V=\hat{\sigma}^2+\frac{B}{m}=\frac{n-1}{n}W+\frac{m+1}{m}B
\ee
for the mixture of chains.
If the initial
distribution is over-dispersed, then $B>\sigma^2$ and $V$ always
overestimates the true variance $\sigma^2$.
For any finite $n$, $W$ is expected to be
less than $\sigma^2$, as individual sequences
in a chain would
not have had the time to explore the full target distribution.
So, initially we expect $V/W>1$.
However, in the limit $n\to \infty$,
the variance $B$ between chains, which is expected to fall off
as $1/n$, goes to 0 and $W$ will approach the true variance
$\sigma^2$, making $V/W$ approach 1.
Therefore the ratio $\hat{R}=\sqrt{V/W}$, also known as the
potential scale reduction factor, can be used to monitor the convergence.
\subsubsection{Thinning}
For making inferences from an MCMC chain,
some algorithms use only the $k$-th iteration of each sequence such that successive draws are approximately independent, a process known as {\it thinning}.
However, there is no additional advantage of thinning
other than savings in storage. Since we are throwing away information,
an estimate from a thinned chain can never
be better than the original chain
\citep{geyer1992practical,maceachern1994subsampling}.
Moreover, it is difficult to choose an appropriate $k$
without studying the autocorrelation of the full chain.
So thinning is useful only in situations where the
autocorrelation is known a priori and is known to be
large. Here again $k$ should be chosen such that it is
smaller than the autocorrelation length, to retain
as much information as possible.
\subsection{Parallel Tempering}\label{sec:partemp}
Multimodal distributions in general pose problems
for all MCMC algorithms. Parallel tempering
is one way to address this problem.
It is a type of ensemble sampler where
multiple chains are simulated in parallel but are allowed to
exchange information. Each chain has a target distribution
different from the other and is controlled by a parameter
$T$ known as the temperature. Let $\pi(x)=\exp(-H(x))$ be the
actual target distribution, then a ladder of distributions
\begin{eqnarray}} \def\ee{\end{eqnarray}
\pi_i(x)=\exp(-H(x)/T_i), i=1,....n
\ee
is created, controlled via the parameter $T_i$, such that
$T_1>T_2>...>T_{n}$. $T_n$ is set to 1. Hence,
$\pi_n$ represents the target distribution.
The temperature broadens the target distribution
and allows a wider exploration of the parameter space
which makes it useful to explore multimodal distributions.
To exchange information between the chains, a state swapping
procedure is used. A swap is proposed between a randomly
chosen chain $i$
and its neighbors $i-1$ and $i+1$ with probability
$q_{i,i-1}=q_{i,i+1}=0.5$ and $q_{1,2}=q(n,n-1)=1$.
Naively, accepting the swap will violate the detailed
balance condition. So the swap proposal is accepted
with probability
\begin{eqnarray}} \def\ee{\end{eqnarray}
p_{ij}= {\rm
min}\left(1,\frac{\pi_i(x_j)\pi_j(x_i)}{\pi_i(x_i)\pi_j(x_j)}\right)
={\rm
min}\left(1,\exp\left(\left[H(x_i)-H(x_j)\right]\left[\frac{1}{T_i}-\frac{1}{T_j}\right]\right)\right)
\ee
which satisfies detailed balance.
In parallel tempering
the temperature ladder needs to be chosen carefully.
If the neighboring temperatures are too far apart, the
acceptance rate will be diminished leading to slow mixing.
If the neighboring temperatures are too close, a large number of elements
in the ladder will be required to explore a wide range in
parameter space, and this can increase the
computational cost significantly. However, by exploiting the trial runs, a suitable
ladder can be constructed \citep{liang2011advanced}.
The idea of parallel tempering can be generalized to
construct evolutionary algorithms that incorporate
features of genetic algorithms into the framework of
MCMC. The basic idea is to have parallel chains as in
parallel tempering and allow exchange of information
while satisfying detailed balance on the product space
defined by the chains. The exchange of information is based
on ideas of mutation and crossover from genetic algorithms
\citep{liang2001evolutionary,liang2001real}.
\subsection{Monte Carlo Metropolis Hastings}\label{sec:mcmh}
\label{sec:MCMH}
\begin{algorithm}
\SetKwInOut{Input}{Input}
\SetKwInOut{Output}{Output}
\Input{$\tilde{p}(\theta|x,Y),\ q(\theta'|\theta),\ T$ and $x$}
\Output{A set of points $({\theta_1,\theta_2,...\theta_{N}})$
sampled approximately from $p(\theta|x)$}
\For{ $t=1$ \KwTo $T-1$}
{
Generate $\theta'$ from $q(.|\theta)$\;
Generate $m$ auxiliary samples $Y=(y_1,..y_m)$
conditioned on $\theta$\;
Obtain a Monte Carlo estimate
$\tilde{r}(x,\theta,\theta',Y)$ of the MH ratio
$\tilde{r}=\tilde{p}(\theta'|x,Y)/\tilde{p}(\theta|x,Y)$ \;
\lIf{$U<{\rm Min}(1,\tilde{r})$}
{
$\theta_{t+1}=\theta'$
} \lElse
{
$\theta_{t+1}=\theta$
}\;
}
\caption{Monte Carlo Metropolis Hastings\label{MCMH}}
\label{alg:mcmh}
\end{algorithm}
In MCMC based Bayesian inference, we are concerned with
simulating samples from some pdf
$p(\theta|x)=p(x|\theta)p(\theta)$. However, there are
situations when $p(x|\theta)$ cannot be easily evaluated
or is not available in an analytically tractable form.
In such situations one can make
use of Monte Carlo based techniques to approximately evaluate
$p(\theta|x)$. More generally, the Metropolis
Hastings ratio $r=p(\theta'|x)/p(\theta|x)$ is used
to update an MCMC chain. In such techniques, typically, one generates
a set of auxiliary samples $Y=\{y_1,...,y_m\}$ conditioned on
$\theta$ and then uses them to compute
$\tilde{p}(\theta|x,Y)$ (an approximation of $p(\theta|x)$)
or $\tilde{r}$ (an approximation of the ratio $r$).
However, Monte Carlo based
estimates are stochastic and special care is needed when
working with them in an MCMC scheme.
An algorithm to
make use of Monte Carlo based estimates inside a Metropolis
Hastings algorithm is given in \alg{mcmh}
(it is implemented in the software that we provide).
There are many variants of \alg{mcmh}, depending
upon how and at what stage the auxiliary sample is generated $-$ see
Chapter 4 in \citet{liang2011advanced}.
The invariant
stationary distribution of such Markov chains is not necessarily the
target density $p(\theta|x)$. The
characteristics of such chains and their convergence
properties are discussed by \citet{beaumont2003estimation}
and \citet{andrieu2009pseudo}.
In \alg{mcmh}, the auxiliary sample is refreshed in each iteration and
the same sample $Y$ is used to estimate both $\tilde{p}(\theta'|x,Y)$
and $\tilde{p}(\theta|x,Y)$. This makes \alg{mcmh}
more robust compared to other similar alternatives.
In classical MCMC, one can reuse the previous estimate of
$p(\theta|x)$ when computing $r$. However, when the
Metropolis Hastings ratio $\tilde{r}$ is
stochastic, if $\tilde{p}(\theta|x,Y)$ is not evaluated in each
iteration using a fresh sample of $Y$, then the
MCMC chain tends to get stuck at a stochastic
maxima of the estimated likelihood
\citep{2014ApJ...793...51S}.
The smaller the size of the auxiliary sample, or the more inaccurate the
Monte Carlo estimate of $\tilde{r}$, the worse is this
problem. Using the same sample $Y$ to estimate
both $\tilde{p}(\theta'|x)$ and $\tilde{p}(\theta'|x)$
leads to lower noise in the estimated ratio of $\tilde{r}$.
This property was also noticed by \citet{2013MNRAS.433.1411M}
in the context of fitting models of the gravitational
potential of the Milky Way to spatio-kinematic data
of stars orbiting inside it. Two specific cases where the
above algorithm can be used are given below.
\subsubsection{Unknown normalization constant}
In fitting a model to data,
we are interested in sampling $p(\theta|x)=p(x|\theta)p(\theta)$.
To do this, the function $p(x|\theta)$ should be
properly normalized over the data space, in the sense that
$\int p(x|\theta)dx=1$. However, on many occasions, we have
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(x|\theta) =\frac{1}{Z(\theta)}\exp(-U(x,\theta))=\frac{1}{Z(\theta)}f(x|\theta),
\ee
where $f(x|\theta)$ is known but the normalization
constant $Z(\theta)$ is not known. An example is
the problem of fitting a density profile
$\rho(r|\theta)$ ($r$ being the Galactocentric distance)
to a sample of stars with Galactic latitude
$b>30^{\circ}$, longitude $l>30^{\circ}$ and heliocentric
distance $s<50$ kpc. Here we have
$Z(\theta)=\int_{b=\pi/6}^{\pi/2} db \int_{0}^{50} ds \int_{\pi/6}^{2\pi}
dl \rho(l,b,s|\theta) s^2 cos(b)$.
Our aim is to compute the Metropolis Hastings ratio
$r=p(\theta'|x)/p(\theta|x)=[Z(\theta)/Z(\theta')][f(x|\theta')/f(x|\theta)]$
that is used to advance an MCMC chain, and it is the ratio
$R=Z(\theta)/Z(\theta')$ that is unknown.
If one can sample exactly from $f(x|\theta)$,
then it is possible to cancel the normalization constant
using ingenious algorithms by \citet{moller2006efficient} and
\citet{murray2012mcmc}. However exact sampling is not
always feasible. In such cases
a Monte Carlo estimate of the ratio of the unknown normalization
constant $R=Z(\theta)/Z(\theta')$ can be done using samples
$Y=(y_1,..., y_m)$\ generated from density $f(y|\theta)$, such that
\begin{eqnarray}} \def\ee{\end{eqnarray}
\tilde{R}(\theta,\theta';Y)=\frac{1}{m}\sum_{i=1}^{m} \frac{f(y_i|\theta')}{f(y_i|\theta)}.
\ee
This sampling can be done by various means, e.g., exact
sampling, MCMC, and rejection sampling. If $f(y|\theta')$ is
difficult to sample
from, one can use so-called
``importance sampling'' by drawing samples from
a distribution $g(y|\theta)$ that is easy to sample
from. The required ratio of normalization constants is then given by
\begin{eqnarray}} \def\ee{\end{eqnarray}
\tilde{R}(\theta,\theta';Y)= \frac{\frac{1}{m}\sum_{i=1}^{m}
f(y_i|\theta')/g(y|\theta)}{\frac{1}{m}\sum_{i=1}^{m} f(y_i|\theta)/g(y|\theta)},
\ee
and the MH ratio is given by $\tilde{r}=\tilde{R}(\theta,\theta';Y)[f(x|\theta')/f(x|\theta)]$.
\subsubsection{Marginal inference}
Here we are interested in the marginal density
$p(\theta|x)=\int p(\theta,y|x) dy$, but the integral may not be analytically tractable and
may also be difficult to do by deterministic schemes. In such
situations, the integration can be done by Monte Carlo
importance sampling, using auxiliary samples $Y$ generated from some density
$g(y|\theta)$ that is easy to sample from. Thus we have
\begin{eqnarray}} \def\ee{\end{eqnarray}
\tilde{p}(\theta|x,Y)=\frac{1}{m}\sum_{i=1}^{m} \frac{p(\theta,y_i|x)}{g(y_i|\theta)}.
\ee
\subsection{Hamiltonian Monte Carlo}\label{sec:hmc}
One of the attractive features of MCMC
for sampling pdfs
is its better performance for higher
dimensions. However, for very large dimensions, traditional
MCMC algorithms start running into problems.
While for lower dimensions, a typical set of the posterior
(e.g. region encompassing 99\% of the total probability)
lies close to the center, for higher dimensions, a typical
set lies in a shell that has a very large volume. Since, a
shell cannot be traversed with large step sizes, it takes a long time to explore the posterior.
Hamiltonian Monte Carlo (HMC) tries to address this problem by
introducing an auxiliary variable called momentum $u$ for
each real variable $x$ called position \citep{1987PhLB..195..216D,neal1993probabilistic}. The log of posterior
(target density) $\pi(x)$ is assumed to define the potential
energy $U(x)=-\ln \pi(x)$,
and the momenta define the kinetic energy $K(u)$. Together
they define the Hamiltonian $H(x,u)=U(x)+K(u)$, where
$K(u)=u^2/2$. The distribution to be explored is
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(x,u)=\exp\left[-H(x,u)\right]=\exp\left[-\ln\pi(x)-\frac{1}{2}u^2\right]
\ee
Next, principles of Hamiltonian dynamics are used to advance
a given point to a new location. The point is then
accepted or rejected based on the MH algorithm. The use
of Hamiltonian dynamics to advance a given point allows
the point to travel to locations which are
far from its current location. This allows faster
exploration of the parameter space.
There are two major obstacles involved with using HMC, and this has
prevented its widespread use.
First, it requires the gradient of the target density.
Secondly, it requires two extra parameters to be tuned by the user:
a step size $\epsilon$ to advance from the current state
and the number of steps over which
to evolve the Hamiltonian system. Considerable progress has
been made to address both these issues.
The automatic/algorithmic differentiation can be used
to accurately compute the derivatives of a given function
without any user
intervention \citep{griewank2008evaluating}.
The idea is that
any function written as a computer program can be described as
a sequence of elementary arithmetic operations, and then
by applying the chain rule of derivatives repeatedly on
these operations, the derivatives can be computed.
Alternatively, one can create analytical functions to
approximate the target density and use these to compute the
derivatives.
This is because the exact Hamiltonian is only required
when computing the acceptance probability
and this does not require derivatives.
For simulating the trajectory, one needs derivatives
and here one can use an approximate Hamiltonian \citep{neal2011mcmc}.
An application of HMC for fitting cosmological
parameters is given by
\citet{hajian2007efficient} and
\citet{2008MNRAS.389.1284T}.
\citet{homan2014no} provide additional algorithms for
automatic tuning of step size $\epsilon$ and the number
of steps $L$, known as the No-U-Turn Sampler.
This is used in the open-source Bayesian inference
package Stan (available at \verbatim{http://www.mc-stan.org}).
\subsection{Population Monte Carlo}
Population Monte Carlo is an iterative importance sampling
technique that adapts itself at each iteration and
produces a sample approximately simulated from
the target distribution. The sample along with its
importance weights
can be used to construct unbiased estimates of quantities
integrated over the target distribution.
Suppose $h(x)$ is a quantity of interest.
One of the major applications for
MCMC applications is to compute integrals like
$J = \int h(x) \pi(x) dx$.
In importance sampling, this is replaced by
\begin{eqnarray}} \def\ee{\end{eqnarray}
J = \frac{1}{N}\sum_{i=1}^{N} h(x_i) \frac{\pi(x_i)}{q(x_i)},
\ee
where $(x_1,...,x_n)$ are sampled from a distribution $q(x)$
which is easier to sample than $\pi(x)$. The closer the
importance function to the target distribution, the better
the quality of the estimate (lower variance).
In practise it is difficult to guess a good importance
function.
The main idea in population Monte Carlo is to
start with a reasonable guess of the importance function
$q_0$ and then iteratively improve $q_t$ by making use
of the past set of samples $(x_1^{t-1},...,x_N^{t-1})$.
The importance function can adapt not only in time (with
each iteration), but also in space, and can be
written in general as $q_{t}(.|x_{i}^{t-1})$. Suppose
$X^t=\{x_1^t,...,x_N^t\}$ are the set of points at iteration
$t$. Let $x_i^t$ be produced from importance distribution
$q_t(x|x_i^{t-1})$.
An estimate of $J$ is then given by
\begin{eqnarray}} \def\ee{\end{eqnarray}
J^{t}=\sum_{i=1}^{n} w_i^t h(x_i^t) \textrm{ where } \rho_i^t=\frac{\pi(x_i^t)}{q_{it}(x_i^t)} \textrm{ and }
w_i^t=\frac{\rho_i^t}{\sum_{i=1}^{n} \rho_i^t}.
\ee
Thus the expectation
value of any function $h(x)$ computed using importance
sampling is unbiased, i.e.
\begin{eqnarray}} \def\ee{\end{eqnarray}
\mathbb{E}\left[h(X^t)\frac{\pi(X^t)}{q_t(X^t|X^{t-1})}\right]
& = & \int h(x) \frac{\pi(x)}{q_t(x|y)}q_t(x|y) g(y) {\rm
d}x {\rm d}y = \int h(x) \pi(x) dx
\ee
Here $g$ is distribution of $X^{t-1}$ and the equality is
valid for any $g$.
A simple choice for the importance function is to have
set it as a mixture
of normal or $t$-distributions, e.g., $q^t(x)=\sum_{d=1}^{D}
\alpha_d^t \mathcal{N}(x|\mu_d^t,\Sigma_d^t)$
\citep{cappe2008adaptive}. This has been used
for cosmological parameter estimation
\citep{2009PhRvD..80b3507W} and model comparison
\citep{2010MNRAS.405.2381K}.
\subsection{Nested Sampling}\label{sec:nested}
In \sec{modelcomp}, we saw that computing the evidence is
computationally challenging. Nested sampling
\citep{skilling2006nested} is designed to
ease this computation.
To compute the evidence,
we are interested in computing quantities like
\begin{eqnarray}} \def\ee{\end{eqnarray}
Z = \int L(\theta)\pi(\theta)\: {\rm d}\theta =\int L(\theta)\: {\rm d} \pi (\theta).
\ee
Integration is basically chopping up the full space into
small volume elements and summing the contribution of the
integrand over these cells. We are free to chop up the volume
and order or label the cells as we wish. So we divide the
space by iso-likelihood contours and define a variable
$X$ to label them. A convenient choice is the
prior probability mass enclosed by an iso-likelihood
contour, i.e.
\begin{eqnarray}} \def\ee{\end{eqnarray}
X(L)=\int_{L(\theta)>L} \pi(\theta) d\theta
\ee
If the the prior probability is normalized, then it
ranges from 0 for the highest likelihood, to 1 for the lowest
likelihood. Given the above definition, we can also define
an inverse function $L(X)$, which is the likelihood that
encloses a probability mass of $X$. So the integral for $Z$
can now be written as $Z=\int L(X)dX$.
Suppose we generate $N$ samples uniformly from the prior
distribution. Next, we sort them in decreasing sequence of $L$
to give prior mass $X_i=i/N$. Then using trapezoidal rule,
one can easily perform the numerical integration.
However, a significant contribution to the integral comes from a region
with small prior mass $X$. So, the integral should be done
in equal steps in $\ln(X)$ rather than $X$. This can
be done using an iterative procedure. We start
with a set $A$ of $N$ points drawn from the prior.
At each iteration, let $L_i$ be the point with lowest $L$;
we replace it in set $A$ with a new point drawn uniformly from the
prior but satisfying $L>L_i$. This generates a sequence of
$L_{i}$ for which the expected $X_i=\exp(-i/N)$.
Nested sampling is widely used for cosmological
model selection and parameter estimation. Three publicly
available packages based on nested sampling are CosmoNest
\citep[][see \verbatim{https://github.com/dparkins/CosmoNest}]{2006PhRvD..73l3523P,2006ApJ...638L..51M},
MultiNest \citep[][see \verbatim{https://ccpforge.cse.rl.ac.uk/gf/project/multinest}]{2009MNRAS.398.1601F} and
DNEST \citep[][see \verbatim{https://github.com/eggplantbren/DNest4}]{brewer2011diffusive}.
\section{Bayesian hierarchical modelling (BHM)}\label{sec:bhm}
In the simplest setting, we have some observed data $Y$ generated
by some model having parameters $\theta$ which can be
inferred using the Bayes theorem as
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta|Y) \propto p(Y|\theta)p(\theta),
\ee
where $p(\theta)$ denotes our prior knowledge or belief
about $\theta$.
If the model parameters $\theta$ depend upon another
set of parameters $y$ through $p(\theta|\phi)p(\phi)$, then
$\theta$ and $\phi$ can be inferred using
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta,\phi|Y) \propto p(Y|\theta)p(\theta|\phi)p(\phi).
\ee
The variable $\phi$ is known as the hyperparameter and $p(\phi)$,
the distribution of the hyperparameter, as a hyperprior.
Alternatively, the observed data $Y$ may depend upon another set of
hidden variables $X$, which in turn depend on $\theta$. The
inference of $\theta$ and $X$ can then be established
using
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta,X|Y) \propto p(Y|X)p(X|\theta)p(\theta).
\ee
Such situations lead to hierarchies and Bayesian models of
this type are known as hierarchical models.
It turns out that hierarchies are quite common in real world
applications, often where more than two levels
exist, and Bayesian hierarchical modelling provides
a framework for capturing this.
\begin{figure}
\centering \includegraphics[width=0.95\textwidth]{hbayes_demo1.pdf}
\caption{Analysis of group mean using a hierarchical
Bayesian model. The dashed line is the global mean from
all data points. The blue dots are group means
computed from the data points in the group. The green
points are estimates of group mean using a hierarchical
Bayesian model, which makes use of the full
information available. The green points have smaller error bars and are systematically
closer to the global mean than the blue points.
\label{fig:hbayes_demo1}}
\end{figure}
Let us consider a simple example, for details see
\citet{gelman2013bayesian}. Suppose
we observe some data $Y$ (a set of measurements of some
variable $y$) with uncertainty
$\sigma$, and we are interested in the mean
$\alpha=\bar{y}$. Now suppose that the data
$Y=\{y_{ij}|0<j<J, 0<i<n_j\}$
are grouped into $J$ independent groups, and we have reason to believe
that the group mean $\alpha_j$ varies from group to group.
For observations within a group $j$, our model is
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(y_{.j}|\alpha_j,\sigma)=\mathcal{N}(y_{.j}|\alpha_j,\sigma^2),
\ee
where we denote by $y_{.j}$ an observation belonging to
group $j$. We now compute the group mean $\bar{y}_{.j}$, instead of
global mean $\bar{y}$, to capture
the variation of mean across groups.
A global mean is certainly an inaccurate description
of data, whenever the group mean is far away from the
global mean. However, if the number of data points in a group
is very small, e.g., $n_j=2$, then the uncertainty in the group
mean is large and it is much better to
trust the global mean than the group mean.
Bayesian hierarchical modelling provides a natural way
to handle the above problem of group means.
It can act like a middle ground
between the two extremes, global mean versus group mean.
To demonstrate this,
we set up the above problem using a Bayesian
hierarchical model.
Suppose the group means are distributed according
to a normal distribution
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\alpha_j|\mu,\omega)=\mathcal{N}(\alpha_j|\mu,\omega^2)
\ee
where $\mu$ and $\omega$ are unknown parameters of the
model. The $\mu$, $\omega$ and the group means
$\alpha=\{\alpha_1,...,\alpha_J\}$ can then be inferred
from data $Y$ using
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\alpha,\mu,\omega|Y) & \propto &
p(Y|\alpha,\sigma)p(\alpha|\mu,\omega)p(\mu,\omega)
\propto p(\mu,\omega) \prod_{j=1}^J p(\alpha_j|\mu,\omega)\prod_{i=1}^{n_j} p(y_{ij}|\alpha_j,\sigma)
\ee
We generated synthetic data with
$\mu=0$, $\omega=1$, $J=40$, $\sigma=1$ and $2<n_j<10$;
we then estimated $\alpha$, $\mu$ and $\omega$ (assuming
flat priors for $\mu$ and $\omega$).
The results are shown in \fig{hbayes_demo1}.
The BHM based group mean estimates are systematically
shifted with respect to standard group mean estimates (
computed from the data points in a group). The BHM estimates
are closer to the global mean than the standard
estimates. The shift between the two estimates is more for
cases where the error bars are large. The BHM estimates also
have smaller error bars. This is because , when estimating
the group mean, in addition to points within a group the BHM
model also makes use of information available from other
groups.
\subsection{Expectation maximization, data
augmentation and Gibbs sampling} \label{sec:gsem}
The easiest way to analyze a Bayesian hierarchical
model is via Gibbs sampling,
and the motivation for doing this was provided by the
the expectation maximization (EM) algorithm.
In fact, the EM algorithm
led to the development of the DA
algorithm, which in turn provided the idea to use
Gibbs sampling to solve Bayesian hierarchical
models.
Hence we begin by exploring the
EM algorithm \citep{dempster1977maximum} which
is one of the most influential algorithm in the field of
statistics. Let us suppose that we have some observed data
$x=\{x_1,...,x_N\}$ generated by some model $p(x|\theta)$
having parameters $\theta$.
We want to
compute the most likely parameters of the model given the
data, i.e., $\hat{\theta} = {\rm
argmax}_{\theta}[p(x|\theta)]$. The full model is specified
by $p(x,z|\theta)$ with $p(x,z|\theta)=\prod_{i=1}^{N}
p(x_i,z_i|\theta)$, where $z$ are variables which are
either missing or hidden or unobserved.
The EM algorithm solves this problem as follows.
The algorithm has two steps. It starts with a
fiducial value of $\theta_0$, then does the following
at every iteration $t$.
\begin{itemize}
\item E-step: Compute $Q(\theta|\theta_t,x)= \int dz\ p(z|\theta_t,x) \log
p(x,z|\theta)$. In other words, it computes the
expectation of the log likelihood $ \log p(x,z|\theta)$
with respect to $p(z|\theta_t,x)$.
\item M-step: Find the value of $\theta$ that maximizes
$Q(\theta|\theta_t,x)$ and set $\theta_{t+1}={\rm argmax}_{\theta}[Q(\theta|\theta_t,x)]$.
\end{itemize}
These steps are repeated iteratively until $\theta_{t+1} \sim \theta_{t}$.
The proof that the EM algorithm increases the likelihood
$p(x|\theta)$ at each stage is as follows.
The conditional density of the missing data $z$ given the
observed data $x$ and the model parameter $\theta$ is given
by
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(z|\theta,x)=\frac{p(x,z|\theta)}{p(x|\theta)}.
\ee
Taking the $\log$ and then the expectation with respect to
$p(z|\theta_t,x)$, we get
\begin{eqnarray}} \def\ee{\end{eqnarray}
\log p(x|\theta) & = & \int dz\ p(z|\theta_t,x) \log
p(x,z|\theta) -\int dz\ p(z|\theta_t,x) \log
p(z|\theta,x), \\
& = & Q(\theta|\theta_{t},x)+S(\theta|\theta_{t},x),
\ee
which is valid for any $\theta$. Using this result, we can compute
the difference
\begin{eqnarray}} \def\ee{\end{eqnarray}
\log p(x|\theta_{t+1}) -\log p(x|\theta_{t}) = Q(\theta_{t+1}|\theta_{t})-Q(\theta_{t}|\theta_{t})+S(\theta_{t+1}|\theta_{t})-S(\theta_{t}|\theta_{t})
\ee
Due to the M-step,
$Q(\theta_{t+1}|\theta_{t},x)-Q(\theta_{t}|\theta_{t},x) \geq
0$. Also, from Gibbs' inequality,
$S(\theta_{t+1}|\theta_{t})-S(\theta_{t}|\theta_{t})\geq
0$.
This means that each EM iteration is guaranteed to increase the marginal
likelihood $p(x|\theta)$.
This guarantees a convergence towards a maximum,
but not necessarily a global maximum. The algorithm can still get stuck at a
saddle point, or a local maximum.
The EM algorithm as presented above is deterministic.
In general, it is not always easy to compute the expectation
value, as it involves integrals over high dimensions.
A general way to compute the $Q(\theta|\theta_t,x)$, would be to draw
$m$ random samples of $z$ from distribution $k(z|\theta_t,x)$
and take its mean. We label this stochastic estimate
$Q_S(\theta|\theta_t,x)$, which in the limit $m \to \infty$
is same as $Q(\theta|\theta_t,x)$. Having computed $Q_S$, the
M-step can proceed as usual to maximize it and
compute a new $\theta_{t+1}$. In fact $m$ can be set to 1.
This is the stochastic version of EM
(SEM) as given by \citet{celeux1985sem}.
Because of stochasticity, one does not get a unique answer but
instead a distribution. In fact, SEM
generates a Markov chain, which under mild regularity
conditions converges to a stationary distribution. The
algorithm has an additional advantage in that it is less
likely to get stuck at a local maximum.
If we now replace the M-step with a draw of $\theta$
from the $Q_S(\theta|\theta_t,x)$,
this becomes a fully stochastic method
; this is, as previously mentioned
the DA algorithm of \citet{tanner1987calculation}.
This is equivalent to a
two-step Gibbs Sampler for sampling from
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta,Z|X) \propto p(X,Z|\theta)p(\theta)
\ee
\begin{enumerate}
\item Sample $Z_{t+1}$ from $p(Z|\theta_{t},X)$.
\item Sample $\theta_{t+1}$ from $p(\theta|Z_{t+1},X)$.
\end{enumerate}
From the properties of the
Gibbs sampler, we know that the sequence of
$(\theta_{t},Z_{t})$ forms a Markov chain that samples
$p(\theta,Z|X)$. Although Gibbs sampling requires sampling from the
conditional distribution, the inner step can be
replaced by MH sampling, leading to the
Metropolis-within-Gibbs method as discussed in \sec{gibbs}.
This provides a completely general scheme for handling
missing data.
Finally, the DA algorithm is not limited to just
missing variables of the data, but can also be applied to
unknown parameters of the model, e.g., $\alpha$ in
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta,\alpha|X) \propto p(X|\theta,\alpha)p(\theta)p(\alpha)
\mathrm{\ \ \ \ \ or\ \ \ \ \ }
p(\theta,\alpha|X) \propto p(X|\theta,\alpha)p(\theta|\alpha)p(\alpha).
\ee
Such dependencies are common in Bayesian hierarchical modeling.
In general, the Bayesian hierarchical modeling
provides a framework for handling
marginalization in Bayesian data analysis, i.e.,
handling parameters or variables that are either unknown
or missing but are necessary to model the data.
\subsection{Handling uncertainties in observed data}
Marginalization is not limited to handling missing data. It
can also be used to handle data $X=\{x_i| i=1,...,N\}$ with uncertainty
$\sigma_X=\{\sigma_{x,i}| i=1,...,N\}$. Consider
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta|X,\sigma_X) & \propto & p(\theta) \prod_i \int
p(x|\theta)p(x_i|x,\sigma_{x,i}) dx \mathrm{\ \ and\ }
\label{equ:pthetax1} \\
p(\theta,X^t|X,\sigma_X) & \propto & p(\theta)\prod_i p(x_{i}^t|\theta)p(x_{i}|x_i^t,\sigma_{x,i}),
\label{equ:pthetax2}
\ee
where $X_t=\{x_{i}^t|i=1,...,N\}$ is the true values of
the observed data $X$. Here again, instead of doing an
integration, one treats the true values as unknowns and
sample them using the Gibbs scheme. We demonstrate this with a simple example
where $p(x_i^t|\theta) \sim \mathcal{N}(x_i^t|\mu,\sigma^2)$ is the model
that generates the data, and $\theta=(\mu,\sigma)$ are the
unknowns which we wish to evaluate. The data has uncertainty
described by another Gaussian function
$p(x_i|x_i^t,\sigma_{x,i}) \sim \mathcal{N}(x_i|x_i^t,\sigma_{x,i}^2)$.
\begin{figure}
\centering \includegraphics[width=0.75\textwidth]{da_demo1.pdf}
\caption{Comparison of two methods to handle nuisance
parameters. Here the nuisance parameter is the
true coordinate which is related to the observed
coordinate via a given uncertainty. In the DA algorithm, the
nuisance parameter is sampled alongside other parameters
using Gibbs sampling on a Bayesian hierarchical model.
In the other method, the nuisance
parameter is marginalized via integration; an analytical form
of the marginalized likelihood is used. The estimated
parameters are $\mu$ and $\sigma$.
The panels (a) and (b) show the probability distribution
function of the parameters given the data.
The panels (c) and (d)
show the autocorrelation function of the two parameters in
their respective Markov chains.
\label{fig:da_demo}}
\end{figure}
For this simple case, the integral in \equ{pthetax1} leads
to an analytical expression
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta|X,\sigma_X) \propto p(\theta) \prod_i \mathcal{N}(x_i|\mu,\sigma_{{\rm
tot},i}^2), \textrm{ where } \sigma_{{\rm tot},i}=\sqrt{\sigma^2+\sigma_{x,i}^2}.
\label{equ:pthetax3}
\ee
We used $(\mu,\sigma)=(0.0,1.0)$ and $\sigma_x=0.5$ to
generate test data and then estimated $\mu$ and $\sigma$
using two schemes: (1) DA algorithm which uses \equ{pthetax2}
and treats $X_t$ as unknown and samples from it, and (2)
explicit integration scheme which uses \equ{pthetax3}
where the variable $x_i^t$ has been integrated out of the
equation. The Markov chain was run for 100,000 iterations.
\fig{da_demo}$a,b$ shows the pdf of the
estimates of the two parameters. Both schemes give
identical results. The autocorrelation function
for the two parameters are shown in
\fig{da_demo}$c,d$. The DA algorithm has a slightly higher
autocorrelation time $\tau$ as it has to sample an
extra parameter for each data point.
\section{Case studies in astronomy}\label{sec:casestudy}
In this section, we study a range of cases in astronomy where
MCMC based Bayesian analysis is making a significant impact.
The emphasis is on showing how to set up
a diverse range of problems within the Bayesian framework
and how to solve them using MCMC techniques.
The examples are intentionally chosen from different areas
of astronomy so as to demonstrate the ubiquity of the techniques
reviewed here. There is a long history of applying
such techniques
in the field of cosmology, and excellent reviews and books
already exist here:
\citet{2008ConPh..49...71T,hobson2010bayesian,2013arXiv1302.1721P}.
\subsection{Exoplanets and binary systems using radial
velocity measurements}
\begin{figure}
\centering
\begin{minipage}{6cm}
\centering
\includegraphics[width=6cm]{rv_mcmc.pdf}
\end{minipage}
\centering
\begin{minipage}{6cm}
\centering
\includegraphics[width=6cm]{rv_mcmc_params.pdf}
\end{minipage}
\caption{Left: Radial velocity as a function of time for a star in
a binary system.
The parameters of the binary system are listed on the top.
The green line is the best fit solution
obtained using an MCMC simulation. The red and the green curves are generated from
\equ{radial_vel} and differ only in the eccentricity $e$.
The plot shows that the shape of the radial velocity curve
depends sensitively upon the eccentricity $e$ of the
orbit. Right: The posterior probability distribution of parameters
obtained using the MCMC simulation.}
\label{fig:rv_mcmc}
\end{figure}
The presence of a planet or a companion star results in temporal
variations in the radial velocity of the host star. By
analyzing the radial velocity data, one can draw inferences
about the ratio of masses between the host and the companion,
and orbital parameters like the period and eccentricity.
We now describe how to set up the above inference problem in a
Bayesian framework. We begin by describing the predictive
model for the radial velocity of a star in a binary system.
The radial velocity of a star of mass $M$ in a binary system
with companion of mass $m$ in an orbit with time period
$T$, inclination $I$ and eccentricity $e$ is given by
\begin{eqnarray}} \def\ee{\end{eqnarray}
v(t)=\kappa \left[\cos(f+\omega)+e\cos
\omega\right]+v_{0},\:\:{\rm with\ } \kappa=\frac{(2\pi G)^{1/3}m \sin I
}{T^{1/3}(M+m)^{2/3}\sqrt{1-e^2}}.
\label{equ:radial_vel}
\ee
The true anomaly $f$ is a function of time, but depends upon $e$,
$T$, and $\tau$ via,
\begin{eqnarray}} \def\ee{\end{eqnarray}
\tan(f/2)=\sqrt{\frac{1+e}{1-e}}\tan(u/2), \quad u-e\sin u=\frac{2\pi}{T}(t-\tau).
\ee
An example of radial velocity data is shown in
\fig{rv_mcmc} which shows the radial velocity for two
binary systems (the green and the red line) that differ in $e$
but have same values for all other parameters
$\kappa,T,\tau,\omega$ and $v_0$. The figure demonstrates that the radial
velocity is sensitive to the eccentricity of the orbit.
\begin{marginnote}
\entry{$v_{0}$}{the mean velocity of the center of mass of
the binary system}
\entry{$I$}{the inclination of the orbital plane with
respect to the sky (angle between orbital angular momentum and
line of sight)}
\entry{$\omega$ }{the angle of the pericenter
measured from the ascending node (the point where the orbit
intersects the plane of the sky)}
\entry{$\tau$ }{time of passage through the pericentre}
\end{marginnote}
The actual radial velocity data will differ from the
perfect relationship given in \equ{radial_vel} due to
observational uncertainty (variance $\sigma_v^2$) and
intrinsic variability of a star (variance $S^2$) and we can
model this by a Gaussian function
$\mathcal{N}(.|v,\sigma_v^2+S^2)$.
For radial velocity data $D$ defined as a set of radial velocities $\{v_1,...,v_M\}$ at various
times $\{t_1,...,t_M\}$, one can fit and constrain seven
parameters, $\theta=(v_{0}, \kappa, T, e, \tau, \omega, S)$,
using the Bayes theorem as shown below
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta|D) \propto p(D|\theta) p(\theta) \propto p(\theta) \prod_{i=1}^{M} \mathcal{N}(v_i|v(t_i;\theta),\sigma_v^2+S^2).
\ee
We generated test data using \equ{radial_vel} and then,
using the above equation, we tried to recover the parameters
$\theta$ (available in the supplied software).
The posterior distribution $p(\theta|D)$ was
sampled using MCMC, and the results are shown in
\fig{rv_mcmc}. Panel $a$ shows the test data along with
the best fit curve. It also shows
the radial velocity for the case with $e=0$. Panel $b$
shows the posterior distribution of the parameters
$\kappa, T$ and $e$.
If we have data for a large number of binary systems,
we can use it to explore the distribution of
orbital parameters. A naive way to do this would be
to get a ``maximum {\it a posteriori}'' (MAP) estimate of the
orbital parameters for each star and then
study the population distribution by constructing
histograms out of it. Such a scheme will give incorrect
estimates of the population distribution as the
uncertainty associated with the parameter estimates
is ignored. In addition to this,
as discussed by \cite{2010ApJ...725.2166H}, the MAP
estimates are in general biased. In the context of
radial velocity data, the estimates of $e$ are biased
high. The problem is especially
acute if the uncertainty associated with the parameters
is large, which is often the case with
radial velocity data from barycentric motions.
All of these problems can be avoided by setting up the
problem of estimation of population distributions as
a hierarchical Bayesian model.
Let us suppose we have radial velocity
data for $N$ binary star systems, and denote by $y_i$ the radial
velocity data set for the $i$-th system. Let $x_i=(v_{0i},
\kappa_i, T_i, e_i, \tau_i, \omega_i, S_i)$ be the orbital
parameters for the $i$-th system. Finally, let $\alpha$
be the set of hyperparameters that govern the
population distribution of the parameters $x$. The problem
to determine $\alpha$ can be set up as
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\alpha,\{x_i\}|\{y_i\}) \propto p(\alpha) \prod p(y_i|x_i)
p(x_i|\alpha)
\ee
This is a BHM and
can be sampled using the Metropolis-within-Gibbs scheme
discussed in \sec{gsem}. The parameters $x_i$ can be
estimated alongside $\alpha$, and to get the marginal
distribution $p(\alpha|\{y_i\})$, one can simply ignore
the computed $x_i$.
However, the above scheme is not well suited
to explore a variety of population models,
especially if sampling from $p(y_i|x_i) p(x_i|\alpha)$ is
computationally demanding.
We now show a computationally efficient
scheme by \citet{2010ApJ...725.2166H}
that can in general be applied to BHMs of two levels.
The marginal distribution of hyperparameters that
we are interested in is given by
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\alpha|\{y_i\}) \propto p(\alpha) \prod \int dx_i
p(y_i|x_i) p(x_i|\theta)
\ee
The integral on the right hand side can be estimated using
a Monte Carlo integration scheme as follows:
\begin{eqnarray}} \def\ee{\end{eqnarray}
\int dx_i p(y_i|x_i) p(x_i|\alpha) = \int dx_i
p(y_i|x_i) p(x_i) \frac{p(x_i|\alpha)}{p(x_i)} =
\frac{1}{K}\sum_{k=1}^{K}
\frac{p(x_{ik}|\alpha)}{p(x_{ik})},
\ee
with $x_{ik}$ sampled from $p(x_i|y_i) \propto p(y_i|x_i)
p(x_i)$, which can be done by an MCMC scheme.
\subsection{Data driven approach to estimation of stellar parameters from a spectrum}
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{data_model_cyan_cannon1.pdf}
\caption{The APOGEE spectra of four stars (black line) along with the best
model spectra generated by {\it The Cannon} algorithm
(cyan line) along with scatter around the fit. Each row
shows the spectra of a single star in
two wavelength intervals (left and right). Image from \citet{2015ApJ...808...16N}.}
\label{fig:data_model_cannon}
\end{figure}
The spectrum of a star contains information about its
properties like temperature, gravity and the abundance of
different chemical elements that make up the star.
Decoding information about stellar parameters
from a stellar spectrum is a problem of great significance
for astronomy. With the advent of large spectroscopic
stellar surveys having several hundred thousand spectra,
the need for fast and accurate methods to analyze the
stellar spectra has gained prominence.
Let us denote the stellar parameters
(e.g., $T_{\rm eff},\log g,{\rm [Fe/H]},$ and ${\rm [X/Fe]}$)
by label vector ${\bf x}=(x_1,...,x_K)$ and
the observed spectrum by vector ${\bf
y}=\{y_1,...,y_L\}$, denoting normalized flux at
specific wavelengths indexed by ${\bf \lambda}=(1,...,L)$
(see \fig{data_model_cannon}).
The problem is to find ${\bf x}$ given ${\bf y}$, which
using the Bayes
theorem can be written down as
\begin{eqnarray}} \def\ee{\end{eqnarray}
p({\bf x}|{\bf y},\theta) \propto p({\bf y}|{\bf x},\theta)p({\bf
x}).
\label{equ:cannon_bayes}
\ee
Here
$p({\bf y}|{\bf x},\theta)$ denotes a probabilistic generative model
for the data, with $\theta$ being the parameters of the
model. If we denote by $f_{\lambda}({\bf x},\theta_{\lambda})$ the flux
predicted by the model at wavelength $\lambda$ and by
$s_{\lambda}^2$ the variance or scatter about this relation
(assuming Gaussian noise), then the probabilistic generative
model for the full spectrum can be written as
\begin{eqnarray}} \def\ee{\end{eqnarray}
p({\bf y}|{\bf x},\theta)=\prod_{\lambda=1}^L p(y_{\lambda}|{\bf x},\theta_{\lambda},s_{\lambda})=\prod_{\lambda=1}^L\mathcal{N}(y_{\lambda}|f_{\lambda}({\bf x},\theta_{\lambda}),s_{\lambda}^2)
\label{equ:cannon_genmod}
\ee
Traditionally, $f_{\lambda}({\bf x},\theta_{\lambda})$ is calculated
from first principles using a physical theory for the
formation of spectral lines in a stellar atmosphere
specified by stellar parameters ${\bf x}$. Frequently,
$f_{\lambda}({\bf x},\theta_{\lambda})$ is evaluated on a grid defined
on $x$ and then interpolation is used to get the spectrum
for any arbitrary value of ${\bf x}$.
The $f_{\lambda}({\bf x},\theta_{\lambda})$ can also be computed
by interpolating over a library of empirical spectra with predefined
stellar parameters. A more refined data
driven approach to the problem
using machine learning techniques was presented in
\citet{2015ApJ...808...16N}.
In this approach, $f_{\lambda}({\bf
x},\theta_{\lambda})$ is approximated by a simple
(linear or quadratic) function of label vector ${\bf x}$. Therefore
\begin{eqnarray}} \def\ee{\end{eqnarray}
f_{\lambda}({\bf x},\theta_{\lambda})=\theta_{\lambda
0}+\sum_{i=1}^{K} \theta_{\lambda i}x_{i} +\sum_{i=1}^{K}\sum_{j=1}^{K}
\theta_{\lambda ij}x_ix_j.
\ee
Let us consider a training set of $N$ stars
with label vectors $X=\{{\bf x}^1,...,{\bf x}^N\}$
and corresponding set of fluxes at wavelength $\lambda$
by $Y_{\lambda}=\{y_{\lambda}^1,...,y_{\lambda}^N\}$.
One can estimate $\theta_{\lambda}$ by sampling within
MCMC such that
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta_{\lambda},s_{\lambda}|X,Y_{\lambda}) \propto
p(Y_{\lambda}|X,\theta_{\lambda},s_{\lambda})p(\theta_{\lambda})p(s_{\lambda})
\propto p(\theta_{\lambda})p(s_{\lambda})\prod_{i=1}^{N}
p(y_{\lambda}^{i}|{\bf x}^{i},\theta_{\lambda}s_{\lambda}).
\ee
Having obtained the model parameters
$\theta=\{\theta_1,...,\theta_L,s_1,...,s_L\}$, one can now
estimate stellar parameters ${\bf x}$ of a new star with
given spectrum ${\bf y}$ using \equ{cannon_bayes}.
This is the basis of {\it The Cannon} algorithm
(Ness et al. 2015) which is already widely used by
the stellar community.
The ability of the algorithm to model the spectra is
demonstrated in \fig{data_model_cannon} which shows the
spectra of fours stars along with the best-fit spectra for
each of them.
\subsection{Solar-like oscillations in stars}
\begin{figure}
\centering
\includegraphics[width=0.9\textwidth]{15263fg2.pdf}
\caption{Left: The power density of four stars observed with
{\it Kepler} showing solar-like
oscillations along with best fit model (solid black). The
hump is the approximate Gaussian-like envelope that
modulates the power spectrum. The dashed lines
are the individual super-Lorentzian profiles. The dotted
line is the background model without the Gaussian component. Right: The
spectrum region around central frequency $\nu_0$ after subtracting the
background model. Individual modes are clearly visible. The degree $l$
of the modes is labelled in the top panel. The frequency
separations $\Delta \nu, \delta_{01}$ and $\delta_{02}$
are shown using dotted lines in the bottom panel. Image
from \citet{2010A&A...522A...1K}.}
\label{fig:kallinger}
\end{figure}
Solar-like oscillations, which are excited and damped in
the outer convective envelopes of a star, are seen in stars
like the Sun and red giants. With the advent of space-based
missions like {\it Kepler} and COROT that provide high
quality photometric data over a long time series, it has now
becomes feasible to detect solar-like oscillations in
tens of thousands of stars \citep{2013ApJ...765L..41S, 2015ApJ...809L...3S}.
Typically, the power spectrum of a star with solar-like
oscillations (\fig{kallinger}) shows a regular pattern of
modes, characterized by a large
frequency separation $\Delta \nu$. The overall amplitude
is modulated by a Gaussian envelope and this is
characterized by the frequency of maximum oscillation $\nu_{\rm max}$.
Theory suggests that $\Delta \nu$ for a given star is related to its density
\citep{1986ApJ...306L..37U}, whereas the $\nu_{\rm
max}$ is related to its surface gravity and
temperature \citep{1991ApJ...368..599B,1995A&A...293...87K}.
Using the above two relations, the mass and the radius of a
star can be constrained.
The mass of a red giant is sensitive to its age
and this makes asteroseismology very useful for understanding
Galactic evolution \citep{2011Sci...332..213C, 2016ApJ...822...15S}.
For further details on solar type oscillations see review
by \citet{2013ARA&A..51..353C}.
Bayesian-MCMC based techniques are increasingly being
adopted to extract seismic properties, e.g., $\Delta \nu$
and $\nu_{\rm max}$, by analyzing the power spectrum
generated from the time series photometry of a star
\citep{2009A&A...506.1043G,2010A&A...522A...1K,2011A&A...527A..56H}.
The probability that an observed power spectrum
$\mathbf{\Gamma}=\{\Gamma_1,...,\Gamma_N\}$ at frequencies $\mathbf{\nu}=\{\nu_1,...,\nu_N\}$ is
produced by a model spectrum $\Gamma(\nu;\theta)$ (specified by a set of
parameters $\theta$), is given by
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\mathbf{\Gamma}|\theta)=\prod_{i=1}^{N} \frac{1}{\Gamma(\nu_i;\theta)}\exp\left(-\frac{\Gamma_i}{\Gamma(\nu_i;\theta)}\right)
\ee
as shown by \citet{duvall1986solar}. This forms the basis
for the Bayesian treatment of the problem of estimation of
parameters $\theta$ by
$p(\theta|\{\Gamma_i\})=p(\{\Gamma_i\}|\theta)p(\theta)$.
The power density is modelled as a sum of super-Lorentzian
functions
\begin{eqnarray}} \def\ee{\end{eqnarray}
\Gamma(\nu;\theta)=\Gamma_{\rm
wn}+\sum_k\frac{A_k}{1+(2\pi\nu\tau_k)^{c_k}}+P_{\rm
g}\exp\left(\frac{-(\nu_{\rm max}-\nu)^2}{2\sigma_{\rm g}^2}\right)
\ee
To fit the individual modes, one assumes Lorentzian
profiles. Spherical harmonics are used to describe the
oscillations; the modes are characterized
by three wave numbers, $n,l$ and $m$.
In \citet{2010A&A...522A...1K},
eight main modes are fitted
(three $l=0$ and $l=2$ and two $l=1$), parameterized by
the mode lifetime $\tau$, the central frequency $\nu_0$, three
spacings $\Delta \nu, \delta\nu_{01}$ and
$\delta\nu_{01}$, and the amplitudes $A_i,A_j$ and $A_k$.
\begin{eqnarray}} \def\ee{\end{eqnarray}
\Gamma(\nu)&=&P_{\rm wn}+\sum_{i=-1}^{1}
\frac{A_i^2\tau}{1+4[\nu-(\nu_0+i\Delta\nu)]^2(\pi\tau)^2}
+\sum_{j=-1}^{1}\frac{A_j^2\tau}{1+4[\nu-(\nu_0+j\Delta\nu-\delta\nu_{02})]^2(\pi\tau)^2}\nonumber\\
& & +\sum_{k=-1,1} \frac{A_j^2\tau}{1+4[\nu-(\nu_0+k\Delta\nu/2-\delta\nu_{01})]^2(\pi\tau)^2}
\ee
\fig{kallinger} shows the result of fitting the above model
to power spectra of fours stars observed by the {\it Kepler}
mission.
\subsection{Extinction mapping and estimation of intrinsic
stellar properties}
Given the mass $m$ and initial composition (e.g., metallicity
[M/H]) of a star, we can use the
theory of stellar evolution to predict its state and
composition at a later time (age $\tau$).
However, the
intrinsic parameters like mass $m$, [M/H] and
$\tau$ are not directly observable. For most stars
we only have photometric information, apparent
magnitudes in different photometric bands (for example $J,
Ks, u, g, r$ and $i$). The photometry
of a star depends upon temperature $T_{\rm eff}$,
gravity $g$, [M/H], distance $s$ and extinction $E$
(proportional to the dust density integrated along the line
of sight to the location of the star).
If we have spectroscopy, then we can get temperature $T_{\rm eff}$,
$g$ and even composition, but with uncertainties. From asteroseismology,
we can get average seismic parameters like $\Delta \nu$
and $\nu_{\rm max}$, which are sensitive to the mass, radius
and temperature of a star. Given this state of affairs, it
is quite common to ask the question that, given a
certain set of observables of a star, what are the intrinsic
parameters of a star or even some other set of observables.
For example, given the photometry of a star, what is the
distance, temperature and gravity of a star; or given
photometry and distance, what is the temperature and gravity
of a star; or given photometry and spectroscopy, what is the
distance? And so on.
Knowing the intrinsic parameters of a star is also important
for understanding the formation and evolution of the Galaxy,
for example, the star formation rate,
the age-metallicity relation and the distribution of dust
in the Galaxy.
The problem of estimating intrinsic stellar parameters
of a star given some observables can be formulated as
follows. Let
${\bf y}=(J,J-Ks,J-H,T_{\rm eff},\log
g,[M/H]_{\rm obs},l,b)$
be a set of observables associated with a star
and $\sigma_{\bf y}$ their uncertainties.
Let us denote the intrinsic variable of a star
that we are interested in by
${\bf x}=([M/H],\tau,m,s,l,b,E)$.
To specify prior probabilities on ${\bf x}$
we need a Galactic model, and we denote by $\theta$ the
parameters of such a model.
Typically, real catalogs have selection effects,
e.g., stars selected to lie in some apparent magnitude
and color range, or a set of stars with parallax error
less than 10\%, or stars with missing information in
certain bands. To specify selection effects,
we denote the event that a star exists in a catalog
by $S$.
From theoretical isochrones we can predict ${\bf y}$
given ${\bf x}$, in other words a function
${\bf y(x)}$ exists.
However, we are interested in the inverse problem of
estimating ${\bf x}$ given ${\bf y}$. A Bayesian
introduction to solving such a problem was given by
\citet{2004MNRAS.351..487P} and \citet{2005A&A...436..127J} in the
context of estimating ages. The method was further
improved and refined by
\citet{2010MNRAS.407..339B,Binney_Burnett_2011}
and \citet{2014MNRAS.437..351B} in the context of the estimation of
distances, with a better treatment of priors and selection
effects \citep[see also][]{2012MNRAS.427.2119S,2015MNRAS.452.2960S}.
From the Bayes theorem we have
\begin{eqnarray}} \def\ee{\end{eqnarray}
p({\bf x| y,\sigma_y},S,\theta) \propto p(S,{\bf
y| x,\sigma_y}) p({\bf x}|\theta) \propto
p(S|{\bf y})p({\bf y}|{\bf x,\sigma_y})p({\bf x}|\theta).
\ee
We now explain each of the terms in detail.
\begin{enumerate}
\item $p({\bf x|y ,\sigma_y},S,\theta)$ is the {\it
posterior} distribution of intrinsic parameters given the
observables, the selection function and a Galactic model.
\item $p(S|{\bf y,x,\sigma_y})$ is the {\it selection function}. This says
given the observables what is probability that a star was
observed. Typically this can be expressed as $p(S|{\bf
y})p(S|{\bf x})$. The term $p(S|{\bf x})$ enters in
situations where the value of an observable $y'$ is not known
but constraints on it are. Then $p(S|{\bf x})=\int
p(S|y')p(y'|x)dy'$. For example, a parallax of a star is
known to be greater than a certain limit, or the apparent
magnitude of a star may be missing in a band because the
star is too bright or faint
\citep{2010MNRAS.407..339B, 2012MNRAS.427.2119S}.
\item $p({\bf y|x,\sigma_y})$ is the {\it likelihood} of the data given
the uncertainty and the intrinsic parameters. This can be
described by a Gaussian function $\mathcal{N}(y|y({\bf
x}),\sigma_y^2)$ for each $y\in {\bf y}$.
\item $p({\bf x}|\theta)$ is the {\it prior}. This describes the
distribution of mass, metallicity, age and spatial distribution of
stars in the Galaxy. More specifically it can be
written as
$p(x|\theta)=\sum_k p_k(m) p_k([M/H])p_k(\tau)p_k(r)$,
where the sum is over different Galactic components, e.g.,
thin disc, thick disc, bulge and stellar halo.
\end{enumerate}
We now focus on the problem of estimating distance and
extinction. For simplicity, we ignore the selection effects;
for an in depth discussion, see
\citet{2015MNRAS.452.2960S}.
By marginalizing over stellar parameters $\tau,m$ and
$[M/H]$ one obtains $p(s,E|{\bf y,\sigma_y},\theta)$. If we have $N$ stars along a line of sight, we can estimate the distance-extinction relationship $E(s_i;\alpha)$ parameterized by $\alpha$ as
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\alpha|\{{\bf y}\},\theta) & \propto & p(\alpha) \prod_{i=1}^{N}
\int d E_{i} ds_i\ p(s_i,E_i|{\bf y}_i,\sigma^i_{\bf y},\theta)\:
p(E_i|s_i,\alpha).
\ee
The above method is used by
\citet{2014ApJ...783..114G,2015ApJ...810...25G}, to
construct three dimensional maps of interstellar dust
reddening using Pan-STARRS 1 and 2MASS photometry
(\fig{3d_ext_green}). To estimate $p(s,E|{\bf
y,\sigma_y},\theta)$, \citet{2015ApJ...810...25G}
do a kernel density estimate over samples generated by
MCMC, while \citet{2015MNRAS.448.1738S}
present a method based on the Gaussian mixture model.
As described in \citet{2012MNRAS.427.2119S}, we can
also directly estimate $\alpha$ and intrinsic
parameters ${\bf x}$ of each star along a line of sight by setting
up the problem as a BHM and sampling
from the following posterior:
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\alpha,\{\bf x\}|\{{\bf y}\},\sigma_{\bf y},\theta) & \propto & p(\alpha) \prod_{i=1}^{N}
p({\bf y}_i|{\bf x}_i,\theta,\sigma_{\bf y}^i,\alpha) p({\bf
x}_i|\theta) p(S|{\bf y}_i).
\ee
The Metropolis-within-Gibbs scheme is used to
accomplish this sampling.
\begin{figure}
\centering
\begin{minipage}{6cm}
\centering
\includegraphics[width=6cm]{f07_green15.png}
\end{minipage}
\begin{minipage}{6cm}
\centering
\includegraphics[width=6cm]{f08_green15.png}
\end{minipage}
\caption{A three-dimensional map of interstellar dust
reddening in the Milky Way based on Pan-STARRS 1 and
2MASS photometry. Shown are the mean differential reddening in different
heliocentric distance ranges. The map is produced by \citet{2015ApJ...810...25G} and is available at http://argonaut.skymaps.info.
}
\label{fig:3d_ext_green}
\end{figure}
\subsection{Kinematic and dynamical modelling of the Milky Way}
Understanding the origin and evolution of the Milky Way
has received significant boost due to the emergence of large
data sets that catalog the properties of stars in the Milky
Way \citep{2011Prama..77...39B,2012MNRAS.419.2251M,2013MNRAS.433.1411M, 2013A&ARv..21...61R, 2013NewAR..57...29B,
2016arXiv160207702B1}.
Bayesian methods and MCMC based schemes are
now playing a prominent role in the analysis and interpretation
of such large and complex data sets from, e.g., the GCS
survey \citep{2010MNRAS.403.1829S}, the SEGUE
survey \citep{2012ApJ...753..148B}, the APOGEE survey
\citep{2012ApJ...759..131B,2013ApJ...779..115B},
and the RAVE survey \citep{2014ApJ...793...51S,2014MNRAS.445.3133P,
2015MNRAS.449.3479S}. We focus on the problem of
determining the mass distribution, or equivalently the
gravitational potential of the Milky Way, using
halo stars \citep{2014ApJ...794...59K} and disc masers \citep{2017MNRAS.465...76M}.
The observational data of stars in the Milky Way is in
heliocentric coordinates and is in the
form of angular positions on sky
(Galactic longitude $\ell$ and latitude $b$), heliocentric
distance ($s$), heliocentric line of sight velocity
($v_{\rm los}$), and proper motion (tangential motion on the
sky, $\mu_\ell$ and $\mu_b$).
The velocity of halo stars can be described by a
simple Gaussian model of the following form
\begin{eqnarray}} \def\ee{\end{eqnarray}
p({\bf v}|\theta_v,\ell,b,s)=\mathcal{N}(v_r|0,\sigma_{vr})
\mathcal{N}(v_{r}|0,\sigma_{v\theta})
\mathcal{N}(v_{\phi}|v_{\rm rot},\sigma_{v\phi})
\ee
for which $\theta_v$ is the set of parameters that govern the
velocity dispersion profiles $\sigma_{vr},\sigma_{v\theta}$
and $\sigma_{v\phi}$.
The coordinates $(r, \theta, \phi)$ are in the Galactocentric
reference frame. The observed heliocentric coordinates
can be converted to Galactocentric coordinates using
prior estimates of the location and the motion of the sun.
For the stellar halo stars, tangential velocities
cannot be accurately determined.
The distance also has some uncertainty, $\sigma_s$.
Hence we marginalize over unknown tangential velocities and
true distance $s'$, to obtain
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(v_{\rm los}|\theta_v,\ell,b,s,\sigma_s) = \int \int \int
p(v_\ell,v_b,v_{\rm los}|\theta_v,\ell,b,s') p(s'|s,\sigma_s)\: d v_\ell d v_b d
s'
\label{equ:vlos_sh}
\ee
The parameters $\theta_v$ can now be estimated using the data $D$
of multiple stars by
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta_v|D) \propto \left( \prod_i p(v_{\rm los}|\theta_v,\ell_i,b_i,s_i,\sigma_{s,i})\right) p(\theta_v)
\ee
The marginalization in \equ{vlos_sh} can be handled in
various ways. One can make use of deterministic numerical
integration techniques (Gaussian quadrature) or one can
achieve marginalization
via Monte Carlo schemes making use of importance sampling.
For Monte Carlo based integration one can make use of the
MCMH algorithm discussed in \sec{MCMH}.
Alternatively, one
can treat $v_\ell,v_b$ and $s$ as unknowns by setting them
up as a BHM and estimate
them alongside $\theta$ by making use of the
Metropolis-within-Gibbs scheme discussed in \sec{gsem}.
The radial velocity dispersion profile of halo stars
computed using blue horizontal branch and red giant stars
in the SEGUE survey is
shown in \fig{mvir_c_kafle} \citep{2014ApJ...794...59K}.
We now proceed to estimating the potential $\Phi$.
Given $\Phi$, density of halo stars $\rho$ and anisotropy
$\beta=1-(\sigma_{v \theta}^2+\sigma_{v \phi}^2)/(2
\sigma_{v r}^2)$ as function of distance $r$
from the Galactic center, one can
solve for $\sigma_{vr}(r)$.
Let $\theta$
be the set of parameters used to define the above profiles.
So for given $\theta$, the model makes a prediction for
radial velocity dispersion $\sigma_{vr}(r_i;\theta)$ at a
location $r_i$. This can be compared with the
$\sigma_{vr}(r_i)$ estimated from the observed data.
The probability of model parameters $\theta$ is then given
by
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta|D) \propto p(\theta) \prod_{i=1}^{M} \mathcal{N}(\sigma_{vr}(r_i)|\sigma_{vr}(r_i;\theta),\gamma_i).
\ee
The posterior distribution for the virial mass and the concentration
parameter of the Milky Way halo using BHB and giant stars is shown in
\fig{mvir_c_kafle} \citep{2014ApJ...794...59K}.
\begin{figure}
\centering
\begin{minipage}{6cm}
\centering
\includegraphics[width=5cm]{kafle_sigma_vr.pdf}
\end{minipage}
\begin{minipage}{6cm}
\centering
\includegraphics[width=5cm]{kafle_mvir_c.pdf}
\end{minipage}
\caption{$(a)$ Radial velocity dispersion as a function of radius
for halo stars in the Milky Way. $(b)$
Posterior distribution of virial mass and the
concentration parameter of the Milky Ways dark mater
halo. Adapted from \citep{2014ApJ...794...59K} with
permission. Abbreviation: BHB, blue horizontal branch}
\label{fig:mvir_c_kafle}
\end{figure}
We now discuss ways to incorporate prior
information into the analysis. For example,
the angular velocity of the Sun with respect to the
Galactic Center $\omega$ is well constrained to
be within $30.24\pm0.12\: {\rm km\ s}^{-1} {\rm kpc}^{-1}$ \citep{2004ApJ...616..872R}. The vertical force at $1.1$ kpc above the Sun, in terms of surface mass
density, is given by $\Sigma_{1.1,\odot}=72\pm6$ \citep{1991ApJ...367L...9K}.
Let us denote such constraints by
$p(g_j(\theta)|\theta)$.
Additional data sets $D_k$, constraining a certain subset of parameters
can also exist. For example, the tangent point velocities
or terminal velocities as a function of Galactic longitude
$v_{\rm term}(\ell)$ help to constrain the shape of the circular velocity curve
$v_{\rm circ}(R)=\sqrt{|Rd\Phi/dR|}$. The additional priors and data
all enter as multiplicative factors in the posterior, which
is given by
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\theta|D_1,...,D_K) \propto p(\theta) \prod_{k=1}^{K}
p(D_k|\theta) \prod_{j=1}^{J}p(g_j(\theta)|\theta).
\label{equ:mass_post}
\ee
The halo stars carry little information about the
mass distribution close to the center and in the disc of the
Milky Way.
Galactic masers associated with high mass
star forming regions are very good tracers of the Milky Way
disc which makes them excellent candidates for studying the
potential of the Milky Way
\citep{2009ApJ...700..137R,2011MNRAS.414.2446M,2014ApJ...783..130R,2017MNRAS.465...76M}. Due to extremely accurate
astrometric information using very long baseline
interferometry, one has very accurate parallax ($\varpi$) and proper
motion measurements. When combined with line of sight
velocities from Doppler shift of spectral lines, one ends up
with full 6D phase space information for these
sources. Maser sources, are young and have
very little random motion which means their orbits are
highly circularized. The distribution of velocities
can be described by a simple three dimensional Gaussian function, i.e.
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(v_R,v_{\phi},v_{z}|\theta) = \mathcal{N}(v_{\phi}|v_{\rm
circ}(R;\theta)+v_{\phi,{\rm
M}},\sigma_{v{\rm M}}) \mathcal{N}(v_{R}|v_{R,{\rm
M}},\sigma_{v{\rm M}}) \mathcal{N}(v_{z}|v_{z,{\rm
M}},\sigma_{v{\rm M}})
\label{equ:maser_vel}
\ee
Here, ${\bf v}_{\rm M}=(v_{R,{\rm M}},v_{\phi,{\rm M}},v_{z,{\rm M}})$
is any systematic streaming velocity associated with the masers
and $\sigma_{v{\rm M}}$ is the velocity dispersion about the mean motion.
Now, we have
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(\mu_{\alpha},\mu_{\delta},v_{\rm los}|\theta) &= & \int
d \varpi' p(\varpi'|\varpi) \int \int
\int d\mu_{\alpha}'
\ p(\mu_{\alpha}'|\mu_{\alpha}) d\mu_{\delta}' p(\mu_{\delta}'|\mu_{\delta})
\nonumber \\
& & dv_{\rm los}' p(v_{\rm los}'|v_{\rm los})p(\mu_{\alpha}',\mu_{\delta}',v_{\rm los}'|\theta,\varpi')
\ee
The last term is evaluated using \equ{maser_vel}, by
converting from heliocentric coordinates $(\mu_{\alpha}',\mu_{\delta}',v_{\rm
los}',\alpha,\delta, \varpi')$ to Galactocentric
coordinates $(v_R,v_{\phi},v_{z},R,\phi,z)$.
Let $D_1$ denote the full data of $N$ stars then
\begin{eqnarray}} \def\ee{\end{eqnarray}
p(D_1|\theta)=\prod_{i=1}^{N} p(\mu_{\alpha,i},\mu_{\delta,i},v_{\rm los,i}|\theta)
\ee
This when put in \equ{mass_post} gives
the posterior distribution of model parameters.
\section{Concluding remarks}
The power of the Bayesian probability theory lies in the
fact that it is mathematically simple, being based on just
two elementary rules, and yet it is broadly applicable.
However, Bayesian calculations can be computationally
demanding, and this has acted as a major bottleneck in the
past. But with the increase of computational
power, we have witnessed a sharp increase in the adoption
of Bayesian techniques.
More recently, free availability of black-box computer
packages
to efficiently sample from Bayesian posterior distributions
has further accelerated the adoption of Bayesian
techniques in astronomy.
Robust algorithms are now available to sample
multidimensional and complex pdfs.
The MH algorithm
is still the main workhorse of MCMC methods.
Good solutions now exist for the issue of application
specific tuning of the proposal distribution in the
MH algorithm, e.g., adaptive Metropolis schemes
and the affine invariant samplers.
The MH algorithm
when combined with parallel tempering allows one
to sample a wide variety of commonly occurring
distributions. Situations, in which the posterior
is not analytically tractable, can also
now be solved using the Monte Carlo version
of the MH algorithm.
Bayesian methods also provide a framework for
model comparison via the use of Bayesian evidence.
However, efficient computing of evidence still remains
a challenge. Various alternate criteria for comparing
models exist and importantly these can make use of the
computed MCMC chain.
Bayesian hierarchical models further increase the usefulness
of the Bayesian framework. They can solve missing data
problems, marginalization over variables, convolution
with observational uncertainties and so on. This makes
a wide class of complex problems suddenly solvable.
We showed that the Metropolis-within-Gibbs scheme is ideally
suited for sampling posteriors generated by
Bayesian hierarchical models and also provide
a software for doing this.
Multimodal distributions still pose a problem for most
MCMC algorithms. Parallel tempering can overcome them but
requires more computational time and a careful choice of
ladder. If dimensionality of the space being explored
is very high and the distribution is complex, efficient
exploration is not easy. Techniques are being developed to
solve such problems that make use of derivatives of the
posterior distribution, e.g., Hamiltonian Monte Carlo.
However, more work is required in this area.
Efficient exploration of multi-level hierarchical models
will play an increasingly important role in future
studies.
Communication of Bayesian results is also an area where we
anticipate improvements. Traditionally, the estimates are
reported by means of confidence intervals. However,
there is much more information in the MCMC chain, in
particular,
the correlation between different variables.
Also, there is an increasing need to feed results of one
MCMC simulation into another. Such requirements are best
addressed by reporting the full pdfs or the thinned samples
from it. Other alternatives that are economical in terms
storage space are to approximate the pdf by
analytical functions or to employ Gaussian mixture models.
We also need better tools to visualize the Bayesian-MCMC
output, specially for high dimensional and complex hierarchical models.
Such tools will allow us to understand as to
why a model fails and how we should improve it.
There are key topics which we have not addressed
here. Non-parametric Bayesian methods are
becoming increasingly important, e.g., Gaussian processes
\citep{beaumont2002approximate} and Dirichlet
process mixture models \citep{neal2000markov}.
\citet{2014MNRAS.437.2230M} uses this method
to estimate the gravitational potential
of the Milky Way.
Astronomy is no longer a data-starved science. With projects like
the Large Synoptic Survey Telescope and the Square
Kilometre Array, the quality
and quantity of data are going to increase dramatically in the
coming years.
Better quality and larger quantity of data means that we can
expect our data to answer more difficult questions,
which in
turn means more complex models (e.g. multi-level hierarchies
and a higher dimensional parameter space).
Given that
MCMC is a computationally expensive scheme, there will
be an increasing demand for such techniques that can
make full use of the vast quantity of data on offer
and deliver results in an affordable amount of time.
Equivalently,
MCMC schemes that make use of computing environments
with multiple processor and graphic processor units
would also be useful.
An MCMC chain is serial by nature and it
requires special care to parallelize an MCMC algorithm, e.g., use of an ensemble of chains
\citep{2013PASP..125..306F} or parallelizing the
posterior computation by splitting up the data.
Relaxing the condition of reversibility can lead to MCMC
algorithms with faster mixing properties
\citep{chen1999lifting,diaconis2000analysis,girolami2011riemann}.
Finally, the development of approximate methods,
both application specific and general, that can
reduce the computational cost without significantly
compromising the quality of results also hold great promise
for analyzing large data sets. Approximate Bayesian
computation is one such framework
\citep{beaumont2002approximate}; see
\citet{2016ApJ...817...49B} for its use in astronomy
to study the chemical homogeneity of stars in open clusters.
\section*{DISCLOSURE STATEMENT}
The author is not aware of any affiliations, memberships,
funding, or financial holdings that
might be perceived as affecting the objectivity of this review.
\section*{ACKNOWLEDGMENTS}
I am indebted to my colleague Joss Bland-Hawthorn for
suggesting this article and for supervising its
development over the past year. I am thankful
to James Binney, Jo Bovy, Brendon Brewer, Prajwal Kafle
and Prasenjit Saha for numerous suggestions and
discussions from which the review has benefited
significantly. I am also thankful to David Hogg
for words of encouragement on the draft.
I acknowledge
funding from a University of Sydney Senior Fellowship
made possible by the office of the Deputy Vice
Chancellor of Research, and partial funding from
Bland-Hawthorn's Laureate Fellowship from the
Australian Research Council.
\bibliographystyle{ar-style2}
| acfc2b4ace11b8de7eeab075175f598774429dac | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The study of roots and level lines of random functions is a central topic in mathematics, at the crossroad between Algebra, Analysis and Probability theory, which has been extensively studied since the mid 20th century. Being at the very definition of algebraic varieties, the investigation of the geometry of nodal sets associated to random algebraic or analytic functions is naturally of primer importance. In the pioneering work \cite{little1}, the authors considered the expected number of real zeros of random univariate algebraic polynomials with uniform, Gaussian, and discrete entries. Since then, lots of developments were made to estimate the asymptotic behavior, as their degree goes to infinity, of the real/complex level sets associated to such polynomials, under various assumptions on the law of the random entries, see for example \cite{kac1, kac2, erdos, ibragimov1,farah1, kostlan} and the references therein.
\newpage
Among the class of random functions, of particular interest are random trigonometric polynomials of the form $\sum_{k=1}^n a_k \cos (k t)$ or $\sum_{k=1}^n a_k \cos (k t)+b_k\sin (k t)$, where $(a_k)_{k \geq 1}$ and $(b_k)_{k \ge 1}$ are random coefficients, since the distribution of the zeros of such polynomials occurs in a wide range of problems in science and engineering. The asymptotics of the mean number of real zeros of random trigonometric polynomials with independent standard and centered Gaussian coefficients was first explicited by Dunnage in \cite{dunnage}, where it is shown that this number is asymptotically proportional to the degree $n$ of the considered polynomial. Since then, the level sets of random trigonometric polynomials have been intensively investigated in various directions. For example, still in the case of independent standard and centered Gaussian entries, the variance and the fluctuations around the mean were studied in the serie of papers \cite{faravar,wigman,azais,azais2016}. Beyond the purely Gaussian case but still considering independent standard and centered entries, the universality of the asymptotic local/global behavior of the number of real zeros was recently established in \cite{angst2015universality,AZ15,flasche2017,iksanov2016local}.
\par
\bigskip
To the best of our knowledge, the case of dependent Gaussian entries in trigonometric models has only been considered in \cite{samba,renganathan1984average} which focus on the two particular and somehow ``extreme'' cases of a constant correlation $\mathbb{E}(a_i a_j)=\rho \in ]0,1[$ and a geometric correlation $\mathbb{E}(a_i a_j)= \rho^{|i-j|}$. In both cases, it is nonetheless shown that the expected number of real roots lying in $[0,2\pi]$ obeys the same asymptotics
\begin{equation}\label{Meanbehavior}
\lim_{n \to +\infty} \frac{\mathbb E\left [N_n([0,2\pi])\right]}{n} = \frac{2}{\sqrt{3}}.
\end{equation}
The latter result naturally raises the question of ascertaining the sharp conditions on the correlation function $\rho$ of the random coefficients ensuring this universal asymptotic behavior. The main result of this article, i.e. Theorem 1 p. 9 below, provides a significant step in that direction by exhibiting mild conditions on the spectral density $\psi_\rho$ guaranteeing that (\ref{Meanbehavior}) indeed holds. Our proofs exploit the Kac--Rice formula and more specifically the interpretation of its underlying integrand in terms of convolutions with respect to suitable trigonometric kernels. In a spirit close to \cite{ibragimov1971expected}, that is to say by investigating sign changes of piecewise linear approximations of the underlying random functions, the analoguous question for Kac polynomials ($P_n(x)= \sum_{k\le n} a_k x^k$) has been tackled in \cite{glendinning1989growth}. Under some assumptions on the spectral density, yet not covering the case of the increments of a fractional Brownian motion, it is shown that
\[
\lim_{n \to +\infty} \frac{\mathbb E\left [N_n(\mathbb{R})\right]}{\log(n)} = \frac{2}{\pi}.
\]
\par
\bigskip
The plan of the article is the following. In the next Section 2, we introduce the considered model of random trigonometric polynomials with dependent coefficients and we explicit the hypotheses made on their correlation function. In Section 3, we introduce a suitable renormalization allowing us to express the covariance of the process and its derivative as convolutions with positive kernels which approximate unity. In the last Section 4, we use the celebrated Kac--Rice formula to express the expected number of real zeros and deduce its asymptotics.
\section{The model and the hypotheses}
We consider here random trigonometric polynomials of the form
\[
f_n(t):=\sum_{1\le k \le n} a_{k} \cos(kt) + b_{k} \sin(kt), \quad t \in \mathbb R,
\]
where $(a_{k})_{k\ge 1}$ and $(b_{k})_{k\ge 1}$ are two independent sequences of standard centered Gaussian variables with correlation function $\rho : \mathbb N \to \mathbb R$, namely
$\mathbb E[a_k a_{\ell} ]= \mathbb E[b_k b_{\ell}] =:\rho(|k-\ell|)$ and $\mathbb E[a_k b_{\ell} ] =0$, $ \forall k, \ell \in \mathbb N^*.$
We will suppose that the spectral function $\psi_{\rho}$
\[
\psi_{\rho}(x) := \sum_{k \in \mathbb Z} \rho(|k|) e^{i k x}, \quad x \in ]0, 2\pi[,
\]
{{} exists} and satisfies the following hypotheses: \par
\vspace{-0.5cm}
\begin{equation}\label{eq.hypo}
\psi_{\rho} \in \mathbb L^1([0, 2\pi],dx), \quad \psi_{\rho} \; \text{is continuous on} \; ]0, 2\pi[ \;\; \text{and} \;\; \gamma_{\rho}:=\inf_{t \in [0, 2\pi]} \psi_{\rho}(t) >0.
\end{equation}\par
\vspace{-0.2cm}
\noindent
By Riemann--Lebesgue Lemma, the above integrability condition ensures that the correlation coefficient $\rho(k)$ goes to zero as $k$ goes to infinity, but no condition is required on the speed of the decay, allowing us in particular to consider long-range correlations.
For instance, the assumptions \eqref{eq.hypo} are satisfied by the fractional Gaussian noise with Hurst index $1/2<H<1$. Indeed, in this case, the correlation function $\rho_H$ is defined as \par
\vspace{-0.1cm}
\[
\rho_H({{}k}) := \frac{1}{2} \left( |1+{{}k}|^{2H} +|1-{{}k}|^{2H} -2|{{}k}|^{2H} \right).
\]\par
\vspace{-0.1cm}
\noindent
Setting $c_H:=\sin(\pi H) \Gamma(2H+1)$ and as shown for example in Proposition 2.1 and Corollary 2.1 of \cite{beran1994}, , the associated spectral function is then \par
\vspace{-0.1cm}
\[
{{}\psi}_{\rho_H}(x) := 2 c_H ( 1-\cos(x) ) \sum_{j \in \mathbb Z}^{+\infty} \frac{1}{(2\pi j +x)^{2H+1}}.
\]\par
\vspace{-0.1cm}
\noindent
This function admits a pole at the origin in accordance with the long-range correlation feature, with ${{} \psi}_{\rho_H}(x) \sim c_H |x|^{1-2H}$, and it admits a global positive minimum at $\pi$, as illustrated in Figure 1 below, for different values of the Hurst parameter $H$.\par
\vspace{-0.1cm}
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=0.45]{spectral.png}
\end{center}\par
\vspace{-0.5cm}
\caption{The graph of the function ${{}\psi}_{\rho_H}$ on $]0, 2\pi[$ for $H=0.6$, $H=0.75$, $H=0.9$.}
\end{figure}
\newpage
\section{Normalization and convergence}
Let us now introduce the following normalized version $F_n(t)$ of the process $f_n(t)$
\[
F_n(t):=\frac{1}{\sqrt{n}}\, f_n(t) = \frac{1}{\sqrt{n}} \left( \sum_{1\le k \le n} a_{k} \cos(kt) + b_{k} \sin(kt) \right),
\]
and its derivative
\[
F_n'(t):= \frac{1}{\sqrt{n}} \left( \sum_{1\le k \le n} -k a_{k} \sin(kt) + k b_{k} \cos(kt) \right).
\]
Naturally, the zeros of $f_n$ and its normalized version $F_n$ coincide.
Moreover, the variance of $F_n(t)$ and $F_n'(t)$ have particularly nice expressions in terms of convolutions of the spectral function with non-negative kernels.
\begin{lem}\label{lem.var}
There exist two regular $2\pi-$periodic functions $K_n$ and $L_n$ such that for all $t \in [0,2\pi]$
\begin{equation}
\mathbb E [F_n(t)^2] = K_n \ast \psi_{\rho}(t), \qquad \mathbb E [F_n'(t)^2] = \frac{(n+1)(2n+1)}{6} L_n \ast \psi_{\rho}(t).
\end{equation}
The two functions $K_n$ and $L_n$ are non-negative kernels with $||K_n||_1 = ||L_n||_1 =1$ and
\[
\forall \varepsilon>0, \;\; \lim_{n \to +\infty} \int_{\varepsilon}^{2\pi-\varepsilon} K_n(x) dx = \lim_{n \to +\infty} \int_{\varepsilon}^{2\pi-\varepsilon} L_n(x) dx =0.
\]
In particular, under the hypotheses of Section 2, both functions $K_n \ast \psi_{\rho}$ and $L_n \ast \psi_{\rho}$ converge uniformly to $\psi_{\rho}$ on any compact subset of $]0,2\pi[$.
\end{lem}
\begin{proof}
For fixed $t$ and $n$, we have
\[
\begin{array}{ll}
\mathbb E[F_n(t)^2] & \displaystyle{=\frac{1}{n} \sum_{k,l=1}^n \rho(k-l) \cos((k-l)t) = 1 + \frac{2}{n} \sum_{l>k} \rho(k-l) \cos((k-l)t) }\\
& \displaystyle{= 1 + \frac{2}{n} \sum_{k=1}^{n-1} \sum_{l=k+1}^n \rho(k-l) \cos((k-l)t)= 1 + \frac{2}{n}\sum_{k=1}^{n-1} \sum_{r=1}^{n-k} \rho(r) \cos(r t) } \\
& \displaystyle{= 1 + \frac{2}{n}\sum_{r=1}^{n} (n-r) \rho(r) \cos(r t) = \sum_{r=-n}^{n} \left(1-\frac{|r|}{n}\right) \rho(r) e^{irt} = K_n \ast \psi_{\rho}(t),}
\end{array}
\]
where $K_n$ is the celebrated Fej\'er kernel, namely
\[
K_n(x) := \sum_{r=-n}^{n} \left(1-\frac{|r|}{n}\right) e^{irx} = \frac{1}{n} \left(\frac{\sin(n x/2)}{\sin(x/2)}\right)^2.
\]
It is well known that $K_n$ has unit $\mathbb L^1-$norm and is an approximation of unity. By hypothesis, the spectral function $\psi_{\rho}$ is continuous on $]0, 2\pi[$, in particular it is uniformly continuous on any compact subset $K$ of $]0, 2\pi[$, so that the convolution $K_n \ast \psi_{\rho}$ uniformly converges to $\psi_{\rho}$ on $K$ as $n$ goes to infinity.
In the same way, we have
\[\label{e:var}
\begin{array}{ll}
\mathbb E[F_n'(t)^2] & \displaystyle{=\frac{1}{n} \sum_{k,l=1}^n k l \rho(k-l) \cos((k-l)t) = \frac{1}{n} \sum_{k=1}^n k^2 + \frac{2}{n} \sum_{l>k} k l \rho(k-l) \cos((k-l)t) }\\
& \displaystyle{= \frac{1}{n} \sum_{k=1}^n k^2 + \frac{2}{n}\sum_{k=1}^{n-1} \sum_{r=1}^{n-k} k(r+k)\rho(r) \cos(r t) } \\
& \displaystyle{= \frac{1}{n} \sum_{k=1}^n k^2 + \frac{2}{n}\sum_{r=1}^{n-1} \rho(r) \cos(r t) \left( \sum_{k=1}^{n-r} k(r+k) \right) } \\
& \displaystyle{= \sum_{r=-(n-1)}^{n-1} \rho(r) e^{i rt} \left( \frac{1}{n} \sum_{k=1}^{n-|r|} k(|r|+k) \right).}
\end{array}
\]
With the convention that $\sum_{\emptyset} = 0$, and setting $\alpha_n:=6/((n+1)(2n+1))$ for $n \geq 1$, we have thus
\[
\mathbb E[F_n'(t)^2] = \frac{1}{\alpha_n} L_n \ast \psi_{\rho}(t),
\]
where
\[
L_n(x):= \alpha_n \sum_{r=-n}^{n} \left( \frac{1}{{n}} \sum_{k=1}^{n-|r|} k(|r|+k) \right) e^{i r x}.
\]
Alternatively, the function $L_n(x)$ can be written as
\[
\begin{array}{ll}
L_n(x)& =\displaystyle{\alpha_n \sum^{n}_{k,l=1}kl\cos((k-l)x)
=\alpha_n \left( \left[\sum^{n}_{k=0}k\cos(kx)\right]^2 +\left[\sum^{n}_{k=0} k\sin(kx)\right]^2 \right)} \\
\\
& = \displaystyle{\alpha_n \left| \sum^{n}_{k=0} k e^{i kx}\right|^2 = \alpha_n \left| \frac{(n+1)e^{i (n+1)x}}{1-e^{ix}} - \frac{i e^{ix} \left(1-e^{i(n+1)x}\right)}{\left( 1-e^{ix} \right)^2}\right|^2}.
\end{array}
\]
It is therefore non-negative and satisfies the folllowing inequality
\[
L_n(x) \leq \alpha_n \left( (n+1) |1-e^{ix}|^{-1} + 2|1-e^{ix}|^{-2}\right), \quad \forall x \in ]0, 2\pi[.
\]
For all small $\varepsilon>0$, we have thus
\[
\int_{\varepsilon}^{2\pi-\varepsilon} L_n(x) dx \leq\alpha_n \left( (n+1) \sin(\varepsilon/2)^{-1} + 2\sin(\varepsilon/2)^{-2}\right) = O(1/n).
\]
Finally, we have
\[
||L_n||_1=\frac{1}{2\pi} \int_{-\pi}^{\pi} L_n(x)dx = \alpha_n \times \frac{1}{n} \sum_{k=1}^n k^2 =1.
\]
The next figure illustrates the behavior of the kernel $L_n$ for different values of $n$.
\begin{figure}[ht]
\begin{center}
\includegraphics[scale=0.4]{helene.png}
\end{center}
\caption{The graph of the function $L_n$ on $]0, 2\pi[$ for $n=2,5,10,20$.}
\end{figure}
\end{proof}
\begin{rem}\label{r:n-d}
The positivity of the Fej\'er kernel $K_n$ combined with the lower bound hypothesis on the spectral function ensures that the variance of $F_n(t)$ is bounded below, namely
\begin{equation}\label{eq.lbound}
\mathbb E[F_n(t)^2] = K_n \ast \psi_{\rho}(t) \geq \gamma_{\rho} >0, \quad \forall t\in [0, 2\pi].
\end{equation}
\end{rem}
\par
\bigskip
Let us now describe the behavior as $n$ goes to infinity of the covariance between the process $F_n(t)$ and its derivative $F_n'(t)$.
\begin{lem}\label{lem.covar}
The covariance between $F_n(t)$ and $F_n'(t)$ is given by
\[
\mathbb E [F_n(t) F_n'(t)] = \frac{1}{2} \, K_n' \ast \psi_{\rho}(t),
\]
and under the hypotheses of Section 2, for any compact subset $K$ of $]0, 2\pi[$, it satisfies
\[
\lim_{n \to +\infty} \frac{\sup_{t \in K} \left| \mathbb E [F_n(t) F_n'(t)] \right| }{n} = 0.
\]
\end{lem}
\begin{proof}The expression of the covariance as a convolution is simply obtained by differentiating the one of $\mathbb E[F_n(t)^2]$. The derivative of the Fej\'er kernel is given by the explicit formula
\[
K_n'(x) = \frac{\sin(nx/2)\cos(nx/2)}{\sin^2(x/2)}-\frac{\cos(x/2)\sin^2(nx/2)}{n \sin^3(x/2)},
\]
from which it is clear that for any small $\eta>0$, there exists a finite constant $R_{\eta}$ such that
\begin{equation}\label{eq.primeup}
\sup_{x \in [\eta, 2\pi-\eta]}|K_n'(x)| \leq R_{\eta}.
\end{equation}
Moreover, using Bernstein inequality, see e.g. Theorem 3.16 p.11 of \cite{zygmund2003}, we have
\begin{equation}\label{eq.bernstein}
||K_n'||_1 \leq n ||K_n||_1 = n.
\end{equation}
Since the integral of $x \mapsto K_n'(x)$ over a period vanishes, for all $t \in K$, we can write
\begin{equation}\label{eq.prime}
\frac{1}{n} K_n' \ast \psi_{\rho}(t) = \frac{1}{2\pi} \int_{0}^{2\pi} \frac{K_n'(x)}{n} \left[ \psi_{\rho}(t-x) - \psi(t) \right] dx.
\end{equation}
For $\alpha>0$, let us denote by $K_{\alpha}$ the compact $\alpha-$neighboorhood of $K$ and let us fix $\alpha$ small enough so that $K_{\alpha} \subset ]0, 2\pi[$. The function $\psi_{\rho}$ is uniformly continuous on $K_{\alpha}$ and for all $\varepsilon>0$, there exists $0<\eta \leq \alpha$ such that $| \psi_{\rho}(t-x) - \psi(t) | \leq \varepsilon$ as soon as $|x|<\eta$. We can then decompose the right hand side of Equation \eqref{eq.prime} as the sum
\[
\frac{1}{n} K_n' \ast \psi_{\rho}(t) = A_n(t) + B_n(t), \]
where
\[
A_n(t) := \frac{1}{2\pi} \int_{|x|<\eta} \frac{K_n'(x)}{n} \left[ \psi_{\rho}(t-x) - \psi(t) \right] dx
\]
and
\[
B_n(t):=\frac{1}{2\pi} \int_{\eta}^{2\pi-\eta} \frac{K_n'(x)}{n} \left[ \psi_{\rho}(t-x) - \psi(t) \right] dx.
\]
On the one hand, using the upper bound \eqref{eq.bernstein}, we have $|A_n(t)| \leq \varepsilon$ uniformly in $t \in K$. On the other hand, using this time the upper bound \eqref{eq.primeup}, we get that uniformly in $t$
\[
|B_n(t)| \leq \frac{R_{\eta}}{n} \times \left( ||\psi_{\rho}||_1 + \sup_{t \in K} |\psi_{\rho}(t)| \right),
\]
hence the result.
\end{proof}
\section{Asymptotics of the expected number of real zeros}
Thanks to the estimates for the variance and covariance established in the last section, we can now explicit the asymptotic behavior of the expected number of real roots of our random trigometric polynomial. We first consider the real zeros at a positive distance from the origin. Here $N_n(K)$ denotes the random number of real zeros of $F_n$ in a set $K$ whose volume i.e. Lebesgue measure is denoted by $\mathrm{vol}(K)$.
\begin{lem}\label{lem.bulk}
Let $K$ be a compact subset of $]0, 2\pi[$, under the hypotheses of Section 2, as $n$ goes to infinity, we have
\[
\mathbb E \left[ N_n(K) \right] = \frac{n}{\sqrt{3}\pi} \left( \mathrm{vol}(K) +o(1)\right).
\]
\end{lem}
\begin{proof}
The process $(F_n(t))_{t \geq 0}$ is a centered Gaussian process with $C^1-$paths. Besides, Remark
\ref{r:n-d} above implies that for each $t\in[0,2\pi]$ the distribution of $F_n(t)$ is non-degenerated.
Hence, we can use the celebrated Kac--Rice formula as in Theorem 3.2 in \cite{azaisW} to compute the expectation of $N_n$. So let $K$ be a compact subset of $]0, 2\pi[$,
the expected number of real zeros of $F_n$ in $K$ is then given by
\[
\mathbb E \left[ N_n(K) \right] =\frac{1}{\pi} \int_{K} \sqrt{I_n(t)}dt,
\]
where
\[
I_n(t):=\frac{\mathbb E[F_n(t)^2]\mathbb E[F_n'(t)^2]-\mathbb E[F_n(t)F_n'(t)]^{2} }{\mathbb E[F_n(t)^2]^2 }.
\]
From Lemma 1, we have
\begin{equation}\label{eq.In}
I_n(t)= \frac{(n+1)(2n+1)}{6}\frac{L_n \ast \psi_{\rho}(t)}{K_n \ast \psi_{\rho}(t) } - \frac{\mathbb E[F_n(t)F_n'(t)]^2 }{(K_n \ast \psi_{\rho}(t))^2},
\end{equation}
and combining Lemma \ref{lem.var} and \ref{lem.covar}, we get that uniformly on $K$, as $n$ goes to infinity
\[
I_n(t)= \frac{(n+1)(2n+1)}{6}( 1 +o(1)),
\]
so that we have indeed
\[
\mathbb E \left[ N_n(K) \right]=n \, \frac{\text{vol}(K) }{\sqrt{3}\pi} \left( 1 +o(1)\right).
\]
\end{proof}
The expected number of real zeros in the neighborhood of the origin is handled thanks to the following Lemma.
\begin{lem}\label{lem.bord}
Under the hypotheses of Section 2, there exists a finite constant $C$ such that, for $\varepsilon>0$ small enough and for all $n \geq 1$
\[
\frac{\mathbb E \left[ N_n([0,\varepsilon]) \right]}{n} \leq C \sqrt{\varepsilon}, \qquad \frac{\mathbb E \left[ N_n([2\pi-\varepsilon,2\pi]) \right]}{n} \leq C \sqrt{\varepsilon}.
\]
\end{lem}
\begin{proof}
Again, thanks to Kac--Rice formula, we have
\[
\mathbb E \left[ N_n([0,\varepsilon]) \right] = \frac{1}{\pi} \int_{0}^{\varepsilon} \sqrt{I_n(t)}dt.
\]
Using Cauchy--Schwarz inequality, we have then
\begin{equation}\label{eq.CS}
\mathbb E \left[ N_n([0,\varepsilon]) \right]^2 \leq \frac{1}{\pi^2} \left( \varepsilon \int_{0}^{\varepsilon} |I_n(t)|dt\right).
\end{equation}
Starting from the expression \eqref{eq.In} of $I_n(t)$, using the lower bound \eqref{eq.lbound} on the variance, we have for all $t \in [0, \varepsilon]$
\[
|I_n(t)| \leq \frac{(n+1)(2n+1)}{6 \gamma_{\rho}} L_n \ast \psi_{\rho}(t) ,
\]
so that
\[
\begin{array}{ll}
\displaystyle{ \int_{0}^{\varepsilon} |I_n(t)|dt } & \displaystyle{\leq \frac{(n+1)(2n+1)}{6 \gamma_{\rho}} \int_0^{2\pi} L_n \ast \psi_{\rho}(t) dt} \\
\\
& \leq \displaystyle{\frac{2\pi(n+1)(2n+1)}{6 \gamma_{\rho}} || L_n \ast \psi_{\rho} ||_1 }\\
\\
& \leq \displaystyle{\frac{\pi(n+1)(2n+1)}{3 \gamma_{\rho}} || L_n ||_1 \times || \psi_{\rho} ||_1. }
\end{array}
\]
Remembering that $|| L_n ||_1=1$ and injecting this last estimate in Equation \eqref{eq.CS}, we get that for all $n\geq 1$
\[
\frac{ \mathbb E \left[ N_n([0,\varepsilon]) \right]^2}{n^2} \leq \frac{(n+1)(2n+1)}{3\pi \gamma_{\rho}n^2} \varepsilon \leq C^2 \varepsilon, \;\; \text{where} \;\; C:=\sqrt{\frac{2 || \psi_{\rho} ||_1}{\pi \gamma_{\rho}}}.
\]
The proof of the analogue estimate on $[2\pi-\varepsilon,2\pi]$ is similar.
\end{proof}
We can finally combine Lemma \ref{lem.bulk} and Lemma \ref{lem.bord} to deduce the asymptotic behavior of the expected number of real roots on the whole interval $[0, 2\pi]$.
\begin{thm}Under the hypotheses of Section 2, as $n$ goes to infinity, we have
\[
\mathbb E \left[ N_n([0,2\pi]) \right] = \frac{2n}{\sqrt{3}} \left( 1 +o(1)\right).
\]
\end{thm}
\newcommand{\etalchar}[1]{$^{#1}$}
| 4273cae6e2717f71cae7f9aeab842e2cc7d5b72d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\subsection{Details on the time-resolved X-ray diffraction experiment}
The femtosecond X-ray experiment is performed on the FEMTO slicing source of the Swiss Light Source (PSI, Villigen, Switzerland)\cite{Beaud2007}. The femtosecond X-ray pulse used is superimposed on a $\sim50$~ps background pulse, which contributes to about 21$\%$ of the intensity. This picosecond background is systematically recorded looking at X-ray pulse from the single electron bunch arriving just before the sliced bunch (1 $\mu$s before). For the scans as a function of rotation angle $\varphi$, we also perform pump-probe scans using as a probe the x-ray pulse after a time when the femtosecond portion of has relaxed, leaving only the long-time background pulse. These scans are subtracted from the rotation scans in order to avoid any artifact arising from the picosecond background.
The sample was previously oriented at the Material Science beamline of the Swiss Light Source \cite{Willmott2013}, where the orientation matrix was computed using the Bragg reflections (202), (-202), (0-22), and the (2-42). During the time-resolved experiment the orientation matrix was verified using the reflections (202) and (2-q 0 2-q), where q = 0.341.
\subsection{Excitation density calculations}
The X-ray, the optical 800 nm pump, and the 400 nm probe do not have the same penetration depth, therefore instead of using the absorbed fluence we use the average excitation density. We divide the sample into n layers of thickness $dz$ starting at the surface with i ranging from 0 to n. The excitation density for the i$^{th}$ layer is:
\begin{equation}
n(i)=\frac{F(1-R)(e^{\frac{-i dz}{\alpha_L}}-e^{\frac{-(i+1) dz}{\alpha_L}})}{dz}
\end{equation}
where F is the incoming fluence, R is the reflectivity of the pump laser at the specific angle of the experiment (R = 0.06 for trXRD and R = 0.11 for MOKE \cite{Lee2002}), $\alpha_L$ is the penetration depth of the laser (23 nm for MOKE and 21 nm for trXRD). Here $F(1-R)$ is the absorbed fluence.\\
We then compute the average excitation density seen by the probe:
\begin{equation}
\left< n_0 \right>=\sum_{i=0}^{n}n(i)\frac{1}{\alpha_p} e^{-idz/\alpha_p}
\end{equation}
where $\alpha_p$ is the penetration depth of the probe (55 nm for trXRD and 16 nm for MOKE).
\subsection{Static magnetic measurement}
In order to verify that the sample magnetization is saturated for the pump-probe experiment, we perform static magnetic optical Kerr effect with no pump pulse. Fig. \ref{hyst} shows that the field of 0.7 T is enough to saturate completely the bulk sample. Therefore we have performed the pump-probe experiment using 0.7 T.
\begin{figure}
\includegraphics[angle=0,width=1\linewidth,clip=true]{hyst.pdf}
\caption{Magnetization of bulk sample using static magnetic optical Kerr effect.}
\label{hyst}
\end{figure}
\subsection{Temperature calculations}
Neglecting transport effects, we estimate the temperature after achieving local thermal equilibrium as :
\begin{equation}
T_f=T_i+\frac{\left< n_0 \right>}{C_p}
\end{equation}
Where $T_f$ and $T_i$ are the final and initial temperature, and $C_P$ is the sample's heat capacity (95 J K$^{-1}$ mol$^{-1}$ \cite{Uijttewaal2009}). The estimated final temperatures are mentioned in the main text. We neglect heat diffusion since it has a usual timescale of hundreds of picoseconds to nanoseconds, which is outside our measurement window.
\end{document} | 4506a958c6765d41322caa57d5679448bc7c38a1 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Exploratory Analysis}\label{sec:streetstyle_analysis}
A key goal of our work is to take automatic predictions from machine
learning and derive statistics from which we can find trends. Some of
these may be commonsense trends that validate the approach, while
others may be unexpected. This section describes several ways to
explore our fashion data, and presents insights one can derive through
this exploratory data analysis.
\subsection{Analyzing trends}
\noindent{\bf Color.} How does color vary over time?
Figure~\ref{fig:streetstyle_color} shows plots of the frequency of
appearance (across the entire world) of several colors over time.
White and black clothing, for example, exhibit a highly periodic
trend; white is at its maximum ($>20\%$ frequency) in September,
whereas black is nearly reversed, much more common in January.
However, if we break down these plots per city, we find that cities in
the Southern Hemisphere flip this pattern, suggesting a correlation
between season and clothing color.
Figure~\ref{fig:streetstyle_color_cities} shows such plots for two
cities, along with a visualization of all pairs of cities correlated
by their white color temporal profiles, with cities reordered by
latitude to highlight this seasonal effect. Cities in the same
hemisphere tend to be highly correlated with one another, while cities
in opposite hemispheres are highly negatively correlated. In the
United States, there is an oft-cited rule that one should not wear
white after the Labor Day holiday (early September). Does this rule
comport with actual observed behavior? Using our method, we can
extract meaningful data that supports the claim that there is a
significant decrease in white clothing that begins mid-September,
shortly after Labor Day, shown in
Figure~\ref{fig:streetstyle_labor_day}.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=\columnwidth]{streetstyle/figures/major_color_global.pdf}
\caption{{\bf Mean major color score for several different colors.} The means are
computed by aggregating measurements per 1 week intervals and error bars
indicate 95\%\ confidence interval. Colors such as brown and black are
popular in the winter whereas colors such as blue and white are popular in the
summer. Colors such as gray are trending up year over year whereas red, orange
and purple are trending down.\label{fig:streetstyle_color}}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{streetstyle/figures/major_color_cities.pdf}\\
\includegraphics[width=\columnwidth]{streetstyle/figures/white_corr.pdf}
\caption{Top: Mean major color frequency for the color white for two
cities highlighting flipped behavior between the Northern and
Southern Hemisphere. Means are computed by aggregating measurements
per 1 week intervals and error bars indicate a
95\%\ confidence interval. Bottom: Pearson's correlation
coefficients for the mean prediction score of the color white across
all pairs of major world cities. Cities have been ordered by
latitude to highlight the relationship between the northern and
southern hemispheres and their relationship to the seasonal trend of
the color white.}\label{fig:streetstyle_color_cities}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{streetstyle/figures/nyc_white.pdf}
\caption{{\bf Mean frequency for the color white in NYC.} Means are
computed by aggregating measurements per 1 week intervals and error
bars indicate a 95\%\ confidence interval. Dashed lines mark
Labor Day.}\label{fig:streetstyle_labor_day}
\end{center}
\end{figure}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{streetstyle/figures/china_red.pdf}
\caption{{\bf Mean frequency for the color red in China.} Means are
computed by aggregating measurements per 1 week intervals and error
bars indicate a 95\%\ confidence interval. Dashed lines mark
Chinese New Year.}\label{fig:streetstyle_chinese_new_year}
\end{center}
\end{figure}
Other colors exhibit interesting, but more subtle trends. The color
red is much less periodic, and appears to have been experiencing a
downward trend, down approximately 1\%\ since a few years
ago. There are also several spikes in frequency that appear each year,
including one small spike near the end of each October and a much
larger spike near the end of each December---that is, near Halloween
and Christmas. We examined our style clusters and found one that
contained images with red clothing and had spikes at the same two
times of the year. This cluster was not just red, but red hats. An
assortment of winter knit beanies and red jacket hoods were present,
but what stood out were a large assortment of Santa hats as well as an
unexpected assortment of red Halloween costumes with red hats or
hoods. It appears that red hats are uncommon enough that when they do
appear they are highly correlated with the exceptional sorts of
costumes worn on these two holidays. Finally, the Chinese New Year
sees a similar spike in the color red in 2015 and 2016, as shown in
Figure~\ref{fig:streetstyle_chinese_new_year} (frequency of red
clothing in China).
\begin{figure}[tp]
\begin{center}
\includegraphics[width=\columnwidth]{streetstyle/figures/usa_scarves.pdf}
\caption{{\bf Scarf trends in the USA.} Our Instagram-based
measurements are compared to Google Trends. Instagram measurements
are derived from mean scarf frequency aggregated by week with error
bars indicating a 95\%\ confidence interval. Google Trends
signal is per-week ``q=scarf, geo=US'' query volume normalized by
total query volume.\label{fig:streetstyle_scarf}}
\end{center}
\end{figure}
\smallskip \noindent{\bf Clothing types.} We can also explore more
functional clothing attributes. For example, we know people wear
scarves in cold weather, so one would expect to see a seasonal trend
in this attribute. Figure~\ref{fig:streetstyle_scarf} shows such a
trend for visual occurrence of scarves in the United States.
Moreover, we can also compare with a secondary source of information
to estimate this signal, Google
Trends.\footnote{\url{https://www.google.com/trends/explore\#q=scarf\&geo=US}}
Google Trends provides historical trend data for Google Search
queries. By comparing these two signals, we can make several
observations. First, the shapes of the signals are similar. Both
feature a sudden attack and decay rather than a purely sinusoidal
shape. This correlation increases our confidence that our vision-based
signal is measuring something reasonable. Second, the onset of the
signal derived from Google Trends tends to come prior to the onset
from the Instagram photos. One explanation for this behavior is that
people are searching for scarves on Google in order to purchase them
and Instagram allows us to measure when they are actually wearing
them. Third, while Google Trends suggests that scarves are in a
significant downward annual trend, the Instagram results serve to
constrast this trend (scarves are {\em up} in the winter of
2014-2015). While more study would be needed to determine the cause of
such a phenomenon, this comparison illustrates the utility of our
approach for devising new hypotheses for follow-up investigation. The
visual data and the Google search data combine to give us a richer
sense of what is happening in the world, and allow us to formulate new
views about the underlying human behavior.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=\columnwidth]{streetstyle/figures/wearing_jacket__Yes_map.png}
\caption{{\bf Mean jacket frequency per country.} Countries with $\ge$
1,000 photos included (other countries shown in
gray).}\label{fig:streetstyle_jackets}
\end{center}
\end{figure}
\begin{figure}[tp]
\begin{center}
\includegraphics[width=\columnwidth]{streetstyle/figures/wearing_hat__Yes_map.png}\\
\includegraphics[width=0.75\columnwidth]{streetstyle/figures/oman_hats.jpg}\\
\caption{Top: Mean hat frequency per country. Countries with $\ge$ 1,000 photos
included (other countries shown in gray). Bottom: Several top-ranked examples
of hats in Oman. While not strictly hats, it is clear that the classifier has
learned to identify more generally, head coverings.
}\label{fig:streetstyle_hats}
\end{center}
\end{figure}
So far we have focused on temporal trends, but we can also explore
geographic trends. Can we validate our attribute classifications
using prior knowledge of climate and weather-related clothing? For
example, where in the world do people wear jackets?
Figure~\ref{fig:streetstyle_jackets} shows the percentage of detected
people who are wearing jackets, broken down by country. This map
accords with intuition. Countries further north tend to feature more
jackets. In South America, more jackets are found as you go further
west (e.g., in Boliva or Colombia)---winter clothes will tend to be
worn more often at higher elevation (in this case in the Andes) than
at lower elevations.
What about more unexpected geographic trends? Where in the world are
people wearing hats? Figure~\ref{fig:streetstyle_hats} shows the
percentage of photos per country that contain people wearing hats.
Once again, people wear hats in colder places, but interestingly in
Oman, hats are evidently extremely popular.
Figure~\ref{fig:streetstyle_hats} shows several examples of images
from Oman that are highly ranked as wearing hats. In particular, the
kuma and massar are popular in Oman, as they are an important element
of the men's national dress.
\subsection{Visualizing styles}\label{sec:vis_styles}
We also wish to visualize commonly occuring combinations of
attributes, or ``styles.'' To visualize styles, we use the method in
Section~\ref{subsec:clusters} to compute style clusters from the
embedded photos. Figure~\ref{fig:streetstyle_gmm_clusters} shows
several such clusters, in no particular order. Given these clusters,
then for every photo in our dataset, we compute the style cluster
centroid to which it is closest. We can then perform additional
analyses using these style clusters, such as plotting the frequency of
visual occurrence of each cluster across space and time to mine for
styles that are well localized in one or both.
\begin{figure*}[tbp]
\begin{center}
\includegraphics[width=\textwidth]{streetstyle/figures/city_entropy_clusters.jpg}
\caption{{\bf Style clusters sorted by entropy across cities.}
Clusters that are more distinctive for given cities are ranked
higher. Each cluster contains thousands of people so we visually
summarize them by the people closest to the cluster center (left),
and the histogram of all people in the cluster across each city
(right).\label{fig:streetstyle_rank_cities}}
\end{center}
\end{figure*}
\begin{figure*}[p]
\centering
\begin{tabular}{c}
\includegraphics[width=0.95\textwidth]{streetstyle/figures/top_citymonth.jpg}\\
Top five clusters sorted by space-time entropy.\\ \\
\includegraphics[width=0.95\textwidth]{streetstyle/figures/bottom_citymonth.jpg}\\
Bottom five clusters sorted by space-time entropy.
\end{tabular}
\caption{{\bf Style clusters sorted by space-time entropy.} Each
cluster contains thousands of people so we visually summarize them
by the people closest to the cluster center (left), the average
image of the 200 people closest to the center (middle), and the
space-time histogram of all people in the cluster (right). The
histogram lists 36 cities along the X-axis and months along the
Y-axis. These histograms exhibit several types of patterns: vertical
stripes (e.g., rows 1 and 2) indicate that the cluster appears in
specific cities throughout the year. Horizontal stripes correspond
to specific months or seasons. (Note that summer and winter months
will be reversed in the Southern Hemisphere.) For example, the fifth
row is a winter fashion. A single bright spot can correspond to a
special event (e.g., the soccer jerseys evident in row 3).
}\label{fig:sorted_clusters}
\end{figure*}
This raw visualization of style clusters is not particularly easy to
explore, as there are hundreds of such clusters. Therefore, it is
useful to define a ranking function to help surface the most
interesting signals. One way to rank clusters is according to entropy
over the different cities in the dataset---high-entropy clusters will
tend to be specific to a few cities, while low-entropy clusters tend
to be universal. Figure~\ref{fig:streetstyle_rank_cities} shows the
top clusters for such a ranking. This ranking helps provide a sense
of how geography affects clothing: headscarves are common in, for
instance, Jakarta and Cairo, cities in Europe tend to have a higher
percentage of winter clothing than other countries, etc.
Figure~\ref{fig:sorted_clusters} shows a set of top and bottom ranked
clusters according to entropy computed across both city and month, as
well as an average image of people in each cluster. This method for
ordering clusters is very revealing:
\begin{packed_item}
\item Regional clothing is very evident as distinct vertical
structures in the histogram. The {\em gele} (Nigerian head-tie) is
very distinctive of Lagos, Nigeria.
\item Yellow sports jerseys are incredibly well-concentrated at a
specific time and place (Bogota, Colombia, during the World Cup).
\item Certain styles are common around the world and throughout the
year. For instance, blue collared shirts, plaid shirts, and black
t-shirts.
\end{packed_item}
\begin{figure}[h]
\begin{center}
\includegraphics[width=\columnwidth]{streetstyle/figures/cluster5.pdf}
\caption{High-ranking spatio-temporal cluster illustrating that yellow
t-shirts with graphic patterns were very popular for a very short
period of time, June-July 2014, in specifically BogotΓ‘, Karachi, Rio
de Janeiro, and SΓ£o Paulo.\label{fig:streetstyle_cluster4}}
\end{center}
\end{figure}
\begin{figure}[h]
\begin{center}
\includegraphics[width=\columnwidth]{streetstyle/figures/cluster7.pdf}
\caption{High-ranking spatio-temporal cluster illustrating that black polo
shirts with glasses are popular in Singapore and throughout the Indian
subcontinent.\label{fig:streetstyle_cluster6}}
\end{center}
\end{figure}
Figures~\ref{fig:streetstyle_cluster4} and
\ref{fig:streetstyle_cluster6} illustrate two other interesting top
ranked clusters. Figure~\ref{fig:streetstyle_cluster4} shows that
clusters can capture styles that become extremely popular at a very
specific time and place, and is a variant of the yellow jersey cluster
above. In this case, June and July 2014 capture the 2014 World Cup,
and the contents of these clusters contain many examples of soccer
fans wearing team colors. Finally,
Figure~\ref{fig:streetstyle_cluster6} shows that a style that one
might consider commonplace (black polo shirts with glasses) can be
much more popular throughout a specific region (in this case the
Indian subcontinent).
\section{Conclusions and Discussion}\label{sec:streetstyle_conclusion}
In this work we presented a framework for analyzing clothing, style,
and fashion across the world using millions of photos. Towards that
end, we developed {\sc{StreetStyle-27k}}\xspace, a large-scale, worldwide dataset of
people in the wild with clothing annotations, and used this dataset to
enable the analysis of a massive 15 million person photo corpus
through the application of CNNs. Our work illustrates the use of
machine learning and big data to perform visual discovery at scale.
\smallskip \noindent{\bf Limitations.} CNN-based embeddings are a very
powerful approach for organizing imagery in a high-dimensional feature
space. Furthermore, unsupervised methods such as GMMs can be used to
further explore this space. However, there is a limit to the
granularity of styles that we obtain. For instance, we do not learn a
clean separation between eyeglasses and sunglasses, as those were not
labeled as distinct attributes, and our style embedding does not
separate those two categories. One way to discover finer-grained
styles would be to incorporate active learning, where users are asked
to rank similarity of styles. Another would be to apply
weakly-supervised learning using the spatio-temporal metadata directly
in the embedding objective function rather than using it for post hoc
analysis. Finally, our method is limited to analyzing the upper body
of the person. As computer vision techniques for human pose
estimation mature, it would be useful to revisit this design to
normalize articulated pose for full body analysis.
\smallskip \noindent{\bf Future work.} There are many areas for future
work. We would like to more thoroughly study dataset bias. Our current
dataset measures a particular population, in a particular context
(Instagram users and the photos they take). We plan to extend to
additional image sources and compare trends across them, but also to
study aspects such as: can we distinguish between the (posed) subject
of a photo and people who may be incidentally in the background? Does
that make a difference in the measurements? Can we identify and take
into account differences between tourists and locals? Are there
certain types of people that tend to be missing from such data? Do our
face detection algorithms themselves exhibit bias? We would also like
to take other difficult areas of computer vision, such as pose
estimation, and characterize them in a way that makes them amenable to
measurement at scale.
Ultimately, we would like to apply automatic data exploration
algorithms to derive, explain, and analyze the significance of
insights from machine learning (e.g., see the ``Automatic
Statistician''~\cite{lloyd:aaai:2014}). It would be also be
interesting to combine visual data with other types of information,
such as temperature, weather, and textual information on social media
to explore new types of as-yet unseen connections. The combination of
big data, machine learning, computer vision, and automated analysis
algorithms, would make for a very powerful analysis tool more broadly
in visual discovery of fashion and many other areas.
\section{Data}\label{sec:streetstyle_data} Out dataset consists of three key
parts: (1) photos, (2) the people in those photos, and (3) the
clothing attributes of those people.
\smallskip \noindent{\bf Photos.} The photos we use in our analysis
were acquired via Instagram, a popular mobile photography social
network. We chose Instagram due to the sheer volume of uploads---an
advertised 95 million photos/videos per
day\footnote{\url{http://time.com/4375747/instagram-500-million-users/}}---as
well as their providing a public photo search
API.\footnote{\url{https://www.instagram.com/developer/}} Users can
query this API to retrieve images that have been uploaded within a 5
kilometer radius of a specified latitude and longitude and within 5
days of a specified date.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=\columnwidth]{streetstyle/figures/person_detections_map.pdf}
\caption{{\bf Geo-spatial distribution of images collected from Instagram}
(after filtering for images of people). Images were collected
according to two distributions: a set of the world cities (yellow
dots), and a distribution over the entire Earth (blue). The shade of
blue shows the number of photos downloaded from a particular region
(according to a logarithmic
scale).}\label{fig:streetstyle_distribution}
\end{center}
\end{figure}
\input{streetstyle/cities}
How should we sample photos geographically? To answer this question,
we consider two types of experiments we wish to conduct: (1) a
comparison of major world population centers,
and (2) a nation-, continent-, or world-wide analysis of styles and
trends. To sample photos for goal (1) (analysis of urban centers), we
made a list of 44 major cities spanning six continents, each with a
lat/long approximating the city center, shown as yellow dots in
Figure~\ref{fig:streetstyle_distribution}. The complete list of cities
is shown in Table~\ref{tab:streetstyle_cities}.
For each city, we sampled photos centered on those coordinates from
Instagram.
To sample photos for goal (2) (globally distributed photos), we used
the Flickr 100M photo dataset~\cite{thomee:arxiv:2015} to derive a
distribution of photo uploads over the globe. In particular, we
extracted the approximately 48 million geotags from Flickr 100M, and
used these geotags to compute a geographic distribution from which to
sample photos from Instagram. This distribution (after filtering by
person detection, as discussed below) is shown in blue in
Figure~\ref{fig:streetstyle_distribution}.
For both sampling strategies, we uniformly sampled a 5-day window from
June 2013---the earliest date Instagram provided results---until June
2016. For each photo, we record its geolocation and timestamp. In
total, we retrieved over 100 million photos.
\smallskip \noindent {\bf People.} To find people in photos, we ran
two out-of-the-box vision algorithms on each downloaded photo, one
to detect and localize faces, and the second to estimate the
visibility of the rest of the body. To detect faces, we used the API
provided by Face++.\footnote{\url{http://www.faceplusplus.com/}} The
output of the Face++ detector is a bounding box for the face, as well
as facial landmark keypoints. We use the face bounding box to
geometrically align people in our analysis algorithms. To determine
body visibility, we use the Deformable Part
Model~\cite{felzenszwalb:pami:2010,girshick:code:2012} and the
provided person detector trained on
VOC2010~\cite{everingham:site:voc2010}.
\begin{figure}[tp]
\begin{center}
\includegraphics[width=\columnwidth]{streetstyle/figures/crop_examples.jpg}
\caption{Left: Given an approximate person detection and a precise
face localization, we generate a crop with canonical position and
scaling to capture the head and torso of the person. Right: Example
people from our dataset.}\label{fig:streetstyle_crop}
\end{center}
\end{figure}
Given a set of faces and visible bodies in a photo, we pair them up
using a simple distance-based heuristic. For each face, we compute a
canonical image crop based on the position and scale of the detected
face, illustrated in Figure~\ref{fig:streetstyle_crop}. If the crop
extends beyond the visible body bounding box, we discard the
detection.
In total, out of the more than 100 million photos we retrieved from
Instagram, 37.7 million had at least one successful person
detection. For our work, we chose to keep detections with at least the
face and torso visible. We do not analyze the lower part of the body
since the legs are often occluded in online photos. After generating
canonical crops from the face detections and filtering on face and
torso visibility, we gathered a total of 14.5 million images of
people.
\begin{table}
\begin{center}
\resizebox{\columnwidth}{!}{\begin{tabular}{|c|c|c|} \hline & {\bf No}
& {\bf Yes} \\
\hline Wearing Jacket & 18078 & 7113 \\ Collar Presence & 16774 & 7299 \\
Wearing Scarf & 23979 & 1452\\ Wearing Necktie & 24843 & 827\\ Wearing Hat &
23279 & 2255\\ Wearing Glasses & 22058 & 3401\\ Multiple Layers & 15921 & 8829\\
&&\\ &&\\ &&\\ &&\\ &&\\ &&\\ &&\\ \hline
\end{tabular}
{\hskip 0.1cm}
\begin{tabular}{|c|}
\hline {\bf Major} \\ {\bf Color} \\ \hline Black (6545) \\ White (4461) \\ 2+ colors (2439)
\\ Blue (2419) \\ Gray (1345)\\ Red (1131)\\ Pink (649)\\ Green (526)\\ Yellow
(441)\\ Brown (386)\\ Purple (170)\\ Orange (162)\\ Cyan (33)\\ \hline
\end{tabular}
\begin{tabular}{|c|}
\hline {\bf Clothing} \\ {\bf Category} \\ \hline Shirt (4666) \\ Outerwear (4580) \\
T-shirt (4580) \\ Dress (2558) \\ Tank top (1348) \\ Suit (1143) \\ Sweater
(874) \\ \hline \hline {\bf Sleeve} \\ {\bf Length} \\ \hline Long sleeve (13410) \\ Short
sleeve (7145) \\ No sleeve (3520) \\ \\ \hline
\end{tabular}
\begin{tabular}{|c|}
\hline {\bf Neckline} \\ {\bf Shape} \\ \hline Round (9799) \\ Folded (8119) \\ V-shape
(2017) \\ \hline \hline {\bf Clothing} \\ {\bf Pattern} \\ \hline Solid (15933) \\ Graphics
(3832) \\ Striped (1069) \\ Floral (885) \\ Plaid (532) \\ Spotted (241)\\ \\ \\
\hline
\end{tabular}
}
\caption{{\bf Fashion attributes in our dataset}, along with the number of
MTurk annotations collected for each attribute. Left: Binary
attributes. Right: Attributes with three or more
values.}\label{table:streetstyle_attributes}
\end{center}
\end{table}
\smallskip \noindent{\bf Clothing annotations.} For each person in our
corpus of images, we want to extract fashion information. To do so, we
first collect annotations related to fashion and style for a subset of
the data, which we later use to learn attribute classifiers that can
be applied to the entire corpus. Example attributes include: {\em What
kind of pattern appears on this shirt? What is the neckline? What
is the pattern? Is this person wearing a hat?} We created a list
of several clothing attributes we wished to study a priori, such as
clothing type, pattern, and color. We also surveyed prior work on
clothing attribute recognition and created a consolidated list of
attributes that we observed to be frequent and identifiable in our
imagery. We began with several of the attributes presented by Chen
{\em et~al.}~\cite{chen:eccv:2012} and made a few adjustments to more clearly
articulate neckline type as in the work of Di
{\em et~al.}~\cite{di:cvprw:2013}. Finally, we added a few attributes of our
own, including \emph{wearing-hat} and \emph{wearing-jacket}, as well
as an attribute indicating whether or not more than one layer of
clothing is visible (e.g., a suit jacket over a shirt or an open
jacket showing a t-shirt
underneath). Table~\ref{table:streetstyle_attributes} shows our full
list of attributes.
We annotated a 27K-person subset of our person image corpus with these
attributes using Amazon Mechanical Turk. Each Mechanical Turk user
was given a list of 40 photos, presented one-by-one, and asked to
select one of several predefined attribute labels. In addition, the
user could indicate that the person detection was faulty in one of
several ways, such as containing more than one person or containing no
person at all. Each set of 40 photos contained two photos with known
ground truth label as sentinels. If a worker failed at least five
sentinels and their sentinel failure rate was above 20\%, we
prevented them from performing any more work for our tasks.
Compensation for a single attribute on a set of 40 photos was 0.10
USD. Average wage was approximately 4 USD / hour. An attribute for
an image was labeled by five separate workers, and a different set of
five workers labeled each attribute for an image. A label was
accepted if 3 / 5 workers agreed, otherwise it was excluded from the
data. In total these annotations cost approximately 4K USD. We refer
to our annotated dataset as {\sc{StreetStyle-27k}}\xspace.
\smallskip \noindent{\bf Dataset bias and limitations.} Any dataset
has inherent bias, and it is important to be cognizant of these biases
when drawing conclusions from the data. Instagram is a mobile social
network, so participants are more likely to be in regions where
broadband mobile Internet is available, and where access to Instagram
is not censored or blocked. In some areas, people might be more
likely to upload to a competitor service than to Instagram. People of
certain ages might be more or less likely to upload to any social
network. Cultural norms can affect who, when, where, and what people
photograph. The face detector we use is an off-the-shelf component
for which no reported statistics regarding bias for age, gender, and
race are made available. The person detector has also not been
evaluated to determine bias in these factors either. These factors
impact who will and will not be properly added to our dataset, which
could introduce bias. As vision methods mature, analyses for such bias
will become increasingly important.
\section{Introduction} \label{sec:streetstyle_intro}
Our age of big data presents us with the compelling opportunity to use
all available information to measure the world in ways that were never
before possible. Large amounts of data---for instance, OCRed scans of
centuries worth of books---coupled with tools for exploring this data
have yielded powerful new mechanisms for scientists to study our
history, culture, and behavior~\cite{michel:science:2010}. This
opportunity is magnified by the massive scale at which humans are
generating cultural artifacts on social media, and by the increasing
power of machine learning techniques. For instance, by applying
natural language processing to millions of Twitter messages, we can
discover relationships between time of day and mood that leverage
sample sizes much larger than those of any traditional
study~\cite{golder:science:2011}.
To date, most of these new kinds of big data analyses have been
limited to structured data, such as user interactions on social
networks, or to textual data, such as books and tweets. However, a
tremendous cache of unstructured visual information about our world is
locked in images, particularly in images of people, including the
billions of photos uploaded to photo-sharing services each
day. Imagine a future anthropologist with access to trillions of
photos of people---taken over centuries and across the world---and
equipped with effective tools for analyzing these photos to derive
insights. What kinds of new questions can be answered? This problem
area of {\em data-driven visual discovery} is still new, but is
beginning to gain attention in computer vision and
graphics~\cite{doersch:siggraph:2012,wang:iccvw:2013,zhu:siggraph:2014,ginosar:iccvw:2015,gebru:arxiv:2017}.
Our work takes a step towards this vision by analyzing geo-spatial
trends in fashion style across tens of millions of images from social
media.
In this paper, we focus on clothing, as it is a critical aspect of the
visual world and of our daily lives. Individuals make fashion choices
based on many factors, including geography, weather, culture, and
personal preference. The ability to analyze and predict trends in
fashion is valuable for many applications, including analytics for
fashion designers, retailers, advertisers, and manufacturers. This
kind of analysis is currently done manually by analysts by, for
instance, collecting and inspecting photographs from relevant
locations and times.
We aim to extract meaningful insights about the geo-spatial and
temporal distributions of clothing, fashion, and style around the
world through social media analysis at scale. We do so through the
combination of (1) millions of photos spanning the world---photos
uploaded to social media services by everyday users, (2) a new
dataset, {\sc{StreetStyle-27k}}\xspace, consisting of a subset of these images
annotated with fashion attributes, and (3) powerful machine learning
methods based on deep learning that leverage our dataset. We explore
two kinds of machine learning methods: supervised learning methods
that are trained on our annotated dataset to predict clothing
attributes in new images, followed by unsupervised clustering methods
that can automatically detect visual correlations in our data (such as
particular types of headwear, as in Figure~\ref{fig:teaser}(b), top
row).
These machine learning methods allow us to measure clothing features
across millions of photos, and then to use these measurements to
produce analyses in the form of trend reports and map-based
visualizations, enabling new types of visual insight. For instance,
our framework can help answer questions such as:
\begin{packed_item}
\item How is the frequency of scarf use in the US changing over time?
(Figure~\ref{fig:streetstyle_scarf})
\item What styles are most specific to particular regions of the world
or time of the year? Conversely, which styles are popular across the
world? (Figures~\ref{fig:streetstyle_rank_cities} and
\ref{fig:sorted_clusters})
\item For a given city, such as Los Angeles, what styles are most
characteristic of that city (popular in LA, but rare elsewhere)?
(Figures~\ref{fig:teaser}(c) and \ref{fig:streetstyle_rank_cities}.
\end{packed_item}
Our approach also demonstrates the utility of machine learning methods
for vastly simplifying the process of making real-world measurements
from large-scale visual data.
In summary, our work makes the following contributions:
\begin{packed_enum}
\item {\sc{StreetStyle-27k}}\xspace, an annotated dataset of people containing 27K
images, each with 12 clothing attributes, to be made publicly
available,
\item a methodology for analyzing millions of photos to produce a
visual clothing embedding (shown in Figure~\ref{fig:teaser}(a)),
then using this embedding to predict clothing attributes in new
photos,
\item the use of unsupervised clustering methods to automatically
predict visual correlations between clothing attributes (in the form
of {\em style clusters}, Figure~\ref{fig:teaser}(b)), and
\item a first-of-its-kind analysis of global and per-city fashion
choices and trends using our machine learning methods.
\end{packed_enum}
We provide additional visualizations of our results at
\url{http://streetstyle.cs.cornell.edu}.
\section{Machine Learning Methodology}\label{sec:streetstyle_method}
To make measurements across the millions of photos in our dataset, we
wish to generalize labeled data from {\sc{StreetStyle-27k}}\xspace to the entire
corpus. For instance, to measure instances of glasses across the
world, we want to use all of the {\em wearing-glasses} annotations as
training data and predict the {\em wearing-glasses} attribute in
millions of unlabeled images. To do so we use Convolutional Neural
Networks (CNNs)~\cite{krizhevsky:nips:2012}. A CNN maps an input
(e.g., the pixels of an image), to an output (e.g., a binary label
representing whether or not the person is wearing glasses), by
automatically learning a rich hierarchy of internal image features
given training data. In this section, we describe the architecture of
our particular CNN, how we trained it using our labeled attribute
dataset, and how we evaluated the classifier's performance on the
clothing attribute prediction task. We then discuss how and when it
is appropriate to generate predictions from this CNN across millions
of photos and use these predictions to estimate real-world fashion
statistics.
\subsection{Learning attributes} A key design decision in creating a CNN is
defining its network architecture, or how the various types of
internal image processing layers are composed. The penultimate layer
of a typical CNN outputs a multi-dimensional (e.g., 1024-D) feature
vector, which is followed by a ``linear layer'' (matrix product)
operation that outputs one or more prediction scores (e.g., values
indicating the probability of a set of attributes of interest, such as
{\em wearing-glasses}). The base architecture for our CNN is the
``GoogLeNet'' architecture~\cite{szegedy:cvpr:2015}. While several
excellent alternatives exist (such as VGG~\cite{simonyan:arxiv:2014}),
GoogLeNet offers a good tradeoff between accuracy on tasks such as
image classification~\cite{russakovsky:ijcv:2015} and speed. As with
other moderate-sized datasets (fewer than 1 million training
examples), it is difficult to train a CNN on our attribute prediction
task from scratch without overfitting. Instead, we start with a CNN
pretrained on image classification (on ImageNet ILSVRC2012) and
fine-tune the CNN on our data. After ImageNet training, we discard
the last linear layer of the network to expose the 1024-dimensional
feature output and then append several $1024\times N_i$ linear layers
in parallel where $N_i$ is the number of class labels for attribute
$i$ (for instance, $N_i = 2$ for a binary attribute such as {\em
wearing-glasses}, and $N_i = 3$ for neckline shape as we consider
three possible shapes).
\smallskip \noindent{\bf Training details.} We train the CNN as
follows. For each attribute, we use stratified sampling to first
select an attribute label and then a particular example image with
that label. This sampling strategy counteracts the implicit imbalance
within each attribute (e.g., wearing a hat is much less common than
not wearing a hat). We do this 32 times to build a single mini-batch
of images. Using this mini-batch, we apply the forward and backward
passes of the CNN to compute the parameter gradient with respect to
the cross-entropy loss. We do this for each attribute, accumulating
the gradients. Then we use stochastic gradient descent with momentum
$= 0.9$, learning rate $= 10^{-2}$, and weight decay $= 10^{-4}$ to
update the parameters of the CNN. We fine-tune the CNN for 6,000
iterations ($\times 12$ attributes) after which the mean class
accuracy on a validation set stopped increasing. 80\%\ of
{\sc{StreetStyle-27k}}\xspace\ was used for training, 10\%\ was used as
validation to determine when to stop training as well as for the
additional analysis in this section, and 10\%\ was used as a
test set for a final evaluation of the attribute classification task.
\begin{table}[t]
\begin{center}
\resizebox{\columnwidth}{!}{\begin{tabular}{|c|c|c|c|} \hline {\bf
Attribute} &
\begin{tabular}{@{}c@{}}{\bf Our CNN} \\(acc. / mean class. acc.)\end{tabular} &
{\bf Random guessing}& {\bf Majority guessing} \\ \hline \hline wearing jacket & 0.869 /
0.848 &0.5 / 0.5 & 0.698 / 0.5 \\ clothing category & 0.661 / 0.627 &0.142 /
0.142 & 0.241 / 0.142 \\ sleeve length &0.794 / 0.788 & 0.333 / 0.333 & 0.573 /
0.333 \\ neckline shape & 0.831 / 0.766 &0.333 / 0.333 & 0.481 / 0.333 \\ collar
presence & 0.869 / 0.868 & 0.5 / 0.5 & 0.701 / 0.5 \\ wearing scarf & 0.944 /
0.772 & 0.5 / 0.5 & 0.926 / 0.5 \\ wearing necktie & 0.979 / 0.826 & 0.5 / 0.5 &
0.956 / 0.5\\ clothing pattern & 0.853 / 0.772 & 0.166 / 0.166 & 0.717 / 0.166
\\ major color & 0.688 / 0.568 & 0.077 / 0.077 & 0.197 / 0.077 \\ wearing hat &
0.959 / 0.917 & 0.5 / 0.5 & 0.904 / 0.5\\ wearing glasses & 0.982 / 0.945 & 0.5
/ 0.5 & 0.863 / 0.5 \\ multiple layers & 0.830 / 0.823 & 0.5 / 0.5 & 0.619 /
0.5\\ \hline
\end{tabular}}
\caption{{\bf Test set performance for the trained CNN.} Since each
attribute has an unbalanced set of test examples, we report both
accuracy and mean classification accuracy. For reference, we also
include baseline performance scores for random guessing and majority
class guessing.}\label{table:streetstyle_accuracy}
\end{center}
\end{table}
\begin{figure}[t]
\begin{center}
\includegraphics[width=\columnwidth]{streetstyle/figures/confusion.pdf}
\caption{{\bf Confusion matrices for each clothing attribute.} Given class
labels are vertical and predicted class labels are horizontal. Each
row has been normalized such that a cell indicates the probability
of a classification given some true
label.\label{fig:streetstyle_confusion}}
\end{center}
\end{figure}
Table~\ref{table:streetstyle_accuracy} summarizes the accuracy and
mean class accuracy of our learned classifiers for each attribute on
the held-out test set. Figure~\ref{fig:streetstyle_confusion} shows a
set of confusion matrices, one per attribute, illustrating which
attributes tend to get confused for one another (e.g., spotted shirts
can be mistaken for floral shirts). While the classifiers perform
quite well in general (e.g., {\em wearing-hat} has an accuracy over
90\%), even state-of-the-art CNNs make mistakes, and so all of
the classifiers have an inherent error level. We now discuss how we
take this error into account when using these classifiers.
\subsection{Measuring attributes at scale} Our aim is not just to build a
classifier for each clothing attribute and maximize some performance
criterion on a test set, but to take that learned classifier, apply it
to a much larger corpus of images spanning years worth of the world's
photos, and provide a tool for discovering interesting trends.
However, the classifiers are not perfect, and any real-world
measurement is going to have some noise. While it is inevitable that
a classifier will have some non-zero variance, the hope is that the
estimation is unbiased, that is, as we increase the sample size, a
computed statistic, such as the detected percentage of people wearing
hats in photos, will converge on the true percentage of people wearing
hats. However, by examining the confusion matrices in
Figure~\ref{fig:streetstyle_confusion} it becomes clear that this will
not be the case. For example, the proportion of images featuring
people wearing neckties is much less than 50\%. For this
attribute, the classifier predicts \textbf{No} correctly
99\%\ of the time, but \textbf{Yes} correctly only
65\%\ of the time. Therefore, the estimate of the percentage of
people wearing neckties in photos will be biased towards \textbf{No}.
\begin{figure*}[h]
\begin{center}
\includegraphics[width=\textwidth]{streetstyle/figures/calibration.pdf}
\caption{{\bf Reliability curves for several clothing attributes.} A
well-calibrated reliability curve has fraction of positives equal to
mean predicted value. Neural networks tend to produce
well-calibrated probabilities, but significant improvement is made
using Isotonic regression for additional
calibration.\label{fig:streetstyle_reliability}}
\end{center}
\end{figure*}
One way to mitigate this bias is to calibrate the scores of a
classifier using a function that models the posterior probability of
the classification being correct given the classification score.
There are several methods for this type of calibration, including
Platt scaling (an application of logistic
regression)~\cite{platt:book:1999} and isotonic regression. Others
have noted that neural networks tend not to require calibration as
severely as classifiers such as SVMs and boosted decision
trees~\cite{niculescu:icml:2005}. Nevertheless, we found that our
networks benefited from calibration by applying isotonic regression to
our validation set. Figure~\ref{fig:streetstyle_reliability} shows
the generalization performance of calibration for several attribute
labels. Note that after isotonic regression, the curves are very
close to the identify function ($y=x$), which means that these curves
are properly calibrated.
In our experiments, half of the validation set was used for training the
regressor and the other half was used to generate reliability curves.
Isotonic regression typically requires more data to avoid overfitting,
yet is more flexible when the calibration required is not sigmoidal,
an underlying assumption of Platt scaling. The generalization shows
that these estimates do not exhibit severe overfitting and therefore
we opt to use isotonic regression trained on the entire validation set
to calibrate the scores.
\subsection{Visually consistent style clusters} \label{subsec:clusters}
\begin{figure}[tp]
\begin{center}
\includegraphics[width=\columnwidth]{streetstyle/figures/tsne.pdf}
\caption{{\bf t-SNE visualization of our learned fashion feature
space.} The style embedding is 1024-dimensional and is visualized
in 2D here. Several combinations of attributes are highlighted.
For example, white shirts with text, plaid shirts, black v-neck
shirts, etc., are styles discoverable in this embedding. Zoom to see
detail.} \label{fig:tsne}
\end{center}
\end{figure}
\begin{figure*}[tp]
\begin{center}
\includegraphics[width=\textwidth]{streetstyle/figures/gmm_clusters.jpg}
\caption{{\bf Visualization of 12 style clusters}, showing images
closest to the centroid of each cluster. Clusters are shown in no
particular order. Each style cluster reveals a visual theme as a
combination of attributes, such as ``white graphic tees and
glasses'' (top-left cluster). \label{fig:streetstyle_gmm_clusters}}
\end{center}
\end{figure*}
The predicted attributes are immediately useful in plotting and
analyzing trends, such as ``how is the proportion of people wearing
black in Los Angeles changing over time?'' as we demonstrate in
Section~\ref{sec:streetstyle_analysis}. However, we would like to
analyze clothing styles beyond these attributes in a number of ways:
\begin{packed_item}
\item Identify common, visually correlated combinations of these basic
attributes (e.g., {\em blue sweater with jacket and wool hat}).
\item Identify styles that appear more frequently in one city versus
another or more frequently during particular periods of time.
\item Identify finer-grained, visually coherent versions of these
elements (e.g., sports jerseys in a particular style).
\end{packed_item}
To achieve these goals, we use clustering to identify recurring {\em
visual themes} in the embedded space of people. Recall that the
penultimate layer of our network is a 1024-dimensional feature space
where distinct fashion attributes are linearly
separable. Figure~\ref{fig:tsne} shows a visualization of this feature
space projected to 2D using t-SNE~\cite{maaten:jmlr:2008}. Within this
1024-D feature space, people images are organized according to visual
fashion attributes, through the mechanism of training on our
{\sc{StreetStyle-27k}}\xspace dataset. The ideas is that by clustering images in this
1024-D embedded feature space, we can discover and reveal visual
themes such as those described above. We refer to such clusters as
{\bf style clusters}.
Once we identify style clusters, we can further characterize cities
and times in terms of these clusters---for instance, we might discover
that some clusters are heavily correlated with one or two particular
cities,
as we explore in Section~\ref{sec:streetstyle_analysis}.
Figure~\ref{fig:streetstyle_gmm_clusters} shows an example of several
style clusters found by our approach.
To find style clusters, we ran a clustering algorithm on a subset of
the full dataset, for efficiency and to achieve balance between
different times and regions of the world. In particular, we divided
our images into bins by city and by week (e.g., Paris in week 26 of
2015). For bins with fewer than $N$ images, we selected all images,
and for bins with more than $N$ images, we randomly selected $N$
images. For our experiments, we set $N = 4000$, for a total of 5.4M
sample images for clustering in total.
For each cropped person image in this set, we compute its 1024-D CNN
feature vector and $L_2$ normalize this vector. We run PCA on these
normalized vectors, and project onto the top principal components that
retain 90\%\ of the variance in the vectors (in our case, 165
dimensions). To cluster these vectors, we used a Gaussian mixture
model (GMM) of 400 components with diagonal covariance matrices. Each
person is assigned to the mixture component (style cluster) which
maximizes the posterior probability. The people assigned to a cluster
are then sorted by their Euclidean distance from the cluster center,
as depicted in Figure~\ref{fig:streetstyle_gmm_clusters}. Although
more clusters would increase the likelihood of the data under the
model, it would also lead to smaller clusters, and so we selected 400
as a compromise between retaining large clusters and maximizing the
likelihood of the data.
The resulting 400 style clusters have sizes ranging from 1K to
115K. Each style cluster tends to represent some combination of our
chosen attributes, and together form a global visual vocabulary for
fashion. We assign further meaning to these clusters by ranking them
and ranking the people within each cluster, as we demonstrate in
Section~\ref{sec:vis_styles}, with many such ranked clusters shown in
Figures~\ref{fig:streetstyle_rank_cities} and
\ref{fig:sorted_clusters}.
\section{Related Work}\label{sec:streetstyle_related}
\noindent{\bf Visual discovery.} Researchers are beginning to mine
world-wide imagery to (1) discover visual trends and (2) measure
aspects of the world over space and time. Much of this work has looked
at {\em places}. Doersch~{\em et~al.}\ pioneered the idea of {\em
computational geography} and explored the distinctive visual
characteristics of cities through Street View
imagery~\shortcite{doersch:siggraph:2012}, driven by weakly supervised
methods for finding discriminative
elements~\cite{singh:eccv:2012,doersch:nips:2013}. Such visual
characteristics can also be correlated to other properties of cities,
such as perceived
safety~\cite{arietta:tvcg:2014,naik:cvprw:2014,dubey:eccv;2016}. These
correlations can then be used to automatically predict such properties
across a city simply from ground level images. Beyond analyses of
cities, computer vision techniques have been used to answer questions
such as ``is there snow in this photo?'' or ``are there clouds in this
webcam?'' which in turn can be used at scale to estimate maps of snow
or cloud cover over large regions of the
world~\cite{zhang:www:2012,murdock:iccv:2015}.
Other work has used large image collections to study people, including
explorations of how people move through
cities~\cite{crandall:www:2009}, the relationship between facial
appearance and
geolocation/time~\cite{islam:eurosip:2015,salem:wacv:2016}, analysis
of expressions and styles over a century in high school yearbook
photos~\cite{ginosar:iccvw:2015}, and tools for discovering visual
patterns that distinguish two
populations~\cite{matzen:iccv:2015}. Finally, Gebru {\em et~al.}\ study
demographics across the US by detecting and analyzing cars (along with
their makes and models) in Street View
imagery~\shortcite{gebru:arxiv:2017}. Compared to this prior work, we
focus on a different domain, fashion and style, and apply our work to
a much broader sample of people by using worldwide images on social
media.
\smallskip \noindent{\bf Visual understanding of clothing.} Clothing
is a rich, complex visual domain from the standpoint of computer
vision, because clothing analysis combines a detailed attribute-level
understanding with social context. We build on previous work that
describes people in terms of clothing and other fashion attributes,
such as ``wearing glasses,'' or ``short sleeves,'' and recognizes
these attributes in new
images~\cite{chen:eccv:2012,bourdev:iccv:2011,bossard:accv:2013,zhang:cvpr:2014}.
Beyond classifying attributes, other work also produces full
pixel-level clothing
segmentations~\cite{yamaguchi:cvpr:2012,yamaguchi:iccv:2013,yang:cvpr:2014},
or recognizes specific (or similar) products rather than general
attributes~\cite{di:cvprw:2013,vittayakorn:wacv:2015,kiapour:iccv:2015}.
Other work categorizes clothing in terms of explicit, named
styles. For instance, Kiapor~{\em et~al.}\ used a game to analyze coarse
fashion styles such as ``hipster'' and ``goth'' to build inter-style
classifiers (i.e., is this photo goth or hipster?) and intra-style
ranking functions (i.e., how hipster is this person's
fashion?)~\shortcite{kiapour:eccv:2014}. In our work, we train
classifiers using coarse-grained attributes, but then leverage this
training to organize images visually in order to perform more
fine-grained clustering.
\smallskip \noindent{\bf Clothing trends.} Spatial and temporal
fashion trends have been explored in computer vision, but usally on
small samples sizes and with limited statistical analysis.
Hidayati~{\em et~al.}\ analyze catwalk images from NYC fashion shows to find
style trends in high-end fashion~\shortcite{hidayati:acmmm:2014}.
Simo-Serra~{\em et~al.}\ analyzed \url{chictopia.com} to study correlations
between fashionability and other attributes such as wealth, as well as
temporal trends (e.g., a sudden spike in popularity of heels in
Manila)~\shortcite{simoserra:cvpr:2015}. \cite{vittayakorn:wacv:2015}
showed that seasonal trends related to styles such as ``floral'',
``pastel'', and ``neon'' can be found using fashion classifiers, with
the peak occurring in the springtime. \cite{he:www:2016} modeled
per-user fashion taste over time and found certain styles had a
resurgence in the late 2000s. However, these trends were not
evaluated for statistical significance, and so it is challenging to
conclusively distinguish signal from noise. We argue that without the
massive amounts of data we advocate, it is difficult to establish such
significance.
\smallskip \noindent{\bf Clothing datasets.} Most previous clothing style
datasets have been very limited in scale, and biased towards images of the
fashion-conscious. For example, several efforts have made use of
fashion-centric social networking sites such as \url{fashionista.com} and
\url{chictopia.com}~\cite{yang:cvpr:2014,simoserra:cvpr:2015,yamaguchi:bvmc:2015,yamaguchi:cvpr:2012,yamaguchi:iccv:2013}.
These sites enable users to upload photos and annotate articles of clothing for
the express purpose of modeling their style. Other work draws on online clothing
retailers such as Amazon, eBay, and ModShop to build similar
datasets~\cite{liu:acmmm:2012,di:cvprw:2013,kiapour:iccv:2015}. Images from
these websites are relatively clean and organized according to tags,
facilitating the creation of annotated datasets on the scale of hundreds of
thousands of people~\cite{chen:eccv:2012,bossard:accv:2013}. However, our goal
is to measure spatio-temporal trends over the entire world from real-world
images, and to obtain highly certain statistics, so very large data is
key---100K images is insufficient, once the data is sliced in time and space.
Furthermore, fashion sites are designed to engage the fashion-conscious, whereas
our goal is to analyze the world's populace at large. Our approach is to build
a massive dataset from photos on social media, which can be gathered at larger
scale and are more representative of everyday fashions (hence the name {\it
StreetStyle}). However, these images are also much more noisy and are often not
tagged according to clothing. Therefore, we annotate a small subset of our large
dataset manually, and use machine learning to generalize the result to the rest
of the dataset.
| 2547094e266ea6626efbc3d3414b2202d2b225a9 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Recent interest in layered ferromagnetic materials is driven by the desire to develop functional van der Waals heterostructures \cite{Geim-2013, Zhong-2017}. For this purpose materials must be cleavable down to very thin, ideally monolayer, specimens. This is possible when the layers composing the crystal are held together by weak van der Waals bonds so they can be mechanically separated or exfoliated by chemical intercalation routes \cite{Nicolosi-2013}. This led to the identification and study of CrSiTe$_3$ and CrGeTe$_3$ \cite{Lebegue-2013, Li-2014, Casto-2015, Sivadas-2015}. These are small band gap semiconductors with Curie temperatures of 33\,K for CrSiTe$_3$ \cite{Carteaux-1991} and 61 K for CrGeTe$_3$ \cite{Carteaux-1995}. The cleavage energy of these materials is calculated to be 0.35-0.38 J/m$^2$ \cite{Li-2014}, similar to graphite (0.43 J/m$^2$) and MoS$_2$ (0.27 J/m$^2$) \cite{Bjorkman-2012}. Studies of nanosheets of CrSiTe$_3$ suggest ferromagnetism may persist in ultrathin specimens \cite{Lin-2016}, and ferromagnetic few-layer-thick crystals of CrGeTe$_3$ have recently been reported \cite{Gong-2017}. The tin analogue CrSnTe$_3$ is also predicted to be a ferromagnet \cite{Zhuang-2015}. Other recently discovered materials include Fe$_3$GeTe$_2$ \cite{Deiseroth-2006}, a metal with itinerant ferromagnetism below $T_C$=220-230 K \cite{Deiseroth-2006, Chen-2013, May-2016}. Chromium triiodide, and transition metal halides in general \cite{McGuire-Review}, have also been put forth as candidates for cleavable magnetic materials \cite{McGuire-2015, Zhang-2015, Liu-2016, Wang-2016}. Experimentally, CrI$_3$ is a ferromagnet with $T_C$\,=\,61\, K \cite{Hansen-1959, Dillon-1965, McGuire-2015}, CrBr$_3$ is a ferromagnet with $T_C$ = 37 K \cite{Tsubokawa-1960}, and CrCl$_3$ is an antiferromagnet with an ordering temperature near 17 K \cite{Hansen-1958, Cable-1961}. Several theoretical studies have been recently published on chromium trihalides addressing bulk magnetic properties and behavior in monolayer form \cite{Wang-2011, McGuire-2015, Zhang-2015, Liu-2016, Wang-2016, Lado-2017}, and also on VCl$_3$ and VI$_3$, which are predicted to be ferromagnetic and Dirac half-metals \cite{He-2016}. All of the calculations predict low cleavage energies for these compounds, similar to those noted above for graphite and MoS$_2$. Recently ferromagnetism has been experimentally confirmed in monolayer CrI$_3$ with intriguing thickness dependent magnetic phases \cite{Huang-2017}, and heterostructures incorporating ultrathin CrI$_3$ have enabled remarkable control of the spin and valley pseudospin properties in monolayer WSe$_2$ through a large exchange field effect \cite{Zhong-2017}.
It is interesting to note that the magnetic ordering temperatures of the chromium trihalides increase as the halogen size increases from Cl to Br to I. Since the Cr-Cr distances increase with increasing halogen size, the direct exchange is expected to be weakened along this series. This indicates that superexchange, which is expected to favor ferromagnetic alignment, is the more important magnetic interaction \cite{Lado-2017}. As the electronegativity is decreased from Cl to Br to I, it is expected that the Cr-halogen bonding becomes more covalent, strengthening superexchange interactions and raising ordering temperatures. Indeed all three of these chromium trihalides exhibit ferromagnetic ordering of Cr moments within the layers at low temperature \cite{deHaas-1940, Cable-1961, Tsubokawa-1960, McGuire-2015}. Moving from Cl to Br to I is also expected to increase spin orbit coupling associated with the halogen, which has been identified as a source of magnetic anisotropy in these materials \cite{Lado-2017}.
The present work focuses on chromium trichloride. Initial studies of magnetism in this material date back nearly 100 years \cite{Woltjer-1925}. The antiferromagnetic ground state consists of moments lying in the plane defined by the Cr layers. Those moments are aligned ferromagnetically within a layer, and the layers stack antiferromagnetically, as demonstrated by neutron diffraction \cite{Cable-1961}. The transition temperature of 17\,K was identified by heat capacity measurements \cite{Hansen-1958, Starr-1940}. Kuhlow performed Faraday rotation measurements and concluded that upon cooling the magnetic order appeared to develop by first forming two-dimensional ferromagnetic order within the layers and then, at slightly lower temperature, forming long range three-dimensional order through antiferromagnetic coupling between the layers \cite{Kuhlow-1982}. Faraday rotation, magnetization, and neutron diffraction measurements show that the ordered state has very little anisotropy, and fields of only a few kOe are required to fully polarize the magnetization in or out of the plane \cite{Bizette-1961, Cable-1961, Kuhlow-1982}. Weak magnetic anisotropy and weakly antiferromagnetic interlayer interactions are also reported from spin wave analysis \cite{Narath-1965}. In addition, CrCl$_3$ is known to undergo a crystallographic phase transition near 240\,K, similar to CrBr$_3$ and CrI$_3$, corresponding to a change in the layer stacking arrangement and a transition from monoclinic ($C2/m$) at high temperature to rhombohedral ($R\overline{3}$) at low temperature with little change in the intralayer structure \cite{Morosin-1964, McGuire-2015}.
There have been several recent theoretical studies of CrCl$_3$ based on first principles electronic structure calculations. Wang \textit{et al.} employed all-electron calculations for bulk CrCl$_3$, and the experimentally observed rhombohedral-antiferromagnetic ground state was reproduced for one of the methods used for incorporating U, the on-site Coulomb repulsion for Cr \cite{Wang-2011}. The authors noted that the energetics of the different crystal and magnetic structures are sensitive enough to the specific approach used that it is difficult to draw definitive conclusions about their relative stability, consistent with the experimentally observed temperature induced crystallographic phase transition and the weak magnetic anisotropy in the antiferromagnetic state. Other theoretical studies focus mainly on CrCl$_3$ monolayers. Liu \textit{et al.} observed ferromagnetic order below 66\,K in Monte Carlo simulations of monolayers, and suggested that hole doping should increase the Curie temperature \cite{Liu-2016}. The cleavage energy in the bulk crystal was calculated using a van der Waals density functional (vdW-DF) method to account for the dispersion forces in the bulk crystal. This gave a very low cleavage energy of 0.10$-$0.13\,J/m$^2$. Zhang \textit{et al.} calculate a cleavage energy of 0.3\,J/m$^2$ using a different vdW-DF \cite{Zhang-2015}. Monte Carlo simulations using their first principles calculations results for monolayer CrCl$_3$ suggest a Curie temperature of 49\,K, which is predicted to increase with applied strain. The results of those two studies indicate that CrCl$_3$ monolayers should be mechanically stable.
Here we report results from a thorough study of bulk CrCl$_3$\ crystals, including x-ray diffraction, magnetization, and heat capacity data, along with van der Waals density functional studies of the structure and magnetism. In addition we include optical images and atomic force microscopy measurements on ultrathin specimens cleaved from the bulk crystals. Our observations for the bulk crystals are in agreement with previously published literature, and in addition: (1) we note an anomaly in the magnetic susceptibility at the crystallographic phase transition indicating some coupling of the magnetism to the crystal lattice, (2) we find good structural agreement between theory and experiment only when magnetism is included in the calculations, again indicating spin-lattice coupling, (3) we obtain an estimate of the in-plane spin-flop field from magnetization measurements, (4) we observe two separate heat capacity anomalies associated with magnetic ordering indicative of short-ranged or two-dimensional ferromagnetism evolving into long-range antiferromagnetism, (5) we present the evolution of isothermal magnetization curves that is consistent with this scenario and construct a temperature-field phase diagram for CrCl$_3$, and (6) we demonstrate stable monolayer and few-layer specimens can be cleaved from bulk crystals. Overall, it is apparent that CrCl$_3$ is a promising material for the study of magnetism in ultrathin crystals and for incorporating magnetism into van der Waals heterostructures.
\section{Procedures}
Crystal growth is described in the following Section. A PANalytical X-Pert Pro MDP diffractometer equipped with an Oxford PheniX cryostat was used for x-ray diffraction measurements (Cu-K$_{\alpha1}$ radiation). Heat capacity and ac magnetization measurements were performed using a Quantum Design PPMS, and dc magnetization was measured using a Quantum Design MPMS. Optical absorbance \textit{A} was determined from the measured optical transmittance \textit{T} by $A = -log(T)$. Transmission measurements were made using a home-built system specifically designed for small samples. The light source was an Energetiq LDLS 99, which was collimated coming from the fiber. The detector was an Ocean Optics USB4000 0.1 meter spectrometer attached to a 600 micron multimode fiber. The sample was placed directly in front of the 600 micron entrance of the fiber. The transmission was determined by the ratio of the spectroscopic intensity with the sample in place and the intensity with the sample removed. Exfoliation of bulk CrCl$_3$ was achieved using mechanical exfoliation with scotch tape onto 90\,nm SiO$_2$ substrates. Optical microscopy images were obtained using an Olympus BX51M microscope, while atomic force microscopy (AFM) was carried out using a Bruker Edge Dimension atomic force microscope.
DFT calculations were performed using the Vienna Ab-initio Simulation Package (VASP) \cite{Kresse-1996} with the projector-augmented wave (PAW) \cite{Blochl-1994} potentials in order to understand the crystallographic and electronic properties of bulk CrCl$_3$. The exchange-correlation was approximated with generalized gradient approximation (GGA) of Perdew-Burke-Ernzerhof (PBE) functionals \cite{Perdew-1996} as well as van der Waals correction using vdW-DF-optB86b functional \cite{Dion-2004, Klimes-2010, Thonhauser-2007, Thonhauser-2015}. The pseudopotentials used explicitly treat $3p^6$, $4s^1$, $3d^5$, and $3s^2$, $3p^5$ electrons as valence electrons for Cr and for Cl, respectively. The Brillouin zone (BZ) integration was performed using the Monkhorst-Pack \cite{Monkhorst-1976} sampling method with a 4$\times$4$\times$2 ($R\overline{3}$) and 4$\times$2$\times$2 ($C2/m$) k-meshes for structural optimization.The energy cutoff was 500\,eV and the criteria for energy and force convergence were set to be $1\times10^{-4}$\,eV and 0.01\,eV/{\AA}, respectively.
\section{Results and Discussion}
\begin{figure}
\begin{center}
\includegraphics[width=3.0in]{crystal-figure.eps}
\caption{\label{fig:crystals}
(a) CrCl$_3$ single crystals grown for this study. (b) The optical absorbance spectrum showing strong absorption of green and red resulting in the observed violet color. (c) Temperature dependence of the layer spacing in CrCl$_3$ measured using x-ray diffraction during the first cooling of the crystal. (d) Relative amount of the monoclinic phase present during the first cooling through the monoclinic to rhombohedral phase transition. The two crystal structures are depicted on this panel as well. (e) Contour plot of the diffracted intensity from the monoclinic 0\,0\,5 and rhombohedral 0\,0\,15 reflections on the second cooling and warming cycles.
}
\end{center}
\end{figure}
\subsection{Crystal growth, structure, and spin-lattice coupling}
Thin, plate-like single crystals of CrCl$_3$ with lateral dimensions up to several millimeters are easily grown by recrystallizing commercial CrCl$_3$ using chemical vapor transport \cite{Starr-1940}. Here CrCl$_3$ from Alpha Aesar with a metals-basis purity of 99.9\,\% was used. The as-received material, in the form of small platelets, was sealed inside evacuated silica tubes. A typical growth used about 1\,g of CrCl$_3$ in a 15\,cm long tube with 16\,mm inner diameter and 1.5\,mm wall thickness. The tubes were placed in a horizontal tube furnace so that the starting material was at the end of the tube near the center of the furnace and the other end of tube was near the opening at the end of the furnace. The temperature at the center of the furnace was set to 700\,$^{\circ}$C, and the temperature at the cooler end of the tube was measured to be 550\,$^{\circ}$C. The growth begins rapidly and slows as the starting material is consumed. Large, violet-colored, transparent platelets like those shown in Fig. \ref{fig:crystals}a are present after 48\,h at temperature, while it may take up to a week for all of the CrCl$_3$ to be transported. After the recrystallizations are complete a green powder, presumably Cr$_2$O$_3$, is left at the hot end of the tubes. This is attributed to oxide impurity in the starting material.
The optical absorbance of a CrCl$_3$ crystal grown for the present study is shown in Fig. \ref{fig:crystals}b. A color scale showing the approximate color of the visible light for photon energies between 1.7 and 3.3\,eV is included on the plot. The results are consistent with the data reported in the thorough study of the optical properties of this material by Pollini and Spinolo \cite{Pollini-1970}. A band gap of 3.1\,eV is approximated by the onset of strong absorption at higher energies. Below the band edge, there are two broad absorptions in the red and green centered near 1.7 and 2.3\,eV, to which the violet color of the crystal can be attributed. These bands, which have fine structure not resolved here, arise from Cr$^{3+}$ \textit{d}-\textit{d} transitions as described in detail in Ref. \cite{Pollini-1970}.
As noted above, CrCl$_3$ is known to undergo a crystallographic phase transition below room temperature. Using single crystal diffraction, Morosin and Narath \cite{Morosin-1964} demonstrated that CrCl$_3$ adopts the monoclinic space group $C2/m$ at room temperature and the rhombohedral space group $R\overline{3}$ at 225\,K. They used nuclear quadrupole resonance to identify the phase transformation temperature as 240\,K. The crystallographic phase transitions involves mainly a change in the layer-to-layer shift in the stacking sequence. The intralayer structure remains nearly the same, with an ideal honeycomb net of Cr in the rhombohedral structure and a slightly distorted honeycomb net in the monoclinic structure \cite{Morosin-1964, McGuire-2015}. For example, in CrCl$_3$ the in-plane Cr-Cr distances are 3.431\,{\AA} at 225\,K (rhombohedral) and 3.440 and 3.441\,{\AA} at 298\,K (monoclinic) \cite{Morosin-1964}.
Results of x-ray diffraction measurements through this phase transition are shown in Figure \ref{fig:crystals}c-e. During the first cooling the high and low temperature phases coexist over a relatively wide temperature range (Fig. \ref{fig:crystals}c). The fraction of the monoclinic phase present during this initial cool down, determined by the ratio of the peak intensities, is shown in Figure \ref{fig:crystals}d. More than 85\% of the crystal transforms between 240 and 230\,K, and the remaining small fraction of the high temperature phase essentially vanishes between 140 and 130\,K. The two crystal structures are shown on this figure.
During subsequent cooling and heating cycles the phase transformation is much sharper (Fig. \ref{fig:crystals}e), as was found to the case in CrI$_3$ crystals \cite{McGuire-2015}. This suggests that the retained high temperature phase present in the first cool down is likely associated with defects or strain frozen into the crystals during the growth at elevated temperature and subsequent cooling to room temperature. There is still significant thermal hysteresis after the first thermal cycling, consistent with the first order nature of the phase transition.
\begin{table*}
\begin{center}
\caption{\label{tab:dft}
Structural parameters from DFT calculations using PBE/GGA and vdW-DF-optB86b for bulk CrCl$_3$ in the monoclinic and rhombohedral crystal structures, with corresponding experimental values from Ref. \citenum{Morosin-1964}. Energies given in eV/atom can be converted to eV/Cr by multiplying by four.}
\setlength{\tabcolsep}{2mm}
\begin{tabular}{c|cc|cc|cc|c}
\toprule
\multicolumn{8}{c}{Monoclinic, $C2/m$} \\
\hline
& \multicolumn{2}{c|}{non-magnetic} & \multicolumn{2}{c|}{ferromagnetic} & \multicolumn{2}{c|}{antiferromagnetic} & exp. \cite{Morosin-1964} \\
& PBE & vdW-DF & PBE & vdW-DF & PBE & vdW-DF & $T = 298$\,K \\
\hline
a ({\AA}) & 5.992 & 5.825 & 6.060 & 5.944 & 6.039 & 5.945 & 5.959 \\
b ({\AA}) & 10.372 & 9.404 & 10.490 & 10.288 & 10.457 & 10.290 & 10.321 \\
c ({\AA}) & 6.517 & 6.086 & 6.656 & 6.043 & 6.571 & 6.036 & 6.114 \\
$\beta$ (deg.) & 107.3 & 107.5 & 107.1 & 108.6 & 107.3 & 108.7 & 108.5 \\
vdW gap ({\AA}) & 3.775 & 3.100 & 3.869 & 3.245 & 3.772 & 3.205 & 3.299 \\
Energy (eV/atom) & -- & -3.503 & -- & -3.785 & -- & -3.785 & -- \\
\toprule
\multicolumn{8}{c}{Rhombohedral, $R\overline{3}$} \\
\hline
& \multicolumn{2}{c|}{non-magnetic} & \multicolumn{2}{c|}{ferromagnetic} & \multicolumn{2}{c|}{antiferromagnetic} & exp. \cite{Morosin-1964} \\
& PBE & vdW-DF & PBE & vdW-DF & PBE & vdW-DF & $T = 225$\,K \\
\hline
a ({\AA}) & 5.652 & 5.563 & 6.054 & 5.942 & 6.053 & 5.930 & 5.942 \\
c ({\AA}) & 19.641 & 17.288 & 19.239 & 17.088 & 19.027 & 17.184 & 17.333 \\
vdW gap ({\AA}) & 3.816 & 3.035 & 3.746 & 3.037 & 3.676 & 3.070 & 3.093 \\
Energy (eV/atom) & -- & -3.454 & -- & -3.786 & -- & -3.786 & -- \\
\toprule
\end{tabular}
\end{center}
\end{table*}
Results of our DFT calculations for CrCl$_3$ in both the monoclinic and rhombohedral structures are summarized in Table \ref{tab:dft}. We investigated non-magnetic (NM), ferromagnetic (FM), and antiferromagnetic (AFM) configurations for both crystal structures. The AFM configuration refers to the experimental magnetic structure, with ferromagnetic planes stacked antiferromagnetically. Results are presented for both PBE and vdW-DF-optB86b functionals. For one of these cases, the FM rhombohedral structure, results have been previously reported by Zhang \textit{et al.} using PBE and the van der Waals density functional optB88 \cite{Zhang-2015}.The results shown in Table \ref{tab:dft} for this configuration are within 1\% of the values reported there. Our calculations indicate the monoclinic structure to be favored by 0.05\,eV/atom in the non-magnetic calculations, and essentially degenerate monoclinic and rhombohedral structures for the magnetically ordered cases, with a very small 0.001\,eV/atom preference for the experimentally observed rhombohedral structure. The presence of nearly degenerate crystallographic ground states is consistent with the observed temperature induced structural phase transition.
The cleavage energy calculated in the rhombohedral structure with vdW-DF-optB86B is 0.3\,J/m$^2$, in agreement with the calculated value for CrCl$_3$ given in Ref. \citenum{Zhang-2015}. This is similar to values determined using similar approaches for the other chromium trihalides \cite{Zhang-2015, McGuire-2015}, as well as other related, easily-cleaved materials noted in the Introduction. The low cleavage energy enables monolayer and few-layer specimens to be cleaved from the CrCl$_3$ crystals, as demonstrated below.
The optimized lattice parameters for the FM and AFM unit cells are similar to one another, and those determined from vdW-DF-optb86b calculations agree well with the experimental structure (Table \ref{tab:dft}). As expected the PBE results show an elongation along the stacking directions, the \textit{c} axes. It is particularly interesting to note the significant difference between the optimized lattice parameters for the NM structure and those determined for the magnetically ordered structures. In each case the NM unit cell is significantly smaller. This is true for both the monoclinic and rhombohedral structures. For the vdW-DF-optb86b results in the AFM and FM structures, the optimized unit cell volumes are within 2\% of the experimental values. For the NM case they are smaller by more than 10\% (Table \ref{tab:dft}). The reason for this is not completely clear, but it does suggest a strong coupling between the spins and the crystal lattice in CrCl$_3$. The difference in unit cell volume arises primarily from variation in the in-plane lattice constants, where magnetic exchange is expected to be the strongest; however, it is important to note that the method used here includes spin only in the local part of the exchange-correlation functional, and not in the non-local part \cite{Thonhauser-2015}, which may be more important in the interplanar interactions. Crystallographic studies through the magnetic ordering temperature would be helpful in probing the spin-lattice coupling identified in these calculations. Although our diffraction measurements do not extend to low enough temperature to examine this in CrCl$_3$, a structural response at the magnetic ordering temperature was noted in isostructural CrI$_3$ \cite{McGuire-2015}.
\begin{figure*}
\begin{center}
\includegraphics[width=6.5in]{Mag-figure.eps}
\caption{\label{fig:MT}
Measured magnetic behavior of CrCl$_3$ crystals. (a) Temperature dependence of the inverse susceptibility (H/M) measured with a 10\,kOe field applied in the plane showing the small anomaly at the structural phase transition (inset) and Curie-Weiss fits to data above and below the transition. (b) Moment per Cr atom vs temperature near the magnetic ordering measured upon cooling in the indicated applied magnetic fields both in and out of the plane. (c) Isothermal magnetization curves near the magnetic ordering temperatures. Curves were measured at temperature intervals of 1\,K and data sets at 10, 14, 17, and 20\,K are labeled. (d) Moment per Cr atom measured at 2\,K. Due to the thin plate-like shape of the crystals, a demagnetization factor of 4$\pi$ (cgs units) was used to determine the internal field for the field out of the plane measurements. The inset shows both uncorrected (open circles) and corrected data (solid circles).
}
\end{center}
\end{figure*}
Additional evidence of spin-lattice coupling is seen in the magnetization data near the structural phase transition plotted in Fig. \ref{fig:MT}a. Temperature dependent magnetic susceptibility data, plotted as inverse susceptibility $H/M$, reveal a small discontinuity at 256\,K. The data was collected on warming, and the temperature of the discontinuity agrees well with the transition temperature determined by x-ray diffraction on warming (Fig. \ref{fig:crystals}e). Similar evidence of coupling between the magnetism and crystal lattice was previously noted for CrI$_3$ \cite{McGuire-2015}. Above and below this temperature Curie-Weiss behavior is observed, down to about 100\,K and up to 380\,K, the highest temperature investigated here. The results are in agreement with other analyses of high temperature data \cite{Starr-1940, Hansen-1959a}, although the anomaly at the structural phase transition was not noted in previous measurements due to relatively sparse data. Both the high and low temperature fits in Fig. \ref{fig:MT}a give an effective moment that is close to the expected value for spin-only Cr$^{3+}$ (3.87\,$\mu_B$).
The primary difference between the paramagnetic behavior above and below the structural transition is in the Weiss temperature $\theta$, which is smaller in the high temperature monoclinic phase than in the low temperature rhombohedral phase. The Weiss temperatures are positive, indicating that the intraplanar ferromagnetic interactions dominate the magnetic behavior in the paramagnetic state. The strength of the magnetic interactions, primarily superexchange, are expected to be sensitive to the details of the chemical coordination, allowing coupling of the magnetic behavior to the crystal structure details. In particular, the antiferromagnetic interplanar superexchange interactions will depend upon how the CrCl$_3$ layers stack. Since the crystallographic phase transition amounts essentially to a change in the layer stacking, this is identified as a likely source of the coupling between the magnetism and the lattice that produces the magnetic anomaly near 250\,K.
While both CrCl$_3$ and CrI$_3$ {\cite{McGuire-2015} show evidence of spin-lattice coupling, key differences can be noted in their behaviors and magnetic structures. The most obvious difference is the orientation of the ordered moments in the two phases, in-plane for the chloride and out of plane for the iodide (and bromide). In addition, the response of the magnetic susceptibility to the structural phase transition, as quantified by fits using a Curie-Weiss model, differ in the two materials. In CrCl$_3$, the Weiss temperature changes by nearly a factor of three, while in CrI$_3$ this parameter changes little and a response is seen mainly in the effective moment. As noted above, the crystallographic layer spacing was seen to decrease upon cooling into the magnetically ordered state CrI$_3$ \cite{McGuire-2015}. A similar type of coupling between the crystal structure and magnetic order may be expected in CrCl$_3$, though the present diffraction measurements do not extend to low enough temperature to test this hypothesis. A detailed crystallographic study of chromium trihalides through the magnetic ordering temperatures as well as a comparative theoretical study incorporating non-collinear magnetic structures and spin-orbit coupling would be desirable to help quantify and better understand the coupling between the magnetism and the lattice in these compounds.
\subsection{Magnetic behavior and phase transitions}
Magnetization data from our crystals near the magnetic ordering transitions are summarized in Figure \ref{fig:MT}b-c. The temperature dependent data were collected upon cooling, and the isothermal magnetization curves were collected upon decreasing the applied field. The results are in general agreement with previous studies employing a variety of techniques, which find below about 14\,K moments in each layer ferromagnetically aligned and lying in the \textit{ab}-plane with antiferromagnetic stacking between planes along the \textit{c}-axis and weak magnetic anisotropy \cite{Bizette-1961, Cable-1961, Narath-1965, Kuhlow-1982}. The temperature dependence of the magnetic moment per Cr atom measured near the Curie temperature is shown in Figure \ref{fig:MT}b. At 4\,K and 2\,T a moment of 3.0\,$\mu_B$ per Cr is measured, as expected for S\,=\,$\frac{3}{2}$ trivalent Cr. Data are shown for multiple applied magnetic fields (no correction for demagnetization) with the field both in the plane of the CrCl$_3$ layers and out of the plane. The curves are very similar to those reported in Ref. \citenum{Bizette-1961}. Kuhlow reported similar behavior in Faraday-rotation measurements \cite{Kuhlow-1982}. In that study, the magneto-optical measurements indicated that the magnetic order developed in two steps upon cooling, corresponding to the upturn in the measured magnetization (Fig. \ref{fig:MT}b) below about 20\,K and the sharp cusp apparent in low field measurements near 14\,K. This was interpreted as 2D ferromagnetic order within the layers developing first, with interlayer long-range antiferromagnetic order setting in at lower temperature.
This two-step development of magnetic order can also be inferred from the isothermal magnetization curves of Bizette \textit{et al}. \cite{Bizette-1961}, as pointed out by Kuhlow, who saw similar behavior in isothermal Faraday rotation curves \cite{Kuhlow-1982}. This behavior is demonstrated for the crystals used in the present study in Figure \ref{fig:MT}c. Magnetization is shown as a function of field for temperatures between 10 and 20\,K, with the field applied out of the plane defined by the CrCl$_3$ layers and with the field applied out of this plane. Data collected with the field normal to the thin platelet crystals was corrected for demagnetization effects by assuming a demagnetization factor of 4$\pi$ (cgs units). The curves in Figure \ref{fig:MT}c can be divided into three families based on their curvature, and are delineated by different colors and symbols in the figure. The distinctions are most apparent in the out of plane data. Between 20 and 18\,K the curves are linear. At 17\,K a negative curvature develops, indicating a ferromagnetic-like response. This behavior is enhanced upon cooling to 14\,K. At 13\,K the response changes to nearly linear at low fields, characteristic of the behavior expected for polarizing an antiferromagnet with low anisotropy. Thus, the data suggests the onset of ferromagnetic correlations at 17\,K and antiferromagnetic order at 13\,K, in general agreement with Kuhlow's observations \cite{Kuhlow-1982}.
Results of heat capacity measurements on our CrCl$_3$ crystals are summarized in Figure \ref{fig:hc}. The data clearly show two thermal anomalies, a broad feature centered at 17.2\,K and a sharp lambda-like peak centered at 14.1\,K. Data from two samples are shown in Figure \ref{fig:hc}a and they are nearly identical. These well-resolved anomalies are not reported in previous heat capacity studies, which only observe a broad peak near 17\,K \cite{Hansen-1958, Kostryukova-1972}. The presence of two well-separated thermal signatures in the heat capacity data at low magnetic fields shown in Figure \ref{fig:hc}a strongly supports the interpretation of the magnetic data in terms of the evolution of magnetic order proposed by Kuhlow \cite{Kuhlow-1982} and described above.
The magnetic heat capacity $c_{mag}$ was estimated by subtracting a smooth curve fitted to the zero field data at temperatures between 2 and 50\,K excluding the 7.5$-$33\,K range. Figure \ref{fig:hc}c shows results for data collected in applied magnetic fields of 0 and 5\,kOe. Integrating $c_{mag}/T$ to extract the magnetic entropy released up to 30\,K gives about 2\,J\,K$^{-1}$\,mol-Cr$^{-1}$ at both fields. This is a small fraction of the expected value of $R\rm{ln}(2S+1)$\,=\,11.5\,J\,K$^{-1}$\,mol-Cr$^{-1}$, and suggests that most of the magnetic entropy is released as magnetic correlations develop well above the onset of long range order. This is consistent with previous observations for isostructural CrBr$_3$ and CrI$_3$ \cite{Jennings-1965, McGuire-2015}.
Similarly small lambda anomalies have been reported at the long range ordering temperatures in honeycomb layered compounds, for example MnPS$_3$ \cite{Takano-2004, Wildes-2006} and BaCo$_2$(AsO$_4$)$_2$ \cite{regnault-1990}, in which magnetic fluctuations are expected to be responsible for the ``missing'' entropy. Development of these fluctuations produces a very broad contribution to the temperature dependence of the heat capacity extending to well above the ordering temperature. Thus, while significant magnetic fluctuations are expected to be present in CrCl$_3$, they are likely not the origin of the ferromagnetic-like response seen just above 14\,K, since a relatively sharp thermal anomaly is seen at the onset of that behavior.
\begin{figure}
\begin{center}
\includegraphics[width=3.0in]{hc-figure.eps}
\caption{\label{fig:hc}
Heat capacity of CrCl$_3$ crystals. (a) Data collected on two different samples near the magnetic ordering, with position of the two observed peaks noted. Data from 2 to 300 K are shown in the inset. (b) Magnetic field dependence of the heat capacity at low fields (left) and higher fields (right). The field was applied out of the plane. (c) Estimated magnetic heat capacity plotted as $c_{mag}/T$ (symbols) determined for data collected in zero field and at 20 kOe, and the associated magnetic entropies (lines).
}
\end{center}
\end{figure}
To further probe the nature of the magnetic phase transitions, heat capacity data were collected in applied magnetic fields (normal to the CrCl$_3$ layers) and the data are presented in Figure \ref{fig:hc}b. For small applied magnetic fields, up to 2\,kOe, the feature near 17\,K is nearly unchanged while the feature at 14\,K is suppressed in magnitude and temperature. At applied fields of 5\,kOe and higher the low temperature feature is completely suppressed and the 17\,K feature begins to broaden and move to higher temperature. The latter behavior is typical of heat capacity anomalies associated with ferromagnetic order, which is reinforced by an applied magnetic field. The heat capacity anomaly at 14\,K is suppressed with field much more rapidly than is seen in typical antiferromagnets. In this case, such behavior can be attributed to the competing ferromagnetic order at low applied fields in this weakly anisotropic magnetic system.
Two heat capacity anomalies have been observed in data from the closely related compound RuCl$_3$, and have been associated with different types of magnetic order occurring in different parts of the sample that have different crystal structures or stacking fault densities \cite{Banerjee-2016, RuCl3-Cao-2016}. Several observations suggest that this is not the case for the CrCl$_3$ data shown in Fig. \ref{fig:hc}. First, identical behavior is observed in two samples selected from separate growths (Fig. \ref{fig:hc}a). Second, the evolution of the two heat capacity anomalies with applied field are coupled as described above. Third, the magnetization data (Fig. \ref{fig:MT}c) show that the FM-like behavior onsetting at 17\,K vanishes at the 14\,K transition indicating that the magnetic behavior of the entire sample is affected. Fourth, the ferromagnetic in-plane magnetic structure of CrCl$_3$ should result in less sensitivity to stacking faults than the more complex zig-zag antiferromagnetic order in RuCl$_3$.
Using heat capacity data like that shown in Figure \ref{fig:hc}b, where the local maxima can be easily identified and tracked as the applied magnetic field is changed, the magnetic phase diagram for CrCl$_3$ shown in Figure \ref{fig:PD} can be constructed. Three regions are delineated: (1) the paramagnetic (PM) phase at high temperatures and low fields, (2) the antiferromagnetic (AFM) region where the magnetic moments are aligned ferromagnetically within layers and the layers are stacked antiferromagnetically, and (3) a ferromagnetic-like phase (``FM''), which is either field polarized in the case of the high field region or short ranged or low dimensional in the case of the low field region between the PM and AFM phases. Note that the boundaries are not as well defined as this depiction might suggest, especially at higher magnetic fields.
\begin{figure}
\begin{center}
\includegraphics[width=3.25in]{PD-figure.eps}
\caption{\label{fig:PD}
(a) Magnetic phase diagram of CrCl$_3$ crystals. $H$ is the applied external field, directed out of the plane of the Cr layers. The points were determined from heat capacity measurements like those shown in Figure \ref{fig:hc}b. PM = paramagnetic, AFM = antiferromagnetic, ``FM'' = ferromagnetic-like. (b) Magnetic structures of CrCl$_3$ accessible with relatively small applied magnetic fields.
}
\end{center}
\end{figure}
\subsection{Magnetic anisotropy and potential for monolayer and heterostructure studies}
As noted above, the observed magnetic anisotropy in the ordered state is small. Figure \ref{fig:MT}d shows the magnetic moment as a function of magnetic field measured at 2\,K with the field in the plane and out of the plane. The out of plane data is shown both as a function of internal field (demag. corr.) and applied field (no demag. corr.) and is consistent with the data of Narath and Davis \cite{Narath-1965}. Demagnetization effects are significant as expected. An applied field of 6\,kOe is required to essentially fully polarize the magnetization out of the plane, consistent with neutron diffraction results \cite{Cable-1961}. This corresponds to an internal field of only 2.5\,kOe. Interestingly, this is the same field required to polarize the magnetization in the plane. Since within each plane the moments are ferromagnetically aligned and lie in the plane, it is expected that polarization out of the plane is realized through a coherent rotation of moments within each layer \cite{Kuhlow-1982}. When the field is applied in the plane, different field induced behaviors are expected depending on whether the field is parallel or perpendicular to the moment direction. In the latter case, a coherent rotation occurs, while in the former case a spin flop can be expected to occur first. The sample used for the in plane measurement comprised several thin crystals that were not co-aligned with respect to the \textit{ab}-plane, so the direction of the applied field relative to the moment direction in the zero field magnetic structure is not well defined. As a result the signature of the spin flop on the magnetization data is somewhat diluted, but it is still observable in Figure \ref{fig:MT}d. The spin flop is seen to occur at a field between 100 and 200\,Oe at \textit{T}\,=\,2\,K, consistent with the value of 163\,Oe estimated from magneto-optical measurements at 7\,K from which an in-plane anisotropy field of about 10\,Oe is determined \cite{Kuhlow-1982}. This is small relative to the field required to rotate the moments out of the plane, suggesting an XY-Heisenberg description for the moments in CrCl$_3$.
Note that the energies of the FM and AFM configurations in Table \ref{tab:dft} are the same within the precision of the calculations; the magnetically ordered structures are about 0.3\,eV/atom lower in energy than the NM states in both the monoclinic and rhombohedral crystal structures in the vdW-DF-optb86b calculations. This is similar to the 0.4\,eV/atom stability found in the calculations of Ref. \citenum{Wang-2011}. This appears consistent with the observation that the moments can be fully polarized with relatively small applied fields. Note that spin-orbit coupling, a key source of magnetocrystalline anisotropy, is not included in the calculations. However, this effect is expected to be weak in trivalent chromium with electronic configuration $3d^3$ and nearly octahedral coordination due to quenching of the orbital moment. Thus single ion anisotropy should be weak for Cr, as was demonstrated recently for CrI$_3$ by Lado and Fern{\'{a}}ndez-Rossier \cite{Lado-2017}. In that study, spin-orbit coupling on the heavy iodine ions was found to contribute the majority of the anisotropy through Cr-I-Cr superexchange. In CrCl$_3$ this effect is expected to be much weaker, since the spin-orbit coupling strength varies as the fourth power of the atomic number. Thus, spin-orbit coupling is expected to be negligible for both elements in CrCl$_3$, and this is likely the reason for the much smaller anisotropy field observed in CrCl$_3$ than CrI$_3$.
The low magnetic anisotropy and ferromagnetic in-plane order make CrCl$_3$ potentially very attractive for van der Waals heterostructure studies. Figure \ref{fig:PD}b shows the magnetic structure of two adjacent layers of CrCl$_3$ in zero field, a relatively small out-of-plane field, and a relatively small in-plane field. These distinct magnetic orderings would produce distinct magnetic proximity effects on neighboring materials. Importantly, tuning among these and any intermediate magnetic states should be possible, which would enable exquisite control over interactions and associated functionalities when coupled to appropriate electronic or optical materials within a heterostructured device.
Finally, we show here that monolayer CrCl$_3$ can indeed be realized by exfoliation of bulk crystals. Figure \ref{fig:layers}a shows an AFM image of a flake of CrCl$_3$ on a SiO$_2$ substrate. Several step edges are observed across the flake. Figure \ref{fig:layers}b shows the height profile across an edge measured along the path indicated by the red line in Figure \ref{fig:layers}a. The single layer spacing in CrCl$_3$ determined from the bulk crystal structure is $c~\rm{sin}\beta = 0.58$\,nm, indicated on the figure. The measured step height of $\approx 0.6$\,nm indicates that this step is one monolayer high. By correlating changes in optical contrast with the step height measured using AFM, optical contrast can be used to qualitatively assign layer number. An optical micrograph of an ultrathin specimen cleaved from a bulk crystal is shown in Figure \ref{fig:layers}c on a 90\,nm SiO$_2$ substrate. Measured optical contrast is used to identify the thicknesses of different regions, which is shown on the image in units of CrCl$_3$ layers. The inset shows a separate specimen determined to be one monolayer thick. It is expected that the single-layer CrCl$_3$ may display ferromagnetism at low temperature, as observed in CrI$_3$ \cite{Huang-2017} and as predicted by first principles calculations for all of the chromium trihalides \cite{Liu-2016, Zhang-2015}.
\begin{figure}
\begin{center}
\includegraphics[width=3.4in]{layers.eps}
\caption{\label{fig:layers}
(a) AFM micrograph showing several step edges on a cleaved surface of a CrCl$_3$ crystal. (b) Height profile across the step edge indicated by the red line in (a), with the CrCl$_3$ layer spacing of 0.58\,nm determined from the bulk crystal structure \cite{Morosin-1964} indicated on the plot. (c) Optical micrograph of an ultrathin specimen cleaved from a bulk CrCl$_3$ crystal. Regions of varying optical contrast are labeled by the corresponding thickness in number of CrCl$_3$ layers. A separate monolayer specimen is shown in the inset.
}
\end{center}
\end{figure}
\section{Summary and Conclusions}
We have presented a thorough investigation of the bulk properties of the layered antiferromagnet CrCl$_3$. The magnetic behavior of our crystals is in general agreement with literature reports, indicating ferromagnetic correlations emerging upon cooling before entering the long range antiferromagnetically ordered ground state. Low magnetic anisotropy is observed in the ordered state. A spin flop is seen near 150\,Oe when the field is applied in plane, and fields of several kOe are sufficient to polarize the moments in any direction. In addition, we find that the magnetic susceptibility shows a small anomaly at the crystallographic transition temperature, indicating some coupling of the magnetism to the lattice in the paramagnetic state. Spin-lattice coupling is also apparent in our first principles calculations, in which the experimental structure is matched only when magnetism is included in the calculations. Our heat capacity measurements confirm the two-step nature of the magnetic ordering transition. A broad feature appears near 17\,K indicating the development of ferromagnetism and a sharp, lambda-like anomaly is seen at 14\,K as the antiferromagnetic long range order forms. The observed evolution of the heat capacity with applied magnetic field supports this scenario, and allows the construction of the magnetic phase diagram for CrCl$_3$ shown in Figure \ref{fig:PD}. Neutron scattering experiments would be highly desirable to gain further insight into the development of the magnetic order, specifically to elucidate the potential roles of magnetic fluctuations, short range order, or 2 dimensional order. We have demonstrated that cleaving of CrCl$_3$ crystals into stable monolayer specimens is possible, and note several properties that make CrCl$_3$ a promising material for incorporating magnetism into van der Waals heterostructures at low temperatures: bulk crystals are easily grown, stable in air, electrically insulating, and have magnetic order that can be manipulated with relatively low external fields.
\section*{Acknowledgements}
Research sponsored by the US Department of Energy, Office of Science, Basic Energy Sciences, Materials Sciences and Engineering Division, and by the Laboratory Directed Research and Development Program of Oak Ridge National Laboratory, managed by UT-Battelle, LLC, for the U.S. Department of Energy (WMC). Computational resources were provided by the National Energy Research Scientific Computing Center (NERSC), which is supported by the Office of Science of the U.S. DOE under Contract No. DE-AC02-05CH11231.
| 7dd7fa2140d1a5cfe405afcfb65c25b20a989f13 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Cognitive radio networks (CRNs) have shown great promise in alleviating the acute shortage of spectrum. In such networks, secondary (unlicensed) users are allowed to share the spectrum of the primary (licensed) users. Underlay CRNs in particular, have been shown to result in great improvement in spectral utilization efficiency. In these networks, the secondary transmitters transmit simultaneously with the primary transmitters in the same frequency band, but with powers carefully constrained to limit interference to the primary receiver below a specified interference temperature limit.
\par In the recent years, the use of energy harvesting (EH) is being studied to prolong battery life of nodes. Although energy can be harvested from natural sources, it is wireless EH that shows the greatest promise. In particular, the possibility of simultaneous wireless information and power transfer has spurred research interest in this area. Since practical circuits cannot simultaneously harvest energy and perform information processing, time-switching and power-splitting relaying protocols have been proposed \cite{Nasir2013}. In the former, the relay is first charged by the source for a fraction of the signalling interval prior to two-hop relaying. In the latter, the received signal is split into two parts, with a fraction (referred to as the power-splitting parameter $\rho$) being used for energy harvesting, while the rest is used for information processing. It is well known that optimization of this parameter is crucial to maximize throughput.
\par While the use of EH relays in CRNs is well motivated, analysis of performance of such networks has attracted attention only over the past few years
\cite{Janghel2014,Liu2015,Bhowmick2016,Yang2015,Miridakis2016,Wang2016,He2016,Janghel2016}.
Two types of EH methodologies have been proposed in the context of CRNs In the first type, energy is harvested from the primary signal \cite{Janghel2014,Liu2015,Bhowmick2016,Miridakis2016,
Wang2016,He2016}. Note that \cite{Bhowmick2016,Miridakis2016} use the interweave cognitive radio principles, \cite{Janghel2014,Liu2015,Janghel2016} use the underlay signalling mode, while \cite{Wang2016,He2016} use overlay principles. In the second type, which is of primary interest in this paper, energy is harvested from the secondary source \cite{Yang2015,Janghel2016} (this requires secondary EH nodes to be in close proximity to the secondary transmitter). Under these circumstances, even in the non-cognitive context, the importance of considering the direct channel from source to destination in addition to the signal relayed by the EH relay has been recognized by only a few authors \cite{Lee2017,Van2016}. In {\em all existing literature} on two-hop EH relaying in underlay CRNs \cite{Liu2015,Yang2015,Janghel2016}, the direct channel from source to destination has been ignored. In underlay cognitive radio, powers used at secondary transmitters is a random quantity. For this reason, the secondary nodes need to be in close proximity to each other to ensure any reasonable quality of service in the secondary links. In underlay signalling with EH relays, the nodes need to be even closer since energy harvested is typically small. Ignoring the direct channel from source to destination is therefore not reasonable in such situations. In this paper, we demonstrate this fact.
\par The contributions of this paper are as follows:
\begin{enumerate}
\item We derive closed-form expressions for the throughput of an incremental relaying scheme in which the destination combines the signal relayed by the EH relay with the direct signal from the source.
\item Using approximated throughput expressions, we derive an expression for the power-splitting parameter that maximizes throughput.
\item We demonstrate that relaying with an EH relay results in larger throughput than direct signalling only when the destination combines the direct and relayed signals.
\end{enumerate}
{\em Notations}: $\E_{\C}(\cdot)$ denotes the expectation over the condition/conditions $\C$. $\exp(\lambda)$ represents the exponential distribution with parameter $\lambda$, and ${\cal CN}(0,a)$ denotes the circular normal distribution with mean $0$ and variance $a$. $E_1(\cdot)$ and $\text{Ei}(\cdot)$ represent the exponential integrals defined in \cite[5.1.1]{Abramowitz1964} and \cite[5.1.2]{Abramowitz1964} respectively.
\section{System Model}\label{sec:system_model}
\par We consider a two-hop decode-and-forward (DF) relay network as depicted in Fig.~\ref{fig:sys_model}.
\begin{figure}[h]
\vspace*{-0.3cm}
\centering
\includegraphics[width=0.5\textwidth]{system_model.eps}
\caption{System Model}
\label{fig:sys_model}
\vspace*{-0.3cm}
\end{figure}
The primary network consists of the primary transmitter (not depicted in the figure) and the primary receiver P. The secondary network (SN) consists of three nodes - a source (S), a relay (R) and a destination (D). Each node equipped with one antenna. Denote the channel between S and R by $h_{sr}\sim {\cal CN}(0,\lambda_{sr}^{-1})$, and that between the R and D by $h_{rd}\sim {\cal CN}(0,\lambda_{rd}^{-1})$. Similarly, denote the channel between S and P by $g_{sp}\sim {\cal CN}(0,\lambda_{sp}^{-1})$, and that between R and P by $g_{rp}\sim {\cal CN}(0,\lambda_{rp}^{-1})$. We assume that R is equipped with a super-capacitor, and acts as an EH node with EH factor $\eta$. The energy harvested in the first phase is used by R to relay the signal to D. As in most literature on underlay CRN, we neglect the primary signal at R and D. This is reasonable because of the large distance between the primary transmitter and the secondary nodes \cite{Tourki2013} (this assumption has been justified on information theoretic grounds \cite{Jovicic2009}). All the channels are reversible and quasi-static.
\section{Transmission Protocol}\label{sec:transmission_protocol}
We assume fixed-rate transmission at rate $R_s$ by all nodes. Signalling based on the incremental protocol is completed in two time-slots as depicted in Fig.\ref{fig:transmission_schemeInc}. Message transmission is based on the incremental relaying protocol \cite{Ikki2011}, and EH is based on the power-splitting protocol \cite{Nasir2013}.
\begin{figure}[h]
\centering
\includegraphics[width=0.48\textwidth]{transmission_schemeInc.eps}
\caption{Transmission scheme for incremental relaying}
\label{fig:transmission_schemeInc}
\end{figure}
\par In the first transmission time-slot, S transmits information symbols to D and R as depicted in Fig.\ref{fig:transmission_schemeInc}. R harvests energy from this signal using power splitting, and attempts to decode the information symbols. Meanwhile, D attempts to decode the symbols. If successful, it sends feedback to the source and the relay\footnote{We assume that feedback time is extremely small and can be neglected in the analysis without loss of generality.}. The relay discards the decoded symbols, and the source re-transmits a new block of symbols to D as depicted in Fig.\ref{fig:transmission_schemeInc}a\footnote{The relay does not harvest energy in this phase. Energy stored in the super-capacitor is assumed to be lost since it lacks the ability to store charge over long intervals.}. Else, R relays the symbols using the energy harvested, and D combines the signals in the first and the second time-slots as depicted in Fig.\ref{fig:transmission_schemeInc}b.
\par In the first time-slot, S transmits a message signal $x$ with power $P_s$ in underlay mode to R and D. In PS-EH case, a component $y_{d_1}$ of the received signal with $\rho$ fraction of the power is utilized for EH, while the remaining signal with fraction $1-\rho$ of the power is used to decode the symbols. Clearly, $y_r$ at R and and $y_{d_1}$ at D in the first phase are given by:
\begin{IEEEeqnarray}{rcl}
y_r &=& \sqrt{(1-\rho) P_{s}}\,x\,|h_{sr}|^2+n_{r}\;\quad\text{and}\label{eq:yr_PS}\\
y_{d_1} &= &\sqrt{P_{s}}\,x\, |h_{sd}|^2+n_{d_{1}} \label{eq:yd1_PS}
\end{IEEEeqnarray}
respectively, where $n_r,n_{d_1}\sim{\cal CN}\left(0,N_o\right)$ are noise samples at R and D. Clearly, the SNR $\Gamma_{d_1}$ at D in the first time-slot is $\frac{P_s|h_{sd}|^2}{N_o}$. Energy harvested at R is $\eta \rho\,P_s |h_{sr}|^2/2$ (ignoring noise) so that the power available for relaying is given by:
\begin{equation}
^hP_{r}=\rho\,\eta P_s |h_{sr}|^2 = \beta\,P_s |h_{sr}|^2\,,\label{eq:Prh-PS}
\end{equation}
where $\beta= \eta\, \rho$.
Let $I$ denote the interference temperature limit. In order to ensure that the interference caused to primary receiver P is limited to $I$, the power $P_s$ at S is chosen to be:
\begin{equation}
P_s= {I}/{|g_{sp}|^2}\,.\label{eq:PsOnlyInterference}
\end{equation}
We consider only the peak interference constraint at S, and ignore the peak power constraint for the following reason:
\begin{enumerate}
\item It is well known that performance of CRNs exhibits an outage floor, and does not improve with increase in peak power (it is in this low outage and high throughput region that CRNs are typically operated) \cite{Lee2011,Duong2012} (and references therein). In this paper, we discuss optimization of PS EH parameter, which is of interest in this high throughput regime.
\item Since performance of CRNs is typically limited by interference, and sufficient peak power is typically available, this assumption is quite reasonable.
\end{enumerate}
\par In the second time-slot, R is used to forward the decoded symbol $\hat{x}$ to D. In order to ensure that the interference at P is constrained to $I$, the total transmit power $ P_{r} $ at R is chosen to be:
\begin{equation}
P_{r} = \min\left(^hP_{r}, \frac{I}{|g_{rp}|^2} \right). \label{eq:Pr_PS}
\end{equation}
Received signal $y_{d_2}$ at D can be expressed as
\begin{equation}
y_{d_2} = \sqrt{P_{r}}\, \hat{x}\,h_{rd}+n_{d_2}\;, \label{eq:yd2_PS}
\end{equation}
where $n_{d_2}\sim{\cal CN}\left(0,N_o\right)$ is the additive white Gaussian noise sample. We assume that D uses MRC to combine the signals obtained in the first phase (\ref{eq:yd1_PS}) and second phase (\ref{eq:yd2_PS}).
Signal-to-noise ratios (SNRs) at R and D are expressed using (\ref{eq:yr_PS}), (\ref{eq:yd1_PS}) and (\ref{eq:yd2_PS}) as:
\begin{IEEEeqnarray}{rcl}
\Gamma_{r} &=&\frac{(1-\rho)P_s|h_{sr}|^2}{N_o}\qquad \text{and} \label{eq:SNR_rps}\\
\Gamma_{d} &=& \underbrace{\frac{P_s|h_{sd}|^2}{N_o}}_{\Gamma_{d_1}}+\underbrace{\frac{P_{r}|h_{rd}|^2}{N_o}}_{\Gamma_{d_2}}.\label{eq:SNR_dps}
\end{IEEEeqnarray}
\section{Performance Analysis}\label{sec:perfromance}
In this section, we analyze throughput performance of the described incremental relaying protocol. When $\Gamma_{d_1}\geq\gamma_{th}$, the direct link is successful, so that rate achieved is $R_s$. When $\Gamma_{d_1}<\gamma_{th}$, and the relay can decode successfully ($\Gamma_r\geq\gamma_{th}$), and the SNR at the destination after combining is sufficient ($\Gamma_d=\Gamma_{d_1}+\Gamma_{d_2}\geq \gamma_{th}$), signalling is completed in two hops (rate $R_s/2$). Clearly, throughput $\tau$ is given by:
{\begin{IEEEeqnarray}{rcl}
\tau&=& 0.5\,R_s\underbrace{\Pr\left(\Gamma_{d_1} < \gamma_{th},\, \Gamma_{r}\geq\gamma_{th},\, \Gamma_{d_1}+\Gamma_{d_2}\geq\gamma_{th}\right)}_{q_1}\nonumber\\
&&\hspace{1.5in}+\underbrace{\ R_s\Pr(\Gamma_{d_1}\geq\gamma_{th})}_{q_2}\,,\label{eq:tauIn}
\end{IEEEeqnarray}}
where $\gamma_{th}=2^{R_s}-1$. We note that $q_1$, the probability that the relayed link is successful while the direct link is not successful, can alternatively be represented as:
\begin{IEEEeqnarray}{rcl}
q_1&=&1-\Big(\underbrace{\Pr\left(\Gamma_r< \gamma_{th}\right)}_{p_1}+\underbrace{\Pr\left(\Gamma_{d_1}\geq\gamma_{th},\Gamma_r \geq\gamma_{th} \right)}_{p_2}\nonumber\\
&&\hspace{0.8in}+\underbrace{\Pr\left(\Gamma_d< \gamma_{th},\;\Gamma_r \geq \gamma_{th}\right)}_{p_3}\Big).\label{eq:q}
\end{IEEEeqnarray}
We evaluate each of the terms in what follows. From \eqref{eq:SNR_rps}, $p_1$ can be evaluated as:
\begin{eqnarray}
p_1=\Pr\left(\Gamma_r< \gamma_{th}\right) = \Pr\left(\frac{(1-\rho) P_s|h_{sr}|^2}{N_o}< \gamma_{th}\right). \label{eq:p1}
\end{eqnarray}
Using $P_s=I/|g_{sp}|^{2}$, it can be shown that:
\begin{equation}
p_1=1-\frac{1}{1+{\lambda_{sr}\psi}/({\lambda_{sp}(1-\rho)})},\label{eq:p1Final}
\end{equation}
where $\psi=\frac{\gamma_{th}}{I/N_o}$.
Similarly, $q_2$ is given by:
\begin{eqnarray}
q_2=\Pr(\Gamma_{d_1}\geq\gamma_{th})=\frac{1}{1+{\lambda_{sd} \psi}/{\lambda_{sp}}}.\label{eq:q2}
\end{eqnarray}
We note that $\Gamma_r$ and $\Gamma_{d_1}$ are not independent due to their dependence on the random variable $P_s=I/|g_{sp}|^{2}$. By first conditioning on $|g_{sp}|^{2}$, exploiting the independence of $\Gamma_{r\big||g_{sp}|^2}$ and $\Gamma_{d\big||g_{sp}|^{2}}$, and then averaging over $|g_{sp}|^{2}$, we can show that $p_2$ can be written as:
\begin{IEEEeqnarray*}{rcl}
p_2& =& \int_{0}^{\infty} \lambda_{sp} e^{\lambda_{sp}|g_{sp}|^2} \int_{\psi |g_{sp}|^2}^{\infty} \lambda_{sd} e^{-\lambda_{sd} |h_{sd}|^2} \int_{\frac{\psi|g_{sp}|^2}{(1-\rho)}}^{\infty}\lambda_{sr} \nonumber\\
&&\hspace*{2cm}e^{-\lambda_{sr}|h_{sr}|^2} d|h_{sr}|^2 d|h_{sd}|^2 d|g_{sp}|^2. \nonumber
\end{IEEEeqnarray*}
After integrating over $|h_{sr}|^2$ and $|h_{sd}|^2$, $p_2$ can be expressed in terms of $|g_{sp}|^{2}$ as:
\begin{equation*}
p_2=\int_{0}^{\infty} \lambda_{sp} e^{\lambda_{sp}|g_{sp}|^2} e^{-\lambda_{sd}\psi |g_{sp}|^2-\frac{\lambda_{sr} \psi |g_{sp}|^2}{(1-\rho)}}d |g_{sp}|^2. \nonumber
\end{equation*}
Now after averaging over $|g_{sp}|^2$, the resultant expression can be found out to be:
\begin{eqnarray}
p_2=\frac{1}{1+ \frac{\psi}{\lambda_{sp}}\left(\lambda_{sd}+\frac{\lambda_{sr}}{(1-\rho)}\right)}.\label{eq:p2Final}
\end{eqnarray}
An approximate closed-form expression for $p_3$ is derived in Appendix-A. The final expression is presented in \eqref{eq:p3_Final}.
\setcounter{equation}{14}
\begin{figure*}[t!]
\begin{IEEEeqnarray}{rcl}
&&p_3\approx t \Bigg(\left(\frac{a}{c+\lambda_{sp}}\right)^2\frac{ \lambda_{sp} \lambda_{sr}}{ (a+b+d \lambda_{sp})} e^{\frac{a \lambda_{sr}}{c+\lambda_{sp}}} \left(E_1\left(\frac{a \lambda_{sr}}{c+\lambda_{sp}}\right)-E_1\left(\frac{a \lambda_{sr}}{c+\lambda_{sp}}+a s\right)\right)+\frac{d^2 \lambda_{sp} \lambda_{sr}}{a+b+d \lambda_{sp}} e^{s (-(a+b))-d (\lambda_{sp} s+\lambda_{sr})} \nonumber\\
&&\times(\text{Ei}(d (\lambda_{sr}+\lambda_{sp} s))-\text{Ei}(d \lambda_{sr}+b s+d \lambda_{sp} s))-\frac{\lambda_{sr} (a+b)^2e^{\frac{\lambda_{sr} (a+b)}{\lambda_{sp}}}}{\lambda_{sp} (a+b+d \lambda_{sp})} \left(E_1\left(\frac{(a+b) \lambda_{sr}}{\lambda_{sp}}\right)-E_1\left(\frac{(a+b) \lambda_{sr}}{\lambda_{sp}}+(a+b) s\right)\right)\nonumber\\
&&-\frac{\lambda_{sr} e^{s (-(a+b))}}{\lambda_{sp} s+\lambda_{sr}}+\frac{\lambda_{sp} \lambda_{sr} e^{-a s}}{(c+\lambda_{sp}) (c s+\lambda_{sp} s+\lambda_{sr})}+\frac{c}{c+\lambda_{sp}}\Bigg)+\frac{b^2 \lambda_{sr}}{\lambda_{sp} (b+d \lambda_{sp})} e^{\frac{b \lambda_{sr}}{\lambda_{sp}}} \left(E_1\left(\frac{b \lambda_{sr}}{\lambda_{sp}}\right)-E_1\left(\frac{b (\lambda_{sr}+\lambda_{sp} s)}{\lambda_{sp}}\right)\right)\nonumber\\
&&-\frac{d^2 \lambda_{sp} \lambda_{sr}}{b+d \lambda_{sp}} e^{-b s-d (\lambda_{sp} s+\lambda_{sr})} (\text{Ei}(d (\lambda_{sr}+\lambda_{sp} s))-\text{Ei}(d \lambda_{sr}+b s+d \lambda_{sp} s))+\frac{\lambda_{sr} e^{-b s}}{\lambda_{sp} s+\lambda_{sr}}+\frac{c \lambda_{sp} s^2}{(\lambda_{sp} s+\lambda_{sr}) (c s+\lambda_{sp} s+\lambda_{sr})}\nonumber\\
&&-\frac{\lambda_{sp} \lambda_{sr}}{(c+\lambda_{sp}) (c s+\lambda_{sp} s+\lambda_{sr})}-\frac{c}{c+\lambda_{sp}}\,,
\label{eq:p3_Final}\\
&& {\normalsize\text{where $a=\frac{\lambda_{rp}}{\beta}$, $b=\frac{\psi \lambda_{rd} }{\beta}$, $c={\psi\lambda_{sd}}$, $d=\frac{\lambda_{rd}}{\beta \lambda_{sd}}$, and $s=\frac{(1-\rho)}{\psi}$.\label{eq:param}}}
\end{IEEEeqnarray}
\hrulefill
\vspace*{-0.5cm}
\end{figure*}
\setcounter{equation}{16}
Resultant expression of $\tau$ can be found out by substituting for $p_1$, $p_2$, $p_3$ and $q_2$ into \eqref{eq:tauIn}.\\
\subsection{Value of EH-factor for optimal throughput}
Throughput $\tau$ is small for $\rho=0$ (no energy harvested at the relay) and $\rho=1$ (no decoding is possible at the relay). In both these cases, the relayed signal is not available. When $0<\rho<1$, the throughput is larger since the relayed signal is not always in outage. It is clear that the
following optimal value of EH parameter ($\rho^{*}$) that maximizes throughput is of interest:
\begin{eqnarray}
\rho^* = \arg\displaystyle\max_{\rho} \quad \tau \label{eq:rhoOptimum}.
\end{eqnarray}
It is difficult to find an exact solution for $\rho^*$ since the lengthy expression for $\tau$ contains several nonlinear functions.
\par To obtain an expression for $\tau$ in a simplified form (say $\tau_{sim}$), we use the high-SNR approximation $\left({I\lambda_{sp}}/{N_o}\gg\gamma_{th}\right)$ and also neglect the R-P link\footnote{We note that the R-P link is ignored only for the simplified analysis to obtain insights into the optimum power-splitting parameter. The relay needs to apply power control as in any other underlay system. We note that all computer simulations are performed with the interference channel from relay to primary receiver.}. We show through simulations in Fig. \ref{fig:TauVsRho} of Section~\ref{sec:simulations} that throughput of a practical system that imposes the peak interference constraint at the relay is indistinguishable from one that neglects it. In other words, SN performance does not depend (or at best very loosely depends) on the statistical parameter ($\lambda_{rp}$) of the R-P channel for most practical range of parameters. Intuitively, this is because of the fact that the harvested energy is very small (less than $I/|g_{rp}|^2$) with very high probability.
\par From the above discussion, throughput can be represented in most simplified form as given in \eqref{eq:TauSimplified} (please refer the Appendix-B for derivation).
\begin{figure*}[t!]
\begin{eqnarray}
\tau_{sim}\approx0.5R_s\left(1-\left(\frac{1}{\left(1+\frac{\lambda_{sp}}{\lambda_{sd}\psi}\right)\left(1+\frac{\eta \lambda_{sp} \rho}{\psi \lambda_{rd} \lambda_{sr} }\right)}+\frac{1}{\frac{\lambda_{rd} \lambda_{sp} (1-\rho)}{\eta \rho \lambda_{sr} \psi}}+\frac{1}{\frac{\psi \lambda_{sd} }{ \lambda_{sp}}+1}\right)\right) + R_s q_2\label{eq:TauSimplified}
\end{eqnarray}
\hrulefill
\vspace*{-0.4cm}
\end{figure*}
We omit proof of concavity due to space constraints. Solving
$\frac{d\tau_{sim}}{d\rho}=0$.
results in a quadratic equation which has two roots, out of which the one between $0$ to $1$ is given by\footnote{Since $\lambda_{sr}\ll\lambda_{sd}$, it can be shown that $\rhoΛ{*}\in [0,1]$.}:
\begin{equation}
\rho^* \approx \frac{1-\sqrt{\left(1+\frac{\lambda_{sp}}{\lambda_{sd}\psi}\right)}\frac{ \psi \lambda_{sr}\,}{ \lambda_{sp}}}{1+\sqrt{\left(1+\frac{\lambda_{sp}}{\lambda_{sd}\psi}\right)}\frac{ \eta}{\lambda_{rd}}}.\label{eq:OptRho}
\end{equation}
\par As a special case, value of $\rho$ which maximizes the throughput if the S-D link ignored \footnote{In this case, there is no involvement of direct path and a case of two-hop transmission between nodes S to D via R.} (i.e $\lambda_{sd}\rightarrow\infty$) becomes:
\begin{eqnarray}
\rho^*_{nd} \approx \lim\limits_{d_{sd}\rightarrow\infty} \rho^* = \frac{1-\frac{ \psi \lambda_{sr} \,}{ \lambda_{sp}}}{1+\frac{ \eta}{\lambda_{rd}}},\label{eq:OptRhoWSD}
\end{eqnarray}
where subscript $nd$ is used to emphasize the fact that no direct path is present. In this case, the throughput is derived from \eqref{eq:tauIn} by using $\lambda_{sd}\rightarrow\infty$, and given by:
\begin{eqnarray}
\tau_{nd} = \lim\limits_{\lambda_{sd}\rightarrow\infty} \, \tau.\label{eq:TauWSD}
\end{eqnarray}
\section{Simulation Results}\label{sec:simulations}\
In this section, we validate the derived expressions by computer simulations. The normalized S-R, R-D, and S-D distances are assumed to be $1.2$, $1.8$ and $3$ respectively. The normalized S-P and R-P distances are assumed to be $3$. Path loss exponent $\epsilon$ is assumed to be $4$. We assume $\eta = 0.7$ and $I/N_o=6$ dB unless stated otherwise.
\begin{figure}[h]
\vspace*{-0.3cm}
\centering
\includegraphics[width=0.5\textwidth]{TauVsRho.eps}
\caption{Throughput vs. $\rho$ for different $d_{sr}$}
\label{fig:TauVsRho}
\vspace*{-0.25cm}
\end{figure}
\par The importance of optimizing $\rho$ for maximizing throughput is clearly brought out in Fig.\ref{fig:TauVsRho} (concavity is clear from the plots). The graph depicts a plot of $\tau$ versus $\rho$ for $R_s=3$ bpcu for different $d_{sr}$.
\begin{figure}[h]
\vspace*{-0.3cm}
\centering
\includegraphics[width=0.5\textwidth]{TauVsRsConf.eps}
\caption{Throughput vs. $R_s$ for different I}
\label{fig:TauVsRsConf}
\vspace*{-0.5cm}
\end{figure}
\par The value of $\rho^*$ indicated by \eqref{eq:OptRho} for $d_{sr}$ of $1.2$, and $1.7$ are $0.87$ and $0.62$ respectively, which are in close agreement with simulations. Similarly, in the absence of the direct link, $\rho^*_{nd}$ of $0.89$ and $0.68$ are indicated by (\ref{eq:OptRhoWSD}), which are in close agreement with simulations. We note that $\rho^*_{nd}>\rho^*$, as can be intuitively expected. It is clear that a) incremental relaying results in higher throughput than relay-less signalling from S to D, b) two-hop relaying that ignores the S-D link results in throughput that is quite poor as compared to direct S-D signalling without the relay, and c) a smaller $d_{sr}$ results in larger throughput, especially for large $\rho$ (note that $d_{sr}+d_{rd}=d_{sd}$ so that a smaller $d_{sr}$ implies a larger $d_{rd}$).
\par In Fig.\ref{fig:TauVsRsConf}, throughput is plotted versus $R_s$ for two different values of $I$. For each point, the optimum value of $\rho^{*}$ is computed and used. The superiority of incremental relaying over direct point-to-point transmission is apparent. Moreover, the gap between the two is higher for larger value of $I$. This happens because relayed signalling has a higher chance of non-outage for larger value of $I$.
It be observed that an optimum value of $R_s$ exists which maximizes throughput. It is apparent from ~\eqref{eq:tauIn}, that throughput is limited by $R_s$ when it is small, and by outage when $R_s$ is large. The optimum value needs to be obtained by numerical search.
\section{Conclusion} \label{sec:conclusion}
In this paper, we derived a closed-form expression for the throughput performance of an underlay two-hop network with a power-splitting based energy harvesting relay. We present a closed-form expression for the throughput maximizing power-splitting parameter.
\vspace{-0.5cm}
\section*{Appendix}
\vspace{-0.5cm}
\setcounter{equation}{25}
\begin{figure*}[t!]
\begin{IEEEeqnarray}{rcl}
p_3\Big|_{|h_{sr}|^2} &=& \frac{\lambda_{sp}}{\lambda_{rd}-\beta |h_{sr}|^2 \lambda_{sd}} \Big(\frac{\lambda_{rd} \lambda_{sd} \psi}{\lambda_{sd} \lambda_{sp} \psi+\lambda_{sp}^2}-\beta |h_{sr}|^2 \lambda_{sd} \Big(\frac{1}{\lambda_{sp}}-\beta |h_{sr}|^2 \Big(t \Big(\frac{1}{\beta |h_{sr}|^2 (\lambda_{sd} \psi+\lambda_{sp})+\lambda_{rp}}\nonumber\\
&&-\frac{1}{\beta |h_{sr}|^2 \lambda_{sp}+\lambda_{rd} \psi+\lambda_{rp}}\Big)+\frac{1}{\beta |h_{sr}|^2 \lambda_{sp}+\lambda_{rd} \psi}\Big)\Big)\Big)\label{eq:p3_hsr}
\end{IEEEeqnarray}
\vspace*{-0.6cm}
\hrulefill
\end{figure*}
\setcounter{equation}{27}
\begin{figure*}[t!]
{
\begin{IEEEeqnarray}{rcl}
p^{d_{rp}\rightarrow\infty}_{3}&=&\frac{\left({\lambda_{sr} \psi}/{\beta}\right)^2}{\lambda_{sp} \left(\frac{\lambda_{rd} \lambda_{sp}}{\beta \lambda_{sd}}+\frac{\lambda_{rd} \psi}{\beta}\right)} \left(\lambda_{sr} e^{\frac{\lambda_{rd} \lambda_{sr} \psi}{\beta \lambda_{sp}}} \left(E_1\left(\frac{\lambda_{rd} \lambda_{sr} \psi}{\beta \lambda_{sp}}\right)-E_1\left(\frac{\lambda_{rd} \psi}{\beta \lambda_{sp}} \left(\lambda_{sr}+\frac{\lambda_{sp} (1-\rho)}{\psi}\right)\right)\right)\right)\label{eq:p3_WithoutLRP}\\
&&-\frac{\lambda_{sr} \psi \left(1-e^{-\frac{\lambda_{rd} (1-\rho)}{\beta}}\right)}{\lambda_{sp} (1-\rho)+\lambda_{sr} \psi} -\frac{\lambda_{sp} \lambda_{sr}}{\frac{\lambda_{rd} \lambda_{sp}}{\beta \lambda_{sd}}+\frac{\lambda_{rd} \psi}{\beta}} \left(\frac{\lambda_{rd}}{\beta \lambda_{sd}}\right)^2 e^{-\frac{\lambda_{rd}}{\beta \lambda_{sd}} \left(\frac{\lambda_{sp} (1-\rho)}{\psi}+\lambda_{sr}\right)-\frac{\lambda_{rd} (1-\rho)}{\beta}}\nonumber\\
&&\times \left(\text{Ei}\left(\frac{\lambda_{rd}}{\beta \lambda_{sd}} \left(\lambda_{sr}+\frac{\lambda_{sp} (1-\rho)}{\psi}\right)\right)-\text{Ei}\left(\frac{\lambda_{rd}}{\beta \lambda_{sd}} \left(\lambda_{sr}+\frac{\lambda_{sp} (1-\rho)}{\psi}\right)+\frac{\lambda_{rd} (1-\rho)}{\beta}\right)\right)\nonumber
\end{IEEEeqnarray}}
\hrulefill
\vspace*{-0.4cm}
\end{figure*}
\setcounter{equation}{29}
\begin{figure*}[t!]
\begin{eqnarray}
\tau_{sim} = 0.5R_s\left(1-\left(\frac{1}{\left(1+\frac{\lambda_{sp}}{\lambda_{sd}\psi}\right)\left(1+\frac{\eta \lambda_{sp} \rho}{\psi \lambda_{rd} \lambda_{sr} }\right)}+\frac{1}{\left(\frac{\lambda_{rd}}{\eta \rho}+1\right) \left(\frac{ \lambda_{sp} (1-\rho)}{\psi \lambda_{sr} }+1\right)}+\frac{1}{\frac{\psi}{ \lambda_{sp}} \left(\lambda_{sd}+\frac{\lambda_{sr}}{1-\rho}\right)+1}\right)\right)\label{eq:TauApproxApp1}
\end{eqnarray}
\hrulefill
\vspace*{-0.4cm}
\end{figure*}
\setcounter{equation}{21}
\subsection{Derivation of $p_3$}\label{eq:p3Derivation}
In this Appendix, we derive an expression for $p_3$.
To this end, we first define $X$ as:
\begin{eqnarray}
X\triangleq \min\left(^hP, \frac{I}{|g_{rp}|^2}\right)|h_{rd}|^2 \label{eq:X}
\end{eqnarray}
Its Cumulative distribution function (CDF) conditioned on $^hP$ can be derived as:
\begin{equation}
\hspace{-0.04in}F_{X\big|_{^hP}}\hspace*{-0.3cm}(x)=1-e^{\frac{\lambda_{rd}x}{\beta\,P_s\,|h_{sr}|^2}}\Big(1-e^{-\frac{\lambda_{rp}I}{\beta\,P_s\,|h_{sr}|^2}}\Big(\frac{\frac{\lambda_{rd} x}{I\lambda_{rp}}}{1+\frac{\lambda_{rd}x}{I\lambda_{rp}}}\Big)\Big)\label{eq:CDFofX}
\end{equation}
From \eqref{eq:q}, $p_3$ can be expressed as:
{\small\begin{IEEEeqnarray*}{rcl}
p_3=\Pr\Big(\frac{P_s |h_{sd}|^2}{N_o}+\frac{X\big|_{P_s|h_{sr}|^2}}{N_o}\leq \gamma_{th},\frac{(1-\rho) P_s |h_{sr}|^2}{N_o} > \gamma_{th}\Big).\nonumber
\end{IEEEeqnarray*}}
We derive $p_3$ by successive averaging over each random variable and keeping other r.vs. in terms of condition. We first average w.r.t. $X$ by using \eqref{eq:CDFofX} to get:
\begin{IEEEeqnarray*}{rcl}
\hspace{-0.1in}p_3&=&\E_{\C_1 }\left[ F_{X\big|{^hP}}\left(\gamma_{th}-\frac{P_s |h_{sd}|^2}{N_o}\right)\right]\nonumber\\
&=&\E_{\C_1 }\Big[1-e^{-\frac{\lambda_{rd}}{\beta\,P_s\,|h_{sr}|^2}\left(\gamma_{th}-\frac{P_s |h_{sd}|^2}{N_o}\right)}\Big(1-e^{-\frac{\lambda_{rp}I}{\beta\,P_s\,|h_{sr}|^2}}\nonumber\\
&&\hspace{0.75in}\times\Big(1-\frac{1}{1+\frac{\lambda_{rd}}{I\lambda_{rp}}\left(\gamma_{th}-\frac{P_s |h_{sd}|^2}{N_o}\right)}\Big)\Big)\Big]\,,
\end{IEEEeqnarray*}
where the condition $\C_1=\{\frac{(1-\rho) P_s |h_{sr}|^2}{N_o}>\gamma_{th},\, \frac{P_s |h_{sd}|^2}{N_o}\leq\gamma_{th}\}$.
Unfortunately, averaging the above with respect to random variables $|h_{sr}|^2$ and $|g_{sp}|^2$ results in intractable expressions. We need to make use of the following fact: $1\gg\frac{\lambda_{rd}}{I\lambda_{rp}}\left(\gamma_{th}N_o-P_s |h_{sd}|^2\right)$ (since $P_s |h_{sd}|^2/N_o <\gamma_{th}$ and $I\lambda_{rp}\gg\lambda_{rd}$). Hence for tractability, $\frac{\lambda_{rd}}{I\lambda_{rp}}\left(\gamma_{th}N_o-P_s |h_{sd}|^2\right)$ can be replaced by its mean i.e. $\frac{\lambda_{rd}}{I\lambda_{rp}}\left(\gamma_{th}N_o-\E_{P_s|h_{sd}|^2\leq \gamma_{th}}(P_s |h_{sd}|^2)\right)$. $\E_{P_s|h_{sd}|^2\leq \gamma_{th}}(P_s |h_{sd}|^2)$ can be derived as:
{\small\begin{equation}
\E_{\C_2}(P_s |h_{sd}|^2)=\frac{I \lambda_{sp}}{\lambda_{sd}} \Big(\log \Big(\frac{\gamma_{th} \lambda_{sd} N_o}{I \lambda_{sp}}+1\Big)+\frac{\frac{\gamma_{th} \lambda_{sd} N_o}{I \lambda_{sp}}}{\frac{\gamma_{th} \lambda_{sd} N_o}{I \lambda_{sp}}+1}\Big),\label{eq:EC2}
\end{equation}}
where $\C_2=\{P_s|h_{sd}|^2\leq N_o \gamma_{th}\}$.
Now, $p_3$ can further represented in approximated form as:
\begin{IEEEeqnarray}{rcl}
p_3&\approx&\E_{\C_1}\Big[1+t e^{-\frac{\lambda_{rd} (\gamma_{th} N_o-|h_{sd}|^2 P_s)}{\beta |h_{sr}|^2 P_s}-\frac{I \lambda_{rp}}{\beta |h_{sr}|^2 P_s}}\nonumber\\
&&\hspace*{2cm}-e^{-\frac{\lambda_{rd} (\gamma_{th} N_o-|h_{sd}|^2 P_s)}{\beta |h_{sr}|^2 P_s}}\Big], \label{eq:p3hSR}
\end{IEEEeqnarray}
where $t=1-1/({1+\frac{\lambda_{rd}}{I\lambda_{rp}}\left(\gamma_{th} N_o-\E_{\C_2}\left[P_s|h_{sd}|^2\right]\right)})$. The above approximation is tight for $\lambda_{rd}<<I\lambda_{rp}$. This is true for general system settings in underlay-CRN.
\par Now averaging the above expression over $|h_{sr}|^{2}$ and using the expression for $P_s$ in \eqref{eq:PsOnlyInterference}, results in the following expression for $p_3$:
\begin{IEEEeqnarray*}{rcl}
&&p_3\approx\E_{\C_3}\Big[\frac{\lambda_{rd} \left(1-e^{-\frac{\gamma_{th} \lambda_{sd} N_o}{P_s}}\right)}{\lambda_{rd}-\beta |h_{sr}|^2 \lambda_{sd}}-\frac{(\beta |h_{sr}|^2 \lambda_{sd})}{\lambda_{rd}-\beta |h_{sr}|^2 \lambda_{sd}}\Big(1+\nonumber \\
&&t e^{-\frac{\gamma_{th} \lambda_{rd} N_o+I \lambda_{rp}}{\beta |h_{sr}|^2 P_s}}-t e^{-\frac{\beta \gamma_{th} |h_{sr}|^2 \lambda_{sd} N_o+I \lambda_{rp}}{\beta |h_{sr}|^2 P_s}}-e^{-\frac{\gamma_{th} \lambda_{rd} N_o}{\beta |h_{sr}|^2 P_s}}\Big)\Big]\,,\nonumber
\end{IEEEeqnarray*}
where condition $\C_3=\{\frac{(1-\rho) |h_{sr}|^2}{\psi}>|g_{sp}|^2\}$. Now let $\psi = \frac{\gamma_{th}}{I/N_o}$ , $p_3$ becomes:
\begin{IEEEeqnarray*}{rcl}
p_3\approx&\E_{\C_3}&\Big[\frac{\lambda_{rd} \left(1-e^{-{|g_{sp}|^2\gamma_{th} \lambda_{sd} \psi }}\right)}{\lambda_{rd}-\beta |h_{sr}|^2 \lambda_{sd}}-\frac{(\beta |h_{sr}|^2 \lambda_{sd})}{\lambda_{rd}-\beta |h_{sr}|^2 \lambda_{sd}}\Big(1\nonumber \\
&&+t e^{-\frac{|g_{sp}|^2}{\beta |h_{sr}|^2}(\lambda_{rd}\psi+\lambda_{rp})}-t e^{-|g_{sp}|^2\left(\psi\lambda_{sd}+\frac{\lambda_{rp}}{\beta|h_{sr}|^2}\right)}\nonumber\\
&&-e^{-\frac{|g_{sp}|^2 \lambda_{rd} \psi}{\beta |h_{sr}|^2 }}\Big)\Big]\nonumber
\end{IEEEeqnarray*}
\par Now averaging the above equation over $|g_{sp}|^2$ and then using some straightforward manipulations, the resultant expression is presented in \eqref{eq:p3_hsr}. From \eqref{eq:p3_hsr}, $p_3$ is now expressed as:
\setcounter{equation}{26}
\begin{equation}
p_3 = \int_{0}^{\infty} p_3\big|_{|h_{sr}|^2} \lambda_{sr} e^{-\lambda_{sr} |h_{sr}|^2} d|h_{sr}|^2
\end{equation}
The above integral can be simplified using the integral presented in \cite[5.1.1]{Abramowitz1964}:
{\begin{IEEEeqnarray*}{rcl}
\int_r^s \frac{e^{-p x}}{q x+1} \, dx = \frac{e^{\frac{p}{q}}}{q} \left(E_1\left(\frac{(q r +1)p}{q}\right)-E_1\left(\frac{(q s +1)p}{q}\right)\right)\nonumber.
\end{IEEEeqnarray*}}
We omit the manipulations due to space limitations and present the approximated $p_3$ as in \eqref{eq:p3_Final} (top of page-4).
\subsection{Derivation of $\tau_{sim}$} \label{AppB:ApproxP3}
\par In the expression for $p_3$ in \eqref{eq:p3_Final}, $t=1-1/({1+\frac{\lambda_{rd}}{I\lambda_{rp}}\left(\gamma_{th} N_o-\E_{\C_2}\left[P_s|h_{sd}|^2\right]\right)})$ as defined in \eqref{eq:p3hSR}, with $\E_{\C_2}\left[P_s|h_{sd}|^2\right]$ given by \eqref{eq:EC2}. Clearly, $t=0$ when $\lambda_{rp}\rightarrow\infty$, which enables us to simplify $p_3$. Resultant expression is given in \eqref{eq:p3_WithoutLRP} (second equation on the top of the next page).
\par Further simplification is possible when $\frac{I\lambda_{sp}}{N_o}\gg\gamma_{th}$, which is the commonly encountered situation. In this case, the arguments of $e^{-x}\text{Ei}(x)$ and $e^{x} E_1(x)$ terms in $p_3$ increase and decrease respectively. Using the fact that $e^{-x}\text{Ei}(x)\rightarrow 0$ for $x\rightarrow\infty$ (here $x\propto\frac{\lambda_{sp}}{\psi}$), the term associated with $\text{Ei}(x)$ vanishes from the expression for $p_3$. We further use the fact that $E_1(x)\gg E_1((x+A))$, for $A>0$ when $A$ is very large (as it is in this case since $(1-\rho) I\lambda_{sp}/(\gamma_{th} N_o)>0$) . The term with $e^x E_1((x+A))$ can be neglected in the high SNR region. From~\eqref{eq:p3_WithoutLRP}, the resultant approximated expression can be written as:
\setcounter{equation}{28}
{\small\begin{align}
p^{d_{rp}\rightarrow\infty}_{3}\overset{A\gg0}{\approx}&\left(\frac{\psi \lambda_{rd} }{\beta \lambda_{sp}}\right)^2\frac{(\lambda_{sp} \lambda_{sr} )}{\frac{\psi \lambda_{rd} }{\beta}+\frac{\lambda_{rd} \lambda_{sp}}{\beta \lambda_{sd}}}e^{\frac{\psi \lambda_{rd} \lambda_{sr} }{\beta \lambda_{sp}}} E_1\left(\frac{\psi \lambda_{rd} \lambda_{sr} }{\beta \lambda_{sp}}\right)\nonumber\\
&+\frac{\lambda_{sr} }{\frac{\lambda_{sp} (1-\rho)}{\psi }+\lambda_{sr} } \left( e^{-\frac{ \lambda_{rd} (1-\rho)}{\beta}}-1\right). \label{eq:p3_HSNR}
\end{align}}
By using the following very tight lower and upper bounds $e^{x}E_1(x)\geq ({1+x})^{-1}$ and
$e^{-{\lambda_{rd} (1-\rho)}/{\rho}}\leq({1+{\lambda_{rd} (1-\rho)}/{\rho}})^{-1}$
respectively into \eqref{eq:p3_HSNR}, and from \eqref{eq:p1Final}, \eqref{eq:p2Final} and \eqref{eq:q2}, $\tau$ can be approximated as $\tau_{sim}$ in \eqref{eq:TauApproxApp1}.
\par It is worth noting that the above approximation is valid for $\frac{I\lambda_{sp}}{N_o}\gg\gamma_{th}$. However, it continues to follow the throughput $\tau$ in other cases. To get a closed form expression, we further approximate the above by utilizing the fact $\frac{\lambda_{rd}}{\eta\rho}\gg 1-\frac{\lambda_{rd}}{\eta}$ (since $\eta, \rho\leq 1$) and $\frac{I \lambda_{sp} (1-\rho)}{\gamma_{th} N_o\lambda_{sr}}\gg1$ except when $\rho\approx1$ and represented in \eqref{eq:TauSimplified} (top of the page-4). Please note that $\rho\approx1$ is unlikely, since it results in outage of the relayed link.
\vspace*{-0.15cm}
\bibliographystyle{ieeetr}
| 378a002f58f7863968f41fa933f15fffa2edb51b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
Recent progress in unmanned aerial vehicle (UAV) technology has led to exponential growth in the use of drones with onboard sensing. Originally designed for military applications, camera-equipped UAVs are now commonplace in domains such as commercial surveillance, photography, disaster management, product delivery, and mapping \cite{colomina2014unmanned}. An essential step toward better utilization of aerial videos and autonomous drones is real-time localization of the vehicle and accurate mapping of the observed world. Localization relying solely on onboard inertial and GPS measurements, however, cannot achieve pixel-level accuracy due to error accumulation, latency, and relatively coarse precision; furthermore, onboard sensing can be unreliable in GPS-denied environments. Visual simultaneous localization and mapping (SLAM) attempts to address these challenges using camera images to augment or replace other sensors.
Our proposed method, nicknamed \textit{Global-Local Airborne Mapping} (GLAM), approaches large-scale visual SLAM by partitioning the camera's video stream into small local \textit{submaps} created using point features, epipolar geometry between keyframes, point triangulation, and incremental bundle adjustment (BA). Submaps are aligned globally using a graph-based least squares optimization that minimizes the distance between corresponding 3D points. Small temporal overlap ensures that sequential submaps observe common scene content and hence provides correspondences for alignment, while restricting submap size to a small number of frames mitigates drift accumulation. Associations between non-overlapping submaps, for instance due to loop closures, are detected via fast bag-of-visual-words recognition. This paper demonstrates that such a system can run in near real-time on videos of the scale of a hundred thousand frames.
\section{Related Work}
Monocular SLAM is a long-researched subject that has seen evolution from filter based \cite{thrun2005probabilistic, davison2007monoslam} to keyframe based \cite{strasdat2012visual, resch2015scalablesfm} approaches.
PTAM \cite{klein07parallel}, originally developed for augmented reality applications, was one of the first widely used and practical real-time SLAM systems, but was limitated in terms of scale and robustness. Subsequent research improved upon this point feature and bundle adjustment methods \cite{strasdat2010scale, resch2015scalablesfm}. More sophisticated monocular SLAM approaches like DTAM \cite{newcombe2011dtam} and LSD-SLAM \cite{engel12iros} perform semi-dense reconstruction by optimizing directly on image intensities rather than discrete features. SVO \cite{Forster2014ICRA} is a semi-direct approach monocular SLAM approach designed specifically for Micro Aerial Vehicles (MAVs); although SVO operates at very high frame rates, it does not perform loop closure and has been evaluated only on small datasets. ORB-SLAM \cite{mur2015orb} is a recent feature-based approach that has shown good performance on a wide variety of datasets, but largely untested for large aerial datasets.
City-scale 3D reconstruction has received some attention in the structure-from-motion research community \cite{musialski2013survey}. Pollefeys \textit{et al.} \cite{pollefeys2008detailed} reconstruct parts of a city from a hundred thousand frames using INS and metadata to simplify the computation. Agarwal \textit{et al.} \cite{agarwal2011building} and Heinly \textit{et al.} \cite{heinly2015_reconstructing_the_world} perform city and world scale reconstruction respectively from large collections of photographs utilizing cloud computing. Google Earth provides 3D models of a few selected cities, but the technology behind its reconstructions is largely unknown. Leotta \textit{et al.}\cite{leotta2016open} have published open-source software for aerial SfM focused on videos of a scale of a thousand frames but do not address scaling beyond that.
Real-time reconstruction of a city from aerial videos using purely visual data has remained largely unexplored in the literature: traditional visual SLAM approaches do not address the challenges of very long image sequences and large-scale maps, while structure-from-motion methods require substantial offline computation and/or information from additional sensors. Our system incorporates elements of both visual SLAM and large-scale 3D reconstruction, demonstrating fast and accurate city-scale mapping on aerial videos nearly 90,000 frames in length on a single consumer grade computer.
\begin{figure*}
\begin{center}
\includegraphics[width=1.0\linewidth]{GLAM_WACV.png}
\end{center}
\caption{\textbf{GLAM} system overview, showing the main steps performed by submap creation and submap alignment threads.}
\label{fig:overview}
\end{figure*}
\section{Proposed Approach}
Like many existing real-time SLAM systems, we adopt a multi-threaded strategy for computational efficiency, partitioning the problem into tasks that run in parallel. However, rather than defining the tasks as \textit{tracking} and \textit{mapping} \cite{klein07parallel, Forster2014ICRA, mur2015orb}, we define tasks as \textit{local submap creation} and \textit{global submap alignment}. Figure \ref{fig:overview} shows an overview of GLAM's building blocks.
The submap creation thread operates similarly to previous point features and keyframe based approaches \cite{klein07parallel, mur2015orb, strasdat2010scale} that extract point features from video frames, determine relative pose, triangulate points, and perform incremental bundle adjustment. Building the Hessian matrix for incremental bundle adjustment is the most computationally expensive step \cite{sibley2009adaptive}: as the number of frames processed increases, the cost of building the Hessian at every BA iteration becomes prohibitive for real-time operation. We therefore limit each submap to a small number of keyframes in order to maintain a relatively consistent and bounded processing rate.
Each submap consists of a set of 3D landmarks. The submap alignment thread creates and optimizes a pose-graph to determine the set of 7-DoF similarity transformations, one per submap, that minimizes the distances between corresponding landmarks. In contrast to existing approaches that close loops by creating a global database of keyframes (e.g.\, \cite{mur2015orb}), our approach closes loops by building a visual vocabulary of submaps.
GLAM has been evaluated on both synthetic and real datasets. The former are created by simulating drone flight trajectories over an aerial LIDAR point cloud \cite{rigis}, while the latter consists of a large continuous aerial video (labeled ``Downtown'') with nearly $90,000$ 1-megapixel frames captured at $30$ fps.
The main contribution of this paper is a novel visual SLAM pipeline that (1) partitions work into parallel threads of fast local submap creation and large-scale global submap alignment; (2) has the ability to close large loops (3) reconstructs accurate city-size maps from aerial videos; and (4) operates in near real-time at the scale of a hundred thousand images.
\subsection{Local Submap Creation}
The following sections describe keyframe-based submap creation in greater detail. The central idea of keyframe-based SLAM is to use only frames that have sufficiently distinct information for 3D reconstruction. Each video frame is processed in sequence, with keyframes created periodically to reduce redundancy and improve efficiency---only the keyframes are used in intra-submap bundle adjustment.
\textbf{Feature Extraction.} Point features are extracted from each video frame. We use SIFT features \cite{lowe2004distinctive} throughout our system due to their proven robustness to viewpoint, scale and orientation changes. SIFT extraction and matching can be slow, so we use a GPU implementation \cite{wu2007siftgpu} for greater efficiency.
\textbf{Tracking.}
Existing 3D landmarks are projected into the current frame using the pose of the previous frame, since inter-frame motion is assumed to be small. Points that fall outside the frame or that were originally viewed from a substantially different angle are removed. The SIFT descriptors of the remaining points are then matched to those of the current image features to obtain initial 2D-3D correspondences. The 3D pose of the current frame is then estimated using Perspective-n-Point (PnP) localization \cite{lepetit2009epnp} embedded in a RANSAC outer loop.
PnP can fail if the video is discontinuous or if the viewpoint has changed drastically between frames. If the number of PnP inliers is too low, the submap is terminated at the previous frame and a new submap is initiated at the current frame.
\textbf{Keyframes.}
As new parts of the world come into view, the number of PnP inliers continually decreases. When this number falls below a threshold $\tau_{resection}$, the system adds a new keyframe and triangulates new 3D points. Selection of $\tau_{resection}$ is key in ensuring a balance between speed and stability: a high value causes frequent keyframe additions and thus slows the overall system, while a low value reduces visual overlap between keyframes and may lead to unstable or failed bundle adjustment.
To improve stability further, an additional keyframe is chosen halfway between the previous keyframe and the current frame. This \textit{middle} keyframe's pose is initialized to its PnP estimate, and its 2D PnP inliers are added as observations of their counterpart 3D landmarks for BA. Geometrically consistent 2D matches are obtained between the middle keyframe and the current frame via RANSAC-based essential matrix estimation, and these matches are triangulated using the PnP pose estimates to form new 3D landmarks and corresponding 2D observations for BA. The SIFT descriptor from the \textit{middle} keyframe is assigned to each landmark.
Since the triangulation of new landmarks does not take into account the landmarks that have already been constructed, a subset may be duplicated. Therefore, any new landmark whose descriptor matches any existing landmark is removed. The current frame is finally added as a keyframe and its pose and observations are added to BA.
This selection strategy depends on neither physical nor temporal distance between keyframes; instead, keyframes are added only when sufficient new visual information is available, allowing the system to process videos acquired at arbitrary speed. In the Downtown dataset, for instance, the system selected keyframes spaced roughly 25 to 100 frames apart.
\begin{figure*}
\setlength{\tabcolsep}{1pt}
\begin{tabular}{llll}
\includegraphics[width=4.3cm]{005000.jpg}
\includegraphics[width=4.3cm]{022000.jpg}
\includegraphics[width=4.3cm]{024000.jpg}
\includegraphics[width=4.3cm]{079000.jpg}
\\
\includegraphics[width=4.3cm]{037000.jpg}
\includegraphics[width=4.3cm]{049500.jpg}
\includegraphics[width=4.3cm]{066000.jpg}
\includegraphics[width=4.3cm]{054000.jpg}
\\
\end{tabular}
\caption{\textbf{Example images from the Downtown dataset.} A few scenes from the dataset showing challenging views with illumination changes, planar regions, high 3D relief, and water.}
\label{fig:frames}
\end{figure*}
\textbf{Incremental Bundle Adjustment.}
Similar to PTAM \cite{klein07parallel} and ORB-SLAM \cite{mur2015orb}, GLAM runs BA every time a new keyframe is added, optimizing the poses of all keyframes and the positions of all landmarks while holding fixed the intrinsic camera parameters. Different libraries including g2o \cite{kuemmerle11icra} and ceres-solver \cite{agarwal2013ceres} were evaluated, but pba \cite{wu2011multicore} performed best due to its specialization for BA problems.
\textbf{Bootstrapping}. When a new submap is initiated, no keyframes or 3D landmarks yet exist, so a strategy different from the steady-state process described above must be employed to establish initial geometry. The first frame of the submap becomes the first keyframe, with its pose fixed at the origin. Features from subsequent frames are matched against those of the first keyframe to find epipolar geometry via RANSAC-based essential matrix estimation \cite{Hartley00}. The current frame becomes the second keyframe if the number of RANSAC inliers falls below $\tau_{stereo}$ or if the average triangulation angle exceeds $\alpha_{stereo}$. (for the experiments described in this paper, we used $\tau_{stereo} = 1000$ and $\alpha_{stereo} = 30^{\circ}$). The inlier correspondences are triangulated to form the initial 3D landmarks.
\textbf{Completion.} A submap is sucessfully completed when the number of keyframes exceeds a threshold. We found that 20 keyframes were generally sufficient to form a stable reconstruction and minimize internal loop closures. Upon completion, all frames are re-localized to the final landmark set via PnP, which requires only that 3D-2D matches and not features to be recomputed. The next submap is initiated to have a degree of overlap (in our experiments, 10\%) with the current submap so that the two share a set of 3D points for alignment.
\textbf{Outlier filtering.}
Incorrectly matched point features lead to outliers that can dramatically affect bundle adjustment accuracy. Our system therefore filters outliers at several stages. First, after every bundle adjustment step, any point whose reprojection error exceeds a threshold is removed. Some incorrect matches may exhibit small reprojection error, such as when triangulation rays are nearly parallel and the reconstructed landmark is very far from the rest of the scene. To remove these points, the \textit{k} (30) nearest neighbours of each point are computed using a kd-tree, and a point is removed if its average distance from its neighbours is more than 2 standard deviations. Finally, upon submap completion, any point with large reprojection error in all frames is removed.
\textbf{Focal length estimation.}
Use of accurate camera intrinsics is instrumental in full-scale reconstruction, since there is no single rectifying transformation for incorrect intrinsics. The published focal length of 1800 for the lens used to collect the Downtown dataset was imprecise and led to reconstruction failures. We therefore included per-keyframe focal length in PBA optimization, observing that the median over all frames' estimated focal length converged to 1751 as the number of frames increased, as shown in Figure \ref{fig:focal}.
\begin{figure}[t]
\begin{tabular}{cc}
\includegraphics[width=8.2cm]{focal_orig.png}\\
(a)\\
\includegraphics[width=8.2cm]{focal2_orig.png}\\
(b)\\
\end{tabular}
\caption{\textbf{Estimation of focal length and radial distortion.} (a) Focal length of a new keyframe after it is added to bundle adjustment. (b) Median focal length of all keyframes after adding a new keyframe to bundle adjustment.}
\label{fig:focal}
\end{figure}
\subsection{Global Submap Alignment}
As local submaps are created, a parallel thread uses them as the unit of processing for full map reconstruction. This thread discovers correspondences between submaps, including loop closures, and optimizes pose and structure globally and efficiently.
\textbf{Submap matching.}
Although temporally adjacent submaps are known to be associated with one another by construction, it is required to discover visual links between submaps that view common scene content, regardless of the time at which they were created. This forms a connectivity graph in which submaps constitute nodes and commonly observed points constitute edges.
A naive implementation would identify visual links through brute-force search over every possible pair of submaps. This requires $O(n^2)$ time for $n$ submaps, making the process computationally intractable as $n$ grows. To form a connectivity graph efficiently, we employ a bag-of-words technique. A vocabulary tree is created offline using randomly-sampled SIFT descriptors drawn from aerial images; a series of k-means clustering operations partition the descriptors into a tree of visual words \cite{nister2006scalable}. As each submap is completed, its landmarks' SIFT descriptors are added to a database formed over the tree, and visual words are incrementally associated with the submap via an inverted index. The set of landmark descriptors in this submap is then used to query the database, which returns a ranked set of matching submaps and their weighted match scores. This mechanism reduces the complexity of submap matching from quadratic to linear time, as shown in Figure \ref{fig:realanalysis}.
The result is a graph whose edges represent potential visual matches based on an unordered bag of indexed visual words with no geometric constraints. Some graph edges may be incorrect, sharing a number of similar-looking features but not actually viewing common areas of the scene (e.g.\, due to repeated urban structures such as windows). The geometric consistency of correspondences is verified by estimating a closed form similarity transform between 3D landmarks in a RANSAC loop, accounting for both pose and scale differences. Edges with too few inliers are removed, as are all outlier landmark correspondences.
\begin{figure}
\begin{center}
{\includegraphics[width=6.0cm]{submap-graph2.png}}
\end{center}
\caption{\textbf{Representation of a submap pose-graph.} The nodes $S_1$, $S_2$, $S_3$,... represent a 7 DoF Sim(3) transformation for each submap. The nodes $l_1$, $l_2$, $l_3$,... represent a location in $\mathbb{R}^3$ for each 3D landmark point. An edge between a submap and 3D point represents the observation of the point in the submap. The nodes $p_1$, $p_2$, $p_3$,... are scale prior nodes for every submap that prevent collapse to zero scale during optimization. }
\label{fig:submapgraph}
\end{figure}
\textbf{Submap pose-graph}
After establishing temporal and spatial correspondences between submaps, the next step is to create a pose-graph and optimize the pose of the submaps such that the 3D distance between corresponding points is minimized (Figure \ref{fig:submapgraph}). Scale-drift \cite{strasdat2010scale} within submaps is addressed implicitly by the keyframe addition strategy, while scale-drift across submaps is adressed by using a 7 DoF similarity transform to represent each submap's pose.
The distance between corresponding 3D points of the submaps is minimized by using non-linear Gauss-Newton (GN) optimization \cite{Hartley00}, which involves finding a Jacobian w.r.t. to all free parameters. In this case, the 7 DoF pose of a submap and the $\mathbb{R}^3$ location of a landmark are the free parameters of the cost function. Differentiating the cost function w.r.t. the Lie group elements \textit{(R,s,t)} yields the Lie algebra space sim(3) \cite{eade2014lie}. sim(3) is a 7 dimensional vector space ($\omega$, $\sigma$, $\mu$), where $\omega$, $\sigma$ and $\mu$ respresent the Lie algebra components of rotation, translation and scale respectively.
Given a set of submaps $S$ and landmarks $L$, a 7 DoF similarity transform $(R, s, t)$ is associated with every submap $i$ and represented in the pose-graph with a Sim(3) node. A world point $X_j$ is represented with a $\mathbb{R}^3$ node in the pose-graph. If submap $i$ observes 3D point $X_j$, it is represented by edge $x_{ij}$ in the pose-graph.
Scale is constrained by multiplication of a small constant $\lambda_{a}$ ($<$1) with the Lie algebra of scale, $\mu_i$.
The scale prior prevents the solver from converging to a trivial solution, i.e zero scale.
The resulting cost function is:
\begin{equation}
f(S,L) = \sum_{i \in S} \sum_{j \in L} ||s_i R_i x_{ij} + t_i - X_j||^2 + \sum_{i \in S}|| \lambda_{a} \mu_{i} ||^2
\end{equation}
Initially, the scale prior used was $\sum_{i \in S} ||c - s_{i}||^2$, where c was a small constant, typically 1. When added to the remaining cost function, this scale prior prevented the trivial solution by evaluating to a non-zero value. However, it is an invalid operation since it is subtracting the Lie group $s_{i}$ from 1, and addition / subtraction operations are undefined for Lie groups. Hence, a valid scale prior instead constraints the scale by multiplying the Lie Algebra of scale $\mu$ with a small constant $\lambda$. This scale prior has a similar effect on the overall cost function at $\mu$ equals zero. It was also observed experimentally that a cost function with the former scale prior resulted in submaps that didn't align properly to each other but the latter scale prior fixed those alignment issues.
Every step of a GN optimization updates the sim(3) vector of each submap vertex with a $\upsilon_{up}$ vector. If $\upsilon$ and $G$ are the Lie algebra and Lie group of the current state of a Sim(3) vertex, then the vector update $\upsilon_{up}$ is applied as follows:
\begin{align}
G_{up} &= exp(\upsilon_{up}) \\
\upsilon &= log(G_{up} * exp(\upsilon)) \\
G &= G_{up} * G
\end{align}
The g2o \cite{kuemmerle11icra} framework is used to form and optimize the factor graph via Levenberg-Marquardt (LM), with robustness added via Huber loss \cite{Hartley00}. Preconditioned conjugate gradient is used to efficiently solve the sparse linear system of equations arising at each step of LM, and analytic Jacobians were dervied to further speed up the computation. and are as follows:
\begin{align}
\frac{\partial{y}}{\partial{R}} = -y_{\times}, \hspace{2mm}
\frac{\partial{y}}{\partial{t}} = I, \hspace{2mm}
\frac{\partial{y}}{\partial{s}} = y
\end{align}
where,
\begin{align}
y = (y_1, y_2, y_3) = f(S,L)
\\
y_{\times} =
\begin{bmatrix}
0 & -y_3 & y_2 \\
y_3 & 0 & -y_1 \\
-y_2 & y1 & 0
\end{bmatrix}
\end{align}
\begin{figure*}
\begin{center}
\begin{tabular}{cc}
\includegraphics[width=.5\textwidth]{google-earth.png}&
\includegraphics[width=.5\textwidth]{google-earth2.png}\\(a) & (b)\\
\includegraphics[width=.4\textwidth]{synthRI.png}&
\includegraphics[width=.4\textwidth]{synth1_3.png}
\\(c) & (d)\\
\end{tabular}
\end{center}
\caption{\textbf{Synthetic datasets.} (a) Two camera trajectories and viewing direction trajectories drawn in green and red respectively in Google Earth. (b) LIDAR points and the synthetic camera and viewing direction trajectories of synthetic dataset synth1. (c), (d) 3D reconstruction of synth1 using GLAM. The cameras from each submap are represented with a different colour. Results from datasets synth2 and synth3 are shown in the supplementary material.}
\label{fig:synthfig}
\end{figure*}
\section{Experiments}
We evaluated GLAM on both synthetic and real datasets. Our experiments demonstrate centimeter level accuracy on synthetic datasets, and accuracy comparable to VisualSFM \cite{wu2011visualsfm} on real datasets, while providing a much faster run-time.
\subsection{Synthetic Data}
We created synthetic datasets using Google Earth to simulate camera trajectories and using a LIDAR point cloud \cite{rigis} to serve as ground truth for simulated image projections. The results of 3D reconstruction on three such datasets is shown in Figure \ref{fig:synthfig} and summarized in Table \ref{tab:synthetic}.
\begin{table}[!htb]
\begin{center}
\begin{tabular}{|l|c|c|c|c| }
\hline
& Frames & Submaps & Points & RMSE(cm)\\
\hline
synth1 & 1100 & 14 & 239727 & 0.3 \\ \hline
synth2 & 3408 & 15 & 328337 & 3.4 \\ \hline
synth3 & 40000 & 53 & 505000 & 0.6 \\
\hline
\end{tabular}
\end{center}
\caption{\textbf{Results on synthetic datasets.} Comparison of reconstructed maps to ground truth demonstrates errors on the order of centimeters and millimeters. The number of points in synth3 was made intentionally smaller than those in synth1 and synth2. Figures of the reconstruction from the three synthetic datasets can be found in the supplmentary material.}
\label{tab:synthetic}
\end{table}
\begin{figure*}
\setlength{\tabcolsep}{.16667em}
\begin{tabular}{ccc}
\includegraphics[width=5.7cm]{frame_BA_submap.png}&
\includegraphics[width=5.7cm]{results_plot2_2.png}&
\includegraphics[width=5.7cm]{results_plot3.png}\\
(a)&(b)&(c)
\end{tabular}
\caption{\textbf{Analysis of the system on the Downtown dataset.} (a) The final bundle adjustment reprojection error within each submap (blue) remains bounded as new submaps are created every few thousand frames (red). (b) Submap creation time (green) also remains bounded, while submap alignment time (red) increases linearly but maintains acceptable latency for real-time operation. Spikes in the red curve are due to convergence failure when the graph has multiple disconnected components that are resolved later when missing correspondences are established. (c) At 88,000 frames, the total number of 3D points in the map are 1.7 million but the number of vertices and edges in the submap pose-graph are just 64k and 146k respectively, which allows an optimization of the full graph in real-time. A naive method that does incremental bundle adjustment with g2o on all the points in the dataset would have to potentially deal with BA on 1.7 million points, which would make each step of bundle adjustment infeasible in real-time and the required memory would potentially not fit on a single consumer-grade machine.}
\label{fig:realanalysis}
\end{figure*}
\subsection{Real Data}
The Downtown dataset (Figure \ref{fig:frames}) is a video of 88,100 frames shot over Providence, USA. We applied GLAM to the entire video; Figure \ref{fig:realanalysis} illustrates runtime performance and scale, and Figure \ref{fig:realresults} visualizes the reconstruction results. The average run-time for submap creation is 5.8 fps, while the latency of the submap alignment and reconstruction thread is typically a few seconds.
We registered the landmark cloud reconstructed in the first 5,000 frames to a LIDAR point cloud of the same area using ICP \cite{besl1992method} to quantify accuracy with respect to ground truth. We also compared performance with VisualSFM \cite{wu2011visualsfm}, run on every $10^{th}$ frame of the first 5,000 frames (Table \ref{tab:realresults}).
\begin{figure*}
\begin{tabular}{ccc}
\includegraphics[width=8.2cm]{downtown_cameras.png}&
\includegraphics[width=8.2cm]{downtown_points4.png}\\
(a)&(b)\\
\includegraphics[width=8.2cm, angle=0]{downtown_points2.png}&
\includegraphics[width=8.2cm]{earth2.jpg}\\
(c)&(d)
\end{tabular}
\caption{ \textbf{Final results.} (a), (b) GLAM created 93 submaps for the full Downtown sequence, but 6 submaps failed in early stages due to bad frames. The above figure shows the 87 successful submaps aligned over the Downtown dataset. The cameras of each submap are represented with a different color. (c) The reconstructed 3D point cloud. (d) Google Earth image of Providence Downtown.}
\label{fig:realresults}
\end{figure*}
\begin{table}[!htb]
\begin{center}
\begin{tabular}{|l|c|c| }
\hline
& RMSE & Runtime\\
\hline
Visual SFM & \textbf{2.88 m} & 3.96 hrs \\ \hline
GLAM & 2.93 m & \textbf{11.79 min} \\ \hline
\end{tabular}
\label{tab:realresults}
\end{center}
\caption{\textbf{Comparison with Visual SFM.} ICP results on the first 5k frames of Downtown datset. GLAM is much faster than Visual SFM with comparable accuracy. Note that the LIDAR points are quantized to one meter spacing.}
\end{table}
\section{Conclusion}
This paper introduces a new visual SLAM system that can process long videos in near real-time on a single computer. Local submaps are constructed using incremental bundle adjustment over keyframes, and the global scene is reconstructed using factor graph optimization over submaps rather than keyframes, which allows closure of large loops. Submap creation has constant runtime complexity throughout, while submap alignment time increases linearly at a low rate, indicating that the system can be run on even larger datasets while maintaining acceptable latency.
The system's main performance bottleneck is currently SIFT feature extraction and matching. Future work will focus on improving run time by using binary descriptors. We also plan to investigate reducing the total number of constraints, incorporation of global averaging techniques, testing the approach on even larger datasets, and applying the system to terrestrial videos.
\section{Acknowledgements}
The authors would like to thank Vishal Jain his helpful advice and guidance on different parts of the project. This research is based upon work partly supported by Air Force Research Laboratory(AFRL) under contract number FA8650-14-C-1826. This document is approved for public release via 88ABW-2017-2724.
{\small
\bibliographystyle{ieee}
| 3fbae3a50e8674e929d8db0825ec9930debe9c86 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Different types of astrophysical objects have been proposed as production sites of high-energy neutrinos through the decay of charged pions, previously produced in the interactions of nuclei with ambient matter or radiation \cite{Sources1, Sources2, Sources3, Sources4, Sources5}.
In contrast to charged cosmic rays, neutrinos are not deflected by (inter-)galactic magnetic fields and point straight back to their production sites.
Finding sources of cosmic neutrinos would identify sources of cosmic rays, whose origin and acceleration processes are a long-standing astrophysical question.
The results of the latest search for point-like sources using the ANTARES neutrino telescope are presented in this paper. For the first time, events based on the signal
induced by electromagnetic and/or hadronic showers are included. This has been achieved by using a new reconstruction algorithm \cite{TMichael_proc,TANTRA} which allows a median pointing accuracy between \unit{2}{\degree} and \unit{3}{\degree} for electron neutrinos which interact via CC with energies in the 10$^3$ -- 10$^6$ GeV range.
This paper is structured as follows: in Sec. \ref{sec:ANTARES}, the ANTARES neutrino telescope and the event selection for this analysis are introduced. The search method is explained in Sec. \ref{sec:METHOD}, the results of the analysis shown in Sec.\ref{sec:RESULTS} and the discussion on the possible effects due to systematic uncertainties described in Sec. \ref{sec:SYST}. Finally, the conclusions are summarized in Sec. \ref{sec:CONCL}.
\section{ANTARES neutrino telescope and event selection}\label{sec:ANTARES}
The ANTARES telescope \cite{antares}, located in the Mediterranean Sea, is the largest neutrino detector in the Northern Hemisphere.
The detector comprises a three-dimensional array of 885 optical modules (OMs), each one housing a 10'' photomultiplier tube (PMT), and distributed over 12 vertical strings anchored at the sea floor at a depth of about 2400 m.
The detection of light from upgoing charged particles is optimized with the PMTs facing \unit{45}{\degree} downward.
Simulations are used to evaluate the performance of the detector. Atmospheric muons and neutrinos are simulated with the MUPAGE \cite{MuPara,mupage} and GENHEN \cite{simtools, simtools2} packages, respectively. For the simulation of atmospheric muons, the total amount of simulated events corresponds to 1/3 of the total livetime of the data set.
The Bartol flux \cite{bartol} is considered to represent the atmospheric neutrino flux.
The simulation of the amount of optical background (\ce{^{40}K} and bioluminescence) is performed according to the collected data in order to account for the variations of the environmental conditions \cite{RBR}.
The data used for the analysis was recorded between 2007 January 29th and 2015 December 31st. During this period, which includes the commissioning phase, the detector operation with at least five lines corresponds to a total livetime of 2423.6 days.
The event selection is optimized following a blind procedure on pseudo-data sets of data randomized in time (pseudo-experiments) before performing the analysis. A more detailed description of the pseudo-experiments can be seen in section \ref{sec:TS}.
The selection criteria for tracks and showers are explained in Sec. \ref{muon_sel} and \ref{show_sel}, respectively. These criteria have been optimized to minimize the neutrino flux needed for a $5\sigma$ discovery of a point-like source in $50\%$ of the pseudo-experiments. The effective area for $\parenbar{\nu}\kern-0.4ex_{\mu}$\footnote{The notation $\parenbar{\nu}\kern-0.4ex$ refers to both neutrinos and anti-neutrinos.} CC after the track selection cuts, and for $\parenbar{\nu}\kern-0.4ex_{e}$ CC and $\parenbar{\nu}\kern-0.4ex_{\mu}$ NC events after the shower selection cuts can be seen in Fig. \ref{fig:Aeff}.
The simulations produced to account for the interaction of $\parenbar{\nu}\kern-0.4ex_{\tau}$ cover a limited livetime of the data. Because of this, the contribution due to the leptonic and hadronic channels after the interaction of $\parenbar{\nu}\kern-0.4ex_{\tau}$ are estimated by scaling the contribution of other flavor neutrino events. A neutrino flux with an E$^{-2}$ energy is assumed in this scaling. The obtained rates are the following: to take into account the decay of the outgoing $\tau$ into a $\mu$ (branching ratio of $\sim$ 17\%) after a $\parenbar{\nu}\kern-0.4ex_{\tau}$ Charged Current (CC) interaction, the number of events predicted by CC $\parenbar{\nu}\kern-0.4ex_{\mu}$ interactions is increased by 9\%;
to take into account the decay of the outgoing $\tau$ into an electron (branching ratio of $\sim$ 17\%), the number of CC $\parenbar{\nu}\kern-0.4ex_{e}$ is increased by 12\%;
finally, to take into account the $\tau$ decaying into hadrons (branching ratio of $\sim$ 64\%) after a $\parenbar{\nu}\kern-0.4ex_{tau}$ CC interaction and the $\parenbar{\nu}\kern-0.4ex_{\tau}$ NC interactions, the small number of events predicted by $\parenbar{\nu}\kern-0.4ex_{\mu}$ NC interactions is increased by 374\%.
\begin{figure}[!ht]
\centering
\includegraphics[width=.49\linewidth]{Aeff_d-90}
\includegraphics[width=.49\linewidth]{Aeff_d-45}
\includegraphics[width=.49\linewidth]{Aeff_d0}
\caption{Effective area for ${\nu_\mu}$+$\bar{\nu}_\mu$ CC events after the track selection cuts (solid line) and for $\nu_{e}$+$\bar{\nu}_{e}$ CC and $\nu_\mu$+$\bar{\nu}_\mu$ NC events after the shower selection cuts (dashed lines) considering three declination ranges.}
\label{fig:Aeff}
\end{figure}
\subsection{Muon track selection}
\label{muon_sel}
Muon tracks are reconstructed using a multi-step procedure that concludes with a maximum likelihood method \cite{antPS4y}. This likelihood takes into account the so-called \textit{hits}. A hit is defined as the digital information on the time and amplitude of a PMT signal, where the latter is proportional to the number of detected photons.
As in the previous publication \cite{lastPS}, muon events are selected applying cuts on the reconstructed zenith angle ($\cos\theta_{tr} > -0.1$), the estimated angular error ($\beta_{tr} < \unit{1}{\degree}$) and the parameter that describes the quality of the reconstruction ($\Lambda > -5.2$).
An approximated evaluation of the energy deposited per unit of path length is used to estimate the muon energy. An energy estimator, $\rho$, is defined using the hit charge, recorded by all PMTs used to reconstruct the track, and the length of the muon path in the detector \cite{dEdXcuts,dEdX}.
The energy estimator fails for events for which the muon energy is below the value of the critical energy to produce significant energy losses due to radiative processes ($\sim$500 GeV), and for tracks with estimated track length, $L_\mu$, below \unit{380}{\metre}, yielding small values of $\rho$.
Such events are excluded from the analysis. Table~\ref{tab:TrackSel} gives an overview of the selection cuts applied for the simulated track sample. A total of 7622 neutrino candidates in the track channel are selected in data for this search.
\begin{table*}[!h]
\centering
\caption{Selection cuts for the track sample and number of remaining simulated events after each step for atmospheric muons ($n_{\mu}^{atm}$), atmospheric neutrinos ($n_{\nu}^{atm}$) and cosmic neutrinos ($n_{\nu}^{E^{-2}}$) reconstructed as a track in the detector. For cosmic neutrinos, a flux according to $d\varPhi/dE = 10^{-8}\,(E / \si{GeV})^{-2} \, \si{GeV^{-1}\,cm^{-2}\,s^{-1}}$ is assumed.
\medskip }
\label{tab:TrackSel}
\begin{tabular}{l c c c c}
Criterion & Condition & $n_{\mu}^{atm}$ & $n_{\nu}^{atm}$ & $n_{\nu}^{E^{-2}}$ \\
\hline
Trigger & & $4.9\times 10^{8}$ & $ 6.3 \times 10^{4}$ & 204 \\
Up-going & $\cos\theta_{tr} > -0.1$ & $4.3\times 10^{7}$ & $5.0\times 10^{4}$ & 151 \\
Angular error estimate
& $\beta_{tr} < \unit{1}{\degree}$ & $2.2\times 10^{7} $ & $3.3\times 10^{4}$ & 105 \\
Track reconstruction quality & $\Lambda > -5.2$ & $1513$ & 7475 & 44 \\\smallskip
Track length and energy cut & $ L_{\mu} > 380$ m, $\log_{10}(\rho) > 1.6$ & 1117 & 7086 & 41 \\
\end{tabular}
\end{table*}
\subsection{Shower selection}
\label{show_sel}
Shower events are reconstructed with a new algorithm based on a two-step procedure. In the first step, the interaction vertex is obtained by the maximization of an M-estimator, $M_\mathrm{est}$, which depends on the time and charge of the hits. The direction of the event is estimated with a maximum likelihood method, using the information of the reconstructed interaction vertex and the detected amplitude of the OMs.
Shower events are required to be reconstructed as up-going or coming from close to the horizon ($\cos\theta_{sh} > -0.1$) with a restriction on the angular error estimate ($\beta_{sh} < \unit{30}{\degree}$). The interaction vertex of each event is also required to be reconstructed inside or close to the instrumented volume.
To further reduce the background from mis-reconstructed atmospheric muons, additional selection cuts are imposed.
These cuts are based on the $M_\mathrm{est}$ value, on a Random Decision Forest classifier value, $RDF$, made with parameters provided by an alternative shower reconstruction \cite{Dusj-Paper}, and on a likelihood, $\mathscr L_\mathrm{\mu}$ or muon likelihood, that uses information of the hits in the event.
A description of these cuts is given in Appendix \ref{appendix}. Events passing the muon track selection are excluded from the shower channel making the two samples mutually exclusive. The full list of selection cuts is summarized in Table~\ref{tab:ShowSel}. The selection yields 180 shower events.
\begin{table*}[!h]
\centering
\caption{Selection cuts for the shower sample and number of remaining simulated events after each step for atmospheric muons ($n_{\mu}^{atm}$), atmospheric neutrinos ($n_{\nu}^{atm}$) and cosmic neutrinos ($n_{\nu}^{E^{-2}}$) reconstructed as a shower in the detector. For cosmic neutrinos, a flux according to $d\varPhi/dE = 10^{-8}\,(E / \si{GeV})^{-2} \,\si{GeV^{-1}\,cm^{-2}\,s^{-1}}$ is assumed.
Refer to Appendix \ref{appendix} for more details.\medskip }
\label{tab:ShowSel}
\begin{tabular}{l c c c c c}
Criterion & Condition & $n_{\mu}^{atm}$ & $n_{\nu}^{atm}$ & $n_{\nu}^{E^{-2}}$ \\
\hline
Track Veto & not selected as muon track & $4.9 \times 10^{8}$ & $5.6 \times 10^{4}$ & 160 \\
Up-going & $\cos\theta_{sh} > -0.1$ & $1.5 \times 10^{8}$ & $2.3 \times 10^{4} $ & 90 \\
Interaction vertex & $R_{sh} < \unit{300}{\metre}$,
$|Z_{sh}| < \unit{250}{\metre}$ & $7.7 \times 10^{7}$ & $2.1 \times 10^{4} $ & 80 \\
M-estimator & $M_\mathrm{est} < 1000$ & $7.2 \times 10^{7}$ & $2.0 \times 10^{4} $ & 80 \\
RDF & $ RDF > 0.3$ & $8.0 \times 10^{4}$ & 2044 & 24 \\
Muon likelihood & $\mathscr L_\mathrm{\mu} > 50$ & 90 & 109 & 12 \\
\end{tabular}
\end{table*}
\subsection{Comparison between data and simulations}
Figure~\ref{fig:data_mc}-left compares the distributions of the quality parameter $\Lambda$ for different types of simulated events with the data set for the track channel. Figure~\ref{fig:data_mc}-right shows the comparison in the reconstructed zenith angle for the shower channel. It is estimated that about 13\% of the selected muon tracks and $52\,\%$ of the selected shower events are atmospheric muons mis-reconstructed as up-going.
The simulation overestimates the number of events by $8\,\%$ ($17\,\%$) in the track (shower) channel for the final set of cuts. This difference is well within the overall systematic uncertainty on the atmospheric neutrino flux normalization \cite{dEdX}. A larger overestimation of events is observed in the region where the background of mis-reconstructed atmospheric muons is dominant.
On the other hand, in the cascade channel, an underestimation of events is observed for zenith angles larger than $cos(\theta_{sh} ) >$ 0.4, which can be explained due to the large uncertainties of the atmospheric neutrino flux. The searches for an excess from a point-like direction in the sky are, compared to other searches, less influenced by higher levels of background contamination. Referring to the right panel of Fig. \ref{fig:data_mc}, even if there is a large contamination of atmospheric muons between -0.1 and 0.1, according to our pseudo-experiment simulations the current event selection produces better sensitivities compared to considering only events with $cos(\theta_{sh}) >$ 0.1. A global good agreement between data and Monte Carlo justifies the procedure to optimize the selection criteria for the separation of signal and background using simulated events.
\begin{figure*}
\includegraphics[width=0.48\linewidth]{lambda_Int_v3}
\includegraphics[width=0.48\linewidth]{Costh_v3}
\caption{Comparison of the data with Monte Carlo (MC) simulations as a function of the quality parameter $\Lambda$ (left). This figure corresponds to the event distribution after a cut on the estimated angular error ($\beta_{tr} < \unit{1}{\degree}$) and on the reconstructed zenith angle ($\cos\theta_{tr} > -0.1$). The dashed vertical line marks the cut value. Right: comparison of the data with the simulations in the zenith $\theta_{sh}$ of the reconstructed shower direction. This figure corresponds to the event distribution after all shower selection cuts presented in Table~\ref{tab:ShowSel}. For the cosmic neutrinos, a flux according to $d\varPhi/dE = 10^{-8}\,(E / \si{GeV})^{-2} \, \si{GeV^{-1}\,cm^{-2}\,s^{-1}}$ is assumed in both figures. The two bottom plots show the data to MC ratio, where the number of MC events is the sum of atmospheric muons and neutrinos.}
\label{fig:data_mc}
\end{figure*}
\section{Search method}\label{sec:METHOD}
While atmospheric neutrino events are randomly distributed, neutrinos from point-like sources are expected to accumulate in spatial clusters. To find these clusters, a maximum likelihood ratio approach is followed. The likelihood used describes the data in terms of signal and background probability density functions (PDFs) and is defined as
\begin{align} \label{eq:pslik}
\log \mathscr L_\mathrm{s+b} = &\sum_{\mathcal S} \sum_{i\in\mathcal S} \log \Big[ \mu_\mathrm{sig}^{\mathcal S}
\mathscr {F}^{\mathcal S}_{i} {\mathscr P}^{\mathcal S}_\mathrm{sig, i} + {\mathscr N}^{\,\mathcal S} {\mathscr B}^{\,\mathcal S}_{i} {\mathscr P}^{\mathcal S}_\mathrm{bkg, i}\Big]
- \mu_\mathrm{sig} .
\end{align}
In this equation, $\mathcal S$ denotes the sample ($\it{tr}$ for tracks, $\it{sh}$ for showers), $i$ indicates the event of the sample $\mathcal S$, $\mu_\mathrm{sig}^{\mathcal S}$ is the number of signal events fitted to in the $\mathcal S$ sample, $\mathscr {F}^{\mathcal S}_{i}$ is a parameterization of the point spread function, ${\mathscr {P}}^{\mathcal S}_\mathrm{sig, i}$ is derived from the probability density function of the energy estimator, yielding the probability of measuring the signal with the reconstructed energy of the event $i$, ${\mathscr N}^{\,\mathcal S}$ is the total number of events in the $\mathcal S$ sample, ${\mathscr B}^{\,\mathcal S}_{i}$ is the background rate obtained from the distribution of the observed background events at the declination of event $i$, ${\mathscr {P}}^{\mathcal S}_\mathrm{bkg, i}$ is the probability density function of the energy estimator for background and $\mu_\mathrm{sig} = {\mu}^{tr}_\mathrm{sig} + {\mu}^{sh}_\mathrm{sig}$ is the total number of fitted signal events.
More details on the components of the PDFs are given below.
\subsection{Point spread function}
The distribution of signal events around a hypothetical point-like source is described by the point spread function (PSF) $\mathscr F$. The PSF is defined as the probability density to find a reconstructed event at an angular distance $\Delta\Psi$ around the direction of the source. It depends on the angular resolution of the event sample.
Figure~\ref{fig:psf_cum} shows the cumulative distributions of the angular distance between the reconstructed and true neutrino direction for track and shower events. The PSFs are determined from Monte Carlo simulations of neutrinos with an $E^{-2}$ energy spectrum.
The figure shows that about $50\,\%$ of the track (shower) events are reconstructed within \unit{0.4}{\degree} (\unit{3}{\degree}) of the parent neutrino.
\begin{figure}[]
\centering
\includegraphics[width=.8\linewidth]{angular_resolution_both}
\caption{
Probability to obtain a reconstructed angle within $\Delta\Psi$ between the reconstructed direction of track-like events (blue) and shower-like events (red) with respect to the true Monte Carlo neutrino direction. A neutrino flux with an $E^{-2}$ energy spectrum is assumed.
}
\label{fig:psf_cum}
\end{figure}
\subsection{Background rate}
The background rate $\mathscr B$ is described as a function of the declination, $\delta$. Given the small expected contribution of a cosmic signal in the overall data set, the background rate is estimated directly from the measured data.
Due to the Earth's rotation and a sufficiently uniform exposure, the background is considered independent of right-ascension, $\alpha$.
The rate of selected events as a function of declination is shown in Fig.~\ref{fig:background}.
\begin{figure*}[h]
\includegraphics[width=0.48\linewidth]{BGrate_tracks.pdf}
\includegraphics[width=0.48\linewidth]{BGrate_cascades.pdf}
\caption{Number of selected track-like (left) and shower-like (right) events as a function of the reconstructed declination. The red and blue lines are different spline parametrisations (see Sec.~\ref{sec:SYST}). The different shape for showers is mainly due to a higher relative contamination of atmospheric muons in the sample.
}
\label{fig:background}
\end{figure*}
\subsection{Energy estimator}
Neutrinos generated in the atmosphere have a much softer energy spectrum ($\propto E^{-3.7}$) than neutrinos from the expected astrophysical flux proportional to $E^{-2}$.
For this reason, the energy estimator information is used in the likelihood to further distinguish between cosmic signal and atmospheric background.
For the shower channel, the number of hits ($N_{sh}$) used by the reconstruction algorithm is employed as energy estimator.
A different and more elaborate approach is assumed for the track channel.
In this case the estimator $\rho$ is used as a proxy for the energy of the neutrino event. The information of the event angular error estimate $\beta_{tr}$ is also included. Moreover, the dependence of the energy estimator on the declination of the event is taken into account by generating both the signal and the background PDF in steps of 0.2 over $\sin\delta$.
\subsection{Implementation}
\label{sec:TS}
The significance of any observation is determined by a test statistic denoted as $\mathcal Q$ which is defined from the likelihood as
\begin{equation}
\mathcal Q = \log \mathscr L_\mathrm{s+b} - \log \mathscr L_\mathrm{b}.
\label{eq:teststat}
\end{equation}
The $\mathcal Q$ distributions for different signal strengths are determined from pseudo-experiments. In these, $\BigO{10^{4}\sim10^{5}}$ random sky maps are generated with a number of background events that follow the declination-dependent event distribution as seen in the actual data, and a uniform right ascension. In addition, signal events are injected according to the investigated spectrum by assuming either a point or extended source profile.
In equation~(\ref{eq:teststat}) $\mathscr L_\mathrm{b}$ corresponds to the definition of $\mathscr L_\mathrm{s+b}$ in equation~(\ref{eq:pslik}) evaluated with the same parameters as the maximum likelihood estimate but with the numbers of signal events set to zero: $\mu^\mathrm{tr}_\mathrm{sig} = \mu^\mathrm{sh}_\mathrm{sig} = 0$ (background-only case).
In the likelihood maximization, the position in the sky of the fitted source is either kept fixed or allowed to be fitted within specific limits depending on the type of search (see Sec. \ref{sec:RESULTS}). Furthermore, the values of $\mu^\mathrm{tr}_\mathrm{sig}$ and $\mu^\mathrm{sh}_\mathrm{sig}$ are left free to vary, and can indeed go below zero to reflect the degree of absence of events around the probed coordinates.
The declination-dependent acceptance for a given sample, $A_{S}(\delta)$, is defined as the proportionality constant between a given flux normalization $\varPhi_0 = E^2 d\varPhi/dE$ and the expected number of signal events for this particular flux. It can be expressed in terms of the effective area, $\mathcal{A}_{\mathrm{eff},S}(E_\nu, \delta)$:
\begin{equation}
A_{S}(\delta) = \varPhi_0^{-1} \iint dtdE_\nu\ \mathcal{A}_{\mathrm{eff}, S}(E_\nu, \delta) \frac {d\varPhi} {dE_\nu},
\end{equation}
where the integral is over the livetime of all selected runs (2423.6 days) and over an energy range large enough to include all potential events within the sensitivity of ANTARES. With the assumed $E^{-2}$ spectrum, 90\% of the events are found in an energy range between 2$\cdot 10^{3}$ and 3$\cdot 10^{6}$ GeV for the track channel (between 5$\cdot 10^{3}$ and 4$\cdot 10^{6}$ GeV for the shower channel).
Figure~\ref{fig:acceptance} shows how the acceptances for tracks and showers depend on the declination.
\begin{figure}[h]
\centering
\includegraphics[width=0.7\linewidth]{Acceptance.pdf}
\caption{The acceptance as a function of the source declination for an $E^{-2}$ energy spectrum with a flux normalization factor of $\varPhi_0 = \unit{e-8}{\giga\electronvolt\,\centi\metre\tothe{$-2$}\,\second\tothe{$-1$}}$ for the track (blue) and shower (red) samples. For better visibility, the acceptance for showers is scaled up by a factor 3.}
\label{fig:acceptance}
\end{figure}
\section{Search for neutrino sources}\label{sec:RESULTS}
The search for astrophysical neutrino sources presented in this paper is performed with four approaches.
\begin{description}
\item [1] {\it Full sky search.} In the first method, the whole visible sky of ANTARES is scanned in an unbinned way to search for spatial clustering of events with respect to the expected background.
\item [2] {\it Candidate list search.} In the second approach, the directions of a pre-defined list of known objects which are neutrino source candidates are investigated to look for an excess or (in the case of null observation) to determine an upper limit on their neutrino fluxes.
\item[3] {\it Galactic Center region.} The third search is similar to the full sky search but restricted to a region centered in the origin of the galactic coordinate system ($\alpha$, $\delta$) = (266.40$^\circ$,--28.94$^\circ$) and defined by an ellipse with a semi-axis of 15$^\circ$ in the direction of the galactic latitude and a semi-axis of 20$^\circ$ in galactic longitude.
The motivation relies on the number of high-energy neutrino events observed by the IceCube (IC) detector \cite{IC3years, IC4yproc} that appear to cluster in this region. Furthermore, the HESS Collaboration recently discovered an accelerator of PeV protons in the Galactic Center \cite{HESSPeV} that could produce high-energy neutrinos.
\item[4] {\it Sagittarius A*.} Finally, the fourth approach tests the location of Sagittarius A* as an extended source by assuming a Gaussian emission profile of various widths.
\end{description}
Figure \ref{fig:SkyMap} represents the event sample in equatorial coordinates in the ANTARES visible sky. The considered neutrino source candidates and the search region around the Galactic Center are also indicated.
\begin{figure}[h]
\centering
\includegraphics[width=1.0\linewidth]{skymap.pdf}
\caption{Sky map in equatorial coordinates of the 7622 track (blue crosses) and the 180 shower (red circles) events passing the selection cuts. Green stars indicate the location of the 106 candidate neutrino sources, and green squares indicate the location of the 13 considered tracks from the IceCube high energy sample events or HESE (see Sec. \ref{sec:CL}). The black solid ellipse indicates the search region around the Galactic Center, in which the origin of the galactic coordinates is indicated with a black star. The black dashed line indicates the galactic equator.}
\label{fig:SkyMap}
\end{figure}
\subsection{Full sky search}
\label{fullsky}
In the full sky search, the whole visible sky of ANTARES is divided on a grid with boxes of $\unit{1}{\degree} \times \unit{1}{\degree}$ in right ascension and declination for the evaluation of the $\mathcal Q$-value defined in Eq.~(\ref{eq:teststat}). This value is maximised in each box by letting the location of the fitted cluster free between the $\unit{1}{\degree} \times \unit{1}{\degree}$ boundaries. Since an unbinned search is performed, events outside the grid boxes are indeed considered in each $\mathcal Q$-value maximisation.
The pre-trial p-value of each cluster is calculated by comparing the $\mathcal Q$-value obtained at the location of the fitted cluster with the background-only $\mathcal Q$ obtained from simulations at the corresponding declination.
Figure~\ref{fig:Pretrial} shows the position of the cluster and the pre-trial p-values for all the directions in the ANTARES visible sky.
The most significant cluster of this search is found at a declination of $\delta = \unit{23.5}{\degree}$ and a right-ascension of $\alpha = \unit{343.8}{\degree}$ and with a pre-trial p-value of $\num{3.84e-6}$.
To account for trial factors, this pre-trial p-value is compared to the distribution of the smallest p-values found anywhere in the sky when performing the same analysis on many pseudo-data sets. It is found that 5.9\% of pseudo-experiments have a smaller p-value than the one found in the final sample, corresponding to a post-trial significance of $1.9\sigma$ (two-sided convention). The upper limit on the neutrino flux coming from this sky location is $E^2 d\varPhi/dE = \unit{3.8e-08}{GeV\,cm^{-2}\,s^{-1}}$.
The location of this cluster is found at a distance of 1.1$^\circ$ from event ID 3 of the 6 year Northern Hemisphere Cosmic Neutrino flux sample of IceCube \cite{IC6years_HE}. A rough estimate of the significance of this coincidence is performed. 26 out of the 29 of these events are found in the declination range between -5$^\circ$ and 30$^\circ$. The remaining events were excluded since the event density in the selected region is larger, and therefore the estimation is slightly more conservative. By assuming a random distribution of 26 events within this declination range, the probability of a random coincidence within 1$^\circ$ between at least one event and the most significant cluster of the full sky search is $\sim$1\%.
The distribution of events of this cluster is shown in Fig.~\ref{fig:AllClusters}-top-left. It contains 16(3) tracks within $\unit{5}{\degree}(\unit{1}{\degree})$ and 1 shower event within $\unit{5}{\degree}$. The upper limits of the highest significant cluster in bands of 1$^\circ$ in declination at a 90\% Confidence Level (C.L.) obtained using the Neyman method \cite{neyman} are shown in Fig. \ref{fig:LimitsFix}. The limits computed in this analysis are set on the one-flavor neutrino flux assuming equipartition at Earth of the three neutrino flavors.
\begin{figure}[h]
\centering
\includegraphics[width=\linewidth]{skymap2_def_w0.png}
\caption{%
Sky map in equatorial coordinates of pre-trial p-values for a point-like source of the ANTARES visible sky. The red circle indicates the location of the most significant cluster of the full sky search. For this map, a smaller grid size of 0.2$^\circ \times$ 0.2$^\circ$ was used.
}
\label{fig:Pretrial}
\end{figure}
\begin{figure}[p]
\centering
\includegraphics[width=0.75\textwidth]{AllClusters_Scatter.pdf}
\caption{\scriptsize Distribution of events in the ($\alpha$, $\delta$) (RA, DEC) coordinates for the most significant clusters found in the full sky search (top left), candidate list search (HESSJ0632+057) (top right), search over the track events from the IceCube HESE sample (track with ID = 3) (middle left), search around the Galactic Center for an $E^{-2}$ point-like source (middle right), search around the Galactic Center for an $E^{-2.5}$ point-like source (bottom left) and at the location of Sagittarius A* (bottom right). In all figures, the inner (outer) green line depicts the one (five) degree distance from the position of the best fit or known location, indicated as a gray star. The red points denote shower-like events, whereas the blue points indicate track-like events. Different tones of red and blue correspond to the values assumed by the energy estimators: the number of hits (shower-like events) and the $\rho$ parameter (track-like events) as shown in the legend. The dashed circles around the events indicate the angular error estimate. }
\label{fig:AllClusters}
\end{figure}
\begin{figure}[h!]
\centering
\includegraphics[width=1\linewidth]{sensitivity_discovery_fluxes_FS}
\caption{Upper limits at a $90\,\%$ C.L. on the signal flux from the investigated candidates assuming an $E^{-2}$ spectrum (red circles). The dashed red line shows the ANTARES sensitivity (defined as the median upper limit at 90\% C.L. for a background-only case) and the blue dashed line the sensitivity of the seven years point-like source analysis by the IceCube Collaboration for comparison \cite{icecubePS}. The upper-limits obtained in this analysis are also included (blue dots). The ANTARES 5$\sigma$ pre-trial discovery flux is a factor 2.5 to 2.9 larger than the sensitivity. The curve for the sensitivity for neutrino energies under 100 TeV is also included (solid red line). The IceCube curve for energies under 100 TeV (solid blue line) is obtained from the 3 years MESE analysis \cite{icecubeMESE}. The limits of the most significant cluster obtained in bands of 1$^\circ$ in declination (dark red squares) are also shown.}
\label{fig:LimitsFix}
\end{figure}
\subsection{Candidate list}
\label{sec:CL}
The candidate list used in the last ANTARES point-like source analysis \cite{lastPS} contained neutrino source candidates both from Galactic and extra-Galactic origin listed in the TeVCat catalog \cite{tevcat}. These sources had been observed by gamma-ray experiments before July 2011 in the 0.1--100 TeV energy range and with declinations lower than $\unit{20}{\degree}$. Furthermore, since the energy of high energy gamma-rays of extra-galactic origin can degrade before they reach the Earth, extra-Galactic candidates were selected also among the sources observed by gamma-ray satellites in the 1--100 GeV energy range.
This paper updates the neutrino search for the 50 objects considered in \cite{lastPS} with additional 56 galactic and extragalactic sources. The newly considered sources include those detected in the 0.1--100 TeV energy range by gamma-ray experiments after July 2011 and some bright sources with declinations between 20$^\circ$ and 40$^\circ$ not considered in the past. Additionally, the reconstructed direction of the IceCube multi-PeV track event \cite{IC6years_HE} and the 2HWC sources which are not coincident with any known source \cite{HAWC} have been included.
Finally, seven more sources are added: the three blazars with highest intensity observed by the TANAMI Collaboration that coincide with three events from the IceCube HESE sample \cite{tanami, tanami2, antares_blazars}, and the four gravitationally lensed Flat Spectrum Radio Quasars with the highest magnification factor analyzed in a previous work \cite{quasar}.
The list of the astronomical candidates is shown in Table~\ref{tab:LimitsFix} along with their equatorial coordinates, fitted number of signal events and upper limits on the flux.
The most signal-like cluster is found at the location of HESSJ0632+057 at $(\alpha,\delta) = (\unit{98.24}{\degree},\unit{5.81}{\degree})$, with a pre-trial p-value of 0.16\%. The second and third most significant sources correspond to PKS1440-389 and PKS0235+164, with pre-trial p-values of 0.5\% and 5\%, respectively. To account for trial factors, the search is performed on the same list of sources using pseudo data-sets, from which the distribution of the smallest p-value for a background-only case is obtained. It is found that 13\% of the pseudo-experiments have a smaller p-value for any source compared to the one obtained for this location, corresponding to a post-trial significance of $1.5\sigma$ (two-sided convention). The cluster contains 11(1) tracks within $\unit{5}{\degree}(\unit{1}{\degree})$ and 2 shower events within $\unit{5}{\degree}$ around the source candidate.
The distribution of events around this source is shown in Fig.~\ref{fig:AllClusters}-top-right.
The sensitivities and limits calculated with the Neyman method at a 90\% C.L. and the 5$\sigma$ discovery flux for this search (assuming an $E^{-2}$ spectrum) are shown in Fig.~\ref{fig:LimitsFix} as a function of the declination. To prevent undesired effects of the Neyman construction \cite{Neyman-problems}, the maximum between the sensitivity (i.e., the median upper limit at a 90\% C.L. for a background-only case) and the limit for the particular location of the source are reported, as is customary in the field \cite{lastPS, icecubePS}.
The 13 track candidates from the IceCube HESE sample classified as muon tracks \cite{IC3years, IC4yproc} are considered in a separate candidate list search.
Since those events have a non-negligible angular error estimate, the direction parameters are not fixed but fitted within a cone of twice their angular error estimate around the direction given by the IceCube tracks.
The coordinates of these events are shown in Table~\ref{tab:LimitsHESE} together with their angular uncertainty (provided by the IceCube Collaboration), fitted number of signal events and upper limits on the flux derived from this analysis.
The muon track candidate from the HESE sample with the largest excess in fitted signal is the IceCube track with ID 3 and $\mu_\mathrm{sig} =5.3$.
The fitted cluster is located at $(\alpha,\delta) = (\unit{130.1}{\degree},\unit{-29.8}{\degree})$, which is at a distance of 1.5$^\circ$ from the original HESE track at $(\alpha,\delta) = (\unit{127.9}{\degree},\unit{-31.2}{\degree})$. The observed post-trial p-value is 20\% (significance of 1.2$\sigma$). The upper limit on the signal from this candidate is $\varPhi_0^{90\,\%} = \unit{2.1e-8}{\giga\electronvolt\,\centi\metre\tothe{$-2$}\,\second\tothe{$-1$}}$.
The cluster is shown in Fig.~\ref{fig:AllClusters}-middle-left.
\setlength{\tabcolsep}{.3em}
\begin{table*}[p]
\caption{\scriptsize List of astrophysical objects used in the candidate list search. Presented are the object's coordinates in declination ($\delta$) and right-ascension ($\alpha$). The first column reports the type of source: \textit{Binary} means X-Ray binary, \textit{GC} means Galactic Center, \textit{Radio} means Radio Galaxy, \textit{Sey2} means Seyfert 2 Galaxy, \textit{UNID} means unidentified. The last two columns show the sum of the fitted number of signal track and shower events $\mu_\mathrm{sig} = {\mu}^{tr}_\mathrm{sig} + {\mu}^{sh}_\mathrm{sig}$, and the $90\,\%$ C.L. upper limits on the flux normalization factor $\varPhi_0^{90\,\%}$ (in units of $10^{-8}$ GeV cm$^{-2}$ s$^{-1}$). Candidates of the same type are sorted by declination.
\medskip}
\label{tab:LimitsFix}
\resizebox{\linewidth}{!}{
\begin{minipage}{1.05\textwidth}
\centering
{\def0.54{0.54}
\footnotesize
\begin{tabular}{l crrcc | l crrcc}
Type & Name & $\delta [\si{\degree}]$ & $\alpha [\si{\degree}]$ & $\mu_\mathrm{sig}$ & $\varPhi_0^{90\,\%}$ & Type & Name & $\delta [\si{\degree}]$ & $\alpha [\si{\degree}]$ & $\mu_\mathrm{sig}$ & $\varPhi_0^{90\,\%}$ \\
\hline
BLLac & PKS2005-489 & -48.82 & 302.37 & 0.3 & 0.93 & & PKS1406-076 & -7.90 & 212.20 & -- & 0.92 \\
& PKS0537-441 & -44.08 & 84.71 & 0.6 & 0.96 & & QSO2022-077 & -7.60 & 306.40 & 1.0 & 1.64 \\
& PKS1440-389 & -39.14 & 220.99 & 2.9 & 1.56 & & 3C279 & -5.79 & 194.05 & 0.8 & 1.59 \\
& PKS0426-380 & -37.93 & 67.17 & -- & 0.70 & & B1030+074 & 7.19 & 158.39 & -- & 1.01 \\
& PKS1454-354 & -35.67 & 224.36 & 1.2 & 1.28 & & PKS1502+106 & 10.52 & 226.10 & -- & 1.03 \\
& TXS1714-336 & -33.70 & 259.40 & 0.8 & 1.31 & & 3C454.3 & 16.15 & 343.50 & -- & 1.10 \\
& PKS0548-322 & -32.27 & 87.67 & -- & 0.85 & & 4C+21.35 & 21.38 & 186.23 & -- & 1.37 \\
& H2356-309 & -30.63 & 359.78 & -- & 0.79 & & B1422+231 & 22.93 & 216.16 & -- & 1.12 \\
& PKS2155-304 & -30.22 & 329.72 & -- & 0.80 & & PKS1441+25 & 25.03 & 220.99 & -- & 1.38 \\
& 1ES1101-232 & -23.49 & 165.91 & -- & 0.85 & Radio & PKS0625-35 & -35.49 & 96.78 & -- & 0.74 \\
& 1ES0347-121 & -11.99 & 57.35 & -- & 0.92 & SNR & LHA120-N-157B & -69.16 & 84.43 & -- & 0.63 \\
& RGBJ0152+017 & 1.79 & 28.17 & -- & 1.14 & & RCW86 & -62.48 & 220.68 & -- & 0.62 \\
& RBS0723 & 11.56 & 131.80 & -- & 1.03 & & MSH15-52 & -59.16 & 228.53 & -- & 0.68 \\
& PKS0235+164 & 16.61 & 39.66 & 2.1 & 1.93 & & SNRG327.1-01.1 & -55.08 & 238.65 & -- & 0.63 \\
& RGBJ2243+203 & 20.35 & 340.98 & -- & 1.29 & & RXJ0852.0-4622 & -46.37 & 133.00 & -- & 0.65 \\
& VERJ0521+211 & 21.21 & 80.44 & 1.2 & 1.84 & & RXJ1713.7-3946 & -39.75 & 258.25 & -- & 0.67 \\
& S20109+22 & 22.74 & 18.02 & -- & 1.30 & & W28 & -23.34 & 270.43 & 0.8 & 1.43 \\
& PKS1424+240 & 23.79 & 216.75 & -- & 1.12 & & SNRG015.4+00.1 & -15.47 & 274.52 & 0.2 & 1.34 \\
& MS1221.8+2452 & 24.61 & 186.10 & -- & 1.13 & & W44 & 1.38 & 284.04 & -- & 0.97 \\
& 1ES0647+250 & 25.05 & 102.69 & -- & 1.65 & & HESSJ1912+101 & 10.15 & 288.21 & -- & 1.03 \\
& S31227+25 & 25.30 & 187.56 & -- & 1.14 & & W51C & 14.19 & 290.75 & -- & 1.07 \\
& WComae & 28.23 & 185.38 & -- & 1.20 & & IC443 & 22.50 & 94.21 & -- & 1.12 \\
& 1ES1215+303 & 30.10 & 184.45 & -- & 1.26 & Sey2 & ESO139-G12 & -59.94 & 264.41 & -- & 0.82 \\
& 1ES1218+304 & 30.19 & 185.36 & -- & 1.21 & & CentaurusA & -43.02 & 201.36 & -- & 0.62 \\
& Markarian421 & 38.19 & 166.08 & -- & 1.59 & UNID & HESSJ1507-622 & -62.34 & 226.72 & -- & 0.62 \\
Binary & CirX-1 & -57.17 & 230.17 & -- & 0.84 & & HESSJ1503-582 & -58.74 & 226.46 & -- & 0.62 \\
& GX339-4 & -48.79 & 255.70 & -- & 0.63 & & HESSJ1023-575 & -57.76 & 155.83 & 1.5 & 1.08 \\
& LS5039 & -14.83 & 276.56 & -- & 1.19 & & HESSJ1614-518 & -51.82 & 243.58 & 0.7 & 0.96 \\
& SS433 & 4.98 & 287.96 & -- & 0.99 & & HESSJ1641-463 & -46.30 & 250.26 & -- & 0.78 \\
& HESSJ0632+057 & 5.81 & 98.24 & 2.7 & 2.40 & & HESSJ1741-302 & -30.20 & 265.25 & 0.6 & 1.29 \\
FSRQ & S30218+35 & 35.94 & 35.27 & 0.7 & 2.15 & & HESSJ1826-130 & -13.01 & 276.51 & -- & 1.07 \\
& B32247+381 & 38.43 & 342.53 & -- & 1.54 & & HESSJ1813-126 & -12.68 & 273.34 & -- & 0.90 \\
GC & GalacticCentre & -29.01 & 266.42 & 1.1 & 1.36 & & HESSJ1828-099 & -9.99 & 277.24 & 0.7 & 1.45 \\
PWN & HESSJ1356-645 & -64.50 & 209.00 & 0.4 & 0.98 & & HESSJ1834-087 & -8.76 & 278.69 & -- & 0.92 \\
& HESSJ1303-631 & -63.20 & 195.75 & -- & 0.64 & & 2HWCJ1309-054 & -5.49 & 197.31 & -- & 0.92 \\
& HESSJ1458-608 & -60.88 & 224.54 & 1.2 & 1.05 & & 2HWCJ1852+013* & 1.38 & 283.01 & -- & 0.97 \\
& HESSJ1616-508 & -50.97 & 243.97 & 0.5 & 0.96 & & 2HWCJ1902+048* & 4.86 & 285.51 & -- & 0.99 \\
& HESSJ1632-478 & -47.82 & 248.04 & -- & 0.73 & & MGROJ1908+06 & 6.27 & 286.99 & -- & 1.22 \\
& VelaX & -45.60 & 128.75 & -- & 0.62 & & 2HWCJ1829+070 & 7.03 & 277.34 & -- & 1.01 \\
& HESSJ1831-098 & -9.90 & 277.85 & -- & 0.95 & & 2HWCJ1907+084* & 8.50 & 286.79 & -- & 1.02 \\
& HESSJ1837-069 & -6.95 & 279.41 & -- & 1.30 & & ICPeV & 11.42 & 110.63 & -- & 1.03 \\
& MGROJ2019+37 & 36.83 & 304.64 & 0.4 & 2.08 & & 2HWCJ1914+117 & 11.72 & 288.68 & -- & 1.16 \\
Pulsar & PSRB1259-63 & -63.83 & 195.70 & -- & 0.64 & & 2HWCJ1921+131 & 13.13 & 290.30 & -- & 1.05 \\
& Terzan5 & -24.90 & 266.95 & -- & 1.09 & & 2HWCJ0700+143 & 14.32 & 105.12 & -- & 1.48 \\
& Geminga & 17.77 & 98.47 & 0.9 & 1.75 & & VERJ0648+152 & 15.27 & 102.20 & -- & 1.57 \\
& Crab & 22.01 & 83.63 & 0.1 & 1.64 & & 2HWCJ0819+157 & 15.79 & 124.98 & -- & 1.06 \\
Quasar & PKS1424-418 & -42.10 & 216.98 & 1.1 & 1.04 & & 2HWCJ1928+177 & 17.78 & 292.15 & -- & 1.26 \\
& SwiftJ1656.3-3302 & -33.04 & 254.07 & -- & 1.10 & & 2HWCJ1938+238 & 23.81 & 294.74 & -- & 1.24 \\
& PKS1622-297 & -29.90 & 246.50 & -- & 0.80 & & 2HWCJ1949+244 & 24.46 & 297.42 & -- & 1.60 \\
& PKS0454-234 & -23.43 & 74.27 & -- & 0.84 & & 2HWCJ1955+285 & 28.59 & 298.83 & -- & 1.18 \\
& PKS1830-211 & -21.07 & 278.42 & -- & 0.86 & & 2HWCJ1953+294 & 29.48 & 298.26 & -- & 1.20 \\
& QSO1730-130 & -13.10 & 263.30 & -- & 0.94 & & 2HWCJ1040+308 & 30.87 & 160.22 & -- & 1.42 \\
& PKS0727-11 & -11.70 & 112.58 & 1.3 & 1.59 & & 2HWCJ2006+341 & 34.18 & 301.55 & -- & 1.38 \\
\end{tabular}
}
\end{minipage}
}
\end{table*}
\begin{table}
\centering
\caption{The 13 IceCube muon track candidates from the IceCube HESE sample \cite{IC3years, IC4yproc} that are in the field of view of the ANTARES detector. The table gives the equatorial coordinates, the angular error estimate $\beta_\mathrm{IC}$ of the event and the $90\,\%$ C.L. upper limits on flux $\varPhi_0^{90\,\%}$ (in units of $10^{-8}$ GeV cm$^{-2}$ s$^{-1}$).\medskip}
\label{tab:LimitsHESE}
\begin{tabular}{crrcc}
HESE ID &$\delta [\si{\degree}]$ & $\alpha [\si{\degree}]$ & $\beta_\mathrm{IC} [\si{\degree}]$ & $\varPhi_0^{90\,\%}$ \\
\hline
3 & -31.2 & 127.9 & 1.4 & 2.1 \\
5 & -0.4 & 110.6 & 1.2 & 1.5 \\
8 & -21.2 & 182.4 & 1.3 & 1.7 \\
13 & 40.3 & 67.9 & 1.2 & 2.4 \\
18 & -24.8 & 345.6 & 1.3 & 2.0 \\
23 & -13.2 & 208.7 & 1.9 & 1.7 \\
28 & -71.5 & 164.8 & 1.3 & 1.2 \\
37 & 20.7 & 167.3 & 1.2 & 1.7 \\
38 & 14.0 & 93.3 & 1.2 & 2.1 \\
43 & -22.0 & 206.6 & 1.3 & 1.3 \\
44 & 0.0 & 336.7 & 1.2 & 1.8 \\
45 & -86.3 & 219.0 & 1.2 & 1.2 \\
53 & -37.7 & 239.0 & 1.2 & 1.6 \\
\end{tabular}%
\end{table}
\subsection{Galactic Center region}
The restricted search region is defined as an ellipse around the Galactic Center with semi-axes of \unit{20}{\degree} in galactic longitude and \unit{15}{\degree} in galactic latitude.
Due to the smaller search area, the search for astrophysical sources is more sensitive than a full sky search because it is less probable for background events to randomly cluster together, mimicking the signature of a signal.
Assuming the usual $E^{-2}$ spectrum, the most significant cluster found in this restricted region is located at $(\alpha,\delta) = (\unit{257.4}{\degree},\unit{-41.0 }{\degree})$ with a pre-trial p-value of 0.09\% and a fitted number of signal events of 2.3. The post-trial significance of this cluster, calculated as in the full sky search but in the restricted region around the Galactic Center, is 60\%. Other spectral indices ($\gamma$ = 2.1, 2.3, 2.5) and source extensions ($\sigma = \unit{0.5}{\degree}$, $\unit{1.0}{\degree}$, $\unit{2.0 }{\degree}$) are considered, yielding different most significant clusters. The source extension is quantified by the $\sigma$ of the gaussian distribution.
For a spectral index of $\gamma$ = 2.5 and a point-source, the most significant cluster is found at $(\alpha,\delta) = (\unit{273.0}{\degree},\unit{-42.2 }{\degree})$, with a pre-trial p-value of 0.02\% and a post-trial significance of 30\%. The distribution of events for these two clusters is shown in Fig.~\ref{fig:AllClusters}-middle-right and bottom-left. The positions of the most significant clusters found for the remaining spectral indices and source extensions considered are within 1$^\circ$ from the position of the latter.
The declination-dependent limit of such a restricted point-like source search is shown in Fig.~\ref{fig:SensGal}, both for different energy spectral indices $\gamma$ and different source extensions. The upper limits increase with increasing values of $\gamma$ and with the source extension. A softer energy spectrum of cosmic neutrinos (larger values of the spectral index $\gamma$) is less distinguishable from the spectrum of atmospheric neutrinos, as is a source with a larger extension.
For a softer spectrum, fewer neutrinos are emitted by the source within an energy range in which they can be statistically separated from atmospheric neutrinos. The flux required at the normalization point for a significant detection is therefore larger.
\begin{figure}
\includegraphics[width=0.48\textwidth]{GC_limits.pdf}
\includegraphics[width=0.48\textwidth]{GC_limits_width.pdf}
\caption{90\% C.L. upper limits of a search restricted to the region around the origin of the galactic coordinates at ($\alpha$, $\delta$) = (266.40$^\circ$,--28.94$^\circ$) assuming different spectral indices for the neutrino flux (left) and different source extensions for $\gamma$ = 2 (right). }
\label{fig:SensGal}
\end{figure}
\subsection{Sagittarius A*}
Super-massive black holes are strong candidates to be accelerators of very-high energy cosmic rays and therefore for cosmic neutrino production \cite{LightHouseSgrA}.
Additionally, due to the high concentration of candidate sources and gas around the Galactic Center (GC), it is probable that an extended signal from that region will be detected before identifying individual point-like sources.
For this reason, Sagittarius A*, located at $(\alpha,\delta) = (\unit{266.42}{\degree},\unit{-29.01}{\degree}$), is investigated as an extended source with widths between \unit{0.5}{\degree} and \unit{5}{\degree}. The cluster of events around Sagittarius A* reconstructed by ANTARES is shown in Fig.~\ref{fig:AllClusters}-bottom-right.
The sensitivity and upper limits for the assumption of different source extensions can be seen in Fig.~\ref{fig:SensExt}. The sensitivity degrades with increasing extension but an improvement of up to a factor of $2.7$ can be achieved by assuming an extended source with the simulated extension. The largest excess above the background is found at an extension of \unit{0}{\degree} with a pre-trial p-value of 22\%.
\begin{figure}[!h]
\centering
\includegraphics[width=0.6\textwidth]{SgrAStar.pdf}
\caption{Discovery flux (dotted red), median sensitivity (dotted blue) and 90\,\% C.L. upper limits (green) for a search for an extended source at Sagittarius A* at $(\alpha,\delta) = (\unit{266.42}{\degree},\unit{-29.01}{\degree}$) assuming different angular extensions $\sigma$. The dashed lines correspond to the point-like source assumption.
}
\label{fig:SensExt}
\end{figure}
\section{Systematic uncertainties}
\label{sec:SYST}
The effects of systematic uncertainties on the absolute pointing accuracy, angular resolution, acceptance and the background rate distribution of events are evaluated.
\emph{Absolute Pointing Accuracy Uncertainty.} An uncertainty of 0.13$^\circ$ and 0.06$^\circ$ on the horizontal ($\phi$) and vertical ($\theta$) directions, respectively, was established in a previous study \cite{PointingAcc}. To take this into account, randomly generated offsets have been added to the $\phi$ and $\theta$ variables of the simulated events. The offsets are generated according to two Gaussian distributions with the aforementioned uncertainties as sigmas.
\emph{Angular Resolution Uncertainty.} The angular resolution of the track reconstruction algorithm can be affected by the accuracy of the detected hit times. A smearing of these times was performed in simulations leading to a 15\% degradation on the angular resolution in the track channel \cite{antPS4y}. For neutrinos of the shower sample, the reconstruction of the direction depends most significantly on the recorded charge. A smearing
in the measured charges \cite{ChargeResol} leads to a 12\% degradation of the angular resolution for the shower channel.
\emph{Acceptance Uncertainty.} A 15\% uncertainty on the acceptance has been considered for the calculation of the reported fluxes. This uncertainty was calculated after performing simulations with a reduction of the OM efficiency by 15\% \cite{antPS4y}.
\emph{Background Uncertainty.} In order to account for possible systematic uncertainties on the background, the distribution of the background rates in Fig.~\ref{fig:background} are parametrized by two different spline functions, $R(\delta)$ and $B(\delta)$ (the red and blue lines). The declination-dependent distribution of background events of the pseudo-experiments is determined as $\mathscr B(\delta) = B(\delta) + r \cdot ( R(\delta) - B(\delta) )$, with $r$ being a random number drawn for each pseudo-experiment from a uniform distribution between -1 and 1.
It is found that not considering these uncertainties would improve the median sensitivity at $90\,\%$ C.L. and the $5 \sigma$ discovery potential by less than $5\,\%$.
\section{Conclusion and outlook}\label{sec:CONCL}
Various searches for cosmic neutrino sources using combined information from the track and shower channels have been presented.
These searches provide the most sensitive limits for a large fraction of the Southern Sky, especially at neutrino energies below 100 TeV. No significant evidence of cosmic neutrino sources has been found. The IceCube HESE accumulation reported near the Galactic Center could neither be totally attributed to a point-like source nor to an extended source.
The most significant cluster in the full sky search is located at $(\alpha,\delta) = (\unit{343.8}{\degree}, \unit{23.5}{\degree})$ with a post-trial significance of $5.9\,\%$ or $1.9\sigma$.
Upper limits on the neutrino flux from 106 astrophysical candidates and 13 IceCube muon tracks have been presented.
The most significant source candidate is HESSJ0632+057 -- located at $(\alpha,\delta) = (\unit{98.24}{\degree},\unit{5.81}{\degree})$ -- with a post-trial significance of $1.5\sigma$. The upper limit on the signal from this candidate is $E^2 d\varPhi/dE = \unit{2.40e-8}{\giga\electronvolt\,\centi\metre\tothe{$-2$}\,\second\tothe{$-1$}}$.
The most significant cluster of events close to the Galactic Center when assuming a point-like source with an $E^{-2}$ energy spectrum is located at $(\alpha,\delta) = (\unit{-102.6}{\degree},\unit{-41.0 }{\degree})$ with a post-trial p-value of 60\%.
Sagittarius A* as a possible extended source has been investigated. Upper limits for the flux and number of events assuming a Gaussian morphology with different extensions have been presented. The largest excess over the background is observed at an angular extension of \unit{0}{\degree} with a pre-trial p-value value of 22\%.\\
The KM3NeT/ARCA neutrino telescope \cite{km3netLOI}, which is currently under construction, will combine a cubic kilometer-sized detector with the same high visibility towards the Galactic Center as ANTARES. It is expected that this detector will be able to make definite statements about a neutrino flux from several Galactic candidates within a few years of operation.
\section*{Acknowledgements}
The authors acknowledge the financial support of the funding agencies:
Centre National de la Recherche Scientifique (CNRS), Commissariat \`a
l'\'ener\-gie atomique et aux \'energies alternatives (CEA),
Commission Europ\'eenne (FEDER fund and Marie Curie Program),
Institut Universitaire de France (IUF), IdEx program and UnivEarthS
Labex program at Sorbonne Paris Cit\'e (ANR-10-LABX-0023 and
ANR-11-IDEX-0005-02), Labex OCEVU (ANR-11-LABX-0060) and the
A*MIDEX project (ANR-11-IDEX-0001-02),
R\'egion \^Ile-de-France (DIM-ACAV), R\'egion
Alsace (contrat CPER), R\'egion Provence-Alpes-C\^ote d'Azur,
D\'e\-par\-tement du Var and Ville de La
Seyne-sur-Mer, France;
Bundesministerium f\"ur Bildung und Forschung
(BMBF), Germany;
Istituto Nazionale di Fisica Nucleare (INFN), Italy;
Stichting voor Fundamenteel Onderzoek der Materie (FOM), Nederlandse
organisatie voor Wetenschappelijk Onderzoek (NWO), the Netherlands;
Council of the President of the Russian Federation for young
scientists and leading scientific schools supporting grants, Russia;
National Authority for Scientific Research (ANCS), Romania;
Mi\-nis\-te\-rio de Econom\'{\i}a y Competitividad (MINECO):
Plan Estatal de Investigaci\'{o}n (refs. FPA2015-65150-C3-1-P, -2-P and -3-P, (MINECO/FEDER)), Severo Ochoa Centre of Excellence and MultiDark Consolider (MINECO), and Prometeo and Grisol\'{i}a programs (Generalitat
Valenciana), Spain;
Ministry of Higher Education, Scientific Research and Professional Training, Morocco.
We also acknowledge the technical support of Ifremer, AIM and Foselev Marine
for the sea operation and the CC-IN2P3 for the computing facilities.
\clearpage
| 83bd71f982646c3f2f2e7cc690218ba8d1f1acf9 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Introduction}
Considering non-commutative (nc for short) polynomials (elements of the
\emph{free associative algebra}) as elements in its
\emph{universal field of fractions} (free field)
seems ---at a first glance--- like to take a sledgehammer
to crack a nut. It is maybe more common to view them
as \emph{nc rational series} (or \emph{nc formal power series}).
Therefore one could skip the rather complicated theory behind
free fields, briefly introduced in
\cite[Section~9.3]{Cohn2003b
. For details we refer to
\cite[Section~6.4]{Cohn1995a
.
However, as we shall see, the \emph{normal form}
of \ifJOURNAL\else Cohn and Reutenauer \fi
\cite{Cohn1994a
\ provides new insights.
For a given element (in the \emph{free field}) \emph{minimal}
linear representations can be transformed into each other
by invertible matrices over the ground field.
In our case we are interested in the (finite set of)
transformation matrices which yield \emph{all} possible
factorizations (modulo insertion of units and permutation of
commuting factors)
of a given polynomial.
Here we use the concept of \emph{admissible linear systems}
(ALS for short) of
\cite{Cohn1972a
, closely related to linear representations.
At any step, an ALS (for a nc polynomial)
can be easily transformed into a \emph{proper linear system}
\cite[Section~II.1]{Salomaa1978a
.
A more general concept for the factorization of arbitrary
elements in free fields (for example non-commutative
rational formal power series) is considered in future work.
There (left and right) divisors will be defined
on the level of linear representations. In this way
additional structure compensates what is missing in fields
(when every non-zero element is invertible).
\medskip
Section~\ref{sec:pf.faa} provides the notation and the basic
(algebraic) setup.
Section~\ref{sec:pf.fnp} contains the main result,
Theorem~\ref{thr:pf.factorization}, which describes a correspondence
between factorizations and upper right blocks of zeros
(in the system matrix). To be able to formulate it,
Proposition~\ref{pro:pf.minmul} (``minimal'' polynomial multiplication)
is necessary. The main idea of its proof can be further developed
to Algorithm~\ref{alg:pf.minals} (minimizing ``polynomial'' admissible
linear systems).
Section~\ref{sec:pf.ncm} generalizes the concept
of companion matrices to provide immediately \emph{minimal}
linear representations for a broader class of elements
in the free associative algebra.
Section~\ref{sec:pf.ex} illustrates the concept step by
step.
\medskip
To our knowledge the literature on the factorization of elements in
free associative algebras is rather sparse.
\ifJOURNAL\else Caruso \fi \cite{Caruso2010a
\ describes ideas of J.~Davenport using homogenization.
We have not yet investigated how their approach compares to
ours. Some special cases (for example \emph{variable disjoint}
factorizations) are treated in \cite{Arvind2015b
.
Here we do not consider factorizations in skew polynomial rings
(or rather \emph{domains}), or ---more general--- rings satisfying
the Ore condition
\cite[Section~7.1]{Cohn2003b
\ or
\cite[Section~0.8]{Cohn1985a
.
A starting point in this context is
\cite{Heinle2013a
\ or
\cite{Levandovskyy2018a
. Factorizations of skew polynomials have various
connections to other areas, see for example
\cite{Retakh2010a
\ and
\cite{Gelfand2001a
, just to mention two.
\section{Representing Non-Commutative Polynomials}\label{sec:pf.faa}
There are much simpler ways for representing elements
in free associative algebras (in such a way that addition and multiplication
can be defined properly) than the following general presented here.
However, including the inverse we keep the full flexibility,
which could be used later to compute the left (or right)
greatest common divisor of two polynomials $p,q$ by minimizing
the linear representation for $p^{-1} q$ in which common left factors
would cancel.
\begin{notation}
The set of the natural numbers is denoted by $\mathbb{N} = \{ 1,2,\ldots \}$,
that including zero by $\mathbb{N}_0$.
Zero entries in matrices are usually replaced by (lower) dots
to stress the structure of the non-zero entries
unless they result from transformations where there
were possibly non-zero entries before.
We denote by $I_n$ the identity matrix of size $n$
respectively $I$ if the size is clear from the context.
\end{notation}
\smallskip
Let $\field{K}$ be a \emph{commutative} field,
$\aclo{\field{K}}$ its algebraic closure and
$X = \{ x_1, x_2, \ldots, x_d\}$ be a \emph{finite} alphabet.
$\freeALG{\field{K}}{X}$ denotes the \emph{free associative
algebra} or \emph{free $\field{K}$-algebra}
(or ``algebra of non-commutative polynomials'')
and $\freeFLD{\field{K}}{X}$ denotes the \emph{universal field of
fractions} (or ``free field'') of $\freeALG{\field{K}}{X}$,
see
\cite{Cohn1995a
,
\cite{Cohn1999a
. In our examples the alphabet is usually $X=\{x,y,z\}$.
Including the algebra of nc \emph{rational} series $\ncRATS{\field{K}}{X}$
we have the following chain of inclusions:
$\field{K}\subsetneq \freeALG{\field{K}}{X}
\subsetneq \ncRATS{\field{K}}{X}
\subsetneq \freeFLD{\field{K}}{X}$.
The \emph{free monoid}\index{free monoid} $X^*$ generated by $X$
is the set of all
\emph{finite words}\index{finite word}
$x_{i_1} x_{i_2} \cdots x_{i_n}$ with $i_k \in \{ 1,2,\ldots, d \}$.
An element of the alphabet is called \emph{letter}\index{letter},
an element of $X^*$ is called \emph{word}\index{word}.
The multiplication on $X^*$ is just the \emph{concatenation}\index{concatenation}
of words, that is,
$(x_{i_1} \cdots x_{i_m})\cdot (x_{j_1} \cdots x_{j_n})
= x_{i_1} \cdots x_{i_m} x_{j_1} \cdots x_{j_n}$,
with neutral element $1$, the \emph{empty word}.
The \emph{length} of a word $w=x_{i_1} x_{i_2} \cdots x_{i_m}$ is $m$,
denoted by $\length{w} = m$ or $\ell(w) = m$.
For detailled introductions see
\cite[Chapter~1]{Berstel2011a
\ or
\cite[Section~I.1]{Salomaa1978a
.
\begin{definition}[Inner Rank, Full Matrix, Hollow Matrix,
see
\cite{Cohn1985a
, \cite{Cohn1999a
]\label{def:pf.full}
\index{inner rank}\index{full matrix}\index{hollow matrix}%
Given a matrix $A \in \freeALG{\field{K}}{X}^{n \times n}$, the \emph{inner rank}
of $A$ is the smallest number $m\in \mathbb{N}$
such that there exists a factorization
$A = T U$ with $T \in \freeALG{\field{K}}{X}^{n \times m}$ and
$U \in \freeALG{\field{K}}{X}^{m \times n}$.
The matrix $A$ is called \emph{full} if $m = n$,
\emph{non-full} otherwise.
It is called \emph{hollow} if it contains a zero submatrix of size
$k \times l$ with $k+l>n$.
\end{definition}
\begin{definition}[Associated Matrices,
\cite{Cohn1995a
]\label{def:pf.ass}
\index{associated matrix}%
Two matrices $A$ and $B$ over $\freeALG{\field{K}}{X}$ (of the same size)
are called \emph{associated} over a subring $R\subseteq \freeALG{\field{K}}{X}$
if there exist (over $R$) invertible matrices $P,Q$ such that
$A = P B Q$.
\end{definition}
\begin{lemma}[%
\protect{\cite[Corollary~6.3.6]{Cohn1995a
}]\label{lem:pf.cohn95.636}
A linear square matrix over $\freeALG{\field{K}}{X}$
which is not full is associated over $\field{K}$ to a linear
hollow matrix.
\end{lemma}
\begin{remark}
A hollow square matrix cannot be full
\cite[Section~3.2]{Cohn1985a
, illustrated in an example:
\begin{displaymath}
A =
\begin{bmatrix}
z & . & . \\
x & . & . \\
y & -x & 1
\end{bmatrix}
=
\begin{bmatrix}
z & 0 \\
x & 0 \\
0 & 1
\end{bmatrix}
\begin{bmatrix}
1 & 0 & 0 \\
y & -x & 1
\end{bmatrix}.
\end{displaymath}
\end{remark}
\begin{definition}[Linear Representations, Dimension, Rank%
\ifJOURNAL
, \cite{Cohn1994a,Cohn1999a
\else
\ \cite{Cohn1994a,Cohn1999a
\fi]\label{def:pf.rep}
\index{linear representation}%
Let $f \in \freeFLD{\field{K}}{X}$.
A \emph{linear representation} of $f$ is a triple $\pi_f$ = $(u,A,v)$ with
$u \in \field{K}^{1 \times n}$,
full $A = A_0 \otimes 1 + A_1 \otimes x_1 + \ldots
+ A_d \otimes x_d$, $A_\ell \in \field{K}^{n\times n}$ and
$v \in \field{K}^{n\times 1}$ such that $f = u A^{-1} v$.
The \emph{dimension} of the representation is $\dim \pi_f = n$.
It is called \emph{minimal} if $A$ has the smallest possible dimension
among all linear representations of $f$.
A minimal one $\pi_f$ defines the \emph{rank}\index{rank} of $f$,
$\rank f = \dim \pi_f$.
The ``empty'' representation $\pi = (,,)$ is the minimal one
of $0 \in \freeFLD{\field{K}}{X}$ with $\dim \pi = 0$.
\end{definition}
\begin{definition}[Left and Right Families
\cite{Cohn1994a
]\label{def:pf.family}
\index{left family}\index{right family}%
Let $\pi=(u,A,v)$ be a linear representation of $f \in \freeFLD{\field{K}}{X}$
of dimension $n$.
The families $( s_1, s_2, \ldots, s_n )\subseteq \freeFLD{\field{K}}{X}$
with $s_i = (A^{-1} v)_i$
and $( t_1, t_2, \ldots, t_n )\subseteq \freeFLD{\field{K}}{X}$
with $t_j = (u A^{-1})_j$
are called \emph{left family} and \emph{right family} respectively.
$L(\pi) = \linsp \{ s_1, s_2, \ldots, s_n \}$ and
$R(\pi) = \linsp \{ t_1, t_2, \ldots, t_n \}$
denote their linear spans (over $\field{K}$).
\end{definition}
\begin{proposition}[%
\cite{Cohn1994a
, Proposition 4.7]
A representation $\pi=(u,A,v)$ of an element $f \in \freeFLD{\field{K}}{X}$
is minimal if and only if both, the left family
and the right family are $\field{K}$-linearly independent.
\label{pro:pf.cohn94.47}
\end{proposition}
\begin{definition}[Admissible Linear Systems%
\ifJOURNAL ,\fi\
\cite{Cohn1972a
. Admissible Transformations]\label{def:pf.als}
\index{admissible linear system}\index{admissible transformation}%
A linear representation $\als{A} = (u,A,v)$ of $f \in \freeFLD{\field{K}}{X}$
is called \emph{admissible linear system} (for $f$),
denoted by $A s = v$,
if $u=e_1=[1,0,\ldots,0]$. The element $f$ is then the first component
of the (unique) solution vector $s$.
Given a linear representation $\als{A} = (u,A,v)$
of dimension $n$ of $f \in \freeFLD{\field{K}}{X}$
and invertible matrices $P,Q \in \field{K}^{n\times n}$,
the transformed $P\als{A}Q = (uQ, PAQ, Pv)$ is
again a linear representation (of $f$).
If $\als{A}$ is an ALS,
the transformation $(P,Q)$ is called
\emph{admissible} if the first row of $Q$ is $e_1 = [1,0,\ldots,0]$.
\end{definition}
Transformations can be done by elementary row- and column operations,
explained in detail in
\cite[Remark~1.12]{Schrempf2017a
. For further remarks and connections to the related concepts
of linearization and realization see
\cite[Section~1]{Schrempf2017a
.
For rational operations on ALS level see the following proposition.
If an ALS is \emph{minimal} then more refined versions of an
inverse give a minimal ALS again. For a detailled discussion
we refer to
\cite[Section~4]{Schrempf2017a
.
\begin{proposition}[Rational Operations%
\ifJOURNAL ,\fi\
\cite{Cohn1999a
]\label{pro:pf.ratop}
Let $0\neq f,g \in \freeFLD{\field{K}}{X}$ be given by the
admissible linear systems $\als{A}_f = (u_f, A_f, v_f)$
and $\als{A}_g = (u_g, A_g, v_g)$ respectively
and let $0\neq \mu \in \field{K}$.
Then admissible linear systems for the rational operations
can be obtained as follows:
\smallskip\noindent
The scalar multiplication
$\mu f$ is given by
\begin{displaymath}
\mu \als{A}_f =
\bigl( u_f, A_f, \mu v_f \bigr).
\end{displaymath}
The sum $f + g$ is given by
\begin{displaymath}
\als{A}_f + \als{A}_g =
\left(
\begin{bmatrix}
u_f & .
\end{bmatrix},
\begin{bmatrix}
A_f & -A_f u_f^{\!\top} u_g \\
. & A_g
\end{bmatrix},
\begin{bmatrix} v_f \\ v_g \end{bmatrix}
\right).
\end{displaymath}
The product $fg$ is given by
\begin{displaymath}
\als{A}_f \cdot \als{A}_g =
\left(
\begin{bmatrix}
u_f & .
\end{bmatrix},
\begin{bmatrix}
A_f & -v_f u_g \\
. & A_g
\end{bmatrix},
\begin{bmatrix}
. \\ v_g
\end{bmatrix}
\right).
\end{displaymath}
And the inverse $f^{-1}$ is given by
\begin{displaymath}
\als{A}_f^{-1} =
\left(
\begin{bmatrix}
1 & .
\end{bmatrix},
\begin{bmatrix}
-v_f & A_f \\
. & u_f
\end{bmatrix},
\begin{bmatrix}
. \\ 1
\end{bmatrix}
\right).
\end{displaymath}
\end{proposition}
\begin{definition}\label{def:pf.reg}
An element in $\freeFLD{\field{K}}{X}$ is called \emph{regular},
if it has a linear representation $(u,A,v)$ with $A = I - M$,
where $M = M_1 \otimes x_1 + \ldots + M_d \otimes x_d$
with $M_i \in \field{K}^{n \times n}$ for some $n\in \mathbb{N}$,
that is, $A_0 = I$ in Definition~\ref{def:pf.rep},
or equivalently, if $A_0$ is regular (invertible).
\end{definition}
\begin{remark}
$A = I - M$ is also called a \emph{monic pencil}\index{monic pencil}%
\ifJOURNAL ,\fi\
\cite{Helton2007b
. A regular element can also be represented by a
\emph{proper linear system} $s = v + M s$
\cite[Section~II.1]{Salomaa1978a
.
\end{remark}
\begin{remark}
For a polynomial $p \in \freeALG{\field{K}}{X}
\subseteq \ncRATS{\field{K}}{X} \subseteq \freeFLD{\field{K}}{X}$
the rank of $p$ is just the \emph{Hankel rank}\index{Hankel rank},
that is, the rank of its Hankel matrix\index{Hankel matrix}
$\mathcal{H}(p) = (h_{w_1, w_2})$ ---rows and columns are
indexed by words in the free monoid---
where $h_{w_1, w_2} \in \field{K}$ is the coefficient of the monomial
$w = w_1 w_2$ of $p$. See
\cite{Fliess1974a
\ and
\cite[Section~II.3]{Salomaa1978a
.
\end{remark}
\begin{example}
The Hankel matrix for $p = x(1-yx) = x- xyx$, with
row indices $[1, x, xy, xyx]$ and column indices $[1, x, yx, xyx]$,
is ---without zero rows $\{y,x^2,yx,y^2,$ $x^3,x^2 y, y x^2, \ldots\}$
and columns $\{y,x^2,xy,y^2,x^3,x^2 y, y x^2, \ldots\}$---
\begin{displaymath}
\mathcal{H}(p) =
\begin{bmatrix}
. & 1 & . & -1 \\
1 & . & -1 & . \\
. & -1 & . & . \\
-1 & . & . & . \\
\end{bmatrix}.
\end{displaymath}
Its rank is $4$. Thus $\rank p =4$ and therefore the dimension
of any \emph{minimal} admissible linear system
is~4, as will be seen later
in the ALS~\eqref{eqn:pf.minmul.1}
in Example~\ref{ex:pf.minmul},
where we show minimality by
Proposition~\ref{pro:pf.cohn94.47}.
\end{example}
\medskip
The following definitions follow mainly
\cite{Baeth2015a
\ and are streamlined to our purpose.
We do not need the full generality here.
While there is a rather uniform factorization theory
in the commutative setting
\cite[Section~1.1]{Geroldinger2006a
, even the ``simplest'' non-commutative case, that is,
a ``unique'' \emph{factorization domain} like the
\emph{free associative algebra}, is not straightforward.
For a general (algebraic) point of view we recommend the survey
\cite{Smertnig2015a
. Factorization in \emph{free ideal rings} (FIRs)
is discussed in detail in
\cite[Chapter~3]{Cohn1985a
. FIRs play an important role in the construction of free fields.
More on ``non-commutative'' factorization can
be found in
\cite{Jordan1989a
\ and
\cite{Bell2017a
\ (just to mention a few)
and the literature therein.
\begin{definition}[Similar Right Ideals, Similar Elements
\protect{\cite[Section~3.2]{Cohn1985a
}]
\index{similar right ideals}%
Let $R$ be a ring.
Two right ideals $\ideal{a},\ideal{b} \subseteq R$ are called
\emph{similar}, written as $\ideal{a} \sim \ideal{b}$,
if $R/\ideal{a} \cong R/\ideal{b}$ as right $R$-modules.
\index{similar elements}%
Two elements $p,q\in R$ are called \emph{similar} if
their right ideals $pR$ and $qR$ are similar,
that is, $pR \sim qR$.
See also
\cite[Section~4.1]{Smertnig2015a
.
\end{definition}
\begin{definition}[Left and Right Coprime Elements
\protect{\cite[Section~2]{Baeth2015a
}]\label{def:pf.lrcop}
Let $R$ be a domain and $H = R^\bullet = R \setminus \{ 0 \}$.
An element $p \in H$ \emph{left divides}\index{left divides} $q \in H$,
written as $p \!\mid_{\mkern-1mu\text{l}}\! q$, if $q \in pH = \{ ph \mid h \in H \}$.
Two elements $p,q$ are called \emph{left coprime}\index{left coprime}
if for all $h$ such that $h\!\mid_{\mkern-1mu\text{l}}\! p$ and $h\!\mid_{\mkern-1mu\text{l}}\! q$
implies $h \in H^\times = \{ f \in H \mid f \text{ is invertible} \}$,
that is, $h$ is an element of the \emph{group of units}\index{unit group}.
Right division $p \!\mid_{\mkern-1mu\text{r}}\! q$ and the notion of
\emph{right coprime} is defined in a similar way.
Two elements are called \emph{coprime}\index{coprime}
if they are left and right coprime.
\end{definition}
\begin{definition}[Atomic Domains, Irreducible Elements
\protect{\cite[Section~2]{Baeth2015a
}]\label{def:pf.atoms}
Let $R$ be a domain and $H = R^\bullet$.
An element $p\in H \setminus H^\times$, that is,
a non-zero non-unit (in $R$), is called an \emph{atom}\index{atom}
(or \emph{irreducible}\index{irreducible element})
if $p = q_1 q_2$ with $q_1,q_2 \in H$ implies that
either $q_1 \in H^\times$ or $q_2 \in H^\times$.
The set of atoms in $R$ is denoted by
$\mathbf{A}(R)$.
The (cancellative) monoid $H$ is called \emph{atomic}\index{atomic domain}
if every non-unit can be written as a finite product of atoms of $H$.
The domain $R$ is called \emph{atomic} if the monoid $R^\bullet$ is atomic.
\end{definition}
\begin{remark}
Similarity of two elements $a,a'$ in a \emph{weak Bezout ring} $R$
is equivalent to the existence of $b,b' \in R$ such that
$ab' = ba'$ with $ab'$ and $ba'$ coprime,
that is, $a$ and $b$ are left coprime and
$b'$ and $a'$ are right coprime.
The free associative algebra $\freeALG{\field{K}}{X}$ is a weak Bezout ring
\cite[Proposition~5.3 and Theorem~6.2]{Cohn1963b
.
\end{remark}
\begin{example}
The polynomials $p = (1-xy)$ and $q=(1-yx)$ are similar,
because $p x = (1-xy) x = x - xyx = x(1-yx) = x q$.
See also Example~\ref{ex:pf.minmul}.
\end{example}
\begin{example}
In the \emph{free monoid} $X^*$ the atoms are just the letters
$x_i \in X$.
\end{example}
\begin{definition}[Similarity Unique Factorization Domains
\protect{\cite[Definition~4.1]{Smertnig2015a
}]\label{def:pf.sfd}
\index{similarity factorization domain}\index{similarity-UFD}%
A domain $R$ is called \emph{similarity factorial}
(or a \emph{similarity-UFD}) if
$R$ is atomic and it satisfies the property that
if $p_1 p_2 \cdots p_m = q_1 q_2 \cdots q_n$ for atoms\index{atom}
(irreducible elements)\index{irreducible element}
$p_i,q_j \in R$,
then $m=n$ and there exists a permutation $\sigma \in \mathfrak{S}_m$
such that $p_i$ is similar to $q_{\sigma(i)}$ for all $i\in 1,2, \ldots, m$.
\end{definition}
\begin{proposition}[%
\protect{\cite[Theorem~6.3]{Cohn1963b
}]\label{pro:pf.cohn63b}
The free associative algebra $\freeALG{\field{K}}{X}$
is a similarity (unique) factorization domain.
\end{proposition}
\section{Factorizing non-commutative Polynomials}\label{sec:pf.fnp}
Our concept for the factorization of nc polynomials
(Theorem~\ref{thr:pf.factorization})
relies on \emph{minimal} linear representations.
Beside the ``classical'' algorithms from
\cite{Cardon1980a
\ and
\cite{Fornasini1980a
, there is a naive one
illustrated in Section~\ref{sec:pf.ex}.
The latter is not very efficient in general for an
alphabet with more than one letter
but it preserves the form defined in the following.
After formulating a \emph{minimal polynomial multiplication}
in Proposition~\ref{pro:pf.minmul}
and illustrating it in an example
we present an algorithm which works directly
on the system matrix (of the admissible linear system).
Using \emph{proper linear systems}
\cite[Section~II.1]{Salomaa1978a
\ would be slightly too restrictive
because the only possible admissible transformations are
conjugations (with respect to the system matrix).
That these are not sufficient can be seen in
Example~\ref{ex:pf.irred}.
On the other hand, \emph{admissible linear systems} are too general.
Therefore we define a form that is suitable for our purpose.
\begin{remark}
Although we use admissible linear systems,
restricting the application of the rational operations
(excluding the inverse) in Proposition~\ref{pro:pf.ratop}
to (systems for) polynomials only, one gets again polynomials.
If the inverse is restricted to
(systems for) rational formal power series
with \emph{non-vanishing} constant coefficient
one gets again rational formal power series.
Since we are using multiplication only,
Theorem~\ref{thr:pf.factorization} does \emph{not}
rely on the (construction of the) free field.
\end{remark}
\begin{definition}[Pre-Standard ALS, Pre-Standard Admissible Transformation]%
\label{def:pf.psals}
An admissible linear system $\als{A} = (u,A,v)$
of dimension $n$
with system matrix $A = (a_{ij})$
for a non-zero polynomial $0 \neq p \in \freeALG{\field{K}}{X}$
is called
\emph{pre-standard}\index{pre-standard ALS}, if
\begin{itemize}
\item[(1)] $v = [0,\ldots,0,\lambda]^{\!\top}$ for some $\lambda \in\field{K}$ and
\item[(2)] $a_{ii}=1$ for $i=1,2,\ldots, n$ and $a_{ij}=0$ for $i>j$,
that is, $A$ is upper triangular.
\end{itemize}
A pre-standard ALS is also written as $\als{A} = (1,A,\lambda)$
with $1,\lambda \in \field{K}$.
An admissible transformation $(P,Q)$ for an ALS $\als{A}$
is called \emph{pre-standard}\index{pre-standard admissible transformation},
if the transformed system $P\als{A} Q$ is (again) pre-standard.
\end{definition}
\subsection{Minimal Multiplication}\label{sec:pf.fnp.mm}
To be able to prove that the construction
in Proposition~\ref{pro:pf.minmul} leads to a \emph{minimal}
linear representation (for the product of two nc polynomials)
some preparation is necessary.
One of the main tools is Lemma~\ref{lem:pf.cohn95.636}
\cite[Corollary~6.3.6]{Cohn1995a
.
Although we are working with regular elements only,
invertibility of the constant coefficient matrix $A_0$
(in the system matrix) does not have to be assumed in
Lemma~\ref{lem:pf.rt1}.
\begin{lemma}\label{lem:pf.rt1}
Let $\als{A} = (u,A,v)$ be an ALS
of dimension $n \ge 1$ with $\field{K}$-linearly independent
left family $s = A^{-1} v$ and
$B = B_0 \otimes 1 + B_1 \otimes x_1 + \ldots + B_d \otimes x_d$
with $B_\ell\in \field{K}^{m\times n}$, such that
$B s = 0$. Then there exists a (unique) $T \in \field{K}^{m \times n}$
such that $B = TA$.
\end{lemma}
\begin{proof}
The trivial case $n=1$ implies $B=0$ and therefore $T=0$.
Now let $n \ge 2$.
Without loss of generality assume that $v = [0,\ldots,0,1]^{\!\top}$
and $m=1$.
Since $A$ is full and thus invertible over the free field,
there exists a unique $T$ such that $B = TA$, namely
$T = B A^{-1}$ in $\freeFLD{\field{K}}{X}^{1 \times n}$.
The last column in $T$ is zero because
$0 = Bs = BA^{-1} v = Tv$.
Now let $A'$ denote the matrix $A$ whose last row is removed
and $A'_B$ the matrix obtained from $A$ when the last row is replaced by
$B$. $A'_B$ cannot be full since $s \in \ker A'_B$
would give a contradiction: $s = (A'_B)^{-1} 0 = 0$.
We claim that there is only one possibility to transform $A'_B$
to a hollow matrix, namely with zero last row. If we cannot produce a
$(n-i) \times i$ block of zeros (by invertible transformations) in the
first $n-1$ rows of $A'_B$, then we cannot get
blocks of zeros of size $(n-i+1) \times i$ and we are done.
Now assume to the contrary that there are invertible matrices
$P' \in \field{K}^{(n-1) \times (n-1)}$
and (admissible) $Q\in\field{K}^{n \times n}$
with $(Q^{-1} s)_1 = s_1$,
such that $P'A'Q$ contains a zero block of size $(n-i) \times i$
for some $i=1,\ldots, n-1$. There are two cases. If the first $n-i$
entries in column~1 cannot be made zero, we construct an upper right
zero block:
\begin{displaymath}
\hat{A} =
\begin{bmatrix}
A_{11} & . \\
A_{21} & A_{22}
\end{bmatrix},
\quad \hat{s} = Q^{-1} s
\quad\text{and}\quad \hat{v} = P v = v
\end{displaymath}
where $A_{11}$ has size $(n-i) \times (n-i)$.
If $A_{11}$ \emph{were not} full,
then $A$ would not be full (the last row is not involved
in the transformation). Hence this pivot block is
invertible over the free field.
Therefore $\hat{s}_1 = \hat{s}_2 = \ldots = \hat{s}_{n-i} = 0$.
Otherwise we construct an upper left zero block in $PAQ$.
But then
$\hat{s}_{i+1} = \hat{s}_{i+2} = \ldots = \hat{s}_{n} = 0$.
Both contradict $\field{K}$-linear independence of the left family.
Hence, by Lemma~\ref{lem:pf.cohn95.636},
$A'_B$ is associated over $\field{K}$
to a linear hollow matrix with a $1 \times n$ block of zeros,
say in the last row (columns are left untouched):
\begin{displaymath}
\begin{bmatrix}
I_{n-1} & . \\
T' & 1
\end{bmatrix}
\begin{bmatrix}
A' \\
B
\end{bmatrix}
I_n
=
\begin{bmatrix}
A' \\
.
\end{bmatrix}.
\end{displaymath}
The matrix $T = [-T', 0] \in \field{K}^{1 \times n}$ satisfies
$B = TA$.
\end{proof}
\begin{remark}
Although the ALS in Lemma~\ref{lem:pf.rt1}
does not have to be minimal,
$\field{K}$-linear independence of the left family
is an important assumption for two reasons.
One is connected to ``pathological'' situations,
compare with
\cite[Example~2.5]{Schrempf2017a
.
An entry corresponding
to some $s_j=0$, say for $j=3$, could be arbitrary:
\begin{displaymath}
\begin{bmatrix}
1 & -x & . \\
. & . & z \\
. & 1 & -1
\end{bmatrix}
s
=
\begin{bmatrix}
. \\ . \\ 1
\end{bmatrix}.
\end{displaymath}
For $B = [2, -2x, y]$ the transformation $T$ has non-scalar entries: $T = [2, y z^{-1}, 0]$.
The other reason concerns the exclusion of other possibilities for non-fullness
except the last row. For $B= [0, 0, 1]$ the matrix
\begin{displaymath}
A'_B =
\begin{bmatrix}
1 & -x & . \\
. & . & z \\
. & . & 1
\end{bmatrix}
\end{displaymath}
is hollow.
However, the
transformation we are looking for here is $T = [0, z^{-1}, 0]$.
\end{remark}
\begin{lemma}\label{lem:pf.min1}
Let $\als{A} = (u,A,v)$ be
a pre-standard ALS of dimension $n\ge 2$
with $\field{K}$-linearly dependent left family $s=A^{-1} v$.
Let $A = (a_{ij})$.
Let $m \in \{ 2, 3, \ldots, n \}$ be the minimal index
such that the left subfamily $\underline{s} = (A^{-1} v)_{i=m}^n$
is $\field{K}$-linearly independent.
Then there exist matrices $T,U \in \field{K}^{1 \times (n+1-m)}$
such that
\begin{displaymath}
U + (a_{m-1,j})_{j=m}^n - T(a_{ij})_{i,j=m}^n =
\begin{bmatrix}
0 & \ldots & 0
\end{bmatrix}
\quad\text{and}\quad
T(v_i)_{i=m}^n = 0.
\end{displaymath}
\end{lemma}
\begin{proof}
By assumption, the left subfamily $(s_{m-1}, s_m, \ldots, s_n)$
is $\field{K}$-linearly dependent.
Thus there are $\kappa_j \in \field{K}$ such that
$s_{m-1} = \kappa_m s_m + \kappa_{m+1} s_{m+1} + \ldots + \kappa_n s_n$.
Let $U = [\kappa_m, \kappa_{m+1}, \ldots, \kappa_n ]$.
Then $s_{m-1} - U \underline{s} = 0$.
Since $\als{A}$ is pre-standard, $v_{m-1}=0$.
Now we can apply Lemma~\ref{lem:pf.rt1} with
$B = U + [ a_{m-1,m}, a_{m-1,m+1}, \ldots, a_{m-1,n}]$
(and $\underline{s}$). Hence, there exists
a matrix $T\in \field{K}^{1 \times (n+1-m)}$ such that
\begin{equation}\label{eqn:pf.min1}
U +
\begin{bmatrix}
a_{m-1,m} & \ldots & a_{m-1,n}
\end{bmatrix}
- T
\begin{bmatrix}
1 & a_{m,m+1} & \ldots & a_{m,n} \\
. & \ddots & \ddots & \vdots \\
. & . & 1 & a_{n-1,n} \\
. & . & . & 1
\end{bmatrix}
=
\begin{bmatrix}
0 & \ldots & 0
\end{bmatrix}.
\end{equation}
Recall that the last column of $T$ is zero,
whence $T(v_i)_{i=m}^n = 0$.
\end{proof}
\begin{proposition}[Minimal Polynomial Multiplication]\label{pro:pf.minmul}
Let $0\ne p,q \in \freeALG{\field{K}}{X}$ be given by the
\emph{minimal} pre-standard admissible linear systems
$A_p = (u_p, A_p, v_p) = (1, A_p, \lambda_p)$ and
$A_q = (u_q, A_q, v_q) = (1, A_q, \lambda_q)$ of dimension $n_p,n_q \ge 2$ respectively.
Then a \emph{minimal} ALS for $pq$ has dimension $n = n_p + n_q -1$
and can be constructed in the following way:
\begin{itemize}
\item[(1)] Construct the following ALS $\als{A}'=(u',A',v')$ for the product $pq$:
\begin{displaymath}
\begin{bmatrix}
A_p & -v_p u_q \\
. & A_q
\end{bmatrix}
s' =
\begin{bmatrix}
. \\ v_q
\end{bmatrix}.
\end{displaymath}
\item[(2)] Add $\lambda_p$-times column~$n_p$ to column~$n_p+1$
(the entry $s_{n_p}'$ becomes zero).
\item[(3)] Remove row~$n_p$ of $A'$ and $v'$ and column~$n_p$ of $A'$ and $u'$ and
denote the new (pre-standard) ALS of dimension $n_p + n_q -1$
by $\als{A} = (u,A,v) = (1,A,\lambda)$.
\end{itemize}
\end{proposition}
\begin{proof}
The left family of $\als{A}'$ is
\begin{displaymath}
s' =
\begin{bmatrix}
A_p^{-1} & A_p^{-1} v_p u_q \\
. & A_q^{-1}
\end{bmatrix}
\begin{bmatrix}
. \\ v_q
\end{bmatrix}
=
\begin{bmatrix}
s_p q \\ s_q
\end{bmatrix}.
\end{displaymath}
Clearly, $\als{A}$ is again pre-standard with $\lambda = \lambda_q$.
Both systems $\als{A}_p$ and $\als{A}_q$ are minimal.
Therefore their left and right families
are $\field{K}$-linearly independent.
Without loss of generality assume that $\lambda_p = 1$.
Then the last entry $t_{n_p}^p$ of the right family of $\als{A}_p$
is equal to $p$. Let $k = n_p$.
We have to show that both, the left and the right family
\begin{align*}
s &= (s_1, s_2, \ldots, s_n)
= (s_1^p q,\ldots, s_{k-1}^p q, q, s_2^q, \ldots, s_{n_q}^q), \\
t &= (t_1, t_2, \ldots, t_n)
= (t_1^p, \ldots, t_{k-1}^p, p, p t_2^q, \ldots, p t_{n_q}^q)
\end{align*}
of $\als{A}$ are $\field{K}$-linearly independent
respectively.
Assume the contrary for $s$, say there
is an index $1< m \le k$ such that $(s_{m-1}, s_m, \ldots, s_n)$
is $\field{K}$-linearly dependent while $(s_m, \ldots, s_n)$
is $\field{K}$-linearly independent.
Then, by Lemma~\ref{lem:pf.min1} there exist
matrices $T,U \in \field{K}^{1 \times (n-m+1)}$
such that $\eqref{eqn:pf.min1}$ holds and therefore
invertible matrices $P,Q \in \field{K}^{n \times n}$,
\begin{displaymath}
P =
\begin{bmatrix}
I_{m-2} & . & . \\
. & 1 & T \\
. & . & I_{n-m+1}
\end{bmatrix}
\quad\text{and}\quad
Q =
\begin{bmatrix}
I_{m-2} & . & . \\
. & 1 & U \\
. & . & I_{n-m+1}
\end{bmatrix},
\end{displaymath}
that yield equation $s_{m-1} = 0$ (in row~$m-1$) in $P\als{A}Q$.
Let $\tilde{P}$ (respectively $\tilde{Q}$) be the upper
left part of $P$ (respectively $Q$) of size $k \times k$.
Then the equation in row~$m-1$ in $\tilde{P} \als{A}_p \tilde{Q}$
is $s^p_{m-1} = \alpha \in \field{K}$,
contradicting $\field{K}$-linear independence of the left
family of $\als{A}_p$ since $s^p_{k} = \lambda_p \in \field{K}$.
A similar argument (and a variant of Lemma~\ref{lem:pf.min1})
for the right family yields its
$\field{K}$-linear independence.
Hence, by Proposition~\ref{pro:pf.cohn94.47}, the admissible linear
system $\als{A}$ (for $pq$) is minimal.
\end{proof}
\begin{corollary}
Let $0\neq p,q \in \freeALG{\field{K}}{X}$.
Then $\rank(pq) = \rank(p) + \rank(q) - 1$.
\end{corollary}
\begin{Example}\label{ex:pf.minmul}
The polynomials $p = x \in\freeALG{\field{K}}{X}$
and $q = 1-yx \in\freeALG{\field{K}}{X}$ have the minimal
pre-standard admissible linear systems
\begin{displaymath}
\als{A}_p = \left(
1,
\begin{bmatrix}
1 & - x \\
. & 1
\end{bmatrix},
1
\right)
\quad\text{and}\quad
\als{A}_q = \left(
1,
\begin{bmatrix}
1 & y & -1 \\
. & 1 & -x \\
. & . & 1
\end{bmatrix},
1
\right)
\end{displaymath}
respectively. Then a pre-standard ALS for $pq = x(1-yx)$ is given by
\begin{displaymath}
\begin{bmatrix}
1 & -x & . & . & . \\
. & 1 & -1 & . & . \\
. & . & 1 & y & -1 \\
. & . & . & 1 & -x \\
. & . & . & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ . \\ . \\ 1
\end{bmatrix},
\quad
s =
\begin{bmatrix}
x(1-yx) \\
1-yx \\
1-yx \\
x \\
1
\end{bmatrix}.
\end{displaymath}
Adding column~2 to column~3 (and subtracting $s_3$ from $s_2$)
yields
\begin{displaymath}
\begin{bmatrix}
1 & -x & -x & . & . \\
. & 1 & 0 & . & . \\
. & . & 1 & y & -1 \\
. & . & . & 1 & -x \\
. & . & . & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ . \\ . \\ 1
\end{bmatrix},
\quad
s =
\begin{bmatrix}
x(1-yx) \\
0 \\
1-yx \\
x \\
1
\end{bmatrix},
\end{displaymath}
thus the pre-standard ALS
\begin{equation}\label{eqn:pf.minmul.1}
\als{A} = \left(
\begin{bmatrix}
1 & . & . & .
\end{bmatrix},
\begin{bmatrix}
1 & -x & 0 & 0 \\
. & 1 & y & -1 \\
. & . & 1 & -x \\
. & . & . & 1
\end{bmatrix},
\begin{bmatrix}
. \\ . \\ . \\ 1
\end{bmatrix}
\right).
\end{equation}
Since also the right family $t = [1, x, -xy, x(1-yx)]$ is
$\field{K}$-linearly independent, this system is
\emph{minimal} by Proposition~\ref{pro:pf.cohn94.47}.
Note the upper right $1 \times 2$ block of zeros in the system matrix
of $\als{A}$.
\end{Example}
\subsection{Minimizing a pre-standard ALS}\label{sec:pf.fnp.min}
A close look on the proof of Proposition~\ref{pro:pf.minmul}
reveals a surprisingly simple algorithm
for the construction of a \emph{minimal}
pre-standard admissible linear system, provided that one
in pre-standard form is given. It can be used for minimizing
the ALS for the sum in Proposition~\ref{pro:pf.ratop}.
``Simple'' means that it can easily be done manually
also for quite large sparse systems. Depending on the data structure,
the implementation itself is somewhat technical.
One has to be very careful if the scalars (from the ground
field $\field{K}$) cannot be represented exactly,
especially when systems of linear equations (see below)
have to be solved.
To illustrate the main idea we (partially) minimize
a \emph{non-minimal} ``almost'' pre-standard
ALS $\als{A} = (u,A,v)$ of dimension $n=6$
for $p = -xy + (xy + z)$. Note that we do not need knowledge
of the left and right family at all.
Let
\begin{equation}\label{eqn:pf.min0}
\als{A} = \left(
\begin{bmatrix}
1 & . & . & . & . & .
\end{bmatrix},
\begin{bmatrix}
1 & -x & . & -1 & . & . \\
. & 1 & -y & . & . & . \\
. & . & 1 & . & . & . \\
. & . & . & 1 & -x & -z \\
. & . & . & . & 1 & -y \\
. & . & . & . & . & 1
\end{bmatrix},
\begin{bmatrix}
. \\ . \\ -1 \\ . \\ . \\ 1
\end{bmatrix}
\right).
\end{equation}
First we do one ``left'' minimization step,
that is, we remove (if possible)
one element of the $\field{K}$-linearly
dependent left family $s = A^{-1} v$
and construct a new system.
We fix a $1 \le k < n$, say $k=3$.
If we find a (pre-standard admissible)
transformation $(P,Q)$ of the form
\begin{equation}\label{eqn:pf.ltrf}
(P,Q) = \left(
\begin{bmatrix}
I_{k-1} & . & . \\
. & 1 & T \\
. & . & I_{n-k}
\end{bmatrix},
\begin{bmatrix}
I_{k-1} & . & . \\
. & 1 & U \\
. & . & I_{n-k}
\end{bmatrix}
\right)
\end{equation}
such that row~$k$ in $PAQ$ is $[0,0,1,0,0,0]$
and $(Pv)_k = 0$, we can eliminate row~$k$ and
column~$k$ in $P \als{A} Q$ because $(Q^{-1} s)_k = 0$.
How can we find these blocks $T,U \in \field{K}^{1 \times (n-k)}$?
We write $\als{A}$ in block form
---block row and column indices are
underlined to distinguish them from component indices---
(with respect to row/column~$k$)
\begin{equation}\label{eqn:pf.bloals}
\als{A}^{[k]} = \left(
\begin{bmatrix}
u_{\block{1}} & . & .
\end{bmatrix},
\begin{bmatrix}
A_{1,1} & A_{1,2} & A_{1,3} \\
. & 1 & A_{2,3} \\
. & . & A_{3,3}
\end{bmatrix},
\begin{bmatrix}
v_{\block{1}} \\ v_{\block{2}} \\ v_{\block{3}}
\end{bmatrix}
\right)
\end{equation}
and apply the transformation $(P,Q)$:
\begin{align*}
P A Q &=
\begin{bmatrix}
I_{k-1} & . & . \\
. & 1 & T \\
. & . & I_{n-k}
\end{bmatrix}
\begin{bmatrix}
A_{1,1} & A_{1,2} & A_{1,3} \\
. & 1 & A_{2,3} \\
. & . & A_{3,3}
\end{bmatrix}
\begin{bmatrix}
I_{k-1} & . & . \\
. & 1 & U \\
. & . & I_{n-k}
\end{bmatrix} \\
&=
\begin{bmatrix}
A_{1,1} & A_{1,2} & A_{1,2} U + A_{1,3} \\
. & 1 & U + A_{2,3} + T A_{3,3} \\
. & . & A_{3,3}
\end{bmatrix}, \\
Pv &=
\begin{bmatrix}
I_{k-1} & . & . \\
. & 1 & T \\
. & . & I_{n-k}
\end{bmatrix}
\begin{bmatrix}
v_{\block{1}} \\ v_{\block{2}} \\ v_{\block{3}}
\end{bmatrix}
=
\begin{bmatrix}
v_{\block{1}} \\ v_{\block{2}} + T v_{\block{3}} \\ v_{\block{3}}
\end{bmatrix}.
\end{align*}
Now we can read of a \emph{sufficient} condition
for $(Q^{-1} s)_k = 0$, namely
the \emph{existence} of $T,U \in \field{K}^{1 \times (n-k)}$ such that
\begin{equation}\label{eqn:pf.lmsys}
U + A_{2,3} + T A_{3,3} = 0
\quad\text{and}\quad
v_{\block{2}} + T v_{\block{3}} = 0.
\end{equation}
Let $d$ be the number of letters in our alphabet $X$.
The blocks $T= [\alpha_{k+1}, \alpha_{k+2}, \ldots, \alpha_n]$ and
$U = [\beta_{k+1}, \beta_{k+2}, \ldots, \beta_n]$ in the transformation $(P,Q)$
are of size $1 \times (n-k)$, thus we have
a \emph{linear} system of equations (over $\field{K}$) with
$2(n-k)$ unknowns (for $k>1$) and $(d+1)(n-k) + 1$ equations:
\begin{align*}
\begin{bmatrix}
\beta_{k+1} & \beta_{k+2} & \beta_{k+3}
\end{bmatrix}
+
\begin{bmatrix}
0 & 0 & 0
\end{bmatrix}
+
\begin{bmatrix}
\alpha_{k+1} & \alpha_{k+2} & \alpha_{k+3}
\end{bmatrix}
\begin{bmatrix}
1 & -x & -z \\
. & 1 & -y \\
. & . & 1
\end{bmatrix}
&=
\begin{bmatrix}
0 & 0 & 0
\end{bmatrix}, \\
\begin{bmatrix}
-1
\end{bmatrix}
+
\begin{bmatrix}
\alpha_{k+1} & \alpha_{k+2} & \alpha_{k+3}
\end{bmatrix}
\begin{bmatrix}
. \\ . \\ 1
\end{bmatrix}
&=
\begin{bmatrix}
0
\end{bmatrix}.
\end{align*}
One solution is $T=[0,0,1]$ and $U=[0,0,-1]$.
We compute $\tilde{\als{A}}_1 = P\als{A}Q$
and remove block row~$\block{2}$
and column~$\block{2}$,
that is, row~$k$ and column~$k$, to get the new ALS
\begin{displaymath}
\als{A}_1 = (u,A,v) = \left(
\begin{bmatrix}
1 & . & . & . & .
\end{bmatrix},
\begin{bmatrix}
1 & -x & -1 & . & . \\
. & 1 & . & . & y \\
. & . & 1 & -x & -z \\
. & . & . & 1 & -y \\
. & . & . & . & 1
\end{bmatrix},
\begin{bmatrix}
. \\ . \\ . \\ . \\ 1
\end{bmatrix}
\right).
\end{displaymath}
Next we do one ``right'' minimization step, that is, we remove (if possible)
one element of the $\field{K}$-linearly dependent right family $t = u A^{-1}$
and construct a new system.
We fix a $1 < k \le n=5$, say $k=3$. Note that $t_1 = 1$.
Now we are looking for a transformation $(P,Q)$ of the form
\begin{equation}\label{eqn:pf.rtrf}
(P,Q) = \left(
\begin{bmatrix}
I_{k-1} & T & . \\
. & 1 & . \\
. & . & I_{n-k}
\end{bmatrix},
\begin{bmatrix}
I_{k-1} & U & . \\
. & 1 & . \\
. & . & I_{n-k}
\end{bmatrix}
\right)
\end{equation}
such that column~$k$ in $PAQ$ is $[0,0,1,0,0]^{\!\top}$ (for an
admissible transformation, that is,
$U$ has a zero first row, the corresponding
entry $u_k$ in $u$ is zero).
A sufficient
condition for $(t P^{-1})_k=0$ is the existence of
$T,U \in \field{K}^{(k-1) \times 1}$ such that
\begin{equation}\label{eqn:pf.rmsys}
A_{1,1} U + A_{1,2} + T = 0.
\end{equation}
(In Remark~\ref{rem:pf.complexity}
there is a less ``compressed'' version of this linear system
of equations.)
One solution is $T=[1,0]^{\!\top}$ and $U=[0,0]^{\!\top}$.
We compute $\tilde{\als{A}}_2 = P \als{A}_1 Q$
and remove
row~$k$ and column~$k$ to get the new (not yet minimal) ALS
\begin{displaymath}
\als{A}_2 = (u,A,v) = \left(
\begin{bmatrix}
1 & . & . & .
\end{bmatrix},
\begin{bmatrix}
1 & -x & -x & -z \\
. & 1 & . & y \\
. & . & 1 & -y \\
. & . & . & 1
\end{bmatrix},
\begin{bmatrix}
. \\ . \\ . \\ 1
\end{bmatrix}
\right).
\end{displaymath}
If a left (respectively right) minimization step with $k=1$
(respectively $k=n$ and $v = [0,\ldots,0,\lambda]^{\!\top}$)
can be done, then the ALS represents zero and we can stop immediately.
The following is the only non-trival observation:
Recall that, if there exist row (respectively column) blocks $T,U$ such that
\eqref{eqn:pf.lmsys} (respectively \eqref{eqn:pf.rmsys})
has a solution then the left (respectively right) family
is $\field{K}$-linearly dependent.
To guarantee \emph{minimality} by Proposition~\ref{pro:pf.cohn94.47}
we need the other implication, that is,
the existence of appropriate row \emph{or} column blocks
for non-minimal pre-standard admissible linear systems.
Although the arguments can be found in the proof of
Proposition~\ref{pro:pf.minmul}, we repeat them here
because this is the crucial part of the minimization
algorithm:
Let $\als{A} = (u,A,v)$ be a pre-standard ALS of
dimension $n \ge 2$ with left family $s = (s_1, s_2, \ldots, s_n)$
and assume that there exists a $1 \le k < n$ such that
the subfamily $(s_{k+1}, s_{k+2}, \ldots, s_n)$ is $\field{K}$-linearly
independent while $(s_k, s_{k+1}, \ldots, s_n)$ is
$\field{K}$-linearly dependent.
Then, by Lemma~\ref{lem:pf.min1}, there exist
matrices $T,U \in \field{K}^{1 \times (n-k)}$
such that \eqref{eqn:pf.lmsys} holds.
In other words: We have to start with $k_s=n-1$ for a left
and $k_t=2$ for a right minimization step.
If we apply one minimization step, we must check
the other family ``again'', illustrated in the following
example, which is \emph{not} constructed out of two
minimal systems:
\begin{displaymath}
\als{A} = (u,A,v) = \left(
\begin{bmatrix}
1 & . & . & . & .
\end{bmatrix},
\begin{bmatrix}
1 & -x & -y & x+y & . \\
. & 1 & . & . & -z \\
. & . & 1 & . & -z \\
. & . & . & 1 & -y \\
. & . & . & . & 1 \\
\end{bmatrix},
\begin{bmatrix}
. \\ . \\ . \\ . \\ 1
\end{bmatrix}
\right).
\end{displaymath}
Clearly, the left subfamily
$(s_3,s_4,s_5)$
and the right subfamily $(t_1, t_2, t_3)$
of $\als{A}$
are $\field{K}$-linearly independent respectively.
If we subtract row~3 from row~2 and add column~2 to column~3,
we get the ALS
\begin{displaymath}
\als{A}' = (u',A',v') = \left(
\begin{bmatrix}
1 & . & . & . & .
\end{bmatrix},
\begin{bmatrix}
1 & -x & -x-y & x+y & . \\
. & 1 & 0 & . & 0 \\
. & . & 1 & . & -z \\
. & . & . & 1 & -y \\
. & . & . & . & 1 \\
\end{bmatrix},
\begin{bmatrix}
. \\ . \\ . \\ . \\ 1
\end{bmatrix}
\right).
\end{displaymath}
The right subfamily $(t_1'',t_2'',t_3'')$
of $\als{A}'' = \als{A}'\rule[-0.5ex]{0pt}{2.2ex}^{[-2]}$
is (here) \emph{not} $\field{K}$-linearly independent anymore,
therefore we must check for a right minimization step
for $k=3$ again.
\begin{definition}[Minimization Equations, Minimization Transformations]
\label{def:pf.meqn}
Let $\als{A} = (u,A,v)$ be a pre-standard ALS of dimension
$n \ge 2$.
Recall the block decomposition \eqref{eqn:pf.bloals}
\begin{displaymath}
\als{A}^{[k]} = \left(
\begin{bmatrix}
u_{\block{1}} & . & .
\end{bmatrix},
\begin{bmatrix}
A_{1,1} & A_{1,2} & A_{1,3} \\
. & 1 & A_{2,3} \\
. & . & A_{3,3}
\end{bmatrix},
\begin{bmatrix}
v_{\block{1}} \\ v_{\block{2}} \\ v_{\block{3}}
\end{bmatrix}
\right).
\end{displaymath}
By $\als{A}^{[-k]}$ we denote the ALS $\als{A}^{[k]}$
without row/column~$k$
(of dimension $n-1$):
\begin{displaymath}
\als{A}^{[-k]} = \left(
\begin{bmatrix}
u_{\block{1}} & .
\end{bmatrix},
\begin{bmatrix}
A_{1,1} & A_{1,3} \\
. & A_{3,3}
\end{bmatrix},
\begin{bmatrix}
v_{\block{1}} \\ v_{\block{3}}
\end{bmatrix}
\right).
\end{displaymath}
For $k = \{ 1,2,\ldots, n-1 \}$ the equations
$U + A_{2,3} + T A_{3,3} = 0$ and $v_{\block{2}} + T v_{\block{3}} = 0$,
see \eqref{eqn:pf.lmsys},
with respect to the block decomposition $\als{A}^{[k]}$
are called \emph{left minimization equations}%
\index{left minimization equations},
denoted by $\mathcal{L}_k = \mathcal{L}_k(\als{A})$.
A solution by the row block pair $(T,U)$ is denoted by
$\mathcal{L}_k(T,U) = 0$,
the corresponding transformation $(P,Q) = \bigl(P(T), Q(U) \bigr)$,
see \eqref{eqn:pf.ltrf},
is called \emph{left minimization transformation}%
\index{left minimization transformation}.
For $k = \{ 2,3,\ldots, n \}$ the equations
$A_{1,1} U + A_{1,2} + T = 0$, see \eqref{eqn:pf.rmsys},
with respect to the block decomposition $\als{A}^{[k]}$
are called \emph{right minimization equations}%
\index{right minimization equations},
denoted by $\mathcal{R}_k = \mathcal{R}_k(\als{A})$.
A solution by the column block pair $(T,U)$ is denoted by
$\mathcal{R}_k(T,U) = 0$,
the corresponding transformation $(P,Q) = \bigl(P(T), Q(U) \bigr)$,
see \eqref{eqn:pf.rtrf},
is called \emph{right minimization transformation}%
\index{right minimization transformation}.
\end{definition}
Now there is only one important detail left, namely
that we cannot apply Lemma~\ref{lem:pf.min1} in the following
(first) left minimization step. For $0 \neq \alpha \in \field{K}$
consider the ALS $\als{A}$
\begin{displaymath}
\begin{bmatrix}
1 & -\alpha \\
. & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ \lambda
\end{bmatrix}
\end{displaymath}
and (admissibly) transform the system matrix $A$ in the following way:
\begin{displaymath}
\underbrace{%
\begin{bmatrix}
. & \alpha \\
1 & .
\end{bmatrix}}_{=:P}
\begin{bmatrix}
1 & -\alpha \\
. & 1
\end{bmatrix}
\underbrace{%
\begin{bmatrix}
1 & . \\
1/\alpha & 1
\end{bmatrix}}_{=:Q}
=
\begin{bmatrix}
. & \alpha \\
1 & -\alpha
\end{bmatrix}
\begin{bmatrix}
1 & . \\
1/\alpha & 1
\end{bmatrix} \\
=
\begin{bmatrix}
1 & \alpha \\
0 & -\alpha
\end{bmatrix}
\end{displaymath}
thus
\begin{equation}\label{eqn:pf.left1}
\begin{bmatrix}
1 & \alpha \\
. & -\alpha
\end{bmatrix}
s =
\begin{bmatrix}
\alpha \lambda \\
0
\end{bmatrix}
\end{equation}
and we can remove the \emph{last} row and column.
Note that we do not have to consider such a special case
for the right family.
\begin{algorithm}[Minimizing a pre-standard ALS]\label{alg:pf.minals}
\ \\
Input: $\als{A} = (u,A,v)$ pre-standard ALS
of dimension $n \ge 2$ (for some polynomial $p$).\\
Output: $\als{A}' = (,,)$ if $p=0$ or
a minimal pre-standard ALS $\als{A}' = (u',A',v')$ if $p \neq 0$.
\begin{algtest}
\hbox{}\\[-3ex]
\lnum{1:}\>$k := 2$ \\
\lnum{2:}\>while $k \le \dim \als{A}$ do \\
\lnum{3:}\>\>$n := \dim(\als{A})$ \\
\lnum{4:}\>\>$k' := n +1 - k$ \\
\lnum{ }\>\>\textnormal{Is the left subfamily
\raisebox{0pt}[0pt][0pt]{%
$(s_{k'},\overbrace{\rule[-0.5ex]{0pt}{2.2ex} s_{k'+1},\ldots, s_{n}}^{\text{lin.~indep.}})$}
$\field{K}$-linearly dependent?} \\
\lnum{5:}\>\>if $\exists\, T,U \in \field{K}^{1 \times (k-1)}
\textnormal{ admissible}: \mathcal{L}_{k'}(\als{A})=\mathcal{L}_{k'}(T,U)=0$ then \\
\lnum{6:}\>\>\>if $k' = 1$ then \\
\lnum{7:}\>\>\>\>return $(,,)$ \\
\lnum{ }\>\>\>endif \\
\lnum{8:}\>\>\>\raisebox{0pt}[0pt][0pt]{%
$\als{A} := \bigl(P(T) \als{A} Q(U)\bigr)\rule[-0.5ex]{0pt}{2.2ex}^{[-k']}$} \\
\lnum{9:}\>\>\>if $k > \max \bigl\{ 2, \frac{n+1}{2} \bigr\}$ then \\
\lnum{10:}\>\>\>\>$k := k-1$ \\
\lnum{ }\>\>\>endif \\
\lnum{11:}\>\>\>continue \\
\lnum{ }\>\>endif \\
\lnum{12:}\>\>if $k = 2$ and $s_{n-1} = \alpha s_n$ \textnormal{(for some $\alpha\in \field{K}$)} then \\
\lnum{13:}\>\>\>find \textnormal{admissible} $(P,Q)$ such that $(Q^{-1} s)_n=0$
and $PAQ$ is \textnormal{upper triangular} \\
\lnum{14:}\>\>\>$\als{A} := (P\als{A} Q)^{[-n]}$ \\
\lnum{15:}\>\>\>continue \\
\lnum{ }\>\>endif \\
\lnum{ }\>\>\textnormal{Is the right subfamily
\raisebox{0pt}[0pt][0pt]{%
$(\overbrace{\rule[-0.5ex]{0pt}{2.2ex} t_1, \ldots,t_{k-1}}^{\text{lin.~indep.}}, t_k)$}
$\field{K}$-linearly dependent?} \\
\lnum{16:}\>\>if $\exists\, T,U \in \field{K}^{(k-1) \times 1}
\textnormal{ admissible} :
\mathcal{R}_k(\als{A}) = \mathcal{R}_k(T,U)=0$ then \\
\lnum{17:}\>\>\>$\als{A} := \bigl(P(T) \als{A} Q(U) \bigr)\rule[-0.5ex]{0pt}{2.2ex}^{[-k]}$ \\
\lnum{18:}\>\>\>if $k > \max \bigl\{ 2, \frac{n+1}{2} \bigr\}$ then \\
\lnum{19:}\>\>\>\>$k := k-1$ \\
\lnum{ }\>\>\>endif \\
\lnum{20:}\>\>\>continue \\
\lnum{ }\>\>endif \\
\lnum{21:}\>\>$k := k+1$ \\
\lnum{ }\>done \\
\lnum{22:}\>return $P\als{A},$
\textnormal{with $P$, such that $Pv = [0,\ldots,0,\lambda]^{\!\top}$}
\end{algtest}
\end{algorithm}
\begin{proof}
The admissible linear system $\als{A}$
represents $p=0$ if and only if $s_1 = (A^{-1} v)_1 = 0$.
Since all systems are equivalent to $\als{A}$,
this case is recognized for $k'=1$ because
by Lemma~\ref{lem:pf.min1} there is an \emph{admissible}
transformation such that the first left minimization equation
is fulfilled.
Now assume $p \neq 0$.
We have to show that both, the left family $s'$ and
the right family $t'$ of $\als{A}' = (u',A',v')$ are
$\field{K}$-linearly independent respectively.
Let $n' = \dim(\als{A}')$ and for $k \in \{ 1, 2, \ldots, n' \}$
denote by
\raisebox{0pt}[0pt][0pt]{%
$s'_{(k)} = (s'_{n'+1-k}, s'_{n'+2-k}, \ldots, s'_{n'})$}
the left and by
\raisebox{0pt}[0pt][0pt]{%
$t'_{(k)} = (t'_1, t'_2, \ldots, t'_k)$}
the right subfamily.
By assumption $\als{A}$ is a pre-standard ALS and
therefore $s'_{n'} \neq 0$ and $t'_1 \neq 0$,
that is, $s'_{(1)}$ is $\field{K}$-linearly independent
and $t'_{(1)}$ is $\field{K}$-linearly independent.
The loop starts with $k=2$.
Only if both, $s'_{(k)}$ and $t'_{(k)}$
are $\field{K}$-linearly indpendent respectively,
$k$ is incremented. For $k=2$ we can solve the special
case by~\eqref{eqn:pf.left1} for the left family.
Otherwise a left (Lemma~\ref{lem:pf.min1}) or a right
(variant of Lemma~\ref{lem:pf.min1}) minimization step
was successful and the dimension of the current ALS is
strictly smaller than that of the previous.
Hence, since $k$ is bounded from below,
the algorithm stops in a finite number of steps.
We just have to make sure that there exists an
admissible transformation, if there exists a
column couple $(T,U)$ such that the right minimization
equations $\mathcal{R}_k(T,U)=0$ are fulfilled.
However, if the first column would be involved to
eliminate the first entry in column~$k$, the
$k$-th row could be used instead.
Clearly, $\als{A}'$ is in pre-standard form.
\end{proof}
\iffalse
Since in \emph{each} step of the loop
either $k$ is incremented or the dimension
of the (current) ALS $\als{A}$ is decremented,
the algorithm stops in a \emph{finite} number of steps.
We show $\field{K}$-linear independence of the
\emph{complete} left (respectively right) family
of $\als{A}'=(u,A,v)$ by induction.
Let $n = \dim \als{A}'$ and
let $s = A^{-1} v$ (respectively $t = u A^{-1}$)
be the left (respectively right) family of $\als{A}'$.
For $k=1,2,\ldots, n$ we define the sequence of
the left subfamilies $F^s_k = (s_{n-k+1}, s_{n-k+2}, \ldots, s_n)$
and the right subfamilies $F^t_k = (t_1, t_2, \ldots, t_k)$.
Since the input ALS $\als{A}$ is pre-standard,
$s_n$ and $t_1$ are non-zero.
Now, assuming $\field{K}$-linear independence of
$F^s_k$ (for $k = 2$ see \eqref{eqn:pf.left1}) and $F^t_k$,
by Lemma~\ref{lem:pf.min1} (and a variant for
the right family) we get an appropriate transformation
and remove row and column $k+1$ until \emph{both}
$F^s_{k+1}$ and $F^t_{k+1}$ are
$\field{K}$-linearly independent respectively.
Thus, by Proposition~\ref{pro:pf.cohn94.47},
the ALS is \emph{minimal}.
All the transformations are admissible
and ``almost'' pre-standard, that is,
other entries than the last in the right hand side $v$
might be non-zero.
However, for $k<n$, a solution to the left minimization equations
guaranties $v_k = 0$.
Finally it is brought to pre-standard form.
\fi
\begin{Remark}\label{rem:pf.complexity}
This algorithm can be implemented very efficiently
provided that row and column transformations are
done directly. Let $d$ be the number of letters in our alphabet $X$.
For $\ell=0,1,\ldots,d$ let $A_{ij}^{(\ell)}$ denote the
submatrix corresponding to letter $x_\ell$ and the (current)
block decomposition of $\als{A}^{[k]}$.
The right minimization equations
$A_{1,1} U + A_{1,2} + T = 0$ can be written as
\begin{displaymath}
\begin{bmatrix}
I & A_{1,1}^{(0)} \\
I & A_{1,1}^{(1)} \\
\vdots & \vdots \\
I & A_{1,1}^{(d)}
\end{bmatrix}
\begin{bmatrix}
T \\ U
\end{bmatrix}
=
\begin{bmatrix}
- A_{1,2}^{(0)} \\
- A_{1,2}^{(1)} \\
\vdots \\
- A_{1,2}^{(d)}
\end{bmatrix}
\end{displaymath}
with $2(k-1)$ unknowns, $k<n$.
By Gaussian elimination one
gets complexity $\mathcal{O}(dn^3)$
for solving such a system,
see \cite[Section~2.3]{Demmel1997a
.
To build such a system and working on a
linear matrix pencil
$\bigl[\begin{smallmatrix} 0 & u \\ v & A \end{smallmatrix}\bigr]$
with $d+1$ square
coefficient matrices of size $n+1$
(transformations, etc.)
has complexity $\mathcal{O}(dn^2)$.
Since there are at most $2(n-1)$ steps,
we get overall (minimization) complexity
$\mathcal{O}(d n^4)$.
The algorithm of \cite{Cardon1980a
\ has complexity $\mathcal{O}(d n^3)$.
One has to be careful with a direct comparison.
The latter works more general for regular elements,
that is, rational formal power series.
However the idea used here generalizes directly for
larger blocks, say for a block decomposition $\als{A}^{[k,k+1,\ldots,k+m]}$
for $k,m < n$ and can be used partially for \emph{non-regular} elements
in the free field, for example to solve the \emph{word problem}
which would have complexity $\mathcal{O}(d n^6)$.
Further details can be found in
\cite{Schrempf2017a
.
\end{Remark}
\subsection{Factorizing non-commutative Polynomials}\label{sec:pf.fnp.fact}
\begin{definition}[Atomic Admissible Linear Systems]\label{def:pf.atom}
A \emph{minimal} pre-standard ALS $\als{A} = (1,A,\lambda)$
of dimension $n\ge 2$ is called \emph{atomic}\index{atomic ALS}
(\emph{irreducible})\index{irreducible ALS},
if there is no pre-standard admissible transformation $(P,Q)$ such that
$PAQ$ has an upper right block of zeros of size $(n-i-1) \times i$ for some
$i=1,2,\ldots,n-2$.
\end{definition}
If we take the minimal ALS \eqref{eqn:pf.minmul.1}
for $pq = x(1-yx)$
from Example~\ref{ex:pf.minmul} and
add column~2 to column~4, we obtain
\begin{displaymath}
\begin{bmatrix}
1 & -x & . & -x \\
. & 1 & y & 0 \\
. & . & 1 & -x \\
. & . & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ . \\ 1
\end{bmatrix},
\quad
s =
\begin{bmatrix}
x(1-yx) \\
-yx \\
x \\
1
\end{bmatrix}.
\end{displaymath}
Subtracting row~3 from row~1 yields
\begin{displaymath}
\begin{bmatrix}
1 & -x & -1 & 0 \\
. & 1 & y & 0 \\
. & . & 1 & -x \\
. & . & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ . \\ 1
\end{bmatrix},
\quad
s =
\begin{bmatrix}
x(1-yx) \\
-yx \\
x \\
1
\end{bmatrix}
\end{displaymath}
with an upper right $2 \times 1$ block of zeros in the system matrix.
We would obtain the same system by \emph{minimal polynomial multiplication}
of $1-xy$ and $x$. This illustrates that we can find the factors of
non-commutative polynomials by searching for
pre-standard admissible transformations
that give a corresponding upper right block of zeros in the transformed
system matrix.
Theorem~\ref{thr:pf.factorization}
will establish the correspondence between a factorization into
atoms and the structure of the upper right blocks of zeros.
Thus, to factorize a polynomial $p$ of rank~$n\ge 3$ into \emph{non-trivial} factors
$p = q_1 q_2$ with $\rank(q_i)= n_i \ge 2$ and $n = n_1 + n_2 - 1$,
we have to look for (pre-standard admissible) transformations of the form
\begin{equation}\label{eqn:pf.factrn}
(P,Q) = \left(
\begin{bmatrix}
1 & \alpha_{1,2} & \ldots & \alpha_{1,n-1} & 0 \\
& \ddots & \ddots & \vdots & \vdots \\
& & 1 & \alpha_{n-2,n-1} & 0 \\
& & & 1 & 0 \\
& & & & 1
\end{bmatrix},
\begin{bmatrix}
1 & 0 & 0 & \ldots & 0 \\
& 1 & \beta_{2,3} & \ldots & \beta_{2,n} \\
& & 1 & \ddots & \vdots \\
& & & \ddots & \beta_{n-1,n} \\
& & & & 1 \\
\end{bmatrix}
\right)
\end{equation}
with entries $\alpha_{ij}, \beta_{ij} \in \field{K}$.
In general this is a \emph{non-linear} problem
with $(n-2)(n-1)$ unknowns.
\begin{definition}[Standard Admissible Linear Systems]\label{def:pf.stdals}
Every
\begin{itemize}
\item ALS $(1, I_1, \lambda)$ for a scalar $0 \neq \lambda \in \field{K}$,
\item atomic ALS and
\item non-atomic \emph{minimal} pre-standard ALS
$\als{A} = (1,A,\lambda)$ for $p \in \freeALG{\field{K}}{X}$
of dimension $n \ge 3$ obtained
from minimal multiplication (Proposition~\ref{pro:pf.minmul})
of atomic admissible linear systems for
its atomic factors $p = q_1 q_2 \cdots q_m$ (with atoms $q_i$)
\end{itemize}
is called \emph{standard}\index{standard ALS}.
\end{definition}
\begin{remark}
We have the following chain of ``inclusions'' of admissible linear systems:
Atomic ALS $\subsetneq$ Standard ALS $\subsetneq$ Pre-Standard ALS
$\subsetneq$ ALS.
\end{remark}
\begin{Example}\label{ex:pf.irred}
Let $p = x^2 -2\in \freeALG{\field{K}}{X}$ be given by the minimal ALS
\begin{displaymath}
\als{A} = \left(
\begin{bmatrix}
1 & . & .
\end{bmatrix},
\begin{bmatrix}
1 & -x & 2 \\
. & 1 & -x \\
. & . & 1
\end{bmatrix},
\begin{bmatrix}
. \\ . \\ 1
\end{bmatrix}
\right).
\end{displaymath}
If $\field{K} = \mathbb{Q}$ then $\als{A}$ is atomic (irreducible).
If $\field{K} = \mathbb{R}$ then there is the pre-standard admissible transformation
\begin{displaymath}
(P,Q) = \left(
\begin{bmatrix}
1 & \sqrt{2} & . \\
. & 1 & . \\
. & . & 1
\end{bmatrix},
\begin{bmatrix}
1 & . & . \\
. & 1 & -\sqrt{2} \\
. & . & 1
\end{bmatrix}
\right)
\end{displaymath}
such that $\als{A}' =P\als{A} Q$ is
\begin{displaymath}
\als{A}'= \left(
\begin{bmatrix}
1 & . & .
\end{bmatrix},
\begin{bmatrix}
1 & -x+\sqrt{2} & 0 \\
. & 1 & -x-\sqrt{2} \\
. & . & 1
\end{bmatrix},
\begin{bmatrix}
. \\ . \\ 1
\end{bmatrix}
\right),
\end{displaymath}
thus $p = x^2 -2 = (x-\sqrt{2})(x+\sqrt{2})$ in $\freeALG{\mathbb{R}}{X}$.
\end{Example}
\begin{remark}
Note that it is easy to check that
$p = xy -2 \in \freeALG{\field{K}}{X}$
is atomic, because both entries, $\alpha_{1,2}$ in $P$
and $\beta_{2,3}$ in $Q$ have to be zero
(otherwise the upper right entry in $A'$ could not
become zero) and hence there is \emph{no} non-trivial pre-standard
admissible transformation, that is, transformation that changes
the upper right block structure.
\end{remark}
\begin{lemma}\label{lem:pf.factorization}
Let $0\ne p,q_1,q_2 \in \freeALG{\field{K}}{X}$ be given by the \emph{minimal}
pre-standard admissible linear systems $\als{A} = (1,A,\lambda)$,
$\als{A}_1 = (1,A_1,\lambda_1)$ and $\als{A}_2 = (1,A_2,\lambda_2)$
of dimension $n=\rank p$ and $n_1,n_2 \ge 2$ respectively
with $p = q_1 q_2$.
Then there exists a pre-standard admissible transformation $(P,Q)$ such
that $PAQ$ has an upper right block of zeros of size
$(n_1-1) \times (n_2-1) = (n-n_2) \times (n-n_1)$.
\end{lemma}
\begin{proof}
Let $\als{A}'=(1,A',\lambda_2)$ be the minimal ALS from
Proposition~\ref{pro:pf.minmul}
for $q_1 q_2$. Clearly, we have $\dim \als{A}' = n_1 + n_2 - 1 = n = \rank p$.
And $A'$ has, by construction, an upper right block of zeros of size
$(n_1-1) \times (n_2-1)$. Both, $\als{A}$ and $\als{A}'$ represent the
same element $p$, thus
---by \cite[Theorem~1.4]{Cohn1999a
--- there exists an admissible transformation $(P,Q)$
such that $P\als{A}Q = \als{A}'$. Since $\als{A}'$ is
pre-standard, $(P,Q)$ is a pre-standard admissible transformation.
\end{proof}
\begin{theorem}[Polynomial Factorization]\label{thr:pf.factorization}
Let $p \in \freeALG{\field{K}}{X}$ be given by the
\emph{minimal} pre-standard admissible linear system $\als{A} = (1,A,\lambda)$
of dimension $n = \rank p \ge 3$.
Then $p$ has a factorization into $p = q_1 q_2$ with $\rank(q_i) = n_i \ge 2$
if and only if there exists a pre-standard admissible transformation $(P,Q)$
such that $PAQ$ has an upper right block of zeros of size $(n_1 - 1) \times (n_2 - 1)$.
\end{theorem}
\begin{proof}
If there is such a factorization, then Lemma~\ref{lem:pf.factorization}
applies.
Conversely, if we have such a (pre-standard admissible) transformation
for a zero block of size $k_1 \times k_2$
we obtain an ALS in \emph{block} form ($p$ is the first entry of $s_{\block{1}}$)
\begin{displaymath}
\begin{bmatrix}
A_{1,1} & A_{1,2} & . \\
. & 1 & A_{2,3} \\
. & . & A_{3,3}
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ v_{\block{3}}
\end{bmatrix},
\quad
s =
\begin{bmatrix}
s_{\block{1}} \\ g \\ s_{\block{3}}
\end{bmatrix},
\end{displaymath}
with
$A_{1,1}$ of size $k_1 \times k_1$
and $A_{3,3}$ of size $k_2 \times k_2$,
in which we duplicate the entry $s_{k_1+1}$ by inserting
a ``dummy'' row (and column) to get the following ALS of
size $k_1 +k_2 +2 = n + 1$:
\begin{displaymath}
\begin{bmatrix}
A_{1,1} & A_{1,2} & 0 & . \\
0 & 1 & -1 & 0 \\
. & . & 1 & A_{2,3} \\
. & . & 0 & A_{3,3}
\end{bmatrix}
s' =
\begin{bmatrix}
. \\ . \\ . \\ v_{\block{3}}
\end{bmatrix},
\quad
s' =
\begin{bmatrix}
s_{\block{1}} \\ g \\ g \\ s_{\block{3}}
\end{bmatrix}.
\end{displaymath}
According to the construction of the multiplication
in Proposition~\ref{pro:pf.ratop} we have $p = f g$
in the first component of $s_{\block{1}}$,
the first block in $s'$, for $f,g \in \freeALG{\field{K}}{X}$
given by the (pre-standard) admissible linear systems
\begin{displaymath}
\begin{bmatrix}
A_{1,1} & A_{1,2} \\
. & 1
\end{bmatrix}
s_f =
\begin{bmatrix}
. \\ 1
\end{bmatrix}
\quad\text{and}\quad
\begin{bmatrix}
1 & A_{2,3} \\
. & A_{3,3}
\end{bmatrix}
s_g =
\begin{bmatrix}
. \\ v_{\block{3}}
\end{bmatrix}
\end{displaymath}
of dimension $n_1 = k_1 + 1$ and $n_2 = k_2+1$ respectively.
\end{proof}
This finishes the algorithm. A simple analogue of
\cite[Theorem~4.1]{Cohn1999a
\ will do the rest, we do not have to worry about
invertibility of the transformation matrices $P$ and $Q$
in~\eqref{eqn:pf.factrn}.
In our case
$\field{K}[\alpha,\beta] = \field{K}[
\alpha_{1,2},\ldots,\alpha_{1,n-1},\alpha_{2,3},\ldots,\alpha_{2,n-1},\ldots,\alpha_{n-2,n-1},$
$\beta_{2,3},\ldots,\beta_{2,n},$
$\beta_{3,4},\ldots,\beta_{3,n},\ldots,\beta_{n-1,n} ]$.
However, a non-trivial ideal does \emph{not} guarantee
a solution over $\field{K}$ although
there are solutions over $\aclo{\field{K}}$.
To test only (without computing it)
\emph{if} there is a solution (over $\field{K}$)
one can use the concept of \emph{resultants}.
An introduction can be found in
\cite[Section~3.6]{Cox2015a
. This book contains also an introduction
to GrΓΆbner bases and an overview about
computer algebra software for computing them.
Additional to the work of
\cite{Buchberger1970a
, the survey on GrΓΆbner--Shirshov bases
of \cite{Bokut2000a
\ could be consulted.
\begin{remark}
Note, that in general in order to reverse the multiplication
(Proposition~\ref{pro:pf.minmul})
to find factors, we also need a lower left block of zeros of appropriate size.
This important fact is hidden in the pre-standard form of an ALS.
A general factorization concept is considered in future work.
\end{remark}
\begin{proposition}\label{pro:pf.ideal}
Let $p\in \freeALG{\field{K}}{X}$ be given by the
minimal pre-standard admissible linear system $\als{A} = (1,A,\lambda)$
of dimension $n = \rank p \ge 3$ and let $(P,Q)$ as in~\eqref{eqn:pf.factrn}.
Fix a $k \in \{ 1, 2, \ldots, n-2 \}$
and denote by $I_k$ the ideal
of $\field{K}[\alpha,\beta]$
generated by the coefficients of each $x\in \{ 1 \} \cup X$ in the
$(i,j)$ entries of the matrix $PAQ$ for $1 \le i \le k$ and
$k+2 \le j \le n$. Then $p$ factorizes
\emph{over} $\freeALG{\aclo{\field{K}}}{X}$
into $p=q_1 q_2$ with $\rank q_1 = k+1$ and $\rank q_2 = n-k$
if and only if the ideal $I_k$ is non-trivial.
\end{proposition}
Given a polynomial $p\in \freeALG{\field{K}}{X}$
by a minimal pre-standard ALS of dimension $n=\rank p \ge 2$
there are at most $\phi(n) = 2^{n-2}$ (minimal) standard admissible linear
systems (with respect to the structure of the upper right blocks
of zeros). For $n=2$ this is clear. For $n > 2$ the ALS could
be atomic or have a block of zeros of size $1 \times (n-2)$
or $(n-2) \times 1$, thus $\phi(n+1) = 1 + 2 \phi(n)- 1 = 2 \phi(n)$
because the system with ``finest'' upper right structure is
counted twice.
\begin{remark}
Recall that ---up to similarity--- each element $p\in \freeALG{\field{K}}{X}$
has only \emph{one} factorization into atoms.
If one is interested in the number of factorizations
(not necessarily irreducible)
up to permutation (and multiplication of the factors by units)
the above estimate $2^{\rank p-2}$ can be used.
However, the number of factorizations in the sense
of \cite[Definition~3.1]{Bell2017a
\ can be bigger.
As an example take the polynomial $p = (x-1)(x-2)(x-3)$
which has $3! = 6$ different factorizations while
the number of standard admissible
linear systems is bounded by
$2^{\rank p-2} = 4$.
\end{remark}
\medskip
Let $p$ be a non-zero polynomial with
the factorization $p=q_1 q_2 \cdots q_m$
into atoms $q_i \in \freeALG{\field{K}}{X}$.
Since $\freeALG{\field{K}}{X}$ is a similarity-UFD
(Proposition~\ref{pro:pf.cohn63b}), \emph{each}
factorization of (a non-zero non-unit) $p$ into atoms has $m$ factors.
Therefore one can define the
\emph{length}\index{length of a polynomial}
of $p$ by $m$, written as $\ell(p) = m$.
For a word $w \in X^* \subseteq \freeALG{\field{K}}{X}$
the length is $\ell(w) = \length{w}$.
By looking at the minimal polynomial multiplication
(Proposition~\ref{pro:pf.minmul})
it is easy to see that the length of an element
$p\in \freeALG{\field{K}}{X}^\bullet$
can be estimated by the rank, namely $\ell(p) \le \rank(p) - 1$.
More on length functions ---and transfer homomorphisms in the
context of non-unique factorizations--- (in the non-commutative setting)
can be found in
\cite[Section~3]{Smertnig2015a
\ or
\cite{Baeth2015a
.
\newpage
\section{Generalizing the Companion Matrix}\label{sec:pf.ncm}
For a special case, namely an alphabet with just one letter,
the companion matrix of a polynomial $p(x)$ yields immediately a
\emph{minimal} linear representation of $p \in \freeALG{\field{K}}{\{x\}}$.
If this is the characteristic polynomial of some (square) matrix
$B \in \field{K}^{m \times m}$,
then its eigenvalues can be computed by the techniques from
Section~\ref{sec:pf.fnp} (if necessary going over to $\aclo{\field{K}}$),
illustrated in Example~\ref{ex:pf.eig}.
For a broader class of (nc) polynomials,
\emph{left} and \emph{right} companion systems can be defined.
In general \emph{minimal} pre-standard admissible linear systems
are necessary to generalize companion matrices,
see Definition~\ref{def:pf.cm}.
\begin{definition}[Companion Matrix, Characteristic Polynomial, Normal Form
\protect{\cite[Section~6.6]{Gantmacher1986a
}]
Let $p(x) = a_0 + a_1 x + \ldots + a_{m-1} x^{m-1} + x^m \in \field{K}[x]$.
The \emph{companion matrix} $L(p)$ is defined as
\begin{displaymath}
L(p) =
\begin{bmatrix}
0 & 0 & \ldots & 0 & -a_0 \\
1 & 0 & \ddots & \vdots & -a_1 \\
& \ddots & \ddots & 0 & \vdots \\
& & 1 & 0 & -a_{m-2} \\
& & & 1 & -a_{m-1}
\end{bmatrix}.
\end{displaymath}
Then $p(x)$ is the \emph{characteristic polynomial}\index{characteristic polynomial}
of $L = L(p)$:
\begin{displaymath}
\det(xI - L) = \det
\begin{bmatrix}
x & 0 & \ldots & 0 & a_0 \\
-1 & x & \ddots & \vdots & a_1 \\
& \ddots & \ddots & 0 & \vdots \\
& & -1 & x & a_{m-2} \\
& & & -1 & x+a_{m-1}
\end{bmatrix}.
\end{displaymath}
Given a square matrix $M \in \field{K}^{m \times m}$,
the \emph{normal form}%
\index{normal form (of a matrix)}
for $M$ can be defined in terms of the
\emph{companion matrix} $L(M)$ of its
\emph{characteristic polynomial} $p(M) = \det (xI - M)$.
\end{definition}
\begin{remark}
In \cite[Section~8.1]{Cohn1995a
, $C(p) = xI - L(p)^{\!\top}$ is also called \emph{companion matrix}.
This is justified by viewing $C(p)$ as \emph{linear matrix pencil}
$C(p) = C_0 \otimes 1 + C_x \otimes x$.
It generalizes nicely for non-commutative polynomials.
\end{remark}
Now we leave $\field{K}[x] = \freeALG{\field{K}}{\{x\}}$
and consider (nc) polynomials $p \in \freeALG{\field{K}}{X}$.
There are two cases where a \emph{minimal} ALS can be
stated immediately, namely if the support (of the polynomial)
can be ``built'' from the left (in the left family) or
from the right (in the right family)
with \emph{strictly} increasing rank.
For example,
a \emph{minimal} pre-standard ALS for
$p = a_0 + a_1 (x+2y) + a_2 (x-z)(x+2y) + y(x-z)(x+2y)$
is given by
\begin{displaymath}
\begin{bmatrix}
1 & -y -a_2 & -a_1 & -a_0 \\
. & 1 & -(x-z) & . \\
. & . & 1 & -(x+2y) \\
. & . & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ . \\ 1
\end{bmatrix},
\quad s =
\begin{bmatrix}
p \\
(x-z)(x+2y) \\
x+2y \\
1
\end{bmatrix}.
\end{displaymath}
\begin{definition}[Left and Right Companion System]\label{def:pf.cs}
For $i=1,2,\ldots, m$ let $q_i \in \freeALG{\field{K}}{X}$
with $\rank{q_i} = 2$ and $a_i \in \field{K}$.
For a polynomial $p \in \freeALG{\field{K}}{X}$
of the form
\begin{displaymath}
p = q_m q_{m-1} \cdots q_1 + a_{m-1} q_{m-1} \cdots q_1 + \ldots + a_2 q_2 q_1
+ a_1 q_1 + a_0
\end{displaymath}
the pre-standard ALS
\begin{equation}\label{eqn:pf.lcs}
\begin{bmatrix}
1 & -q_m -a_{m-1} & -a_{m-2} & \ldots & -a_1 & -a_0 \\
& 1 & -q_{m-1} & 0 & \ldots & 0 \\
& & \ddots & \ddots & \ddots & \vdots \\
& & & 1 & -q_2 & 0 \\
& & & & 1 & -q_1 \\
& & & & & 1
\end{bmatrix}
s =
\begin{bmatrix}
0 \\ 0 \\ \vdots \\ 0 \\ 0 \\ 1
\end{bmatrix}
\end{equation}
is called \emph{left companion system}\index{left companion system}.
And for a polynomial $p \in \freeALG{\field{K}}{X}$ of the form
\begin{displaymath}
p = a_0 + a_1 q_1 + a_2 q_1 q_2 + \ldots + a_{m-1} q_1 q_2 \cdots q_{m-1}
+ q_1 q_2 \cdots q_m
\end{displaymath}
the pre-standard ALS
\begin{equation}\label{eqn:pf.rcs}
\begin{bmatrix}
1 & -q_1 & 0 & \ldots & 0 & -a_0 \\
& 1 & -q_2 & \ddots & \vdots & -a_1 \\
& & \ddots & \ddots & 0 & \vdots \\
& & & 1 & -q_{m-1} & -a_{m-2} \\
& & & & 1 & -q_m -a_{m-1} \\
& & & & & 1
\end{bmatrix}
s =
\begin{bmatrix}
0 \\ 0 \\ \vdots \\ 0 \\ 0 \\ 1
\end{bmatrix}
\end{equation}
is called \emph{right companion system}\index{right companion system}.
\end{definition}
\begin{proposition}\label{pro:pf.minsys}
For $i=1,2,\ldots, m$ let $q_i \in \freeALG{\field{K}}{X}$
with $\rank{q_i} = 2$. Then the polynomials
$p_{\text{l}} = q_m q_{m-1} \cdots q_1 + a_{m-1} q_{m-1} \cdots q_1
+ \ldots + a_2 q_2 q_1 + a_1 q_1 + a_0$
and
$p_{\text{r}} = a_0 + a_1 q_1 + a_2 q_1 q_2
+ \ldots + a_{m-1} q_1 q_2 \cdots q_{m-1} + q_1 q_2 \cdots q_m$
have rank $m+1$.
\end{proposition}
\begin{proof}
Both, the left family $( p, q_{m-1} \cdots q_1, \ldots, q_2 q_1 , q_1, 1 )$
and the right family $( 1, q_m + a_{m-1},$ $(q_m+a_{m-1})q_{m-1} + a_{m-2}, \ldots, p)$
for \eqref{eqn:pf.lcs} are $\field{K}$-linearly independent.
Thus, the left companion system (for $p_{\text{l}}$) is minimal
of dimension $m+1$. Hence $\rank p_{\text{l}} = m+1$.
By a similar argument for the right companion system we
get $\rank p_{\text{r}} = m+1$.
\end{proof}
For a general polynomial $p\in \freeALG{\field{K}}{X}$ with $\rank p = n \ge 2$
we can take any \emph{minimal} pre-standard ALS $\als{A} = (1,A,\lambda)$
to obtain an ALS of the form $(1,A',1)$ by
dividing the last row by $\lambda$ and multiplying the last column by $\lambda$.
Now we can define a (generalized version of the) companion matrix.
Those companion matrices can be used as building blocks to get
companion matrices for products of polynomials.
This is just a different point of view on the ``minimal'' polynomial
multiplication from Proposition~\ref{pro:pf.minmul}.
\begin{remark}
Although nothing can be said in general about minimality
of a linear representation for \emph{commutative} polynomials
(in several variables), Proposition~\ref{pro:pf.minsys}
can be used for constructing minimal linear representations
in the commutative case. Because in this case, the rank is
the maximum of the ranks of the monomials.
For example $p = x^2 y + xyz = xyx + xyz = xy(x+z)$.
\end{remark}
\begin{definition}[Companion Matrices]\label{def:pf.cm}
Let $p \in \freeALG{\field{K}}{X}$ with $\rank p = n \ge 2$ be given
by a \emph{minimal} pre-standard admissible linear system
$\als{A} = (1,A,1)$ and denote by $C(p)$ the upper right
submatrix of size $(n-1) \times (n-1)$.
Then $C(p)$ is called a (nc) \emph{companion matrix}\index{companion matrix}
of $p$.
\end{definition}
\begin{Example}\label{ex:pf.eig}
Let
\begin{displaymath}
B =
\begin{bmatrix}
6 & 1 & 3 \\
-7 & 3 & 14 \\
1 & 0 & 1
\end{bmatrix}.
\end{displaymath}
Then the characteristic polynomial of $B$ is
$p(x) = \det(xI - B) = x^3 - 10 x^2 + 31 x - 30$.
The left companion system of $p$ is
\begin{displaymath}
\begin{bmatrix}
1 & -x+10 & -31 & 30 \\
. & 1 & -x & . \\
. & . & 1 & -x \\
. & . & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ . \\ 1
\end{bmatrix},
\quad
s =
\begin{bmatrix}
p(x) \\
x^2 \\
x \\
1
\end{bmatrix}.
\end{displaymath}
Applying the transformation $(P,Q)$ with
\begin{displaymath}
P =
\begin{bmatrix}
1 & . & . & . \\
& 1 & -3 & . \\
& & 1 & . \\
& & & 1
\end{bmatrix}
\begin{bmatrix}
1 & -5 & 6 & . \\
& 1 & . & . \\
& & 1 & . \\
& & & 1
\end{bmatrix}
=
\begin{bmatrix}
1 & -5 & 6 & . \\
& 1 & -3 & . \\
& & 1 & . \\
& & & 1
\end{bmatrix}
\end{displaymath}
and
\begin{displaymath}
Q =
\begin{bmatrix}
1 & . & . & . \\
& 1 & 5 & . \\
& & 1 & \frac{6}{5} \\
& & & 1
\end{bmatrix}
\begin{bmatrix}
1 & . & . & . \\
& 1 & . & . \\
& & 1 & \frac{9}{5} \\
& & & 1
\end{bmatrix}
=
\begin{bmatrix}
1 & . & . & . \\
& 1 & 5 & 9 \\
& & 1 & 3 \\
& & & 1
\end{bmatrix}
\end{displaymath}
we get the ALS
\begin{displaymath}
\begin{bmatrix}
1 & 5-x & . & . \\
. & 1 & 2-x & . \\
. & . & 1 & 3-x \\
. & . & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ . \\ 1
\end{bmatrix},
\quad
s =
\begin{bmatrix}
p(x) \\
(x-2)(x-3) \\
x-3 \\
1
\end{bmatrix}.
\end{displaymath}
Thus the eigenvalues of $B$ are $2$, $3$ and $5$.
Compare with Example~\ref{ex:pf.irred} for the case when
the polynomial does not decompose in linear factors.
\end{Example}
\newpage
\section{An Example (step by step)}\label{sec:pf.ex}
Let $p = x(1-yx)(3-yx)$ and $q = (xy-1)(xy-3)x$ be given.
Taking companion systems (Definition~\ref{def:pf.cs}) for
their factors respectively,
by Proposition~\ref{pro:pf.minmul}
we get the \emph{minimal} ALS (for $p$):
\begin{equation}\label{eqn:pf.ex.1}
\begin{bmatrix}
1 & - x & . & . & . & . \\
. & 1 & -y & -1 & . & . \\
. & . & 1 & x & . & . \\
. & . & . & 1 & -y & -3 \\
. & . & . & . & 1 & x \\
. & . & . & . & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ . \\ . \\ . \\ 1
\end{bmatrix},
\quad
s =
\begin{bmatrix}
x(1-yx)(3-yx) \\ (1-yx)(3-yx) \\
-x(3-yx) \\ 3-yx \\ -x \\ 1
\end{bmatrix}
\end{equation}
Clearly, $p = xyxyx - 4xyx + 3x = q$
and $\rank p = \rank q = 6$.
\subsection{Constructing a minimal ALS}
If a minimal ALS (for some polynomial) cannot be stated directly by
a left or right companion system (Definition~\ref{def:pf.cs}),
an ALS for $p$
can be constructed using rational operations from
Proposition~\ref{pro:pf.ratop} (except the inverse)
out of monomials (or polynomials).
If the involved systems are pre-standard,
then the resulting system for the sum can be easily transformed
into a pre-standard ALS by row operations
and we can apply Algorithm~\ref{alg:pf.minals} to
obtain a minimal one.
Here we describe an alternative approach which works
directly on the left and right families like
\cite{Cardon1980a
. It preserves the upper triangular form of the system
matrix. However, for an alphabet with more than one letter,
it is far from optimal because then the
dimension of the vector spaces (see below) might
grow exponentially. We leave it to the reader to
use the ideas of \cite{Cardon1980a
\ to improve that significantely.
Recall that an ALS $\als{A} = (u,A,v)$ is minimal if and only if both the
left family $s = A^{-1} v$ is $\field{K}$-linearly independent and the
right family $t = u A^{-1}$ is $\field{K}$-linearly independent
(Proposition~\ref{pro:pf.cohn94.47}).
Let $f = 3x - 4xyx$. An ALS for $f$ of dimension $n=6$ is given by
\begin{displaymath}
\begin{bmatrix}
1 & - x & . & . & -1 & . \\
. & 1 & - y & . & . & . \\
. & . & 1 & -x & . & . \\
. & . & . & 1 & . & . \\
. & . & . & . & 1 & -x \\
. & . & . & . & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ . \\ -4 \\ . \\ 3
\end{bmatrix},
\quad s =
\begin{bmatrix}
3x - 4xyx \\ -4yx \\ -4x \\ -4 \\ 3x \\ 3
\end{bmatrix}.
\end{displaymath}
Note that the free associative algebra $\freeALG{\field{K}}{X}$
is a \emph{vector space} with basis $X^*$,
the free monoid of the alphabet $X$.
Here $X^* = \{ 1, x, y, xx, xy, yx, yy, \ldots \}$.
Therefore,
we can write the left family $s$ as a matrix (of coordinate row vectors)
$S \in \field{K}^{n \times m_s}$
with column indices $\{ 1, x, yx, xyx \}$
and the right family $t$ as a matrix (of coordinate column vectors)
$T \in \field{K}^{m_t \times n}$
with row indices $\{ 1, x, xy, xyx \}$, that is,
\begin{displaymath}
S =
\begin{bmatrix}
. & 3 & . & -1 \\
. & . & -4 & . \\
. & -4 & . & . \\
-4 & . & . & . \\
. & 3 & . & . \\
3 & . & . & .
\end{bmatrix}
\quad\text{and}\quad
T =
\begin{bmatrix}
1 & . & . & . & 1 & . \\
. & 1 & . & . & . & 1 \\
. & . & 1 & . & . & . \\
. & . & . & 1 & . & .
\end{bmatrix}.
\end{displaymath}
Both, $S$ and $T$, have rank~$4 < n = 6$, hence $\als{A}$ cannot be
minimal. The entries of the left (respectively right) family are just rows in $S$,
also denoted by $s_i$ for $i =1,\ldots,n$
(respectively columns in $T$, denoted by $t_j$ for $j=1,\ldots,n$).
Adding a row $s_k$ to $s_i$ results in \emph{subtracting} column~$i$
from column~$k$ in the system matrix $A$. In order to keep the
triangular structure $i < k$ must hold. Similarly,
adding a column $t_k$ to $t_j$ results in \emph{subtracting}
row~$j$ to row~$k$, hence $j > k$ must hold.
If we want to construct zeros in row~$i=3$ (we cannot produce zeros in
rows~1 and~2) in $S$ we need to find
an (invertible) transformation matrix $Q\in \field{K}^{n \times n}$
of the form (note that we will apply $Q^{-1}$ from the left)
\begin{displaymath}
Q =
\begin{bmatrix}
1 & . & . & . & . & . \\
& 1 & . & . & . & . \\
& & 1 & \beta_4 & \beta_5 & \beta_6 \\
& & & 1 & . & . \\
& & & & 1 & . \\
& & & & & 1
\end{bmatrix}
\end{displaymath}
by solving the (underdetermined) linear system with
the system matrix consisting of the rows $i+1,i+2,\ldots,n$
and the right hand side consisting of the row~$i$
of~$S$:
\begin{displaymath}
\begin{bmatrix}
\beta_4 & \beta_5 & \beta_6
\end{bmatrix}
\begin{bmatrix}
-4 & . & . & . \\
. & 3 & . & . \\
3 & . & . & .
\end{bmatrix}
=
\begin{bmatrix}
. & -4 & . & . \\
\end{bmatrix}.
\end{displaymath}
Here $\beta_5 = -4/3$ and we choose $\beta_4 = \beta_6 = 0$.
Now row~3 in $Q^{-1} S$ is zero
\begin{displaymath}
Q^{-1} S =
\begin{bmatrix}
1 & . & . & . & . & . \\
& 1 & . & . & . & . \\
& & 1 & 0 & 4/3 & 0 \\
& & & 1 & . & . \\
& & & & 1 & . \\
& & & & & 1
\end{bmatrix}
\begin{bmatrix}
. & 3 & . & -1 \\
. & . & -4 & . \\
. & -4 & . & . \\
-4 & . & . & . \\
. & 3 & . & . \\
3 & . & . & .
\end{bmatrix}
=
\begin{bmatrix}
. & 3 & . & -1 \\
. & . & -4 & . \\
. & 0 & . & . \\
-4 & . & . & . \\
. & 3 & . & . \\
3 & . & . & .
\end{bmatrix},
\end{displaymath}
that is $s'_3 = (Q^{-1} s)_3 = 0$. Therefore we can remove
row~3 and column~3 from the (admissibly) modified system
$(uQ, AQ, v)$ and get one of dimension $n-1=5$:
\begin{displaymath}
\begin{bmatrix}
1 & - x & . & -1 & . \\
. & 1 & . & \frac{4}{3} y & . \\
. & . & 1 & . & . \\
. & . & . & 1 & -x \\
. & . & . & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ -4 \\ . \\ 3
\end{bmatrix},
\quad s =
\begin{bmatrix}
3x - 4xyx \\ -4yx \\ -4 \\ 3x \\ 3
\end{bmatrix}.
\end{displaymath}
The right family of the new ALS is $\bigl(1,x,0,1-\frac{4}{3} xy, x-\frac{4}{3} xyx\bigr)$.
After adding $4/3$-times row~5 to row~3 we can remove row~3 and column~3
and get a \emph{minimal} ALS for $3x - 4xyx = (1-\frac{4}{3} xy) 3x$:
\begin{displaymath}
\begin{bmatrix}
1 & -x & -1 & . \\
. & 1 & \frac{4}{3} y & . \\
. & . & 1 & -x \\
. & . & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ . \\ 3
\end{bmatrix},
\quad
s =
\begin{bmatrix}
3x - 4xyx \\
-4yx \\
3x \\
3
\end{bmatrix}.
\end{displaymath}
\subsection{Factorizing a Polynomial}
Now we consider the following \emph{minimal} ALS for $p = xyxyx + (3x - 4xyx)$,
constructed in a similar way as shown in the previous subsection
or using the right companion system (Definition~\ref{def:pf.cs}):
\begin{displaymath}
\begin{bmatrix}
1 & - x & . & . & . & -x \\
. & 1 & -y & . & . & . \\
. & . & 1 & -x & . & \frac{4}{3} x \\
. & . & . & 1 & -y & . \\
. & . & . & . & 1 & -\frac{1}{3} x \\
. & . & . & . & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ . \\ . \\ . \\ 3
\end{bmatrix}.
\end{displaymath}
We try to create an upper right block of zeros of size $3 \times 2$.
For that we apply the (admissible) transformation $(P,Q)$ directly
to the coefficient matrices $A_0$, $A_x$ and $A_y$ in
$A = A_0 \otimes 1 + A_x \otimes x + A_y \otimes y$
to get the equations. For $y$ we have
\begin{align*}
P A_y Q &=
P
\begin{bmatrix}
. & . & . & . & . & . \\
& . & -1 & . & . & . \\
& & . & . & . & . \\
& & & . & -1 & . \\
& & & & . & . \\
& & & & & .
\end{bmatrix}
\begin{bmatrix}
1 & 0 & 0 & 0 & 0 & 0 \\
& 1 & \beta_{2,3} & \beta_{2,4} & \beta_{2,5} & \beta_{2,6} \\
& & 1 & \beta_{3,4} & \beta_{3,5} & \beta_{3,6} \\
& & & 1 & \beta_{4,5} & \beta_{4,6} \\
& & & & 1 & \beta_{5,6} \\
& & & & & 1
\end{bmatrix}
\end{align*}
\begin{align*}
\quad\quad\quad
&=
\begin{bmatrix}
1 & \alpha_{1,2} & \alpha_{1,3} & \alpha_{1,4} & \alpha_{1,5} & 0 \\
& 1 & \alpha_{2,3} & \alpha_{2,4} & \alpha_{2,5} & 0 \\
& & 1 & \alpha_{3,4} & \alpha_{3,5} & 0 \\
& & & 1 & \alpha_{4,5} & 0 \\
& & & & 1 & 0 \\
& & & & & 1
\end{bmatrix}
\begin{bmatrix}
. & . & . & . & . & . \\
& . & -1 & -\beta_{3,4} & -\beta_{3,5} & -\beta_{3,6} \\
& & . & . & . & . \\
& & & . & -1 & -\beta_{5,6} \\
& & & & . & . \\
& & & & & .
\end{bmatrix} \\
&=
\begin{bmatrix}
. & . & -\alpha_{1,2} & -\alpha_{1,2}\beta_{3,4} &
-\alpha_{1,2}\beta_{3,5} - \alpha_{1,4} &
-\alpha_{1,4}\beta_{5,6} - \alpha_{1,2} \beta_{3,6} \\
& . & -1 & -\beta_{3,4} & -\beta_{3,5} - \alpha_{2,4} &
-\alpha_{2,4}\beta_{5,6} - \beta_{3,6} \\
& & . & . & -\alpha_{3,4} & -\alpha_{3,4}\beta_{5,6} \\
& & & . & -1 & -\beta_{5,6} \\
& & & & . & . \\
& & & & & .
\end{bmatrix}
\end{align*}
of which we pick the upper right $3 \times 2$ block. Thus we
have the following~6 equations for $y$:
\begin{displaymath}
\begin{bmatrix}
\alpha_{1,2}\beta_{3,5} + \alpha_{1,4} &
\alpha_{1,4}\beta_{5,6} + \alpha_{1,2} \beta_{3,6} \\
\beta_{3,5} + \alpha_{2,4} &
\alpha_{2,4}\beta_{5,6} + \beta_{3,6} \\
\alpha_{3,4} & \alpha_{3,4}\beta_{5,6}
\end{bmatrix}
=
\begin{bmatrix}
0 & 0 \\
0 & 0 \\
0 & 0
\end{bmatrix}.
\end{displaymath}
Similarly, we get the following 6 equations for $x$:
\begin{displaymath}
\begin{bmatrix}
\alpha_{1,3} \beta_{4,5} + \beta_{2,5} &
\alpha_{1,3} \beta_{4,6} + \beta_{2,6} + \frac{1}{3} \alpha_{1,5} - \frac{4}{3} \alpha_{1,3} + 1 \\
\alpha_{2,3} \beta_{4,5} &
\alpha_{2,3} \beta_{4,6} + \frac{1}{3} \alpha_{2,5} - \frac{4}{3} \alpha_{2,3} \\
\beta_{4,5} & \beta_{4,6} + \frac{1}{3} a_{3,5} - \frac{4}{3}
\end{bmatrix}
= 0.
\end{displaymath}
And finally those for $1$ (without terms containing
$\alpha_{3,4} = \beta_{4,5} = 0$):
\begin{displaymath}
\begin{bmatrix}
\alpha_{1,3} \beta_{3,5} + \alpha_{1,2}\beta_{2,5} + \alpha_{1,5} &
\alpha_{1,5} \beta_{5,6} + \alpha_{1,4} \beta_{4,6} +
\alpha_{1,3} \beta_{3,6} + \alpha_{1,2} \beta_{2,6} \\
\alpha_{2,3} \beta_{3,5} + \beta_{2,5} + \alpha_{2,5} &
\alpha_{2,5} \beta_{5,6} + \alpha_{2,4} \beta_{4,6} +
\alpha_{2,3} \beta_{3,6} + \beta_{2,6} \\
\beta_{3,5} + \alpha_{3,5} &
\alpha_{3,5} \beta_{5,6} + \beta_{3,6}
\end{bmatrix}
= 0.
\end{displaymath}
\medskip
A GrΓΆbner basis for the ideal generated by these 18 equations
(computed by \textsc{FriCAS},
\cite{FRICAS2016
\ using \emph{lexicographic order})
is $\bigl(
\alpha_{1,2} - \alpha_{1,4}\beta_{4,6},\;
\alpha_{1,3} - \alpha_{1,5} \beta_{4,6},\;
\alpha_{2,3} - \alpha_{2,5} \beta_{4,6},\;
\alpha_{2,4} + 3 \beta_{4,6} - 4,\;
\alpha_{3,4},\;
\alpha_{3,5} + 3\beta_{4,6} - 4,\;
\beta_{2,5},\;
\beta_{2,6} + 1,\;
\beta_{3,5} - 3 \beta_{4,6} + 4,\;
\beta_{3,6} - 3 \beta_{4,6} \beta_{5,6} + 4 \beta_{5,6},\;
\beta_{4,5},\;
\beta_{4,6}^2 - \tsfrac{4}{3} \beta_{4,6} + \tsfrac{1}{3} \bigr)$.
The last (over $\field{K}$ reducible)
generator is $(\beta_{4,6} -1)(\beta_{4,6}-\frac{1}{3})$,
thus $\beta_{4,6} \in \{ 1, \frac{1}{3} \}$.
We proceed with the case $\beta_{4,6}=1$.
We then have
$\alpha_{2,4} = \alpha_{3,5} = 1$,
$\beta_{2,6} = \beta_{3,5} = -1$
and choose
$\alpha_{1,2} = \alpha_{1,4} = 0$,
$\alpha_{1,3} = \alpha_{1,5} = 3$,
$\alpha_{2,3} = \alpha_{2,5} = 0$,
$\beta_{3,6} = -\beta_{5,6} = 0$.
Note that $\alpha_{4,5}$, $\beta_{2,3}$, $\beta_{2,4}$ and
$\beta_{3,4}$ are not present, so we set them to zero.
Therefore one possible transformation is
\begin{displaymath}
(P,Q) = \left(
\begin{bmatrix}
1 & 0 & 3 & 0 & 3 & . \\
& 1 & 0 & 1 & 0 & . \\
& & 1 & 0 & 1 & . \\
& & & 1 & 0 & . \\
& & & & 1 & . \\
& & & & & 1
\end{bmatrix},
\begin{bmatrix}
1 & . & . & . & . & . \\
& 1 & 0 & 0 & 0 & -1 \\
& & 1 & 0 & -1 & 0 \\
& & & 1 & 0 & 1 \\
& & & & 1 & -0 \\
& & & & & 1
\end{bmatrix}
\right)
\end{displaymath}
leading to the admissible linear system
$\als{A}' = P \als{A} Q$
\begin{displaymath}
\begin{bmatrix}
1 & -x & 3 & -3x & 0 & 0 \\
. & 1 & -y & 1 & 0 & 0 \\
. & . & 1 & -x & 0 & 0 \\
. & . & . & 1 & -y & 1 \\
. & . & . & . & 1 & -\tsfrac{1}{3}x \\
. & . & . & . & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ . \\ . \\ . \\ 3
\end{bmatrix},
\end{displaymath}
compare with the ALS \eqref{eqn:pf.ex.1}.
$\als{A}'$ is the (minimal) product of
\begin{equation}\label{eqn:pf.ex.3}
\begin{bmatrix}
1 & -x & 3 & -3x \\
. & 1 & -y & 1 \\
. & . & 1 & -x \\
. & . & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ . \\ 1
\end{bmatrix},
\quad
s =
\begin{bmatrix}
xyx - x \\
yx - 1 \\
x \\
1
\end{bmatrix}
\end{equation}
and
\begin{equation}\label{eqn:pf.ex.4}
\begin{bmatrix}
1 & -y & 1 \\
. & 1 & -\tsfrac{1}{3}x \\
. & . & 1
\end{bmatrix}
s =
\begin{bmatrix}
. \\ . \\ 3
\end{bmatrix},
\quad
s =
\begin{bmatrix}
yx - 3 \\
x \\
3
\end{bmatrix}.
\end{equation}
Thus $p = (xyx-x)(yx-3)$.
The first factor is \emph{not} atomic because we could (pre-standard
admissibly) construct
either an $1 \times 2$ upper right block of zeros
(by subtracting $3$-times row~3 from row~1)
or an $2 \times 1$ upper right block of zeros
(by subtracting column~2 from column~4
and subtracting $2$-times row~3 from row~1)
in the system matrix of $\eqref{eqn:pf.ex.3}$.
On the other hand,
a brief look at the ALS $\eqref{eqn:pf.ex.4}$
immediately shows that the second factor is irreducible.
\section*{Acknowledgement}
I thank Franz Lehner for the fruitful discussions
and his support, Daniel Smertnig who helped me
to find some orientation in abstract non-commutative factorization
(especially with literature),
Roland Speicher for the creative environment in SaarbrΓΌcken
and the anonymous referees for the valuable comments.
\ifJOURNAL
This work has been supported by the Austrian FWF Project P25510-N26
``Spectra on Lamplighter groups and Free Probability''.
\fi
\ifJOURNAL
\bibliographystyle{elsarticle-harv}
| c7542d96e4cda88f5b1e5fe81bda7f6b98628a25 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{{\bf Introduction}}
\vskip 3mm
Some well known random fixed point theorems are stochastic generalizations of Banach's fixed point theorem and Banach's type fixed point theorems in complete metric spaces. In 1955, Spacek \cite{Spacek} and Hans \cite{Hans1, Hans2} initiated to prove random fixed point theorems for random contraction mappings in separable complete metric spaces. In 1966, Mukherjee \cite{Mukherjee} proved a random fixed point theorem in the sense of Schaduer's fixed point theorem in atomic probability measure spaces. Especially, in 1976, the work of Bharucha-Reid \cite{BharuchaReid} has been developed by various mathematicians. In 1979, Itoh \cite{Itoh} extended some random fixed point theorems of Spacek and Hans to the setting of multi-valued contraction mappings and applied random fixed point theorems to solve some random differential equations in Banach spaces. In 1984, Sehgal and Waters \cite{SehgalWaters} proved some random fixed point theorems including classical results given by Rothe \cite{Rothe}.
Recently, Beg and Shahzad \cite{BegShahzad} showed the existence of random common fixed points and random coincidence points of a pair of compatible random multi-valued mappings in Polish spaces. Especially, Kumam et al. \cite{KumamKumam01, KumamPlubtieng04, KumamPlubtieng08, Kumam02, KumamKumam02} proved many random fixed point theorems for multi-valued nonexpansive nonself-mappings satisfying the inwardness condition in Banach spaces (see \cite{KumamPlubtieng05}). Jung et al. \cite{JungChoKangLeeThakur} proved random fixed point theorems for a certain
class of mappings in banach spaces. Cho et al. \cite{ChoLiHuang} proved random Ishikawa iterative sequence with errors for approximating random fixed points. Likewise, Kumam and Plubtieng \cite{KumamPlubtieng01} showed the existence of a random coincidence point for a pair of reciprocally continuous and compatible single-valued and multi-valued mappings and Saha \cite{Saha}, Saha and Debnath \cite{SahaDebnath} established some random fixed point theorems in separable Hilbert spaces and separable Banach spaces, respectively. On the other hand, Padgett \cite{Padgett}, Achari \cite{Achari}, Saha and Dey \cite{SahaDey} applied some random fixed point theorems to show the existence of solutions of random nonlinear integral equations in Banach spaces.
Recently, Saha and Ganguly \cite{SahaGanguly} proved some random fixed point theorems for a class of contractive mappings in separable Banach spaces equipped with a complete probability measure.
\vskip 2mm
In fact, Banach's contraction principle (\cite{Banach}) is very important to show the existence of solutions of some nonlinear equations, differential and integral equations, and other nonlinear problems. Since Banach's contraction principle, many authors have studied in several ways.
\vskip 2mm
\begin{thm} {\rm (Banach's contraction principle)} If $(X, d)$ is a complete metric space and $T : X \rightarrow X$ be a mapping such that,
for some $\alpha \in [0, 1)$,
\begin{eqnarray} \label{IntroHardyRogerscontraction001}
d(Tx, Ty) \leq \alpha d(x, y)
\end{eqnarray}
for each $x, y \in X$, then $T$ has a unique fixed point in $X$.
\end{thm}
\vskip 2mm
Note that the mapping $T$ satisfying the Banach contraction condition is continuous, but the mappings $T$ satisfying the following contractions conditions are not continuous.
\vskip 2mm
(1)\, In 1968, Kannan's contraction (\cite{Rhoades}): for some $\beta \in [0, \frac{1}{2})$,
\begin{eqnarray} \label{IntroHardyRogerscontraction002}
d(Tx, Ty) \leq \beta [d(x, Tx) + d(y, Ty)]
\end{eqnarray}
for each $x, y \in X$;
\vskip 1mm
(2)\, In 1971, Reich's contraction (\cite{Rhoades}): for some $\alpha, \beta, \gamma \geq 0$ with $\alpha + \beta + \gamma < 1$,
\begin{eqnarray} \label{IntroHardyRogerscontraction003}
d(Tx, Ty) \leq \alpha d(x, Tx) + \beta d(y, Ty) + \gamma d(y, Ty)
\end{eqnarray}
for each $x, y \in X$;
\vskip 1mm
(3)\, In 1971, $\acute{\text{C}}$iri$\acute{\text{c}}$'s contraction (\cite{Rhoades}): for some $\alpha, \beta, \gamma, \delta \geq 0$ with $\alpha + \beta + \gamma + 2\delta < 1$,
\begin{eqnarray} \label{IntroHardyRogerscontraction004}
d(Tx, Ty) \leq \alpha d(x, y) + \beta d(x, Tx) + \gamma d(y, Ty) + \delta [d(x, Ty) + d(y, Tx)]
\end{eqnarray}
for each $x, y \in X$;
\vskip 1mm
(4)\, In 1972, Chatterjea's contraction (\cite{Rhoades}): for some $\beta \in [0, \frac{1}{2})$,
\begin{eqnarray} \label{IntroHardyRogerscontraction005}
d(Tx, Ty) \leq \beta [d(x, Ty) + d(y, Tx)]
\end{eqnarray}
for each $x, y \in X$;
\vskip 1mm
(5)\, In 1972, Zamfirescu contractive conditions (\cite{Rhoades}): there exist real numbers $\alpha, \beta, \gamma, 0 \leq \alpha < 1, 0 \leq \beta < 1, \gamma < \frac{1}{2}$, such that, for each $x, y \in X$, at least one of the following is true:
\begin{enumerate}
\item [(i)] $d(Tx, Ty) \leq \alpha d(x, y)$;
\item [(ii)] $d(Tx, Ty) \leq \beta [d(x, Tx) + d(y, Ty)]$;
\item [(iii)] $d(Tx, Ty) \leq \gamma [d(x, Ty) + d(y, Tx)]$.
\end{enumerate}
For each $x, y \in X, x \neq y$,
\begin{eqnarray} \label{IntroHardyRogerscontraction005b}
d(Tx, Ty) < \max \{ d(x, y), [d(x, Tx) + d(y, Ty)]/2, [d(x, Ty) + d(y, Tx)]/2 \}.
\end{eqnarray}
\vskip 1mm
(6)\, In 1973, Hardy and Rogers's contraction (\cite{Rhoades}): for some $\alpha, \beta, \gamma, \delta, \eta \geq 0$ with $\alpha + \beta + \gamma + \delta + \eta < 1$,
\begin{eqnarray} \label{IntroHardyRogerscontraction006}
d(Tx, Ty) \leq \alpha d(x, y) + \beta d(x, Tx) + \gamma d(y, Ty) + \delta d(x, Ty) + \eta d(y, Tx)
\end{eqnarray}
for each $x, y \in X$:
\vskip 2mm
In 2000, $\acute{\text{C}}$iri$\acute{\text{c}}$ \cite{Ciric01} dealt with a class of mappings (not necessarily continuous) satisfying Gregus type contraction in metric spaces (\cite{Gregus01}) and proved the following fixed point theorem:
\vskip 2mm
\begin{thm}
Let $C$ be a closed convex subset of a complete convex metric space $X$ and $T : C \rightarrow C$ be a mapping satisfying
\begin{eqnarray} \label{IntroHardyRogerscontraction007}
d(Tx, Ty) \leq ad(x, y) + b\max \{d(x, Tx), d(y, Ty)\} + c [d(x, Ty) + d(y, Tx)]
\end{eqnarray}
for all $x, y \in C$, where $0 < a < 1$, $a + b = 1$ and $c \leq \frac{4 - a}{8 - b}$. Then $T$ has a unique fixed point in $X$.
\end{thm}
\vskip 2mm
Moreover, Common fixed points under contractive conditions in cone metric spaces was studied by Radenovi$\acute{\text{c}}$ (see in \cite{Radenovic}).
Recently, Saha and Ganguly \cite{SahaGanguly} proved some random fixed point theorems for a certain class of contractive mappings in a separable Banach space equipped with a complete probability measure as follows:
\vskip 2mm
\begin{thm} \label{MainMotivate}
Let $X$ be a separable Banach space and $(\Omega, \beta, \mu)$ be a complete probability measure space. Let $T : \Omega \times X \rightarrow X$ be a continuous random operator such that
for all $\omega \in \Omega$, $T$ satisfies
\begin{eqnarray} \label{IntroHardyRogerscontraction008}
&& \|T(\omega, x_1)-T(\omega, x_2)\|\nonumber\\
&\leq& a(\omega) \max \{\|x_1 - x_2\|, \nonumber\\
&& \qquad\quad \qquad c(\omega) [\|x_1 - T(\omega, x_2)\| + \|x_2 - T(\omega, x_1)\|]\}\nonumber\\
&&+ b(\omega) \max\{\|x_1 - T(\omega, x_1)\|, \|x_2 - T(\omega, x_2)\|\} \\ \nonumber
\end{eqnarray}
for all random variables $x_1, x_2 \in X$ where $a(\omega), b(\omega), c(\omega)$ are real-valued random variables such that $0 < a(\omega) < 1$, $a(\omega) + b(\omega) = 1$, $c(\omega) \leq \frac{4 - a(\omega)}{8 - a(\omega)}$ almost surely. Then there exist unique random fixed point of $T$ in $X$.
\end{thm}
\vskip 3mm
Note that, if $\beta = 0$ or $\gamma = 0$ and $\delta = \eta$, then fixed point theorems for Hardy and Roger's contraction \eqref{IntroHardyRogerscontraction006} reduced to fixed point theorems for Gregus type contraction \eqref{IntroHardyRogerscontraction007}.
\vskip 2mm
The purpose of this paper is to prove some random fixed point theorems for random Hardy-Rogers self-mappings in separable Banach spaces and, by using our main results, we show the existence of solutions of random nonlinear integral equations.
\vskip 4mm
\section{{\bf Preliminaries}}
\vskip 3mm
Throughout this paper, X will denote a separable Banach over the real. Let $\beta_X$ be a $\sigma$-algebra of Borel subsets of $X$. Let $(\Omega, \beta, \mu)$ denote a complete probability measure space with the measure $\mu$ and $\beta$ be a
$\sigma$-algebra of subsets of $\Omega$. For more details, see Joshi and Bose \cite{JoshiBose}.
\vskip 2mm
\begin{dfn} \label{def2.1}
(1)\, A mapping $x : \Omega \rightarrow X$ is called an {\it $X$-valued random variable} if the inverse image under the mapping $x$ of every Borel set $B$ of $X$ belongs to $\beta$, that is, $x^{-1}(X) \in \beta$ for all $B \in \beta_X$.
(2) A mapping $x : \Omega \rightarrow X$ is called a {\it finitely-valued random variable} if it is constant on each finite number of disjoint sets $A_i \in \beta$ and is equal to $0$ on $\Omega - (\bigcup_{i=1}^n A_i)$. The mapping $x$ is called a {\it simple random variable} if it is finitely valued and $\mu\{\omega : \|x(\omega)\| > 0\} < \infty$.
(3)\, A mapping $x : \Omega \rightarrow X$ is called a {\it strong random variable} if there exists a sequence $\{x_n(\omega)\}$ of simple random variables which converges to $x(\omega)$ almost surely, that is, there exists a set $A_0 \in \beta$ with $\mu(A_0) = 0$ such that $\lim_{n \rightarrow \infty} x_n(\omega) = x(\omega)$ for any $\omega \in \Omega - A_0$.
(4)\, A mapping $x : \Omega \rightarrow X$ is called a {\it weak random variable} if the function $x^*(x(\omega))$ is a real-valued random variable for each $x^* \in X^*$, where $X^*$ denots the first normed dual space of $X$.
\end{dfn}
\vskip 2mm
In a separable Banach space $X$, the notions of strong and weak random variables $x : \Omega \rightarrow X$ (\cite{JoshiBose}) coincide and, in $X$, $x$ is termed as a random variable.
\vskip 2mm
Now, we recall the following:
\vskip 2mm
\begin{thm} \label{thm2.5}
{\rm (\cite{JoshiBose})} Let $x, y : \Omega \rightarrow X$ be strong random variables and $\alpha, \beta$ be constants. Then the following statements hold:
\vskip 1mm
{\rm (1)}\, $\alpha x(\omega) + \beta y(\omega)$ is a strong random variable.
{\rm (2)}\, If $f(\omega)$ is a real-valued random variable and $x(\omega)$ is a strong random variable, then
$f(\omega) x(\omega)$ is a strong random variable.
{\rm (3)}\, If $x_n(\omega)$ is a sequence of strong random variables converging strongly to $x(\omega)$ almost
surely, that is, if there exists a set $A_0 \in \beta$ with $\mu(A_0) = 0$ such that
$$
\lim_{n \rightarrow \infty} \|x_n(\omega) - x(\omega)\| = 0
$$
for any $\omega \notin A_0$, then $x(\omega)$ is a strong random variable.
\end{thm}
\vskip 2mm
\begin{rmk}
If $X$ is a separable Banach space, then every strong and also weak random variable is measurable in the sense of Definition \ref{def2.1}.
\end{rmk}
\vskip 2mm
Let $Y$ be an another Banach space. We also need the following definitions (see Joshi
and Bose \cite{JoshiBose}).
\vskip 2mm
\begin{dfn}
(1)\, A mapping $F : \Omega \times X \rightarrow Y$ is called a {\it random mapping} if $F(\omega, x) = Y(\omega)$ is a $Y$-valued random variable for all $x \in X$.
(2)\, A mapping $F : \Omega \times X \rightarrow Y$ is called a {\it continuous random mapping} if the set of all $\omega \in \Omega$ for which $F(\omega, x)$ is a continuous function of $x$ has measure one.
(3)\, A mapping $F : \Omega \times X \rightarrow Y$ is said to be {\it demicontinuous} at the $x \in X$ if $\|x_n - x\| \rightarrow 0$ implies $F(\omega, x_n) \rightharpoonup F(\omega, x)$ almost surely.
\end{dfn}
\vskip 2mm
\begin{thm}
{\rm (\cite{JoshiBose})} Let $F : \Omega \times X \rightarrow Y$ be a demicontinuous random mapping where a Banach space $Y$ is separable. Then, for any $X$-valued random variable $x$, the function $F(\omega, x(\omega))$ is a $Y$-valued random variable.
\end{thm}
\vskip 2mm
\begin{rmk}
(\cite{JoshiBose}) Since a continuous random mapping is a demicontinuous random mapping, Theorem \ref{thm2.5} is also true for a continuous random mapping.
\end{rmk}
\vskip 2mm
Also, we recall the following definitions (see Joshi and Bose \cite{JoshiBose}):
\vskip 2mm
\begin{dfn}
(1)\, An equation of the type $F(\omega, x(\omega)) = x(\omega)$, where $F : \Omega \times X \rightarrow X$ is a random mapping, is called a {\it random fixed point equation}.
(2)\, Any mapping $x : \Omega \rightarrow X$ which satisfies the random fixed point equation $F(\omega, x(\omega)) = x(\omega)$ almost surely is called
a {\it wide sense solution} of the fixed point equation.
(3)\, Any $X$-valued random variable $x(\omega)$ which satisfies
\begin{center}
$\mu\{\omega : F(\omega, x(\omega)) = x(\omega)\} = 1$
\end{center}
is called a {\it random solution} of the fixed point equation or a random fixed point of $F$.
\end{dfn}
\vskip 2mm
\begin{rmk}
A random solution is a wide sense solution of the fixed point equation. But the converse is not necessarily true. This is evident from an example, under Remark 1, in Joshi and Bose \cite{JoshiBose}.
\end{rmk}
\vskip 4mm
\section{{\bf The main results}}
\vskip 3mm
Motivated and inspired by Theorem \ref{MainMotivate}, we proposed the definition as follows:
\vskip 2mm
\begin{dfn}
Let $T : \Omega \times X \rightarrow X$ be a continuous random mapping. The random mapping $T$ is called {\it Hardy-Rogers' contraction} if, for any $\omega \in \Omega$,
\begin{eqnarray} \label{HardyRogerscontraction001}
&& \|T(\omega, x_1(\omega))-T(\omega, x_2(\omega))\|\nonumber\\
&\leq& \alpha_1(\omega) \|x_1(\omega) - x_2(\omega)\| + \alpha_2(\omega) \|x_1(\omega) - T(\omega, x_1(\omega))\| \\ \nonumber
&& + \alpha_3(\omega) \|x_2(\omega) - T(\omega, x_2(\omega))\| + \alpha_4(\omega) \|x_1(\omega) - T(\omega, x_2(\omega))\| \\ \nonumber
&& + \alpha_5(\omega) \|x_2(\omega) - T(\omega, x_1(\omega))\|
\end{eqnarray}
for all random variables $x_1, x_2 : \Omega \rightarrow X$ and $\alpha_i : \Omega \rightarrow \mathbb{R}_+ \cup \{0\}$ for $i = 1, 2, 3, 4, 5$ such that $\sum^5_{i=1} \alpha_i(\omega) < 1$.
\end{dfn}
\vskip 2mm
\begin{thm} \label{Maintheorem1}
Let $X$ be a separable Banach space and $(\Omega, \beta, \mu)$ be a complete probability measure space. Let $T : \Omega \times X \rightarrow X$ be a continuous random mapping
satisfying Hardy-Rogers' contraction. Then there exists a unique random fixed point of $T$ in $X$.
\end{thm}
\vskip 2mm
\begin{proof}
Let
\begin{eqnarray*}
A = \{\omega \in \Omega : T(\omega, x_1) ~~~~ \text{is a continuous function of} ~~~~ x \},
\end{eqnarray*}
\begin{eqnarray*}
B = \Big\{\omega \in \Omega : \sum^5_{i=1} \alpha_i(\omega) < 1 \Big\}
\end{eqnarray*}
and
\begin{eqnarray*}
C_{x_1, x_2} &=& \{\omega \in \Omega : \|T(\omega, x_1(\omega))-T(\omega, x_2(\omega))\| \leq \alpha_1(\omega) \|x_1(\omega) - x_2(\omega)\| \\
&&\qquad\quad + \alpha_2(\omega) \|x_1(\omega) - T(\omega, x_1(\omega))\| + \alpha_3(\omega) \|x_2(\omega) - T(\omega, x_2(\omega))\| \\
&&\qquad\quad + \alpha_4(\omega) \|x_1(\omega) - T(\omega, x_2(\omega))\| + \alpha_5(\omega) \|x_2(\omega) - T(\omega, x_1(\omega))\| \}
\end{eqnarray*}
Let $S$ be a countable dense subset of $X$. Now, we prove that
\begin{center}
$\bigcap_{x_1, x_2 \in X}(C_{x_1, x_2} \cap A \cap B) = \bigcap_{s_1, s_2 \in S}(C_{s_1, s_2} \cap A \cap B)$.
\end{center}
Now, for all $s_1, s_2 \in S$, we have
\begin{eqnarray} \label{HardyRogerscontraction002}
&& \|T(\omega, s_1(\omega))-T(\omega, s_2(\omega))\|\nonumber\\
&\leq& \alpha_1(\omega) \|s_1(\omega) - s_2(\omega)\| + \alpha_2(\omega) \|s_1(\omega) - T(\omega, s_1(\omega))\| \\ \nonumber
&& + \alpha_3(\omega) \|s_2(\omega) - T(\omega, s_2(\omega))\| + \alpha_4(\omega) \|s_1(\omega) - T(\omega, s_2(\omega))\| \\ \nonumber
&& + \alpha_5(\omega) \|s_2(\omega) - T(\omega, s_1(\omega))\|.
\end{eqnarray}
Since $S$ is dense in $X$, for any $\delta_i(x_i) > 0$, there exist $s_1, s_2 \in S$ such that $\|x_i - s_i\| < \delta_i(x_i)$ for each $i = 1, 2$.
Note that, for any $x_1, x_2 \in X$,
\begin{eqnarray} \label{HardyRogerscontraction003}
\|s_1(\omega) - s_2(\omega)\| \leq \|s_1(\omega) - x_1(\omega)\| + \|x_1(\omega) - x_2(\omega)\| + \|x_2(\omega) - s_2(\omega)\|,
\end{eqnarray}
\begin{eqnarray} \label{HardyRogerscontraction004}
\|s_1(\omega) - T(\omega, s_1(\omega))\| &\leq& \|s_1(\omega) - x_1(\omega)\| + \|x_1(\omega) - T(\omega, x_1(\omega)\| \\ \nonumber
&& + \|T(\omega, x_1(\omega) - T(\omega, s_1(\omega)\|,
\end{eqnarray}
\begin{eqnarray} \label{HardyRogerscontraction005}
\|s_2(\omega) - T(\omega, s_2(\omega))\| &\leq& \|s_2(\omega) - x_2(\omega)\| + \|x_2(\omega) - T(\omega, x_2(\omega)\| \\ \nonumber
&& + \|T(\omega, x_2(\omega) - T(\omega, s_2(\omega)\|,
\end{eqnarray}
\begin{eqnarray} \label{HardyRogerscontraction006}
\|s_1(\omega) - T(\omega, s_2(\omega))\| &\leq& \|s_1(\omega) - x_1(\omega)\| + \|x_1(\omega) - T(\omega, x_2(\omega)\| \\ \nonumber
&& + \|T(\omega, x_2(\omega) - T(\omega, s_2(\omega)\|
\end{eqnarray}
and
\begin{eqnarray} \label{HardyRogerscontraction007}
\|s_2(\omega) - T(\omega, s_1(\omega))\| &\leq& \|s_2(\omega) - x_2(\omega)\| + \|x_2(\omega) - T(\omega, x_1(\omega)\| \\ \nonumber
&& + \|T(\omega, x_1(\omega) - T(\omega, s_1(\omega)\|.
\end{eqnarray}
Suppose that
\begin{eqnarray} \label{HardyRogerscontraction008}
&& \|T(\omega, s_1(\omega))-T(\omega, s_2(\omega))\|\nonumber\\
&\leq& \alpha_1(\omega) \|s_1(\omega) - s_2(\omega)\| + \alpha_2(\omega) \|s_1(\omega) - T(\omega, s_1(\omega))\| \\ \nonumber
&& + \alpha_3(\omega) \|s_2(\omega) - T(\omega, s_2(\omega))\| + \alpha_4(\omega) \|s_1(\omega) - T(\omega, s_2(\omega))\| \\ \nonumber
&& + \alpha_5(\omega) \|s_2(\omega) - T(\omega, s_1(\omega))\|.
\end{eqnarray}
Since
\begin{eqnarray} \label{HardyRogerscontraction009}
&&\|T(\omega, x_1(\omega))-T(\omega, x_2(\omega))\|\nonumber\\
&\leq& \|T(\omega, x_1(\omega))-T(\omega, s_1(\omega))\| + \|T(\omega, s_1(\omega))-T(\omega, s_2(\omega))\| \\ \nonumber
&& + \|T(\omega, s_2(\omega))-T(\omega, x_2(\omega))\|,
\end{eqnarray}
substituting \eqref{HardyRogerscontraction008} in \eqref{HardyRogerscontraction009}, we have
\begin{eqnarray} \label{HardyRogerscontraction010}
&&\|T(\omega, x_1(\omega))-T(\omega, x_2(\omega))\|\nonumber\\
&\leq& \|T(\omega, x_1(\omega))-T(\omega, s_1(\omega))\| + \|T(\omega, s_2(\omega))-T(\omega, x_2(\omega))\| \\ \nonumber
&& + \alpha_1(\omega) \|s_1(\omega) - s_2(\omega)\| + \alpha_2(\omega) \|s_1(\omega) - T(\omega, s_1(\omega))\| \\ \nonumber
&& + \alpha_3(\omega) \|s_2(\omega) - T(\omega, s_2(\omega))\| + \alpha_4(\omega) \|s_1(\omega) - T(\omega, s_2(\omega))\| \\ \nonumber
&& + \alpha_5(\omega) \|s_2(\omega) - T(\omega, s_1(\omega))\|.
\end{eqnarray}
Thus, from \eqref{HardyRogerscontraction003}, \eqref{HardyRogerscontraction004}, \eqref{HardyRogerscontraction005}, \eqref{HardyRogerscontraction006}, \eqref{HardyRogerscontraction007}, \eqref{HardyRogerscontraction010}, it follows that
\begin{eqnarray} \label{HardyRogerscontraction011}
&&\|T(\omega, x_1(\omega))-T(\omega, x_2(\omega))\|\nonumber\\
&\leq& \alpha_1(\omega)\|x_1(\omega)-x_2(\omega)\| + \alpha_2(\omega)\|x_1(\omega)- T(\omega, x_1(\omega))\| \\ \nonumber
&& + \alpha_3(\omega)\|x_2(\omega)- T(\omega, x_2(\omega))\| + \alpha_4(\omega)\|x_1(\omega)- T(\omega, x_2(\omega))\| \\ \nonumber
&& + \alpha_5(\omega)\|x_2(\omega)- T(\omega, x_1(\omega))\| \\ \nonumber
&& + (1 + \alpha_2(\omega) + \alpha_5(\omega))\|T(\omega, x_1(\omega))-T(\omega, s_1(\omega))\| \\ \nonumber
&& +(1 + \alpha_3(\omega) + \alpha_4(\omega))\|T(\omega, x_2(\omega))-T(\omega, s_2(\omega))\| \\ \nonumber
&& +(\alpha_1(\omega) + \alpha_2(\omega) + \alpha_4(\omega))\|s_1(\omega)-x_1(\omega)\| \\ \nonumber
&& +(\alpha_1(\omega) + \alpha_2(\omega) + \alpha_4(\omega))\|s_2(\omega)-x_2(\omega)\| .
\end{eqnarray}
For any $\omega \in \Omega$, sice $T(\omega, x(\omega))$ is a continuous function of $x(\omega)$, for any $\varepsilon > 0$, there exists $\delta_i(x_i(\omega)) > 0 ~~~ (i = 1, 2)$ such that
\begin{eqnarray} \label{HardyRogerscontraction012}
\| T(\omega, x_1(\omega)) - T(\omega, s_1(\omega))\| < \frac{\varepsilon}{8}
\end{eqnarray}
whenever $\| x_1(\omega) - s_1(\omega) \| < \delta_1(x_1(\omega))$ and
\begin{eqnarray} \label{HardyRogerscontraction013}
\| T(\omega, x_2(\omega)) - T(\omega, s_2(\omega))\| < \frac{\varepsilon}{8}
\end{eqnarray}
whenever $\| x_2(\omega) - s_2(\omega) \| < \delta_1(x_2(\omega))$. Now, choosing
\begin{eqnarray} \label{HardyRogerscontraction014}
\delta_1 = \min\Big\{\delta_1(x_1(\omega)), \frac{\varepsilon}{8}\Big\}
\end{eqnarray}
and
\begin{eqnarray} \label{HardyRogerscontraction015}
\delta_2 = \min\Big\{\delta_2(x_2(\omega)), \frac{\varepsilon}{8}\Big\},
\end{eqnarray}
by \eqref{HardyRogerscontraction011}, we have
\begin{eqnarray*}
&&\|T(\omega, x_1(\omega))-T(\omega, x_2(\omega))\|\nonumber\\
&\leq& \alpha_1(\omega)\|x_1(\omega)-x_2(\omega)\| + \alpha_2(\omega)\|x_1(\omega)- T(\omega, x_1(\omega))\| \\ \nonumber
&& + \alpha_3(\omega)\|x_2(\omega)- T(\omega, x_2(\omega))\| + \alpha_4(\omega)\|x_1(\omega)- T(\omega, x_2(\omega))\| \\ \nonumber
&& + \alpha_5(\omega)\|x_2(\omega)- T(\omega, x_1(\omega))\| + (1 + \alpha_2(\omega) + \alpha_5(\omega)) \frac{\varepsilon}{8} \\ \nonumber
&& + (1 + \alpha_3(\omega) + \alpha_4(\omega))\frac{\varepsilon}{8} + (\alpha_1(\omega) + \alpha_2(\omega) + \alpha_4(\omega)) \frac{\varepsilon}{8} \\ \nonumber
&& + (\alpha_1(\omega) + \alpha_2(\omega) + \alpha_4(\omega))\frac{\varepsilon}{8}
\end{eqnarray*}
and so
\begin{eqnarray*}
&&\|T(\omega, x_1(\omega))-T(\omega, x_2(\omega))\|\nonumber\\
&\leq& (2 + 2\sum^5_{i=1}\alpha_i(\omega))\frac{\varepsilon}{8} + \alpha_1(\omega)\|x_1(\omega)-x_2(\omega)\| \\ \nonumber
&& + \alpha_2(\omega)\|x_1(\omega)- T(\omega, x_1(\omega))\| + \alpha_3(\omega)\|x_2(\omega)- T(\omega, x_2(\omega))\| \\ \nonumber
&& + \alpha_4(\omega)\|x_1(\omega)- T(\omega, x_2(\omega))\| + \alpha_5(\omega)\|x_2(\omega)- T(\omega, x_1(\omega))\|. \\ \nonumber
\end{eqnarray*}
Since $\varepsilon > 0$ is arbitrary, it follows that
\begin{eqnarray} \label{HardyRogerscontraction015}
&&\|T(\omega, x_1(\omega))-T(\omega, x_2(\omega))\|\nonumber\\
&\leq& \alpha_1(\omega)\|x_1(\omega)-x_2(\omega)\| + \alpha_2(\omega)\|x_1(\omega)- T(\omega, x_1(\omega))\| \\ \nonumber
&& + \alpha_3(\omega)\|x_2(\omega)- T(\omega, x_2(\omega))\| + \alpha_4(\omega)\|x_1(\omega)- T(\omega, x_2(\omega))\| \\ \nonumber
&& + \alpha_5(\omega)\|x_2(\omega)- T(\omega, x_1(\omega))\|.
\end{eqnarray}
Thus we have $\omega \in \bigcap_{x_1, x_2 \in X}(C_{x_1, x_2} \cap A \cap B)$, which implies that
\begin{center}
$\bigcap_{s_1, s_2 \in S}(C_{s_1, s_2} \cap A \cap B) \subset \bigcap_{x_1, x_2 \in X}(C_{x_1, x_2} \cap A \cap B)$.
\end{center}
Also, we have
\begin{center}
$\bigcap_{x_1, x_2 \in X}(C_{x_1, x_2} \cap A \cap B) \subset \bigcap_{s_1, s_2 \in S}(C_{s_1, s_2} \cap A \cap B)$.
\end{center}
Therefore, we have
\begin{center}
$\bigcap_{s_1, s_2 \in S}(C_{s_1, s_2} \cap A \cap B) = \bigcap_{x_1, x_2 \in X}(C_{x_1, x_2} \cap A \cap B)$.
\end{center}
Let $N' = \bigcap_{s_1, s_2 \in S}(C_{s_1, s_2} \cap A \cap B)$. Then $\mu(N') = 1$, which implies that $T(\omega, x)$ is a deterministic mapping.
Hence $T$ has a unique random fixed point in $X$. This completes the proof.
\end{proof}
\vskip 2mm
If $\alpha_4(\omega) = \alpha_5(\omega) = 0$ in Theorem \ref{Maintheorem1}, then we obtain the following random fixed point theorem for Reich's contraction:
\vskip 2mm
\begin{cor} \label{RandomRiech}
Let $X$ be a separable Banach space and $(\Omega, \beta, \mu)$ be a complete probability measure space. Let $T : \Omega \times X \rightarrow X$ be a continuous random mapping
satisfying the following condition: for any $\omega \in \Omega$,
\begin{eqnarray*}
&& \|T(\omega, x_1(\omega))-T(\omega, x_2(\omega))\| \nonumber\\
&\leq& \alpha_1(\omega) \|x_1(\omega) - x_2(\omega)\| + \alpha_2(\omega) \|x_1(\omega) - T(\omega, x_1(\omega))\| \\ \nonumber
&& + \alpha_3(\omega) \|x_2(\omega) - T(\omega, x_2(\omega))\| \\ \nonumber
\end{eqnarray*}
for all random variables $x_1, x_2 : \Omega \rightarrow X$ and $\alpha_i : \Omega \rightarrow \mathbb{R}_+ \cup \{0\}$ for $i = 1, 2, 3$ such that $\sum^3_{i=1} \alpha_i(\omega) < 1$. Then there exists a unique random fixed point of $T$ in $X$.
\end{cor}
\vskip 2mm
If $\alpha_1(\omega) = \alpha_4(\omega) = \alpha_5(\omega) = 0$ in Theorem \ref{Maintheorem1}, then we obtain the following random fixed point theorem for Kannan's contraction:
\vskip 2mm
\begin{cor} \label{Kannan}
Let $X$ be a separable Banach space and $(\Omega, \beta, \mu)$ be a complete probability measure space. Let $T : \Omega \times X \rightarrow X$ be a continuous random mapping
satisfying the following condition: for any $\omega \in \Omega$,
\begin{eqnarray*}
&&\|T(\omega, x_1(\omega))-T(\omega, x_2(\omega))\|\nonumber\\
&\leq& \alpha_2(\omega) \|x_1(\omega) - T(\omega, x_1(\omega))\| + \alpha_3(\omega) \|x_2(\omega) - T(\omega, x_2(\omega))\| \\ \nonumber
\end{eqnarray*}
for all random variables $x_1, x_2 : \Omega \rightarrow X$ and $\alpha_i : \Omega \rightarrow \mathbb{R}_+ \cup \{0\}$ for $i = 2, 3$ such that $\alpha_2(\omega) + \alpha_3(\omega) < 1$. Then there exists a unique random fixed point of $T$ in $X$.
\end{cor}
\vskip 2mm
If $\alpha_1(\omega) = \alpha_2(\omega) = \alpha_3(\omega) = 0$ in Theorem \ref{Maintheorem1}, then we obtain the following random fixed point theorem for Chatterjea's contraction:
\vskip 2mm
\begin{cor} \label{Chatterjea}
Let $X$ be a separable Banach space and $(\Omega, \beta, \mu)$ be a complete probability measure space. Let $T : \Omega \times X \rightarrow X$ be a continuous random mapping
satisfying the following condition: for all $\omega \in \Omega$,
\begin{eqnarray*}
&&\|T(\omega, x_1(\omega))-T(\omega, x_2(\omega))\|\nonumber\\
&\leq& \alpha_4(\omega) \|x_1(\omega) - T(\omega, x_2(\omega))\| + \alpha_5(\omega) \|x_2(\omega) - T(\omega, x_1(\omega))\| \\ \nonumber
\end{eqnarray*}
for all random variables $x_1, x_2 : \Omega \rightarrow X$ and $\alpha_i : \Omega \rightarrow \mathbb{R}_+ \cup \{0\}$ for $i = 4, 5$ such that $\alpha_4(\omega) + \alpha_5(\omega) < 1$. Then there exists a unique random fixed point of $T$ in $X$.
\end{cor}
\vskip 2mm
\begin{rmk}
The random fixed point theorems for Hardy-Rogers's contraction reduced to the random fixed point theorems for $\acute{\text{C}}$iri$\acute{\text{c}}$'s contraction.
\end{rmk}
\vskip 4mm
\section{{\bf Applications to random nonlinear integral equations}}
\vskip 3mm
In this section, we give an application of Theorem \ref{Maintheorem1} to show the existence and uniqueness of a solution of a nonlinear stochastic integral equation of the Hammerstein type (\cite{Padgett}):
$$
\aligned
x(t;\omega)=h(t;\omega)+\int_Sk(t;s;\omega)f(s;x(s;\omega))d\mu(s),
\endaligned
\eqno{(4.1)}
$$
where
(a)\, $S$ is a locally compact metric space with metric d defined on $S\times S$
and $\mu_0$ is a complete $\sigma$-finite measure defined on the collection of Borel
subsets of $S;$
(b)\, $\omega\in \Omega$ where $\omega$ is the supporting set of the probability measure
space $(\Omega, \beta, \mu);$
(c)\, $x(t ; \omega)$ is the unknown vector-valued random variable for each $t \in S;$
(d)\, $h(t; \omega)$ is the stochastic free term defined for $t \in S;$
(e)\, $k(t, s; \omega)$ is the stochastic kernel defined for $t$ and $s$ in $S;$
(f)\, $f(t, x)$ is a vector-valued function of $t \in S$ and $x.$
\vskip 2mm
Note that the integral in the equation (4.1) is interpreted as a Bochner integral (\cite{Yosida}).
\vskip 2mm
Further, we assume that the union of a countable family $\{C_n\}$ of compact sets with $C_{n+1}\subset C_n$ is defined as $S$ such that, for each other compact set in $S$,
there exists $C_i$ which contains it (see \cite{Arens}).
\vskip 2mm
We define $C=C(S, L_2(\Omega, \beta, \mu))$ as a space of all continuous functions from $S$ into the space $L_2(\Omega, \beta, \mu)$ with the topology of uniform convergence on compact sets of $S$, that is, $x(t; \omega)$ is a vector-valued random variable for each fixed $t \in S$ such that
\begin{center}
$\| x(t; \omega) \|^2_{L_2(\Omega, \beta, \mu)} = \int_\Omega |x(t; \omega)|^2 d\mu(\omega) < \infty.$
\end{center}
Noted that $C(S, L_2(\Omega, \beta, \mu))$ is a space of locally convex (\cite{Yosida}) whose topology is defined by the countable family of semi-norms given by
\begin{center}
$\| x(t; \omega) \|_n = \sup_{t \in C_n} \| x(t; \omega) \|_{L_2(\Omega, \beta, \mu)}$
\end{center}
for each $n\geq 1$. Furthermore, since $L_2(\Omega, \beta, \mu)$ is complete, $C(S, L_2(\Omega, \beta, \mu))$ is complete relative to this topology.
\vskip 2mm
Next, we define $BC=BC(S, L_2(\Omega, \beta, \mu))$ as a Banach space of all bounded continuous functions from $S$ into $L_2(\Omega, \beta, \mu)$ with the norm
\begin{center}
$\| x(t; \omega) \|_{BC} = \sup_{t \in S} \| x(t; \omega) \|_{L_2(\Omega, \beta, \mu)}.$
\end{center}
The space $BC \subset C$ is a space of all second order vector-valued stochastic processes defined on $S$ which are bounded and continuous in mean-square.
\vskip 2mm
Now, we consider the functions $h(t; \omega)$ and $f(t, x(t; \omega))$ to be in the $C(S, L_2(\Omega, \beta, \mu))$ space with respect to the stochastic kernel and assume that,
for each pair $(t, s)$, $k(t, s; \omega) \in L_\infty(\Omega, \beta, \mu)$ and the norm denoted by
\begin{center}
$\|| k(t, s; \omega)|\| = \| k(t, s; \omega)\|_{L_\infty(\Omega, \beta, \mu)} = \mu-ess \sup_{\omega \in \Omega}| k(t, s; \omega) |.$
\end{center}
Also, we suppose that $k(t, s; \omega)\in L_\infty(\Omega, \beta, \mu)$ is such that
$$
\|| k(t, s; \omega)|\| = \| x(s; \omega)\|_{L_2(\Omega, \beta, \mu)}
$$
is $\mu$-integrable with respect to $s$ for each $t \in S$ and $x(s; \omega) \in C(S, L_2(\Omega, \beta, \mu))$ and there exists a real-valued function $G$ $\mu$-a.e. on $S$
such that $G(S) \| x(s; \omega) \|_{L_2(\Omega, \beta, \mu))}$ is $\mu$-integrable and, for each pair $(t, s) \in S \times S$,
\begin{equation*}
\|| k(t, u; \omega) - k(s, u; \omega)|\| \cdot \| x(u; \omega)\|_{L_2(\Omega, \beta, \mu)} \leq G(u)\| x(u; \omega)\|_{L_2(\Omega, \beta, \mu)} ~~~~ \mu-a.e.
\end{equation*}
Forward, assume that, for almost all $s \in S$, $k(t, s; \omega)$ is continuous in $t$ from $S$ into $L_\infty(\Omega, \beta, \mu)$.
\vskip 2mm
Now, we define the random integral operator $T$ on $C(S, L_2(\Omega, \beta, \mu))$ by
$$
\aligned
(Tx)(t; \omega) = \int_Sk(t, s; \omega)x(s; \omega)d\mu(s),
\endaligned
\eqno{(4.2)}
$$
where the integral is a Bochner integral. From the conditions on $k(t, s; \omega)$, it follows that, for each $t \in S$, $(Tx)(t; \omega) \in L_2(\Omega, \beta, \mu)$ and $(Tx)(t; \omega)$ is continuous in mean square by Lebesgue's dominated convergence theorem, that is, $(Tx)(t; \omega) \in C(S, L_2(\Omega, \beta, \mu))$.
\vskip 2mm
\begin{lem} {\rm (\cite{Padgett})}
The linear operator $T$ defined by the equation $(4.2)$ is continuous
from $C(S, L_2(\Omega, \beta, \mu))$ into itself.
\end{lem}
\vskip 2mm
\begin{proof}
See \cite{Padgett}.
\end{proof}
\vskip 2mm
\begin{dfn} (\cite{Achari}, \cite{LeePadgett})
Let $B$ and $D$ be Banach spaces. The pair $(B, D)$ is said to be \emph{admissible} with respect to a linear operator $T$ if $T(B) \subset D$.
\end{dfn}
\vskip 2mm
\begin{lem} {\rm (\cite{Padgett})}
If $T$ is a continuous linear operator from $C(S, L_2(\Omega, \beta, \mu))$ into itself and $B, D \subset C(S, L_2(\Omega, \beta, \mu))$ are Banach spaces stronger than $C(S, L_2(\Omega, \beta, \mu))$ such that $(B, D)$ is admissible with respect to $T$, then $T$ is continuous from $B$ into $D$.
\end{lem}
\vskip 2mm
By a {\it random solution} of the equation (4.1), we mean a function
$$
x(t; \omega) \in C(S, L_2(\Omega, \beta, \mu))
$$
which satisfies the equation (4.1) $\mu-a.e.$
\vskip 2mm
Now, by using Theorem \ref{Maintheorem1}, we prove the following:
\vskip 2mm
\begin{thm} \label{Maintheorem2}
If the stochastic integral equation $(4.1)$ is subject to the following conditions:
{\rm (1)}\, $B$ and $D$ are Banach spaces stronger than $C(S, L_2(\Omega, \beta, \mu))$ such that $(B,D)$ is admissible with respect to the integral operator defined by $(4.2)$;
{\rm (2)}\, $x(t; \omega) \mapsto f(t, x(t; \omega))$ is an operator from the set $Q(\rho) = \{ x(t; \omega) : x(t; \omega) \in D, \| x(t; \omega) \|_D \leq \rho \}$ into the space $B$ satisfying
$$
\aligned
&\quad\|f(t, x_1(t, \omega))-f(t, x_2(t, \omega))\|_B\\
&\leq \alpha_1(\omega) \|x_1(t, \omega) - x_2(t, \omega)\| + \alpha_2(\omega) \|x_1(t, \omega) - f(t, x_1(t, \omega))\| \\
&\quad + \alpha_3(\omega) \|x_2(t, \omega) - f(t, x_2(t, \omega))\| + \alpha_4(\omega) \|x_1(t, \omega) - f(t, x_2(t, \omega))\| \\
&\quad + \alpha_5(\omega) \|x_2(t, \omega) - f(t, x_1(t, \omega))\|
\endaligned
\eqno{(4.3)}
$$
for all $x_1(t, \omega), x_2(t, \omega)\in Q(\rho)$ and $\alpha_i : \Omega \rightarrow \mathbb{R}_+ \cup \{0\}$ for $i = 1, 2, 3, 4, 5$ such that $\sum^5_{i=1} \alpha_i(\omega) < 1$ almost surely;
{\rm (3)}\, $h(t; \omega) \in D$,
\vskip 1mm
\noindent
then there exists a unique random solution of the equation $(4.1)$ in $Q(\rho)$ provided
\begin{eqnarray*}
\|h(t, \omega)\|_D + l(\omega)\|f(t,0)\|_B\Big(\frac{1 + \alpha_3(\omega) + \alpha_4(\omega)}{1 - \alpha_2(\omega) + \alpha_5(\omega)}\Big)
\leq \rho\Big(1 - \frac{l(\omega)}{1 - \alpha_2(\omega) - \alpha_5(\omega)}\Big),
\end{eqnarray*}
where the norm of $T(\omega)$ denoted by $l(\omega)$.
\end{thm}
\vskip 2mm
\begin{proof}
Let a mapping $\mathcal{U}(\omega) : Q(\rho) \rightarrow D$ defined by
\begin{eqnarray*}
(\mathcal{U}x)(t, \omega) = h(t, \omega) + \int_S k(t, s, \omega)f(s, x(s, \omega))d_{\mu_0}(s).
\end{eqnarray*}
Then we have
\begin{eqnarray*}
\|(\mathcal{U}x)(t, \omega)\|_D &\leq& \|h(t, \omega)\|_D + l(\omega)\|f(t, x(t, \omega))\|_B \\ \nonumber
&\leq& \|h(t, \omega)\|_D + l(\omega)\|f(t, 0)\|_B + l(\omega)\|f(t, x(t, \omega)) - f(t, 0)\|_B.
\end{eqnarray*}
Thus it follows from (4.3) that
\begin{eqnarray*}
&& \|f(t, x(t, \omega))-f(t, 0)\|_B \nonumber\\
&\leq& \alpha_1(\omega) \|x(t, \omega)\|_D + \alpha_2(\omega) \|x(t, \omega) - f(t, x(t, \omega))\|_D \\ \nonumber
&& + \alpha_3(\omega) \|f(t, 0)\|_D + \alpha_4(\omega) \|x(t, \omega) - f(t, 0)\|_D + \alpha_5(\omega) \|f(t, x(t, \omega))\|_D \\
&\leq& \alpha_1(\omega) \|x(t, \omega)\|_D + \alpha_2(\omega) \|x(t, \omega)\|_D \\
&& + \alpha_2(\omega) \|f(t, x(t, \omega))-f(t, 0)\|_B + \alpha_2(\omega)\|f(t, 0)\|_D \\
&& + \alpha_3(\omega)\|f(t, 0)\|_D + \alpha_4(\omega) \|x(t, \omega)\|_D + \alpha_4(\omega) \|f(t, 0)\|_D \\
&& + \alpha_5(\omega) \|f(t, x(t, \omega))-f(t, 0)\|_B + \alpha_5(\omega) \|f(t, 0)\|_D
\end{eqnarray*}
and so
\begin{eqnarray*}
&& (1 - \alpha_2(\omega) - \alpha_5(\omega))\|f(t, x(t, \omega))-f(t, 0)\|_B\\
&\leq& (\alpha_1(\omega) + \alpha_2(\omega) + \alpha_4(\omega))\rho + (\alpha_2(\omega) + \alpha_3(\omega) + \alpha_4(\omega) + \alpha_5(\omega))\|f(t,0)\|_D.
\end{eqnarray*}
Hence we have
$$
\aligned
\|f(t, x(t, \omega))-f(t, 0)\|_B &\leq \Big(\frac{\alpha_1(\omega) + \alpha_2(\omega) + \alpha_4(\omega)}{1 - \alpha_2(\omega) - \alpha_5(\omega)}\Big)\rho \\
&\quad + \Big(\frac{\alpha_2(\omega) + \alpha_3(\omega) + \alpha_4(\omega) + \alpha_5(\omega)}{1 - \alpha_2(\omega) - \alpha_5(\omega)}\Big)\|f(t,0)\|_D.
\endaligned
\eqno{(4.4)}
$$
Therefore, by (4.4), we have
$$
\aligned
&\quad \|(\mathcal{U}x)(t, \omega)\|_D \\
&\leq \|h(t, \omega)\|_D + l(\omega)\|f(t, 0)\|_B + \Big(\frac{\alpha_1(\omega) + \alpha_2(\omega) + \alpha_4(\omega)}{1 - \alpha_2(\omega) - \alpha_5(\omega)}\Big)l(\omega)\rho \\
&\quad + \Big(\frac{\alpha_2(\omega) + \alpha_3(\omega) + \alpha_4(\omega) + \alpha_5(\omega)}{1 - \alpha_2(\omega) - \alpha_5(\omega)}\Big)l(\omega)\|f(t,0)\|_B \\
&\leq \|h(t, \omega)\|_D + \Big(\frac{\alpha_1(\omega) + \alpha_2(\omega) + \alpha_4(\omega)}{1 - \alpha_2(\omega) - \alpha_5(\omega)}\Big)l(\omega)\rho \\
&\quad + \Big(1 + \frac{\alpha_2(\omega) + \alpha_3(\omega) + \alpha_4(\omega) + \alpha_5(\omega)}{1 - \alpha_2(\omega) - \alpha_5(\omega)}\Big)l(\omega)\|f(t,0)\|_B \\
&\leq \|h(t, \omega)\|_D + \Big(\frac{\alpha_1(\omega) + \alpha_2(\omega) + \alpha_4(\omega)}{1 - \alpha_2(\omega) - \alpha_5(\omega)}\Big)l(\omega)\rho \\
&\quad + \Big(\frac{1 + \alpha_3(\omega) + \alpha_4(\omega)}{1 - \alpha_2(\omega) - \alpha_5(\omega)}\Big)l(\omega)\|f(t,0)\|_B \\
&< \rho
\endaligned
\eqno{(4.5)}
$$
and so, by (4.5), $(\mathcal{U}x)(t, \omega) \in Q(\rho)$.
Thus, for any $x_1(t, \omega)$, $x_2(t, \omega) \in Q(\rho)$ and, by the condition (2), we have
\begin{eqnarray*}
&& \|(\mathcal{U}x_1)(t, \omega) - (\mathcal{U}x_2)(t, \omega)\|_D\nonumber \\
&=& \Big\|\int_S k(t, s, \omega)[f(s, x_1(s, \omega)) - f(s, x_2(s, \omega))]d\mu_0(s)\Big\|_D \nonumber\\
&\leq& l(\omega)\|f(s, x_1(s, \omega)) - f(s, x_2(s, \omega))\|_B \nonumber\\
&\leq& \alpha_1(\omega)\|x_1(t, \omega) - x_2(t, \omega)\|_D + \alpha_2(\omega)\|x_1(t, \omega) - (\mathcal{U}x_1)(t, \omega)\|_D \nonumber\\
&& + \alpha_3(\omega)\|x_2(t, \omega) - (\mathcal{U}x_2)(t, \omega)\|_D + \alpha_4(\omega)\|x_1(t, \omega) - (\mathcal{U}x_2)(t, \omega)\|_D \nonumber\\
&& + \alpha_5(\omega)\|x_2(t, \omega) - (\mathcal{U}x_1)(t, \omega)\|_D.
\end{eqnarray*}
Consequently, $\mathcal{U}(\omega)$ is a random contractive mapping on $Q(\rho)$. Hence, by Theorem \ref{Maintheorem1}, there exists a random fixed point of $\mathcal{U}(\omega)$, which is the random solution of the equation (4.1). This completes the proof.
\end{proof}
\vskip 5mm
\textbf{Open Problem:} Can Theorems \ref{MainMotivate} and \ref{Maintheorem1} be generalized to non-separable Banach spaces?
\section*{{\bf Acknowledgements}}
The authors are gratefully thankful for referee's valuable comments, which significantly improve materials in this paper.
The first author was supported by Rajamangala University of Technology Lanna (RMUTL) for Ph.D. program at King Mongkut's University of Technology Thonburi (KMUTT). Yeol Je Cho was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Science, ICT and future Planning (2014R1A2A2A01002100).
This work was carried out while the third author (YJ. Cho) was visiting Theoretical and Computational Science Center (TaCS), Science Laboratory Building, Faculty of Science, King Mongkut's University of Technology Thonburi (KMUTT), Bangkok, Thailand, during 15 January-2 march, 2016. He thanks Professor Poom Kumam and the University for their hospitality and support.
Moreover, Poom Kumam was supported by the Thailand Research Fund (TRF) and the King Mongkut's University of Technology Thonburi (KMUTT) under the TRF Research Scholar Award (Grant No. RSA6080047).
\vskip 5mm
\footnotesize
| 64d665565753fb84006ba2450398b102df0be3ae | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:intro}
The normal assumption is the basis of statistical analyses in several
fields, such as medicine and health sciences. Indeed, under this
assumption, standard parametric estimation and testing procedures are
simple and most efficient. However, it is well known that these
procedures are not robust when the normal distribution is just an
approximate parametric model or in the presence of outliers in the
observed data. A challenging example, characterised by the possible presence of outliers,
is given by the study of \cite{tramentozzi2016} discussed in Section 2, which focuses on
the role of a protein in the production of immunoglobines and cytokines. In particular,
the study concerns repeated measures of moderate sample size, for which it is shown that the classical Bayesian analysis through the normal Linear Mixed Model (LMM) is strongly influenced by outliers. In this case, it may be preferable to base inference on procedures that are more resistant, which specifically take into account the fact that the assumed models is only approximate. In order to produce statistical procedures that are stable
with respect to small changes in the data or to small model departures,
robust statistical methods can be considered.
The concept of robustness has been widely discussed in the frequentist literature; see, for instance, \cite{hampel1986,tsou1995robust,lavine1995on} and \cite{markatou1998wieg}. On the other hand, Bayesian robustness literature is also well developed, most of which is focused on sensitivity with respect to the prior distribution; we mention here the overviews in \cite{berger1994overview}, \cite{rios2000robust} and \cite{rios2005robust}. Recently, Bayesian robustness with respect to model misspecification have attracted considerable attention. For instance, \cite{lazar03}, \cite{greco2008robust} and \cite{vcr2010} discuss approaches based on robust pseudo-likelihood functions, i.e.\ the quasi-likelihood and the empirical likelihood, derived from robust $M$-estimation functions, as replacement of the genuine likelihood in Bayes' formula. \cite{lewis2014bayesian} discuss an approach for building posterior distributions from robust $M$-estimators using constrained Markov chain Monte Carlo (MCMC) methods. Approaches based on tilted likelihoods can be found in \cite{miller2015robust}, \cite{watson2016approximate} and \cite{grunwald2014inconsistency}, among others. Finally, approaches based on model embedding through heavy-tailed distributions or on weighted likelihoods are discussed by \cite{andrade2006bayesian} and \cite{agostinelli2013}, respectively.
However, the aforementioned approaches are not without limitations. In particular, robust posterior distributions derived from robust estimating functions have two drawbacks: the empirical likelihood is not computable for very small sample sizes and for moderate sample sizes appears to have always heavy tails; posterior distributions based on the quasi-likelihood can be easily obtained only for scalar parameters. The restricted likelihood approach of \cite{lewis2014bayesian}, as well as the previous approaches based on estimating equations can be computationally cumbersome with robust $M$-estimating functions in the context of LMM (see Sect.~\ref{sec:application}). The solution of embedding in a larger structure has the cost of eliciting a prior distribution for the extra parameters introduced in the analysis. Moreover, the statistical procedures derived under an embedded model are not necessarily robust in a broad sense, since the larger model may still be too restricted in the space of all distributions. Finally, the tilted and the weighted likelihood approaches refer to concepts of robustness that are not directly related to the one considered in this paper.
Concretely, we focus on the robustness approach based on the influence function ($IF$), extensively discussed in book-length treatments by \cite{huber1981robust}, \cite{hampel1986} and \cite{huber2009robust}, and on the derivation of robust posterior distributions that condition on, generally insufficient, statistics with bounded $IF$. Such a conditioning is performed by appealing to the well-developed frequentist robust $M$-estimation theory, i.e. unbiased $M$-estimating equations with bounded $IF$, which provides estimators that target the parameter of interest and satisfy the usual asymptotical properties such as unbiasedness and normality. $M$-estimating functions are general unbiased estimating functions that include as special case the score function and scoring rules \citep[see, e.g., ][and references therein]{dawid2014theory}. To overcome the previous difficulties with current approaches to Bayesian robustness based on unbiased estimating equations, we propose an alternative method based on Approximate Bayesian Computation (ABC) \citep[see, e.g.,][]{pritchard1999population}. The typical motivation for the use of ABC is the intractability of the likelihood. Here, instead, we couple ABC with robust unbiased $M$-estimating functions. The idea can be seen as an extension of the results of \cite{ruli2016approximate} to Bayesian robustness with respect to misspecification. The method is easy to implement and computationally efficient, even when the $M$-estimating functions are cumbersome to evaluate for different parameter values. Theoretical properties, implementation details and simulations results are discussed.
Section 2 describes a motivating dataset on the glucose regulated protein94 (\texttt{GRP94}). Section 3 sets the necessary background. Section 4 describes the proposed method and its theoretical properties. Section 5 investigates the properties of the proposed method in the context of LMM through simulations and applies it to the \texttt{GRP94} dataset. Concluding remarks are given in Section 6.
\section{The \texttt{GRP94} dataset}
\label{sec:amotiv}
The \texttt{GRP94} dataset \citep{tramentozzi2016} concerns the measurement of glucose-regulated protein94 in plasma or other biological fluids and the study of its role as a tumour antigen, i.e. its ability to alter the production of immunoglobines (IgGs) and inflammatory cytokines in the peripheral blood mononuclear cells (PBMCs) of tumour patients. The study involved 28 patients admitted to the division of General Surgery of the Civil Hospital of Padova for ablation of primary, solid cancer of the gastro-intestinal tract. For each patient, gender, age (expressed in years), type and stage of tumour (ordinal scales of four levels) are given. Patient with ID 15 was removed from the study afterwards, for clinical reasons. Patients' plasma and PBMCs were challenged with GRP94 complexes and the level of IgG and of the cytokines: interferon$\gamma$ (IFN$\gamma$), interleukin 6 (IL-6), interleukin 10 (IL-10) and tumour necrosis factor $\alpha$ (TNF$\alpha$) were measured. Owing to time and cost constraints, for patients IDs 17, 27 and 28 only the IgG were measured. The following five treatments were considered: GRP94 at the dose of either 10 ng/ml or 100 ng/ml, GRP94 in complex with IgG (\texttt{GRP94+IgG}) at the doses 10 ng/ml or 100 ng/ml and IgG a the dose 100 ng/ml. Finally, baseline measurements of IgG and of the aforementioned cytokines were taken from untreated PMBCs. Although fresh patient's plasma and PMBCs are taken for each treatment and patient, the resulting measures are likely to be correlated since plasma and PMBCs are taken from the same patient. Hence, a LMM is a suitable model for this data. Using paired Mann-Whitney tests \cite{tramentozzi2016} showed that GRP94 in complex with IgG at the higher dose can significantly inhibit the production of IgG and stimulate the secretion of IL-6 and TNF$\alpha$ from PBMCs of cancer patients. In addition, some of the differences between treatments were significant for a specific gender; see \cite{tramentozzi2016} for full details.
Another feature of these data is the presence of extreme observations, both at baseline and challenged PMBCs-based measurements, as it can be seen from the strip plots in Figure \ref{fig:eda01}. Such extreme observations induce high variability on the response measurements, especially for IFN$\gamma$, IL-6, IL-10 and TNF$\alpha$. Hence, one must be cautious when fitting a LMM to such data; see Section~\ref{sec:application} for the complete analysis of the \texttt{GRP94} data.
\begin{figure}[t!]
\centering
\includegraphics[width=1.0\textwidth]{figures/eda01.eps}
\caption{Strip plots of IgG, IFN$\gamma$, IL-6, IL-10 and TNF$\alpha$ (in logarithmic scale) measured from PBMCs at baseline and after challenging with complexes of GRP94 and IgG. Values on the horizontal axis are (arbitrarily) ordered according to patient ID. Patient ID 5 was removed from the study and cytokines' measurements for patients with ID 17, 27 and 28 are missing.}
\label{fig:eda01}
\end{figure}
\section{Background}
\label{sec:background}
The present section provides an overview on concepts and methods that are needed to make the paper as much self-contained as possible. In particular, Section 3.1 gives background on robust $M$-estimation, Section 3.2 provides an overview on pseudo-likelihoods and Section 3.3 gives background on ABC.
\subsection{Robust $M$-Estimation}
\label{ssec:robustm}
Let $y = (y_1,\ldots , y_n)$ be a random sample of size $n$, having independent and identically distributed components, according to a distribution function $F_\theta =F(y;\theta)$, with $\theta \in \Theta \subseteq {\rm I}\negthinspace {\rm R}^d$, $d \geq 1$. Let $L(\theta;y)$ be the likelihood function based on model $F_\theta$.
Let $\Psi_\theta = \Psi(y;\theta)= \sum_{i=1}^n \psi(y_i;\theta)$ be an unbiased estimating function for $\theta$, i.e.\ such that $E_\theta (\psi(Y; \theta)) = 0$ for every $\theta$, with $\psi(\cdot)$ known function and with $E_\theta(\cdot)$ expectation with respect to $F_\theta$. Typically, $\Psi_\theta$ can be written as
\[
\Psi_\theta = b(\theta)^\T a(y, \theta) - c(\theta)\,,\label{eq:typicalPsi}
\]
where $a(\cdot,\cdot)$ and $c(\cdot)$ are known vector-valued functions and $b(\theta)$ may be a known vector- or matrix-valued function of suitable dimensions. The function $c(\cdot)$ is typically a consistency correction, i.e. it ensures that $E_\theta(\Psi_\theta)=0$. Most of the robust M-estimating functions can be recast in this form.
A general $M$-estimator (see, e.g., \citealp{hampel1986}, \citealp{huber2009robust}) is defined as the root $\tilde\theta$ of the estimating equation
$$
\Psi_\theta = 0
\ .
$$
The class of $M$-estimators is wide and includes a variety of well-known estimators. For example, it includes the maximum likelihood estimator (MLE), the maximum composite likelihood estimator \citep[see, e.g.,][and references therein]{ruli2016approximate} and the scoring rule estimator (see \citealp{basu1998robust}, \citealp{dawid2016minimum}, and references therein).
Under broad conditions assumed throughout this paper (see, e.g., \citealp{hampel1986}, \citealp{huber2009robust}), an $M$-estimator is consistent and approximately normal with mean $\theta$ and variance
\begin{eqnarray}
K (\theta) = H(\theta)^{-1} J(\theta) H(\theta)^{-\T}
\ ,
\label{var}
\end{eqnarray}
where $H(\theta) = -E_\theta (\partial \Psi_\theta/\partial \theta^{\T})$ and $J(\theta) = E_\theta(\Psi_\theta \Psi_\theta^{\T} )$ are the sensitivity and the variability matrices, respectively. The matrix $G(\theta)=K(\theta)^{-1}$ is known as the Godambe information and the form of $K(\theta)$ is due to the failure of the information identity since, in general, $H(\theta) \neq J(\theta)$.
The $IF$ of the estimator $\tilde\theta$ is given by
$$
IF (x;\tilde\theta,F_\theta) = H(\theta)^{-1} \psi (x;\theta)
\ ,
$$
and it measures the effect on the estimator $\tilde\theta$ of an infinitesimal contamination at the point $x$, standardised by the mass of the contamination. The supremum of the $IF$, i.e. the gross-error sensitivity, measures the worst influence on $\tilde\theta$ and a desirable robustness property for a statistical procedure is that the gross-error sensitivity is finite, i.e. that the $IF$ is bounded (B-robustness). Note that the $IF$ of the MLE is proportional to the score function; therefore, in general, MLE has unbounded $IF$, i.e. is not B-robust. On the contrary, if $\psi(x; \theta)$ is bounded, then the corresponding $M$-estimator $\tilde\theta$ is B-robust \citep{hampel1986}. Finally, note that the $IF$ can also be used to evaluate the asymptotic covariance matrix of $\tilde\theta$, since $K (\theta)=E_\theta (IF(x;\tilde\theta,F_\theta) IF(x;\tilde\theta,F_\theta)^{\T})$.
\subsection{Pseudo-likelihoods from unbiased estimating functions}
\label{ssec:pseudolik}
In general, there is not a unique function which has first derivative equal to $\Psi_\theta$, and many efforts have been made in order to derive pseudo-likelihood functions for $\theta$ based on $\Psi_\theta$. Pseudo-likelihoods are functions of the parameter $\theta$ and the data, with properties similar to a genuine likelihood function. Here we focus on three possible pseudo-likelihoods derived from $\Psi_\theta$.
A first pseudo-likelihood may be defined in terms of the conditional density of $\tilde{\theta}$ given $\theta$. In particular, assume that the joint density of $y$ is parametrized in terms of $\theta$ and a nuisance parameter $\lambda$ as
\[
f(y;\theta,\lambda) = g(\tilde{\theta};\theta)
h(y|\tilde{\theta};\theta,\lambda)\,.
\]
Then $g(\tilde\theta;\theta)$ is a pseudo-likelihood for $\theta$ and its use is justified provided the information contained in $h(y|\tilde{\theta};\theta,\lambda)$ about $\theta$, when $\lambda$ is unknown, is small or irrelevant \citep{davison1992bootstrap}.
In practice $g(\tilde{\theta};\theta)$ is unknown, but an approximation can be obtained via the bootstrap method (\citealp[see, e.g.,][]{davison1992bootstrap,boos1986bootstrap}, \citealp{rubin1981bayesian}), or by saddle-point approximations \citep[see, e.g.,][and references therein]{ronchetti1994empirical}. The restricted likelihood method of \cite{lewis2014bayesian} can be seen as a simulation-based approximation of $g(\tilde\theta;\theta)$.
Other pseudo-likelihoods, such as the quasi-likelihood \citep{mccullagh,adimari2002quasi,adimari2002quasim} and the empirical likelihood \citep{owen2001empirical} provide other means of building pseudo-likelihood function directly from $\Psi_\theta$.
The quasi-likelihood is defined as
$$
L_Q(\theta) = \exp \left( \int_k^\theta A(t) \Psi(y;t) \, dt \right)
\ ,
$$
where the matrix $A(\theta)$ is such that $A(\theta)^{\T} = J(\theta)^{-1} H(\theta)$ and $k$ is an arbitrary constant. When $d = 1$, $L_Q(\theta)$ is usually easy to derive, but when $d > 1$ the integrals is path-dependent. The adjustment $\Psi_\theta^a = A(\theta) \Psi(y;\theta)$ of $\Psi_\theta$ is necessary in order to recover the information identity, and thus the correct curvature (see, e.g., \citealp{pacesalvan97}, Chap.\ 4). Since $|A(\theta)| \neq 0$, $\Psi_\theta^a=0$ leads to the same solution $\tilde\theta$ of the original estimating equation $\Psi_\theta=0$ and the robustness properties of $\tilde\theta$ do not depend on $A(\theta)$ because this matrix does not change its $IF$.
The empirical likelihood $L_E(\theta)$ is defined through the empirical likelihood ratio statistic $W_E(\theta)=- 2 \log R (\theta)$, with $R (\theta)=\max_{p_i} \prod_{i=1}^n n p_i$, where the $p_i$ weights satisfy $p_i \geq 0$, $\sum_{i=1}^n p_i=1$ and $\sum_{i=1}^n \psi(y_i;\theta)p_i=0$. A Lagrangian argument leads to
$$
W_E (\theta) = 2 \sum_{i=1}^n \log ( 1+\eta^{\T} \psi(y_i;\theta) )\,,
$$
if $\theta=0$ is inside the convex hull of $\psi(y_1; \theta),\ldots , \psi(y_n; \theta)$; otherwise, it is adequate to set $W_E(\theta)=+\infty$. The Lagrangian multiplier $\eta = \eta(\theta)$ satisfies $\sum_{i=1}^n \psi(y_i ; \theta)/(1 + \eta^{\T} \psi(y_i ; \theta)) = 0$.
\cite{davison1992bootstrap} show that, under suitable regularity conditions, the bootstrap likelihood is asymptotically equivalent to first-order to the empirical likelihood $L_E(\theta)$. Moreover, under standard regularity conditions, it can also be shown that the quasi-likelihood ratio statistic $W_Q(\theta)$ and $W_E(\theta)$ are equivalent to the first term of their Taylor expansions. Lastly, \cite{monti1993relationship} investigate connections between $L_E(\theta)$ and saddlepoint pseudo-likelihoods in the case of $M$-estimating functions.
In the Bayesian framework, the use of $L_Q(\theta)$ in place of the proper likelihood has been discussed by \cite{greco2008robust} and \cite{vcr2010}, while the use of $L_E(\theta)$ has been discussed by \cite{lazar03}, though not for robustness purposes, and by \cite{greco2008robust}. Since $L_Q(\theta)$ and $L_E(\theta)$, as well as the bootstrap and the saddle-point likelihoods, share most of the properties of the genuine likelihood, they can be used as a replacement of the latter in the Bayes' formula \citep{ventura2016pseudo} which, in the case of robust estimating functions, leads to the robust posterior distribution
\begin{equation}
\pi_R(\theta|y) \propto \pi(\theta) \, L_R(\theta)
\ ,
\label{postR}
\end{equation}
where $\pi(\theta)$ is a prior distribution for $\theta$ and $L_R(\theta)$ is a pseudo-likelihood based on a robust $\Psi_\theta$, i.e. $L_Q(\theta)$, $L_E(\theta)$.
The approach based on robust posterior distributions (\ref{postR}) derived from robust $M$-estimating functions has two main drawbacks: the empirical likelihood is not computable for very small values of the sample size and for moderate sample sizes appears to have always heavy tails \citep[see, e.g.,][]{adimari2002quasi,greco2008robust}; the posterior distribution based on the quasi-likelihood can be easily obtained only for scalar parameters. Lastly, efficient use of bootstrap, empirical saddle-point and of restricted likelihood methods requires $M$-estimating functions that are easy to evaluate at different parameter values, which is generally not the case, especially in the context of this paper.
\subsection{Approximate Bayesian Computation}
\label{ssec:approximate}
Given a prior $\pi(\theta)$ and assuming that simulation from $F_\theta$, at a given $\theta$, is possible, the ABC method (\citealp{tavare1997inferring}, \citealp{pritchard1999population}) can provide an approximation of the posterior distribution $\pi(\theta|y)$, by means, for instance, of the accept-reject ABC algorithm (see Algorithm~\ref{alg:abc-ar} and \citealp{tavare1997inferring}). The latter
\begin{algorithm}
\caption[]{\label{alg:abc-ar} ABC accept-reject algorithm.}
\SetAlgoLined
\For {$i = 1 \to m$}{
\Repeat{$\rho(y, y^{*})\leq\epsilon$}{
draw $\theta^*\,\sim\,\pi(\cdot)$\;
draw $y^*\,\sim\, F_{\theta^*}$
}
set $\theta^{(i)}=\theta^{*}$
}
\end{algorithm}
samples from the joint distribution
$$
\pi_{\epsilon}(\theta,y^*|y) = \frac{ \pi(\theta)f(y^*;\theta)\mathbb{I}_{A_{\epsilon,y}(y^*)} }{ \int_{A_{\epsilon,y} \times\Theta }\pi(\theta)f(y^*;\theta)\,dy^* d\theta}
\ ,
$$
where $\mathbb{I}_{A_{\epsilon,y}(y^*)}$ is the indicator function of the set
$A_{\epsilon,y}(y^*) = \{y^*:\,\rho(y^*,y)\leq\epsilon\}$, $\rho(\cdot)$ is a given distance and $\varepsilon>0$ is a fixed tolerance level. The ABC method provides an approximation of the posterior distribution $\pi(\theta|y)$, given by
\[
\pi_{\epsilon}(\theta|y) = \int \pi_{\epsilon}(\theta,y^*|y)\,dy^*
\,.
\]
If $\epsilon\to 0$, then $\pi_{\epsilon}(\theta|y) \to \pi(\theta|y)$ \citep{blum2010approximate}. The threshold $\epsilon$ is generally set to the $\alpha$th quantile of the distance among the simulated and observed data, with $\alpha$ {being} typically very small \citep[see, e.g.,][]{beaumont2002approximate}.
In practice, the distance among the raw data can be rather noisy and a dimensional reduction of $y$ and $y^*$ is often necessary \citep[see, e.g,][]{beaumont2002approximate,fearnhead2012constructing}. This can be performed by considering a low-dimensional set of summary statistics $t(\cdot) = (t_1(\cdot), \ldots,t_q(\cdot))$, for some $q<n$, such as the mean, the median, quantiles etc. If $t(\cdot)$ is sufficient, then the ABC posterior $\pi_\epsilon(\theta|t(y))$ converges to the exact posterior as $\epsilon\to0$ \citep[see][Appendix A]{soubeyrand2013}. However, sufficient statistics are generally available only for models belonging to the exponential families.
The choice of the summary statistic is therefore a crucial point of ABC and this issue has been widely discussed in the statistical literature. In particular, \citet[][Sect. 3.1]{ruli2016approximate} show that, when a full computable likelihood is available, the standardised score function evaluated at the observed MLE $\hat\theta$, is an ideal summary statistic in ABC for a general model. This ideal summary statistic is given by
\begin{equation}
\eta(y;\hat\theta) = B(\hat\theta)^{-1} \ell_\theta(\hat\theta;y)
\ ,
\label{score}
\end{equation}
where $\ell_\theta(\theta;y)=\partial \log L(\theta;y)/\partial \theta$ is the score function and $B(\theta)$ is the square root of the expected information $i(\theta)$, i.e. is such that that $i(\theta)=B(\theta) B(\theta)^{\T}$. When considering the Mahalanobis distance we have
$$
\rho(\eta(y;\hat\theta),\eta(y^*;\hat\theta)) = \ell_\theta(\hat\theta;y^*)^{\T} i(\hat\theta)^{-1} \, \ell_\theta(\hat\theta;y^*)
\ ,
$$
which is the score test statistic computed at $\hat\theta$, based on data $y^*$.
In situations in which the likelihood function is intractable, and a composite likelihood is available, \citet[][Sect. 3.2]{ruli2016approximate} show that the corresponding (suitably standardised) composite score function can be used as a summary statistic in ABC. The resulting ABC procedure is shown to be invariant to reparameterisations and it automatically adjusts the curvature of the composite likelihood and of the corresponding posterior distribution. Lastly, note that in practice the ABC accept-reject algorithm is computationally inefficient and more advanced algorithms, such as ABC coupled with MCMC (ABC-MCMC) \citep{marjoram2003markov}, are preferred.
\section{Robust ABC Bayesian inference}
\label{sec:bayesian}
\subsection{Methodology}
\label{ssec:method}
One possibility to perform robust Bayesian inference is to resort to a pseudo-posterior distribution of the form \eqref{postR}. However, as discussed in Section \ref{ssec:pseudolik}, this approach has several limitations. A further computational limitation is that robust estimating equations $\Psi_\theta$ typically involve consistency corrections which are often intractable functions of $\theta$, defined as suitable expectations under the assumed model. See Section \ref{ssec:atoy} for a simple illustrative example.
Here we propose an alternative method for computing posterior distributions based on robust $M$-estimating functions. The method resorts to the ABC machinery in which a standardised version of $\Psi_\theta$ is used as summary statistic. The idea is an extension of \cite{ruli2016approximate} to the context of general unbiased estimating functions, though from a different perspective.
The rescaled $M$-estimating function is
\begin{equation}
\tilde\eta_R(y^{*}) = \eta_R(y^{*};\tilde\theta) = B_R(\tilde\theta)^{-1} \Psi(y^{*};\tilde\theta)
\ ,
\label{etar}
\end{equation}
where $B_R(\theta)$ is such that $J(\theta)=B_R(\theta) B_R(\theta)^{\T}$ and $\tilde\theta=\tilde\theta(y)$ is based on the observed sample $y$.
To generate posterior samples we propose to use the ABC algorithm with an MCMC kernel (see Algorithm~\ref{alg:abc-mcmc}), similar to Algorithm 2 of \cite{fearnhead2012constructing}; see also \cite{marjoram2003markov}. In Algorithm 2, $K_h(\cdot)$ denotes a density kernel with bandwidth $h$. The latter controls discrepancy between the observed and simulated summary statistics; in particular, the lower is $h$ the lower is the discrepancy.
\vspace{0.2cm}
\IncMargin{1em}
\begin{algorithm}[H]
\SetAlgoLined
\KwResult{A Markov dependent sample $(\theta^{(1)},\ldots,\theta^{(m)})$ from $\pi_{ABC}^R(\theta|\tilde\theta)$}
\KwData{a starting value $\theta^{(0)}$, a proposal density $q(\cdot|\cdot)$}
\For {$i = 1 \to m$}{
draw $\theta^*\,\sim\,q(\cdot|\theta^{(i-1)})$\\
draw $y^*\,\sim\, F_{\theta^*}$\\
draw $u\sim U(0,1)$\\
\eIf{u $\leq\frac{K_h(\tilde\eta_R(y^{*}))}{K_h(\tilde\eta_R(y^{(i-1)}))}\frac{\pi(\theta^{*})q(\theta^{(i-1)}|\theta^{*})}{\pi(\theta^{(i-1})q(\theta^{*}|\theta^{(i-1)})}$}{
set $(\theta^{(i)},\tilde\eta_R^{{(i)}})=(\theta^{*},\tilde{\eta}_R(y^{*}))$
}{
set $(\theta^{(i)},\tilde\eta_R^{(i)})\, =\, (\theta^{(i-1)},\tilde\eta_R(y^{(i-1)}))$
}
}
\caption[]{\label{alg:abc-mcmc} ABC-R algorithm with MCMC.}
\end{algorithm}
\DecMargin{1.em}
\vspace{0.2cm}
The proposed method via Algorithm~\ref{alg:abc-mcmc}
gives Markov-dependent samples from the ABC-R posterior
\begin{eqnarray}
\pi_{ABC}^R(\theta|\tilde\theta) = \frac{\int_{\mathcal{Y}^*} \pi(\theta) \, f(y^*;\theta)K_h(\tilde\eta_R(y^*))\,dy^*}{\int_{\mathcal{Y}^*\times\Theta} \pi(\theta) \, f(y^*;\theta)K_h(\tilde\eta_R(y^*))\,dy^*d\theta}
\ .
\label{postABCR}
\end{eqnarray}
Using similar arguments to \citet[][Appendix A]{soubeyrand2013}, it can be shown that, for $h \to 0$, $\pi_{ABC}^R(\theta|\tilde\theta)$ converges to $\pi(\theta|\tilde{\theta})$ pointwise \citep[see also][]{blum2010approximate}, in the sense that $\pi_{ABC}^R(\theta|\tilde\theta)$ and $\pi(\theta|\tilde{\theta})$
are equivalent for sufficiently small $h$. Since in general (\ref{etar}) is not sufficient, then $\pi(\theta|\tilde\theta)$ may differ from $\pi(\theta|y)$ and information is lost by using (\ref{etar}) instead of $y$. However such a loss of efficiency pays off in terms of robustness in the inference about $\theta$.
Posteriors conditional on partial information have been extensively discussed in the literature (\citealp[see, e.g.,][and reference therein]{doksum1990}). \cite{soubeyrand2015weak} study the properties of the ABC posterior when the summary statistic is the MLE or the pseudo-MLE derived from a simplified parametric model. However, they do not treat the case with the summary statistic given by an estimating function. Indeed, an alternative version of the ABC-R algorithm could be based on $\tilde{\theta}$ used as the summary statistic and a, possibly rescaled, distance among the observed and the simulated statistic as in \cite{soubeyrand2015weak}. Apparently, these two versions of ABC, namely the one based on $\tilde{\theta}$ and that based on (\ref{etar}) seem to be treated in the literature as two separate approaches (see, e.g., \citealp{gleim2013, drovandi2015}). However, both methods use essentially the same statistic $\tilde{\theta}$, but through different distance metrics. In addition, for small tolerances $\epsilon$ and $h$, these two distances converge to zero, and both methods give a posterior distribution conditional on the same statistic $\tilde{\theta}$.
Indeed, let $\tilde{\theta}$ be the summary statistic of the usual ABC posterior and let the tolerance threshold $\epsilon$ be sufficiently small and consider the random draw $\theta^{*}$ and its corresponding simulated summary statistics $\tilde\theta^{*}$ taken with the ABC algorithm. Then, by construction $\tilde\theta^{*}$ will be close to $\tilde{\theta}$.
This implies that also $\tilde\eta_R(y^*) = \eta_R(y^*;\tilde{\theta})$ will be close to $\eta_R(y^*;\tilde{\theta}^*)=0$, and hence $\theta^{*}$ is also a sample from the ABC-R posterior which uses the summary statistic $\tilde\eta_R$.
The use of $\tilde{\theta}$ as summary statistic in ABC requires the solution of $\Psi_\theta=0$ at each iteration, which is clearly computationally cumbersome. Therefore, this alternative approach will not be pursued further. Moreover, the proposed approach, besides sharing the same invariance properties stated by \cite{ruli2016approximate}, i.e. invariance with respect to both monotonic transformation of the data or with respect to reparametrisations, has the advantage of avoiding some computational problems related to the evaluation of $\Psi_\theta$ as shown by the following lemma.
\begin{lemma}
The ABC-R algorithm does not require repeated evaluations of the consistency correction involved $c(\theta)$ in $\Psi_\theta$ as given by \eqref{eq:typicalPsi}.
\end{lemma}
\begin{proof}[Proof.]
Let $\tilde{\theta}$ be the solution of $\Psi_\theta = 0$, with $\Psi_\theta$ of the form \eqref{eq:typicalPsi}. Then, for a given simulated $y^*$ at $\theta^*$, we have
\begin{align*}
\tilde\eta_R(y^*) & = B_R(\tilde{\theta})^{-1}(\Psi(y;\tilde\theta)-\Psi(y^*;\tilde\theta))\\ & = \,\,b(\tilde\theta)^\T\{a(y, \tilde\theta) - a(y^*, \tilde\theta)\}\,.
\end{align*}
This implies that $c(\theta)$, as also $b(\theta)$, is computed only once, at $\tilde{\theta}$.
\end{proof}
\subsection{Theoretical properties}
\label{ssec:theoretical}
Theorem~\ref{th:theo1} below shows that the proposed method gives a robust approximate posterior distribution with the correct curvature, even though $\Psi_\theta$, unlike the full score function, does not satisfy the information identity.
\begin{theorem}\label{th:theo1}
The ABC-R algorithm with rescaled $M$-estimating function $\eta_R(y)$, as $h \to 0$, leads to an approximate posterior distribution with the correct curvature and is also invariant to reparameterisations.
\end{theorem}
\begin{proof}
The proof follows from Theorem 3.2 of Ruli {\em et al.} (2016), by substituting the composite estimating equation with the more general $M$-estimating function $\Psi_\theta$.
\end{proof}
The ABC-R algorithm delivers thus a robust approximate posterior distribution which does not need calibration. This is not the case with (\ref{postR}) for which a calibration is typically required.
The following Theorem~\ref{th:theo2} shows that the proposed ABC approximate posterior distribution is asymptotically normal. In the theorem, the regularity assumptions of \cite{soubeyrand2015weak} are assumed, as well as the Euclidean distance in the ABC-R algorithm. Under these assumptions the density of $\tilde\theta$ is, for large $n$, equivalent to the normal density with mean vector equal to $\theta$ and covariance matrix equal to $K(\theta)$, and the density of $\Psi_\theta$ is, for large $n$, equivalent to the normal density with zero mean vector and covariance matrix equal to $J(\theta)$.
\begin{theorem}\label{th:theo2}
Consider the ABC-R algorithm that samples from the posterior $\pi_{R}^{ABC}(\theta|\tilde{\theta})$. Assume that when $h\to 0$, $\pi_{R}^{ABC}(\theta|\tilde\theta)$ converges pointwise to $\pi_R(\theta|\tilde\theta)$. Moreover, when $n \to \infty$ and $h \to 0$, the posterior $\pi_{R}^{ABC}(\theta|\tilde{\theta})$ is asymptotically equivalent to the density of the normal distribution with mean vector $\tilde\theta$ and covariance matrix $K(\tilde\theta)$:
\begin{eqnarray}
\pi_{R}^{ABC}(\theta|\tilde\theta) \, \dot{\sim} \, N_d (\tilde\theta,K(\tilde\theta))
\ , \quad \mathrm{for} \, \, h \to 0
\ .
\end{eqnarray}
\end{theorem}
\begin{proof}[Proof.]
The proof follows from Lemma 2 and Theorem 1 in \cite{soubeyrand2015weak} and from the asymptotic relation between the Wald-type statistic and the score-type statistic, i.e.
$$
\tilde\eta_R(y)^{\T} \, \tilde\eta_R(y) = \Psi_\theta^{\T} J(\theta)^{-1} \Psi_\theta =
(\tilde\theta - \theta)^{\T} K(\theta)^{-1} (\tilde\theta - \theta) + o_p(1)
\ .
$$
\end{proof}
In view of Theorem~\ref{th:theo2}, the resulting point estimates and credible sets converge to their frequentist analogues.
If $\psi(y;\theta)$ is bounded in $y$, i.e.\ if the estimator $\tilde\theta$ is B-robust, then the ABC-R posterior is resistant with respect to slight violations of the model assumptions. The advantage of the ABC-R posterior with respect to a robust posterior distribution of the form (\ref{postR}) is that the former allows to easily deal with multidimensional parameters, and thus avoids the difficulties of the quasi-likelihood. Moreover, it is computable also for very small sample sizes, provided $\tilde\theta$ is finite, and thus avoids the numerical instabilities of the empirical likelihood.
\begin{theorem}
If $\psi(y;\theta)$ is bounded in $y$, i.e.\ if the estimator $\tilde\theta$ is B-robust, then asymptotically the posterior mode, as well as posterior moments, of $\pi_{R}^{ABC}(\theta|\tilde{\theta})$ have bounded IF.
\end{theorem}
\begin{proof}[Proof.]
From Theorem~\ref{th:theo2}, the asymptotic posterior mode of $\pi_{R}^{ABC}(\theta|\tilde\theta)$ is $\tilde\theta$, which is $B$-robust. Moreover, following results in \cite{greco2008robust}, it can be shown that asymptotic posterior moments have bounded $IF$ if and only if the posterior mode has bounded $IF$.
\end{proof}
Notice that the ABC method can be seen as inherently robust since the resulting posterior conditions on a user-specified subset of the data instead of on the full data, as in the classical Bayesian setting. Nevertheless, through the ABC-R method we are able to obtain a B-robust posterior, that is a posterior having summaries with a bounded influence function.
\subsection{Implementation details}
\label{ssec:implementation}
Provided simulation from $F_\theta$ is fast, the main demanding requirement of the proposed method is essentially the computation of the observed $\tilde{\theta}$ and the scaling matrix $B_R(\theta)$ evaluated at $\tilde{\theta}$.
The ABC-R Algorithm~\ref{alg:abc-mcmc} involves a kernel density $K_h(\cdot)$, which is governed by the bandwidth $h$ and a proposal density $q(\cdot\vert\cdot)$. Given that, for large sample sizes,
$$
\tilde{\eta}_R (y) \sim N_d(0_d, I_d)\,,
$$
where $0_d$ is a $d$-vector of zeros and $I_d$ is the identity matrix of order $d$, it is reasonable to replace $K_h(\cdot)$ with the multivariate normal density centred at zero and with covariance matrix $hI_d$. To choose the bandwidth $h$ we consider several pilot runs of the ABC-R algorithm for a grid of $h$ values, and select the value of $h$ that delivers approximately 1\% acceptance ratio \citep[as done, for instance, by][]{fearnhead2012constructing}.
Contrary to other ABC-MCMC settings in which the proposal requires pilot runs \citep[see,][for building proposal distributions in ABC]{cabras2015}, in our case a scaling matrix for the proposal $q(\cdot|\cdot)$ can be readily build, almost effortlessly, by using the usual sandwich formula \eqref{var} \citep[see also][]{ruli2016approximate} evaluated at $\tilde\theta$. Even in cases in which $H(\theta)$ and $J(\theta)$ are not analytically available, they can be straightforwardly estimated via simulation. Indeed, in our experience, 100-500 samples from the model $F_\theta$, with $\theta =\tilde\theta$, give estimates with reasonably low Monte Carlo variability \citep[see also][]{cattelan2015empirical}. Throughout the examples considered we use the multivariate $t$-density with 5 degrees of freedom as the proposal density $q(\cdot|\cdot)$ and the ABC-R is always started from $\tilde{\theta}$. An \texttt{R} implementation of the proposed method for linear mixed effects models is also provided through the \texttt{robustBLME} package available on \texttt{CRAN}.
As a final remark, we note that robust estimating functions are available for many models of practical interest; see, among others, \cite{rousseeuw2003robust}, \cite{heritier2011}, \cite{huber2009robust}, \cite{farcomeni2012overview} and \cite{farcomeni2015robust} and references therein. Therefore, it is straightforward to exted the application of the proposed method in such model settings.
\subsection{An illustrative example}
\label{ssec:atoy}
We consider an example in which we compare the numerical properties of the ABC-R posterior, with the classical posterior based on the assumed model and the pseudo-posterior \eqref{postR} based on the empirical likelihood. Scenarios with data simulated either from the assumed model or from a slightly misspecified model are considered.
Let $F_\theta$ be a location-scale distribution with location $\mu$ and scale $\sigma>0$, and let $\theta = (\mu, \sigma)$. $M$-estimation through Huber's estimating function is a standard choice for the robust estimation of location and scale parameters. The $M$-estimating function is
\[
\Psi_\theta = (\Psi_{\mu}, \Psi_{\sigma})
\,,\]
with
\begin{eqnarray}
\Psi_\mu = \sum_{i=1}^n \psi_{c_1}(z_i)\,\quad\text{and}\quad
\Psi_\sigma = \sum_{i=1}^{n}\left(\psi_{c_2}(z_i)^2 - k(c_2)\right)
\ ,
\label{huber}
\end{eqnarray}
where $z_i = (y_i-\mu)/\sigma$, $i=1,\ldots,n$, $\psi_{c_1}(z) = \min(1;c_1/|z|)$ is the Huber $\psi$-function, $c_1>0$ and $c_2>0$ are scalar parameters which control the desired degree of robustness of $\tilde{\theta}$, and $k(\cdot)$ is a known function.
As an example let $F_\theta$ be the normal distribution $N(\mu, \sigma^2)$ and assume $\mu$ and $\sigma$ a priori independent with $\mu\sim N(0,10^2)$ and $\sigma\sim \text{halfCauchy}(5)$, where $\text{halfCauchy}(a)$ is the half Cauchy distribution with scale parameter equal to $a$. We consider random samples of sizes $n=\{15, 30\}$ drawn from either the normal distribution with $\theta=(0,1)$ and from a contaminated model. The latter is given by $F_\epsilon = (1-\epsilon)N(0,1) + \epsilon N(0, \sigma_1^2)$, with $\sigma_1^2>0$. For illustration purposes, we set the contamination to 10\%, i.e. $\epsilon=0.1$, and we set $\sigma_1^2=10$. Moreover, we assume $c_1=1.345$ and $c_2=2.07$, which imply that $\tilde{\mu}$ and $\tilde{\sigma}$ are, respectively, 5\% and 10\% less efficient than the corresponding MLE under the central (normal) model. In the simulation study we compute the genuine posterior distribution, e.g. the posterior based on the likelihood function of the normal model, the pseudo-posterior (\ref{postR}) based on the empirical likelihood (EL) and the ABC-R posterior.
The genuine and the EL posteriors are computed by numerical integration. The ABC-R posterior is obtained using Algorithm~\ref{alg:abc-mcmc} (Section~\ref{ssec:implementation}). From the posterior distributions illustrated in Figure~\ref{fig:fig1} we note that, when the data come from the central model (panels (a)-(b)), i.e. for $\epsilon=0$, all the posteriors are in reasonable agreement, even if the EL posterior behaves slightly worse, especially the marginal posterior of $\sigma$ with $n=15$. When the data are contaminated (panels (c)-(d)), the genuine posterior is less trustworthy as the bulk of the posterior drifts away from the true parameter value (vertical and horizontal straight lines). This is not the case however for the ABC-R posterior which remains centered around the true parameter value. The behavior of the EL posterior is less clear-cut here. In particular, it seems to have bimodal shape with the dominant mode laying around the true value and the secondary mode being located distantly. Consequently, the EL posterior has higher spread than the ABC-R posterior.
\begin{figure}[ht!]
\includegraphics[scale=0.65, angle=-90]{figures/full15}
\includegraphics[scale=0.65, angle=-90]{figures/full30}\\
\includegraphics[scale=0.65, angle=-90]{figures/cont15}
\includegraphics[scale=0.65, angle=-90]{figures/cont30}\\
\vspace{-0.2cm}
\caption{First row: genuine (continued), EL (blue dashed) and ABC-R posteriors (shaded image and histogram) for the normal model, when the data come from the central model $N(0,1)$ with (a) $n=15$ and (b) $n=30$. Second row: genuine, EL and ABC-R posteriors for the normal model, when the data come from the contaminated model with $\epsilon=0.1$, (c) $n=15$ and (d) and $n=30$.
\label{fig:fig1}
\vspace{-0.5cm}
\end{figure}
\vspace{1.2cm}
\noindent{\bf Sensitivity analysis}\\
To assess the robustness of the ABC-R posterior, we consider a sensitivity analysis. We simulate a sample $y$ of size $n=31$ from the central model and the contaminated data are given by $y^c = (y_{(j)},y_{(n/2)}+c)$, $j=1,\ldots,n/2-1,n/2+1,\ldots,n$, where $y_{(j)}$ denotes the $j$th ordered value of $y$ and $c$ a contamination scalar with possible values $\{-15,-14,\ldots,15\}.$ The results of the sensitivity analysis, illustrated by means of violin plots in Figure~\ref{fig:fig2}, highlight that the posterior median (the symbol "$\circ$") of the genuine posterior (panel (c)) is substantially driven by $c$. The $y$-axis is left free in order to enhance the of readability of the plot. On the other hand, the ABC-R and the EL posteriors are more resistant. For all the posteriors, the behavior of the posterior median reflects the behavior of the influence function of the posterior mode. Finally, the EL posterior under contamination is much more dispersed than the ABC-R posterior. Hence, even in an example as simple as this, robust inference based on the ABC-R posterior appears to be more precise than inference based on the EL posterior.
\begin{figure}[ht!]
\includegraphics[scale=0.8, angle=-90]{figures/sanalysisLS}
\vspace{-0.2cm}
\caption{Marginal ABC-R (a), EL (b) and genuine (c) posteriors for $\mu$ (left columns) and $\sigma$ (right) represented by means of violin plots for each $y^c$ generated under each value of $c=\{-15,-14,\ldots,15\}$. For each violin plot, the central circle represents the posterior median. The horizontal lines are the corresponding posterior median under $y^c$ with $c=0$.
\label{fig:fig2}
\vspace{-0.5cm}
\end{figure}
\clearpage
\section{Application to LMM}
\label{sec:application}
\subsection{Classical and Robust Estimation of LMM}
\label{ssec:classical}
The linear mixed model is a popular framework for analysing datasets in the context of hierarchical, longitudinal or repeated measures. A general formulation of LMM is by the following regression equation
\begin{equation}
y = X\alpha + \sum_{i= 1}^{c-1} Z_i\beta_i + \varepsilon\,,
\label{eq:model}
\end{equation}
where $y$ is a $n$-vector of the response observations; $X$ and $Z_i$ are known $n\times q$ and $n\times p_i$ design matrices respectively; $\alpha$ is a $q$-vector of unknown fixed effects; the $\beta_i$ are $p_i$-vectors of unobserved random effects $(1\leq i \leq c-1)$; and $\varepsilon$ is an $n$-vector of unobserved errors. The $p_i$ levels of each random effect $\beta_i$ are assumed to be independent with mean zero and variance $\sigma_i^2$. Moreover, each random error $\varepsilon_i$ is assumed to be independent with mean zero and variance $\sigma_c^2$ and $\beta_1,\ldots,\beta_{c-1}$ and $\varepsilon$ are assumed to be independent.
In this paper we focus on the classical normal LMM, which assumes that $\varepsilon \sim N_n(0_n,\sigma_c^2I_n)$ and $\beta_i \sim N(0, \sigma_i^2)$, $i=1, \dots, c-1$. Obviously, the validity and performance of this LMM
requires strict adherence to the model assumption, which is usually chosen
because it simplifies the analyses and not because it fits exactly the
data at hand. The robust procedure discussed in this paper specifically
takes into account the fact that the normal model is only approximate and
then it produces statistical analyses that are stable with respect to
outliers, deviations from the model or model misspecifications.
Under \eqref{eq:model} and the aforementioned assumptions, it follows that $Y$ is multivariate normal with $$E(Y) = X\alpha\quad \text{and} \quad \text{var}(Y) = V = \sum_{i = 1}^{c-1}\sigma_{i}^2Z_iZ_i^\T\,,$$ where $Z_{c} = I_n$. We assume that the set of $d = q + c$ unknown parameters $\theta = (\alpha, \sigma^2) = (\alpha, \sigma_{1}^2,\ldots,\sigma_{c}^2)$ is identifiable.
Although the $n$ observations $y$ are not independent, if the random effects are nested, then independent subgroups of observations can be found. Indeed, in many situations, $y$ can be split into $g$ independent groups of observations $y_j$, $j = 1, \ldots,g$. In this case the log-likelihood function is
\begin{equation}
\ell(\theta) = (-1/2)\sum_{j = 1}^g \left\{\log\vert V_j\vert + (y_j-X_j\alpha)^\T V_j^{-1}(y_j - X_j\alpha)\right\}\,,\label{eq:lmmlik}
\end{equation}
where $(y_1,\ldots,y_g)$, $X$ and $V$ are partitioned conformably.
Classical Bayesian inference for $\theta$ is based on the posterior distribution $\pi(\theta|y)\propto \exp\{\ell(\theta)\} \, \pi(\theta)$, where $\pi(\theta)$ is a prior distribution for $\theta$. However, $\ell(\theta)$ in the Gaussian LMM \eqref{eq:lmmlik} can be very sensitive to model deviations. This is because in the LMM framework a strict adherence to the multivariate normal model is required, and even a mild deviation from this model can have a great impact on the inferential results (\citealp{richardson1995robust}, \citealp{richardson1997bounded}, \citealp{copt2006high}); see also results of the simulation study in Section~\ref{ssec:simulation}.
There is a rich variety of robust inferential procedures for LMM in the presence of model misspecification: $M$-estimators \citep[see, e.g.,][and references therein]{richardson1995robust,richardson1997bounded, welsh1997approaches}, $S$-estimators \citep{copt2006high}, $MM$-estimators \citep{copt2006robust} and multivariate $t$ distributions \citep[see, e.g.,][]{lange1989robust}. The $M$-, $S$- and $MM$-estimation in the LMM framework provide robust estimators derived from estimating equations which are unbiased under the central model. In particular, $S$- and $MM$-estimators provide robust estimators with high breakdown point, but they are generally available only for balanced designs. Robustness and efficiency in $M$-, $S$- and $MM$-type estimators is guided by tuning constants which in practice are fixed by choosing the desired amount of robustness or, equivalently, by setting the desired loss of efficiency with respect to the MLE \citep[see also][]{basu1998robust}. The multivariate $t$ framework deals with robustness by replacing the normal model with the multivariate $t$, but at the cost of introducing the degrees of freedom as an additional unknown parameter.
Here we focus on robust inference in LMM through the $M$-estimation approach as pioneered by \cite{richardson1995robust}, \cite{richardson1997bounded} and \cite{welsh1997approaches} and extend it to the Bayesian framework through the ABC-R method. The aforementioned M-estimation approach is very general as it can deal with a wide variety of situations; for instance it can deal with unbalanced designs and robustness with respect to the design matrix \citep[see][for further details]{richardson1997bounded}.
Motivated by the \texttt{GRP94} dataset of Section~\ref{sec:amotiv}, in the present work we focus on robustness with respect to the response variable. Following \cite{welsh1997approaches} \citep[and][]{richardson1997bounded}, we focus on bounded of M-estimating functions of the form
\begin{dgroup}
\begin{dmath}
X^\T V^{-1/2}\psi_{c_1}\left(V^{-1/2}(y-X\alpha)\right)=0\,,
\label{eq:richalpha}\\
\end{dmath}
\begin{dmath}
(1/2)\left\{
\psi_{c_2}\left(V^{-1/2}(y-X\alpha)\right)^\T V^{-1/2} Z_iZ_i^\T V^{-1/2} \\
\times \psi_{c_2}\left(V^{-1/2}(y-X\alpha)\right)
- \text{tr}(CPZ_iZ_i^\T)\right\}=0 \condition{for $i = 1,\ldots,c$}
\label{eq:richsigma}
\end{dmath}
\end{dgroup}
where $C = E(\psi_{c_2}(R)\psi_{c_2}(R)^\T)$, with $R = V^{-1/2}(Y-X\alpha)$, $P = V^{-1} - V^{-1}X(X^\T V^{-1}X)^{-1}X^\T V^{-1}$ and $\text{tr}(\cdot)$ is the trace operator. The function $\text{tr}(CPZ_iZ_i)$ is a correction factor needed to ensure consistency at the Gaussian model for each $i=1,\ldots,c$.
\cite{welsh1997approaches} call \eqref{eq:richalpha}-\eqref{eq:richsigma} robust REML II estimating equations as they are bounded versions of restricted likelihood equations \citep{patterson1971recovery}. \cite{richardson1997bounded} shows that the estimator of $\theta$ given by the solution of \eqref{eq:richalpha}-\eqref{eq:richsigma} is asymptotically normal with mean the true parameter $\theta$ and variance matrix of the form \eqref{var}.
\subsection{Simulation study}
\label{ssec:simulation}
Let us consider the following two-component nested model
\begin{equation}
y_{ij} = \mu + \alpha_j + \beta_i + \varepsilon_{ij}\,,\label{eq:modsim}
\end{equation}
where $\mu$ is the grand mean; $\alpha_j$ are the fixed effects, constrained such that $\sum_{j=1}^q \alpha_j=0$; $\beta_i\sim N(0,\sigma_1^2)$ are the random effects and $\varepsilon_{ij}\sim N(0,\sigma_2^2)$ is the residual term, for $j = 1,\ldots,q$ and $i = 1,\ldots,g$. This model is a particular case of \eqref{eq:model}, with $c = 2$, a single random effect $\beta = \beta_1$ with $p_1 = g$ levels and $Z_1$ being the unit diagonal matrix. Furthermore, the covariate is a categorical variable with $q$ levels hence the design matrix is given by $q-1$ dummy variables.
We asses bias and efficiency of the proposed method via simulations with 500 Monte Carlo replications. For each Monte Carlo replication, the true values for the variance components $(\sigma_1^2,\sigma^2_2)$ and those of $\alpha$ are drawn uniformly in $(1,10)\times(1,10)$ and $(-5,5)$, respectively. With these parameter values, two datasets each of size $g$ are generated: one from the central model and one from the contaminated model $F_\epsilon = (1-\epsilon)N(X_i^\T\alpha,V_i) + \epsilon N(X_i^\T\alpha, 15V_i)$, with $\epsilon = 0.10$. We consider $q = \{3, 5, 7\}$ and $g = \{30, 50, 70\}$. Prior distributions are $\alpha\sim N_q(0,10^2I_q)$ and $(\sigma_1^2,\sigma^2_2) \sim \text{halfCauchy}(7)\times\text{halfCauchy}(7)$. For each scenario, we fit model \eqref{eq:modsim} in the classical Bayesian way, using an adaptive random walk Metropolis-Hastings algorithm. The same model is fitted by the ABC-R method using the estimating equations \eqref{eq:richalpha}-\eqref{eq:richsigma}. As in \cite{richardson1995robust}, we set $c_1 = 1.345$ and $c_2 = 2.07$. To find $\tilde{\theta}$, we solve \eqref{eq:richalpha}-\eqref{eq:richsigma} iteratively until convergence, as suggested by \cite{richardson1995robust}. The classical REML estimate, computed by the \texttt{lmer} function of the \texttt{lme4} package \citep{bates2015fitting}, is used as starting value. In our experiments, the convergence of the solution is quite rapid, i.e. $\tilde{\theta}$ stabilises within 10 to 15 iterations.
We assess the component-wise bias of the posterior median $\tilde\theta_m$ by the modulus of $\tilde{\theta}_m-\theta_0$ in logarithmic scale, where $\theta_0$ is the true value. Moreover, the efficiency of the classical Bayesian estimator relative to the ABC-R estimator is assessed through the index $MD_{MCMC}/MD_{ABC}$, where $MD = \text{med}(\tilde{\theta}_m-\theta_0)$. A similar index is used also by \cite{richardson1995robust} \citep[see also][]{copt2006high}. In addition, for each Monte Carlo replication we compute the Euclidean distance of $\tilde\theta_m$ from $\theta_0$, which can be considered as a global measure of bias. Contrary to \cite{richardson1995robust}, we consider a different $\theta_0$ for each Monte Carlo replication. For this reason, instead of just summarising the distribution of bias and efficiency by simple statistics we find it more informative to depict them graphically. The bias and efficiency of the classical Bayesian posterior and of the ABC-R posterior for the 500 replications are illustrated in Figures~\ref{fig:fig3} and \ref{fig:fig4}, respectively.
\begin{figure}[t!]
\centering
\includegraphics[width=1.\textwidth, height = 1.1\textwidth]{figures/bias.eps}
\caption{Bias of the ABC-R and classical (\texttt{MCMC}) Bayesian estimation of LMM under either the central (\texttt{Full}) or the contaminated model (\texttt{Mix}) for varying $n$ and $q$. Rows refer to a parameter or combination of parameters (row \texttt{all\_par}); columns within each cell refer to different vales of $q$; e.g. the last two rows (starting from top) have only two boxplots since $\alpha_6$ and $\alpha_7$ are available only with $q = 7$.}
\label{fig:fig3}
\end{figure}
\begin{figure}[ht!]
\centering
\includegraphics[width=1.\textwidth, height = 1.1\textwidth]{figures/efficiency.eps}
\caption{Efficiency of the ABC-R compared to the classical Bayesian estimation of LMM under the central (\texttt{Full}) and the contaminated models (\texttt{Mix}) for varying $n$ and $q$. Rows refer to a parameter and columns within each cell refer to different vales of $q$; e.g. the last two rows (starting from top) have only two boxplots since $\alpha_6$ and $\alpha_7$ are available only with $q = 7$.}
\label{fig:fig4}
\end{figure}
Under the central model, inference with the ABC-R and the classical Bayesian posteriors is roughly similar, i.e.\ both bias (component-wise and global) and efficiency compare equally well across the two methods. This holds for both the fixed effects $\alpha$ and for the variance components $(\sigma_1^2,\sigma^2_2)$. The latters present more bias than the formers. Under the contaminated model, we notice important differences among ABC-R and the classical Bayesian estimation. In particular, $\tilde{\theta}_m$ based on ABC-R is less biased, both globally and on a component by component basis, and more efficient. The gain in efficiency is particularly evident for the variance components.
\subsection{Effects of GRP94-based complexes on IgG and cytokines}
\label{sec:antitumour}
We fit the two-component nested LMM \eqref{eq:modsim} with ABC-R using estimating equations \eqref{eq:richalpha}-\eqref{eq:richsigma} to each of the five response variables: IgG, IFN$\gamma$, IL-6, IL-10 and TNF$\alpha$. Since all measures are positive and some of them are highly skewed, a logarithmic transformation is used in order to alleviate distributional skewness. Furthermore, since \cite{tramentozzi2016} highlight a possible gender effect (especially with respect to the cytokines) we check for gender effects by including an interaction with gender. In the multivariate form, the model with interaction can be written as
\begin{equation}
y_{i} = X_{i}^\T\alpha + X_i^\T\times\text{w}_i\gamma +\beta_i{1}_6 + \varepsilon_{i}\,,\quad i = 1,\ldots, g\label{eq:modinteract}
\end{equation}
where $X_i$ is the matrix of covariates for the $i$th unit, w$_i$ is a dummy variable for gender, $\gamma$ is the fixed effect of the treatment-gender interaction, ${1_6}$ is the unit vector of dimension 6. The number of statistical units $g$ is equal to 27 for the response IgG and 24 for the other response variables. Notice that the interaction model \eqref{eq:modinteract} has $12$ unknown fixed effects $(\alpha,\gamma)$.
As there is no extra-experimental information, we use relatively diffused priors. In particular, we assume $\alpha_j\sim N(0,100)$ and $\gamma_j\sim N(0,100)$, for $j = 1,\ldots,6$. For the variance components, following the recommendation of \cite{gelman2006prior}, we assume $\sigma_{1}^2\sim \text{halfCauchy}(7)$ and $\sigma_{2}^2\sim \text{halfCauchy}(7)$ in both models.
ABC-R posterior samples are drawn using Algorithm~\ref{alg:abc-mcmc} and following the implementation details of Section~\ref{ssec:implementation}. For comparison purposes, we fit also a classical Bayesian LMM with the aforementioned prior; an adaptive random walk Metropolis-Hastings algorithm is used for sampling from this posterior. Figure~\ref{fig:lmmfit} compares the ABC-R and the classical posterior for a subset of the fixed effects of models \eqref{eq:modsim} and \eqref{eq:modinteract} by means of kernel density estimations. The parameters shown are those referring to the treatments based on GRP94 at the dose of 10 ng/ml (\texttt{GRP94\_10}), GRP94 at the dose of 100 ng/ml (\texttt{GRP94\_100}) and GRP94 in complex with IgG at the dose of 100 ng/ml (\texttt{GRP94+IgG\_100}), which according to \cite{tramentozzi2016} are the most prominent. In Figure~\ref{fig:lmmfit}, panels (a1) to (e1) report the marginal posteriors of the fixed effects of \eqref{eq:modsim} (with \texttt{baseline} being the reference category), fitted to the five response variables, respectively; panels (a2) to (e2) give those of \eqref{eq:modinteract} (with \texttt{baseline} and \texttt{female} being the reference categories). Numbers within parenthesis in the plot sub titles give the evidence in favour of the null hypothesis H$_0$ that the parameter is equal to zero, computed under the Full Bayesian Significance Testing (FBST) setting of \cite{pereira2008can}; inside the parenthesis, the first (last) value from left refers to the ABC-R (classical) posterior.
The FBST in favour of $H_0$ has been
proposed by \citep{pereira1999evidence} as an intuitive measure of
evidence, defined as the posterior probability related to the less
probable points of the parametric space. It favours the null hypothesis
whenever it is large and it is based on a specific loss function and thus the
decision made under this procedure is the action that minimises the
corresponding posterior risk \citep[see][and references therein]{pereira2008can}. The FBST solves the drawback of the usual
Bayesian procedure for testing based on the Bayes factor (BF), that is, when the null hypothesis is precise and improper or vague priors are assumed, the BF can be undetermined and it can lead to the so-called Jeffreys-Lindleys paradox.
In panel (a1), the first plot from left reports the marginal distributions of the expected difference in terms of IgG between treatment \texttt{GRP94_10} and baseline. It shows that such a treatment is quite different from baseline measurements, because under the classical Bayesian LMM the evidence is zero to three decimal places, wheres under the ABC-R posterior the evidence is as low as 0.002. In what follows, evidences greater than 0.15 are treated as not worth mentioning and are therefore omitted. From the plots we can conclude the following.
\begin{figure}[ht!]
\includegraphics[width =1\columnwidth, height = 1\columnwidth]{figures/part1.eps}
\end{figure}
\begin{figure}[ht!]
\includegraphics[width =1\columnwidth,height=0.7\columnwidth]{figures/part2.eps}
\vspace{0.3cm}
\caption{Comparison of robust (ABC) and full (MCMC) posterior distributions of the fixed effects of the LMM \eqref{eq:modsim}-\eqref{eq:modinteract}, fitted to: $\log$ IgG (panels (a1)-(a2)), $\log$(INF$\gamma$+1) ((b1)-(b2)), $\log$ IL-6 ((c1)-(c2)), $\log$ IL-10 ((d1)-(d2)) and $\log$ TNF$\alpha$ ((e1)-(e2)). At each row, the left block refers to the posterior of the marginal effects of the treatments against the baseline and the right block refers to the interactions of the treatments with gender (with \texttt{baseline} and \texttt{female} being the reference categories). Numbers within refer to the evidence in favour of the null hypothesis H$_0$ that the parameter is equal to zero, computed under the Full Bayesian Significance Testing (FBST) setting of \cite{pereira2008can}; inside the parenthesis, the first (last) value from left refers to the ABC-R (classical) posterior. Dashed vertical lines corresponds to $\tilde{\theta}$.}
\label{fig:lmmfit}
\end{figure}
\clearpage
\noindent{\bf IgG.} There is a high posterior probability that the marginal treatments are different from the baseline, since the evidence of H$_0$ is rather low. However, the interaction with gender seem not to be supported. Although the density plots reveal some qualitative differences between the classical and the proposed robust Bayesian LMM, the evidences in favour of H$_0$ according to these posteriors, are rather low and in reasonable agreement, hence the message they convey is the same.\\
\noindent{\bf IFN$\gamma$.} No marginal nor interaction effects seem to be supported by the data for this variable according to both the classical and the proposed robust method.\\
\noindent{\bf IL-6.} The classical Bayesian LMM suggests that there is a marginal effect of \texttt{GRP94\_10} and of \texttt{GRP94\_100} as in both cases the evidence is low, i.e. less than 0.05. Furthermore, there is also evidence of an interaction effect of the latter with gender. However, on the basis of the ABC-R posterior, only \texttt{GRP94\_100} shows a marginal effect with respect to the baseline and no interaction effects are worth mentioning. These apparently contradictory results can be explained by the presence of few patients with extreme observations (see Figure~\ref{fig:eda01}) which seem to drive the classical LMM posterior but not the ABC-R posterior.\\
\noindent{\bf IL-10 and TNF$\alpha$.} For both these variables we notice both marginal and interaction effects of \texttt{GRP94\_100} under the classical Bayesian LMM. However, such effects disappear under the robust ABC-R procedure, which again signals that the classical LMM posterior is driven by few extreme observations (see Sect.~\ref{sec:amotiv}).
\section{Final remarks} %
\label{sec:final}
Former approaches on the use of robust unbiased estimating functions in Bayesian inference have been focused mainly on the quasi- and the empirical likelihoods as replacement of the full likelihood in Bayes' formula \citep[see, e.g.,][]{greco2008robust}. However, quasi-likelihood is available only for scalar parameters of interest while empirical likelihood can be unstable for small sample sizes.
In this respect, the present work offers a contribution by providing an alternative approach for building posterior distributions from robust estimating functions using simulated data from the central model and comparing them with the one actually observed, in an ABC fashion. Such a comparison is done through the chosen robust estimating function and this permits us to condition upon a robust subset of the data.
The method shares some limitations of the ABC setting. Firstly, a tolerance threshold has to be chosen. We fix this tolerance in order to give a pre-specified but small acceptance ratio, as frequently done in the ABC literature. Secondly, Bayesian hypothesis testing via BFs or posterior model probabilities is generally infeasible since the ABC model selection algorithm \citep{robert2011lack} requires ad hoc summary statistics \citep{marin2014relevant}, which are generally very difficult to find. The full Bayesian significance testing setting \citep{pereira2008can} is another practical and theoretically sound approach to Bayesian testing. Furthermore, it is easy to implement since for testing scalar parameters it requires only marginal posterior densities, which can be readily build by, say, kernel density estimation.
Motivated by the \texttt{GRP94} dataset, we considered only two-component nested LMM, but more complex models can be fitted since the estimating equations \eqref{eq:richalpha}-\eqref{eq:richsigma} are very general \citep[see][]{richardson1997bounded}. For instance, it is possible to deal with models with many random effects or even with robustness with respect to the design matrix. Moreover, the proposed ABC-R method works with any unbiased estimating equations, hence $S$- and $MM$-estimating functions can be readily used.
\section*{Acknowledgement}
This work was partially supported by University of Padova (Progetti di Ricerca di Ateneo 2015, \texttt{CPDA153257}) and by PRIN 2015 (grant \texttt{2015EASZFS\_003}).
| 2c3feaad83840be05be3a2fabaa27533f996d9a0 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Let $X$ be a 4-dimensional, oriented, smooth, Riemannian manifold and let $Q \rightarrow X$ be a ${\rm Spin}$-structure. A spinor bundle over $X$ is a vector bundle associated to $Q$, with typical fibre $\HH$. The idea for generalisation is to replace the spinor representation with a hyperK\"ahler manifold $(M, g_{\scriptscriptstyle M}, I_1, I_2, I_3)$ equipped with an isometric action of ${\rm Sp}(1)$ (or ${\rm SO}(3)$) which \emph{permutes} the complex structures on $M$. We will often refer to $M$ as the \emph{target hyperK\"ahler manifold}. The sections of the non-linear fibre-bundle now play the role of spinors. The interplay between the ${\rm Sp}(1)$ (or ${\rm SO}(3)$) action and the quaternionic structure on $M$ allows one to define the Clifford multiplication. Composing the Clifford multiplication with the covariant derivative gives the generalised Dirac operator, which we denote by $\Dd$.
In order to define a generalisation of the Seiberg-Witten equations, we need additionally a twisting principal $G$-bundle $P_G \rightarrow X$, with a tri-Hamiltonian action of $G$ on $M$. The action gives rise to a hyperK\"ahler moment map $\mu: M \longrightarrow \mf{sp}(1)^{\ast} \otimes \mf{g}^{\ast}$. For a connection $A$ on $P_G$ and a spinor $u$ the 4-dimensional generalised Seiberg-Witten equations on $X$ are the following system of equations
\eq{
\label{eq: intro gen. sw on X}
\left\{\begin{array}{l}
\Dd_{A} u = 0 \\
F^{+}_{A} - \mu \circ u = 0
\end{array}\right.
}
where $\Dd_{A}$ is a twisted Dirac operator for a connection $A$ on $P_G$.
This non-linear generalisation of the Dirac operator is well-known to physicists and has been used in the study of gauged, non-linear $\sigma$-models \cite{anselmi-fre}. The 3-dimensional version of equations \eqref{eq: intro gen. sw on X} was studied by Taubes \cite{taubes} (see also \cite{martin}). The 4-dimensional generalisation was considered by Pidstrygach \cite{victor}, Schumacher \cite{henrik} and Haydys \cite{haydys}. The moduli spaces of solutions to \eqref{eq: intro gen. sw on X} makes for an interesting study, especially because of its application to gauge theories on manifolds with special holonomies (cf. \cite{haydys15}, \cite{witten11}). Many well-known gauge-theoretic equations like the ${\rm PU}(2)$-monopole equations \cite{feehan-leness1998}, the Vafa-Witten equations \cite{vafa-witten1994}, ${\rm Pin}(2)$-monopole equations \cite{manolescu2016}, the non-Abelian monopole equations \cite{teleman00}, etc. can be treated as special cases of this generalisation.
It is possible to obtain the target hyperK\"ahler manifold with requisite symmetries from \emph{Swann's construction} \cite{swann}, \cite{boyergalickimann2}. Starting with a quaternionic K\"ahler manifold $N$ of positive scalar curvature, Swann constructs a fibration $\swann \rightarrow N$, whose total space admits a hyperK\"ahler structure. Such manifolds are characterised by the existence of a hyperK\"ahler potential. Alternatively, the permuting ${\rm Sp}(1)$-action extends to a homothetic action of $\HH^{\ast}$. The bundle construction commutes with the hyperK\"ahler quotient construction of Hitchin, Karlhede, Lindstr\"om and Ro\v cek \cite{hklr} and the quaternionic K\"ahler quotient construction of Galicki and Lawson \cite{galilaw}. As a result, many examples of (finite dimensional) hyperK\"ahler manifolds with homothetic $\HH^{\ast}$-action can be obtained via hyperK\"ahler reduction of $\HH^{n}$.
With $M = \swann$, we derive a transformation formula for the generalised Dirac operator, under the conformal change of metric on the base manifold. Since $\swann$ admits a natural homothetic action of ${\mathbb{R}}^+$, this setting allows one to make sense of ``weighted spinors".
Let $\pi_1: P_{{\rm CO}(4)} \rightarrow X$ be the bundle of conformal frames with respect to the conformal class $[g_{\scriptscriptstyle X}]$ and $P_G \rightarrow X$ be a principal $G$-bundle over $X$. Assume that the action of $G$ on $M$ is tri-Hamiltonian. Let $\widetilde{\pi}: \widetilde{Q} \rightarrow X$ denote the conformal ${\rm Spin}^G(4)$-bundle, which is a double cover of $P_{{\rm CO}(4)} \times_X P_G$.
\begin{thm} \label{thm: main thm 1}
Let $f$ be a smooth, real-valued function on $X$ and let $u$ be a (generalised) spinor. Consider the metric $g'_{\scriptscriptstyle X} := e^{2f}g_{\scriptscriptstyle X}$ in the conformal class $[g_{\scriptscriptstyle X}]$ and let $\varphi'$ and $\varphi$ be the Levi-Civita connections associated to $g_{\scriptscriptstyle X}$ and $g'_{\scriptscriptstyle X}$ respectively. For a fixed connection $A$ on $P_G$, denote by $A_{\varphi}$ and $A_{\varphi'}$ the corresponding lifts to $\widetilde{Q}$. Then, the associated generalised Dirac operators $\Dd_{A_{\varphi}}$ and $\Dd_{A_{\varphi'}}$ are related as
\eq{
\Dd_{A_{\varphi'}} (\scr{B}u) = \scr{B} \Big( de^{-5/2\pi_1^{\ast}f}\Dd_{A_{\varphi}}(e^{3/2\pi_1^{\ast}f}u) \Big)
}
where, $\scr{B}$ is the lift of the automorphism $ B: P_{{\rm CO}(4)} \longrightarrow P_{{\rm CO}(4)}$, given by $p \longmapsto e^{-f} p,$ and $de^{-5/2\pi_1^{\ast}f}$ is the action of $e^{-5/2\pi_1^{\ast}f}$ by differential on $TM$.
\end{thm}
For $M=\HH$, the result was proved by Hitchin \cite{hitchin1974}.
Assume that $M = \swann$ is a 4-dimensional hyperK\"ahler manifold. Using the above theorem, we show that away from a singular set, the generalised Seiberg-Witten equations can be interpreted in terms of almost-complex geometry of the underlying 4-manifold, as equations for a compatible almost-complex structure and a real-valued function which is associated to a conformal factor. Recall that on a Riemannian 4-manifold $(X, g_{\scriptscriptstyle X})$, the compatible almost-complex structures on $X$ are parametrized by sections of the twistor bundle ${\mathcal{Z}}$, which is a sphere bundle in $\Lambda^+$. Thus the almost-complex structures can be thought of as self-dual, 2-forms $\Omega$ with $\abs{\Omega} = 1$. An almost-complex structure gives a splitting of $\Lambda^+$ into the direct sum of the trivial bundle spanned by $\Omega$ and its orthogonal complement $\overline{K}$, where $K$ is a complex line bundle. Since $\abs{\Omega} = 1$, its covariant derivative is a section of $T^{\ast}X \otimes_{{\mathbb{R}}} \overline{K}$. Using the almost-complex structure, we get the isomorphism
\eqst{
T^{\ast}X \otimes_{{\mathbb{R}}} \overline{K} \cong T^{\ast}X \otimes_{{\mathbb{C}}} K \oplus T^{\ast}X \otimes_{{\mathbb{C}}} \overline{K}.
}
Moreover, the wedge product gives a complex, bi-linear map
\eqst{
T^{\ast}X \times T^{\ast}X \longrightarrow \Lambda^2 T^{\ast}X = K.
}
using which, we can identify $TX \cong T^{\ast}X \otimes_{{\mathbb{C}}} \overline{K}$.
Thus $\nabla \Omega$ has two components: the first component in $T^{\ast}X \otimes_{{\mathbb{C}}} K$ is the Nijenhuis tensor and the second one in $TX$ is $d\Omega$. Let $\langle \cdot, \cdot \rangle$ denote the obvious $\overline{K}$-valued pairing between $TX$ and $T^{\ast}X \otimes \overline{K}$.
Let $G = {\rm U}(1)$ and $M = \swann$ be 4-dimensional hyperK\"ahler manifold, which is total space of a Swann bundle, equipped with a tri-Hamiltonian action of ${\rm U}(1)$ that commutes with the permuting ${\rm Sp}(1)$-action. We will call such an action a permuting action of ${\rm U}(2) \cong {\rm Sp}(1) \times_{\pm} {\rm U}(1)$.
\begin{thm}
\label{thm: main thm 2}
Fix a metric $g_{\scriptscriptstyle X}$ on $X$ and let $[g_{\scriptscriptstyle X}]$ be its conformal class. Assume that $M$ is obtained as a quotient of a flat, quaternionic space and equipped with a residual permuting action of ${\rm U}(2)$ from the flat space. Then, there exists a 1-1 correspondence between the following:
\begin{itemize}
\item pairs consisting of a metric $g'_{\scriptscriptstyle X} \in [g_{\scriptscriptstyle X}]$ and a solution $(u, \ms{A})$ to the generalised Seiberg-Witten equations, such that the image of $u$ does not contain a fixed point of the ${\rm U}(1)$ action on $M$
\item pairs consisting of a metric $g''_{\scriptscriptstyle X} \in [g_{\scriptscriptstyle X}]$ and a self-dual 2-form $\Omega$ satisfying
\eq{
\label{eq: main thm 2}
(\nabla^{\ast}\nabla \Omega)^{\perp} + 2\pair{d\Omega, N_{\Omega}} = 0,~~~
\frac{3}{2}\abs{N_{\Omega}}^2 + \frac{1}{2}\abs{d\Omega}^2 + \frac{1}{2}\, s_{\scriptscriptstyle X}(g''_{\scriptscriptstyle X}) < 0
}
where $s_{\scriptscriptstyle X}(g''_{\scriptscriptstyle X})$ denotes the scalar curvature with respect to the metric $g''_{\scriptscriptstyle X}$.
\end{itemize}
\end{thm}
Theorem \ref{thm: main thm 2} was proved by Donaldson \cite{donaldson} for the usual Seiberg-Witten equations.
\smallskip
Notice that first equation in the second bullet of Theorem \ref{thm: main thm 2} is nothing but a perturbation of Euler-Lagrange equation of the energy functional
\eq{
\label{eq: energy functional twistor section}
\int_{X} \abs{\nabla\Omega}^2.
}
The functional was studied by Wood \cite{wood}. Critical points of the functional correspond to a choice of ``optimal" almost-complex structures, amongst all possible almost-complex structures on $X$.
\section{Acknowledgements}
The author wishes to thank Prof. Clifford Taubes for pointing out a crucial error in the earlier version of the article.
A major part of this article is based on the author's doctoral dissertation, during which he was financially supported by the DFG. The author thanks his supervisor Prof. V. Pidstrygach for his unwavering support and constant encouragement.
The author would also like to thank the anonymous referee for many helpful remarks and suggestions, especially on the second half of the article.
\section{Preliminaries and definitions}
\label{preliminaries}
\subsection{HyperK\"ahler manifolds}
A $4n$-dimensional Riemannian manifold $(M, g_{\scriptscriptstyle M})$ is \emph{hyperK\"ahler} if it admits a triple of almost-complex structures $I_i \in \End(TM) ~ i=1,2,3$ , which are covariantly constant with respect to the Levi-Civita connection and satisfy quaternionic relations $I_i I_j = \delta_{ijk} I_k$.
Let ${\rm Sp}(1)$ denote the group of unit quaternions and $\mf{sp}(1)$ denote its Lie algebra. The quaternionic structure on $M$ induces a covariantly constant endomorphism of $TM$ with values in $\mf{sp}(1)^{\ast} = \left(\mf{Im}(\HH)\right)^{\ast}$.
\eq{
\label{eq:algebra homomorphism}
I \in \Gamma(M, \End(TM) \otimes \mf{sp}(1)^{\ast}), ~~~ I_\xi := \xi_1 I_1 + \xi_2 I_2 + \xi_3 I_3, ~~ \xi \in \mf{sp}(1).
}
Observe that for every $\xi \in S^2 \subset \mf{Im}(\HH)$, the endomorphism $I_{\xi}$ is a complex structure. In other words, $M$ has an entire family of K\"ahler structures parametrized by $S^2$. Define the 2-form
\eqst{
\omega \in \Lambda^2 M \otimes \mf{sp}(1)^{\ast}, ~~ \omega_{\xi}(\cdot, \cdot) = g_{\scriptscriptstyle M}(I_{\xi}(\cdot), \cdot).
}
If $\xi \in S^2$, then $\omega_{\xi}$ is just the K\"ahler 2-form associated to $I_{\xi}$.
\begin{defn}
\label{def: permuting action}
An isometric action of ${\rm Sp}(1)$ on $M$ is said to be \emph{permuting} if the induced action on the 2-sphere of complex structures is the standard action of ${\rm SO}(3) = {\rm Sp}(1)/\pm 1$ on $S^2$:
\eqst{
dq ~ I_{\xi} ~ dq^{-1} = I_{ q\xi \bar{q}}, ~~ \text{for} ~~ q \in {\rm Sp}(1),~~ \xi \in \mf{sp}(1), ~~ \norm{\xi}^2 = 1
}
\end{defn}
\begin{defn}
\label{def: tri-holomorphic action}
An isometric action of a Lie group $G$ on $M$ is \emph{tri-holomorphic} or \emph{hyperK\"ahler}, if it preserves the hyperK\"ahler structure
\eqst{
\eta_{\ast} I_{i} = I_{i} \eta_{\ast} ~~ i=1,2,3, ~~ \eta \in G.
}
In particular, $G$ fixes the 2-sphere of complex structures on $M$. The action is \emph{tri-Hamiltonian} (or hyperHamiltonian) if it is Hamiltonian with respect to each $\omega_i$. The three moment maps can be combined together to define a single, $G$-equivariant map \emph{hyperK\"ahler moment map} $\mu: M \longrightarrow \mf{sp}(1)^{\ast}\otimes \mf{g}^{\ast}$, which satisfies
\eqst{
d(\langle \mu, \xi_{i}\otimes\eta) = \iota_{\scriptscriptstyle K^M_{\eta}}\,\omega_i, ~~~ \eta \in \mf{g}, ~~ \xi_i \in \mf{sp}(1) ~~\text{is the basis}
}
and $K^M_{\eta}$ denotes the fundamental vector-field due to the infinitesimal action of $\eta$.
\end{defn}
\begin{defn}
\label{def: hyperkahler potential}
A \emph{hyperK\"ahler potential} is a smooth function $f: M \longrightarrow {\mathbb{R}}^+$ which is simultaneously a K\"ahler potential for all the three complex structures $I_1, I_2, I_3$.
\end{defn}
\subsection{Target hyperK\"ahler manifold}
Suppose that $M$ is a hyperK\"ahler manifold with a permuting action of ${\rm Sp}(1)$ and a tri-Hamiltonian action of a compact Lie group $G$ which commutes with the ${\rm Sp}(1)$-action. Let $\varepsilon \in G$ be a central element of order two. Let $\Z/2\Z \subset {\rm Sp}(1) \times G$ denote the normal subgroup of order two, generated by the element $(-1, \varepsilon)$. Assume that $\Z/2\Z$ acts trivially on $M$ so that the action of ${\rm Sp}(1) \times G$ descends to an action of ${\rm Spin}^G(3) \, := \, {\rm Sp}(1)\times_{\Z/2\Z} G$. We will refer to this action as a permuting action of ${\rm Spin}^G(3)$. An action of ${\rm Spin}^G(4):= ({\rm Sp}(1)_+ \times {\rm Sp}(1)_-) \times_{\Z/2\Z} G$ is said to be permuting if the action is induced by a permuting action of ${\rm Sp}(1) \cong {\rm Spin}(3)$ via the homomorphism
\eqst{
\rho: {\rm Spin}^G(4) \longrightarrow {\rm Spin}^G(4)/{\rm Sp}(1)_- \cong {\rm Spin}^G(3).
}
Note that ${\rm Sp}(1)_-$ acts trivially on $M$.
\subsection{\texorpdfstring{${\rm Spin}^G(4)-$}~ structure}
From the definition of the group ${\rm Spin}^G(4)$, we have the following exact sequence
\eq{
\label{eq: short exact seq sping str}
0 \longrightarrow \Z/2\Z \longrightarrow {\rm Spin}^G(4)\overset{\gamma}{\longrightarrow} {\rm SO}(4) \times \left(G/\{1, \epsilon \} \right) \longrightarrow 0.
}
For simplicity, put $\overline{G} = G/\{1, \epsilon \}$. Let $P_{{\rm SO}(4)}$ denote the frame-bundle of $X$ and $P \rightarrow X$ be a principal $\overline{G}$-bundle over $X$. A ${\rm Spin}^G(4)$-structure over $X$ is a principal ${\rm Spin}^G(4)$-bundle $\pi: Q \rightarrow X$, which is an equivariant double cover of the bundle $P_{{\rm SO}(4)} \times_X P$, with respect to the map $\gamma$ as defined in \eqref{eq: short exact seq sping str}. We refer to \cite{teleman00} for details.
\subsection{Generalised Dirac operator}
\label{subsec:generalised dirac operator}
We define the space of \emph{generalised spinors} to be the space of smooth, equivariant maps
\eqst{
\Ss \, \, := \, \, C^{\infty}(Q, M)^{{\rm Spin}^G(4)} \, \, \cong \, \, \Gamma(X, Q \times_{{\rm Spin}^G(4)} M).
}
The Levi-Civita connection $\varphi$ on $P_{{\rm SO}(4)}$ and a connection $a$ on the principal $P$ together determine a unique connection on $Q$. Let $\scr{A}$ denote the space of all connections on $Q$, which are the lifts of the Levi-Civita connection. We define the covariant derivative of a spinor $u \in \Ss$, with respect to a connection $\ms{A} \in \scr{A}$ by\footnote{The subscript \emph{hor} implies that $D_{\ms{A}}u$ vanishes on vertical vector fields.}
\eq{
\label{covariant derivative}
D_{\ms{A}} : C^{\infty}(Q,M)^{{\rm Spin}^G(4)} \longrightarrow \Hom(TQ, TM)^{{\rm Spin}^G(4)}_{hor} , ~~ D_{\ms{A}} u = du + K^M_{\ms{A}}|_{u}
}
where $K^M_{\ms{A}}|_{u}: TQ \rightarrow u^{\ast}TM$ is an equivariant bundle homomorphism defined by $K^M_{\ms{A}}|_{u} (v) = K^M_{\ms{A}(v)}|_{u(p)}$ for $v\in T_p Q$. Denote by $\pi_{\scriptscriptstyle {\rm SO}}: Q \longrightarrow P_{{\rm SO}(4)}$ the projection to the frame bundle. Then, alternatively, one can view the covariant derivative as
\al{
\label{eq:covariant derivative alt. def.}
D_{\ms{A}}:~ C^{\infty}(Q,M)^{{\rm Spin}^G(4)} \longrightarrow C^{\infty}(Q, ({\mathbb{R}}^{4})^{*} \otimes TM)^{{\rm Spin}^G(4)}, ~~~ \langle D_{\ms{A}}u(q), w \rangle = du(q)(\widetilde{w})
}
where, $w \in {\mathbb{R}}^{4}$, $\widetilde{w}$ denotes the horizontal lift of $\pi_{\scriptscriptstyle {\rm SO}}(q)(w) \in T_{\pi(q)}X$.
\subsection*{Clifford multiplication}
The second ingredient we need to define the Dirac operator is Clifford multiplication. From \eqref{eq:algebra homomorphism}, we an construct an action of $\mc{C}l^0_4 \cong \mc{C}l_3$ on $TM$ as
\eqst{
{\mathbb{R}}^3\cong \mf{Im}(\HH) \longrightarrow \End(TM), ~~ h \mapsto I_{h}.
}
The map extends to a ${\rm Spin}^G(4)$-equivariant map $\mc{C}l_3 \longrightarrow \End(TM)$. Thus $TM$ is naturally a $\mc{C}l^0_4 $ module. Now consider $W:= \mc{C}l_4 \otimes_{\mc{C}l^0_4} E$, where $E = (TM, I_1)$. Since $W$ is a $\mc{C}l^0_4$-module, we get a $\Z_2$-graded $\mc{C}l_4$-module
\eqst{
W = W^+ \oplus W^-, ~~ W^+ = \mc{C}l^0_4 \otimes_{\mc{C}l^0_4} E, ~~ W^- = \mc{C}l^1_4 \otimes_{\mc{C}l^0_4} E.
}
More precisely, $W^+$ is the ${\rm Spin}^G(4)$-equivariant bundle $TM$ with an action induced by $\rho$, whereas $W^-$ is the ${\rm Spin}^G(4)$-equivariant vector bundle $TM$ equipped the left-action: \[ [q_{+}, q_{-}, g] \cdot w_{-} = I_{q_{-}} I_{\bar{q}_{+}} dq_{+} dg ~ w_{-}.\]
Identify ${\mathbb{R}}^{4}$ with $\HH$ by mapping the standard, oriented basis $(e_{1},e_{2},e_{3},e_{4})$ of ${\mathbb{R}}^{4}$, to $(1, \bar{i}, \bar{j}, \bar{k})$. The ${\rm Spin}^G(4)$-action on $\HH$ is given by $[q_{+}, q_{-}, g] \cdot h = q_{-}h \bar{q}_{+}$. Clifford multiplication is the ${\rm Spin}^G(4)$-equivariant map
\eq{
\label{eq:clifford multiplication}
\bullet: ({\mathbb{R}}^{4})^{\ast} \cong \HH \longrightarrow \End(W^+ \oplus W^-), ~~
g_{{\mathbb{R}}^{4}}(h, \cdot) \longmapsto \begin{bmatrix} 0 & -I_{\bar{h}} \\ I_{h} & 0
\end{bmatrix}.
}
Since $ h \bullet h = -g_{{\mathbb{R}}^{4}}(h,h)\cdot \id_{W^+ \oplus W^-}$, by universality property, the map $\bullet$ extends to a map of algebras $\bullet: \mc{C}l_{4} \longrightarrow \End(W^+ \oplus W^-)$. Composing $\bullet$ with the covariant derivative, we get the \emph{generalised Dirac operator}:
\eq{
\label{eq:gen. dirac operator explicit}
\Dd_{\ms{A}} u \in C^{\infty}(Q, u^{*}W^-)^{{\rm Spin}^G(4)}, ~~~~ \Dd_{\ms{A}}u = \sum_{i=0}^{3} e_{i} \bullet D_{\ms{A}}u(\tilde{e_{i}})
}
where the latter expression follows from equation \eqref{eq:covariant derivative alt. def.}.
\subsection*{Generalised Seiberg-Witten equations}
Let $\mu$ be a hyperK\"ahler moment map for the $G$-action on $M$ and $a$ be a connection on $P$. Then \emph{generalised Seiberg-Witten equations} for a pair $(u, \ms{A}) \in \Ss \times \scr{A}$, in dimension four, are
\begin{equation}
\label{eq:gen. seiberg-witten}
\left\{\begin{array}{l}
\Dd_{\ms{A}} u = 0 \\
F^{+}_{a} - \Phi(\mu \circ u) = 0
\end{array}\right.
\end{equation}
where $F^+_{a} \in \Map(Q, \Lambda^2_+({\mathbb{R}}^4)^{\ast})^{{\rm Spin}^G(4)}$ is the self-dual part of the curvature of $a$ and $\Phi: \mf{sp}(1)^{\ast} \longrightarrow \Lambda^2_+ ({\mathbb{R}}^4)^{\ast}$ is the isomorphism mapping the basis elements $\xi_l \mapsto \beta_l, \, l = 1,2,3$, where
\eq{
\label{eq: basis of self-dual 2-forms}
\beta_{0} = dx_0 \wedge dx_1 + dx_2 \wedge dx_3, ~ \beta_1 = dx_0 \wedge dx_2 + dx_3 \wedge dx_1, ~ \beta_3 = dx_0 \wedge dx_3 + dx_1 \wedge dx_2.
}
We will supress the isomorphism henceforth.
\section{Conformal transformation of generalised Dirac operator} \label{conformalproperty}
This section is divided into three parts. In the first part, subsection \ref{subsec: metric connections}, we study metric connections for metrics in the conformal class of $g_{\scriptscriptstyle X}$. Namely, given the Levi-Civita connection of $g_{\scriptscriptstyle X}$ and a metric $g'_{\scriptscriptstyle X} \in [g_{\scriptscriptstyle X}]$, we explicitly construct the Levi-Civita connection for $g'_{\scriptscriptstyle X}$. In the second part, subsection \ref{subsec: example 2 Swann bundles}, we give a quick review of Swann's construction. In the third part, subsection \ref{subsec: gen dirac and conf. metric}, we use the results from subsection \ref{subsec: metric connections} to obtain a formula for conformal transformation of the generalised Dirac operator when the target hyperK\"ahler manifold obtained via Swann's construction. For details on ideas used in this section, we refer the interested reader to \cite{salamon}.
\subsection{Metric connections on conformal bundle}
\label{subsec: metric connections}
\noindent Fix a metric $g_{\scriptscriptstyle X}$ on $X$ and let $[g_{\scriptscriptstyle X}]$ denote its conformal class. Let $\pi_{1}: P_{{\rm CO}(4)} \longrightarrow X$ denote the bundle of all conformal frames on $(X,[g_{\scriptscriptstyle X}])$. A point $p\in P_{{\rm CO }(4)}$ is a ${\rm CO}(4)$-equivariant, linear isomorphism $p: {\mathbb{R}}^4 \longrightarrow T_{\pi_1(p)}X$. Consider the canonical one-form $\theta: P_{{\rm CO}(4)} \longrightarrow {\mathbb{R}}^4$ defined as
\eqst{
\theta_{p}(v) = p^{-1}\left((\pi_{1})_{\ast}(v)\right), ~~ p \in P_{{\rm CO}(4)}, ~~v \in T_{p}P_{{\rm CO}(4)}.
}
A metric on $X$ is a section $g_{\scriptscriptstyle X} \in \Gamma(X, S^{2}(T^{*}X))$, which can viewed as an equivariant map in $C^{\infty}(P_{{\rm CO}(4)}, S^2({\mathbb{R}}^4)^{\ast})^{{\rm CO}(4)}$
\eqst{
\pi_{1}^{\ast}g_{\scriptscriptstyle X}\left(\cdot, \cdot \right) = g_{{\mathbb{R}}^{4}}\left(\theta_{p}(\cdot), \theta_{p}(\cdot)\right) .
}
For a smooth, real-valued function $f$ on $X$, consider the metric $g'_{\scriptscriptstyle X}=e^{2(\pi_{1}^{*}f)}g_{\scriptscriptstyle X}$ in the conformal class of $g_{\scriptscriptstyle X}$. The metrics $g_{\scriptscriptstyle X}$ and $g'_{\scriptscriptstyle X}$ determine two isomorphic ${\rm SO}(4)$ bundles:
\begin{center}
$P_{{\rm SO}(4)}=\{p\in P_{{\rm CO}(4)}~|~ g_{{\mathbb{R}}^{4}}(\theta_{p}, \theta_{p}) = \pi_{1}^{\ast}g_{\scriptscriptstyle X}(\cdot, \cdot)\}$ \\[0.2cm]
$P'_{{\rm SO}(4)}=\{p\in P_{{\rm CO}(4)}~|~ g_{{\mathbb{R}}^{4}}(\theta_{p}, \theta_{p}) = e^{2(\pi_{1}^{\ast}f)}\pi_{1}^{\ast}g_{\scriptscriptstyle X}(\cdot, \cdot)\}$
\end{center}
where, $ g_{{\mathbb{R}}^{4}}(\cdot, \cdot)$ is the standard metric on ${\mathbb{R}}^{4}$. Let $\varphi$ be a connection on $P_{{\rm CO}(4)}$. Then $\varphi + \theta$ define a 1-form with values in $\mf{co}(4) \oplus {\mathbb{R}}^4$. We can extend the bracket on the Lie algebra $\mf{co}(4)$ to $\mf{co}(4) \oplus {\mathbb{R}}^4$ as
\eqst{
[A, x] = - [x, A] = Ax, ~ [x,y] = 0, ~ \text{for} ~ x,y \in {\mathbb{R}}^4 ~\text{and}~ A \in \mf{co}(4).
}
This defines an affine Lie algebra which is best identified with the frame bundle of ${\mathbb{R}}^4$.
The failure of the 1-form $\varphi + \theta$ to conform with the associated Maurer-Cartan form is measured by
\eqst{
d(\varphi + \theta) + [\varphi + \theta, \varphi + \theta] = \mc{R}(\varphi) + T(\varphi)
}
where
\eqst{
\mc{R}(\varphi) = d\varphi + \frac{1}{2}[\varphi, \varphi], ~~~ T(\varphi) = d\theta + [\varphi, \theta].
}
Here the entities $\mc{R}$ and $T$ are horizontal 2-forms on the conformal frame bundle, which are nothing but the curvature and the torsion tensors, respectively and the Lie bracket operations are carried out simultaneously with wedging of 1-forms.
Suppose that $\varphi$ is a connection on $P_{{\rm CO}(4)}$ satisfying
\eq{
(d+\varphi)\, g_{\scriptscriptstyle X} = 0 ~~~ \text{and} ~~~ (d+\varphi)\, \theta = 0.
}
Then $ \varphi$ is just the Levi-Civita connection for the metric $ g_{\scriptscriptstyle X}$. Let $\varphi'$ denote the Levi-Civita connection for the metric $g'_{\scriptscriptstyle X}$. The difference of the 2-connections is a horizontal 1-form on $P_{{\rm CO}(4)}$ and therefore can be written as contraction of $\theta$ with an equivariant function $\xi \in \Hom({\mathbb{R}}^{4}, \mf{co}(4)) \cong ({\mathbb{R}}^{4})^{*} \otimes \mf{co}(4)$. More precisely,
\[ \pair{\theta_p, \, \xi }(Y) = \pair{\theta_p(Y), \, \xi}, ~~~ Y\in T_{p}P_{{\rm CO}(4)}.\]
Therefore we may write
\eq{
\label{difference of LC connections}
\varphi'-\varphi \, = \, \pair{\theta, \, \xi} ~~\text{for some}~~ \xi \in ({\mathbb{R}}^{4})^{*} \otimes \mf{co}(4).
}
Throughout, we will supress the pairing with $\theta$ and simply write $\varphi'-\varphi=\xi$. Consider the covariant derivative of $g'_{\scriptscriptstyle X}$ with respect to $\varphi$
\eq{
\label{eq: levi civita phi'}
(d+\varphi)\,(g'_{\scriptscriptstyle X}) \, = \, -e^{2(\pi_{1}^{*}f)}\,\, 2\,(\pi_{1}^{*}df)\,g_{\scriptscriptstyle X}.
}
The right hand side of the equation can be understood as follows. Define
\eqst{
f_{i}(p) = \pi_{1}^{*}df \, (\widetilde{p(e_{i})}),
}
where, $e_{i}\in {\mathbb{R}}^{4}$ is the standard basis element of ${\mathbb{R}}^4$ and $\widetilde{p(e_{i})}$ is the horizontal lift of $p(e_{i})$ to $P_{{\rm CO }(4)}$ with respect to $\varphi$. We can write
\eqst{
\pi_{1}^{*}df(p) \, = \, \pair{\sum_{i=1}^4 f_i(p)\, e^i, \, \theta_p}, \,\,\,\,\,\, \sum_{i=1}^4 f_i(p)\, e^i \in ({\mathbb{R}}^4)^{\ast} \hookrightarrow ({\mathbb{R}}^4)^{\ast}\otimes \mf{co}(4)
}
where $e^i$ are the basis for $({\mathbb{R}}^4)^{\ast}$. So the action of $\pi_{1}^{*}df$ is just the (left) action of $\sum_{i=1}^4 f_i\, e^i \in \End({\mathbb{R}}^4)$.
\begin{rmk}
The negative sign in the equation \eqref{eq: levi civita phi'} is due to the left action of $\aut({\mathbb{R}}^{4}) \curvearrowright S^{2}({\mathbb{R}}^{4})^{\ast}$, which is given by \[S^{2}({\mathbb{R}}^{4})^{\ast} \ni g_{\scriptscriptstyle X} \longmapsto b \cdot g_{\scriptscriptstyle X}(\cdot, \cdot) := g_{\scriptscriptstyle X}(b^{-1}, b^{-1}),\] where $b \in \aut({\mathbb{R}}^{4})$.
\end{rmk}
It follows that $\varphi \, + \, \pi_1^{\ast}df$ is a metric connection for $g'_{\scriptscriptstyle X}$. But it has a non-zero torsion. Indeed
\eq{
\label{eq: torsion of phi'}
\left(d \, + \, \varphi \,+ \, \pi_1^{\ast}df \right)\,\theta = \pair{\sum_{i=1}^4 f_i\, e^i, \, \theta} \wedge \theta.
}
Point-wise, the torsion tensor is a map
\eqst{
T(\varphi)(p):\Lambda^{2}D_{p} \cong \Lambda^{2}{\mathbb{R}}^{4}\xrightarrow{d\theta} {\mathbb{R}}^{4}.
}
For the connections $\varphi$ and $\varphi'$ on $P_{{\rm CO}(4)}$, the difference between their torsion tensors is
\eqst{
T(\varphi')_p \, (x\wedge y) - T(\varphi)_p \, (x\wedge y) = \frac{1}{2}(\xi_p(x) \, y -\xi_p(y) \, x), ~~~~ x,y\in{\mathbb{R}}^{4},
}
In terms of the ${\rm CO}(4)$-equivariant homomorphism:
\eqst{
\delta:({\mathbb{R}}^{4})^{*}\otimes \mf{co}(4)\hookrightarrow ({\mathbb{R}}^{4})^{\ast}\otimes({\mathbb{R}}^{4})^{\ast}
\otimes{\mathbb{R}}^{4}\mapsto\Lambda^{2}({\mathbb{R}}^{4})^{*}\otimes{\mathbb{R}}^{4}\cong \Lambda^{2}({\mathbb{R}}^{4})^{*}\otimes({\mathbb{R}}^{4})^{\ast}
}
where, the first map is the inclusion and the second one is the anti-symmeterization, we can write $T(\varphi')_p - T(\varphi)_p = -\delta\xi$. Therefore, it follows from \eqref{eq: torsion of phi'} that
\eqst{
\pair{\sum_{i=1}^4 f_i(p)\, e^i, \, \theta} \wedge \theta \, = \, -\delta\left(\sum_{i=1}^4 f_i(p)\, e^i \right).
}
Identify $\mf{so}(4)\cong \Lambda^{2}$ by associating the skew-symmetric endomorphism, to a pair of vectors $v,w\in {\mathbb{R}}^{n}$,
\eq{ \label{so4isomorphism}
v\wedge w = \langle v,\cdot \rangle w - \langle w,\cdot \rangle v.
}
\begin{lem}[\cite{salamon}, Prop. 2.1]
The restriction \[\delta|_{\mf{so}(4)}:({\mathbb{R}}^{4})^{\ast}\otimes\Lambda^{2}({\mathbb{R}}^{4})^{\ast}\mapsto\Lambda^{2}({\mathbb{R}}^{4})^{\ast}\otimes({\mathbb{R}}^{4})^{\ast}\]
that maps the difference of two connections to the difference of their torsions is an isomorphism.
\end{lem}
\begin{proof}
Let $a_{ijk} \in ({\mathbb{R}}^{4})^{\ast}\otimes\Lambda^{2}({\mathbb{R}}^{4})^{\ast}$ denote the difference of Christoffel symbols of the two connections. Then, $\delta(a_{ijk}) = \frac{1}{2}(a_{ijk}-a_{jik})$. It is easily seen that if $a_{ijk}\in \text{ker}(\delta)$, then $a_{ijk} = 0$ and hence $\delta|_{\mf{so}(4)}$ is an isomorphism.
\end{proof}
Suppose that $A$ is the Levi-Civita connection and $B$ is a metric connection on $P_{{\rm CO}(4)}$. Then using the isomorphism $\delta|_{\mf{so}(4)}$, we obtain the expression for $A$ in terms of $B$. Let $B' = B - \alpha$ where $\alpha = \delta|_{\mf{so}(4)}^{-1} (\delta(\xi))$. Then a straightforward computation shows that $T(B') = 0$. This is the strategy we are going to employ to express $\varphi'$ in terms of $\varphi$ and correction terms.
\noindent Pointwise, we can view $\sum_{i=1}^4 f_i\, e^i$ as a 1-form with values in $({\mathbb{R}}^4)^{\ast}\otimes\mf{co}(4)$, by writing
\eqst{
\sum_{i=1}^4 f_i\, e^i = \sum_{i,j} f_{i}\, e^{i} \otimes e^j \otimes e_j \in ({\mathbb{R}}^4)^{\ast}\otimes ({\mathbb{R}}^4)^{\ast} \otimes {\mathbb{R}}^4.
} Using the isomorphism ${\mathbb{R}}^4 \cong ({\mathbb{R}}^4)^{\ast}$, we can write the right hand side as
$\sum_{i,j} f_{i}\, e^{i} \otimes e^j \otimes e^j.$ So,
\eqst{
\delta \left(\sum_{i, j} f_i \, e^i \otimes e^j \otimes e^j \right) \,
= \, \frac{1}{2} \sum_{i, j} f_i \left( e^i \otimes e^j \otimes e^j - e^j \otimes e^i \otimes e^j \right)
}
and therefore
\alst{
\delta|_{\mf{so}(4)}^{-1} \left[\delta \left(\sum_{i, j} f_i \, e^i \otimes e^j \otimes e^j \right) \right] = \sum_{i, j} f_{i} \, (e^j \otimes e^j \otimes e^i - e^j \otimes e^i \otimes e^j) = -\sum_{i, j} f_{i} \, e^{j}\otimes (e^{i}\wedge e^{j}).
}
It is now easily verified that the torsion
\eqst{
T \left(\varphi + \pi_1^{\ast}df - \delta|_{\mf{so}(4)}^{-1} \left(\delta \left(-\sum_{i=1}^4 f_i\, e^i \right) \right)\right) = -\delta \left(\sum_{i=1}^4 f_i\, e^i \right) - \delta \left(-\sum_{i=1}^4 f_i\, e^i \right) = 0.
}
In conclusion, this is nothing but the Levi-Civita connection for the metric $g'_{\scriptscriptstyle X}$ and therefore
\eqst{
\label{diffconnso}
\varphi' = \varphi + \pi_1^{\ast}df + \pair{\sum_{i, j} f_{i} e^{j}\otimes (e^{i}\wedge e^{j}), \theta}.
}
For simplicity, put $\alpha = \pi_1^{\ast}df + \pair{\sum_{i, j} f_{i} e^{j}\otimes (e^{i}\wedge e^{j}), \theta}$.
\begin{prop}[\cite{lawsonmichelsohn} Prop. 6.2, Chap. I]
The adjoint representation induces the Lie algebra isomorphism $\zeta: \mf{spin}(n)\longrightarrow \mf{so}(n)$ is given by:
\[ \zeta(e_{i}e_{j}) = 2e_{i}\wedge e_{j},\] where, $\{e_{i}e_{j}\}_{i<j}$ are the basis elements of $\mf{spin}(n)$. Consequently for $v,w \in {\mathbb{R}}^{n}$, \[ \zeta^{-1}(v\wedge w) = \frac{1}{4} [v, w] .\]
\end{prop}
Under this isomorphism , $\alpha$ gets mapped to $\sum_{i=1}^4 f_i\, e^i + \frac{1}{4}\sum_{i, j} f_{i} e^{j}\otimes (e^{i}e^{j} - e^{j}e^{i})$. We denote this again by $\alpha$.
\subsection{A review of Swann's construction}
\label{subsec: example 2 Swann bundles}
A quaternionic K\"ahler manifold is a $4n$ dimensional manifold whose holonomy is contained in ${\rm {Sp}}(n){\rm {Sp}}(1):= ({\rm {Sp}}(n)\times {\rm {Sp}}(1))/\pm 1$. Let $N$ be a quaternionic K\"ahler manifold of positive scalar curvature and $F$ be the ${\rm {Sp}}(n){\rm {Sp}}(1)$ reduction of the frame bundle $P_{{\rm SO}(4n)}$ of $N$. Then $\mc{S}(N) := F/{\rm {Sp}}(n)$ is a principal ${\rm SO}(3)$-bundle, which is the frame bundle of the three- dimensional vector sub-bundle of skew symmetric endomorphisms of $TN$. The ${\rm {Sp}}(1)$-action, by left multiplication, descends to an isometric action of ${\rm SO}(3)$ on $\HH^{\ast}/\Z_2$. Swann bundle over $N$ is the principal $\HH^{\ast}/\Z_2$
\eqst{
\swann := \mc{S}(N)\times_{{\rm SO}(3)}(\HH^{\ast}/Z_2) \longrightarrow N
}
\begin{thm}\cite{swann}
\label{thm: swanns theorem}
The manifold $\swann$ is a hyperK\"ahler manifold with a free, permuting action of ${\rm SO}(3)$ and admits a hyperK\"ahler potential given by $\rho_0 = \frac{1}{2} r^2$. The vector field $\euler = -I_{\xi}K^M_{\xi}$ is independent of $\xi\in \mf{sp}(1)$ and $\grad \rho_0 = \euler$. Moreover, if a Lie group $G$ acts on $N$, preserving the quaternionic K\"ahler structure, then the action can be lifted to a tri-Hamiltonian action of $G$ on $\swann$.
\end{thm}
\noindent The Riemannian metric on the total space $\swann$ is given by $g_{\scriptscriptstyle \swann} = g_{\HH^{*}/\Z_{2}} + r^{2}g_{N}$ where $r$ is the radial co-ordinate on $\HH^{*}/\Z_{2}$ and $g_{\HH^{*}/\Z_{2}}$ is the quotient metric obtained from $\HH$. Alternatively, one can write
\eqst{
\swann = (0,\infty) \times \mc{S}(N)
}
with metric $g_{\mc{U}(N)} = dr^{2} + r^{2}(g_{N} + g_{\mbb{RP}^{3}})$, where $g_{\mbb{RP}^{3}}$ is the quotient metric on $\mbb{RP}^{3}$ derived from its double cover $S^{3}$. Thus, $\swann$ is a metric cone over $\mc{S}(N)$. The manifold $\swann$ is equipped with a natural left action of $\HH^{\ast} \cong {\mathbb{R}}^+ \times {\rm {Sp}}(1)$
\eq{
\big( (\lambda,q)\, (r,s) \big) \longmapsto (\lambda \cdot r, \, q \cdot s).
}
\subsection{Generalised Dirac operators for conformally related metrics}
\label{subsec: gen dirac and conf. metric}
Henceforth, fix an $M = \swann$, for some quaternionic K\"ahler manifold $N$ of positive scalar curvature and an action of $G$ that preserves the quaternionic K\"ahler structure on $N$. By Theorem \ref{thm: swanns theorem}, the action lifts to a tri-Hamiltonian action of $G$ on $\swann$. Therefore $M$ carries a permuting action of ${\rm Spin}^G(4)$.
Define the conformal ${\rm Spin}^G(4)$ group ${\rm CSpin}^G(4) := {\mathbb{R}}^{+} \times {\rm Spin}^G(4)$, which is a double cover of ${\rm CO}(4) \times G$
\eq{
\label{csping}
0 \longrightarrow \Z/2\Z \longrightarrow {\rm CSpin}^G(4) \overset{\gamma}{\longrightarrow} {\rm CO}(4) \times G \longrightarrow 0.
}
\begin{defn}
A ${\rm CSpin}^G(4)$-structure over $X$ is a principal ${\rm CSpin}^G(4)$-bundle $\widetilde{\pi}: \widetilde{Q} \rightarrow X$, which is an equivariant double cover of bundle $P_{{\rm CO}(4)} \times_X P$, with respect to the map $\gamma$.
\end{defn}
Let $\varphi$ and $\varphi'$ denote the Levi-Civita connections for metrics $g_{\scriptscriptstyle X}$ and $g'_{\scriptscriptstyle X} \in [g_{\scriptscriptstyle X}]$ respectively. Fix a $\overline{G}$-connection $A$ on $P$. Then $A$ uniquely determines the connections $A_{\varphi}$ and $A_{\varphi '}$, which are lifts of $\varphi$ and $\varphi'$ to $\widetilde{Q}$. Then, as shown in subsection \ref{subsec: metric connections},
\eqst{
A_{\varphi '} - A_{\varphi} = \alpha \in C^{\infty}(\widetilde{Q}, \,({\mathbb{R}}^4)^{\ast}\otimes \mf{g})^{{\rm Spin}^G(4)}.
}
Consequently, the covariant derivative of $u$, with respect to $A_{\varphi '}$ is
\eq{ \label{diffcovop}
D_{A_{\varphi'}}u = D_{A_{\varphi}}u + K^{M}_{\alpha}|_{u} \in C^{\infty}\big(\widetilde{Q}, ({\mathbb{R}}^{4})^{\ast}\otimes u^{\ast}TM \big)^{{\rm CSpin}^G(4)}.
}
Recall that $\swann$ admits a hyperk\"ahler potential $\rho_{0}$ and $\euler = \grad \rho_0$. For $\lambda \in {\mathbb{R}}\setminus \{0\}$,
\eqst{
\rho_{0}(e^{\lambda}x) = \frac{1}{2} g^{M}(\euler|_{e^{\lambda}x}, \euler|_{e^{\lambda}x}) = \frac{1}{2} ~ e^{2\lambda} g^{M}(\euler|_{x}, \euler|_{x}) = e^{2\lambda} \rho_{0}(x).
}
Therefore
\eqst{
\frac{d}{dt} \rho_{0}(e^{2t \lambda}x)|_{t=0} = d\rho_{0} (\frac{d}{dt}(e^{2t \lambda}x)) = 2d\rho_{0} (K_{\lambda}^{M, {\mathbb{R}}^{+}}) |_{x} = g^{M}(\euler|_{x}, K_{\lambda}^{M, {\mathbb{R}}^{+}}|_{x}).
}
On the other hand
\eqst{
\frac{d}{dt} \rho_{0}(e^{2t \lambda}x)|_{t=0} = \frac{d}{dt}(e^{2t\lambda}) \rho_{0}(x) = 2\lambda \rho_{0}(x) = g_{M}(\euler|_{x},\euler|_{x}),
}
which implies that $K_{\lambda}^{M, {\mathbb{R}}^{+}} = \lambda \euler$.
\noindent We are now in a position to give the proof of Theorem \ref{thm: main thm 1}. But first, we need the following Lemma:
\begin{lem}
\label{gendirac derivative property}
For $f \in C^{\infty}(X, {\mathbb{R}})$, we have
\begin{equation}
\Dd_{A}(e^{-\pi_{1}^{*}f}u) = de^{-\pi_{1}^{*}f}~\Dd_{A}u - \pi_1^{\ast}df \bullet \euler \circ u,
\end{equation}
where $de^{-\pi_{1}^{*}f}$ denotes the differential of the action of $e^{-\pi_{1}^{*}f}$ on $TM$.
\end{lem}
\begin{proof}
Let $p \in \widetilde{Q}$ and $v \in T_{p}\widetilde{Q}$. Let $\gamma:[0,1] \longrightarrow \widetilde{Q}$ be a curve in $\widetilde{Q}$ such that $\gamma(0) = p$ and $\dot{\gamma}(0) = v$. Evaluating the covariant derivative of $e^{-\pi_{1}^{*}f} u$ for $v$:
\eqst{
D_{A}(e^{-\pi_{1}^{*}f} u)(v) = d(e^{-\pi_{1}^{*}f}u)\,(v) + K^{M}_{A(v)}|_{e^{-\pi_{1}^{*}f(p)}u(p)}.
}
The first term of the above expression is
\alst{
d(e^{-\pi_{1}^{*}f} u)(v)
&= \frac{d}{dt}\big(e^{-\pi_{1}^{*}f} u \big)(\gamma(t))|_{t=0} \\
&= \frac{d}{dt}\big(e^{-\pi_{1}^{*}f(\gamma(t))} u(\gamma(t)) \big)|_{t=0} \\
&= de^{-\pi_{1}^{*}f(p)}du(v) + K^{M}_{\left(-\pi_{1}^{*}df(v)\right)} |_{u(p)}\\
&= de^{-\pi_{1}^{*}f(p)}\, du(v) - \big\langle \sum_{i=1}^4 f_i\, e^i, \, \theta(v) \big\rangle \euler|_{u(p)}
}
and the second term is
\eqst{
K^{M}_{A(v)}|_{e^{-\pi_{1}^{*}f(p)}u(p)} = de^{-\pi_{1}^{*}f(p)} ~ K^{M}_{A(v)}|_{u(p)}.
}
In conclusion,
\eqst{
D_{A}(e^{-\pi_{1}^{*}f} u) = de^{-\pi_{1}^{*}f}~D_{A}u - \big\langle\sum_{i=1}^4 f_i\, e^i, \, \theta \big\rangle \otimes \euler \circ u.
}
Applying Clifford multiplication, proves the statement of the Lemma.
\end{proof}
\begin{proof}[\textbf{Proof of Theorem \ref{thm: main thm 1}}]
With respect to the metric $e^{2\pi^{*}f}g_{\scriptscriptstyle X}$, the Clifford multiplication is given by $\bullet' = de^{-\pi_{1}^{*}f} \bullet$.
Substituting for $\alpha$ in \eqref{diffcovop} and applying the Clifford multiplication we get:
\eq{
\label{eq: thm.1 eq 1}
\Dd_{A_{\varphi'}}u
= de^{-\pi_{1}^{*}f}\left(\Dd_{A_{\varphi}}u + \pi_1^{\ast}df \bullet \euler\circ u + \frac{1}{4} \langle\sum_{i < j} f_{i} e^{j},\, \theta\rangle \bullet K^{M}_{(e^{i}e^{j} - e^{j}e^{i})} |_{u}\right)
}
Note that in using the identification $({\mathbb{R}}^4)^{\ast} \cong \HH$, the element $(e^{i}e^{j} - e^{j}e^{i})$ belongs to the Lie algebra $\mf{sp}(1) \cong \mf{Im}(\HH)$ and has norm 1. Now recall from Theorem \ref{thm: swanns theorem} the vector field $\euler = - I_{\xi}K^M_{\xi}$ is independent of $\xi \in \mf{sp}(1)$. In particular when $\abs{\xi} = 1$, we get $I_{\xi}\euler = K^M_{\xi}$. Therefore,
\eqst{
K^{M}_{(e^{i}e^{j} - e^{j}e^{i})} |_{u} = I_{(e^{i}e^{j} - e^{j}e^{i})} \euler\circ u = (e^{i}e^{j} - e^{j}e^{i}) \bullet \euler\circ u.
}
Substituting this in \eqref{eq: thm.1 eq 1}, we get
\alst{
\Dd_{A_{\varphi'}}u
&= de^{-\pi_{1}^{*}f}\left(\Dd_{A_{\varphi}}u + \pi_1^{\ast}df \bullet\euler\circ u + \frac{1}{4} \langle\sum_{i < j} f_{i} e^{j}, \, \theta\rangle \bullet (e^{i}e^{j} - e^{j}e^{i}) \bullet \euler\circ u \right) \\
&= de^{-\pi_{1}^{*}f}\Big(\Dd_{A_{\varphi}}u + \pi_1^{\ast}df \bullet \euler\circ u + \frac{1}{4} \, \langle 4\sum_{i} f_{i}e^{i} - 2\sum_{i,j} f_{i}e^{j}\delta_{i,j} + 4\sum_{i} f_{i}e^{i}, \theta \rangle \bullet \euler\circ u \Big) \\
&= de^{-\pi_{1}^{*}f}\Big(\Dd_{A_{\varphi}}u + \pi_1^{\ast}df \bullet \euler\circ u + \frac{3}{2}\,\pi_{1}^{*}df \bullet \euler\circ u \Big).
}
Now observe that
\alst{
\Dd_{A_{\varphi'}}(e^{-\pi_{1}^{*}f}u)
&= de^{-\pi_{1}^{*}f} \left(de^{-\pi_{1}^{*}f} \, \Dd_{A_{\varphi}}u + \frac{3}{2} \, de^{-\pi_{1}^{*}f} \, \pi_{1}^{*}df \bullet \euler\circ u \right) \\
&= de^{-\pi_{1}^{*}f}\left(de^{-\frac{5}{2}\pi_{1}^{*}f} \, \Dd_{A_{\varphi}} \, (e^{\frac{3}{2} \pi_{1}^{*}f}u)\right).
}
Thus, in conclusion
\eq{
\label{conformal transformation of gen dirac operator}
\Dd_{A_{\varphi'}} (\scr{B}u) = \scr{B} \left(de^{-5/2\pi_{1}^{*}f} \, \Dd_{A_{\varphi}} \, (e^{3/2\pi_{1}^{*}f}u) \right).
}
\end{proof}
\section{Almost Hermitian geometry and generalised Seiberg-Witten}
In this section, we give the proof of Theorem \ref{thm: main thm 2}. Let the target hyperK\"ahler manifold $M$ be as in Section \ref{subsec: gen dirac and conf. metric}, but with $G= {\rm U}(1)$, so that $M$ now carries a permuting action of ${\rm Spin}^c(4)$. Moreover, let $\text{dim}\,M = 4$. Fix a ${\rm Spin}^c(4)$-structure $Q \rightarrow X$. In this section we restrict our attention to those $\swann$ which can be obtained by a hyperK\"ahler reduction of a flat, quaternionic space. Examples include nilpotent co-adjoint, orbits of complex semi-simple Lie groups, the moduli spaces of instantons on 4-manifolds, etc. We describe this set-up below.
Let $V$ be a finite-dimensional, Hermitian vector space and $\mathsf{H}:= V \oplus V^{\ast}$. Then $\mathsf{H}$ is a flat-hyperK\"ahler manifold. Identifying $\mathsf{H}$ with $\HH^n$, for some $n$, it is easy to see that $\mathsf{H}$ carries a natural permuting action of ${\rm Sp}(1)$ given by multiplication by conjugate on the right. Consider the left action of ${\rm U}(1)$ on $\mathsf{H}$
\eq{
\label{eq: U(1) action on hk}
z \cdot (v, w) = (z\cdot v, z^{-1} \cdot w).
}
The action is tri-Hamiltonian, with a moment map
\eq{ \label{eq: moment map for S1 action}
\mu_{{\mathbb{R}}}(v,w) = \frac{1}{2} (\norm{v}^2 - \norm{w}^2), ~~ \mu_{{\mathbb{C}}}(v,w) = \pair{v,w}
}
Therefore, $\mathsf{H}$ admits a permuting action of ${\rm U}(2)$. Suppose that another compact Lie group $G \subset U(n) \hookrightarrow Sp(n)$ has a tri-Hamiltonian action on $\mathsf{H}$ that commutes with the ${\rm U}(2)$-action. Assume zero is a regular value of the $G$-moment map $\mu_{\mf{g}}: \mathsf{H} \rightarrow \mf{sp}(1)^{\ast}\otimes \mf{g}^{\ast}$. Then, ${\rm U}(2)$ preserves the zero level set of $\mu_{\mf{g}}$ and therefore descends to a permuting action on the quotient $M:= \mu_{\mf{g}}^{-1}(0)/ G$. Put $\widehat{G}:= {\rm Spin}^c(4) \times G$.
\begin{rmk}
More generally, we can consider $\mathsf{H} = \displaystyle \sum_{i=1}^k V_i \oplus V^{\ast}_i$, where each $V_i$ is a complex representation of ${\rm U}(2) \times G$, equipped with the tri-holomorphic action of ${\rm U}(1)$ by (weighted) left multiplication, so that it may happen that ${\rm U}(1)$ acts non-trivially on the first $\{V_l\}_{l=1}^m, ~ 1<m<k$ and trivially on the rest. However, we require that the image of the spinor be devoid of fixed points of the ${\rm U}(1)$-action. Therefore, we stick to the case where $\mathsf{H} = V \oplus V^{\ast}$ and ${\rm U}(1) \hookrightarrow {\rm Sp}(n) \curvearrowright\mathsf{H}$.
\end{rmk}
\subsection{Modified Seiberg-Witten equations}
By assumption $\mu_{\mf{g}}^{-1}(0)/G = M$. Let $P:=\mu^{-1}_{\mf{g}}(0)$ denote the ${\rm Spin}^c$-equivariant principal $G$-bundle over $M$.
\begin{wrapfigure}{r}{0.35\textwidth}
\begin{center}
\vspace{-6mm}
\hspace{-3mm}
\begin{tikzpicture}[->, node distance=2.25cm, auto, shorten >=1pt, every edge/.style={font=\footnotesize, draw}, fill=blue]
\node (Hn+1) {$Q$};
\node (On+1) [right of=Hn+1] {$M$};
\node (HPn) [above of=Hn+1] {$\widehat{Q}$};
\node (Xn-1) [right of=HPn] {$P \subset \mathsf{H}$};
\node (X) [below of=Hn+1] {$X$};
\draw (Hn+1) -- node [left, midway] {$\pi$} (X);
\draw (Hn+1) -- node [above, midway] {$u$} (On+1);
\draw (HPn) -- node [left, midway] {$\pi_{1}$} (Hn+1);
\draw (Xn-1) -- node [right, midway] {$\pi_{2}$} (On+1);
\draw (HPn) -- node [above] {$\widehat{u}$} (Xn-1);
\end{tikzpicture}
\vspace{0.5cm}
\end{center}
\end{wrapfigure}
Consider a $\widehat{G}$-bundle $\widehat{Q} \rightarrow X$, as in the diagram. Given a smooth, equivariant map $\widehat{u}: \widehat{Q} \longrightarrow \mathsf{H}$, such that $\mu_{\mf{g}} \circ \widehat{u} = 0$, define $u: Q \rightarrow M$ by $u(q) = \pi_2 (\widehat{u}(p)), ~~ q\in Q, ~~ p \in \pi^{-1}_1(q)$.
Clearly then, $u$ is a ${\rm Spin}^c(4)$-equivariant map and the diagram commutes. On the other hand, given a smooth spinor $u: Q \longrightarrow M$, it defines a principal $\widehat{G}$-bundle over $X$, via pull-back of $P$ and canonically defines $\widehat{u}$, making the diagram commutative. In summary,
\begin{lem}
There is a bijective correspondence between
\eqst{\{u \in C^{\infty}(Q, \, M)^{{\rm Spin}^c}\} \Longleftrightarrow \{\widehat{u}\in C^{\infty}(\widehat{Q}, \, \mathsf{H})^{\widehat{G}}~|~ \mu_{\mf{g}}\circ \widehat{u} = 0\}.
}
\end{lem}
\noindent Fix a connection $\ms{A}$ on $Q$. This is uniquely determined by the Levi-Civita connection on $X$ and a connection $\ms{b}$ on the determinant bundle $P_{{\rm U}(1)}$. The bundle $P \rightarrow M$ is a Riemannian submersion and therefore carries a canonical connection $\ms{a}$. This is defined as follows.
For $p \in P$, let $K^{P, G}_{\eta}|_p$ denote the fundamental vector field at $p$ due to $\eta \in \mf{g}$. For $v \in T_p P$, define $\ms{a}_p(v) \in \mf{g}$ be the unique element such that
\eqst{
K^{P, G}_{\ms{a}}|_{p}(v) = K^{P, G}_{\ms{a}(v)}|_{p} = -\pr^{\text{im}K^{P, G}}(v)
}
where $\pr^{\text{im}K^{P, G}}$ denotes the orthogonal projection to the vertical sub-bundle, which is nothing but the image of the map
\eqst{
K^{P, G}: \mf{g} \rightarrow TP, \,\,\,\,\, \eta \longmapsto K^{P, G}(\eta)|_p = K^{P, G}_{\eta}|_p.
}
The pull-back of this connection by $\widehat{u}$, along with the connection $\ms{A}$ on $Q$, uniquely determine a connection $\widehat{\ms{A}}$ on $\widehat{Q}$ (see \cite{victor})
\eq{
\label{eq:unique connection on P_H}
\widehat{\ms{A}} = \pi^{\ast} \ms{A} \oplus \widehat{\ms{A}}_{\mf{g}} \in \Lambda^{1}\left(\widehat{Q}, ~\widehat{\mf{g}}\right)^{\widehat{G}}, ~~ \widehat{\ms{A}}_{\mf{g}} = \widehat{u}^{\ast}\ms{a} - \langle \pi_1^{\ast} \ms{A}, \iota_{\mf{spin}^c} \widehat{u}^{\ast}\ms{a} \rangle.
}
We can define a twisted Dirac operator $\Dd_{\widehat{\ms{A}}}$ acting on maps $\widehat{u}$.
\begin{prop}
\label{prop:1-1 correspondence}
Then, there is a 1-1 correspondence between
\eq{
\label{eq:1-1 correspondence harmonic spinors}
\{(\widehat{u}, \widehat{\ms{A}})~|~ \Dd_{\widehat{\ms{A}}}\widehat{u} = 0, ~~ \mu_{\mf{g}} \circ \widehat{u} = 0\}~~~\text{and}~~~\{(u, \ms{A})~|~ \Dd_{\ms{A}}u = 0\}.
}
Whenever $\Dd_{\widehat{\ms{A}}}\widehat{u} = 0, ~~ \mu_{\mf{g}} \circ \widehat{u} = 0$ and $\pr_{\mf{g}} \widehat{\ms{A}} = \widehat{\ms{A}}_{\mf{g}}$ as in \eqref{eq:unique connection on P_H} and therefore, $\widehat{\ms{A}}$ is uniquely determined by a ${\rm U}(1)$-connection $a$ on $P_{{\rm U}(1)}$.
\end{prop}
\begin{proof}
For $h\in P$ such that $\mu_{\mf{g}}(h) = 0$, define $\Hh_h := \ker d\mu_{\mf{g}}(h) \cap (\im K^{P, G})^{\perp}$. This is just the horizontal subspace over $h$ with respect to the canonical connection $\ms{a}$.
We will prove the proposition in two steps. In what follows, we shall denote the $G$ and $Spin^c$-components of $\widehat{\ms{A}}$ by $\widehat{\ms{A}}_{\mf{g}}$ and $\ms{A}$ respectively.
\vspace{0.3cm}
\paragraph{\bf Step 1:}
\label{para:step 1}
In the first step we will prove that $I_{\xi}D_{\widehat{\ms{A}}}\widehat{u}(v) \in \Hh_{\widehat{u}}$ for every $\xi \in \mf{sp}(1)$ and $v \in \Hh_{\widehat{\ms{A}}} \subset T \widehat{Q}$. Indeed, if $\mu_{\mf{g}} \circ \widehat{u} = 0$, then $d\widehat{u}(v) \in \ker d\mu_{\mf{g}}(\widehat{u}(p))$. Also, $K^{P, G}_{\widehat{\ms{A}}_{\mf{g}}}|_{\widehat{u}} \in \ker d\mu_{\mf{g}}(\widehat{u}(p))$ and $K^{P, {\rm Spin}^c}_{\widehat{\ms{A}}}|_{\widehat{u}} \in \ker d\mu_{\mf{g}}(\widehat{u}(p))$. Therefore, $D_{\widehat{\ms{A}}}\widehat{u}(v) \in \ker d\mu_{\mf{g}}(\widehat{u}(p))$. Consequently
\eqst{
0
=\langle d\mu_{\mf{g}} (D_{\widehat{\ms{A}}}\widehat{u}(v)), \xi \otimes \eta\rangle
= \langle I_{\xi} K^{P, G}_{\eta}|_{\widehat{u}(p)}, D_{\widehat{\ms{A}}}\widehat{u}(v) \rangle
= - \langle K^{P, G}_{\eta}|_{\widehat{u}(p)}, I_{\xi} D_{\widehat{\ms{A}}}\widehat{u}(v) \rangle
}
for $\xi \in \mf{sp}(1), ~ \eta \in {\mf{g}}$ and so $I_{\xi} D_{\widehat{\ms{A}}}\widehat{u}(v) \in (\im K^{P, G})^{\perp}$ for all $\xi \in \mf{sp}(1)$. For $\xi' \in \mf{sp}(1)$,
\eqst{
\langle d\mu_{\mf{g}} (I_{\xi}D_{\widehat{\ms{A}}}\widehat{u}(v)), \xi' \otimes \eta\rangle = \langle d\mu_{\mf{g}} (D_{\widehat{\ms{A}}}\widehat{u}(v)), [\xi, \xi'] \otimes \eta\rangle = 0
}
which implies $I_{\xi} D_{\widehat{\ms{A}}}\widehat{u}(v) \in \ker d\mu_G(\widehat{u}(p))$ for all $\xi \in \mf{sp}(1)$. Thus, $I_{\xi} D_{\widehat{\ms{A}}}\widehat{u}(v) \in \Hh_{\widehat{u}}$.
\vspace{0.3cm}
\paragraph{\bf Step 2:}
\label{para:step 2}
In this step, we prove the equivalence \eqref{eq:1-1 correspondence harmonic spinors}. If $\Dd_{\widehat{\ms{A}}} \widehat{u} = 0$, then from \eqref{eq:gen. dirac operator explicit}, we have
\eqst{
0 = D_{\widehat{\ms{A}}}\widehat{u}(\tilde{e_0}) - \sum^3_{i=1} I_{i}D_{\widehat{\ms{A}}} \widehat{u}(\tilde{e_i})
}
From Step 1, $D_{\widehat{\ms{A}}}\widehat{u}(\tilde{e_0}) \in \Hh_{\widehat{u}}$. It follows that $D_{\widehat{\ms{A}}}\widehat{u}(\tilde{e_i}) \in \Hh_{\widehat{u}}$ for all $i=1,2,3$. Consequently, for any $v \in \Hh_{\widehat{\ms{A}}}, ~~\pr^{\im K^{P, G}}D_{\widehat{\ms{A}}}\widehat{u}(v) = 0$ and we get $\displaystyle K^{P, G}_{\widehat{\ms{A}}_{\mf{g}}(v)} = - \pr^{\im K^{P, G}} d\widehat{u}(v)$. In other words, the ${\mf{g}}$-connection component of $\widehat{\ms{A}}$ is just the pull-back of the canonical connection on $P$. Since the diagram commutes, $d\pi_2(D_{\widehat{\ms{A}}}\widehat{u}) = D_{\ms{A}}u$. Also, as $D_{\widehat{\ms{A}}}\widehat{u}(\tilde{e_i}) \in \Hh_{\widehat{u}}$ for all $i=0, 1,2,3$, we have $\iota^{\ast}I_{i} = \pi^{\ast}_2 \tilde{I_{i}}$ and so,
\eqst{
0
=d\pi_2 (\Dd_{\widehat{\ms{A}}}\widehat{u})
= d\pi_2 \left(D_{\widehat{\ms{A}}}\widehat{u}(\tilde{e_0}) - \sum^3_{i=1} \iota^{\ast}I_{i}~D_{\widehat{\ms{A}}} \widehat{u}(\tilde{e_i})\right)
= \Dd_{\ms{A}}u
}
Thus, $\Dd_{\widehat{\ms{A}}}\widehat{u} = 0$ implies $\Dd_{\ms{A}}u = 0$. On the other hand if $\displaystyle K^{P, G}_{\widehat{\ms{A}}_{{\mf{g}}(v)}} = - \pr^{\im K^{P, G}} d\widehat{u}(v)$ then $D_{\widehat{\ms{A}}}\widehat{u} \in \Hh_{\widehat{u}}$ and so $d\pi_2 (\Dd_{\widehat{\ms{A}}}\widehat{u}) = \Dd_{\ms{A}}u$. Therefore, if $\Dd_{\ms{A}}u = 0$, it implies that $\Dd_{\widehat{\ms{A}}}\widehat{u} \in \im K^{P, G}$. But since,
\eqst{
\Dd_{\widehat{\ms{A}}}\widehat{u}
= D_{\widehat{\ms{A}}}\widehat{u}(\tilde{e_0}) - \sum^3_{i=1} \pi^{\ast}_2 \tilde{I_{i}}~D_{\widehat{\ms{A}}} \widehat{u}(\tilde{e_i}) \in \Hh_{\widehat{u}}
}
it follows that $\Dd_{\widehat{\ms{A}}}\widehat{u} \in (\im K^{P, G})^{\perp}$ and so $\Dd_{\widehat{\ms{A}}}\widehat{u} = 0$. This proves the statement.
\end{proof}
With this observation, it is now easy to construct a ``lift" of the equations as follows.
\begin{prop}
\label{prop: 1-1 correspondence of equations}
Fix a connection $a$ on $P_{{\rm U}(1)}$. There is a 1-1 correspondence between the following systems of equations
\eq{\label{eq: modified sw}
\left\{
\begin{array}{lcl}
\Dd_{\widehat{\ms{A}}}\widehat{u}= 0 \\
F^+_{\ms{b}} - \mu\circ \widehat{u} = 0 \\
\mu_{\mf{g}} \circ \widehat{u} = 0
\end{array}
\right. ~~\text{and}~~~~ \left\{
\begin{array}{lcl}
\Dd_{\ms{A}}u= 0 \\
F^+_{\ms{b}} - \mu\circ u = 0
\end{array}
\right.}
where $\mu: \mathsf{H} \rightarrow \rm{i}{\mathbb{R}}$ denotes the moment map for ${\rm U}(1)$-action on $\mathsf{H}$.
\end{prop}
Since the tri-Hamiltonian action of ${\rm U}(1)$ descends to $M$, we denote the ${\rm U}(1)$-moment map by $\mu$ itself. The above correspondence was independently obtained by Pidstrygach \cite{pidstrygach2006} and also by Haydys \cite{haydys2012} (Prop. 4.5 and Thm. 4.6).
\subsection{Almost-complex geometry and generalised Seiberg-Witten}
In this subsection, we give a proof of Theorem \ref{thm: main thm 2}. It exploits the equivalence \eqref{eq: modified sw} and Theorem \ref{thm: main thm 1}.
Firstly, note that the generalised Seiberg-Witten are not conformally invariant. On the other hand, from Theorem \ref{thm: main thm 1}, we know that the space of harmonic, generalised spinors is conformally invariant. It follows that there is 1-1 correspondence between the solutions $(\widehat{u}', \widehat{\ms{A}}')$ of the system \eqref{eq: modified sw} with respect to the metric $g'_{\scriptscriptstyle X} \in [g_{\scriptscriptstyle X}]$, such that image of $\widehat{u}$ does not contain a fixed point of the ${\rm U}(1)$-action on $\mathsf{H}$, and the triples $(g''_{\scriptscriptstyle X}, \widehat{u}'', \widehat{\ms{A}}'')$ such that $\abs{\mu\circ \widehat{u}''} = 1$ and $(\widehat{u}'', \widehat{\ms{A}}'')$ satisfy the equations
\eq{
\label{eq: conformally rescaled gen sw}
\left\{
\begin{array}{lcl}
\Dd_{\widehat{\ms{A}}''}\widehat{u}'' = 0\\
F^+_{\ms{b}} - \lambda\mu\circ \widehat{u}'' = 0 \\
\mu_{\mf{g}} \circ \widehat{u}'' = 0
\end{array}
\right.}
where is a strictly positive function given by $\lambda = \abs{\mu\circ u}^{-1}$. To see the correspondence, choose $g''_{\scriptscriptstyle X} = \abs{\mu\circ\widehat{u}'}^{-4/3} g'_{\scriptscriptstyle X}$. Then $u'' = \abs{\mu\circ\widehat{u}'}^{-1/2} u'$. By virtue of Theorem \ref{thm: main thm 1}, $u''$ is harmonic and the third equation of \eqref{eq: modified sw} remains invariant under the conformal scaling. Moreover, $\lambda\mu\circ\widehat{u}'' = \mu\circ \widehat{u}'$. The said correspondance follows from the map $(u', \ms{A}') \mapsto (u'', \ms{A}')$.
Suppose we are given a triple $(g''_{\scriptscriptstyle X}, \widehat{u}, \widehat{\ms{A}})$ satisfying \eqref{eq: conformally rescaled gen sw} and $\abs{\mu\circ \widehat{u}} = 1$. Then $\Omega = \Phi(\mu \circ \widehat{u})$ is a non-degenerate, self-dual 2-form on $X$, where $\Phi:\mf{sp}(1)^{\ast} \longrightarrow \Lambda^2_+({\mathbb{R}}^4)^{\ast}$ is the isomorphism, and defines an almost-complex structure on $X$.
\begin{lem}
\label{lem: 4d unique cont.}
Suppose that the target hyperK\"ahler manifold $M$ is 4-dimensional. Let $\ms{A}_0$ be a fiducial connection on $Q$ and $u$ be a spinor such that the range of $u$ does not contain a fixed point of the ${\rm U}(1)$-action on $M$. Then there exists a unique 1-form $\ms{a}_0$ on $X$ such that $\Dd_{\ms{A}}u = 0$, where $\ms{A} = \ms{A}_0 + \mathrm{\mathbf{i}}\ms{a}_0$.
\end{lem}
\begin{proof}
Observe that $\Dd_{\ms{A}}u = \Dd_{\ms{A}_0}u + \sum_{i=0}^3 e^i \bullet K^M_{\mathrm{\mathbf{i}}\ms{a}_0(\widetilde{e_i})}|_u$. At a point $q \in Q$,
\eqst{
K^M_{\mathrm{\mathbf{i}}\ms{a}_0\,(\widetilde{e_i}(q))} \,|_{u(q)} \, = \, \frac{d}{dt} \exp\left(\,\mathrm{\mathbf{i}} \, t\, \ms{a}_0(\widetilde{e_i}(q))\right) u(q)|_{t=0} \, = \, \left(\ms{a}_0(\widetilde{e_i}(q))\right) K^M_{\mathrm{\mathbf{i}}}|_{u(q)}.
}
Therefore
\alst{
\Dd_{\ms{A}}u(q)
& = \Dd_{\ms{A}_0}u(q) + \sum_{i=0}^3 \left(\ms{a}_0(\widetilde{e_i}(q)) e^i \right) \bullet K^M_{\mathrm{\mathbf{i}}}|_{u(q)} \\
& = \Dd_{\ms{A}_0}u(q) + \ms{a}_0(q) \bullet K^M_{\mathrm{\mathbf{i}}}|_{u(q)}.
}
Suppose that $\Dd_{\ms{A}}u = 0$. Then, we need to solve the equation
\eqst{
- \Dd_{\ms{A}_0}u = \ms{a}_0 \bullet K^M_{\mathrm{\mathbf{i}}}|_{u}.
}
Point-wise, we can choose identification of $T_{u(q)} M$ and ${\mathbb{R}}^4$ with quaternions, such that the Clifford multiplication is just the usual quaternionic multiplication. Since the image of $u$ does not contain a fixed point of the ${\rm U}(1)$ action on $M$, $K^M_{\mathrm{\mathbf{i}}}|_{u}$ is a non-vanishing, equivariant section of $u^{\ast}TM \rightarrow Q$. The statement of the Lemma follows.
\end{proof}
In essence, this translates to saying that given a non-vanishing spinor $\widehat{u}$ such that $\mu_{\mf{g}}\circ \widehat{u} = 0$, then there exists a unique 1-form $\ms{a}_0$ on $X$ such that $\Dd_{\widehat{\ms{A}}}\widehat{u} = 0$. Therefore, the connection $\widehat{A}$ is entirely determined by $\widehat{u}$ and hence by the almost complex structure $\Omega = \Phi(\mu\circ \widehat{u})$.
Let $B: \mathsf{H} \times \mathsf{H} \longrightarrow \mf{sp}(1)$ denote the symmetric (real) bi-linear form associated to the ${\rm U}(1)$-moment map and $\widetilde{B}$ denote the induced map on $(T^{\ast}X \otimes \mathsf{H}) \times (T^{\ast}X \otimes \mathsf{H})$, obtained using contraction furnished by the Riemannian metric on $X$. Then, $\Omega = B(\widehat{u}, \widehat{u})$ and so
\alst{
\nabla^{\ast}\nabla \Omega
&= 2 \left(B(D^{\ast}_{\widehat{\ms{A}}}D_{\widehat{\ms{A}}}\widehat{u}, \, \widehat{u}) - \widetilde{B}(D_{\widehat{\ms{A}}}\widehat{u}, \, D_{\widehat{\ms{A}}}\widehat{u})\right)}
Applying the Weitzenb\"ock formula
\eq{ \label{eq: weitzenbock for modified dirac}
\Dd^{\ast}_{\widehat{\ms{A}}}\Dd_{\widehat{\ms{A}}}\widehat{u} = D^{\ast}_{\widehat{\ms{A}}}D_{\widehat{\ms{A}}} \widehat{u} + \frac{s_{\scriptscriptstyle X}(g''_{\scriptscriptstyle X})}{4} \widehat{u} + F^+_{\ms{b}} \bullet \widehat{u} + F^+_{\widehat{\ms{A}}_{\mf{g}}} \bullet \widehat{u}
}
gives
\eqst{\nabla^{\ast}\nabla \Omega
= -\frac{s_{\scriptscriptstyle X}(g''_{\scriptscriptstyle X})}{2} \Omega - B(F^+_{\widehat{\ms{A}}_{\mf{g}}} \bullet \widehat{u}, \widehat{u}) - B(F^+_{\ms{b}} \bullet \widehat{u}, \widehat{u}) - 2 \widetilde{B}(D_{\widehat{\ms{A}}}\widehat{u}, D_{\widehat{\ms{A}}}\widehat{u})
}
We claim that the term $B(F^+_{\widehat{\ms{A}}_{\mf{g}}} \bullet \widehat{u}, \widehat{u})$ vanishes. This follows from the following Lemma:
\begin{lem}
Assume that $\mu_{\mf{g}}(h) = 0$ and let $\xi \in \mf{sp}(1)$ and $\eta \in \mf{g}$. Then
\eqst{
B(\widehat{u}, ~\eta~\widehat{u}~\overline{\xi}) = 0
}
\end{lem}
\begin{proof}
This follows from the fact that the ${\rm U}(1)$-moment map is $G$-invariant. For $\eta \in \mf{g}$, computing $\frac{d}{dt} B \left(u, \, \exp(t\eta)\, u \, \overline{\xi} \right)|_{t=0}$ proves the statement of the Lemma.
\end{proof}
It follows that $B(F^+_{\widehat{\ms{A}}_{\mf{g}}} \bullet \widehat{u}, \widehat{u}) = 0$.
Therefore,
\eq{
\label{eq: laplacian of omega}
\nabla^{\ast}\nabla \Omega
= -\left(\frac{s_{\scriptscriptstyle X}(g''_{\scriptscriptstyle X})}{2} + \lambda \right) \Omega - 2 \widetilde{B}(D_{\widehat{\ms{A}}}\widehat{u}, D_{\widehat{\ms{A}}}\widehat{u})
}
We are now is position to give the proof of Theorem \ref{thm: main thm 2}. The arguments of the proof are essentially the same as those of Donaldson's \cite{donaldson}. Nonetheless, for the sake of completeness, we present them here once again.
\begin{proof}[\textbf{Proof of Theorem \ref{thm: main thm 2}}]
Observe that since $\abs{\Omega} = 1$,
\eqst{
0 = \Delta \abs{\Omega} = 2\pair{\nabla^{\ast}\nabla \Omega, \Omega} - 2\abs{\nabla \Omega}^2.
}
Using \eqref{eq: laplacian of omega}, we get
\eqst{
2\lambda = -s_{\scriptscriptstyle X}(g''_{\scriptscriptstyle X}) -2\abs{\nabla \Omega}^2 - 2\pair{\widetilde{B}(D_{\widehat{\ms{A}}}\widehat{u}, D_{\widehat{\ms{A}}}\widehat{u}), \Omega}.
}
Therefore, re-arranging, we have
\eq{
\label{eq: eq1 for lemma for thm 2}
\abs{\nabla \Omega}^2 + \frac{1}{2}\, s_{\scriptscriptstyle X}(g''_{\scriptscriptstyle X}) + \pair{\widetilde{B}(D_{\widehat{\ms{A}}}\widehat{u}, D_{\widehat{\ms{A}}}\widehat{u}), \Omega} < 0.
}
Also, from \eqref{eq: laplacian of omega} we have that $(\nabla^{\ast}\nabla \Omega)^{\perp_{\Omega}} + \widetilde{B}(D_{\widehat{\ms{A}}}\widehat{u}, D_{\widehat{\ms{A}}}\widehat{u})^{\perp_{\Omega}} = 0$. Thus comparing with the identities \eqref{eq: main thm 2} of Theorem \ref{thm: main thm 2}, to complete our proof, we merely need to show that
\eq{
\label{eq: aim}
\widetilde{B}(D_{\widehat{\ms{A}}}\widehat{u}, D_{\widehat{\ms{A}}}\widehat{u})^{\perp_{\Omega}} = 2\pair{d\Omega, N_{\Omega}}, ~~\pair{\widetilde{B}(D_{\widehat{\ms{A}}}\widehat{u}, D_{\widehat{\ms{A}}}\widehat{u}), \Omega}
= \frac{1}{4}\left(\abs{N_{\Omega}}^2 - \abs{d\Omega}^2\right).
}
The key issue here is to identify the the map $\widetilde{B}$ on kernel of the Clifford multiplication. In order to do this, it suffices to restrict to the standard model when $X = {\mathbb{R}}^4$ and the connection $\widehat{\ms{A}}$ is trivial. This is because at any point $x \in X$, there exists a trivialisation in which the connection matrix $\widehat{\ms{A}}$ vanishes at the point $x$.
Since $\widehat{u} \in \ker \mu_{\mf{g}}$, the derivative $D\widehat{u} \in \Hh_{\widehat{u}} \subset \ker d\mu_{\mf{g}}$. At every point $p \in \ker \mu_{\mf{g}}$, the horizontal subspace $\Hh_p$ can be identified with $T_{\pi_2 (p)}M$. Since $M$ is 4-dimensional, $\Hh_p$ is 4-dimensional and so $\Hh_p \cong \HH$.
Let $(x_0, x_1, x_2, x_3)$ be the standard co-ordinates on ${\mathbb{R}}^4$. Let $s_1, s_2,\cdots\cdot s_{2n}$ denote the complex basis for the spinors and write $\widehat{u}$ as
\eqst{
\widehat{u}: {\mathbb{R}}^4 \longrightarrow \mathsf{H}, \,\,\,\,\, \widehat{u} = \sum^n_{i=1} f_i~s_i + \sum^{2n}_{i=n+1} g_{i-n} ~ s_{i} \,\,\,\,\, \text{where} \,\,\,\,\, f_i, g_i \in C^{\infty}({\mathbb{R}}^4, {\mathbb{C}}).
}
By Step 2 of Proposition \ref{prop:1-1 correspondence}, $D\widehat{u} \in \Hh_{\widehat{u}}$, which means that without loss of generality, at the origin, we can assume that
\eqst{
(f_i)_{x_j} = (g_i)_{x_j} = 0 \,\,\,\,\, \text{for} \,\,\,\,\, i = 2, 3, \cdots n \,\,\,\,\, \text{and} \,\,\,\,\, j = 0, 1, 2, 3.
}
Consequently, in the decomposition \eqref{eq: aim}, the only contributing terms are the 1-jets of $f_1, g_1$ at the origin. Therefore, without loss of generality, we can assume that at the origin, $f_i, g_i = 0$ for $i = 2, 3, \cdots\cdot n$. Let $f_0 = f_1(0)$ and $g_0 = g_1(0)$. Then, at the origin $u = f_0 \, s_1 + g_0 \, s_2$. Moreover, since $\abs{\Omega} = 1$, $\abs{f_0}^2 + \abs{g_0}^2 = 1$ and
\eqst{
B(\widehat{u}, \widehat{u}) = \left(\frac{\abs{f_0}^2 - \abs{g_0}^2}{2}\right) \,\beta_0 \, + \, \text{Re} \pair{f_0,g_0}\, \beta_1 \, + \, \text{Im} \pair{f_0,g_0}\, \beta_2
}
where $\beta_i$ are the basis of self-dual 2-forms on ${\mathbb{R}}^4$, given as in \eqref{eq: basis of self-dual 2-forms}. The group ${\rm Spin}(4)$ acts on the base ${\mathbb{R}}^4$ and also transitively on unit positive spinors. In particular, for a suitable choice of an element in ${\rm Spin}(4)$, we may further assume that at the origin, $f_0 = 1$ and $g_0 = 0$. In particular, $\Omega = \frac{1}{2} \beta_0$ at the origin. Thus $\Omega$ defines the standard complex structure $\frac{1}{2}\,\beta_0$ on ${\mathbb{R}}^4$. This allows us to use the complex co-ordinates
\eqst{
z = x_0 + ix_1, \,\,\,\,\, w = x_2 + ix_3.
}
From the Dirac equation we have
\eq{\label{eq: dirac eq. on R4}
f_{1\,\overline{z}} = g_{1\,w}, \,\,\,\,\, f_{1\,\overline{w}} = -g_{1\,z}.
}
Moreover, since $f_1 = 1$ at the origin, the derivatives of $f_1$ at the origin are purely imaginary. Therefore, at the origin,
\eq{
\label{eq: imaginary derivative}
f_{1\,z} = -\overline{f_{1\,\overline{z}}} \,\,\,\,\, \text{and} \,\,\,\,\, f_{1\,w} = -\overline{f_{1\,\overline{w}}}.
}
Now, the component of $\widetilde{B}(D\widehat{u}, D\widehat{u})$ along $\frac{1}{2}\,\beta_0$ is
\eqst{
\frac{1}{4}\sum^3_{l=0} \abs{\frac{\partial f_1}{\partial x_l}}^2 - \abs{\frac{\partial g_1}{\partial x_l}}^2 = \frac{1}{16}\left(\abs{f_{1\,z}}^2 + \abs{f_{1\,\overline{z}}}^2 + \abs{f_{1\,w}}^2 + \abs{f_{1\,\overline{w}}}^2 - \abs{g_{1\,z}}^2 - \abs{g_{1\,\overline{z}}}^2 - \abs{g_{1\,w}}^2 - \abs{g_{1\,\overline{w}}}^2 \right).
}
Using the identities \eqref{eq: dirac eq. on R4} and \eqref{eq: imaginary derivative}, we get
\eq{\label{eq: component along w}
\left\langle \widetilde{B}(D\widehat{u}, D\widehat{u}), \,\frac{1}{2}\beta_0 \right\rangle = \frac{1}{16} \left(\abs{g_{1\,z}}^2 + \abs{g_{1\,w}}^2 \right) - \frac{1}{16}\left(\abs{g_{1\,\overline{z}}}^2 + \abs{g_{1\,\overline{w}}}^2 \right).
}
The space orthogonal to $\frac{1}{2}\,\beta_0$ is spanned by $\beta_c = d\overline{z}\cdot d\overline{w}$ and therefore the component of $B(D\widehat{u}, D\widehat{u})$ orthogonal to $\frac{1}{2}\,\beta_0$ is
\alst{
\left(B(D\widehat{u}, D\widehat{u})\right)^{\perp_{\beta_0}}
&= \sum^3_{l=0} \left[ \left(\frac{\partial f_1}{\partial x_l}\right)^{\dagger} ~ \frac{\partial g_1}{\partial x_l} \right]\,\beta_c \\
&= \frac{1}{4}\left(f_{1\,z} ~ \overline{g_{1\,z}} + f_{1\,\overline{z}} ~ \overline{g_{1\,\overline{z}}} + f_{1\,w} ~ \overline{g_{1\,w}} + f_{1\,\overline{w}} ~ \overline{g_{1\,\overline{w}}}\right)\,\beta_c = \frac{1}{4} \left(g_{1\,z}\,\overline{g_{1\,\overline{w}}} + g_{1\,w} \, \overline{g_{1\,\overline{z}}} \right)\,\beta_c
}
where, once again, we have used the identities \eqref{eq: dirac eq. on R4} and \eqref{eq: imaginary derivative} in the penultimately step. Now $\Omega$ is a section of the twistor bundle and therefore its covariant derivative at the origin is given by the derivative of $f_1 \, \overline{g}_1$ which is nothing but the derivative of $g_1$. The holomorphic part $(g_{1\,z}, g_{1\,w})$ corresponds to the Nijenhuis tensor $N_{\Omega}$ whereas the anti-holomorphic component $(g_{1\,\overline{z}}, g_{1\,\overline{w}})$ corresponds to $d\Omega$, due to the vanishing of the rest of the partial derivatives.
Recall that there is a natural $\overline{K}$-valued pairing between $TX$ and $T^{\ast}X \otimes \overline{K}$. Applying this to $d\Omega$ and $N_{\Omega}$, the pairing corresponds to $\left(g_{1\,z}\,\overline{g_{1\,\overline{w}}} + g_{1\,w} \, \overline{g_{1\,\overline{z}}} \right)~\beta_c$. Therefore,
\al{
\left(B(D\widehat{u}, D\widehat{u})\right)^{\perp_{\Omega_0}}
&= \frac{1}{4} \times 4\pair{d\Omega, N_{\Omega}}
= \pair{d\Omega, N_{\Omega}} \\
\left\langle \widetilde{B}(D\widehat{u}, D\widehat{u}), \frac{1}{2}\Omega_0 \right\rangle
&= \frac{1}{16} \times 4 \left(\abs{N_{\Omega}}^2 - \abs{d\Omega}^2 \right) = \frac{1}{4} \left(\abs{N_{\Omega}}^2 - \abs{d\Omega}^2 \right)
}
Substituting in equation \eqref{eq: laplacian of omega}, we have
\eq{
\label{eq: laplacian of omega full}
\nabla^{\ast}\nabla \Omega
= -\left(\frac{s_{\scriptscriptstyle X}(g''_{\scriptscriptstyle X})}{2} + \lambda\right) \Omega + \frac{1}{2} \left(\abs{d\Omega}^2 - \abs{N_{\Omega}}^2 \right) \Omega - 2 \pair{d\Omega, N_{\Omega}}
}
Also, observe that $\abs{\nabla\Omega}^2 = \abs{d\Omega}^2 + \abs{N_{\Omega}}^2$. The statement of the theorem follows from eq. \eqref{eq: laplacian of omega full} and eq. \eqref{eq: eq1 for lemma for thm 2}.
\end{proof}
\section{Some Remarks}
For the usual Seiberg-Witten equations, Donaldson remarks that for a fixed metric, the Seiberg-Witten equations are in one-to-one correspondance with solutions to the following equations
\begin{multline}
\label{eq: full ODE of Donaldson}
\nabla^{\ast}\nabla \Omega = - \left(\frac{s}{2} + \abs{\Omega}^2\right) \Omega - 2 \langle d\Omega + \ast d\abs{\Omega}, N_{\Omega} \rangle + \frac{1}{2} \left(\frac{|d\Omega|^2}{\abs{\Omega}^2} - |N_{\Omega}|^2 \right) \Omega \\
+ \frac{1}{2}\left(|d\abs{\Omega}|^2 + 2\langle d\abs{\Omega}, \ast d\Omega \rangle\right) \frac{\Omega}{\abs{\Omega}^2}
\end{multline}
Many examples of hyperK\"ahler manifolds with requisite properties can be obtained via hyperK\"ahler reduction of flat space. Using Prop. \ref{prop: 1-1 correspondence of equations} and applying Donaldson's arguments, one can show that the Abelian, generalised Seiberg-Witten equations, for a 4-dimensional target hyperK\"ahler manifold, can be expressed as \eqref{eq: full ODE of Donaldson}.
Note that the specification of an almost-complex structure $I$ compatible with $\Omega$ imposes a topological restrictions on $X$. Namely, in terms of the Euler characteristic $\chi$ and the signature $\tau$ of $X$,
\eqst{
c_1^2(L) \, = \, 2\,\chi \, + \, 3\, \tau
}
where $L$ is the line-bundle associated to the determinant bundle $P_{{\rm U}(1)}$. For the usual Seiberg-Witten equations, this is precisely the condition under which the expected dimension of the moduli space is zero. Therefore Theorem \ref{thm: main thm 2}, in combination with Donaldson's result \cite{donaldson} delivers a potential candidate to get a compact moduli space.
The arguments in the latter half of the article can be extended for target hyperK\"ahler manifolds of higher dimensions, using similar techniques. However, in this case, one obtains a map from the moduli space of generalised Seiberg-Witten to the usual Seiberg-Witten equations, which is not one-to-one and may not even be surjective.
| 27f1acc8a056738ec0ad99b7ab9a1e60202d39d3 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:intro}\input{tcns-text/intro}
\section{Problem Formulation}\label{sec:pf}\input{tcns-text/pf}
\section{Localization of a Single Target}\label{sec:st}\input{tcns-text/single}
\section{Localization of Multiple Targets}\label{sec:mt}\input{tcns-text/multi}
\section{Optimal Planning and Resource Allocation for Multiple Robots}\label{sec:many}\input{tcns-text/many}
\section{Numerical Simulations}\label{sec:sim}\input{tcns-text/sims}
\section{Experiments}
\label{sec:exp}
\input{tcns-text/exp}
\section{Conclusion}\label{sec:conc}\input{tcns-text/con}
\bibliographystyle{ieeetr}
\subsection{Noise Modeling}
In this paper, we assume that individual measurements are subject to a known zero mean Normal noise distribution $\bbnu \sim \normal{\bb0, \bbQ}$.
In what follows, we estimate $\bbQ$ so that this assumption holds for the stereo rigs used in our laboratory experiments.
This is critical for a variety of reasons:
\begin{itemize}
\item If the mean of $\bbnu$ is biased, then the KF will not converge to the ground truth.
\item If $\bbQ$ is an under approximation to the actual covariance of $\bbnu$, then the KF will become inconsistent and will not converge to the ground truth, if it converges at all.
\item If our choice of $\bbQ$ is too conservative, it may not be informative enough to be useful in the decision process at the core of the controller.
\end{itemize}
Making things even more difficult, we want to test the system in relatively extreme conditions, particularly at long ranges, when triangulation error distributions are known to be heavy tailed, biased away from zero, and highly asymmetric \cite{freundlich15cvpr}.
In our experiments we make use of the physics of stereo vision \cite{ma2012invitation}.
In particular, following \cite{matthies87} we assume that pixel error are Gaussian, and we propagate them to the localization estimates via triangulation equations, which we can use to give us an accurate distribution for $\bbnu$.
Specifically, if the robot registers a correspondence in the left (L) and right (R) cameras at pixel coordinates $[x_L, x_R, y]^\top$, then the coordinates of the 3-D point that generated the match are
\begin{equation}\label{eq:triangulate}
\mat{X\\Y\\Z}
= \frac{b}{d} \mat{\frac{1}{2}(x_L+x_R) \\ y \\ f},
\end{equation}
where $f$ is the focal distance, $b$ is the stereo baseline, $d = x_L - x_R$ is the disparity.
Let $\bbJ$ denote the Jacobian of \eqref{eq:triangulate}.
If the error in $[x_L, x_R, y]^\top$ has covariance matrix $\bbP,$ then the error covariance of $[X,Y,Z]^\top$ is $\bbJ \bbP \bbJ^\top.$
Moreover, if $\bbt$ and $\bbT$ are the translation vector and rotation matrix from the coordinate frame of the camera to a fixed coordinate frame, then the covariance of $\bbnu$ is $\bbT \bbJ \bbP \bbJ^\top \bbT^\top.$
With this noise propagation formula in mind, we now study errors in the observed pixels.
For this we follow a data-driven approach that we have recently developed in \cite{freundlich16nbv}.
Using a set of $n=600$ pairs of training images for each robot at various ranges and viewing angles, we obtain a regression that maps pixel observations $[x_L, x_R, y]^\top$ to a corrected tuple of pixels $[x_L^c, x_R^c, y^c]^\top$ such that the error in the corrected tuple is zero mean.
For every image pair, we project the ground truth landmark onto the image planes using the inverse mapping of \eqref{eq:triangulate}, giving us $\ell = 1, \dots, n$ individual output vectors $\bbY_\ell$, which we stack into an $n \times 3$ matrix of outputs $\bbY$.
The ground truth data for this training set includes a marker placed on top of the ping pong ball, and the pose information of the stereo rig, captured by a $\bbt $ and $\bbT .$
We then compute five features and, because the data are not centered, include one constant, for each raw pixel tuple according to the model
\begin{equation}\label{eq:model}
\bbX_\ell = \left[
1,\, y_\ell,\, d_\ell,\, x_{{L, \ell}}+ x_{{R, \ell}},\, y d_\ell, \frac{ x_{{L, \ell}}+ x_{{R, \ell}}}{d_\ell}
\right].
\end{equation}
These features are taken from the constituent terms in the nonlinear equation \eqref{eq:triangulate}
Stacking the $\bbX_\ell$ into an $n \times 6$ matrix, we have a linear model $\bbY = \bbX \bbbeta + \bbepsilon,$ where $\bbbeta$ is a $6 \times 3$ matrix of coefficients and $\bbepsilon$ is an $n \times 3$ matrix of zero-mean Normal errors.
We experimentally verified that the rows of $\bbepsilon$ are roughly Gaussian, as can be seen in the right panel of Fig.~\ref{fig:px_errors}.
We refer to the raw pixels as \emph{uncorrected}.
The associated error vectors (computed with respect to the uncorrected pixels and the projected ground truth) $\bbepsilon_\ell^\text{uc}$ for $\ell = 1, \dots, n$ are plotted in the left panel of Fig~\ref{fig:px_errors}.
In the scatter plot it can be seen that the mean error is nonzero, contributing average bias to individual measurements.
Also note the apparent skew of the error distribution in the vertical ($y$) direction.
Applying the ordinary least squares estimator, the maximum likelihood estimate of the coefficient matrix is $\hat{\bbbeta} = (\bbX^\top \bbX)^{-1} \bbX^\top \bbY.$
Using $\hat{\bbbeta},$ the residual covariance in the pixel measurements for the two robots, named ${{\rm r1}} \; \and \; {{\rm r2}}$, we obtained are
\agn*{
\bbP_{\rm r1} = \mat{
0.13& 0.09& 0.02\\
0.09& 0.13& -0.03\\
0.02& -0.03& 0.28}\!\!,
\bbP_{\rm r2} = \mat{
0.22& 0.16& 0.04\\
0.16& 0.23& 0.03\\
0.04& 0.03& 0.74}.
}
Note that the standard deviation of the $y$ pixel value, corresponding to the variances in the lower right entries of the above matrices, is 0.53 and 0.86 pixels, respectively.
This corresponds to errors in the height of the ping pong ball center in vertical coordinates.
The right panel of Fig.~\ref{fig:px_errors} shows the residual errors in the training set $\bbepsilon_\ell$ for $\ell = 1, \dots, n$ for the corrected vector $X \hat{\bbbeta}$ on ${\rm r1}$.
For prediction, if ${\rm r1}$ makes a new observation $(x^*_{L}, x^*_{R}, y^*)$, it forms a $1 \times 6$ vector $\bbX^*$.
Then, ${\rm r1}$ calculates the corrected pixels $[x_L^c, x_R^c, y^c]^\top = \bbX^* \hat{\bbbeta},$ which are subject to zero mean Normal errors.
Using the corrected pixels, ${\rm r1}$ triangulates the relative location of the target via \eqref{eq:triangulate}, propagates $\bbP_{\rm r1}$ via the Jacobian $\bbJ,$ rotates and translates the estimates to global coordinates, and thus the assumptions that the error terms $\bbnu$ are zero mean with covariance $\bbQ= \bbT \bbJ \bbP_{\rm r1} \bbJ^\top \bbT^\top$ are approximately satisfied.
The case for ${\rm r2}$ is analogous.
\begin{figure}[t]
\centering
\begin{tabular}{c c}
\includegraphics[height=4cm]{figures/scat-uncorrected.pdf}
&
\includegraphics[height=4cm]{figures/scat-corrected.pdf}
\end{tabular}
\caption{Scatter plots of the residual errors $\bbepsilon_\ell^\text{uc}$ (left panel) and $\bbepsilon_\ell$ (right panel) for the training data for ${\rm r1}$
The plots for ${\rm r2}$ are similar.
}
\label{fig:px_errors}
\end{figure}
Finally, note that the regression takes place in three dimensions, whereas the algorithm is designed only for two dimensions.
For planning purposes, we consider only the components of $\bbQ$ that lie on the plane of the workspace, i.e., we do not use the third row and column for planning.
Of course, the experiments take place in three dimensions, so we need to use the full covariance matrices in the Kalman Filters.
Note that the algorithm proposed in this paper could be implemented in 3D at the expense of increasing the pose and covariance spaces accordingly.
\subsection{Literature Review}\label{sec:lit_review}
A typical approach to active state estimation is to employ gradient descent methods to generate sensor trajectories and sequences of associated state observations that minimize an information theoretic objective of interest, such as the trace of the covariance matrix.
This is the approach followed, e.g., in mobile target tracking \cite{chung06,freundlich13cdc}, sparse landmark localization \cite{freundlich15acc, freundlich13icra,vanderhook15}, and active SLAM \cite{carrillo15,atanasov15,agha15}.
Recently, \cite{carrillo15} have shown that information-theoretic objectives may fail even to be monotonic in many active sensing tasks, removing performance guarantees for greedy control.
When planning multiple observations, the robot may need to reason over the combinatorial set of future probability distributions (pdfs), efficiently represent them, and solve Bellman's equation.
This is known as the ``belief representation problem.''
This problem, in the context of information acquisition, has received a great deal of attention recently, both when the robot state is observable\cite{ryan2010particle,adurthi14,atanasov14,hollinger14sampling,LeNy09trajectory} and when it is only partially observable\cite{atanasov14nonmyopic,spaan14,agha2014firm,agha15,kurniawati2016online,seiler15,bai14,omidshafiei15,oliehoek16}.
The most widely used approach is to grow a tree with a prior distribution of the hidden states at the root, sampling in this way several sequences of possible future observations to obtain the set of reachable belief states at the leaves.
Once the tree has been constructed, one simply selects the leaf with the lowest cost and traces it back to the root to obtain the optimal policy.
The horizon length can be set \emph{a priori} \cite{ryan2010particle,adurthi14,atanasov14,atanasov14nonmyopic} or, e.g., defined implicitly by a budget \cite{hollinger14sampling}.
\input{tcns-text/exploration}
Two approaches that are fundamentally different from choosing a dynamic programming horizon are
\begin{inlinelist}
\item to represent the reachable belief space with a finite set in a clever way, typically by making an assumption about the family of distributions of the hidden states\cite{spaan14,freundlich15acc,atanasov14,agha2014firm,vanderhook15,LeNy09trajectory}, or
\item to avoid this representation problem altogether by working in policy space \cite{kurniawati2016online,seiler15,bai14}.
\end{inlinelist}
With respect to the latter, \cite{bai14} defines a \emph{generalized policy graph}, which nonetheless relies on belief space sampling.
In fact, some kind of sampling is at the core of all point-based approaches stemming from the seminal paper \cite{kurniawati08}.
Non-myopic active sensing for teams of decentralized robots often leads to decentralized Partially Observable Markov Decision Process POMDPs (dec-POMDPs) \cite{omidshafiei15, oliehoek16}.
To our knowledge, decentralization in this context refers to the execution of the planner, as its computation is always done offline at a central location due to the non-separable nature of the value function.
Recently, \cite{atanasov16} used sequential planning for decentralized active SLAM.
Common in the vast majority of approaches discussed above is that mobile sensor planning relies on sampling-based strategies, e.g., forward search, for partially \cite{atanasov14nonmyopic,spaan14,agha2014firm,agha15,omidshafiei15, oliehoek16,kurniawati2016online,seiler15,bai14,spaan14}
and fully observable \cite{atanasov14,LeNy09trajectory} robots, or belief space sampling in the policy domain \cite{kurniawati2016online,seiler15,bai14}.
Sampling-based approaches will typically run into scalability issues due to one or more of the following reasons:
\begin{inlinelist}
\item sparsity of information in the environment, which forces longer planning horizons,
\item high dimensional unknowns, which make observation sampling inefficient, or
\item large teams of robots, which significantly increase the size of the action space.
\end{inlinelist}
To design an algorithm that avoids these pitfalls, in this paper we introduce a hierarchical approach
that decomposes the set of $M$ states into $P\ll M$ clusters, designs optimal controllers for each cluster, and then allocates those controllers among the $N$ robots. Specifically, for every hidden state that needs to be estimated,
we define a local Dynamic Program (DP) in the joint state-space
of robot positions and state uncertainties that determines
robot paths and associated sequences of state observations that
collectively minimize the estimation uncertainty. Then, we divide the
collection of hidden states into clusters based on a prior belief
of their geographic locations, and, for each cluster, we define a
second DP that determines how far along the local optimal trajectories
the robot should travel before transitioning to estimating
the next hidden state within the cluster. Finally, a distributed
assignment algorithm is used to dynamically allocate controllers
to the robot team from the set of optimal control policies for every
cluster.
\input{tcns-text/tsp-complex}
We are not aware of any other non myopic method in the literature that can handle as many hidden states and robot sensors.
We also illustrate these claims with experiments on real robots, which are a contribution in and of themselves as the first to demonstrate experimental multi-robot active sensing using stereo vision.
The paper is organized as follows: In Section~\ref{sec:pf}, we formulate the distributed state estimation problem addressed in this paper.
In Sections~\ref{sec:st} and \ref{sec:mt} we propose a local and a cluster DP to obtain controllers that are optimal at the cluster level.
Section \ref{sec:many} presents the distributed auction mechanism that can efficiently allocate clusters to the robots in real-time.
In Section~\ref{sec:sim}, we simulate large robot teams carrying stereo vision rigs that localize hundreds of sparse landmarks.
In Section~\ref{sec:exp}, we report experiments on a team of two robots localizing eight landmarks.
\subsection{Distributed Auction Mechanism}\label{sec:auctions}
In this section we propose a distributed auction method to dynamically and sequentially allocate clusters to robots as they move to localize the whole scene.
\input{tcns-text/auction}
Specifically, let $s=1,2,\dots$ denote a sequence of time instants when the robots communicate with each other, that is in general different from the times $k$ when the robots move, and let $\ccalN_i^s \triangleq \big\{j \mid \|\bbz_i^s-\bbz_j^s\|<\Delta\big\}$ denote the set of neighbors of robot $i$ at time $s$, where $\Delta>0$ denotes a given communication range. Moreover, assume that every robot $i$ carries two lists: the list of `free' clusters $\ccalI_{i,f}^s$ and the list of `taken' clusters $\ccalI_{i,t}^s$, so that $\ccalI_{i,f}^s \cup \ccalI_{i,t}^s = \set{1, \dots, P}$ and $\ccalI_{i,f}^s \cap \ccalI_{i,t}^s = \emptyset$ for all time $s$.
Initially, $\ccalI_{i,f}^0 = \set{1, \dots, P}$ and $\ccalI_{i,t}^0 = \emptyset$. The list of `free' clusters contains clusters that are available to robot $i$, meaning that robot $i$ can select from those clusters a cluster to visit next. On the other hand, the list of `taken' clusters contains clusters that have been selected by other robots and are, therefore, not available to robot $i$. During operation, robot $i$ coordinates with its neighbors $j \in \ccalN_i^s$ to update its list of `taken' and `free' clusters by $\ccalI_{i,t}^{s+1}=\cup_{j \in \ccalN_i^s } \ccalI_{j,t}^s$ and $\ccalI_{i,f}^{s+1}=\{1,\dots,P\}\setminus \ccalI_{i,t}^{s+1}$, respectively. In other words, with every communication round, robot $i$ removes {\blue from} its list of free clusters those clusters that are considered taken by other robots.
Given the list of `free' clusters $\ccalI_{i,f}^s$ at time $s$, robot $i$ can select any cluster from that list to be the next cluster $c_{i,\text{next}}$ to visit. To minimize the total distance travelled by the robots, we propose a greedy approach where robots select a cluster that is the closest to their current location. In particular, we define
\begin{align}\label{eq:cnext}
&c_{i,\text{next}}^{s+1} = \! \begin{cases}
\argmin_{c \in \ccalI_{i,f}^s} d_{p,j_p}^\text{busy}(\ell^s, c) &\text{if }m_i^s = \text{`busy'} \\
\argmin_{c \in \ccalI_{i,f}^s} d^\text{trans} (\bbz_i^s,c) &\text{if }m_i^s = \text{`transit'}
\end{cases}
\\
&
\label{eq:busy}
d_{p,j_p}^{\text{busy}}(\ell^s, c) \triangleq (L_{p,j_p} \! - \! \ell^s) + \! \! \min_{\bbp\in\partial\ccalZ_c} \|\bbxi_{p,j_p}(L_{p,j_p}) \! - \! \bbp\|\\
&\text{and }
\label{eq:trans}
d^{\text{trans}}(\bbz_i^s, c) \triangleq \min_{\bbp\in\partial\ccalZ_c} \|\bbz_i^s-\bbp\|,
\end{align}
In \eqref{eq:busy} and \eqref{eq:trans}, $d_{p,j_p}^{\text{busy}}(\ell^s, c)$ and $d^{\text{trans}}(\bbz_i^s, c)$ are the distances that robot $i$ needs to travel in order to reach a new cluster $c$ from its current location $\bbz_i^s$ while in modes `busy' and `transit', respectively, and $\ell^s=\bbxi_{p,j_p}^{-1}(\bbz_i^s)\in[0,L_{p,j_p}]$, as in \eqref{eq:busynav}. When robot $i$ selects a new cluster $c_{i,\text{next}}^{s}$, then it also updates its lists of `free' and `taken' clusters by removing $c_{i,\text{next}}^{s}$ form $\ccalI_{i,f}^s$ and adding it to $\ccalI_{i,t}^s$.
Every time $c_{i,\text{next}}$ is updated, robot $i$ also places a bid that indicates how important the selection of the new cluster is. The bids are inversely proportional to the distance robot $i$ needs to travel to reach the new cluster, so that nearby clusters have higher value. Specifically, bids are placed according to
\begin{align}\label{eq:bids}
b_i^{s+1} &= \begin{cases}
\max_{c \in \ccalI_{i,f}^s} \left(1+d_{p,j_p}^\text{busy}(\ell^s, c) \right)^{-1} &\text{if }m_i^s = \text{`busy'}, \\
\max_{c \in \ccalI_{i,f}^s} \Big(1+d^\text{trans} (\bbz_i^s,c) \Big)^{-1} &\text{if }m_i^s = \text{`transit'}
\end{cases}.
\end{align}
If at some point in time there exist neighbors $j\in\ccalN_i^s$ of robot $i$ so that $c_{j,\text{next}}^s = c_{i,\text{next}}^s$, then these robots set up a local auction and compare their bids to resolve the underlying conflict. If $b_i^s>b_j^s$ for all $j\in\ccalN_i^s$ for which $c_{j,\text{next}}^s = c_{i,\text{next}}^s$, then robot $i$ wins the auction and maintains the same next cluster and bid, i.e., $c_{i,\text{next}}^{s+1} = c_{i,\text{next}}^s$ and $b_i^{s+1}=b_i^s$. The robots that lose the auction update their set of `free' clusters by removing cluster $c_{j,\text{next}}^s$, i.e., $\ccalI_{j,f}^s=\ccalI_{j,f}^s\setminus \{c_{j,\text{next}}^s\}$, and select a new next cluster and bid according to \eqref{eq:cnext} and \eqref{eq:bids}. If $\ccalI_{i,f}^s=\emptyset$, i.e., if there are no other available clusters for robot $i$, we set $c_{i,\text{next}}^{s+1} = \textrm{`depot'}$, effectively controlling the robot to return to a depot after it has completed its current (final) task.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=11cm]{figures/automoton}
\caption{
A diagram representing the controller for the $i$-th robot.
The two blocks, Navigation and Coordination, synchronize whenever the updates denoted by $\alpha$ and $\beta$ are triggered.
Functions \emph{transit} and \emph{busy} are defined in \eqref{eq:transitnav} and \eqref{eq:busynav}.
}
\label{fig:automaton}
\end{center}
\vspace{-0.5cm}
\end{figure*}
Fig~\ref{fig:automaton} illustrates the integrated, hybrid, controller.
The labels $\alpha$ and $\beta$ mark events that need to be synchronized across the navigation and coordination control blocks.
In particular, transitions labeled by the letter $\alpha$ are triggered by the navigation block and generate synchronous transitions in the coordination block aimed to produce new bids or update the next cluster that the robot needs to visit.
Similarly, transitions labeled by the letter $\beta$ are triggered by the coordination block when new bids are computed or the next cluster that the robot needs to visit is updated and they generate synchronous transitions in the navigation block aimed to guide the robot its new assigned cluster.
Note that while the $k$ and $s$ time indices used for the navigation and coordination blocks, respectively, can in general be different, the transitions labeled by the letters $\alpha$ and $\beta$ can generate transitions in these blocks that can be off-clock.
For example, a transition at time $k$ in the navigation block labeled by $\alpha$ will generate a transition in the coordination block at a time instant $k\neq s$.
\subsection{Active sensing of a single target}\label{sec:st_results}
The local pose space $\ccalW$ for the single target case is made up of concentric spherical shells with randomly distributed viewpoints on each.
There were 177 total views in $\ccalW$ at six equally spaced radii between 20 and 40 units from the landmark.
To discretize the covariance space $\ccalC$, we set $1 = \max \set{\lambda \in \ccalL} \triangleq \beta $.
In total, the number of possible principal eigenvalues was $N_\ccalL = 6$.
We also use $N_\ccalA = 3$ condition numbers.
We set the number of principal eigenvectors for covariances with $\beta$ as the principal eigenvalue to $N_{\ccalT_\lambda} = 98.$
This means $\ccalT_\beta$ has 98 unit vectors.
The number of samples at other principal eigenvectors $\set{\lambda \in \ccalL \mid \lambda \neq \beta}$ was $
\ceil{
\frac{\lambda}{\beta}
N_{\ccalT_\beta}
}.$
For the logspace discretizations, we set $\kappa_\ccalL=9,$ and $\kappa_\ccalA=3.$
Stereo vision is used as the sensing model; see Appendix~\ref{app:noise}.
The discount factor $\gamma$ was 0.99.
The value of the uncertainty reduction gain $\rho$ is very important, since it controls the tradeoff between traveling cost and uncertainty reduction due to more images.
Fig.~\ref{fig:error} shows the sensitivity of the total reward gained and the two opposing objectives that comprise it as they depend on $\rho$.
Fig.~\ref{fig:rho999} shows example an optimal trajectory in both pose and covariance space with $\rho=0.999.$
\begin{figure}[t]
\centering
\includegraphics[width=8cm]{figures/compare.pdf}
\caption{Showing the effect of the uncertainty gain parameter $\rho$ on the total reward, uncertainty reduction, and total distance traveled by the sensor in the single landmark localization simulations.
Lines are drawn to guide the eye
}
\label{fig:error}
\vspace{-0.0cm}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=2in]{figures/local2.pdf}
\caption{Showing an optimal trajectory through both pose (top panel) and covariance (bottom six panels) space for $\rho = 0.999.$
The indices in the top panel correspond to the time indices in the bottom panel.
Covariance states are shown as 50\% confidence regions.
Robot starts at $\square$ and ends at $\triangle.$
All units in stereo baselines and axis directions equally scaled
}
\label{fig:rho999}
\vspace{-0.0cm}
\end{figure}
In these simulations, we approximate error distributions as state-dependent white noise using the equations given in Appendix~\ref{app:noise}.
White noise is uncorrelated in time, thus, under standard KF assumptions, two observations from the same vantage point will reduce variance by Lemma~\ref{lem:fusion}.
This is why, in the figures, there are more intermediate covariance ellipsoids than steps in the robot trajectories; it is sometimes optimal to remain stationary and take multiple observations.
In real scenarios, noise is seldom well-approximated as white, and duplicate observations will exacerbate bias.
We account for such hidden biases in these simulations by using a real model of stereo vision that is subject to the nonlinear effects of quantization on the image plane; see Appendix~\ref{app:noise}.
Interestingly, the presence of such noise reveals a secondary benefit of sensor mobility:
mobile sensors avoid by default duplicate observations that would plague a static sensor.
Although our controller does not explicitly encourage motion for this reason, a basic heuristic adjustment that does would be simple to implement by, e.g., removing the action ``remain stationary.''
\begin{table}
\centering
\caption{Comparison to Heuristics for Single Landmark}
\label{tab:heur}
\begin{tabular}{| c | c c c c c |}
\hline
Method & optimal & closer & static & circle & random \\ \hline
Reward (m)& 3.15 & 2.75 & 1.31 & 1.66 & 2.74 \\ \hline
\end{tabular}
\end{table}
Table~\ref{tab:heur} compares the total reward gained by a robot starting from the same initial condition under the optimal policy with $\rho=0.999$ found by our (optimal) local controller with some heuristic policies.
The heuristics we chose for comparison were a policy that drives closer to the hidden state (closer), one that circles around the source (circle), one that remains still (static), and one that acts randomly (random).
The fact that that `random' performs about as well as `closer' suggests that a sophisticated controller is needed in order to beat `random.'
Our `optimal' policy obtains 115\% of the reward of the `closer' and `random' policies.
Also note that the `static' and `circle' policies perform very poorly in comparison to the others.
This is because they stay at a constant depth with respect to the landmark.
In triangulation with stereo vision, there is significant bias in the viewing direction, causing poor localization performance given multiple observations at constant depth.
\subsection{Active sensing of Target Clusters}\label{sec:mt_results}
\begin{figure}
\centering
\begin{tabular}{c}
\includegraphics[width=7cm]{figures/clustercompare.pdf}
\\
\includegraphics[width=7cm]{figures/clustercompareerror.pdf}
\end{tabular}
\caption{
Top panel:
Showing trajectories generated by our hierarchical controller and by the method in \cite{atanasov14}.
The trajectories start at the $\square$ and end at the $\triangle$.
The number {\bf k} indicates the number of observations taken.
Bottom Panel: plotting the sum of errors in all landmarks versus the number of observations required found over 100 simulations of observations taken along the trajectories in the top panel.
The midlines represent the empirical mean and error bars represent the standard deviation of the sum of error vectors over the 100 simulations.
All units in meters
}
\label{fig:cluster}
\vspace{-0.0cm}
\end{figure}
We present simulations of the proposed distributed estimation method for a single robot observing a cluster of sparse landmarks.
For the landmarks, we model a famous group of sculptures called the Queens of France and Famous Women, which can be found in the Luxembourg Garden, Paris, France.
The Luxembourg Garden is actually home to hundreds of sculptures, and the Queens of France and Famous Women is a cluster that surrounds a large pool (octagon in Fig.~\ref{fig:cluster}) that is adjacent to the Luxembourg Palace (large rectangle in Fig.~\ref{fig:cluster}).
We chose a subset of eight queens as individual sparse landmarks to comprise a cluster of statues.
The task of the robot in this scenario is to exactly localize each statue to create a precise spatial map of the sculpture garden.
For these simulations, we used a discount factor and gain of 0.999.
We use the same hemispherical pose state-spaces as in Section~\ref{sec:st_results} for each local DP.
We also compare the trajectories generated by our cluster DP with a state-of-the-art algorithm \cite{atanasov14}.
The method in \cite{atanasov14}, called $\epsilon$-$\delta$ Reduced Value Iteration (${\epsilon\text{-}\delta\text{RVI}}$), grows a tree in belief space that computes an optimal sensor trajectory, i.e., it applies the most widely used approach to the same problem that we solve under similar assumptions.
In our implementation of ${\epsilon\text{-}\delta\text{RVI}}$, planning horizons greater than 12 caused our simulation, run on Macbook Air\texttrademark with 4 GB of RAM, to run out of memory.
Because of the sparsity of the scene, 12 step lookahead was sometimes not enough to plan future landmarks to visit.
Therefore, if the robot following ${\epsilon\text{-}\delta\text{RVI}}$ finishes observing a landmark to the threshold, it greedily selects a new landmark.
The top panel in Fig.~\ref{fig:cluster} displays four snapshots of two different trajectories: one produced by the cluster DP optimal policy and one produced using ${\epsilon\text{-}\delta\text{RVI}}$.
The bottom panel of Fig.~\ref{fig:cluster} shows the result of the KF output of 100 simulations of observations taken along the trajectories in the top panel.
The empirical standard deviations for each error vector are also drawn on the figure.
Both methods perform similarly, with our method requiring slightly less observations.
The important distinction, however, is that for every new initial condition, any tree-based planner, including ${\epsilon\text{-}\delta\text{RVI}}$, needs to be run again, whereas we can reuse our optimal policy for any initial condition.
This is important in the next section, as it allows allocation of clusters dynamically in the multi-robot team from a variety of initial conditions along the cluster boundaries.
In other words, our method allows us to compute an \emph{optimal policy}, rather than a single \emph{optimal trajectory}.
\subsection{Multiple Robots and Multiple Clusters}
\begin{figure}[t]
\centering
\includegraphics[width=7cm]{figures/many2.pdf}
\caption{
Cooperative active sensing for 100 sparse landmarks and 15 clusters using 8 robots.
Top panel: Targets are denoted by circles, and clusters are shaded grey.
Squares represent the initial locations of the robots at the bottom of the panel.
Bottom panel: cluster assignment result from the online distributed auction for each of the eight robots.
Colors correspond to robot trajectories in the top panel.
Note that at time 150, the orange robot loses an auction for cluster 3 to the cyan robot, and selects a different cluster 7.
}
\label{fig:mmt}
\end{figure}
We present simulations of the proposed distributed estimation framework for multi-robot multi-target active localization in Fig.~\ref{fig:mmt}.
For these simulations, we uniformly randomly generated targets in a rectangle in $\reals^2.$
The communication range of the robots was set to 1500 m.
\input{tcns-text/clustering}
\subsection{The Uncertainty State-Space and Transition Function}\label{sec:cog}
In this section, we discretize ${\rm Sym}_{++}(n,\reals)$ to design the finite set $\ccalC.$
We emphasize that optimally sampling bounded subsets of ${\rm Sym}(n,\reals)$ is an interesting and deep problem \cite{hardin2004discretizing}, and we do not provide a general framework.
Our method works well for representing a specific bounded region of ${\rm Sym}(n , \reals),$ which we call the \emph{reachable covariance matrices} for the problem described herein.
\begin{figure}[t]
\centering
\begin{tabular}{c c}
\includegraphics[width=2.6cm]{actions4.pdf}
&
\includegraphics[width=3.5cm]{figures/charges.pdf}
\\
(a) & (b)
\end{tabular}
\caption{
(a) An illustration of the state-space and transition for a local DP where $\ccalW$ is a NWSE grid and actions are to take an image and move north, south, east, west, or remain stationary.
Red lines are drawn to represent each action, showing the transition of the state both in $\ccalW$ and in $\ccalC$.
The vertical axis represents the discrete nature three dimensional $\ccalC$ using colored regions, which are each represented by a single matrix in $\ccalC$.
(b) A fifty point discretization of $\mbR\mbP^{2}$ is represented by allowing 100 simulated charges confined to the 2-sphere $\mbS^2$ to come to equilibrium and discarding the points below an arbitrary equator.
Charges are red dots, and the net electric force on each is plotted in blue.
}
\label{fig:discrete_actions}
\vspace{-0.5cm}
\end{figure}
We begin by the following lemmas; proofs are omitted due to space limitations.
\begin{lem}\label{lem:pre-fusion}
Let $\bbI$ denote the $n \times n$ identity matrix and let $\bbC \in {\rm Sym}_{++}(n,\reals)$. Then, $\bbI-(\bbI+\bbC)^{-1} \in {\rm Sym}_{++}(n,\reals)$.
\end{lem}
\begin{lem}[Lemma 2.7, \cite{trace_ineq}] \label{lem:fusion}
Let $n \in \mbN$.
If $\bbA,\bbB \in {\rm Sym}_{++}(n,\reals)$, then
$
{\bf tr} \left( \bbA^{-1} +\bbB^{-1} \right)^{-1} < {\bf tr} \bbA.
$
\end{lem}
The first implication of Lemma~\ref{lem:fusion} is that the trace of the largest instantaneous covariance bounds the maximum eigenvalue of the reachable covariance matrices.
Define this bound as
\input{tcns-text/eig}
Lemma~\ref{lem:fusion} also implies that the trace is decreasing with additional independent measurements.
Therefore, the set of reachable covariances is bounded.
Lemma ~\ref{lem:fusion} and the resulting bound can be used to obtain a discretization of the set of possible maximum eigenvalues for the reachable covariance matrices.
In particular, we define the logspace set of maximum eigenvalues
\begin{equation} \label{eq:mu}
\ccalL \triangleq \set{ {\lambda_\text{max}} e^{\kappa_{\ccalL} \left( i-N_{\ccalL} \right)/N_{\ccalL}} \mid i=1, \dots ,N_{\ccalL} }\subset \reals_{++},
\end{equation}
where $N_{\ccalL}$ is the cardinality of $\ccalL$ and $\kappa_{\ccalL}$ is a sampling gain that controls how clustered the samples are toward zero.
Note that ${\lambda_\text{max}}$ is the maximal element of $\ccalL$.
In \eqref{eq:mu}, we sample in logspace as a heuristic; we have found empirically that the maximum eigenvalues of the filtered covariance matrices accumulate near zero.
To obtain a scalable discretization of the space of covariance matrices ${\rm Sym}_{++}(n , \reals),$ we assume that $\lambda_\text{max} (\bbSigma)$ and its corresponding eigenvector are more important than any one of the other eigenvalues and eigenvectors.
In particular, we assume that, in comparison to $\lambda_\text{max},$
the other eigenvalues are roughly equal,
thus all other eigenvalues can be parameterized by a number in the half open interval $\alpha \in (0,1] \subset \reals$ such that $\lambda_i \approx \alpha \lambda_\text{max}$ for all $i = 2, \dots, n.$
This choice alleviates the need to independently consider all possible combinations of eigenvectors corresponding to the nonprincipal eigenvalues.
Define the set of ratios, which can be thought of as the set of possible inverse condition numbers, as
\begin{equation} \label{eq:ratiovec}
{\ccalA } \triangleq \set{ e^{\kappa_{\ccalA } \left( i-N_{\ccalA } \right)/N_{\ccalA }} \mid i =1, \dots, N_{\ccalA } }\subset (0,1].
\end{equation}
In \eqref{eq:ratiovec}, $N_{\ccalA }$ is the number of eigenvalue ratios we sample and $\kappa_{\ccalA }$ is a sampling gain that controls the ellipticity of the confidence region associated with $(\lambda,\alpha) \in \ccalL\times\ccalA.$
Again, the logspace discretization is a heuristic based on experience; KF error covariance matrices produced using robotic sensors are typically cigar-shaped, i.e., dominated by uncertainty in the direction of the principal eigenvector.
The set of possible principal eigenvectors is equivalent to the set of lines passing through the origin, known as the real projective space $\mbR \mbP^{n-1}\!\!\!\!\!.\;$
Let $N_\ccalT$ be the number of samples needed to capture these possible directions for the principal eigenvalue.
Because $\mbR \mbP^{n-1}$can be formed by identifying antipodal points on any sphere, this problem can be approximately solved by placing $2N_\ccalT$ point charges on a sphere of radius $\sqrt{{\lambda_\text{max}}},$ allowing them to move until the ``electrostatic forces'' among them come to equilibrium, cutting the sphere along any equator, discarding one of the hemispheres, and saving the unit directions to each ``charge'' location on the other hemisphere.
It is straightforward to build such a simulation, and for brevity we do not provide the specifics here.
The result for 100 charges on the unit 2-sphere is shown in Fig.~\ref{fig:discrete_actions}~(b).
To force the sampling density to be consistent, defined in terms of the surface area of the sphere used in the simulation, we create a set of sets $\set{\ccalT_\lambda \mid \lambda \in \ccalL}$; each element $\ccalT_\lambda$ is the set of unit vectors produced by the simulation using $\sqrt{\lambda}$ as the radius of the sphere.
The number of elements in the largest, i.e., the set with the most elements, of these sets $\ccalT_{\lambda_\text{max}}$ is set to a user-specified number $N_{\ccalT_{\lambda_\text{max}}}$.
The remaining sets $\set{ \ccalT_\lambda \mid \lambda \in \ccalL}$ correspond to the other $N_{\ccalL} -1$ possible principal eigenvalues and, since ${\lambda_\text{max}}$ is the maximal element of $\ccalL,$ the sets $\set{\ccalT_\lambda \mid \lambda \neq {\lambda_\text{max}}}$ must have fewer elements than $\ccalT_{\lambda_\text{max}}$ so that the sampling density is the same.
In particular, for some $\lambda \in \ccalL,$ the set $ \ccalT_\lambda$ has
$
\ceil{
\frac{\lambda}{{\lambda_\text{max}}}
N_{\ccalT_{\lambda_\text{max}}}
}$ elements, where the ceiling function is used to ensure that the number of elements in $\ccalT_\lambda$ is positive.
Using the sets $\ccalL, \ccalA,$ and $ \set{\ccalT_\lambda\mid \lambda \in \ccalL},$ we can create a discretization of ${\rm Sym}_{++} (n, \reals)$.
The result, interpreted geometrically, is $N_\ccalL$ concentric sets of cigar-shaped confidence ellipsoids with a variety of major diameters, defined by $\lambda \in \ccalL,$ ellipticities $\alpha \in \ccalA,$ and orientations $\bbu \in \ccalT_\lambda.$
The full covariance space can thus be described with a map $f \colon \reals^{2+n} \to {\rm Sym} (n, \reals)$ given by
\begin{equation}\label{eq:covmap}
f(\lambda, \alpha,\bbu)= \lambda \mat{\bbu & *}
\mat{1 & \bb0 \\ \bb0 & \alpha\bbI_{n-1}} \mat{\bbu & *}^\top,
\end{equation}
where $*$ is any basis completion for $\reals^n$, and $\bbI_{n-1}$ is the identity matrix.
The function $f$ essentially builds a covariance matrix from the parameters supplied by $\ccalA$ and $\set{(\lambda, \ccalT_\lambda) \mid \lambda \in \ccalL}.$
\input{tcns-text/cov-set.tex}
\noindent The $\bb0$ covariance is an artificial state that we include in $\ccalC $ to denote that no more uncertainty remains in the variable being estimated, i.e., estimation is complete to the user-specified tolerance, defined as $\frac{1}{2} \min \set{\lambda \in \ccalL}.$
The projection operator $\mathsf{\Pi}_{\ccalC }$ guarantees that the fusion of the current covariance state $\bbSigma_k$ and the new measurement covariance $\bbQ_k$ is a member of $\ccalC.$
In particular, for some $\bbSigma \in {\rm Sym}_{++} (n ,\reals),\mathsf{\Pi}_{\ccalC}$ first computes the principal eigenvalue $\lambda_\text{max}$ and its corresponding normalized eigenvector $\bbu_\text{max}$.
Then, it rounds $\lambda_\text{max}$ to the closest element in $\ccalL \cup \set{0}.$
Call this map $\mathsf{\Pi}_\ccalL : {\rm Sym}(n,\reals) \to \ccalL \cup\set{0}.$
If $\mathsf{\Pi}_\ccalL(\bbSigma)$ is nonzero,
there will be some $\lambda ' \in \ccalL$ that is closest to $\lambda_\text{max},$
and $\mathsf{\Pi}_\ccalC$ then finds the element $\bbu' \in \ccalT_{\lambda'}$ that forms the largest magnitude inner product with $\bbu_\text{max}.$
Call this map $\mathsf{\Pi}_{\ccalT_{\lambda'}} : {\rm Sym}(n,\reals) \to \mbS^{n-1},$
where $\mbS^{n-1} = \set{\bbu \in \reals^n \mid \left\langle\bbu,\bbu \right\rangle =1}.$
In particular,
$
\bbu' =
\mathsf{\Pi}_{\ccalT_{\lambda'}}(\bbSigma) \triangleq \max_{\bbu \in \ccalT_{\lambda'}} \abs{\left\langle\bbu, \bbu_\text{max} \right\rangle }.
$
Finally, $\mathsf{\Pi}_\ccalC$ computes the ratio of $\lambda_\text{min}(\bbSigma)$ with $\lambda_\text{max}(\bbSigma)$ and finds the closest element $\alpha' \in \ccalA$ to that ratio.
Call this map $\mathsf{\Pi}_\ccalA : {\rm Sym}_{++}(n ,\reals) \to (0,1].$
By this construction, it holds that $(\alpha', \lambda', \bbu') \in \ccalA \times \bigcup\nolimits_{\lambda \in \ccalL}
\left( \lambda \times \ccalT_\lambda \right),$
so that its image of this triplet under $f$ from \eqref{eq:covmap} is guaranteed to be a matrix in $\ccalC$.
In particular, the projection map is given by
$
\mathsf{\Pi}_\ccalC (\bbSigma) = f
\left(
\mathsf{\Pi}_{\ccalL } \left(\bbSigma \right),
\mathsf{\Pi}_{\ccalA} \left(\bbSigma\right),
\mathsf{\Pi}_{\ccalT_{\mathsf{\Pi}_{\ccalL } \left(\bbSigma \right)}}(\bbSigma)
\right),
$
\subsection{Procedure}
\begin{figure}[t]
\centering
\includegraphics[width=7cm]{figures/experiment-above.pdf}
\caption{Overhead photograph of the experimental setup
}
\label{fig:above}
\end{figure}
Fig.~\ref{fig:above} shows the experimental setup for the multi robot, multi landmark localization experiment.
We place eight colored ping pong balls in a square workspace of about 2 m side length for the multi robot, multi landmark localization experiment.
We use three types of local workspaces: section, semicircle, and full circular polar grids.
To avoid collisions between robots and ping pong balls and avoid overlapping local workspaces, the minimum and maximum radii were set to 30 and 50 cm for all local workspaces, respectively.
The available poses within each local workspace were located at five equally spaced radii between these two extrema.
The polar grids were also divided into increments of 15$^\circ$.
We set the covariance space parameters to $N_{\ccalL}=10,N_{\ccalA}=3, \kappa_\ccalL=10,$ and $ \kappa_\ccalA=7.$
We set $N_{\ccalT}$ to be twice the number of viewing angles in the local workspaces
For these experiments, we set $\rho=10^{-2}$ and $\gamma=0.9$.
In `transit' mode, a potential field algorithm guides the robots to the next local workspace and avoids collisions with landmarks.
All navigation and waypoint tracking relies on a PID controller using the next waypoint as the set point.
In the experiment, robots generally came within 2 cm of their target waypoints.
The servo guided the stereo cameras toward the estimate of the target locations with accuracy of $\pm 1^\circ$.
Cooperation of ${{\rm r1}}$ and ${{\rm r2}}$ relies on robots updating individual ROS topics and checking neighbors' ROS topics.
The ROS topic for a given robot contains that robot's current bid value and the subtask at which the bid is directed.
\begin{figure}[t]
\centering
\includegraphics[width=6cm]{figures/exp-errs.pdf}
\caption{
Plotting the filtered target localization error in meters for ${\rm r1}$ (a) and ${\rm r2}$ (b).
The horizontal axis is the number of observations for each. This means that each robot took 28 observations total, localizing four targets each.
}
\label{fig:exptraj-err}
\end{figure}
The filtered estimates generated by observations along the paths followed by the robots in one run of the experiments are compared with the ground truth locations of the ping pong balls in Fig.~\ref{fig:exptraj-err}.
The filtered errors are computed as the Euclidean distances between filtered estimates and the ground truth locations of the ping pong balls.
We note that any non-decreasing aspects of these plots are attributed to stochasticity in the dataset, and the fact that we are only approximating the true noise distribution with a data-driven Gaussian.
Note that, due to the relatively constrained space in our lab, the robots are not able to move as freely as in the simulations, which, in addition to unmodeled noise, is responsible for the smaller error reduction in Fig.~\ref{fig:exptraj-err} compared to the simulations.
| 4e4425d2c08f5f864a44c81ce442e95fd2cb58a9 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The Ehrhart polynomial of a lattice polytope counts the number of lattice points in its integer dilates and is arguably the most fundamental arithmetic invariant of a lattice polytope. It is a cornerstone of geometric combinatorics and appears in various guises in other areas of mathematics such as commutative algebra, optimization, representation theory, or voting theory (see, e.g., \cite{Barvinok,Berenstein,DeLoeraetal,Lepelley,MillerSturmfels}). Concepts from Ehrhart theory have been generalized in various directions; for example, $q$-analogs of Ehrhart polynomials \cite{chapoton2016q}, equivariant versions \cite{Stapledonequivariant}, multivariate extensions \cite{beck2001multidimensional,bihan2016irrational,haase2015mixed}, and generalizations to valuations \cite{jochemkocombinatorial,jochemko2016combinatorial,McMullen77}.
Recently, Ludwig and Silverstein \cite{LS} introduced Ehrhart tensor polynomials based on discrete moment tensors that were defined by B\"or\"oczky and Ludwig~\cite{boroczkyvaluations}. Let $\mathcal{P}({\Z^d})$ denote the family of convex polytopes with vertices in $\mathbb{Z}^d$ called \textbf{lattice polytopes} and let $\mathbb{T}^r$ be the vector space of symmetric tensors of rank $r$ on $\mathbb{R}^d$. For $x,y \in \mathbb{R}^d$, we write $xy$ for $x\otimes y$. In particular, $x^r=x\otimes \cdots \otimes x$ and we set $x^0 :=1$.
\goodbreak
The \textbf{discrete moment tensor of rank $r$} of a polytope $P\in\mathcal{P}({\Z^d})$ is
\begin{equation}\label{eq:discretemoment}
\operatorname{L}^{r}(P) \ = \ \sum_{x\in P\cap\mathbb{Z}^d}x^{r}
\end{equation}
where $r\in\mathbb{N}$ and $\mathbb{N}$ denotes the set of nonnegative integers.
Note that, for our convenience, this definition differs by a scalar from the original definition given in~\cite{boroczkyvaluations}. A version of $\operatorname{L}^{r}(P)$, the discrete directional moment, was studied in~\cite{Schulz}.
For $r=0$, the usual \textbf{discrete volume} or \textbf{lattice point enumerator} $\operatorname{L}(P):=\operatorname{L}^0 (P) = |P\cap \mathbb{Z}^d|$ is recovered. For $r=1$, $\operatorname{L}^1(P)$ equals the discrete moment vector defined in~\cite{boroczky2016minkowski}. Based on results by Khovanski{\u\i} and Pukhlikov~\cite{PK92} and Alesker\cite{Alesker98}, it was identified in~\cite{LS} that $\operatorname{L}^r (nP)$ is given by a polynomial, for any $n\in\mathbb{N}$, extending Ehrhart's celebrated result for the lattice point enumerator~\cite{Ehrhart62}.
\begin{thm*}[{\cite[Theorem 1]{LS}}]
There exist $\operatorname{L} _i ^r \colon \mathcal{P}(\mathbb{Z} ^d) \rightarrow \mathbb{T}^r$ for all $1\leq i\leq d+r$ such that
\[
\operatorname{L} ^r(nP) \ = \ \sum _{i=0}^{d+r} \operatorname{L} _i ^r (P)n^i
\]
for any $n\in\mathbb{N}$ and $P\in \mathcal{P}(\mathbb{Z} ^d)$.
\end{thm*}
\noindent The expansion of $\operatorname{L}^r(nP)$ will be denoted as $\operatorname{L}_P^r(n)$ and is called the \textbf{Ehrhart tensor polynomial} of $P$ in commemoration of this result. Furthermore, the coefficients $\operatorname{L}^r_1,\dots,\operatorname{L}^r_{d+r}$ are the \textbf{Ehrhart tensor coefficients} or \textbf{Ehrhart tensors}.
A fundamental and intensively studied question in Ehrhart theory is the characterization of Ehrhart polynomials and their coefficients.
The only coefficients that are known to have explicit geometric descriptions are the leading, second-highest, and constant coefficients for the classic Ehrhart polynomial (see, e.g.,~\cite{ccd}). For the Ehrhart tensor polynomial, the leading and constant coefficients were given in~\cite{LS} and we give an interpretation for the second-highest coefficient (Proposition~\ref{prop:secondcoeff}) as the weighted sum of moment tensors over the facets of the polytope; the descriptions of all are given in Section~\ref{sec:discretemoment}.
Conversely, for lattice polygons, the coefficients of the Ehrhart polynomial are positive and well-understood. They are given by Pick's Formula~\cite{Pick}. Let $\partial P$ denote the boundary of the polytope $P$, for any $P\in\mathcal{P}({\Z^d})$.
\begin{thm*}[Pick's Formula]
For any lattice polygon $P$, we have
\[
\operatorname{L}(nP) \ = \ \operatorname{L} _0(P)+\operatorname{L} _1(P)n+\operatorname{L} _2(P)n^2 \,
\]
where $\operatorname{L} _0(P)=1$, $\operatorname{L} _1(P)=\tfrac{1}{2}\operatorname{L}(\partial P)$, and $\operatorname{L} _2(P)$ equals the area of $P$.
\end{thm*}
In Section~\ref{sec:pick}, we determine Pick-type formulas for the discrete moment vector and matrix. Our interpretation of the coefficients is given with respect to a triangulation of the respective polygon. The principal tool we use to study Ehrhart tensor polynomials are \textbf{$h^r$-tensor polynomials} which encode the Ehrhart tensor polynomial in a certain binomial basis. Extending the notion of the usual Ehrhart $h^\ast$-polynomial, we consider
\begin{equation}
\operatorname{L}^r(nP) \ = \ h^{r}_0 (P){n+d+r \choose d+r}+h^{r}_1 (P){n+d+r-1 \choose d+r}+ \ \cdots \ + h^{r}_{d+r} (P){n \choose d+r} \,
\end{equation}
for a $d$-dimensional lattice polytope $P$
and define the \textbf{$h^r$-tensor polynomial} of $P$ to be
\[
h^{r}_P (t)\ = \ \sum _{i=0}^{d+r}h^{r}_i(P)t^i.
\]
\goodbreak
We determine a formula for the $h^r$-tensor polynomial of half-open simplices (Theorem~\ref{prop:h_cone}) by using \textbf{half-open decompositions of polytopes}; an important tool which was introduced by K\"oppe and Verdoolaege~\cite{koppeverdoolaege}. From this formula and the existence of a unimodular triangulation, we deduce an interpretation of all Ehrhart vectors and matrices of lattice polygons.
Stanley's Nonnegativity Theorem \cite{RS80} is a foundational result which states that all coefficients of the $h^\ast$-polynomial of a lattice polytope are nonnegative. Stanley moreover proved that the coefficients are monotone with respect to inclusion; that is, for all lattice polytopes $Q \subseteq P$ and all $0\leq i\leq d$, it holds that $h^{\ast}_i(Q)\leq h^{\ast}_i(P)$. Using half-open decompositions, it was proven in \cite{jochemkocombinatorial} that, with regard to translation invariant valuations, monotonicity and nonnegativity are equivalent.
In Section~\ref{sec:positivity}, we discuss notions of positivity for Ehrhart tensors and investigate Ehrhart tensor polynomials and $h^2$-tensor polynomials with respect to \textbf{positive semi\-definiteness}. In contrast to the usual Ehrhart polynomial, Ehrhart tensor coefficients can even be negative definite for lattice polygons (Example~\ref{ex:nonpositiveEhrharttensors}). Moreover, the coefficients of $h^2$-tensor polynomials are not monotone which is demonstrated by Example~\ref{ex:counter_ex_monotonicity}. Therefore, techniques such as irrational decompositions and half-open decompositions that have been used to prove Stanley's Nonnegativity Theorem (see \cite{ccd,jochemkocombinatorial}) can not immediately be applied to $h^2$-tensor coefficients. Nevertheless, considering an intricate decomposition of lattice points inside a polygon, we are able to prove positive semi-definiteness of the coefficients of $h^2$-tensor polynomial in dimension two (Theorem~\ref{thm:main}). We remark here that the theorem holds true for lattice polygons in a higher dimensional ambient space. Furthermore, all of the results given in this article are independent of the ambient space. Based on computational results, we further conjecture positive-semidefiniteness of the $h^2$-tensor coefficients in higher dimensions (Conjecture~\ref{conj:pos}).
In Section~\ref{sec:further}, we prove a generalization of Hibi's Palindromic Theorem~\cite{Hibi91} characterizing reflexive polytopes as having palindromic $h^r$-tensor polynomials for $r\in\mathbb{N}$ of even rank and conclude by discussing possible future research directions.
\goodbreak
\section{Discrete moment tensors}\label{sec:discretemoment}
We introduce some general notions we will use here yet assume basic knowledge of polyhedral geometry and, in particular, lattice polytopes. For further reference, we recommend~\cite{ccd,ziegler}.
\goodbreak
We work in $d$-dimensional Euclidean space, $\mathbb{R}^d$, equipped with the scalar product $u\cdot v$, for any $u,v\in\mathbb{R}^d$. The vector space of symmetric tensors $\mathbb{T}^r$ is then canonically isomorphic to the space of multi-linear functionals from $(\mathbb{R}^d)^r$ to $\mathbb{R}$ that are invariant with respect to permutations of the arguments. We have $\mathbb{T}^0=\mathbb{R}$ and can now identify $\mathbb{T}^1$ with $\mathbb{R}^d$. Given the standard orthonormal basis $e_1,\dots,e_d$, any tensor $T\in\mathbb{T}^r$ can be written uniquely as
\[
T\ = \ \sum_{1\leq i_j\leq d}T_{i_1\dots i_r}e_{i_1}\otimes\dots\otimes e_{i_r}.
\]
For $r=2$, the bilinear form $T\in\mathbb{T}^2$ can then be identified with a symmetric $d\times d$ matrix $T=(T_{ij})$. To that end, we will call the discrete moment tensor~(\ref{eq:discretemoment}) of ranks 1~and~2 the \textbf{discrete moment vector} and \textbf{discrete moment matrix}, respectively. We will also regard their associated coefficients, their Ehrhart tensors, as \textbf{Ehrhart vectors} and \textbf{Ehrhart matrices}.
\goodbreak
Prior to describing the known Ehrhart tensors, we provide some properties of the discrete moment tensor that we will need. Considering $\operatorname{L}^r$ with respect to its coordinates, for any $P\in\mathcal{P}({\Z^d})$, gives
\begin{equation*}
\operatorname{L}^r(P) (e_{i_1}, \dots, e_{i_r})\ = \ \sum_{x\in P\cap\mathbb{Z}^n} (x\cdot e_{i_1}) \cdots (x\cdot e_{i_r}).
\end{equation*}
\noindent Hence the action of $\mathrm{GL} (\mathbb{Z}^d)$, the general linear group over the integers, on $\operatorname{L}^r$ is observed to be
\begin{equation*}
\operatorname{L}^r(\phi P) (e_{i_1}, \dots, e_{i_r})\ = \ \operatorname{L}^r(P) (\phi^t e_{i_1}, \dots, \phi^t e_{i_r})
\end{equation*}
for any $P\in\mathcal{P}({\Z^d})$ and $\phi\in\mathrm{GL} (\mathbb{Z}^d)$; we say that $\operatorname{L}^r$ is \textbf{$\mathrm{GL} (\mathbb{Z}^d)$ equivariant}.
We let $P^o$ denote the relative interior of $P$ with respect to its affine hull, denoted by $\aff (P)$, and for any $P\in\mathcal{P}({\Z^d})$ and $r\in\mathbb{N}$ we set
\[
\operatorname{L}^r(P^o)\ := \ \sum _{x\in P^o} x^r.
\]
For the discrete volume, the Ehrhart-Macdonald reciprocity was a fundamental result in Ehrhart theory that was established by Ehrhart~\cite{Ehrhart62} and first proven by Macdonald~\cite{Macdonald71}.
\begin{thm}{\cite{Ehrhart62,Macdonald71}}
If $P$ is a $d$-dimensional lattice polytope, then
\[
\operatorname{L}(nP^o)\ = \ (-1)^d\operatorname{L}_P(-n).
\]
\end{thm}
\noindent A general version of this result was given for translation invariant valuations by McMullen~\cite{McMullen77}. Unlike the discrete volume, the discrete moment tensor varies under translations by elements in~$\mathbb{Z}^d$. More precisely, for all $r\in\mathbb{N}$, the discrete moment tensor of a translated polytope is
\[
\operatorname{L}^{r}(P+t) \ = \ \sum _{j=0} ^r {r\choose j}\operatorname{L}^{r-j}(P)t^j \,
\]
and we say that the discrete moment tensor is covariant with respect to translations or \textbf{translation covariant}. A \textbf{unimodular transformation} of a polytope $P\in\mathcal{P}({\Z^d})$ is a $\mathrm{GL} (\mathbb{Z}^d)$ transformation of $P$ paired with a translation.
Similar to McMullen, a reciprocity theorem was given for translation covariant valuations in~\cite{LS}. Extending the classical Ehrhart-Macdonald reciprocity, the following reciprocity theorem gives the special case of the discrete moment tensor.
\begin{thm}{\cite[Theorem 2]{LS}}\label{thm:tensorreciprocity}
Let $P$ be lattice polytope. Then
\[
\operatorname{L} ^r_P (-n) \ = \ (-1)^{\dim (P)+r}\operatorname{L}^r (nP^o) \, .
\]
\end{thm}
\noindent We use this theorem in our characterization of the second-highest Ehrhart tensor.
A complete characterization of the Ehrhart coefficients has been inaccessible up to this point. The coefficients can even be negative and, therefore, are difficult to describe combinatorially. However, it is known that the leading coefficient equals the volume, the second highest coefficient is related to the normalized surface area, and the constant coefficient is always $1$.
\goodbreak
More generally, for Ehrhart tensors, it has been proven \cite[Lemma 26]{LS} that the leading coefficient of the discrete moment tensor equals the \textbf{moment tensor of rank $r$} which is defined as
\[
\operatorname{M}^r (P) \ = \ \int _P x^r \, \diff x \, .
\]
It is also clear that, for $r\geq 1$, the constant coefficient vanishes identically by its $\mathrm{GL} (\mathbb{Z}^d)$ equivariance; that is, $\operatorname{L}_0^r(P)=\operatorname{L}^r(0P)=0$ for any $P\in\mathcal{P}({\Z^d})$~\cite{LS}.
We give an interpretation for the second coefficient (Proposition~\ref{prop:secondcoeff}) as the weighted sum of moment tensors over the facets of the polytope. The coefficient $\operatorname{L}_{d-1}(P)$, specifically, was shown to be equal to one half of the sum over the normalized volumes of the facets of $P$ by Ehrhart~\cite{Ehrhart67}. We extend this statement to Ehrhart tensor polynomials by proving the following.
\begin{prop}\label{prop:secondcoeff}
Let $P$ be a lattice polytope. Then
$$
\operatorname{L}^r_{\dim (P)+r-1}(P) \ = \ \sum _{F} \frac{1}{|\det (\aff (F)\cap \mathbb{Z}^d)|}\int _F x^r \, \diff x \, ,
$$
where the sum is over all facets $F \subset P$.
\end{prop}
\goodbreak
\begin{proof}
Theorem~\ref{thm:tensorreciprocity}, on the one hand, implies
\[
\sum _{x\in \partial nP} x^r \ = \ \sum _{F\subsetneq P} \sum _{x\in nF^o} x^r \ = \ \sum _{F\subsetneq P} (-1)^{\dim (F) +r}\operatorname{L} ^r_F (-n) \, ,
\]
where the sum is taken over all proper faces $F\subsetneq P$. On the other hand, we have
\begin{align*}
\sum _{x\in \partial nP} x^r\ &= \ \operatorname{L}^r (nP)-\operatorname{L}^r (nP^o)
\ = \ \operatorname{L}^r (nP)-(-1)^{\dim (P) +r}\operatorname{L}^r_P (-n)\\
&= \ 2\sum _{i\geq 0}\operatorname{L}^r _{\dim (P) +r -1-2i} (nP) \,
\end{align*}
where we set $\operatorname{L}^r_i=0$ for all $i<0$. Using both equations, we obtain
\begin{align*}
\operatorname{L} ^r _{\dim (P) +r-1} (P)\ &= \ \lim _{n\to \infty} \frac{1}{n^{\dim (P) +r -1}} \sum _{i\geq 0}\operatorname{L}^r _{\dim (P) +r -1-2i} (nP)\\
&= \ \frac{1}{2}\sum _{F\subsetneq P} (-1)^{\dim (F) +r}\lim _{n\to \infty} \frac{1}{n^{\dim (P) +r -1}}\operatorname{L} ^r_F (-n)\\
&= \ \frac{1}{2}\sum _{F \text{ facet}} \frac{1}{|\det (\aff (F)\cap \mathbb{Z}^d)|}\int _F x^r\, \diff x \, ,
\end{align*}
where the last equality follows from~\cite{LS}.
\end{proof}
\section{$h^r$-tensor polynomials}\label{chapter:h^r}
Let $P$ be a $d$-dimensional lattice polytope. Since $\operatorname{L} ^r(nP)$ is a polynomial of degree at most $d+r$, it can be written as a linear combination of the polynomials ${n+d+r \choose d+r}, {n+d+r-1 \choose d+r},\ldots,{n \choose d+r}$, that is,
\begin{equation}\label{eq:expansion}
\operatorname{L}^r(nP) \ = \ h^{r}_0 (P){n+d+r \choose d+r}+h^{r}_1 (P){n+d+r-1 \choose d+r}+ \ \cdots \ + h^{r}_{d+r} (P){n \choose d+r} \,
\end{equation}
for some $h^{r}_0(P),\ldots,h^{r}_{d+r}(P)\in \mathbb{T}^r$. Equivalently, in terms of generating functions,
\begin{equation}\label{eq:h_cone}
\sum _{n\geq 0} \operatorname{L}^r (nP)t^n \ = \ \frac{h^{r}_0(P)+h^r_1(P)t+\dots+h^r_{d+r}(P)t^{d+r}}{(1-t)^{d+r+1}} \, .
\end{equation}
We call $h^{r}(P)=(h^{r}_0(P),h^{r}_1(P),\ldots,h^{r}_{d+r}(P))$ the \textbf{$h^r$-vector}, its entries the \textbf{$h^r$-tensor coefficients} or \textbf{$h^r$-tensors} of $P$, and
\[
h^{r}_P (t)\ = \ \sum _{i=0}^{d+r}h^{r}_it^i
\]
the \textbf{$h^r$-tensor polynomial} of $P$. Observe that for $r=0$ we obtain the usual $h^\ast$-polynomial and $h^\ast$-vector of an Ehrhart polynomial. By evaluating equation~\eqref{eq:expansion} at $n=0,1$, we obtain $h_0^{r}=0$ for $r\geq 1$ and $h_1^{r}=\operatorname{L}^r (P)$ for $r\geq 0$. Inspecting the leading coefficient, we obtain
\[
h_1^{r}(P)+h_2^{r}(P)+\ldots+h_{d+r}^{r}(P)\ = \ (d+r)!\int_Px^r\mathrm{d}x \, .
\]
Applying Theorem~\ref{thm:tensorreciprocity} and evaluating at $n=1$, we obtain
\[
h^r_{d+r} (P)\ = \ \operatorname{L}^r (P^o) \, .
\]
\subsection{Half-open polytopes}
We will not only consider relatively open polytopes, but also \textbf{half-open polytopes}. Let $P$ be a polytope with facets $F_1,\ldots,F_k$ and let $q$ be a generic point in its affine span $\aff (P)$. Then a facet $F_i$ is \textbf{visible} from $q$ if $(p,q]\cap P = \emptyset$ for all $p\in F$. If $I_q (P)=\{i\in [k] \colon F_i \text{ is visible from }q\}$ then the point set
\[
H_q (P) \ = \ P\setminus \bigcup_{i \in I_q (P)} F_i
\]
defines a half-open polytope. In particular, $H_q (P)=P$ for all $q\in P$. The following result by K\"oppe and Verdoolaege~\cite{koppeverdoolaege} shows that every polytope can be decomposed into half-open polytopes, and is implicitely also contained in works by Stanley and Ehrhart (see~\cite{RS74}).
\begin{thm}[\cite{koppeverdoolaege}]\label{thm:half-open decomposition}
Let $P$ be a polytope and let $P_1,\ldots,P_m$ be the maximal cells of a triangulation of $P$. Let $q\in \aff (P)$ be a generic point. Then
\[
H_q (P) \ = \ H_q (P_1)\sqcup H_q (P_2)\sqcup \cdots \sqcup H_q (P_m)
\]
is a partition.
\end{thm}
The discrete moment tensor naturally can be defined for half-open polytopes by setting
\[
\operatorname{L}^r (H_q (P)) \ := \ \operatorname{L}^r(P) - \sum _{J\subseteq I_q(P)} (-1)^{\dim P -\dim F_J} \operatorname{L}^r (F_J) \,
\]
where $F_J := \bigcap _{i \in J} F_i$. Then, from Theorem \ref{thm:half-open decomposition} and the inclusion-exclusion principle, we obtain that
\begin{equation}\label{eq:halfopendec}
\operatorname{L}^r (P) \ = \ \operatorname{L}^r (H_q (P_1)) \ + \ \operatorname{L}^r (H_q (P_2)) \ + \ \cdots \ + \ \operatorname{L}^r (H_q (P_m)) \,
\end{equation}
(Compare also~\cite[Corollary 3.2]{jochemkocombinatorial}).
\subsection{Half-open simplices}\label{sec:half-open}
Let $S$ be a $d$-dimensional lattice simplex with vertices $v_1,\ldots, v_{d+1}$. Let $F_1,\ldots, F_{d+1}$ denote the facets of $S$ such that $v_i \not \in F_i$. Let $S^{\ast}=H_q (S)$ be a $d$-dimensional half-open simplex and let $I=I_q (S)$. We define the half-open polyhedral cone
\[
C_{S^{\ast}} \ = \ \left\{ \sum _{i=1}^{d+1} \lambda _i \bar{v}_i : \lambda _i \geq 0\text{ for }i\in[d+1],\lambda_i\neq 0 \text{ if } i\in I\right\} \subseteq \mathbb{R}^{d+1}
\]
where $\bar{v}_i := (v_i, 1)\in\mathbb{R}^{d+1}$ for all $1\leq i\leq d+1$. Then, by identifying hyperplanes of the form $\{x\in \mathbb{R}^{d+1}\colon x_{d+1}=n\}$ with $\mathbb{R}^{d}$ via $p\colon \mathbb{R}^{d+1}\rightarrow \mathbb{R}^d$ which maps $x\mapsto (x_1,\ldots, x_d)$, we have $C_{S^{\ast}} \cap \{x_{d+1}=n\} = nS^{\ast}$. We consider the half-open parallelepiped
\[
\Pi _{S^{\ast}} \ = \ \left\{ \sum _{i=1}^{d+1} \lambda _i \bar{v}_i : 0< \lambda _i \leq 1 \text{ if } i\in I, 0\leq \lambda _i <1 \text{ if } i\not\in I\right\} \, .
\]
Then
\[
C_{S^{\ast}} \ = \ \bigsqcup _{u\in \mathbb{Z}^{d+1}}\Pi _{S^{\ast}} +u_1\bar{v}_1+\cdots +u_{d+1}\bar{v}_{d+1} \, .
\]
Let $S_i = \Pi _{S^{\ast}} \cap \{x_{d+1}=i\}$. Then $S_i$ is a \textit{partially open} hypersimplex; that is, a hypersimplex with certain facets removed.
\goodbreak
Our next result shows that $\operatorname{L}^r (nS^{\ast})$ is given by a polynomial in $n$ by determining its generating series. We follow the line of argumentation in~\cite[Proposition 3.3]{jochemkocombinatorial}. Observe that, together with equation \eqref{eq:halfopendec}, this reproves the polynomiality result of $\operatorname{L}^r(nP)$.
\begin{thm}\label{prop:h_cone}
With the notation given above, the equation
\[
\sum _{n\geq 0} \operatorname{L}^r (nS^{\ast}) t^n = \sum _{k_0,\ldots, k_{d+1} \geq 0 \atop \sum k_j =r} {r\choose k_0,\ldots,k_{d+1}}v_1^{k_1}\cdots v_{d+1}^{k_{d+1}} \frac{(1-t)^{k _0}A_{k_1}(t)\cdots A_{k_{d+1}}(t)}{(1-t)^{d+r+1}}\sum _{i=0}^d\operatorname{L} ^{k_0}(S_i)t^i \, ,
\]
holds true where $A_j(t)$ is the $j$-th Eulerian polynomial.
\end{thm}
\goodbreak
\begin{proof}
The generating function of the discrete moment tensor allows us to consider the discrete moment tensor of $nS^{\ast}$ by cutting the cone $C_{S^{\ast}}$ with the hyperplane $\{x_{d+1}=n\}$. The geometric interpretation of the half-open parallelepipeds tiling the cone, the translation covariance of the discrete moment tensor, and the binomial theorem together yield the equation
\begin{align*}
\sum _{n\geq 0} &\operatorname{L} ^r(nS^{\ast})t^n =\sum _{i=0}^d t^i \sum _{u_1,\ldots, u_{d+1} \geq 0} \operatorname{L} ^r (S_i +u_1\bar{v}_1+\cdots +u_{d+1}\bar{v}_{d+1})t^{u_1+\cdots +u_{d+1}}\\
&=\sum _{i=0}^d t^i \sum _{u_1,\ldots, u_{d+1} \geq 0}\sum _{j=0}^r{r\choose j}\operatorname{L} ^{r-j}(S_i) (u_1\bar{v}_1+\cdots +u_{d+1}\bar{v}_{d+1})^jt^{u_1+\cdots +u_{d+1}}\\
&=\sum _{i=0}^d t^i \sum _{u_1,\ldots, u_{d+1} \geq 0}\sum _{k_0,\dots, k_{d+1}\geq 0\atop \sum k_j =r} {r\choose k_0,\ldots,k_{d+1}}\operatorname{L} ^{k_0}(S_i) (u_1\bar{v}_1)^{k_1}\cdots (u_{d+1}\bar{v}_{d+1})^{k_{d+1}}t^{u_1+\cdots +u_{d+1}}\\
&=\sum _{i=0}^d t^i \sum _{k_0,\dots, k_{d+1}\geq 0\atop \sum k_j =r} {r\choose k_0,\ldots,k_{d+1}}\operatorname{L} ^{k_0}(S_i) \bar{v}_1^{k_1}\cdots \bar{v}_{d+1}^{k_{d+1}}\sum _{u_1,\ldots, u_{d+1} \geq 0}u_1^{k_1}\cdots u_{d+1}^{k_{d+1}}t^{u_1+\cdots +u_{d+1}}
\end{align*}
from which the result follows since
\[
\sum _{n\geq 0}n^jt^n \ = \ \frac{A_j(t)}{(1-t)^{j+1}} \, ,
\]
a known identity of generating functions (see, e.g.,~\cite{ccd}).
\end{proof}
We remark that the results and proofs of this section immediately carry over to general translative polynomial valuations (see~\cite{LS} for a definition). In particular, Theorem~\ref{prop:h_cone} can be generalized to give a new proof of \cite[Corollary 5]{PK92}.
\goodbreak
\section{Pick-type formulas} \label{sec:pick}
Pick's Theorem~\cite{Pick} gives an interpretation for the coefficients of the Ehrhart polynomial of a lattice polygon which establishes a relationship between the area of the polygon, the number of lattice points in the polygon and on its boundary. An analogue in higher dimensions can not exist (see, e.g.,~\cite{gruber}) as it is crucial that every polygon in dimension two has a unimodular triangulation; that is, a triangulation into simplices of minimal possible area $1/d!$. We offer interpretations for the coefficients of the Ehrhart tensor polynomial in the vector and the matrix cases by taking the route over the $h^r$-tensor polynomial.
\goodbreak
Given a polygon $P\in\mathcal{P}({\Z^2})$, we will consider unimodular triangulations of $P$ where such a triangulation will always be denoted by $\mathcal{T}$. The triangulation will be described by the edge graph $G=(V,E)$ of $\mathcal{T}$ where $V$ are the lattice points contained in $P$ and $E$ the edges of $\mathcal{T}$. Furthermore, the notation $x$ will be reserved for elements of $V$ and $y,z$ for endpoints of the edge $\{y,z\}\in E$. We define $V^o=P^o\cap\mathbb{Z}^2$, $\partial V=\partial P\cap\mathbb{Z}^2$, $E^o=\{\{y,z\}\in E : (y,z)\not\subset\partial P\} $, and $\partial E=\{\{y,z\}\in E : (y,z)\subset\partial P\} $.
\begin{figure}
\input{figures/threesimplices}
\caption{Types of half-open unimodular simplices in $\mathbb{R}^2$.}
\label{figure:threesimplices}
\end{figure}
Up to unimodular transformations, there are three types of half-open unimodular simplices in $\mathbb{R}^2$ that we will consider; these are $T_0$, $T_1$, and $T_2$ as given in Figure~\ref{figure:threesimplices}.
\goodbreak
\subsection{A Pick-type vector formula}
To determine the $h^1$-tensors from Theorem~\ref{prop:h_cone}, note that the Eulerian polynomial has a closed form
\begin{equation}\label{eq:A_j}
A_j(t)\ = \ \sum_{n=0}^j\sum_{i=0}^n(-1)^i\binom{j+1}{i}(n-i)^jt^n
\end{equation}
(see, e.g.,~\cite{ccd}). We then observe that $A_0(t)=1$, $A_1(t)=t$, and $A_2(t)=t^2+t$.
A comparison of coefficients of the numerator of~(\ref{eq:h_cone}) and that in Theorem~\ref{prop:h_cone} yields the formula
\begin{align*}
h_{S^{\ast}}^1(t)\ = \ \sum_{i=0}^2\operatorname{L}^1(S_i)t^{i}(1-t)+\operatorname{L}(S_i)t^{i+1}(v_1+v_2+v_3)
\end{align*} implying that
\begin{equation}\label{eq:hi_vector}
h_i^1(S^{\ast})\ = \ \operatorname{L}^1(S_i)-\operatorname{L}^1(S_{i-1})+\operatorname{L}(S_{i-1})(v_1+v_2+v_3)
\end{equation}
for a half-open simplex $S^{\ast}$ where $S_i$ are defined as in Section~\ref{sec:half-open}.
By Theorem~\ref{thm:half-open decomposition}, any lattice polygon can be partitioned into unimodular transformations of half-open simplices. Therefore, to calculate $h^r$-tensors, we will need to understand the half-open parallelepipeds $\Pi_{T_0}$, $\Pi_{T_1}$, and $\Pi_{T_2}$. For ease, we provide skeletal descriptions of these here. By setting $S^{\ast}$ to $T_0$, $T_1$, and $T_2$ with the vertices given in Figure~\ref{figure:threesimplices}, we obtain:
\begin{alignat}{5}\label{eq:piped}
T_0 &\ : \ S_0\cap\mathbb{Z}^2=\{ 0 \};\nonumber \\
T_1 &\ : \ S_1\cap\mathbb{Z}^2=\{v_1\} ;\\
T_2 &\ : \ S_2\cap\mathbb{Z}^2=\{v_2+v_3\} \nonumber
\end{alignat}
\noindent where $S_i\cap\mathbb{Z}^2=\varnothing$ for any combination of $S_i$, $T_j$ not given.
\begin{prop}\label{h* r=1}
For any lattice polygon, we have
\[
h_P^1(t)\ = \ t\sum_{V}x+t^2\left(\sum_{E^o}(y+z) - 2\sum_{V^o}x\right)+t^3\sum_{V^o}x.
\]
\end{prop}
\goodbreak
\begin{proof}
We determine the $h^1$-tensor polynomial of all half-open unimodular simplices, up to a unimodular transformation, with vertices $v_1,v_2,v_3$. Using formula (\ref{eq:hi_vector}) together with the values given in~(\ref{eq:piped}), we obtain the following $h^1$-tensor polynomials for each $T_i$:
\begin{align*}
h_{T_0}^1(t)&\ = \ t(v_1+v_2+v_3)\\
h_{T_1}^1(t)&\ = \ tv_1+t^2(v_2+v_3)\\
h_{T_2}^1(t)&\ = \ t^2((v_1+v_2)+(v_1+v_3)-2v_1)+t^3v_1
\end{align*}
Theorem~\ref{thm:half-open decomposition} together with a careful inspection of the $h^1$-tensor polynomials of the half-open simplices yield the result.
\end{proof}
From Proposition~\ref{h* r=1}, we can deduce formulas for the Ehrhart vectors.
\begin{prop}
For any lattice polygon,
\[
\operatorname{L}^1(nP)\ = \ \frac{n}{6}\left(2\sum_{V}x+4\sum_{V^o}x-\sum_{E^o}(y+z)\right)+\frac{n^2}{2}\sum_{\partial V}x+\frac{n^3}{6}\left(\sum_{\partial V}x + \sum_{E^o}(y+z)\right)
\]
\end{prop}
\begin{proof}
By definition, the Ehrhart vector polynomial equals
\[
\operatorname{L}^{1}(nP)\ = \ h_0^{1}(P) {n+3 \choose 3} + h_1^{1}(P) {n+2 \choose 3}+ h_2^{1}(P) {n+1 \choose 3}+ h_3^{1}(P) {n \choose 3} \, .
\]
A substitution of values from Proposition~\ref{h* r=1} yields
\[
\operatorname{L}^{1}(nP)\ = \ \frac{n^3+3n^2+2n}{6}\sum_V x+\frac{n^3-n}{6}\left(\sum_{E^o}(y+z) - 2\sum_{V^o}x\right)+\frac{n^3-3n^2+2n}{6}\sum_{\interior V}x\, .
\]
The result now follows from a quick comparison of coefficients.
\end{proof}
\goodbreak
\subsection{A Pick-type matrix formula}
We now determine the $h^2$-tensors in order to find a Pick-type formula for the discrete moment matrix.
Similar to the vector case, by comparing coefficients of the numerator of~(\ref{eq:h_cone}) and that in Theorem~\ref{prop:h_cone}, we obtain the formula
\begin{align*}
h_{S^{\ast}}^2(t)\ = \ \sum_{i=0}^2&\operatorname{L}^2(S_i)t^i(1-t)^2+2(v_1+v_2+v_3)\operatorname{L}^1(S_i)t^{i+1}(1-t)\\
&+(v_1^2+v_2^2+v_3^2)\operatorname{L}(S_i)t^{i+1}+(v_1+v_2+v_3)^2\operatorname{L}(S_i)t^{i+2}
\end{align*}
for a half-open simplex $S^{\ast}$ where $S_i$ are defined as in Section~\ref{sec:half-open}. The $h^2$-tensors of a half-open simplex are then found to be
\begin{align}\label{eq:hi_matrix}
\begin{split}
h_i^2(S^{\ast})\ = \ &\operatorname{L}^2(S_i)-2\operatorname{L}^2(S_{i-1})+\operatorname{L}^2(S_{i-2})+2(v_1+v_2+v_3)\left(\operatorname{L}^1(S_{i-1})-\operatorname{L}^1(S_{i-2})\right)\\
&+(v_1^2+v_2^2+v_3^2)\operatorname{L}(S_{i-1})+(v_1+v_2+v_3)^2\operatorname{L}(S_{i-2})\, .
\end{split}
\end{align}
\goodbreak
\begin{prop}\label{h* r=2}
If $P$ is a lattice polygon, then
\[
h_P^2(t)\ = \ t\sum_{V}x^{2}+t^2\left(\sum_{E}(y+z)^{2} - \sum_{V}x^{2}\right)+t^3\left(\sum_{E^o}(y+z)^{2} - \sum_{V^o}x^{2}\right)+t^4\sum_{V^o}x^{2}.
\]
\end{prop}
\begin{proof}
Similar to the $h^1$-tensor polynomial, we determine the $h^2$-tensor polynomial of all half-open unimodular simplices, up to unimodular transformation. Formula~(\ref{eq:hi_matrix}) for each $T_i$ with the values from~(\ref{eq:piped}) yields the following:
\begin{align*}
h_{T_0}^2(t)&\ = \ t(v_1^2+v_2^2+v_3^2)+t^2((v_1+v_2)^2+(v_2+v_3)^2+(v_3+v_1)^2-v_1^2-v_2^2-v_3^2)\\
h_{T_1}^2(t)&\ = \ tv_1^2+t^2((v_1+v_2)^2 + (v_1+v_3)^2 - v_1^2)+t^3(v_2+v_3)^2\\
h_{T_2}^2(t)&\ = \ t^2(v_2+v_3)^2+t^3((v_1+v_2)^2 + (v_1+v_3)^2 - v_1^2)+t^4v_1^2
\end{align*}
The claim now follows from Theorem~\ref{thm:half-open decomposition}.
\end{proof}
\goodbreak
From Proposition~\ref{h* r=2}, we can now deduce formulas for the Ehrhart matrices.
\begin{prop}
Given a lattice polygon $P$, we have
\begin{align*}
\operatorname{L}^2(nP)\ = \ &\frac{n}{12}\sum_{\partial E}(y-z)^2+\frac{n^2}{24}\left(12\sum _V x^2+12\sum _{V^o} x^2-\sum_{E}(y+z)^2-\sum_{E^o}(y+z)^2\right)\\
&+\frac{n^3}{12}\left(2\sum_{\partial V}x^2+\sum_{\partial E}(y+z)^{2}\right)+\frac{n^4}{24}\left(\sum_{E}(y+z)^2+\sum_{E^o}(y+z)^2\right).
\end{align*}
\end{prop}
\begin{proof}
By definition, the Ehrhart matrix polynomial equals
\[
\operatorname{L}^{2}(nP)\ = \ h_0^{2}(P) {n+4 \choose 4} + h_1^{2}(P) {n+3 \choose 4}+ h_2^{2}(P) {n+2 \choose 4}+ h_3^{2}(P) {n+1\choose 4}+ h_4^{2}(P) {n\choose 4} \, .
\]
The result follows now from Proposition~\ref{h* r=2} and comparing coefficients.
For $\operatorname{L}_{1}^{2}(P)$, we further observe that
\[
\operatorname{L}_{1}^{2} (P)\ = \ \frac{1}{12}\left(4\sum_{\partial V}x^2-\sum_{\partial E}(y+z)^{2}\right)\ = \ \frac{1}{12}\sum_{\partial E}(y-z)^{2} \, .
\qedhere
\]
\end{proof}
\goodbreak
\section{Positivity for $h^2$-vectors}\label{sec:positivity}
A fundamental theorem in Ehrhart theory is Stanley's Nonnegativity Theorem \cite{RS80} that states that the $h^\ast$-vector of every lattice polytope has nonnegative entries.
While positivity of real numbers is canonically defined up to sign change, there are many different choices for higher dimensional vector spaces such as $\mathbb{T}^r$; one for every pointed cone (compare, e.g., \cite{jochemkocombinatorial}). An important and well-studied cone inside the vector space of symmetric matrices is the cone of positive semidefinite matrices.
\goodbreak
A matrix $M\in \mathbb{R}^{d\times d}$ is called \textbf{positive semidefinite} if $x^tMx\geq 0$ for all $x\in \mathbb{R}^d$. By the identification of $\mathbb{T}^2$ with $\mathbb{R}^{d\times d}$, we call a tensor $T\in \mathbb{T}^2$ \textbf{positive semidefinite} if its corresponding symmetric matrix $(T_{ij})$ is positive semidefinite. By the spectral theorem, $T$ is a \textbf{sum of squares}; more precisely, if $T$ has eigenvalues $\lambda_1,\dots,\lambda_d \geq 0$ and corresponding normalized eigenvectors $u_1,\dots,u_d$ then
\[
(T_{ij})\ = \ \sum_{k=1}^d\lambda_ku_ku_k^t\,
\]
which is equivalent to $T=\sum_{k=1}^d\lambda_ku_k^2\in \mathbb{T}^2$. Therefore, a tensor is positive semidefinite if and only if it is a sum of squares.
As is the case for usual Ehrhart polynomials, the coefficients of Ehrhart tensor polynomials can be negative. However, in contrast to Ehrhart polynomials, this phenomenon appears already in dimension $2$. For segments, it can be seen that the linear coefficient of the Ehrhart tensor polynomial is $\sum _E (y-z)^2$. Furthermore, by \cite[Lemma 26]{LS} and Proposition~\ref{prop:secondcoeff}, all coefficients for line segments are positive semidefinite. The following example demonstrates negative definiteness in the plane.
\begin{ex}\label{ex:nonpositiveEhrharttensors}
Let $P$ be the triangle spanned by vertices
$v_1 = (0,1)^t$,
$v_2 = (-1,-7)^t$ and
$v_3 = (1,-4)^t$.
The Ehrhart tensor polynomial of $P$ can be calculated to be
\begin{align*}
\operatorname{L}^{2}(nP)=
\begin{pmatrix} \frac{1}{2} & \frac{3}{4} \\[6pt] \frac{3}{4} &
\frac{49}{6} \end{pmatrix} n
+ \begin{pmatrix} -\frac{1}{12} & -\frac{1}{8} \\[6pt] -\frac{1}{8} &
-\frac{23}{12}
\end{pmatrix} n^2
+ \begin{pmatrix} \frac{1}{2} & \frac{3}{4} \\[6pt] \frac{3}{4} & \frac{149}{6}
\end{pmatrix} n^3
+ \begin{pmatrix} \frac{13}{12} & \frac{13}{8} \\[6pt] \frac{13}{8} &
\frac{1079}{12}
\end{pmatrix} n^4.
\end{align*}
We observe that the coefficient of $n^2$ is negative definite.
Lattice triangles for which this coefficient is indefinite also exist; for
example, the triangle with vertices at $(0,-4)^t$, $(0,4)^t$ and $(-1,0)^t$.
\end{ex}
Our main result is the following analogue to Stanley's Nonnegativity Theorem for the $h^2$-tensor polynomial of a lattice polygon.
\begin{thm}\label{thm:main}
The $h^2$-tensors of any lattice polygon are positive semidefinite.
\end{thm}
Before proving Theorem~\ref{thm:main}, we make a few more observations. Positive semidefiniteness of $h^2$-tensors is preserved under unimodular transformations since, from Equation~\eqref{eq:expansion} and comparing coefficients, we have
\[
h^r _i (\phi P)(v,v)=h^r_i (\phi^t v, \phi ^t v)
\]
for all $P\in \mathcal{P}(\mathbb{Z}^d)$, $\phi \in \mathrm{GL} (\mathbb{Z}^d)$, and $v\in \mathbb{R}^d$. However, as the next example shows, positive semidefiniteness of the $h^2$-vector is in general not preserved under translation.
\begin{ex} \label{ex:counter_ex_monotonicity}
Let $S = \mathrm{conv}\{v_1,v_2,v_3\}\setminus\mathrm{conv}\{v_2,v_3\}$ be the half-open simplex with vertices
$v_1 =~(3,-2)^t$,
$v_2 = (2,-2)^t$, and
$v_3 = (2,-1)^t$.
From the formula of the $h^2$-vector of a half-open simplex which can be found in the proof of Proposition~\ref{h* r=2},
we obtain that
\begin{equation*}
h_S^2(t)\ = \ \begin{pmatrix} 4 & -4 \\ -4 & 4
\end{pmatrix}t+\begin{pmatrix} 37 & -28 \\ -28 & 21
\end{pmatrix}t^2+\begin{pmatrix} 25 & -15 \\ -15 & 9
\end{pmatrix}t^3\, .
\end{equation*}
That is, with a determinant of $-7$, the matrix $h^2_2(S)$ is not positive
semidefinite. However, it can be seen that the positive semidefiniteness of $h^2$-tensors is not preserved under translations.
To illustrate, consider the translate $S-v_1$. The $h^2$-vector of the translated simplex
\begin{equation*}
h_{S-v_1}^2(t)\ = \ \begin{pmatrix} 1 & 0 \\ 0 & 1 \end{pmatrix}t^2
+ \begin{pmatrix} 1 & 1 \\ 1 & 1 \end{pmatrix} t^3, \\
\end{equation*}
has positive semidefinite coefficients.
\end{ex}
\goodbreak
Since Example~\ref{ex:counter_ex_monotonicity} shows that $h^2$-tensors of half-open polytopes can be negative, it follows that $h^2$-tensors are not monotone with respect to inclusion in contrast to the coefficients of the $h^\ast$-polynomial~\cite{RS93}. Therefore, techniques such as irrational decomposition or half-open decomposition that succesfully helped prove Stanley's Nonnegativity Theorem (see \cite{ccd,jochemkocombinatorial}) cannot immediately be applied with Theorem \ref{thm:half-open decomposition}; we will have to take a different route.
To prove Theorem \ref{thm:main}, we decompose a lattice polygon into lattice polygons with few vertices for which the $h^2$-vectors can easily be calculated. For the remainder of this article, allow a lattice polygon to always mean a full-dimensional in $\mathbb{R}^2$ although the argument is independent from the chosen ambient space. A \textbf{sparse decomposition} of $P\in\mathcal{P}({\Z^d})$ is a finite set $\mathcal{D}=~\{P_1,P_2,\ldots, P_m\}$ of lattice polygons such that
\begin{enumerate}[i)]
\item $\operatorname{L}(P_i)\in\{3,4\}$ for each $i\in[m]$,
\item $P_i \cap P_j = \emptyset$ or is a common vertex of $P_i$ and $P_j$ for all $i\neq j$, and
\item $P\cap\mathbb{Z}^2 = \bigcup_{i=1}^m P_i\cap\mathbb{Z}^2$.
\end{enumerate}
\begin{figure}
\begin{minipage}{0.3\textwidth}
\input{figures/polygon4_a_tikz}
\end{minipage}
\begin{minipage}{0.3\textwidth}
\input{figures/polygon4_b_tikz}
\end{minipage}
\begin{minipage}{0.3\textwidth}
\input{figures/polygon4_c_tikz}
\end{minipage}
\caption{Lattice polygons with 4 lattice points and their unimodular triangulations.}
\label{figure:polygon4points}
\end{figure}
\begin{lemma}{\cite[Section 4]{LZ11}}\label{lem:fig}
Up to unimodular transformation, there are three different lattice polygons containing exactly four lattice points. They are given in Figure \ref{figure:polygon4points}.
\end{lemma}
The following lemma ensures that every lattice polygon has a sparse decomposition.
\goodbreak
\begin{lemma} \label{ref:lemma_decomp}
Every lattice polygon has a sparse decomposition.
\end{lemma}
\goodbreak
\begin{proof} We proceed by induction on $\operatorname{L}(P)$. The statement is trivially true if $\operatorname{L}(P)\in \{3,4\}$. Hence, we may assume that $\operatorname{L}(P)>4$
and choose a vector $a\in\mathbb{R}^2\setminus\{0\}$ such that
$a^t v \neq a^t w$ for each $v,w \in P\cap\mathbb{Z}^2$ where $v\neq w$. Note
that such an $a$ exists since $\operatorname{L}(P)$ is finite. Let $P\cap\mathbb{Z}^2 = \{ v_1,\ldots, v_n\}$ be such that
\[
a^t v_1 \ > \ a^t v_2 \ > \ \cdots \ > \ a^t v_n
\]
and set $ Q = \mathrm{conv}\{v_3,v_4,\ldots, v_n\}$.
Then, by convexity, we obtain $Q\cap\mathbb{Z}^2 = P\cap\mathbb{Z}^2\setminus\{v_1,v_2\}$.
\goodbreak
If $Q$ is not full-dimensional and all lattice points of $Q$ lie on a
line, then a sparse decomposition of $P$ can
easily be constructed. If $u_1,u_2,$ and $u_3$ are not collinear, then we can construct a sparse decomposition which is illustrated in Figure~\ref{figure:decomp_aff_dep}. Let $P_1=\mathrm{conv}\{u_1,u_2,u_3\}$. Then, by design, the triangle $P_1$ does not contain any other lattice point and at least one of $u_1$ or $u_2$ are visible from all points $u_4,\ldots, u_n$. Without loss of generality, assume $u_1$ is visible. Then for all $2\leq i\leq \lfloor \frac{n}{2} \rfloor -1$ define $P_i =\mathrm{conv} \{u_1,u_{2i},u_{2i+1}\}$, $P_{\lfloor \frac{n}{2} \rfloor} =\mathrm{conv} \{u_1,u_{n-2},u_n\}$ if $n$ is even, and $P_{\lfloor \frac{n}{2} \rfloor} =~\mathrm{conv} \{u_1,u_{n-1},u_n\}$ if $n$ is odd. Then $\{P_1,\ldots, P_{\lfloor \frac{n}{2} \rfloor}\}$ is a sparse decomposition. If $u_1,u_2,$ and $u_3$ are collinear, then a sparse decomposition can be obtained by instead setting $P_1=\mathrm{conv}\{u_2,u_3,u_4\}$.
\begin{figure}
\input{figures/sparsedecomp}
\caption{Sparse decomposition of $P$ for the case of a collinear $Q$.}
\label{figure:decomp_aff_dep}
\end{figure}
\goodbreak
Suppose $Q$ is full-dimensional. Then, by the induction hypothesis, there is a sparse decomposition $\mathcal{D}_Q$ of $Q$. Let $i$ be the smallest index such that the points $u_1,u_2,u_{i}$ do not lie on a common straight line. By construction, the simplex $S=\mathrm{conv} (u_1,u_2,u_{i})$ contains no other lattice points and, thus, $\mathcal{D}_Q\cup\{S\}$ is a sparse decomposition of $P$.
\end{proof}
\goodbreak
\begin{lemma} \label{ref:lemma_h2_small_polytope}
If $P\in\mathcal{P}({\Z^2})$ is a lattice polygon containing exactly three or four lattice points, then $h_2^2(P)$ is positive semidefinite.
\end{lemma}
\begin{proof}
If $\operatorname{L}(P)=3$, then $P=\mathrm{conv}(v_1,v_2,v_3)$ is a unimodular lattice simplex and the statement follows from Proposition~\ref{h* r=2} as
\[
h_2^2(P)=(v_1+v_2)^2+(v_1+v_3)^2+(v_2+v_3)^2-v_1^2-v_2^2-v_3^2=(v_1+v_2+v_3)^2.
\]
Suppose $\operatorname{L}(P)=4$. We have to distinguish between the three possible cases, up to unimodular transformation, given in Figure~\ref{figure:polygon4points}. First, if $P$ contains one interior lattice point $v_4$ and vertices $v_1, v_2,v_3$, then we have $v_4=\tfrac{1}{3}(v_1+v_2+v_3)$ and Proposition~\ref{h* r=2} implies that
\begin{align*}
h_2^2(P) &= (v_1+v_2)^{2} + (v_1+v_3)^{2} + (v_2+v_3)^{2} +
(v_1+v_4)^{2} + (v_2+v_4)^{2} + (v_3+v_4)^{2} \\
&\quad - v_1^{2}- v_2^{2}- v_3^{2}- v_4^{2} \\
&=(v_1+v_2)^{2} + (v_1+v_3)^{2} + (v_2+v_3)^{2} + 2v_4^2+2v_4(v_1+v_2+v_3)\\
&= (v_1+v_2)^{2} + (v_1+v_3)^{2} + (v_2+v_3)^{2} +
\tfrac{8}{9}(v_1+v_2+v_3)^{2}.
\end{align*}
Next, if $P$ is a parallelepiped, then $v_1+v_3 = v_2+v_4$ and thus
\begin{align*}
h_2^2(P) &= (v_1+v_2)^{2} + (v_2+v_3)^{2} + (v_3+v_4)^{2} + (v_1+v_4)^{2}\\
&+\tfrac{1}{2}(v_1+v_3)^{2} +\tfrac{1}{2}(v_2+v_4)^{2}
- v_1^{2}- v_2^{2}- v_3^{2}- v_4^{2} \\
&=\tfrac{1}{2}(v_1+v_2+v_3+v_4)^{2}+\tfrac{1}{2}(v_1+v_2)^{2}
+\tfrac{1}{2}(v_2+v_3)^{2}+\tfrac{1}{2}(v_3+v_4)^{2}+\tfrac{1}{2}(v_1+v_4)^{2}.
\end{align*}
Finally, if $P$ has three vertices and no interior lattice point, then one lattice point of $P$, say $v_2$ as in Figure~\ref{figure:polygon4points}, lies in the relative interior of the edge given by the vertices $v_1$ and $v_3$ implying that $v_2=\frac{1}{2}(v_1+v_3)$. In this case, we obtain
\begin{align*}
h_2^2(P) &= (v_1+v_2)^{2} + (v_2+v_3)^{2} + (v_3+v_4)^{2} + (v_1+v_4)^{2}
+(v_2+v_4)^{2}
- v_1^{2}- v_2^{2}- v_3^{2}- v_4^{2}\\
&=\tfrac{5}{2}v_1^2+\tfrac{5}{2}v_3^2+2v_4^2+3v_1v_4+3v_3v_4+3v_1v_3\\
&=\tfrac{3}{2}(v_1+v_3+v_4)^{2} +v_1^{2} +v_3^{2} +\tfrac{1}{2}v_4^{2}.
\qedhere
\end{align*}
\end{proof}
\goodbreak
We will need the following geometric observation in our proof of Theorem~\ref{thm:main}.
\begin{lemma} \label{ref:lemma_planar_cond}
Let $P\in\mathcal{P}({\Z^2})$ and $v$ be a lattice point in the relative interior of $P$. Then at least one of the following two statements is true:
\begin{enumerate}[(i)]
\item \label{ref:lemma_two_cond_i}
$v=\tfrac{1}{2}(v_1+v_2)$ for lattice points $v_1, v_2 \in P$ such that $v_1\neq v_2$;
\item \label{ref:lemma_two_cond_ii}
$v=\tfrac{1}{3}(v_1+v_2+v_3)$ for pairwise disjoint lattice points $v_1, v_2, v_3 \in P$.
\end{enumerate}
\end{lemma}
\begin{proof}
If $v$ is contained in a segment formed by two lattice points in $P$, then $v$ is easily seen to be of the form given in \emph{(i)}.
Therefore, we may assume that $v$ is not contained in any line segment formed by lattice points in $P$. By Caratheodory's Theorem~(see, e.g., \cite{Schneider:CB2}), there are lattice points $v_1,v_2,v_3\in P$ such that $v$ is contained in the simplex formed by $v_1,v_2,$ and $v_3$. If $v,v_1,v_2,v_3$ are the only lattice points in the simplex, then condition \emph{(ii)} follows from Lemma~\ref{lem:fig}. Otherwise, there is a lattice point $u\in\mathrm{conv}\{v_1,v_2,v_3\}\setminus\{v,v_1,v_2,v_3\}$ and, consequently, $v$ must be contained in one of
the three lattice simplices
\begin{align*}
S_1 = \mathrm{conv}\{v_2,v_3,u\}, \quad
S_2 = \mathrm{conv}\{v_1,v_3,u\},\quad
S_3 = \mathrm{conv}\{v_1,v_2,u\}.
\end{align*}
Without loss of generality, let $v\in S_1\subsetneq \mathrm{conv}\{v_1,v_2,v_3\}$. By reiteration of the above procedure, each time with a replacement of $v_1$ by $u$, we eventually find affinely independent $v_1, v_2, v_3$ such that
$\{v,v_1,v_2,v_3\} = \mathrm{conv}\{v_1,v_2,v_3\}\cap\mathbb{Z}^2$ and condition \emph{(ii)} follows again from Lemma~\ref{lem:fig}.
\end{proof}
We are now equipped to give the proof of our nonnegativity theorem.
\begin{proof}[Proof of Theorem~\ref{thm:main}]
From Proposition~\ref{h* r=2}, it immediately follows that $h_0^{2}(P),h_1^{ 2}(P)$, and $h_4^{2}(P)$ are sums of squares.
Let $\mathcal{D}=\{P_1,P_2,\ldots,P_m\}$ be a sparse decomposition of $P$ which exists by Lemma~\ref{ref:lemma_decomp} and let $\mathcal{S}$ be some triangulation of $\cup_{i=1}^mP_i$. Observe that the closure of $P \setminus (P_1\cup\dots\cup P_m)$ is a union of not necessarily convex lattice polygons and any triangulation of $\cup_{i=1}^mP_i$ can be extended to a triangulation in $P$. Let $\mathcal{T}$ be a triangulation of $P$ such that $\mathcal{S}\subseteq\mathcal{T}$. Let $G=(V,E)$ be the edge graph of $\mathcal{T}$ and $G'=(V',E')$ be the edge graph of $\mathcal{S}$. For every $x\in V$, we define $\alpha_x = |\{i \in [m]: x\in P_i\}|$. Note that $\alpha_x \geq 1$ for all $x\in V$ since $\mathcal{D}$ is a
sparse decomposition.
Proposition~\ref{h* r=2} then implies that
\begin{align*}
h_2^{2}(P) &= \sum_{E}(y+z)^{ 2} - \sum_{V}x^{ 2}\\
&= \sum_{E'} (y+z)^{ 2}
- \sum_{V}\alpha_x \ x^{ 2}
+ \sum_{E\setminus E'}
(y+z)^2
- \sum_{V}(1-\alpha_x) \ x^{ 2} \\
&= \sum_{i=1}^m h_2^{2}(P_i)
+ \sum_{E\setminus E'}
(y+z)^2
+ \sum_{V}(\alpha_x-1) \ x^{ 2} \, ,
\end{align*}
and therefore, by Lemma~\ref{ref:lemma_h2_small_polytope}, $h_2^2(P)$ is a sum of squares.
We have left to show that $h_3^{2}(P)$ is also a sum of squares.
For every $v\in V$, we define
$N(v)=\{u\in V: \{u,v\}\in E\}$
to be the set of vertices adjacent to $v$ in $G$. Let $E_1\subseteq \interior E$ be the set of edges that have exactly one endpoint on the boundary of $P$ and $E_2\subseteq\interior E$ be the set of edges with both endpoints on the boundary of $P$ but relative interior in $\interior P$. By Proposition~\ref{h* r=2}, we obtain
\begin{align*}
h_3^{2}(P) &= \sum_{\interior E}
(y+z)^{ 2} - \sum_{\interior{V}}x^{ 2}\\
&= \sum_{\substack{v \in \interior{V}}}\left(\sum_{u \in N(v)}\left(\frac{1}{2}(v+u)^{ 2}\right) - v^{ 2}\right) + \sum _{E_1} \frac{1}{2}(y+z)^2 + \sum _{E_2} (y+z)^2 \, .
\end{align*}
It is thus sufficient to show that
\begin{equation*}
a(v) := \sum_{u \in N(v)}\left(\frac{1}{2}(v+u)^{ 2} - v^{ 2}\right)
\end{equation*}
is a sum of squares for all $v\in \interior{V}$. In view of Lemma~\ref{ref:lemma_planar_cond}, we
distinguish two cases. First, suppose that there are
$v_1, v_2 \in V \setminus \{v\}$ such that $v=\tfrac{1}{2}(v_1+v_2)$.
Then
\begin{align*}
a(v)&=\tfrac{1}{2}(v + v_1)^{ 2} +\tfrac{1}{2}(v + v_2)^{ 2} - v^{ 2} + \sum _{u\in N(v)\setminus \{v_1,v_2\}} \tfrac{1}{2}(v+u)^{ 2}\\
&= \tfrac{1}{2}(v_1+v_2)^{ 2} + \tfrac{1}{2} v_1^{ 2} + \tfrac{1}{2} v_2^{2}+ \sum _{u\in N(v)\setminus \{v_1,v_2\}} \tfrac{1}{2}(v+u)^{ 2} \, .
\end{align*}
In the second case, there exist pairwise disjoint
$v_1, v_2, v_3 \in V \setminus \{v\}$ such that $v=\tfrac{1}{3}(v_1+v_2+v_3)$.
Therefore
\begin{align*}
a(v)&=\tfrac{1}{2}(v + v_1)^{ 2}
+\tfrac{1}{2}(v + v_2)^{ 2}
+\tfrac{1}{2}(v+v_3)^{ 2} - v^{ 2}+\sum _{u\in N(v)\setminus \{v_1,v_2,v_3\}} \tfrac{1}{2}(v+u)^{ 2}\\
&= \tfrac{7}{18}(v_1+v_2+v_3)^{ 2}
+ \tfrac{1}{2} v_1^{ 2}
+ \tfrac{1}{2} v_2^{ 2}
+ \tfrac{1}{2} v_3^{ 2}+\sum _{u\in N(v)\setminus \{v_1,v_2,v_3\}} \tfrac{1}{2}(v+u)^{ 2} \, .\qedhere
\end{align*}
\end{proof}
\goodbreak
\section{Further results and outlook}\label{sec:further}
It is natural to ask whether Theorem~\ref{thm:main} holds true in higher dimensions. Using the software package polymake~\cite{polymake:2017,polymake:2000} we have calculated the $h^2$-tensor polynomials of several hundred randomly generated polytopes in dimension $3$ and $4$. Based on these computational results, we offer the following conjecture.
\begin{conj}\label{conj:pos}
For $d\geq 1$, the coefficients of the $h^2$-tensor polynomial of any of a lattice polytope in $\mathbb{R}^d$ are positive semidefinite.
\end{conj}
For our proof of Theorem~\ref{thm:main}, it was crucial that every lattice polygon has a unimodular triangulation. Since this no longer holds true in general for higher dimensional polytopes, a proof of Conjecture~\ref{conj:pos} would need to be conceptually different.
Finding inequalities among the coefficients of the $h^\ast$-polynomial of a
lattice polytope, beyond Stanley's Nonnegativity Theorem, is currently of
great
interest in Ehrhart theory. The ultimate goal is a classification of all
possible $h^\ast$-polynomials: a classification of all
$h^\ast$-polynomials
of degree $2$ can be found in \cite[Proposition 1.10]{HM09}. Another fundamental
inequality
is due to Hibi~\cite{MR1275662} who proved that $h_i(P)-h_1(P)\geq 0$ for all
$1\leq i<d$ and full-dimensional lattice polytopes that have an interior
lattice
point. Calculations with polymake again suggest that there
might be a version for matrices motivating the following conjecture.
\begin{conj}\label{conj:hibi}
Let $P$ be a lattice polytope containing a lattice point in its interior.
Then
the matrices $h^2_i(P)-h^2_1(P)$ for $1\leq i < \dim(P)+2$ are positive
semidefinite.
\end{conj}
In recent years, additional inequalities for the coefficients of the
$h^\ast$-polynomial have been shown (see e.g.
~\cite{Athanasiadis04,Stapledon09,Stapledon16})
which raises the question as to whether there are analogous results for Ehrhart tensors.
\begin{ques}\label{ques:1}
Which known inequalities among the coefficients of the $h^\ast$-polynomial
of a lattice polytope can be generalized to $h^r$-tensor polynomials of
higher rank?
\end{ques}
An answer would depend on the notion of positivity that is chosen. A natural choice for higher rank $h^r$-tensors, extending positive semidefiniteness of matrices, is to define $T\in\mathbb{T}^r$ to be positive semidefinite if and only if $T(v,\dots,v)\geq 0$ for all $v\in\mathbb{R}^d$. However, assuming this definition of positivity, there can not be any inequalities that are valid for all polytopes if the rank $r$ is odd since $T(v,\ldots, v) = (-1)^rT(-v,\ldots, -v)$.
In the case that $r$ is even, we are able to extend another classical result, namely Hibi's Palindromic Theorem~\cite{Hibi91} characterizing reflexive polytopes. A lattice polytope $P\in \mathcal{P}(Z^d)$ is called \textbf{reflexive} if
\[
P = \{x\in \mathbb{R}^d \colon Ax \leq 1\}
\]
where $A\in Z^{d\times d}$ is an integral matrix.
\goodbreak
\begin{thm}[Hibi~\cite{Hibi91}]\label{thm:hibioriginal}
A polytope $P\in\mathcal{P}({\Z^d})$ is reflexive if and only if $h_i^{\ast}(P)=h_{d-i}^{\ast}(P)$ for all $0\leq i\leq d$.
\end{thm}
A crucial step in the proof of Theorem~ \ref{thm:hibioriginal} is to observe that a polytope $P$ is reflexive if and only if
\[
nP\cap \mathbb{Z}^d =(n+1)\interior P \cap \mathbb{Z}^d
\]
for all $n\in\mathbb{N}$ (see \cite{ccd}). We use this fact to give the following generalization.
\begin{prop}\label{thm:hibiext}
Let $r\in\mathbb{N}$ be even and $P\in\mathcal{P}({\Z^d})$ be a lattice polytope that contains the origin in its relative interior. The polytope $P$ is reflexive if and only if $h_i^{r}=h_{d+r-i}^{r}$ for all $0\leq i\leq d+r$.
\end{prop}
\begin{proof}
By Theorem~\ref{thm:tensorreciprocity} and comparing coefficients in equation~\eqref{eq:expansion}, it follows that the assertion $h_i^{r}(P)=h_{d+r-i}^{r}(P)$ is equivalent to $\operatorname{L}^{r}((n-1)P)=\operatorname{L}^{r}(n\interior P)$ for all integers $n$.
If $P$ is a reflexive polytope, then $\operatorname{L}^{r}((n-1)P)=\operatorname{L}^{r}(n\interior P)$ for all integers $n$ since, as given above, we have $(n-1)P\cap \mathbb{Z}^d = n\interior P \cap \mathbb{Z}^d$.
Now assume that $P$ is not reflexive. Then there exists an $n\in\mathbb{N}$ such that
\[
(n-1)P\cap \mathbb{Z}^d \subsetneq n\interior P \cap \mathbb{Z}^d \, .
\]
Therefore, for any $v\in \mathbb{R}^d\setminus \{0\}$, we obtain
\[
\sum _{x\in (n-1)P\cap \mathbb{Z}^d} (x^tv)^r \ < \ \sum _{x\in n\interior P\cap \mathbb{Z}^d} (x^tv)^r \,
\]
and, in particular, $\operatorname{L}^{r}((n-1)P)\not =\operatorname{L}^{r}(n\interior P)$ completing the proof.
\end{proof}
Note that the proof of Proposition~\ref{thm:hibiext} shows that for odd rank $r$ palindromicity of the $h^r$-tensor polynomial of a reflexive polynomial is still necessary, but not sufficient, since all centrally symmetric polytopes have a palindromic $h^r$-tensor polynomial; namely the constant zero polynomial.
\section*{Acknowledgements}
Katharina Jochemko was partially supported by the Knut and Alice Wallenberg Foundation. Laura Silverstein was supported by the Austrian Science Fund (FWF) Projects P25515-N25 and I3017-N35.
\bibliographystyle{siam}
| 24de90a45a89f00d9af1b8f323287ddb19f17f86 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
It is believed that betting house always makes money in long run irrespective of their short term loss or gain. In this paper, we make an attempt to understand this phenomenon with the concept of simple `expectation' and `variance' of probability theory. First, we will discuss what is \textbf{fair game}. Let's consider a simple game of English Premier League (EPL) where Manchester United (ManU) is playing against Liverpool. Suppose a betting house offers a game that if ManU wins with probability 0.606, then the player will receive \$0.65 from the betting house. On the other hand, if ManU loses with probability 0.394 then the player has to pay \$1 to the betting house.
Now player's revenue scheme $R_p$ is defined as follows:
\begin{eqnarray*}
R_{p}=\bigg\{\begin{array}{cc}
0.65 & \mbox{with probability }0.606,\\
-1 & \mbox{with probability }0.394,
\end{array}
\end{eqnarray*}
and player's expected revenue is
$$
E(R_p)=0.65\times 0.606 - 0.394=0.
$$
More about expectation and moments can be found in [1,2]. Now, we will look at the revenue scheme and expected revenue for the betting house for the same game.
\begin{eqnarray*}
R_{b}=\bigg\{\begin{array}{cc}
-0.65 & \mbox{with probability }0.606,\\
1 & \mbox{with probability }0.394.
\end{array}
\end{eqnarray*}
Expected revenue for betting house is,
$$
E(R_b)=-0.65\times 0.606 + 0.394=0.
$$
If we compare $R_p$ and $R_b$, then we can see player's loss is the gain of the betting house and vice-versa.
This type of game is known as `zero-sum' game. Also, $E(R_b)=E(R_p)=0$ means if the player and betting house play this game several times, then `on-average' neither betting house nor the player will make or lose money. This is called `fair-game.'
Now we know that the betting house has an establishment cost and most of the time these betting houses are the for-profit organization. Here the question is how they are making money. An easy way of doing it, if the betting house pays the player less than what they suppose to pay. That means, in the fair game, if they think to pay the player \$0.65, but to make money, they will pay the player less than \$0.65. For example, if they pay the player \$0.6 then the revenue scheme for the betting house is,
\begin{eqnarray*}
R_{b}=\bigg\{\begin{array}{cc}
-0.6 & \mbox{with probability }0.606,\\
1 & \mbox{with probability }0.394,
\end{array}
\end{eqnarray*}
and the expected revenue for betting house is,
$$
E(R_b)=-0.6\times 0.606 + 0.394=0.0304.
$$
It means if the player and betting house play this game several times, the on-average betting house will make \$0.03 or 3 cents from the player and the player will lose the same amount because it is a zero-sum game. So to earn money, in the long run, the betting house offers the player less than what is fair. Rest of the paper is organized as follows. In section \ref{sec_betting_strategy_prob_th}, we present the strategy of betting houses using probability theory. In section \ref{sec_strategy_league_football_match}, we discuss the strategy for league football match. In section \ref{sec_data_analysis}, we presented the data analysis and showed how imputed cost of a betting house could be estimated numerically.
\section{Betting Strategy with Probability Theory}\label{sec_betting_strategy_prob_th}
In this section, we present the strategy of the betting house. Suppose $A$ is an event with $P(A)=p$. If $A$ happens then, the betting house will pay \$$r$. Otherwise, the betting house will receive \$1. So the revenue scheme for the betting house is
$$
R_{b}=\bigg\{
\begin{array}{cc}
-r & \mbox{with probability }p,\\
1 & \mbox{with probability }(1-p),
\end{array}
$$
and the expected revenue for betting house is,
$$
E(R_b)=-rp+(1-p).
$$
Now this game is fair game if $E(R_b)=0$, that is $r=\frac{1}{p}-1$,
and $\frac{1}{p}$ is known as `\emph{decimal odds}'. The variance of the revenue scheme is generally considered as risk of a game. So for a fair game $Var(R_b)=E(R_b^2)-[E(R_b)]^2=E(R_b^2)$. Now
\begin{eqnarray*}
E(R_b^2)&=& p \Big(\frac{1}{p}-1\Big)^2+ (1-p)\\
&=& (1-p)\Big[\frac{1-p}{p}+1\Big]\\
&=&\frac{1-p}{p}=\frac{1}{p}-1=r.
\end{eqnarray*}
Interestingly, for the fair game, $r$ is the amount which betting house pays for each dollar they receive. It turns out that $r$ is also the measure of risk for the same strategy as $Var(R_b)=r$, which is known as `\emph{fractional odds}.' To make a profit, in the long run, betting house pays \$$(r-\epsilon)$, where $\epsilon>0$. The revenue scheme is
$$
R_{b}=\bigg\{
\begin{array}{cc}
-\{(\frac{1}{p}-1)-\epsilon\} & \mbox{with probability }p,\\
1 & \mbox{with probability }(1-p),
\end{array}
$$
and the expected revenue of betting house is,
\begin{eqnarray*}
E(R_b)&=&-(\frac{1}{p}-1)p+\epsilon p+ (1-p)\\
&=&-(1-p)+\epsilon p + (1-p)\\
&=&\epsilon p.
\end{eqnarray*}
After simplification $Var(R_{b})=E(R_b^2)-[E(R_b)]^2= r(1-p\epsilon)^2$.
\section{Strategy for League Football Match}\label{sec_strategy_league_football_match}
There are three mutually exclusive outcomes in a league football match. The outcomes are (i) Hometeam win, (ii) Away team win, (iii) Draw. Let us consider the EPL match between ManU Vs.
Liverpool, where ManU is Hometeam. Suppose for this match betting house calculates the probabilities 0.6, 0.15 and 0.25 for three outcomes, namely Hometeam wins, Away team wins and draw respectively. The corresponding decimal odds are $1/0.6=1.66$, $1/0.15=6.66$ and $1/0.25=4.00$.
In the previous section, we noticed that the betting house would never reveal these fair odds. They will always announce odds which are less than the fair value so that they can stay in profit. For instance, if the betting house offers odds of 1.57, 6.57 and 3.87 against the respective events of Hometeam wins, Away team wins and draw, then the revised revenue scheme for ManU to win the match with announced odds of 1.57 is,
$$
R_{b}^H=\bigg\{
\begin{array}{cc}
-(1.57-1) & \mbox{with probability }0.6,\\
1 & \mbox{with probability }0.4.
\end{array}
$$
The expected revenue for the betting house is,
$$
E(R_b^H)=-0.57\times 0.6 + 0.4=0.058.
$$
Similarly, the revised revenue scheme for the Liverpool to win the match with announce odds of 6.57 is,
$$
R_{b}^A=\bigg\{
\begin{array}{cc}
-(6.57-1) & \mbox{with probability }0.15,\\
1 & \mbox{with probability }0.85,
\end{array}
$$
and the expected revenue for betting house is,
$$
E(R_b^A)= -5.57\times 0.15 + 0.85=0.0145.
$$
Likewise, the revised revenue scheme that match will be draw with announce odds of 3.87 is,
$$
R_{b}^D=\bigg\{
\begin{array}{cc}
-(3.87-1) & \mbox{with probability }0.25,\\
1 & \mbox{with probability }0.75,
\end{array}
$$
and the expected revenue for betting house is,
$$
E(R_b^D)= -2.87\times 0.25 + 0.75=.0325.
$$
Expected revenue of betting house for all events are positive. Mathematically, we can show when the announce odds are converted to probabilities; they usually add up to more than 1 to keep the house on benefit.
Let us assume, $\Big(\frac{1}{P_H}-\epsilon\Big)=\frac{1}{P_{H}^*},$ $\Big(\frac{1}{P_A}-\epsilon\Big)=\frac{1}{P_{A}^*}$ and $\Big(\frac{1}{P_D}-\epsilon\Big)=\frac{1}{P_{D}^*}$
Therefore,
\begin{eqnarray*}
\frac{1-\epsilon P_H}{P_H}=\frac{1}{P_H^*}\\
\frac{P_H}{1-\epsilon P_H}={P_H^*}\\
{P_H}<{P_H^*}
\end{eqnarray*}
In the same way we can show ${P_A}<{P_A^*}$ and ${P_D}<{P_D^*}$. Now adding both sides we can conclude,
${P_H}+{P_A}+{P_D}<{P_H^*}+{P_A^*}+{P_D^*}$.
Of course probabilities of a fair game sums up to 1 i.e. ${P_H}+{P_A}+{P_D}=1$, hence evidently ${P_H^*}+{P_A^*}+{P_D^*}>1$.
\section{Data Analysis}\label{sec_data_analysis}
Let us discuss the former analysis above with EPL data. Data is accessible on \textbf{http://www.football-data.co.uk/englandm.php}. In the table (\ref{table_B365}) we present the decimal odds of 10 different EPL matches from the popular betting house Bet365. First three columns, `B365H', `B365D' and `B365A' represent the decimal odds (i.e. $1/p$) of three events that are (i) home team wins, (ii) draw and (iii) away team wins; from the betting house Bet365. It is visible that total probability of the three mutually exclusive events of a given match is greater than one which violates the thumb rule that probabilities add up to one. It is inevitable that sum of the probabilities will be greater than one from our results described in the previous section. However, if there are two betting houses, we can say that one house is offering better odds to the players if its sum of probabilities is closer to one.
Let us consider a match held on 8th August 2015, between Bournemouth and Aston Villa where the former team is Hometeam and later one is Away team. We look at the betting houses for this match, the following table \ref{table_compare_odds} (for the six betting houses, namely Bet365 (B365), Bet\&Win (BW), Interwetten (IW), Ladbrokes (LB), William Hill (WH), VC Bet (VC) ). Here we can compare the announce odds whichever is near to 1 that betting the house is offering fair. As per the table \ref{table_compare_odds}, B365 is offering fair odds compared with others. The betting house Interwetten is offering worst odds for this match.
\subsection{Imputed Cost of a Betting House}
We can show different ways of calculating Imputed cost of the betting house. Here we are using additive model for the same. In this model, revenue of betting house is described as,
$$
R_{b}=\bigg\{
\begin{array}{cc}
-\{(\frac{1}{p}-1)-\epsilon\} & \mbox{with probability }p,\\
1 & \mbox{with probability }(1-p).
\end{array}
$$
As discussed earlier, to make a profit, betting houses will always offer an amount which is less than what they suppose to pay. Here $\epsilon$ is that profitable amount. Suppose the betting houses announce odds are ${P_{H}^*}$, ${P_{A}^*}$ and ${P_{D}^*}$ respectively for Hometeam, Away team and Draw of a match. These announce odds can be defined as below:
$$\Big(\frac{1}{P_H}-\epsilon\Big)=\frac{1}{P_{H}^*},~~
\Big(\frac{1}{P_A}-\epsilon\Big)=\frac{1}{P_{A}^*},~~
\Big(\frac{1}{P_D}-\epsilon\Big)=\frac{1}{P_{D}^*}$$
Now we all know that probabilities of three mutually exclusive events add up to 1 for fair game, i.e. $P_{H}+P_{A}+P_{D}=1$. But ${P_{H}^*}+{P_{A}^*}+{P_{D}^*}>1$. Let us say $\delta$ is the difference between sum of these two probabilities. Then
\begin{eqnarray*}
&&{P_H^*}+{P_A^*}+{P_D^*}-({P_H}+{P_A}+{P_D})=\delta\\
&\Rightarrow
&{P_H^*}+{P_A^*}+{P_D^*}-(\frac {P_H^*}{1+\epsilon P_H^*}+\frac {P_A^*}{1+\epsilon P_A^*}+ \frac {P_D^*}{1+\epsilon P_D^*})=\delta\\
&\Rightarrow
&{P_H^*}-\frac {P_H^*}{1+\epsilon P_H^*}+{P_A^*}-\frac {P_A^*}{1+\epsilon P_A^*}+{P_D^*}-\frac {P_D^*}{1+\epsilon P_D^*}=\delta\\
&\Rightarrow
&{P_H^*}(1-\frac{1}{1+\epsilon{P_H^*}})+{P_A^*}(1-\frac{1}{1+\epsilon{P_A^*}})+{P_D^*}(1-\frac{1}{1+\epsilon {P_D^*}})=\delta\\
&\Rightarrow
&\frac {\epsilon (P_H^*)^2}{1+\epsilon P_H^*} + \frac {\epsilon (P_A^*)^2}{1+\epsilon P_A^*} + \frac {\epsilon (P_D^*)^2}{1+\epsilon P_D^*}= \delta.
\end{eqnarray*}
Clearly we can estimate $\epsilon$ from the above equation as ${P_H^*},{P_A^*},{P_D^*}$ and $\delta$ are already known to us. Let us solve the equation for $\epsilon$ using the given EPL data. Considering the first match for B365 betting house where $\{P_H^*\}=0.50$, $\{P_A^*\}$=0.25, $\{P_D^*\}$=0.28 and $\delta$=1.03-1=0.03. Now solving the above equation with the help of this numerical values,
\begin{eqnarray*}
&\epsilon(\frac {0.50^2}{1+\epsilon (0.50)} + \frac {0.25^2}{1+\epsilon (0.25)} + \frac {0.28^2}{1+\epsilon (0.28)})=0.03.
\end{eqnarray*}
Here we can see the equation is cubic in $\epsilon$. Instead solving analytically which is complicated, we are introducing \texttt{R} programming to optimise $\epsilon$.
Therefore calculate the value of $\epsilon$ is 0.065 or following the revenue model we can say \$0.065 per dollar is the profitable amount for the betting house.
\subsection{Comparative study of Imputed cost among different betting houses: }
Let us consider the result presented in Table (\ref{table_imputed_cost}) and figure (\ref{fig_imputed_cost}).
\begin{enumerate}
\item All betting houses are limited to \$0.10 seems like there is some regulatory policy
\item For $95\%$ of the matches B365 charges between \$.04 to \$.06.
\item For majority games and approximately for all matches BW and IW respectively charge \$0.10.
\item Same behavior we can see for LB also.
\item But VC never charges greater than \$.09. Moreover, for majority matches, they charge between \$.04 to \$.06.
\item B365 and VC has similar kind of behavior.
\end{enumerate}
\section{Conclusion}
In this paper, we presented how the strategy of the betting houses can be explained with the probability odds. From the real data, it is visible that for all betting houses considered here, the sum of the probability for the win, loss, and draw of a match are greater than one. We also presented how the cost of a game can be imputed numerically. Therefore, with this understanding, one realizes how the betting houses are consistently making money and accordingly decides whether to invest one's money on the betting house.
| cac7c5518e8d8ffa4eec5ea5332cbd243f909002 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
In this note, we want to contribute towards the extension theory of dissipative operators $A$ of the form $A=S+iV$, where $S$ and $V\geq 0$ are both symmetric but neither of them needs to be selfadjoint or essentially selfadjoint.
In this sense, we will obtain a more general result than that of Crandall and Phillips \cite{Crandall}, who considered dissipative operators $A$ that were of the form\footnote{In \cite{Crandall}, a densely defined operator is called dissipative if its numerical range is confined to the left half plane $\Pi_-:=\{z\in{\mathbb C}: {\mbox{Re}}(z)\leq 0\}$. Since we will call an operator dissipative if its numerical range is confined to the upper complex plane, we have changed the presentation of the results in \cite{Crandall} accordingly.}
$A=S+iV$,
where $S$ is symmetric, but $V\geq 0$ is assumed to be selfadjoint. However, we have stricter conditions on the domains of our extensions.
\subsection{Dissipative operators and extension theory}
The study of non-selfadjoint operators has proven itself to be a very fruitful field of mathematical research. For an introduction into the many new phenomena and problems that arise if one gives up the condition of selfadjointness, we refer the interested reader to the classic monograph \cite{GK} and the references therein. We mention in particular the work of Brodskii and {Liv\v sic} who addressed questions such as the completeness of root vectors and introduced characteristic matrix functions and triangular models of non-selfadjoint operators \cite{BL58, Livs46, Livs54}.
In what follows, we will call a densely defined operator $A$ on a Hilbert space $\mathcal{H}$ \emph{dissipative} if and only if its numerical range is confined to the upper complex plane, i.e.\ if and only if
$${\mbox{Im}}\langle \psi,A\psi\rangle\geq 0$$
for any $\psi\in\mathcal{D}(A)$. Note that we have defined the sesquilinear form $\langle\cdot,\cdot\rangle$ to be antilinear in the first and linear in the second component. Moreover, we call a dissipative operator $A$ \emph{maximally dissipative} if it has no non-trivial dissipative operator extension, i.e. $A$ being maximally dissipative and $B$ being a dissipative operator extension of $A$ implies that $A=B$. Maximally dissipative operators possess various interesting features, e.g.\ they generate strongly continuous semigroups of contractions \cite{Phillips59} and always have a selfadjoint dilation \cite{NagyFoias}.
Thus, the theory of dissipative extensions of a given operator is an extensively studied problem (for an overview, we recommend the surveys \cite{Arli,AT2009} and all the references therein). Besides the classical results of von Neumann on the theory of selfadjoint extensions of a given symmetric operator \cite{vNeumann} and of Kre\u\i{n}, Birman, Vishik and Grubb on positive selfadjoint and maximally sectorial extensions of a given symmetric operator with positive numerical range \cite{Krein, Vishik, Birman, Grubb68, Alonso-Simon, Grubbfriendly}, let us also mention the results of authors like Arlinski\u\i , Belyi, Derkach, Kovalev, Malamud, Mogilevskii and {Tsekanovski\u\i} \cite{Arlinskii95, ABT2011, Kovalev, AT2005, DM91, DMT88, MM97, MM99, MM02, Tsek80, Tsek81} who have made many contributions using form methods and boundary triples in order to determine maximally sectorial and maximally accretive extensions of a given sectorial operator.\footnote{A densely defined operator $A$ is called (maximally) accretive if $(iA)$ is (maximally) dissipative. If in addition, there exists a $\phi\neq \pi/2$ such that $(e^{i\phi}A)$ is (maximally) dissipative as well, then it is called (maximally) sectorial.} Let us also mention examples, where explicit computations of maximally dissipative (resp.\ accretive) extensions for positive symmetric differential operators \cite{EvansKnowles85}, \cite{EvansKnowles86} and for sectorial Sturm-Liouville operators \cite{BrownEvans} have been made.
For the general problem of finding dissipative extensions of truly dissipative operators, Phillips showed that -- via the Cayley-transform and its inverse -- this is equivalent to finding contractive extensions of a non-densely defined contraction. This problem has been solved by Crandall \cite[Thm.\ 1 and Cor.\ 1]{CrandallContraction} who therefore has provided a full solution to the extension problem (note also the results in \cite{AG}). Crandall established that if $C$ is a contraction defined on a closed subspace $\mathcal{C}$ of a Hilbert space $\mathcal{H}$ and mapping to $\mathcal{H}$, all contractive extensions $\widetilde{C}$ of $C$ can be described via
\begin{equation*}
\widetilde{C}=CP_\mathcal{C}+({\mathbbm{1}}-CP_\mathcal{C}(CP_\mathcal{C})^*)^{1/2}B({\mathbbm{1}}-P_\mathcal{C})\:,
\end{equation*}
where $P_\mathcal{C}$ is the orthogonal projection onto $\mathcal{C}$ and $B$ is an arbitrary contraction on $\mathcal{H}$. However, for concrete applications, the operators involved in this construction are often very difficult to compute. Thus, in \cite{Crandall}, Crandall and Phillips made extra assumptions on the structure of the considered dissipative operator $A$ and required that it could be written in the form
\begin{equation} \label{eq:crandallrequire}
A=S+iV\:,
\end{equation}
where $S$ is symmetric, $V\geq 0$ is selfadjoint and $\mathcal{D}(A)=\mathcal{D}(S)=\mathcal{D}(V)$. Let us briefly describe their approach in the next section.
\subsection{The construction of Crandall and Phillips}
For the case \eqref{eq:crandallrequire} considered by Crandall and Phillips, it follows from non-negativity and selfadjointness of $V$ that the operator $({\mathbbm{1}}+V)$ is a boundedly invertible bijection from $\mathcal{D}(V)$ onto $\mathcal{H}$. They then introduce the weighted Hilbert space $\mathcal{H}_{+1}$ which is the linear space $\mathcal{D}(V^{1/2})$ equipped with the inner product $\langle f,g\rangle_{+1}:=\langle ({\mathbbm{1}}+V)^{1/2}f,({\mathbbm{1}}+V)^{1/2}g\rangle$. Using standard ideas of the construction of Gel'fand triples, they associate every element $f$ of $\mathcal{H}$ with an element $\ell_f$ of the dual space $\mathcal{H}_{+1}^*$ of $\mathcal{H}_{+1}$ via
\begin{equation*}
\ell_f(g):=\langle f,g\rangle\quad\text{for any}\quad g\in\mathcal{H}_{+1}\:,
\end{equation*}
which has norm equal to
\begin{align*}
\|\ell_f\|&=\|({\mathbbm{1}}+V)^{-1/2}f\|=:\|f\|_{-1}\:.
\end{align*}
The space $\mathcal{H}_{-1}$ is then obtained as the completion of $\mathcal{H}$ in $\mathcal{H}_{+1}^*$ with respect to $\|\cdot\|_{-1}$. Since for any $f\in\mathcal{D}(V^{1/2})$ and for any $g\in\mathcal{H}$ we have that $\|f\|\leq\|({\mathbbm{1}}+V)^{1/2}f\|=\|f\|_{+1}$ and $\|g\|\geq \|({\mathbbm{1}}+V)^{-{1/2}}g\|=\|g\|_{-1}$, we obtain the following inclusions:
\begin{equation*}
\mathcal{H}_{+1}\subset\mathcal{H}\subset\mathcal{H}_{-1}\:.
\end{equation*}
In particular, this implies that $V$ is bounded as an operator from $\mathcal{H}_{+1}$ to $\mathcal{H}_{-1}$ --- a feature which Crandall and Phillips use in order to determine all maximally dissipative extensions of $A$ as an operator from $\mathcal{H}_{+1}$ to $\mathcal{H}_{-1}$ \cite[Thm.\ 1.1]{Crandall}. Having obtained a maximally dissipative operator $\widehat{A}$ from $\mathcal{H}_{+1}$ to $\mathcal{H}_{-1}$, they then construct a dissipative extension $\widehat{A}^0$ of $A$ (as an operator in $\mathcal{H}$) via
\begin{align*}
\widehat{A}^0:\quad\mathcal{D}(\widehat{A}^0)=\{f\in\mathcal{D}(\widehat{A}): \widehat{A}f\in\mathcal{H}\},\quad\widehat{A}^0f:=\widehat{A}f\:.
\end{align*}
If $V$ is bounded, this provides a full characterization of all maximally dissipative extensions of $A$, since the spaces $\mathcal{H}_{+1},\mathcal{H}$ and $\mathcal{H}_{-1}$ are equivalent in this case. For the unbounded case, this construction yields dissipative extensions of $A$ that have domain contained in $\mathcal{D}(V^{1/2})$, which does not always provide a full description of all maximally dissipative extensions of $A$ (cf.\ \cite[Example 2]{Crandall}). Also, even if $\widehat{A}$ is a maximally dissipative operator from $\mathcal{H}_{+1}$ to $\mathcal{H}_{-1}$, it is possible that $\widehat{A}^0$ is not a maximally dissipative operator in $\mathcal{H}$ \cite[Example 1]{Crandall}. However, Crandall and Phillips prove a necessary and sufficient condition for when all maximally dissipative extensions $\widehat{A}$ from $\mathcal{H}_{+1}$ to $\mathcal{H}_{-1}$ induce also a maximally dissipative extension $\widehat{A}^0$ in $\mathcal{H}$ \cite[Thm. 3.3]{Crandall}.
\subsection{Our approach}
In a previous note \cite{FNW}, we considered so called dual pairs of operators $(A,\widetilde{A})$, where $A$ and $(-\widetilde{A})$ were assumed to be dissipative and to possess a common core, which means that there exists a linear space $\mathcal{D}\subset\mathcal{D}(A)\cap\mathcal{D}(\widetilde{A})$ such that $A=\overline{A\upharpoonright_\mathcal{D}}$ and $\widetilde{A}=\overline{\widetilde{A}\upharpoonright_\mathcal{D}}$. Given $A=S+iV$, we will define $\widetilde{A}:=S-iV$ and show that $(A,\widetilde{A})$ is such a dual pair (with common core). We then show that considering such dual pairs is an equivalent point of view to assuming that $A$ is of the form $A=S+iV$ (Lemma \ref{thm:ccore}). In this sense, our results are going to be an extension of \cite{FNW}, where we gave a criterion to determine whether an extension $\widehat{A}$ with the property that $A\subset\widehat{A}\subset\widetilde{A}^*$ is dissipative, since we will drop the requirement that $\widehat{A}\subset\widetilde{A}^*$, while keeping the condition that $\mathcal{D}(\widehat{A})\subset\mathcal{D}(\widetilde{A}^*)$, i.e.\ the restrictions on the domain of $\widehat{A}$ remain but the action of the extensions $\widehat{A}$ may differ from that of $\widetilde{A}^*$.
As it turns out, the square-roots of the selfadjoint Friedrichs and Kre\u\i n-von Neumann extensions of $V$ -- denoted by $V_F^{1/2}$, respectively by $V_K^{1/2}$, will play an important part in the presentation of our main result (Theorem \ref{thm:alter}). In particular, we will single out three cases in which it will be possible to simplify the result of Theorem \ref{thm:alter} and express the necessary and sufficient condition for an extension $\widehat{A}$ to be dissipative in terms that only involve $V_F^{1/2}$ and $V_K^{1/2}$ in terms of the quadratic forms $\psi\mapsto\|V_*^{1/2}\psi\|^2$, where $*\in\{F,K\}$ -- a feature which makes it accessible to direct calculation. These three cases are given by (i) an additional restriction on the action of $\widehat{A}$, (ii) $V\geq \varepsilon > 0$, and (iii) $V$ has only one non-negative selfadjoint extension (i.e. $V_F=V_K$). We will also discuss the interplay between boundary conditions (determined by the choice of $\mathcal{D}(\widehat{A})$) and the ``deviation" of $\widehat{A}$ from being a ``proper" extension of $(A,\widetilde{A})$ -- (determined by $(\widetilde{A}^*-\widehat{A})\upharpoonright_{\mathcal{D}(\widehat{A})}$). We will show that there is a fundamentally different behavior between the case that $V$ has only one non-negative selfadjoint extension $V_F=V_K$ (Corollary \ref{coro:ausmachen}) and the case that $V_F\neq V_K$ (Example \ref{ex:1493}).
\section{Some definitions and previous results}
We start with a few basic definitions and results on dissipative operators.
Firstly, let us state a lemma on by how many linearly independent vectors the domain of a given closed dissipative operator with finite defect index has to increase in order to obtain a maximally dissipative extension.
\begin{lemma}[Mentioned in \cite{Crandall}, see also \cite{thesis} for a proof.] \label{prop:dampfnudel}
Let $A$ be a closed and dissipative operator on a separable Hilbert space $\mathcal{H}$ such that $\dim\text{\emph{ker}}(A^*-i)<\infty$. Moreover, let $\widehat{A}$ be a dissipative extension of $A$. Then, $\widehat{A}$ is maximally dissipative if and only if
\begin{equation*}
\dim \mathcal{D}(\widehat{A})/{\mathcal{D}(A)}=\dim\text{\emph{ker}}(A^*-i)\:.
\end{equation*}
\end{lemma}
Next, let us introduce some convenient notation for complementary subspaces:
\begin{definition} Let $\mathcal{N,M}$ be (not necessarily closed) linear spaces such that $\mathcal{M}\subset\mathcal{N}$. With the notation $\mathcal{N}//\mathcal{M}$ we mean any subspace of $\mathcal{N}$, which is complementary to $\mathcal{M}$, i.e.
\begin{equation*}
(\mathcal{N}//\mathcal{M})+\mathcal{M}=\mathcal{N}\quad\text{and}\quad(\mathcal{N}//\mathcal{M})\cap\mathcal{M}=\{0\}\:.
\end{equation*}
\end{definition}
Finally, we will need the characterization of the Kre\u\i n-von Neumann extension of a given non-negative symmetric operator $V$, which has been shown by Ando and Nishio.
\begin{proposition}[{\cite[Thm.\ 1]{Ando-Nishio}}]
Let $V$ be a non-negative closed symmetric operator. The selfadjoint and non-negative square root of the Kre\u\i n--von Neumann extension of $V$, which we denote by $V_K^{1/2}$, can be characterized as follows:
\begin{align*}
\mathcal{D}(V_K^{1/2})=\left\{h\in\mathcal{H}: \sup_{f\in\mathcal{D}(V):Vf\neq 0}\frac{|\langle h,Vf\rangle|^2}{\langle f,Vf\rangle}<\infty\right\}\:,\\
\text{for any}\:\: h\in\mathcal{D}(V_K^{1/2}):\quad\|V_K^{1/2}h\|^2= \sup_{f\in\mathcal{D}(V):Vf\neq 0}\frac{|\langle h,Vf\rangle|^2}{\langle f,Vf\rangle}\:.
\end{align*} \label{thm:nett}
\end{proposition}
For our purposes, it will be more convenient to use the following characterization of $\mathcal{D}(V_K^{1/2})$ and $\|V_K^{1/2}h\|$:
\begin{corollary} \label{coro:andonishio}
Let $V$ be a non-negative closed symmetric operator on a Hilbert space $\mathcal{H}$. Then, the square root of its Kre\u\i n--von Neumann extension can be characterized as follows
\begin{align} \label{eq:mjunitsch}
\mathcal{D}(V_K^{1/2})=\left\{h\in\mathcal{H}: \sup_{g\in\text{\emph{ran}}(\widehat{V}^{1/2}\upharpoonright_{\mathcal{D}(V)}):\|g\|=1}{|\langle h,\widehat{V}^{1/2}g\rangle|}<\infty\right\}\:,\\
\text{for any}\:\: h\in\mathcal{D}(V_K^{1/2}):\quad\|V_K^{1/2}h\|^2=\sup_{g\in\text{\emph{ran}}(\widehat{V}^{1/2}\upharpoonright_{\mathcal{D}(V)}):\|g\|=1}{|\langle h,\widehat{V}^{1/2}g\rangle|^2}\:,
\end{align}
where $\widehat{V}$ is any non-negative selfadjoint extension of $V$.
\end{corollary}
\begin{proof}
Let us consider any $f\in\mathcal{D}(V)$ such that $Vf\neq 0$. Since $Vf=\widehat{V}f=\widehat{V}^{1/2}\widehat{V}^{1/2}f$, we then get
\begin{equation*}
\frac{|\langle h,Vf\rangle|^2}{\langle f,Vf\rangle}=\frac{|\langle h,\widehat{V}^{1/2}\widehat{V}^{1/2}f\rangle|^2}{\|\widehat{V}^{1/2}f\|^2}=\left|\left\langle h,\widehat{V}^{1/2}\left(\frac{\widehat{V}^{1/2}f}{\|\widehat{V}^{1/2}f\|}\right)\right\rangle\right|^2\:.
\end{equation*}
Now, observe that $\frac{\widehat{V}^{1/2}f}{\|\widehat{V}^{1/2}f\|}$ is a normalized element of ${\mbox{ran}}(\widehat{V}^{1/2}\upharpoonright_{\mathcal{D}(V)})$. Conversely, for any normalized $g\in{\mbox{ran}}(\widehat{V}^{1/2}\upharpoonright_{\mathcal{D}(V)})$, there exists a $f\in\mathcal{D}(V)$ with $Vf\neq 0$ such that $g=\frac{\widehat{V}^{1/2}f}{\|\widehat{V}^{1/2}f\|}$. This implies that
\begin{equation*}
\sup_{f\in\mathcal{D}(V):Vf\neq 0}\left|\left\langle h,\widehat{V}^{1/2}\left(\frac{\widehat{V}^{1/2}f}{\|\widehat{V}^{1/2}f\|}\right)\right\rangle\right|^2=\sup_{g\in\text{ran}(\widehat{V}^{1/2}\upharpoonright_{\mathcal{D}(V)}):\|g\|=1}|\langle h,\widehat{V}^{1/2}g\rangle|^2\:,
\end{equation*}
which --- together with Proposition \ref{thm:nett} --- yields the corollary.
\end{proof}
\section{The common core property}
Given any bounded operator $A$, the decomposition into its selfadjoint real part $S:=(A+A^*)/2$ and selfadjoint imaginary part $V:=(A-A^*)/(2i)$ allows us to always write $A$ as $A=S+iV$. For the unbounded case, this is generally not possible as one has to be careful with the domains. However, in the case that it is possible to decompose $A$ as
\begin{equation} \label{eq:decomp}
A=S+iV\:,
\end{equation}
where both $S$ and $V\geq 0$ are symmetric and $\mathcal{D}(A)=\mathcal{D}(S)= \mathcal{D}(V)$, one can use the framework of dual pairs $(A,\widetilde{A})$ of operators to decompose $A$ analogously as in the bounded case. To this end, let us firstly recall their definition (see also \cite{Edmunds-Evans, LyantzeStorozh} for more details):
\begin{definition} Let $(A,\widetilde{A})$ be a pair of densely defined and closable operators. We say that they form a {\bf{dual pair}} if
\begin{equation*}
A\subset \widetilde{A}^*\quad\text{resp.}\quad \widetilde{A}\subset A^*\:.
\end{equation*}
In this case, $A$ is called a {\bf{formal adjoint}} of $\widetilde{A}$ and vice versa. Moreover, an operator $\widehat{A}$ such that $A\subset\widehat{A}\subset\widetilde{A}^*$ is called a {\bf{proper extension}} of the dual pair $(A,\widetilde{A})$.
\end{definition}
It is then not hard to see that with the choice $\widetilde{A}:=S-iV$ ($\mathcal{D}(\widetilde{A})=\mathcal{D}(S)=\mathcal{D}(V)$), we have that $(A,\widetilde{A})$ is a dual pair since
\begin{equation} \label{eq:kirchedorf}
\langle f,\widetilde{A}g\rangle=\langle f, (S-iV)g\rangle=\langle (S+iV)f,g\rangle=\langle Af,g\rangle
\end{equation}
for any $f\in\mathcal{D}(S+iV)$ and any $g\in\mathcal{D}(S-iV)$.
For the presentation of our results in \cite{FNW}, the notion of dual pairs $(A,\widetilde{A})$ satisfying the so called common core property was particularly useful. Let us restate the definition.
\begin{definition}
Let $(A,\widetilde{A})$ be a dual pair of closed operators. We say that it has the {\bf{common core property}} if there exists a subset $\mathcal{D}\subset\mathcal{D}(A)\cap\mathcal{D}(\widetilde{A})$ such that it is a core for $A$ as well as for $\widetilde{A}$: $$\overline{A\upharpoonright_{\mathcal{D}}}=A\quad\text{and}\quad\widetilde{A}=\overline{\widetilde{A}\upharpoonright_{\mathcal{D}}}\:.$$
\end{definition}
We are now prepared to show the link between dual pairs $(A,\widetilde{A})$ satisfying the common core property and dissipative operators $A$ that can be decomposed according to \eqref{eq:decomp}.
\begin{lemma} \label{thm:ccore} Let $(A,\widetilde{A})$ be a dual pair of closed operators satisfying the common core property with a common core $\mathcal{D}$, where $A$ is dissipative. Then there exist two symmetric operators $S$ and $V\geq 0$ with $\mathcal{D}=\mathcal{D}(S)=\mathcal{D}(V)$ such that
\begin{equation*}
A\upharpoonright_\mathcal{D}=S+iV\quad\text{and}\quad\widetilde{A}\upharpoonright_\mathcal{D}=S-iV\:.
\end{equation*}
Conversely, let $A$ be a dissipative operator of the form $A=S+iV$, where $S$ and $V\geq 0$ are symmetric operators and $\mathcal{D}(A)=\mathcal{D}(S)=\mathcal{D}(V)$. If we define $\widetilde{A}:=S-iV$, where $\mathcal{D}(\widetilde{A})=\mathcal{D}$, then their closures $(\overline{A},\overline{\widetilde{A}})$ form a dual pair that has the common core property.
\end{lemma}
\begin{proof} If $(A,\widetilde{A})$ is a dual pair satisfying the common core condition, with $\mathcal{D}$ being a common core, we may define
\begin{equation}
S:=\frac{A+\widetilde{A}}{2}\upharpoonright_\mathcal{D}\quad\text{and}\quad V:=\frac{A-\widetilde{A}}{2i}\upharpoonright_\mathcal{D}\;.
\end{equation}
Firstly, observe that $A\upharpoonright_\mathcal{D}=S+iV$ and $\widetilde{A}\upharpoonright_\mathcal{D}=S-iV$. Next, let us show that $S$ and $V$ are symmetric and also that $V\geq 0$. To this end, let $\psi\in\mathcal{D}$ and consider
\begin{align*}
\langle \psi,S\psi\rangle&=\frac{1}{2}\langle \psi,(A+\widetilde{A})\psi\rangle=\frac{1}{2}(\langle \psi,A\psi\rangle+\langle \psi,A^*\psi\rangle)={\mbox{Re}}(\langle\psi,A\psi\rangle)\in{\mathbb R}\\
\langle \psi,V\psi\rangle&=\frac{1}{2i}\langle \psi,(A-\widetilde{A})\psi\rangle=\frac{1}{2i}(\langle \psi,A\psi\rangle-\langle \psi,A^*\psi\rangle)={\mbox{Im}}(\langle\psi,A\psi\rangle)\geq 0\:,
\end{align*}
where the inequality follows from dissipativity of $A$.
Now, let $A$ be a dissipative operator of the form $A=S+iV$, where $S$ and $V\geq 0$ are symmetric and $\mathcal{D}(A)=\mathcal{D}(S)=\mathcal{D}(V)$. In \eqref{eq:kirchedorf}, we have already shown that $A$ and $\widetilde{A}:=S-iV$ form a dual pair and so do their closures $(\overline{A},\overline{\widetilde{A}})$, which therefore is a dual pair that has the common core property with common core $\mathcal{D}=\mathcal{D}(S)=\mathcal{D}(V)$.
\end{proof}
\section{The main result}
We are now prepared to prove our main result. Before we proceed, we need to show the following two lemmas:
\begin{lemma} \label{lemma:friedrichsdense}
Let $V$ be a non-negative symmetric operator. Then
\\
i) $\text{\emph{ran}}(V_F^{1/2}\upharpoonright_{\mathcal{D}(V)})$ is dense in $\overline{\text{\emph{ran}}(V_F^{1/2})}$
ii) $\text{\emph{ran}}(V_K^{1/2}\upharpoonright_{\mathcal{D}(V)})$ is dense in $\overline{\text{\emph{ran}}(V_K^{1/2})}$.
\end{lemma}
\begin{proof}
i) By construction of the Friedrichs extension, we know that for any $\psi\in\mathcal{D}(V_F^{1/2})$, there exists a sequence $\{\psi_n\}\subset\mathcal{D}(V)$, such that
\begin{equation*}
\lim_{n\rightarrow\infty}(\|\psi-\psi_n\|^2+\|V_F^{1/2}(\psi-\psi_n)\|^2)=0\:,
\end{equation*}
which implies in particular that $\lim_{n\rightarrow\infty} V_F^{1/2}\psi_n=V_F^{1/2}\psi$, i.e. ${\mbox{ran}}(V_F^{1/2})\subset\overline{{\mbox{ran}}(V_F^{1/2}\upharpoonright_{\mathcal{D}(V)})}$. On the other hand, since ${\mbox{ran}}(V_F^{1/2}\upharpoonright_{\mathcal{D}(V)})\subset{\mbox{ran}}(V_F^{1/2})$, the assertion follows from taking closures.
ii) Any element of ${\mbox{ran}}(V_K^{1/2})$ is of the form $V_K^{1/2}h$, where $h\in\mathcal{D}(V_K^{1/2})$. By Corollary \ref{coro:andonishio} with the choice $\widehat{V}=V_K$, we have that
\begin{align} \label{eq:jeremy}
\|V_K^{1/2}h\|^2=\sup\left\{|\langle V_K^{1/2}h,g\rangle|^2, g\in{\mbox{ran}}(V_K^{1/2}\upharpoonright_{\mathcal{D}(V)}):\|g\|=1\right\}\:.
\end{align}
But this implies that ${\mbox{ran}}(V_K^{1/2}\upharpoonright_{\mathcal{D}(V)})$ is dense in $\overline{{\mbox{ran}}(V_K^{1/2})}$. To see why, assume that there exists a $\varphi\in \overline{{\mbox{ran}}(V_K^{1/2})}$ such that $\|\varphi\|=1$ and
$\langle \varphi,g\rangle=0$ for all $g\in{\mbox{ran}}(V_K^{1/2}\upharpoonright_{\mathcal{D}(V)})$. Take a $V_K^{1/2}h\in{\mbox{ran}}(V_K^{1/2})$, with $\|V_K^{1/2}h\|=1$ such that $\|V_K^{1/2}h-\varphi\|^2<\varepsilon$ for some $0<\varepsilon<1$ small enough. Then, for any $g\in{\mbox{ran}}(V_K^{1/2}\upharpoonright_{\mathcal{D}(V)})$, we get
\begin{equation*}
|\langle V_K^{1/2}h,g\rangle|^2=|\langle V_K^{1/2}h-\varphi,g\rangle|^2\leq \|V_K^{1/2}h-\varphi\|^2\|g\|^2\leq \varepsilon\|g\|^2\:.
\end{equation*}
Taking the supremum over all $g\in{\mbox{ran}}(V_K^{1/2}\upharpoonright_{\mathcal{D}(V)})$ with $\|g\|=1$, we arrive at a contradiction, since the supremum of the left hand side is $1$ whereas the supremum of the right hand side is $\varepsilon<1$. This shows the lemma.
\end{proof}
\begin{lemma} \label{lemma:frischerfisch} Let $V$ be a non-negative symmetric operator and let $V_F$ and $V_K$ denote its Friedrichs, resp. its Kre\u\i n-von Neumann extension. Then there exists a partial isometry $\mathcal{U}$ on $\mathcal{H}$ such that
\begin{equation} \label{eq:kuckuck}
V_K^{1/2}h=\mathcal{U}{V}_F^{1/2}h
\end{equation}
for all $h\in\mathcal{D}(V_F^{1/2})$. The map $\mathcal{U}$ is an isometry on $\overline{{\mbox{ran}}(V_F^{1/2})}$ and its range ${\mbox{ran}}(\mathcal{U})$ is contained in $\overline{{\mbox{ran}}(V_K^{1/2})}$.
\end{lemma}
\begin{proof} Since we have that $V_K\leq V_F$, it is clear that $\mathcal{D}(V_F^{1/2})\subset\mathcal{D}(V_K^{1/2})$. Moreover, by Proposition \ref{thm:nett}, for any $h\in\mathcal{D}(V_F^{1/2})\subset\mathcal{D}(V_K^{1/2})$, we have that
\begin{align*}
\|V_K^{1/2}h\|^2&=\sup_{f\in\mathcal{D}(V):Vf\neq 0}\frac{|\langle h,Vf\rangle|^2}{\langle f,Vf\rangle}=\sup_{f\in\mathcal{D}(V):Vf\neq 0}\frac{|\langle h,V_F^{1/2}V_F^{1/2}f\rangle|^2}{\langle f,V_F^{1/2}V_F^{1/2}f\rangle}\\&=\sup_{f\in\mathcal{D}(V):Vf\neq 0}\frac{|\langle V_F^{1/2} h,V_F^{1/2}f\rangle|^2}{\|V_F^{1/2}f\|^2}=\|V_F^{1/2}h\|^2\:,
\end{align*}
where we have used that ${\mbox{ran}} (V_F^{1/2}\upharpoonright_{\mathcal{D}(V)})$ is dense in $\overline{{\mbox{ran}}(V_F^{1/2})}$ by Lemma \ref{lemma:friedrichsdense}.
This implies that the linear map
\begin{align*}
\mathcal{U}_0: \quad {\mbox{ran}}(V_F^{1/2})&\rightarrow{\mbox{ran}}\left(V_K^{1/2}\upharpoonright_{\mathcal{D}(V_F^{1/2})}\right)\\
V_F^{1/2}h&\mapsto V_K^{1/2}h
\end{align*}
is isometric. Since, trivially, ${\mbox{ran}}(V_F^{1/2})$ is dense in $\overline{{\mbox{ran}}(V_F^{1/2})}$, there exists a unique isometric extension $\mathcal{U}_0\subset\mathcal{U}$ on $\overline{{\mbox{ran}}(V_F^{1/2})}$. Setting
$\mathcal{U}k=0$ for all $k\in\ker(V_F^{1/2})={\mbox{ran}}(V_F^{1/2})^\perp$ defines $\mathcal{U}$ as a partial isometry on the whole Hilbert space $\mathcal{H}$. Here, $\mathcal{M}^\perp$ denotes the orthogonal complement of a linear space $\mathcal{M}$. Moreover, since $${\mbox{ran}}(\mathcal{U}_0)={\mbox{ran}}\left(V_K^{1/2}\upharpoonright_{\mathcal{D}(V_F^{1/2})}\right)\subset{\mbox{ran}}(V_K^{1/2})\:,$$ this implies that ${\mbox{ran}}(\mathcal{U})$ is contained in $\overline{{\mbox{ran}}(V_K^{1/2})}$ and thus the lemma.
\end{proof}
Given a dual pair $(A,\widetilde{A})$, let us introduce the following convenient way of parametrizing all extensions of $A$ which have domain contained in $\mathcal{D}(\widetilde{A}^*)$:
\begin{definition} \label{def:quantz}
Let $(A,\widetilde{A})$ be a dual pair, where $A$ is dissipative and $\widetilde{A}$ is antidissipative. Let $\mathcal{V}\subset\mathcal{D}(\widetilde{A}^*)//\mathcal{D}(A)$ be a linear space, which means that $\mathcal{V}\subset\mathcal{D}(\widetilde{A}^*)$ and $\mathcal{V}\cap\mathcal{D}(A)=\{0\}$. Moreover, let $\mathcal{L}$ be a linear operator from $\mathcal{V}$ into $\mathcal{H}$. Then, the operator $A_{\mathcal{V},\mathcal{L}}$ is given by
\begin{align*}
A_{\mathcal{V},\mathcal{L}}:\qquad\mathcal{D}(A_{\mathcal{V},\mathcal{L}})&=\mathcal{D}(A)\dot{+}\mathcal{V}\\
(f+v)&\mapsto \widetilde{A}^*(f+v)+\mathcal{L}v\:,
\end{align*}
where $f\in\mathcal{D}(A)$ and $v\in\mathcal{V}$. If $\mathcal{L}$ is the zero-operator, i.e. $\mathcal{L}=0$, we define $A_{\mathcal{V},0}=:A_\mathcal{V}$.
\end{definition}
Note that the operator $\mathcal{L}$ can be interpreted as the deviation of $A_{\mathcal{L,V}}$ from $\widetilde{A}^*$, since for any $v\in\mathcal{V}$, we get that
\begin{equation*}
(\widetilde{A}^*-A_{\mathcal{V,L}})v=\mathcal{L}v\:.
\end{equation*}
Let us now show the main theorem:
\begin{theorem} \label{thm:alter} Let $(A,\widetilde{A})$ be a dual pair that has the common core property, where $A$ is dissipative.
Moreover, assume that
$$v\in\mathcal{D}(V_K^{1/2}) \qquad\text{and}\qquad \mathcal{L}v\in \text{\emph{ran}}(V_F^{1/2})=\mathcal{D}(V_F^{-1/2})$$
for all $v\in\mathcal{V}$.
Then, $A_{\mathcal{V},\mathcal{L}}$ is dissipative if and only if for all $v\in\mathcal{V}$ we have
\begin{equation} \label{eq:gstinkert}
{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle\geq\frac{1}{4}\|\mathcal{U}V_F^{-1/2}\mathcal{L}v+2iV_K^{1/2}v\|^2\:.
\end{equation}
Here, $V^{-1/2}_F$ denotes the inverse of $V_F^{1/2}$ as an operator in $\overline{\text{\emph{ran}}(V_F^{1/2})}$, which is given by
\begin{align}
V_F^{-1/2}: \qquad \mathcal{D}(V_F^{-1/2})=\text{\emph{ran}} V_F^{1/2}&\rightarrow\mathcal{D}(V_F^{1/2})\cap\overline{\text{\emph{ran}}(V_F^{1/2})}\notag\\
V_F^{1/2}f&\mapsto f\:,
\label{eq:krakau}
\end{align}
which a well-defined non-negative selfadjoint operator on the Hilbert space $\overline{{\mbox{ran}}(V_F^{1/2})}$.
The operator $\mathcal{U}$ is the partial isometry as defined in Lemma \ref{lemma:frischerfisch}.
\end{theorem}
\begin{proof}
Let us start be showing that the above conditions are sufficient. To this end, let $\mathcal{D}\subset\mathcal{D}(A)\cap\mathcal{D}(\widetilde{A})$ denote a common core for $A$ and $\widetilde{A}$. For any $f\in\mathcal{D}$ and any $v\in\mathcal{V}$, we then get
\begin{align*}
{\mbox{Im}}\langle f&+v,A_{\mathcal{V},\mathcal{L}}(f+v)\rangle={\mbox{Im}}\langle f+v,\widetilde{{A}}^*(f+v)\rangle+{\mbox{Im}} \langle f+v,\mathcal{L}v\rangle\\&=\langle f,Vf\rangle+{\mbox{Im}}\langle v,2iVf\rangle+{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle+{\mbox{Im}}\langle f,\mathcal{L}v\rangle\\
&=\|V_K^{1/2}f\|^2+{\mbox{Im}}\langle v,2iV_K^{1/2}V_K^{1/2}f\rangle+{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle+{\mbox{Im}}\langle V_F^{-1/2}V_F^{1/2}f,\mathcal{L}v\rangle\\
&=\|V_K^{1/2}f\|^2+{\mbox{Im}}\langle V_K^{1/2}v,2iV_K^{1/2}f\rangle+{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle+{\mbox{Im}}\langle \mathcal{U}V_F^{1/2}f, \mathcal{U}V_F^{-1/2}\mathcal{L}v\rangle\\
&=\|V_K^{1/2}f\|^2+{\mbox{Im}}\langle V_K^{1/2}v,2iV_K^{1/2}f\rangle+{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle+{\mbox{Im}}\langle V_K^{1/2}f, \mathcal{U}V_F^{-1/2}\mathcal{L}v\rangle\\
&=\|V_K^{1/2}f\|^2+{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle+{\mbox{Im}}\langle V_K^{1/2}f,(\mathcal{U}V_F^{-1/2}\mathcal{L}+2iV_K^{1/2})v\rangle\\
&\geq \|V_K^{1/2}f\|^2+\frac{1}{4}\|\mathcal{U}V_F^{-1/2}\mathcal{L}v+2iV_K^{1/2}v\|^2+{\mbox{Im}}\langle V_K^{1/2}f,(\mathcal{U}V_F^{-1/2}\mathcal{L}+2iV_K^{1/2})v\rangle\\\
&\geq \|V_K^{1/2}f\|^2+\frac{1}{4}\|\mathcal{U}V_F^{-1/2}\mathcal{L}v+2iV_K^{1/2}v\|^2-\|V_K^{1/2}f\| \|(\mathcal{U}V_F^{-1/2}\mathcal{L}+2iV_K^{1/2})v\|\\
&=\left(\|V_K^{1/2}f\|-\frac{1}{2}\|\mathcal{U}V_F^{-1/2}\mathcal{L}v+2iV_K^{1/2}v\|\right)^2\geq 0\:.
\end{align*}
Let us now show that Condition \eqref{eq:gstinkert} is also necessary. Assume that it is not satisfied, i.e. that there exists a $v\in\mathcal{V}$ such that
\begin{equation} \label{eq:schlechti}
{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle-\frac{1}{4}\|\mathcal{U}V_F^{-1/2}\mathcal{L}v+2iV_K^{1/2}v\|^2\leq -\varepsilon
\end{equation}
for some $\varepsilon>0$. By Lemma \ref{lemma:frischerfisch}, we have that
$(\mathcal{U}V_F^{-1/2}\mathcal{L}v+2iV_K^{1/2}v)\in\overline{{\mbox{ran}}(V_K^{1/2})}$. Thus, by Lemma \ref{lemma:friedrichsdense} ii), there exists a sequence $\{f_n\}\subset{\mathcal{D}(V)}$ such that
$$ V_K^{1/2}f_n\overset{n\rightarrow\infty}{\longrightarrow}\frac{-i}{2}(\mathcal{U}V_F^{-1/2}\mathcal{L}v+2iV_K^{1/2}v)\:,
$$
which means by \eqref{eq:schlechti} that
\begin{align*}
&{\mbox{Im}}\langle f_n+v,A_{\mathcal{V,L}}(f_n+v)\rangle\\&=\|V_K^{1/2}f_n\|^2+{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle+{\mbox{Im}}\langle V_K^{1/2}f_n,\mathcal{U}V_F^{-1/2}\mathcal{L}v+2iV_K^{1/2}v\rangle \overset{n\rightarrow\infty}{\longrightarrow}-\varepsilon<0\:,
\end{align*}
which means that $A_{\mathcal{V,L}}$ is not dissipative in this case. This shows the theorem.
\end{proof}
Note that for the proof of Theorem \ref{thm:alter}, we have assumed that
$$v\in\mathcal{D}(V_K^{1/2}) \qquad\text{and}\qquad \mathcal{L}v\in \text{\emph{ran}}(V_F^{1/2})=\mathcal{D}(V_F^{-1/2})\:.$$
Let us now show that given either condition, the other is necessary for (we will comment on the case that neither condition is satisfied after the proof of the following theorem).
\begin{theorem} Let $(A,\widetilde{A})$ be a dual pair satisfying the common core condition, where $A$ is dissipative.
i) If ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(V_F^{1/2})$, then it is necessary that $\mathcal{V}\subset\mathcal{D}(V_K^{1/2})$ for $A_{\mathcal{V,L}}$ to be dissipative.
ii) If $\mathcal{V}\subset\mathcal{D}(V_K^{1/2})$, then it is necessary that ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(V_F^{1/2})$ for $A_{\mathcal{V,L}}$ to be dissipative.
\label{thm:kuckuckswalzer}
\end{theorem}
\begin{proof}
i) If ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(V_F^{1/2})$, this means that for any $v\in\mathcal{V}$ there exists a $\phi_v\in\mathcal{D}(V_F^{1/2})$ such that $\mathcal{L}v=V_F^{1/2}\phi_v$. Thus, for any $f\in\mathcal{D}$ we can write
\begin{equation} \label{eq:trompete}
{\mbox{Im}}\langle f+v,A_{\mathcal{V,L}}(f+v)\rangle= \|V_F^{1/2}f\|^2+{\mbox{Im}}\langle v,2iV_F^{1/2}V_F^{1/2}f\rangle+{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle-{\mbox{Im}}\langle \phi_v,V_F^{1/2}f\rangle\:.
\end{equation}
Now, assume that there exists a $v\in\mathcal{V}$ such that $v\notin\mathcal{D}(V_K^{1/2})$. By Corollary \ref{coro:andonishio} with the choice $\widehat{V}=V_F$ and Lemma \ref{lemma:friedrichsdense} i), this means that there exists a sequence $\{f_n\}\subset\mathcal{D}(V)$ with $\|V_F^{1/2}f_n\|=1$ for any $n$ and a sequence of complex phases $\{e^{i\varphi_n}\}$ such that
\begin{equation*}
\lim_{n\rightarrow\infty}{\mbox{Im}}\langle v,2ie^{i\varphi_n}V_F^{1/2}V_F^{1/2}f_n\rangle=-\lim_{n\rightarrow\infty}\left|\langle v,2V_F^{1/2}V_F^{1/2}f_n\rangle\right|=-\infty\:.
\end{equation*}
Since all other terms in \eqref{eq:trompete} stay bounded, this shows that $A_{\mathcal{V,L}}$ cannot be dissipative in this case.\\
ii) We start by showing that in this case, it is necessary that $\mathcal{L}v\perp\ker V_F^{1/2}$ for all $v\in\mathcal{V}$. Assume this is not the case, i.e. that there exists a $v\in\mathcal{V}$ and a $k\in\ker (V_F^{1/2})=\ker (V_F)$ such that $\langle \mathcal{L}v,k\rangle\neq 0$. Without loss of generality we may assume that ${\mbox{Im}}\langle \mathcal{L}v,k\rangle=1$. Now, since $\mathcal{D}(V)$ is a core for $V_F^{1/2}$, we can pick a sequence $\{f_n\}\subset\mathcal{D}(V)$ such that $f_n\rightarrow \lambda k$ and $V_F^{1/2}f_n\rightarrow \lambda V_F^{1/2}k=0$, where $\lambda\in{\mathbb C}$ is an arbitrary complex number. We then get
\begin{align*}
&\lim_{n\rightarrow\infty}{\mbox{Im}}\langle (f_n+v,A_{\mathcal{V,L}}(f_n+v)\rangle\\=&\lim_{n\rightarrow\infty}\left(\|V_F^{1/2}f_n\|^2+{\mbox{Im}}\langle v,2iV_K^{1/2}V_K^{1/2}f_n\rangle+{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle-{\mbox{Im}}\langle \mathcal{L}v,f_n\rangle\right)\\
\overset{\eqref{eq:kuckuck}}{=}&\lim_{n\rightarrow\infty}\left(\|V_F^{1/2}f_n\|^2+{\mbox{Im}}\langle V_K^{1/2}v,2i\mathcal{U}V_F^{1/2}f_n\rangle+{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle-{\mbox{Im}}\langle \mathcal{L}v,f_n\rangle\right)\\
=&\lim_{n\rightarrow\infty}\left(\|V_F^{1/2}f_n\|^2+{\mbox{Im}}\langle \mathcal{U}^*V_K^{1/2}v,2iV_F^{1/2}f_n\rangle+{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle-{\mbox{Im}}\langle \mathcal{L}v,f_n\rangle\right)\\
=&\,{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle-{\mbox{Im}}\lambda\:,
\end{align*}
which is negative if we choose ${\mbox{Im}}\lambda$ large enough. This contradicts the dissipativity of $A_{\mathcal{V,L}}$. Hence ${\mbox{ran}}(\mathcal{L})\subset(\ker V_F^{1/2})^\perp=\overline{{\mbox{ran}}(V_F^{1/2})}$. Now, since $\ker V_F^{1/2}$ is a reducing subspace for $V_F^{1/2}$, we have that the operator $V_F^{-1/2}$ is a well-defined non-negative selfadjoint operator on the Hilbert space $\overline{{\mbox{ran}}(V_F^{1/2})}$. Also, note that $\overline{{\mbox{ran}}(V_F^{1/2})}$ reduces $V_F^{1/2}$. Now, assume that there is a $v\in\mathcal{V}$, such that
$\mathcal{L}v\notin{\mbox{ran}}(V_F^{1/2})=\mathcal{D}(V_F^{-1/2})$. This means that we can pick a sequence $\{f_{n}\}\subset{\mathcal{D}(V)}$, where $\|V_F^{1/2}f_{n}\|=1$ for all $n$, such that $$\lim_{n\rightarrow\infty}{\mbox{Im}}\langle \mathcal{L}v,V_F^{-1/2}V_F^{1/2}f_n\rangle=+\infty\:,$$
since otherwise the map $g\mapsto\langle \mathcal{L}v,V_F^{-1/2}g\rangle$ would be a bounded linear functional on ${{\mbox{ran}}(V_F^{1/2}\upharpoonright_{\mathcal{D}(V)})}$, which -- by Lemma \ref{lemma:friedrichsdense} (i) -- is dense in $\overline{{\mbox{ran}}(V_F^{1/2}})$ --- a contradiction to $\mathcal{L}v\notin\mathcal{D}(V_F^{-1/2})$.
Thus, we get
\begin{align*}
&{\mbox{Im}}\langle (f_n+v,A_{\mathcal{V,L}}(f_n+v)\rangle\\
=&\|V_F^{1/2}f_n\|^2+{\mbox{Im}}\langle \mathcal{U}^*V_K^{1/2}v,2iV_F^{1/2}f_n\rangle+{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle-{\mbox{Im}}\langle \mathcal{L}v,f_n\rangle\\
\leq & 1+2\|\mathcal{U}^*V_K^{1/2}v\|+{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle-{\mbox{Im}}\langle \mathcal{L}v,V_F^{-1/2}V_F^{1/2}f_n\rangle\overset{n\rightarrow\infty}{\longrightarrow}-\infty\:,
\end{align*}
which means that $A_{\mathcal{V,L}}$ cannot be dissipative in this case either. This finishes the proof.
\end{proof}
\begin{remark} For the case of proper extensions, i.e. for $\mathcal{L}=0$, Theorems \ref{thm:alter} and \ref{thm:kuckuckswalzer} readily imply our previous result \cite[Thm. 4.7]{FNW}.
\end{remark}
\begin{remark} \normalfont Since for any $f_n\in\mathcal{D}(A)$ and $v\in\mathcal{V}$ we get
\begin{equation} \label{eq:pachelbel}
{\mbox{Im}}\langle f_n+v,A_{\mathcal{V,L}}(f_n+v)\rangle={\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle+\|V_K^{1/2}f_n\|^2+{\mbox{Im}}\langle v,2iV_K^{1/2}V_K^{1/2}f_n\rangle+{\mbox{Im}}\langle V_F^{-1/2}V_F^{1/2}f_n,\mathcal{L}v\rangle\:,
\end{equation}
the condition $\mathcal{V}\subset\mathcal{D}(V_K^{1/2})$ controls the term ${\mbox{Im}}\langle v,2iV_K^{1/2}V_K^{1/2}f_n\rangle$ while the condition ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(V_F^{1/2})$ ensures that the term ${\mbox{Im}}\langle V_F^{-1/2}V_F^{1/2}f_n,\mathcal{L}v\rangle$ can be controlled when minimizing ${\mbox{Im}}\langle f_n+v,A_{\mathcal{V,L}}(f_n+v)\rangle$.
But in the case that neither condition is satisfied it could happen that the last term in \eqref{eq:pachelbel} does not stay bounded either and instead ``competes" against the ${\mbox{Im}}\langle v,2iV_K^{1/2}V_K^{1/2}f_n\rangle$ that would go to $-\infty$ for a suitable choice of a sequence $\{f_n\}$. Thus, in the situation $v\notin\mathcal{D}(V_K^{1/2})$ and $\mathcal{L}v\notin{\mbox{ran}} (V_F^{1/2})$ it is not clear whether it is in general possible that $A_{\mathcal{V,L}}$ is dissipative. Moreover, since it is difficult to compute $V_F^{1/2}$, $V_F^{-1/2}$ and $V_K^{1/2}$ explicitly, we were not able to construct such an example. (The elementary case of $V$ being a multiplication operator or --- more generally --- an essentially selfadjoint operator will be discussed in Lemma \ref{lemma:singulariter}.)
\end{remark}
\section{Applications of the main theorem}
Despite its theoretical merit, Theorem \ref{thm:alter} does not seem to be very useful for practical applications, since it is in general very difficult to explicitly compute the square-roots $V_K^{1/2}$ and $V_F^{-1/2}$ as they occur in the statement of the theorem. Moreover, we do not have explicit knowledge of the partial isometry $\mathcal{U}$. In this section, we are therefore going to single out three situations in which Condition \eqref{eq:gstinkert} can be simplified and made accessible to direct computations.
\subsection{An additional restriction on ${\mbox{ran}}(\mathcal{L})$} \label{subsec:additionalreq} For the statement of Theorem \ref{thm:alter}, we have assumed that ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(V_F^{1/2})$. If we make the even stricter assumption that ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(V_F)$, we can simplify the result of Theorem \ref{thm:alter}:
\begin{corollary} \label{coro:wolfratshausen} Let $(A,\widetilde{A})$ be dual pair satisfying the common core property, where $A$ is dissipative. Moreover, assume that ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(V_F)$. In this case, we write $\mathcal{L}v=V_F\phi_v$, where $\phi_v\in\mathcal{D}(V_F)$, which is determined up to elements in $\ker(V_F)$. Then, $A_\mathcal{V,L}$ is dissipative if and only if $\mathcal{V}\subset\mathcal{D}(V_K^{1/2})$ and for all $v\in\mathcal{V}$, we have that
\begin{equation} \label{eq:scan3}
{\mbox{Im}}\langle v,\widetilde{A}^*v\rangle+{\mbox{Im}}\langle v,V_F\phi_v\rangle\geq\frac{1}{4}\|V_K^{1/2}(\phi_v+2iv)\|^2\:.
\end{equation}
\end{corollary}
\begin{proof} Since ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(V_F)\subset{\mbox{ran}}(V_F^{1/2})$ by assumption, it follows from Theorem \ref{thm:kuckuckswalzer} i), that it is necessary that $\mathcal{V}\subset\mathcal{D}(V_K^{1/2})$ for $A_\mathcal{V,L}$ to be dissipative. Condition \eqref{eq:scan3} follows from \eqref{eq:gstinkert}, where we substitute $\mathcal{L}v=V_F\phi_v$ to get
\begin{equation*}
{\mbox{Im}}\langle v,\widetilde{A}^*v\rangle+{\mbox{Im}}\langle v,V_F\phi_v\rangle\geq\frac{1}{4}\|\mathcal{U}V^{-1/2}_FV_F\phi_v+2iV_K^{1/2}v\|^2=\frac{1}{4}\|V_K^{1/2}(\phi_v+2iv)\|^2\:,
\end{equation*}
which is the desired result.
\end{proof}
\begin{example} \normalfont \label{ex:potsdam}
Let $\mathcal{H}=L^2(0,\infty)$, assume that the real potential $W\in L^2(0,\infty)$ and consider the dual pair of closed operators $(A,\widetilde{A})$ given by
\begin{align*}
A:\qquad\mathcal{D}(A)=H^2_0(0,\infty),\quad (Af)(x)&= -if''(x)+W(x)f(x)\\
\widetilde{A}:\qquad\mathcal{D}(\widetilde{A})=H^2_0(0,\infty),\quad (\widetilde{A}f)(x)&=if''(x)+W(x)f(x)\:,
\end{align*}
which has the common core property since $\mathcal{D}(A)=\mathcal{D}(\widetilde{A})$ and $(A,\widetilde{A})$ are closed.
Their adjoints are given by
\begin{align*}
\widetilde{A}^*:\qquad\mathcal{D}(\widetilde{A}^*)=H^2(0,\infty),\quad (\widetilde{A}^*f)(x)&= -if''(x)+W(x)f(x)\\
A^*:\qquad\mathcal{D}(A^*)=H^2(0,\infty),\quad (A^*f)(x)&= if''(x)+W(x)f(x)\:.
\end{align*}
Moreover, the ``imaginary part" $V$ and its adjoint $V^*$ are given by
\begin{align*}
V:\qquad\mathcal{D}(V)&=H^2_0(0,\infty),\quad f\mapsto -f''\\
V^*:\qquad\mathcal{D}(V^*)&=H^2(0,\infty),\quad f\mapsto -f''\:.
\end{align*}
Since $$\ker(V^*\pm i)={\mbox{span}}\left\{\exp\left(-\frac{1\pm i}{\sqrt{2}}x\right)\right\}\:,$$ and $$\mathcal{D}(\widetilde{A}^*)=\mathcal{D}(V^*)=\mathcal{D}(V)\dot{+}\ker(V^*+i)\dot{+}\ker(V^*-i)=\mathcal{D}(A)\dot{+}\ker(V^*+i)\dot{+}\ker(V^*-i)\:,$$ we may choose
\begin{align*}
\mathcal{D}(V^*)//\mathcal{D}(V)&=\mathcal{D}(\widetilde{A}^*)//\mathcal{D}(A)\\&={\mbox{span}}\left\{\exp\left(-\frac{1+ i}{\sqrt{2}}x\right),\exp\left(-\frac{1- i}{\sqrt{2}}x\right)\right\}={\mbox{span}}\{\sigma,\tau\}\:.
\end{align*}
The functions $\sigma$ and $\tau$ are suitable linear combinations of the elements of $\mathcal{D}(\widetilde{A}^*)//\mathcal{D}(A)$ such that
$\sigma(0)=\tau'(0)=1$ and $\sigma'(0)=\tau(0)=0$. For $\rho\in{\mathbb C}$, define the function $\zeta_\rho(x):=\sigma(x)+\rho\tau(x)$ and let $\zeta_\infty(x):=\tau(x)$. In order to be able to use Corollary \ref{coro:wolfratshausen}, we will only consider $\mathcal{L}\zeta_\rho\in{\mbox{ran}}(V_F)$, i.e. we can write $\mathcal{L}\zeta_\rho=V_F\phi$ for some $\phi\in\mathcal{D}(V_F)=\{f\in H^2(0,\infty), f(0)=0\}$. Let us therefore use the parameter $\rho\in{\mathbb C}\cup\{\infty\}$ and the function $\phi\in\mathcal{D}(V_F)$ to describe all extensions $A_{\rho,\phi}$ of the form
\begin{align*}
A_{\rho,\phi}:\qquad\mathcal{D}(A_{\rho,\phi})&=\mathcal{D}(A)\dot{+}{\mbox{span}}\{\zeta_\rho\}\\
\left(A_{\rho,\phi}(f+\lambda\zeta_\rho)\right)(x)&= -i(f''(x)+\lambda\zeta_\rho''(x))+W(x)(f(x)+\lambda\zeta_\rho(x))-\lambda\phi''(x)\:,
\end{align*}
where $f\in\mathcal{D}(A)$ and $\lambda\in{\mathbb C}$. Next, let us use Corollary \ref{coro:wolfratshausen} to find the conditions on $\rho$ and $\phi$ for $A_{\rho,\phi}$ to be dissipative. Firstly, observe that $V_K$ is the Neumann-Laplacian on the half-line. This can be seen from
\begin{equation} \label{eq:king}
\langle f,V^*f\rangle=\overline{f(0)}f'(0)+\int_0^\infty|f'(x)|^2\text{d}x
\end{equation}
for all $f\in\mathcal{D}(V^*)$. In order to find the selfadjoint restrictions of $V^*$, observe that any additional selfadjoint boundary condition has to be of the form $f'(0)=rf(0)$, where $r\in{\mathbb R}$. The additional choice $r=\infty$ corresponds to a Dirichlet condition at $0$, i.e. $f(0)=0$ and describes the Friedrichs extension of $V$. For any $r<0$, we get that $\langle f,V^*f\rangle$ can be made negative, which therefore does not describe a non-negative selfadjoint extension of $V$. For $r\geq 0$, it is obvious that $r=0$ describes the smallest non-negative extension of $V$. Hence, the Kre\u\i n--von Neumann extension is given by the Neumann-Laplacian with domain
$\mathcal{D}(V_K)=\{f\in H^2(0,\infty), f'(0)=0\}$. It is also not hard to see that if we close $\mathcal{D}(V_K)$ with respect to the norm induced by \eqref{eq:king}, we get $\mathcal{D}(V_K^{1/2})=H^1(0,\infty)$. Now, since $\mathcal{D}(\widetilde{A}^*)//\mathcal{D}(A)\subset H^1(0,\infty)=\mathcal{D}(V_K^{1/2})$, we get that the first necessary condition from Corollary \ref{coro:wolfratshausen}, which requires that ${\mbox{span}}\{\zeta_\rho\}\subset\mathcal{D}(V_K^{1/2})$ in order for $A_{\rho,\phi}$ to be dissipative is satisfied for any $\rho\in{\mathbb C}\cup\{\infty\}$. Next, let us determine for which $\rho\in{\mathbb C}\cup\{\infty\}$ and $\phi\in\mathcal{D}(V_F)$ Condition \eqref{eq:scan3} is satisfied. For $\rho\in{\mathbb C}$, it reads as
\begin{align*}
&{\mbox{Im}}\langle \zeta_\rho,\widetilde{A}^*\zeta_\rho\rangle+{\mbox{Im}}\langle \zeta_\rho,V_F\phi\rangle\geq\frac{1}{4}\|V_K^{1/2}(\phi+2i\zeta_\rho)\|^2\\
\Leftrightarrow\:\:&{\mbox{Im}}\langle\zeta_\rho,-i\zeta_\rho''\rangle+{\mbox{Im}}\langle \zeta_\rho,-\phi''\rangle\geq\frac{1}{4}\|\phi'+2i\zeta_\rho'\|^2=\frac{1}{4}\|\phi'\|^2+\|\zeta_\rho'\|^2+{\mbox{Re}}\langle \phi',i\zeta_\rho'\rangle\\
\Leftrightarrow\:\: & {\mbox{Im}}(\overline{\zeta_\rho(0)}i\zeta_\rho'(0))+\|\zeta_\rho'\|^2+{\mbox{Im}}\langle\phi'',\zeta_\rho\rangle\geq\frac{1}{4}\|\phi'\|^2+\|\zeta_\rho'\|^2+{\mbox{Im}}\langle\phi'',\zeta_\rho\rangle+{\mbox{Im}}(\overline{\phi'(0)}\zeta_\rho(0))\\
\Leftrightarrow\:\: & {\mbox{Re}} \rho\geq \frac{1}{4}\|\phi'\|^2-{\mbox{Im}}(\phi'(0))\:.
\end{align*}
For $\rho=\infty$, we get the condition that
\begin{equation*}
0\geq\frac{1}{4}\|\phi'\|^2\:,
\end{equation*}
which means that the only allowed choice is $\phi(x)\equiv 0$ in this case.
\end{example}
\subsection{The strictly positive case} \label{subsec:strictly} Next, let us consider the case when the imaginary part $V$ is strictly positive, i.e.\ when there exists a positive number $\varepsilon>0$ such that
$\langle f,Vf\rangle\geq \varepsilon\|f\|^2$ for all $f\in\mathcal{D}(V)$. We introduce the notation $V\geq\varepsilon>0$ in this case.
\begin{corollary} Let $(A,\widetilde{A})$ be a dual pair satisfying the common core property, where $A$ is dissipative. Moreover, let the imaginary part $V$ be strictly positive. Then, $A_{\mathcal{V,L}}$ is dissipative if and only if $\mathcal{V}\subset\mathcal{D}(V_K^{1/2})$ and for all $v\in\mathcal{V}$ we have that
\begin{equation} \label{eq:scan}
{\mbox{Im}}\langle v,\widetilde{A}^*v\rangle+{\mbox{Im}}\langle \mathcal{P}v,\mathcal{L}v\rangle\geq\frac{1}{4}\|V_F^{-1/2}\mathcal{L}v\|^2+\|V_K^{1/2}v\|^2\:.
\end{equation}
Here, $\mathcal{P}$ denotes the unbounded projection onto $\ker V^*$ along $\mathcal{D}(V_F^{1/2})$, according to the decomposition $\mathcal{D}(V_K^{1/2})=\mathcal{D}(V_F^{1/2})\dot{+}\ker V^*$.
\end{corollary}
\begin{proof} Since $V\geq\varepsilon>0$, we have that ${\mbox{ran}}(V_F)={\mbox{ran}}(V_F^{1/2})=\mathcal{H}$, which means that the condition ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(V_F^{1/2})$ is always satisfied. Thus, by Theorem \ref{thm:kuckuckswalzer}, it is necessary that $\mathcal{V}\subset\mathcal{D}(V_K^{1/2})$ for $A_{\mathcal{V,L}}$ to be dissipative.\\
Since $V\geq\varepsilon>0$, we have that $\mathcal{D}(V_K)=\mathcal{D}(\overline{V})\dot{+}\ker V^*$ with $V_K=V^*\upharpoonright_{\mathcal{D}(V_K)}$. This implies that $\ker V^*=\ker V_K$ and since $V_K$ is non-negative, we also get that $\ker V_K^{1/2}=\ker V^*$. It is known that $\mathcal{D}(V_K^{1/2})=\mathcal{D}(V_F^{1/2})\dot{+}\ker V^*$ (cf. \cite{Alonso-Simon}). Thus, we can rewrite
\begin{align*}
&\frac{1}{4}\|\mathcal{U}V_F^{-1/2}\mathcal{L}v+2iV_K^{1/2}v\|^2=\frac{1}{4}\|\mathcal{U}V_F^{-1/2}\mathcal{L}v+2iV_K^{1/2}({\mathbbm{1}}-\mathcal{P})v\|^2\\\overset{\eqref{eq:kuckuck}}{=}&\frac{1}{4}\|\mathcal{U}(V_F^{-1/2}\mathcal{L}v+2iV_F^{1/2}({\mathbbm{1}}-\mathcal{P}))v\|^2=\frac{1}{4}\|V_F^{-1/2}\mathcal{L}v+2iV_F^{1/2}({\mathbbm{1}}-\mathcal{P})v\|^2\\
=&\frac{1}{4}\|V_F^{-1/2}\mathcal{L}v\|^2+\|V_F^{1/2}({\mathbbm{1}}-\mathcal{P})v\|^2+{\mbox{Im}}\langle V_F^{1/2}({\mathbbm{1}}-\mathcal{P})v,V_F^{-1/2}\mathcal{L}v\rangle\\
=&\frac{1}{4}\|V_F^{-1/2}\mathcal{L}v\|^2+\|V_K^{1/2}v\|^2+{\mbox{Im}}\langle({\mathbbm{1}}-\mathcal{P})v,\mathcal{L}v\rangle\:.
\end{align*}
With this, Condition \eqref{eq:gstinkert} from Theorem \ref{thm:alter} can be rewritten as
\begin{equation*}
{\mbox{Im}}\langle v,\widetilde{A}^*v\rangle+{\mbox{Im}}\langle \mathcal{P}v,\mathcal{L}v\rangle\geq \frac{1}{4}\|V_F^{-1/2}\mathcal{L}v\|^2+\|V_K^{1/2}v\|^2\:,
\end{equation*}
which is the desired result.
\end{proof}
\begin{remark} \normalfont From $V\geq\varepsilon$ it follows that both $V_F$ and $V_F^{1/2}$ are boundedly invertible and thus ${\mbox{ran}}(V_F)={\mbox{ran}}(V_F^{1/2})=\mathcal{H}$. Hence, the strictly positive case is a special case of Section \ref{subsec:additionalreq}, since ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(V_F)=\mathcal{H}$ is always satisfied. As in Corollary \ref{coro:wolfratshausen}, it is thus helpful to write $\mathcal{L}v=V_F\phi_v$ for any $v\in\mathcal{V}$, where $\phi_v$ is uniquely determined by $\mathcal{L}v$ since $V_F\geq\varepsilon$. Then, we can rewrite \eqref{eq:scan} as follows
\begin{equation} \label{eq:scanbesser}
{\mbox{Im}}\langle v,\widetilde{A}^*v\rangle+{\mbox{Im}}\langle \mathcal{P}v,V_F\phi_v\rangle\geq \frac{1}{4}\|V_F^{1/2}\phi_v\|^2+\|V_K^{1/2}v\|^2\:,
\end{equation}
which is more accessible to explicit computations.
\end{remark}
\begin{example} \label{ex:1493} \normalfont Let $\mathcal{H}=L^2(0,1)$, assume that $\gamma\geq\sqrt{3}$ and consider the dual pair $(A_0,\widetilde{A}_0)$, given by
\begin{align*}
A_0:\qquad\mathcal{D}(A_0)=\mathcal{C}_c^\infty(0,1),\quad\left(A_0f\right)(x)=-if''(x)-\gamma\frac{f(x)}{x^2},\\
\widetilde{A}_0:\qquad\mathcal{D}(\widetilde{A}_0)=\mathcal{C}_c^\infty(0,1),\quad\left(\widetilde{A}_0f\right)(x)=if''(x)-\gamma\frac{f(x)}{x^2}\:.
\end{align*}
Define the dual pair $(A,\widetilde{A})$, where $A:=\overline{A_0}$ and $\widetilde{A}:=\overline{\widetilde{A}_0}$. By construction, $(A,\widetilde{A})$ has the common core property, where we choose $\mathcal{C}_c^\infty(0,1)=:\mathcal{D}$ to be the common core. The ``imaginary part" $V$ is given by
\begin{align*}
V:\qquad\mathcal{D}(V)&=\mathcal{C}_c^\infty(0,1)\\
f&\mapsto-f''\:,
\end{align*}
which is a strictly positive operator, since its closure is a restriction of the Dirichlet-Laplacian on the unit interval: $$\langle f,Vf\rangle\geq \pi^2\|f\|^2\:\:\text{for all}\:\: f\in\mathcal{C}_c^\infty(0,1)\:.$$ Moreover, its adjoint $V^*$ is given by
\begin{equation*}
V^*:\qquad\mathcal{D}(V^*)=H^2(0,1)\:,\quad f\mapsto-f''
\end{equation*}
and its kernel is $\ker(V^*)={\mbox{span}}\{1,x\}$. Thus, observe that for any $f\in H^2(0,1)$, the projection $\mathcal{P}$ onto $\ker(V^*)$ along $\mathcal{D}(V_F^{1/2})$ is given by
\begin{equation} \label{eq:unbddproj}
(\mathcal{P}f)(x)=(1-x)f(0)+xf(1)\:.
\end{equation}
The choice $\gamma\geq\sqrt{3}$ ensures that $\dim \ker \widetilde{A}^*=\dim \ker A^*=1$, which keeps the extension problem simpler. It can be shown by straightforward calculation that $\mathcal{D}(\widetilde{A}^*)$ can be written as
\begin{equation*}
\mathcal{D}(\widetilde{A}^*)=\mathcal{D}(A)\dot{+}{\mbox{span}}\{x^{\omega},x^{\overline{\omega}+2}\}\:,
\end{equation*}
where we have defined $\omega:=(1+\sqrt{1+4i\gamma})/2$. We therefore choose $\mathcal{D}(\widetilde{A}^*)//\mathcal{D}(A)={\mbox{span}}\{x^{\omega},x^{\overline{\omega}+2}\}$. Let us now parametrize all proper ``one-dimensional" extensions of $(A,\widetilde{A})$, with the family of operators $\{A_\rho\}_{\rho\in{\mathbb C}\cup\{\infty\}}$ given by
\begin{equation*}
A_\rho:\qquad\mathcal{D}(A_\rho)=\mathcal{D}(A)\dot{+}{\mbox{span}}\{\xi_\rho\},\quad A_\rho=\widetilde{A}^*\upharpoonright_{\mathcal{D}(A_\rho)}\:,
\end{equation*}
where $${\mbox{span}}\{x^{\omega},x^{\overline{\omega}+2}\}\ni\xi_\rho(x):=\begin{cases}\rho\left(\frac{(2+\overline{\omega})x^{\omega}-\omega x^{\overline{\omega}+2}}{2+\overline{\omega}-\omega}\right)-\frac{x^{\omega}-x^{\overline{\omega}+2}}{2+\overline{\omega}-\omega}\quad&\text{for}\:\:\rho\in{\mathbb C}\\\frac{(2+\overline{\omega})x^{\omega}-\omega x^{\overline{\omega}+2}}{2+\overline{\omega}-\omega}&\text{for}\:\:\rho=\infty\end{cases}$$ satisfies the boundary conditions
\begin{align*}
\xi_\rho(0)=\xi_\rho'(0)=0\:\:\:&\text{for}\:\:\:\rho\in{\mathbb C}\cup\{\infty\}\\
\xi_\rho(1)=\rho,\:\: \xi_\rho'(1)=1\:\:\:&\text{for}\:\:\:\rho\in{\mathbb C} \:\:\:\text{and}\:\:\:\xi_\rho(1)=1,\:\: \xi'_\rho(1)=0\:\:\:\text{for}\:\:\:\rho=\infty\:.
\end{align*}
Next, \eqref{eq:unbddproj} implies that for $\rho\in{\mathbb C}$, we get $\mathcal{P}\xi_\rho (x)=\rho x$, whereas for $\rho=\infty$, we get $\mathcal{P}\xi_\infty(x)=x$. This follows from the fact that $\mathcal{D}(V_F^{1/2})=H^1_0(0,1)$ and for any $\rho\in{\mathbb C}$, we have $\xi_\rho(0)=\xi_\infty(0)=0$ as well as $\xi_\rho(1)=\rho$ and $\xi_\infty(1)=1$. Now, since $V$ is strictly positive, we know that its Friedrichs extension $V_F$ is bijective, which means that any function $\mathcal{L}\xi_\rho\in L^2(0,1)$ can be written as $\mathcal{L}\xi_\rho=V_F\phi=-\phi''$ for some unique $\phi\in\mathcal{D}(V_F)=\{\phi\in H^2(0,1), \phi(0)=\phi(1)=0\}$.
Hence, let us use the parameter $\rho\in{\mathbb C}\cup\{\infty\}$ and the arbitrary function $\phi\in\mathcal{D}(V_F)$ to label all one-dimensional extensions of $\mathcal{D}(A)$ that have domain contained in $\mathcal{D}(\widetilde{A}^*)$. They are given by
\begin{align*}
A_{\rho,\phi}:\qquad\qquad\mathcal{D}(A_{\rho,\phi})&=\mathcal{D}(A)\dot{+}{\mbox{span}}\{\xi_\rho\}\\
[A_{\rho,\phi}(f+\lambda\xi_\rho)](x)&=(-if''(x)-\lambda i\xi_\rho''(x))-\gamma\frac{f(x)+\lambda\xi_\rho(x)}{x^2}-\lambda \phi''(x)\:,
\end{align*}
where $f\in\mathcal{D}(A)$ and $\lambda\in{\mathbb C}$. By \eqref{eq:scanbesser}, we have that $A_{\rho,\phi}$ is dissipative if and only if
\begin{align*}
{\mbox{Im}}\langle \xi_\rho,\widetilde{A}^*\xi_\rho\rangle-\|V_K^{1/2}\xi_\rho\|^2\geq\frac{1}{4}\|V_F^{1/2}\phi\|^2-{\mbox{Im}}\langle \mathcal{P}\xi_\rho,V_F\phi\rangle\:.
\end{align*}
is satisfied. Using that for any $v\in \mathcal{D}(V_K^{1/2})=H^1(0,1)$, we have
\begin{equation*}
\|V_K^{1/2}v\|^2=\|v'\|^2-|v(1)-v(0)|^2\:,
\end{equation*}
it can be shown that for any $v\in{\mbox{span}}\{x^{\omega},x^{\overline{\omega}+2}\}$, we have $${\mbox{Im}}\langle v,\widetilde{A}^*v\rangle-\|V_K^{1/2}v\|^2=-{\mbox{Re}}\left(\overline{v(1)}v'(1)\right)+|v(1)|^2\:,$$
which means that
\begin{equation*}
{\mbox{Im}}\langle\xi_\rho,\widetilde{A}^*\xi_\rho\rangle-\|V_K^{1/2}\xi_\rho\|^2=\begin{cases} |\rho|^2-{\mbox{Re}}(\rho)\quad&\text{if}\quad \rho\in{\mathbb C}\\ 1\quad&\text{if}\quad \rho=\infty\:.\end{cases}
\end{equation*}
Moreover, since $\|V_F^{1/2}\phi\|=\|\phi'\|$ and
\begin{equation}
{\mbox{Im}}\left(\int_0^1 x\phi''(x)\text{d}x\right)={\mbox{Im}}(\phi'(1))
\end{equation}
for any $\phi\in\mathcal{D}(V_F)$, the above yields the conditions on $\rho$ and $\phi$ for $A_{\rho,\phi}$ to be dissipative:
\begin{align*}
&\frac{1}{4}\|\phi'\|^2+{\mbox{Im}}(\overline{\rho}\phi'(1))\leq |\rho|^2-{\mbox{Re}}\rho\quad\text{for}\quad\rho\in{\mathbb C}\\
&\frac{1}{4}\|\phi'\|^2+{\mbox{Im}}(\phi'(1))\leq 1\qquad\qquad\quad\:\:\text{for}\quad\rho=\infty\:.
\end{align*}
For the case of proper extensions, where $\phi=0$, i.e.\ for $A_{\rho,0}$ we therefore have the condition that either $\rho=\infty$ or $|\rho|^2-{\mbox{Re}}\rho\geq 0$ for $A_{\rho,0}$ to be dissipative. In the non-proper case, for a suitable choice of $\phi$, it is no longer necessary that $\rho$ satisfies this condition. For instance, let $\phi(x):=(x^2-x)\in\mathcal{D}(V_F)$. We then get the condition
\begin{equation*}
\frac{1}{4}\|\phi'\|^2+{\mbox{Im}}(\overline{\rho}\phi'(1))=\frac{1}{12}-{\mbox{Im}}(\rho)\leq |\rho|^2-{\mbox{Re}}\rho
\end{equation*}
for $A_{\rho,(x^2-x)}$ to be dissipative. This condition is for example satisfied by $\rho=\frac{1}{2}+\frac{3}{8}i$, i.e. $A_{\left(\frac{1}{2}+\frac{3}{8}i\right),(x^2-x)}$ is dissipative, while $A_{\left(\frac{1}{2}+\frac{3}{8}i\right),0}$ is not. In Corollary \ref{coro:ausmachen}, we will show that the phenomenon that we have a dissipative non-proper extension, defined on a domain on which the corresponding proper extension would not be dissipative, can only occur if the Friedrichs and Kre{\u\i}n-von Neumann extensions of $V$ do not coincide, as it is the case in this example.
\label{ex:shirley}
\end{example}
\begin{remark} \normalfont The choice of the highly singular $x^{-2}$-potential allowed us to compute everything explicitly. It is however not very difficult to add a ``small" extra potential.
\end{remark}
\subsection{The case of coinciding Friedrichs and Kre{\u\i}n-von Neumann extension} \label{subsec:essentials}
Let us now consider the case that the Friedrichs and Kre{\u\i}n-von Neumann extensions of $V$ coincide: $V_F=V_K=:\widehat{V}$. Before we simplify Theorem \ref{thm:alter} with the help of this extra assumption, let us prove that in this case, both conditions, $\mathcal{V}\subset\mathcal{D}(\widehat{V}^{1/2})$ and ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}( \widehat{V}^{1/2})$ are \emph{independently} necessary for $A_{\mathcal{V,L}}$ to be dissipative.
\begin{lemma} \label{lemma:singulariter} Let $(A,\widetilde{A})$ be a dual pair satisfying the common core condition, where $A$ is dissipative and assume in addition that for the imaginary part $V$ we have $V_F=V_K=:\widehat{V}$. Then, for $A_{\mathcal{V,L}}$ to be dissipative it is necessary that $\mathcal{V}\subset\mathcal{D}(\widehat{V}^{1/2})$ and ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(\widehat{V}^{1/2})$.
\end{lemma}
\begin{proof} We only need to show that $\mathcal{V}\subset\mathcal{D}(\widehat{V}^{1/2})$ is necessary for $A_{\mathcal{V,L}}$ to be dissipative. The condition ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(\widehat{V}^{1/2})$ will then just follow from Theorem \ref{thm:kuckuckswalzer}, ii).
Thus, assume that there exists a $v\in\mathcal{V}$ such that $v\notin\mathcal{D}(\widehat{V}^{1/2})$.
Since for any $f\in\mathcal{D}(A)$, $v\in\mathcal{V}$ we have
\begin{equation*}
{\mbox{Im}}\langle f+v,A_\mathcal{V,L}(f+v)\rangle=\langle f,Vf\rangle+{\mbox{Im}}\langle v,2iVf\rangle+{\mbox{Im}}\langle v,(\widetilde{A}^*+\mathcal{L})v\rangle+{\mbox{Im}}\langle f,\mathcal{L}v\rangle\:,
\end{equation*}
showing that
\begin{equation} \label{eq:irish} \inf_{f\in\mathcal{D}({V})}(\langle f,{V}f\rangle+{\mbox{Im}}(\langle v,2i{V}f\rangle-\langle \mathcal{L}v,f\rangle))=-\infty\:,
\end{equation}
will imply that $A_{\mathcal{V,L}}$ cannot be dissipative. We will proceed to show that
\begin{equation} \label{eq:irish2} \inf_{f\in\mathcal{D}(\overline{V})}(\langle f,\overline{V}f\rangle+{\mbox{Im}}(\langle v,2i\overline{V}f\rangle-\langle \mathcal{L}v,f\rangle))=-\infty
\end{equation}
and using that $\mathcal{D}(V)$ is a core for $\overline{V}$, this implies that for each $\widetilde{f}_n\in\mathcal{D}(\overline{V})$ we can choose a sequence $\{\widetilde{f}_{n,m}\}_{m=1}^\infty\subset\mathcal{D}(V)$ such that $\widetilde{f}_{n,m}\overset{m\rightarrow\infty}{\longrightarrow}\widetilde{f}_n$ and $V\widetilde{f}_{n,m}\overset{m\rightarrow\infty}{\longrightarrow}\overline{V}\widetilde{f}_n$. A diagonal sequence argument then shows \eqref{eq:irish} and thus the lemma.
Let us thus now show \ref{eq:irish2}. To this end, let $P$ denote the projection-valued spectral measure corresponding to $\widehat{V}$ and define $P_1:=P([0,1))$ and $P_2:=P([1,\infty))$ as well as $\mathcal{H}_{1,2}:=P_{1,2}\mathcal{H}$. Since $\widehat{V}\geq 0$, we have $P_1+P_2={\mathbbm{1}}$, resp. $\mathcal{H}_1\oplus\mathcal{H}_2=\mathcal{H}$.
Now, observe that $v\notin\mathcal{D}(\widehat{V}^{1/2})$ if and only if $P_2v\notin\mathcal{D}(\widehat{V}^{1/2})$.
Since $V_F^{1/2}=V_K^{1/2}=\widehat{V}^{1/2}$, we get that ${\mbox{ran}} (V_F^{1/2}\upharpoonright_{\mathcal{D}(V)})={\mbox{ran}}(\widehat{V}^{1/2}\upharpoonright_{\mathcal{D}(V)})$. Hence, if $P_2v\notin\mathcal{D}(\widehat{V}^{1/2})=\mathcal{D}(V_K^{1/2})$, we have by Corollary \ref{coro:andonishio} that there exists a sequence $\{f_n\}\subset\mathcal{D}(\overline{V})$ such that $\|\widehat{V}^{1/2}f_n\|=1$ for all $n\in{\mathbb N}$ and
\begin{equation} \label{eq:kuzel}
\lim_{n\rightarrow\infty}|\langle P_2v,\widehat{V}^{1/2}\widehat{V}^{1/2}f_n\rangle|=+\infty\:.
\end{equation}
We now claim that the sequence $\{\widehat{V}^{1/2}P_2{f}_n\}$ satisfies
$$ \|\widehat{V}^{1/2}P_2{f}_n\|\leq 1\quad\text{and}\quad\lim_{n\rightarrow\infty}|\langle v,\widehat{V}^{1/2}\widehat{V}^{1/2}P_2{f}_n\rangle|=+\infty\:.$$
The first statement follows immediately from
\begin{equation} \label{eq:estimate2}
\|\widehat{V}^{1/2}P_2{f}_n\|=\|P_2\widehat{V}^{1/2}{f}_n\|\leq\|\widehat{V}^{1/2}f_n\|=1\:,
\end{equation}
while the second statement follows from \eqref{eq:kuzel}.
Next, observe that
\begin{equation} \label{eq:biscuits}
\|P_2{f}_n\|^2=\int_{[1,\infty)}\text{d}\|P(\lambda){f}_n\|^2\leq\int_{[1,\infty)}\lambda\text{d}\|P(\lambda){f}_n\|^2=\|\widehat{V}^{1/2}P_2{f}_n\|^2\overset{\eqref{eq:estimate2}}{\leq}1\:.
\end{equation}
For any $n\in{\mathbb N}$ choose $\phi_n\in[0,2\pi)$ such that $${\mbox{Im}}\langle v,\widehat{V}^{1/2}\widehat{V}^{1/2}e^{i\phi_n}P_2{f}_n\rangle=-|\langle v,\widehat{V}^{1/2}\widehat{V}^{1/2}P_2{f}_n\rangle|\:.$$ Choosing $g_n=e^{i\varphi_n}P_2f_n$ for any $n\in{\mathbb N}$ now yields \eqref{eq:irish2} since
\begin{align*}
&\langle g_n,Vg_n\rangle+{\mbox{Im}}(\langle v,2iVg_n\rangle-\langle\mathcal{L}v,g_n\rangle)=\|\widehat{V}^{1/2}P_2f_n\|^2-2|\langle v, \widehat{V}^{1/2}\widehat{V}^{1/2}P_2f_n\rangle|-{\mbox{Im}}\langle \mathcal{L}v,e^{i\varphi_n}P_2f_n\rangle\\
\overset{\eqref{eq:estimate2}}{\leq} & 1-2|\langle v, \widehat{V}^{1/2}\widehat{V}^{1/2}P_2f_n\rangle|+\|\mathcal{L}v\|\|P_2f_n\|\overset{\eqref{eq:biscuits}}{\leq} 1-2|\langle v, \widehat{V}^{1/2}\widehat{V}^{1/2}P_2f_n\rangle|+\|\mathcal{L}v\|\overset{n\rightarrow\infty}{\longrightarrow}-\infty\:.
\end{align*}
This finishes the proof.
\end{proof}
\begin{remark} \normalfont This result applies in particular to the case of $V$ being essentially selfadjoint, where we have $\overline{V}=V_F=V_K$. However, note that the previous lemma and the following corollary cover a wider class of imaginary parts $V$ than just the essentially selfadjoint ones. For example, let $\mathcal{H}=L^2({\mathbb R}^+)$ and consider the imaginary part $V$ given by
\begin{align*}
V:\mathcal{D}(V)=\mathcal{C}_c^\infty({\mathbb R}^+),\qquad\left(Vf\right)(x)=-f''(x)-\frac{1/4}{x^2}f(x)\:.
\end{align*}
It is a well-known fact that $V$ is not essentially selfadjoint but that its Friedrichs and Kre{\u\i}n-von Neumann extension coincide (cf.\ e.g.\ \cite[Prop.\ 4.21]{BDG11}). For an abstract criterion as to whether a non-negative symmetric and non-essentially selfadjoint operator has a unique non-negative selfadjoint extension, we refer to Kre{\u\i}n's result in \cite{Krein} and its presentation in \cite[Thm.\ 2.12]{Alonso-Simon}. \label{rem:referee}
\end{remark}
Let us now simplify Theorem \ref{thm:alter} for the case that the Friedrichs and Kre{\u\i}n-von Neumann extension $V_F$ and $V_K$ of $V$ coincide: $V_F=V_K=:\widehat{V}$. Note that the situation of $V$ being essentially selfadjoint is a special case of this. We also want to show that $A_{\mathcal{V,L}}$ can only be dissipative if $A_{\mathcal{V}}$ already is, i.e. there necessarily needs to be a dissipative boundary condition -- described by a suitable choice of $\mathcal{V}$ -- before one can consider deviations from the action of $\widetilde{A}^*$ via non-zero operators $\mathcal{L}$. This is fundamentally different to the case $V_F\neq V_K$, where we have found an example of an extension $A_{\mathcal{V,L}}$, which was dissipative while $A_\mathcal{V}$ was not (Example \ref{ex:1493}).
\begin{corollary} \label{coro:ausmachen} Let $(A,\widetilde{A})$ be dual pair satisfying the common core property, where $A$ is dissipative. Moreover, let the imaginary part $V$ be such that its Friedrichs and Kre{\u\i}n-von Neumann extension coincide, i.e.\ $V_F=V_K=:\widehat{V}$. Then, $A_{\mathcal{V,L}}$ is dissipative if and only if $\mathcal{V}\subset\mathcal{D}(\widehat{V}^{1/2})$, ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(\widehat{V}^{1/2})$ and for all $v\in\mathcal{V}$ we have that
\begin{equation} \label{eq:scan2}
{\mbox{Im}}\langle v,\widetilde{A}^*v\rangle\geq\frac{1}{4}\|\widehat{V}^{\,-1/2}\mathcal{L}v\|^2+\|\widehat{V}^{1/2}v\|^2\:.
\end{equation}
In particular, this implies that for $A_{\mathcal{V,L}}$ to be dissipative, it is necessary that $A_\mathcal{V}$ is dissipative.
\end{corollary}
\begin{proof} The conditions that $\mathcal{V}\subset\mathcal{D}(\widehat{V}^{1/2})$ and ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(\widehat{V}^{1/2})$ for $A_{\mathcal{V,L}}$ to be dissipative follow from Lemma \ref{lemma:singulariter}. Condition \eqref{eq:scan2} follows from \eqref{eq:gstinkert} using that $V_K=V_F=\widehat{V}$, which implies that $\mathcal{U}$ acts like the identity on $\overline{{\mbox{ran}}(\widehat{V}^{1/2})}$. Moreover, for $\mathcal{L}=0$, we get that $A_\mathcal{V}$ is dissipative if and only if $\mathcal{V}\subset\mathcal{D}(\widehat{V}^{1/2})$ and ${\mbox{Im}}\langle v,\widetilde{A}^*v\rangle\geq\|\widehat{V}^{1/2}v\|^2$ for all $v\in\mathcal{V}$. Thus, if $A_\mathcal{V}$ is not dissipative then it is either true that $\mathcal{V}\not\subset\mathcal{D}(\widehat{V}^{1/2})$ or we have $\mathcal{V}\subset\mathcal{D}(\widehat{V}^{1/2})$ but there exists a $v\in\mathcal{V}$ such that
$$ {\mbox{Im}}\langle v,\widetilde{A}^*v\rangle-\|\widehat{V}^{1/2}v\|^2<0\:,$$
both implying that $A_{\mathcal{V,L}}$ cannot be dissipative either.
This shows the corollary.
\end{proof}
\begin{example} \label{ex:konzert} \normalfont Let $0<\gamma<1/2$ and consider the dual pair of operators
\begin{align*}
A_0:\quad\mathcal{D}(A_0)&=\mathcal{C}_c^\infty(0,1),\quad (A_0f)(x):=if'(x)+\frac{i\gamma}{x}f(x)\\
\widetilde{A}_0:\quad\mathcal{D}(\widetilde{A}_0)&=\mathcal{C}_c^\infty(0,1),\quad (\widetilde{A}_0f)(x):=if'(x)-\frac{i\gamma}{x}f(x)\:,
\end{align*}
where $A_0$ is dissipative and $\widetilde{A}_0$ is antidissipative. We denote their closures by $A=\overline{A_0}$ and $\widetilde{A}=\overline{\widetilde{A}_0}$. It can be shown that $$\mathcal{D}(\widetilde{A}^*)=\mathcal{D}(A)\dot{+}{\mbox{span}}\{x^{-\gamma},x^{\gamma+1}\}$$ and we therefore choose $\mathcal{D}(\widetilde{A}^*)//\mathcal{D}(A)={\mbox{span}}\{x^{-\gamma},x^{\gamma+1}\}$. Moreover, it is easy to see that the imaginary part $V$ is the essentially selfadjoint multiplication operator by the function $\frac{\gamma}{x}$ with domain $\mathcal{C}_c^\infty(0,1)$ which has closure to the selfadjoint maximal multiplication operator by $\frac{\gamma}{x}$, which we denote by $\overline{V}$. Since $x^{-\gamma}\notin\mathcal{D}(\overline{V}^{1/2})$, this means that the only choice for $\mathcal{V}\subset{\mbox{span}}\{x^{-\gamma},x^{\gamma+1}\}$ in order to have a chance for $A_{\mathcal{V,L}}$ to be dissipative is $\mathcal{V}:={\mbox{span}}\{x^{\gamma+1}\}$. Let us define $v(x)=:x^{\gamma+1}$ and $\mathcal{L}v=:\ell\in\mathcal{H}$ and let us use the functions $v$ and $\ell$ to label $A_{\mathcal{V,L}}=:A_{v,\ell}$. Since $\langle f,\overline{V}f\rangle\geq\gamma\|f\|^2$ for all $f\in\mathcal{D}(\overline{V})$, we get that $\overline{V}$ and $\overline{V}^{1/2}$ are both boundedly invertible, in particular that ${\mbox{ran}}(\overline{V}^{1/2})=\mathcal{H}$. Thus, by Corollary \ref{coro:ausmachen}, it only remains to check whether Condition \eqref{eq:scan2} is satisfied, which reads as
\begin{equation*}
{\mbox{Im}}\langle v,\widetilde{A}^*v\rangle-\|\overline{V}^{1/2}v\|^2\geq\frac{1}{4}\|\overline{V}^{-1/2}\ell\|^2\:.
\end{equation*}
It can be easily shown that
\begin{equation*}
{\mbox{Im}}\langle v,\widetilde{A}^*v\rangle-\|\overline{V}^{1/2}v\|^2{=}\frac{1}{2}\left(|v(1)|^2-|v(0)|^2\right)=\frac{1}{2}\:.
\end{equation*}
Hence, $A_{v,\ell}$ is dissipative if and only if
\begin{equation*}
\|\overline{V}^{\:-1/2}\ell\|^2=\frac{1}{\gamma}\int_0^1 x|\ell(x)|^2\text{d}x\leq 2\:.
\end{equation*}
This means that all dissipative extensions of $A$ that have domain contained in $\mathcal{D}(\widetilde{A}^*)$ are given by
\begin{align} \label{eq:referee}
A_{v,\ell}:\qquad\qquad\qquad\mathcal{D}(A_{v,\ell})&=\mathcal{D}(A)\dot{+}{\mbox{span}}\{v\}\notag\\
\left(A_{v,\ell}(f+\lambda v)\right)(x)&= if'(x)+i\lambda v'(x)+i\gamma\frac{f(x)+\lambda v(x)}{x}+\lambda \ell(x)\:,
\end{align}
where $f\in\mathcal{D}(A)$ and $\lambda\in{\mathbb C}$. The function $\ell\in L^2(0,1)$ has to satisfy
\begin{equation} \label{eq:erdkreis}
\int_0^1x|\ell(x)|^2\text{d}x\leq 2\gamma\:.
\end{equation}
Moreover, by Lemma \ref{prop:dampfnudel}, we have that $A_{v,\ell}$ is maximally dissipative since it is a one-dimensional extension of $A$.
\begin{remark} As in Example \ref{ex:potsdam}, one could easily add a sufficiently ``small" real potential $W$ to the operators $A_0$ and $\widetilde{A}_0$.
\end{remark}
\end{example}
\section{Operators with bounded imaginary part}
In this section, we will apply the result of Corollary \ref{coro:ausmachen} in order to construct all dissipative extensions of a dissipative operator with bounded imaginary part, where $\mathcal{D}(A)=\mathcal{D}(S)=\mathcal{D}(V)$. While this is not a new result (it can for example essentially be found in \cite[Theorem 1]{Crandall} with a different way of proof), we want to give more attention to the interplay between boundary conditions and bounded dissipative perturbations. In particular, we will show that if an operator with non-dissipative boundary condition is considered, it is impossible to add a bounded dissipative perturbation such that the result is a dissipative operator (Corollary \ref{coro:jsbach}, ii). On the other hand, we will show that the ``more dissipative" a boundary condition is, the more freedom one has in describing dissipative extensions of a given operator (Corollary \ref{coro:jsbach}, iii).
To start with, let us show that it is sufficient to only consider operators of the form $S+iV$, where $S$ is symmetric and $V\geq 0$ is bounded:
\begin{lemma} \label{lemma:uab} Let $A$ be a dissipative operator and assume that the quadratic form $q$ given by
\begin{equation*}
q:\quad\mathcal{D}(q)=\mathcal{D}(A),\quad f\mapsto{\mbox{Im}}\langle f,Af\rangle
\end{equation*}
is bounded. Then there exists a symmetric operator $S$ with $\mathcal{D}(S)=\mathcal{D}(A)$ and a essentially selfadjoint bounded operator $V\geq 0$ with $\mathcal{D}(V)=\mathcal{D}(A)$ such that $A=S+iV$.
\end{lemma}
\begin{proof}
Since $q$ is bounded, it is closable. Let $V'$ denote the bounded selfadjoint operator associated to it and let $V=V'\upharpoonright_{\mathcal{D}(A)}$. It is not hard to see that $S:=A-iV$ is symmetric and trivially $A=S+iV$.
\end{proof}
Next, let us show that for \emph{any} dissipative extension of $S+iV$, it is necessary that its domain is contained in $\mathcal{D}(S^*)$. Since $V$ is assumed to be bounded, note that $A=S+iV$ is closed if and only if $S$ is closed. Also note that we are describing general extensions of $A$ that need \emph{not} be of the form $\widehat{S}+iV$, where $\widehat{S}$ is a symmetric extension of $S$.
\begin{lemma} \label{lemma:allextensions}
Let $A:=S+iV$, where $S$ is closed and symmetric and $V\geq 0$ is bounded. Then, for an extension $A\subset B$ to be dissipative, it is necessary that $\mathcal{D}(B)\subset\mathcal{D}(S^*)$.
\end{lemma}
\begin{proof} Assume that $\mathcal{D}(B)\not\subset\mathcal{D}(S^*)$, i.e. that there exists a $v\in\mathcal{D}(B)$ such that $v\notin\mathcal{D}(S^*)$. For any $f\in\mathcal{D}(A)=\mathcal{D}(S)$, consider
\begin{align} \label{eq:animal}
{\mbox{Im}}\langle f+v,B(f+v)\rangle&={\mbox{Im}}\langle f,(S+iV)f\rangle+{\mbox{Im}}\langle v,(S+iV)f\rangle+{\mbox{Im}}\langle f+v,Bv\rangle\notag\\
&=\langle f,Vf\rangle+{\mbox{Im}}\langle v,Sf\rangle+{\mbox{Im}}\langle v,iVf\rangle+{\mbox{Im}}\langle f+v,Bv\rangle\notag\\
&\leq \|V\|\|f\|^2+{\mbox{Im}}\langle v,Sf\rangle+\|V\|\|v\|\|f\|+\|f\|\|Bv\|+\|v\|\|Bv\|\:.
\end{align}
Since $v\notin\mathcal{D}(S^*)$, there exists a normalized sequence $\{f_n\}\subset\mathcal{D}(S)$ such that $$\lim_{n\rightarrow\infty}{\mbox{Im}}\langle v,Sf_n\rangle=-\infty\:.$$ Using \eqref{eq:animal}, we therefore get
\begin{align*}
{\mbox{Im}}\langle f_n+v,B(f_n+v)\rangle\leq \|V\|+\|V\|\|v\|+\|Bv\|+\|v\|\|Bv\|+{\mbox{Im}}\langle v,Sf_n\rangle\overset{n\rightarrow\infty}{\longrightarrow}-\infty\:,
\end{align*}
which shows that $B$ cannot be dissipative in this case. This finishes the proof.
\end{proof}
We are now able to describe all dissipative extensions of $A=S+iV$:
\begin{theorem} \label{thm:bern} Let $A=S+iV$ be a dissipative operator with bounded imaginary part. Then $S_\mathcal{V,L}+iV$, where $S_{\mathcal{V,L}}$ is defined as in Definition \ref{def:quantz}, is a dissipative extension of $S+iV$ if and only if for all $v\in\mathcal{V}\subset\mathcal{D}(S^*)//\mathcal{D}(S)$ we have that $\mathcal{L}v\in{\mbox{ran}}(\overline{V}^{1/2})$ and the condition
\begin{equation} \label{eq:ragout}
{\mbox{Im}}\langle v,S^*v\rangle\geq\frac{1}{4}\|\overline{V}^{-1/2}\mathcal{L}v\|^2
\end{equation}
is satisfied.
As before, $\overline{V}^{-1/2}$ denotes the inverse of $\overline{V}^{1/2}$ on the reducing subspace $\overline{\text{\emph{ran}}(\overline{V}^{1/2})}$ as described in \eqref{eq:krakau}. Moreover, \emph{all} dissipative extensions of $S+iV$ are of this form.
\end{theorem}
\begin{remark} For the case that $V\equiv 0$, this means in particular that Condition \eqref{eq:ragout} simplifies to the condition that ${\mbox{Im}}\langle v,S^*v\rangle\geq 0$ for all $v\in\mathcal{V}$.
\end{remark}
\begin{proof} Since $V$ is bounded, $S_{\mathcal{V,L}}$ is an extension of $S$ if and only if $A_{\mathcal{V,L}}=S_{\mathcal{V,L}}+iV$ is an extension of $A:=S+iV$. Clearly, for $A:=S+iV$ and $\widetilde{A}:=S-iV$, we have that $(A,\widetilde{A})$ is a dual pair and we get that $\mathcal{D}(A)=\mathcal{D}(\widetilde{A})=\mathcal{D}(S)$, which means that it has the common core property. Moreover, by boundedness of $V$, we get that $\widetilde{A}^*=S^*+iV$, where $\mathcal{D}(\widetilde{A}^*)=\mathcal{D}(S^*)$. Also, observe that $V\upharpoonright_{\mathcal{D}(S)}$ is essentially selfadjoint, which means that we can apply Corollary \ref{coro:ausmachen}. Since $V$ is bounded, we have that $\mathcal{D}(\overline{V}^{1/2})=\mathcal{D}(\overline{V} )=\mathcal{H}$, which means that the Condition that $\mathcal{V}\subset\mathcal{D}(\overline{V}^{1/2})$ is always satisfied. Thus, by Corollary \ref{coro:ausmachen}, it is necessary that ${\mbox{ran}}(\mathcal{L})\subset{\mbox{ran}}(\overline{V}^{1/2})$ for $A_{\mathcal{V,L}}$ to be dissipative. Condition \eqref{eq:scan2} reads as
\begin{align*} {\mbox{Im}}\langle v,(S^*+iV)v\rangle\geq\|\overline{V}^{1/2}v\|^2+\frac{1}{4}\|\overline{V}^{-1/2}\mathcal{L}v\|^2
\Leftrightarrow\:{\mbox{Im}}\langle v,S^*v\rangle\geq\frac{1}{4}\|\overline{V}^{-1/2}\mathcal{L}v\|^2\:,
\end{align*}
which is the desired result. Let us finish by showing that \emph{all} dissipative extensions of $S+iV$ are parametrized by the operators $S_{\mathcal{V,L}}+iV$. By Lemma \ref{lemma:allextensions}, we know that all dissipative extensions have domain contained in $\mathcal{D}(S^*)=\mathcal{D}(\widetilde{A}^*)$. On the other hand, since $\mathcal{V}$ is an arbitrary subspace of $\mathcal{D}(S^*)//\mathcal{D}(S)$, the extensions $S_{\mathcal{V,L}}$ describe all possible extensions of $S$ that have domain contained in $\mathcal{D}(S^*)$. As they are dissipative if and only if $\mathcal{V}$ and $\mathcal{L}$ satisfy the assumptions of this theorem, we have found all dissipative extensions of $(S+iV)$.
\end{proof}
Let us now investigate the relation between the choice of $\mathcal{V}$ and $\mathcal{L}$:
\begin{corollary} \label{coro:jsbach} Let $\mathcal{V}\subset\mathcal{D}(S^*)//\mathcal{D}(S)$.\\
i) If $S_\mathcal{V}$ is symmetric, then $(S_\mathcal{V}+iV)$ is the only dissipative extension of $(S+iV)$ with domain equal to $\mathcal{D}(S_\mathcal{V})$. Moreover, the imaginary part of any other extension of the form $(S_{\mathcal{V,L}}+iV)$ is not bounded from below, i.e. for $\mathcal{L}\neq 0$, there exists no $\gamma\in{\mathbb R}^+$ such that
\begin{equation} \label{eq:semibounded} \inf_{\psi\in\mathcal{D}(S_{\mathcal{V,L}}):\|\psi\|=1}{\mbox{Im}}\langle \psi,(S_{\mathcal{V,L}}+iV)\psi\rangle\geq -\gamma\|\psi\|^2\:.
\end{equation}\\
ii) If $S_\mathcal{V}$ is not dissipative, i.e. if there exists a $v\in\mathcal{V}$ such that
\begin{equation*}
{\mbox{Im}}\langle v,S_\mathcal{V}v\rangle<0\:,
\end{equation*}
then there exists no extension $S_{\mathcal{V,L}}$ and no bounded non-negative operator $V\geq 0$ such that $S_{\mathcal{V,L}}+iV$ is dissipative.\\
iii) If there exists an $\varepsilon>0$ such that
$${\mbox{Im}}\langle v,S_\mathcal{V}v\rangle \geq \varepsilon \|v\|^2$$
for all $v\in\mathcal{V}$ and if the operator $\mathcal{L}$ is bounded, we get that
\begin{equation} \label{eq:vondenleyen}
{\mbox{Im}}\langle \psi, S_{\mathcal{V,L}}\psi\rangle\geq-\frac{\|\mathcal{L}\|^2}{4\varepsilon}\|\psi\|^2
\end{equation}
for all $\psi\in\mathcal{D}(S_{\mathcal{V,L}})$.
This implies in particular that for any bounded $V\geq \frac{\|\mathcal{L}\|^2}{4\varepsilon}$, we get
\begin{equation*}
{\mbox{Im}}\langle \psi, (S_{\mathcal{V,L}}+iV)\psi\rangle\geq 0
\end{equation*}
for all $\psi\in\mathcal{D}(S_{\mathcal{V,L}})$.
\end{corollary}
\begin{proof} i) By Theorem \ref{thm:bern}, Condition \eqref{eq:ragout}, it is necessary that
\begin{equation*}
{\mbox{Im}}\langle v,S^*v\rangle\geq\frac{1}{4}\|V^{-1/2}\mathcal{L}v\|^2
\end{equation*}
for all $v\in\mathcal{V}$. But since $S_\mathcal{V}=S^*\upharpoonright_{\mathcal{D}(S_\mathcal{V})}$ is symmetric, we get ${\mbox{Im}}\langle v,S^*v\rangle= 0$ for all $v\in\mathcal{V}$, which makes it necessary that $\mathcal{L}v=0$ for all $v\in\mathcal{V}$ for $(S_{\mathcal{V,L}}+iV)$ to be dissipative. In other words, only for $\mathcal{L}\equiv 0$ do we have that $A_{\mathcal{V,L}=0}=(S_{\mathcal{V,L}=0}+iV)$ is dissipative. For the second part of i), assume that the imaginary part of $A_{\mathcal{V,L}}$ is semibounded with semibound $-\gamma$ (cf.\ \eqref{eq:semibounded}). This would mean that the operator $S_{\mathcal{V,L}}+i(V+\gamma)$ is dissipative, which by Condition \eqref{eq:ragout} would imply that for all $v\in\mathcal{V}$, the condition
\begin{equation*}
0= {\mbox{Im}}\langle v,S^*v\rangle\geq\frac{1}{4}\|(V+\gamma)^{-1/2}\mathcal{L}v\|^2\:,
\end{equation*}
is satisfied, which is impossible if $\mathcal{L}\neq 0$.
ii) Let $v$ be an element of $\mathcal{V}$ such that ${\mbox{Im}}\langle v,S_\mathcal{V}v\rangle<0$. Thus, by Condition \eqref{eq:ragout} from Theorem \ref{thm:bern}, the operator $(S_{\mathcal{V,L}}+iV)$ cannot be dissipative for any choice of $\mathcal{L}$ or $V$.
iii) Assume now that there exists an $\varepsilon> 0$ such that ${\mbox{Im}}\langle v,S_{\mathcal{V}}v\rangle={\mbox{Im}}\langle v,S^*v\rangle\geq\varepsilon\|v\|^2$ for all $v\in\mathcal{V}$. If $\mathcal{L}=0$, \eqref{eq:vondenleyen} clearly holds with $\|\mathcal{L}\|=0$. Now, let $\mathcal{L}\neq 0$. Again, by Condition \eqref{eq:ragout} of Theorem \ref{thm:bern}, the operator $S_{\mathcal{V,L}}+i\frac{\|\mathcal{L}\|^2}{4\varepsilon}$ is dissipative if and only if
\begin{equation} \label{eq:farage}
{\mbox{Im}}\langle v,S^*v\rangle\geq\frac{1}{4}\left\|\left(\frac{\|\mathcal{L}\|^2}{4\varepsilon}\right)^{-1/2}\mathcal{L}v\right\|^2
\end{equation}
for all $v\in\mathcal{V}$. Since for all $v\in\mathcal{V}$ we may estimate
$$ \frac{1}{4}\left\|\left(\frac{\|\mathcal{L}\|^2}{4\varepsilon}\right)^{-1/2}\mathcal{L}v\right\|^2= \frac{4\varepsilon}{4\|\mathcal{L}\|^2}\|\mathcal{L}v\|^2\leq \varepsilon\|v\|^2\leq{\mbox{Im}}\langle v,S^*v\rangle\:,$$
this proves that \eqref{eq:farage} is satisfied. Hence the operator $S_{\mathcal{V,L}}+i\frac{\|\mathcal{L}\|^2}{4\varepsilon}$ is dissipative, which is equivalent to
$${\mbox{Im}}\langle \psi,S_{\mathcal{V,L}}\psi\rangle\geq -
\frac{\|\mathcal{L}\|^2}{4\varepsilon}\|\psi\|^2 $$
for all $\psi\in\mathcal{D}(S_{\mathcal{V,L}})$. This finishes the proof.
\end{proof}
\begin{example}[Schr\"odinger operator on the half-line] \normalfont Let $\mathcal{H}=L^2({\mathbb R}^+)$ and consider the closed symmetric operator $S$ given by:
\begin{align*}
S:\qquad\mathcal{D}(S)=\{f\in H^2({\mathbb R}^+): f(0)=f'(0)=0\},\quad f\mapsto -f''\:.
\end{align*}
Its adjoint is given by
\begin{align*}
S^*:\qquad\mathcal{D}(S^*)= H^2({\mathbb R}^+),\quad f\mapsto -f''\:,
\end{align*}
where in both cases, $f''$ denotes the second weak derivative of $f$. Since for any $f\in\mathcal{D}(S^*)$ we have
\begin{equation*}
{\mbox{Im}}\langle f,S^*f\rangle=-{\mbox{Im}}\left(\int_0^\infty \overline{f(x)}f''(x)dx\right)={\mbox{Im}} \left(\overline{f(0)}f'(0)\right)\:,
\end{equation*}
from which it can be easily shown that all maximally dissipative extensions of $S$ are parametrized by the boundary condition
\begin{align*}
S_h:\qquad\mathcal{D}(S_h)&=\{f\in H^2({\mathbb R}^+): f'(0)=h f(0)\}\\
f&\mapsto -f''\:,
\end{align*}
where ${\mbox{Im}}(h)\geq 0$. Since $S$ is symmetric, we may choose $$\mathcal{D}(S^*)//\mathcal{D}(S)=\ker(S^*+i)\dot{+}\ker(S^*-i)\:.$$ Now pick $\eta_h\in\mathcal{D}(S^*)//\mathcal{D}(S)$ such that $\eta'_h(0)=h$ and $\eta_h(0)=1$, which means that $\mathcal{D}(S_h)=\mathcal{D}(S)\dot{+}{\mbox{span}}\{\eta_h\}$ with the understanding that $h=\infty$ corresponds to Dirichlet boundary conditions at the origin. This implies that
\begin{equation} \label{eq:labyrinth}
{\mbox{Im}} \langle \eta_h,S^*\eta_h\rangle={\mbox{Im}} h\:,
\end{equation}
where we introduce the convention that ${\mbox{Im}}(\infty)=0$ since $S_\infty$ is selfadjoint. By Theorem \ref{thm:bern}, Condition \eqref{eq:ragout}, we get that for $h=\infty$ the only linear map $\mathcal{L}$ that describes a dissipative extension $S_{\mathcal{V_\infty,L}}$ is given by $\mathcal{L}\equiv 0$, which corresponds to a proper dissipative extension. Here, $\mathcal{V}_\infty:={\mbox{span}}\{\eta_\infty\}$. Hence, we will not treat this case anymore from now on.
Now, for $h\neq\infty$, the map $\mathcal{L}$ from $\mathcal{V}={\mbox{span}}\{\eta_h\}$ has to be of the form $\mathcal{L}\eta_h=k$ for some $k\in\mathcal{H}$. Thus, any $f\in\mathcal{D}(S_h)$ can be written as $f=(f-f(0)\eta_h)+f(0)\eta_h$, where $(f-f(0)\eta_h)\in\mathcal{D}(S)$. This means that the operator $S_{\mathcal{V,L}}$ is given by
\begin{align*} S_{\mathcal{V,L}}:\qquad\mathcal{D}(S_{\mathcal{V,L}})&=\mathcal{D}(S_h)\\
S_{\mathcal{V,L}}f&=-f''+f(0)k\:.
\end{align*}
Since $S_{\mathcal{V,L}}$ only depends on our choice of $h\in{\mathbb C}$ and $k\in\mathcal{H}$, let us use these two parameters to label $S_{\mathcal{V,L}}=S_{h,k}$. Let us now consider a two different bounded dissipative perturbations:
\begin{itemize}
\item Let us start with a rank-one perturbation of the form $V=\alpha|\varphi\rangle\langle\varphi|$, where $\alpha>0$ and $\|\varphi\|=1$. Since ${\mbox{ran}}\, V={\mbox{ran}}\, V^{1/2}={\mbox{span}}\{\varphi\}$, the first condition of Theorem \ref{thm:bern} yields that $k\in{\mbox{span}}\{\varphi\}$. Moreover, on ${\mbox{span}}\{\varphi\}$, the operator $V^{-1/2}$ is given by $\varphi\mapsto \alpha^{-1/2}\varphi$. Thus, the second condition of Theorem \ref{thm:bern} reads as
\begin{equation}
\frac{1}{4}\|\alpha^{-1/2}\lambda\varphi\|^2\leq {\mbox{Im}} h\:\:\Leftrightarrow\:\: |\lambda|^2\leq 4\alpha{\mbox{Im}} h\:,
\end{equation}
where we have parametrized $k=\lambda\varphi$.
Thus, all (maximally) dissipative extensions of the operator
\begin{align*}
A:\quad\mathcal{D}(A)&=\{f\in H^2({\mathbb R}^+):f(0)=f'(0)=0\}\\
f&\mapsto -f''+i\alpha \langle \varphi,f\rangle\varphi
\end{align*}
are given by the family of operators $A_{h,\lambda}$, where $|\lambda|^2\leq 4\alpha{\mbox{Im}} h$:
\begin{align*}
A_{h,\lambda}:\qquad\mathcal{D}(A_{h,\lambda})&=\{f\in H^2({\mathbb R}^+): f'(0)=h f(0)\}\\
f&\mapsto -f''+f(0)\lambda\varphi+i\alpha\varphi\langle\varphi,f\rangle\:.
\end{align*}
\item Now, let $V$ be the multiplication operator by an a.e.\ non-negative function $V(x)\in L^\infty({\mathbb R}^+)$. Moreover for any function $h\in L^2({\mathbb R}^+)$, let $\mathcal{E}_h:=\{x:h(x)\neq 0\}$, which is defined up to a set of Lebesgue measure zero. Clearly, $\overline{{\mbox{ran}}\,V}=L^2(\mathcal{E}_V)$. Hence, the first condition of Theorem \ref{thm:bern} yields the requirement that $\mathcal{E}_k\subset\mathcal{E}_V$ up to a set of Lebesgue measure zero. Next, $k\in\mathcal{D}(V^{-1/2})$ implies that $k$ has to be such that
\begin{equation*}
\int_{\mathcal{E}_V}\frac{|k(x)|^2}{V(x)}dx<\infty\:.
\end{equation*}
Lastly, the second condition of Theorem \ref{thm:bern} reads as
\begin{equation*}
\int_{\mathcal{E}_V}\frac{|k(x)|^2}{V(x)}dx\leq 4{\mbox{Im}} h\:.
\end{equation*}
Thus, all (maximally) dissipative extensions of the operator
\begin{align*}
A:\quad\mathcal{D}(A)&=\{f\in H^2({\mathbb R}^+):f(0)=f'(0)=0\}\\
(Af)(x)&= -f''(x)+iV(x)f(x)
\end{align*}
are given by the family of operators $A_{h,k}$, where $k\in\mathcal{H}$ such that ${\mathcal{E}_k}\subset\mathcal{E}_V$ (up to a set of Lebesgue measure zero) and
\begin{equation*}
\int_{\mathcal{E}_V}\frac{|k(x)|^2}{V(x)}dx\leq 4{\mbox{Im}} h\:.
\end{equation*}
They are given by:
\begin{align*}
A_{h,k}:\qquad\mathcal{D}(A_{h,\lambda})&=\{f\in H^2({\mathbb R}^+): f'(0)=h f(0)\}\\
(A_{h,k}f)(x)&= -f''(x)+f(0)k(x)+iV(x)f(x)\:.
\end{align*}
\end{itemize}
\end{example}
\subsection*{Acknowledgements} The main part if this research was carried out during the author's PhD studies at the University of Kent in Canterbury, UK. It is thus a pleasure to acknowledge the support and guidance of his supervisors Ian Wood and Sergey Naboko. Also, he would like to thank them for carefully reviewing this manuscript and making useful suggestions. Moreover, he is indebted to the UK Engineering and Physical Sciences Research Council (Doctoral Training Grant Ref.\ EP/K50306X/1) and the School of Mathematics, Statistics and Actuarial Science at the University of Kent for a PhD studentship.
He is also very grateful to the referees for helpful remarks and valuable suggestions, in particular for pointing out that the proofs of Lemma \ref{lemma:singulariter} and Corollary \ref{coro:ausmachen} cover more than just the essentially selfadjoint case and also for suggesting the example given in Remark \ref{rem:referee}.
\bigskip
| 28e6e8112592911b87c53ddc7bf063d44bf9e2e8 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{}
\tableofcontents
\newpage
\section{Introduction}
The celebrated singularity theorems in general relativity~\cite{singularities} suggest that the complete description of a black hole requires a quantum theory of gravity that can treat singularities in a consistent manner.
One would expect that singularities in classical general relativity are cured by quantum gravity in much the same way that the classical singularity is resolved in a simplified model of the quantized hydrogen atom.
This suggests a number of natural questions about properties of quantum black holes, such as their mass spectrum and interior structure, that are related to the end state of Hawking radiation. Apart from the important theoretical issues it raises, black-hole evaporation provides one of the few possible ultra-high-energetic astrophysical events that might one day be used to test a given quantum theory of gravity.
Unfortunately, we do not as yet have a complete quantum theory of gravity to provide us with precise answers to such questions.
Suggestive results may none the less be obtained using a midi-superspace approach in which one quantizes only a restricted class of highly symmetric spacetimes~\cite{midisuperspace}.
There are two different canonical methods for quantizing general relativity as a constrained dynamical system: Dirac quantization~\cite{Dirac} and reduced phase-space quantization.
In the former method, the constraints in the Einstein field equations become operators acting on the wave function(al) of spacetime. The resulting Wheeler-DeWitt equation provides the basis for Dirac quantized gravity.
In reduced phase-space quantization, on the other hand, one first chooses a gauge, solves the classical constraint equations, and then puts the solutions to the constraints and gauge fixing conditions back into the action, which is then quantized on the resulting reduced phase space. One variation of this approach is to first do a suitable canonical transformation to a set of phase space coordinates that separate out the gauge invariant modes and decouple them from the modes that are ``pure gauge". The gauge fixing procedure then becomes straightforward. This is the approach that we take in the following.
Although reduced phase-space quantization is a very natural way to quantize constrained dynamical systems, it is technically hopeless to implement with full generality in Einstein's theory.
It is nonetheless possible, {and highly instructive}, to carry out the reduced phase-space quantization of spherically symmetric vacuum spacetimes~\cite{kuchar94}. Classically, the Birkhoff theorem guarantees that the Schwarzschild solution is the unique one-parameter family of solutions in this system.
Following Kucha\v{r}~\cite{kuchar94}, we take the Misner-Sharp quasi-local mass $M$~\cite{ms1964}, the areal radius ${R}$, and their conjugate momenta, $P_M$ and $P_R$, respectively, as canonical variables. The resulting constraint equations are trivial to solve: ${M}={\bf m}(t) $ is independent of the spatial coordinate, and $P_R= 0 $ on the constraint equations. In particular, they solve the constraint equations, resulting in a two-dimensional reduced phase space.
The resulting reduced action is simply
\begin{align}
I[{\bf m},{\bf p}]=\int {\rm d} t \biggl({\bf p}(t) \dot{{\bf m}}(t) - (N_+-N_-){\bf m}(t)\biggl), \label{intro}
\end{align}
where a dot denotes a derivative with respect to $t$~\cite{kuchar94}.
Here the dynamical variable is the spatially constant mode ${\bf m}(t)$ of the Misner-Sharp mass; i.e., its value on the constraint surface. Its conjugate momentum ${\bf p}$ is the Schwarzschild time separation between the two ends of the spatial slices at fixed time $t$, and the Hamiltonian is the mass ${\bf m}$ itself~\cite{kuchar94}.
The prescribed functions $N_\pm(t)$ are the values of the lapse at either end of the spatial slice and are not varied.
The Hamilton equation for ${\bf m}$ is simply ${\dot {\bf m}}=0$, which requires the spatially constant mode of the Misner-Sharp mass, ${\bf m}$, to be constant in time as well.
This Kucha\v{r} reduction has been generalized in arbitrary dimensional spacetime with spherical, plane, or hyperbolic symmetry in the presence of a cosmological constant~\cite{dl2009}, and also in vacuum spherically symmetric Lovelock gravity~\cite{lovelock}, the most general metric theory of gravity yielding second-order field equations in arbitrary dimensions~\cite{kmt2013}.
The Kucha\v{r} action (\ref{intro}) in fact provides a firm {geometrical} foundation for studying the quantum mechanics of the Schwarzschild-Tangherlini-type vacuum black holes in any theory admitting a Birkhoff's theorem.
Based in part on Kucha\v{r}' work, Louko and M\"akel\"a pioneered the method of throat quantization of the Schwarzschild black hole~\cite{LM96}.
They performed a canonical transformation from the Kucha\v{r} action (\ref{intro}) with $N_+=1$ and $N_-=0$ to a new action that describes the dynamics of the radius of the wormhole throat on a maximal slicing of the maximally extended Schwarzschild black-hole spacetime.
While $t$ is interpreted in the new action as the proper time on the wormhole throat, as pointed out by Louko and M\"akel\"a, one can choose the slicings so that $t$ is equal to the proper time in one of the asymptotically flat regions in the Schwarzschild spacetime.
Because the Hamiltonian for the throat dynamics is the Misner-Sharp mass, they quantized this system and identified its energy eigenvalues as mass eigenvalues of the Schwarzschild black hole~\cite{LM96}.
In a previous paper~\cite{km2014}, we generalized results of Ref.~\cite{LM96} to arbitrary-dimensional vacuum black holes with spherical, planar, or hyperbolic symmetry, with or without a cosmological constant.
As in standard quantum mechanics, there are ambiguities in the choice of the operator ordering and boundary conditions.
In our study, we adopted Laplace-Beltrami operator ordering as a natural choice and imposed boundary conditions that preserve unitarity. As expected, we obtained discrete mass spectra for Schwarzschild-Tangherlini-type black holes in the stationary state~\cite{km2014}.
While the spectrum was obtained in the WKB approximation in most of the cases, we obtained an exact mass spectrum with a positive lower bound for asymptotically AdS toroidal black holes.
This exact spectrum was bounded below by a positive Planck scale number, suggesting that the final state after the Hawking radiation of a toroidal black hole in a system with less symmetry would also be a Planck mass relic.
In quantum mechanics, it is natural to expect that an isolated system settles down to a stationary state in the far future due to dissipation of energy to the surroundings.
Nonetheless, for suitably isolated systems, non-stationary exact quantum states also can have physical relevance. The coherent states of the simple harmonic oscillator are a prime example. In quantum gravity, which is as yet not well understood at a fundamental level, the study of such exact dynamical states is particularly useful because it may shed insight into the expected semi-classical behaviour.
One key difference between the stationary and non-stationary states is that, while the expectation value of the mass is constant for both, mass uncertainty exists only in the non-stationary case. This raises the following questions:
\begin{itemize}
\item How does the mass uncertainty change in time?
\item Does it have a maximum or minimum value?
\item How do we define the horizon of a quantum black hole in the dynamical case?
\item Once defined, how large are the fluctuations of the quantum horizon, and how far from the classical value is it located?
\item In the large-mass limit, do the quantum and classical horizons coincide, and do the quantum fluctuations become small?
\end{itemize}
We will address these questions in the framework of throat quantization of toroidal AdS black holes using two families of exact time-dependent wave functions.
One interesting result is that for highly non-stationary states of large-mass black holes, quantum fluctuations are not negligible in one family, while they are greatly suppressed in the other. Such states therefore provide candidates for describing the dynamics of semi-classical black holes.
The outline of the present paper is as follows.
In Sec.~\ref{sec:throat}, we review the throat quantization method and summarize our previous results for stationary states.
We also introduce the concepts of classical and quantum horizons as well as the semi-classical limit.
In Sec.~\ref{sec:wave} the wave functions for the toroidal black hole are presented.
In Secs.~\ref{sec:exact1} and \ref{sec:exact2}, we study the properties of quantum toroidal black holes described by two distinct families of wave functions.
Our conclusions and discussions are summarized in Sec.~\ref{sec:summary}.
Integral formulae involving Hermite polynomials that are used in the main text are given in Appendix A, while the details of computations are presented in Appendices B and C.
Our basic notation follows~\cite{wald}.
The convention for the Riemann curvature tensor is $[\nabla _\rho ,\nabla_\sigma]V^\mu ={R^\mu }_{\nu\rho\sigma}V^\nu$ and $R_{\mu \nu }={R^\rho }_{\mu \rho \nu }$.
The Minkowski metric is taken as diag$(-,+,\cdots,+)$, and Greek indices run over all spacetime indices.
We use a symbol $\kappa_n:=\sqrt{8\pi G_n}$, where $G_n$ is the $n$-dimensional Newton constant.
The Planck length and the Planck mass are defined by $\ell_{\rm p}:=(\hbar\kappa_n^2/c^3)^{1/(n-2)}$ and $m_{\rm p}:=(\hbar^{n-3}/\kappa_n^2c^{n-5})^{1/(n-2)}$, respectively.
We will use the Planck area and the Planck volume defined by $A_{\rm p}:=V_{n-2}^{(0)}\ell_{\rm p}^{n-2}$ and $V_{\rm p}:=V_{n-2}^{(0)}\ell_{\rm p}^{n-1}/(n-1)$, respectively, where $V_{n-2}^{(0)}$ represents volume of the $(n-2)$-dimensional subspace with toroidal symmetry.
Here we have kept the speed of light $c$ explicitly, but in the main text we set $c=1$.
In these units, we have $\hbar\kappa_n^2=\ell_{\rm p}^{n-2}$ and $\hbar=m_{\rm p}\ell_{\rm p}$.
\section{Throat quantization of vacuum symmetric black holes}
\label{sec:throat}
\subsection{Preliminaries}
We consider general relativity with a cosmological constant $\Lambda$ in arbitrary $n(\ge 3)$ dimensions. The action is given by
\begin{align}
I=&\frac{1}{2\kappa_n^2}\int {\rm d} ^nx\sqrt{-g}(R-2\Lambda)+I_{\partial{\cal M}},\label{action}
\end{align}
where $I_{\partial{\cal M}}$ is the York-Gibbons-Hawking boundary term.
The above action gives the following vacuum Einstein field equations:
\begin{align}
R_{\mu\nu}-\frac12 g_{\mu\nu}R+\Lambda g_{\mu\nu}=0. \label{beqL}
\end{align}
We assume an $n$-dimensional warped product spacetime $({\cal M}^n,g_{\mu\nu}) \approx ({M}^2,g_{AB})\times ({K}^{n-2},\gamma_{ab})$, of which the most general metric is given by
\begin{align}
g_{\mu\nu}(x){\rm d} x^\mu {\rm d} x^\nu=g_{AB}({\bar y}){\rm d} {\bar y}^A {\rm d} {\bar y}^B+r({\bar y})^2\gamma_{ab}(z){\rm d} z^a{\rm d} z^b,
\label{eq:structure}
\end{align}
where indices run as $A,B=0,1$ and $a,b=2,3,\cdots,n-1$.
Here $({M}^2,g_{AB})$ is the most general two-dimensional Lorentzian manifold, and $({K}^{n-2},\gamma_{ab})$ is the $(n-2)$-dimensional maximally symmetric space with its curvature $k=1, 0, -1$; namely, the Riemann tensor ${}^{(n-2)}R^{ab}_{~~cd}$ on $({K}^{n-2},\gamma_{ab})$ is given by
\begin{eqnarray}
{}^{(n-2)}R^{ab}_{~~cd}=k(\delta^a_c \delta^b_d-\delta^a_d \delta^b_c).
\end{eqnarray}
Note that, in three dimensions, $({K}^{n-2},\gamma_{ab})$ is one-dimensional and $k=0$ necessarily holds.
We assume that $({K}^{n-2},\gamma_{ab})$ is compact.
$r({\bar y})$ is a scalar on $({M}^2,g_{AB})$, called the areal radius because $r^{n-2}$ is proportional to the area of a symmetric $(n-2)$-surface generated by the spatial Killing vectors on $({K}^{n-2},\gamma_{ab})$.
The generalized Misner-Sharp quasi-local mass~\cite{ms1964,mn2008} in this system is defined by
\begin{align}
M:=& \frac{(n-2)V_{n-2}^{(k)}}{2\kappa_n^2}r^{n-3}\biggl(\frac{r^2}{l^2}+(k-(Dr)^2)\biggl),\label{qlm}
\end{align}
where $D_A$ is the covariant derivative on $({M}^2,g_{AB})$, $(Dr)^2:=(D_A r)(D^A r)$, and the AdS radius $l$ is defined by
\begin{align}
l^2:=&-\frac{(n-1)(n-2)}{2\Lambda}. \label{lambdatil}
\end{align}
The constant $V_{n-2}^{(k)}$ represents the volume of $({K}^{n-2},\gamma_{ab})$.
$M$ reduces to the ADM mass at spacelike infinity in the asymptotically flat spacetime~\cite{mn2008,hayward1996}.
\subsection{Throat dynamics of vacuum black holes}
\label{sec:ClassicalThroatDynamics}
In the vacuum case, the Einstein equations show that the mass function $M$ is constant.
Then, in the case of $k\Lambda \le 0$, the general vacuum solution is the following generalized Schwarzschild-Tangherlini solution:
\begin{align}
{\rm d} s^2=&-f(r){\rm d} t^2+f(r)^{-1}{\rm d} r^2+r^2\gamma_{ab}{\rm d} z^a {\rm d} z^b, \label{f-vacuum} \\
f(r) =&k-\frac{2\kappa_n^2M}{(n-2)V_{n-2}^{(k)}r^{n-3}}+\frac{r^2}{l^2}.
\end{align}
For $k\Lambda > 0$, by contrast, the general vacuum solution consists of the solution (\ref{f-vacuum}) and the Nariai ($k=1$) or anti-Nariai ($k=-1$) direct product solution in which the areal radius $r$ is constant.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.5\linewidth]{SinglehorizonBHallPRD}
\caption{\label{SingleBH} Penrose diagrams for the Schwarzschild-Tangherlini-type black hole (\ref{f-vacuum}) with a single horizon with (a) $\Lambda=0$ and (b) $\Lambda<0$. While a zigzag line represents a curvature singularity at $r=0$, a dashed curve in each portion of the spacetime represents a constant $t$ hypersurface.
In (a), a double line represents null infinity. In (b), it represents both null and spacelike infinities.}
\end{center}
\end{figure}
In the generalized Schwarzschild-Tangherlini spacetime (\ref{f-vacuum}), there is a central curvature singularity at $r=0$ and a Killing horizon located at $r=r_{\rm h}$ such that $f(r_{\rm h})=0$.
Hence, the mass-horizon relation is given by
\begin{align}
M=\frac{(n-2)V_{n-2}^{(k)}r_{\rm h}^{n-3}}{2\kappa_n^2}\biggl(k+\frac{r_{\rm h}^2}{l^2}\biggl). \label{mass-horizon}
\end{align}
The Penrose diagrams for the generalized Schwarzschild-Tangherlini black hole (\ref{f-vacuum}) with a single horizon are drawn in Fig.~\ref{SingleBH}.
(The black hole does not possess multiple horizons in the case of $k=1$ or $k=0$.)
The maximally extended black-hole spacetime drawn in Fig.~\ref{SingleBH} has the structure of a wormhole whose throat is located inside the event horizon $r=r_{\rm h}$.
The areal radius of this wormhole throat evolves with time. We represent it by $r=a(\tau)$, where $\tau$ is the proper time on the throat. Its time evolution is generated by the following Hamiltonian $H[a(\tau),p_a(\tau)]$:
\begin{align}
H=&\frac{(n-2)V_{n-2}^{(k)}a^{n-3}}{2\kappa_n^2}\biggl\{\biggl(\frac{(n-2)V_{n-2}^{(k)}}{\kappa_n^2}\biggl)^{-2}p_a^2a^{-2(n-3)}+k+\frac{a^2}{l^2}\biggl\}, \label{hamiltonian}
\end{align}
where $p_a$ is the momentum conjugate of $a$.
The behaviour of the throat radius as described by the resulting solutions $a(\tau)$ is as follows: the throat radius starts to increase from the white-hole singularity at $a=0$ until it reaches its turning point at the bifurcation $(n-2)$-surface $a=r_{\rm h}$. The throat radius then decreases from its maximum value $a=r_{\rm h}$, and finally reaches the black-hole singularity at $a=0$.
The maximum value $r_{\rm h}$ of $a$ coincides with the horizon radius for any given value of $M$.
Since the horizon radius as given in Eq.~(\ref{mass-horizon}) increases indefinitely as $M$ increases, the domain of $a$ is $0\le a <\infty$, while the domain of $p_a$ is $-\infty<p_a<\infty$.
Incidentally, this provides a fundamental motivation for quantization using throat variables, rather than the Misner-Sharp mass, as done by Kucha\v{r}. As described above, the physics prescribes the range of the phase-space variables $a$ and $p_a$, which in turn determine the spectrum of the energy. If one tries to quantize the energy directly, its conjugate variable is effectively time, which presumably is a continuous variable in $(-\infty,+\infty)$, so that one would naturally obtain a continuous spectrum of plane wave solutions with mass unbounded below. In this context, a continuous spectrum $(-\infty,+\infty)$ for the mass $M$ is natural also, because the Kucha\v{r} reduced action (\ref{intro}) does not retain any information about the fact that the spacetime is that of a black hole. Moreover, if one tries to restrict the range of $M$ to ${\mathbb R}^+$, then as with a free particle on the half line, its conjugate would not exist as a self-adjoint operator.
A fundamental feature of throat quantization is that, as shown for Schwarzschild black holes in Ref.~\cite{LM96}, the corresponding Hamiltonian (\ref{hamiltonian}) can be obtained by a canonical transformation from the Kucha\v{r} action (\ref{intro}) with $N_+=1$ and $N_-=0$, in which the Hamiltonian is the mass $M$ of the black hole and the time $t$ coincides with the proper time $\tau$ of an observer at rest on the throat of the wormhole.
One then quantizes the throat dynamics, whose action is given by
\begin{align}
I=\int {\rm d} t\biggl(p_a{\dot a}-H[a,p_a]\biggl). \label{newaction}
\end{align}
We emphasize that $t$ in the new action (\ref{newaction}) is originally the Kucha\v{r}' asymptotic time variable, which is the proper time of an observer at spacelike infinity in the spacetime (\ref{f-vacuum}), so the expectation values of physical quantities are those measured by such an observer.
The existence of such a canonical transformation provides the required connection between the throat variables and the geometrodynamics of the full diffeomorphism invariant spacetime that underlies the Kucha\v{r} action.
\subsection{Hamiltonian operator}
We now proceed to quantize (\ref{newaction}) using Schr\"odinger quantization.
Replacing $p_a$ in the Hamiltonian (\ref{hamiltonian}) by ${\hat p}_a=-i\hbar \partial/\partial a$, we obtain the Schr\"odinger equation for the wave function $\Psi(t,a)$:
\begin{align}
{\hat H}\Psi =i\hbar\frac{\partial\Psi}{\partial t}. \label{Schro}
\end{align}
Adopting Laplace-Beltrami operator ordering~\cite{zanelli1986}, we obtain the following Hamiltonian operator~\cite{km2014}:
\begin{align}
{\hat H} =&\frac{(n-2)V_{n-2}^{(k)}}{2\kappa_n^2}\biggl\{\frac{a^{n-1}}{l^2}+ka^{n-3}-\biggl(\frac{\kappa_n^2}{(n-2)V_{n-2}^{(k)}}\biggl)^2\frac{\hbar^2}{a^{(n-3)/2}}\frac{\partial}{\partial a}\biggl(\frac{1}{a^{(n-3)/2}}\frac{\partial }{\partial a}\biggl)\biggl\}.
\label{H_LB}
\end{align}
This operator is Hermitian symmetric with respect to the inner product defined by
\begin{align}
\langle \Psi|\Phi\rangle:=\int^\infty_0\Psi^*\Phi\mu(a){\rm d} a.
\end{align}
for two arbitrary wave functions $\Psi$ and $\Phi$, with measure $\mu(a)=a^{(n-3)/2}$. Louko and M\"akel\"a considered the more general measure $\mu(a) = a^\beta$~\cite{LM96}, {and corresponding Hermitian ordering of the momentum term in the Hamiltonian}.
Defining $x:=a^{(n-1)/2}$, we rewrite ${\hat H}$ as
\begin{align}
{\hat H} =&\frac{(n-2)V_{n-2}^{(k)}}{2\kappa_n^2}\biggl\{\frac{x^2}{l^2}+kx^{2(n-3)/(n-1)} -\hbar^2\biggl(\frac{(n-1)\kappa_n^2}{2(n-2)V_{n-2}^{(k)}}\biggl)^2\frac{\partial ^2}{\partial x^2}\biggl\}. \label{H-x}
\end{align}
Note that $x^2 = a^{n-1}$ is proportional to the Euclidean (or special relativistic) volume of a throat with radius $a$.
Comparing Eq.~(\ref{Schro}) with
\begin{align}
\biggl(-\frac{\hbar^2}{2m}\frac{\partial ^2}{\partial x^2}+V(x)\biggl)\psi=i\hbar\frac{\partial\Psi}{\partial t}, \label{Schro-eq}
\end{align}
we identify the effective mass and potential as
\begin{align}
V(x)\equiv &\frac{(n-2)V_{n-2}^{(k)}}{2\kappa_n^2}\biggl(\frac{x^2}{l^2}+kx^{2(n-3)/(n-1)}\biggl),\label{p-potential}\\
m\equiv &\frac{4(n-2)V_{n-2}^{(k)}}{(n-1)^2\kappa_n^2}. \label{p-mass}
\end{align}
In terms of $x$, the inner product becomes simple:
\begin{align}
\langle \Psi|\Phi\rangle:=\int^\infty_0\Psi^*\Phi{\rm d}{x}.
\end{align}
It is important to note that since the exponent in the second term of (\ref{p-potential}) is less than 2 for all values of $n$, the potential $V(x)$ is bounded below for $k=-1$ with $\Lambda<0$.
\subsection{Self-adjointness of the Hamiltonian operator}
We require the quantum system to obey unitarity:
\begin{align}
\frac{{\rm d}}{{\rm d} t}\langle \Psi|\Phi\rangle=0\,.
\end{align}
This requires the Hamiltonian operator ${\hat H}$ to be self-adjoint on the domain of $0 \le x<\infty$ or to have a self-adjoint extension, which guarantees reality of the eigenvalues.
In our previous paper~\cite{km2014}, we showed that the Hamiltonian operator (\ref{H-x}) on the half line $x\in [0,\infty)$ admits an infinite number of self-adjoint extensions. We now summarize the arguments that lead to this conclusion.
(See Ref.~\cite{robin bcs} for general discussions of self-adjoint extensions of operators.)
Starting with Eqs.~(\ref{Schro}) and (\ref{H-x}) and integrating by parts with fall-off condition $\Psi\to 0$ as $x\to \infty$, one obtains
\begin{align}
\frac{{\rm d}}{{\rm d} t}\langle \Psi|\Phi\rangle=\frac{i\hbar}{2m} \biggl(\frac{\partial\Psi^*}{\partial x}\Phi-\Psi^*\frac{\partial\Phi}{\partial x}\biggl)\biggl|_{x=0}.
\end{align}
The right-hand side of the above equation must vanish for all states in the Hilbert space in order to ensure unitarity of the system.
Thus, the wave functions must obey the following boundary condition:
\begin{align}
\Psi(t,0)+L\frac{\partial\Psi}{\partial x}(t,0)=0,
\label{eq:robin bc}
\end{align}
where $L$ is a real constant.
$L=0$ and $L=\infty$ correspond to Dirichlet and Neumann boundary conditions, respectively.
The remaining values of $L$ correspond to Robin boundary conditions~\cite{robin bcs}.
Since one obtains inequivalent but well-defined quantum theories for different real values of $L$, the Hamiltonian operator (\ref{H-x}) on the half line $x\in [0,\infty)$ admits an infinite number of self-adjoint extensions.
We note that in the case of quantization on the whole line $x\in (-\infty,\infty)$, Ehrenfest's theorem holds;
\begin{align}
m\frac{{\rm d}^2\langle x\rangle}{{\rm d} t^2}=-\biggl\langle \frac{{\rm d} V}{{\rm d} x}\biggl\rangle,
\end{align}
so that $\langle x\rangle$ follows quantum corrected classical orbits (since $\langle \frac{{\rm d} V}{{\rm d} x}\rangle\neq \left.\frac{{\rm d} V}{{\rm d} x}\right|_{x=\langle x\rangle}$).
The above equation is modified in the case of the half-line $x\in [0,\infty)$.
Under the assumption that the surface terms at infinity vanish, and again using the Schr\"odinger equation~(\ref{Schro-eq}) and integration by parts, we obtain
\begin{align}
m\frac{{\rm d}^2\langle x\rangle}{{\rm d} t^2}=&-\frac{\hbar^2}{4m} \biggl(\frac{\partial^2\Psi^*}{\partial x^2}\Psi+\Psi^*\frac{\partial^2\Psi}{\partial x^2}-2\frac{\partial\Psi^*}{\partial x}\frac{\partial\Psi}{\partial x}\biggl)\biggl|_{x=0} -\biggl\langle \frac{{\rm d} V}{{\rm d} x}\biggl\rangle.
\end{align}
Therefore, in general, the expectation value of $x$ does not follow classical orbits due to possible boundary effects\footnote{This is related to the fact that the conjugate operator $\hat{p}=-i\hbar \partial/\partial x$ does not exist on the half line as a self-adjoint operator, nor does it have self-adjoint extensions.}.
(See Ref.~\cite{maeda2015} for the case of $\langle x^q\rangle$, where $q$ is a positive constant.)
{We note, however, that for $q=2$, one has
\begin{align}
m\frac{{\rm d}^2\langle x^2\rangle}{{\rm d} t^2}=& \frac{\hbar^2}{2m}
\biggl\{x\biggl( \frac{\partial^2 \Psi^*}{\partial x^2}\Psi + \Psi^* \frac{\partial^2 \Psi}{\partial x^2} \biggl)
-2\biggl(\frac{\partial \Psi^*}{\partial x}\Psi + \Psi^*\frac{\partial \Psi}{\partial x}\biggl)\biggl\}\biggl|_{x=0}\nonumber\\
& + \int^\infty_0 {\rm d} x\biggl(\frac{2\hbar^2}{m} \frac{\partial \Psi^*}{\partial x}\frac{\partial \Psi}{\partial x}
-2 x \Psi^* \frac{{\rm d} V}{{\rm d} x}\Psi\biggl).
\end{align}
Assuming that the wave function and its first and second derivatives are finite at $x=0$, the boundary terms vanish for Dirichlet and Neumann boundary conditions. These are the two boundary conditions that we consider in what follows. }
\subsection{{Stationary states of the quantum black hole}}
While the classical relation between the mass parameter $M$ and the horizon radius $r_{\rm h}$ is given by Eq.~(\ref{mass-horizon}), {the mass spectrum of the quantum black hole is discrete in throat quantization}.
In a previous paper~\cite{km2014}, we studied the mass spectrum for asymptotically flat and AdS black holes in stationary states.
An intriguing result in Ref.~\cite{km2014} is that the presence of a negative cosmological constant $\Lambda$ drastically changes the spectrum: While the mass of the black hole is equally spaced in the asymptotically AdS case, entropy is equally spaced in the asymptotically flat case.
In this analysis, the entropy spectrum is obtained from the classical relation between the mass $M$ and entropy $S$ of a black hole, which is
\begin{align}
S=\frac{2\pi}{\kappa_n^2}V_{n-2}^{(1)}\biggl(\frac{2\kappa_n^2\langle M\rangle_N}{(n-2)V_{n-2}^{(1)}}\biggl)^{(n-2)/(n-3)} \label{entropy-k=1}
\end{align}
for $k=1$ and $\Lambda=0$, and
\begin{align}
S=&\frac{2\pi}{\kappa_n^2}V_{n-2}^{(0)}\biggl(\frac{2\kappa_n^2l^2\langle M\rangle_N}{(n-2)V_{n-2}^{(0)}}\biggl)^{(n-2)/(n-1)} \label{entropy-k=0}
\end{align}
for $k=0$ and $\Lambda<0$.
The relation (\ref{entropy-k=0}) also holds for black holes with $k=\pm 1$ and $\Lambda<0$ in the large-mass limit.
In the following subsections, we will summarize the results in Ref.~\cite{km2014}.
\subsubsection{Asymptotically flat black hole}
For asymptotically flat black holes ($k=1$) with $\Lambda=0$, the WKB approximation yields a mass spectrum that is not equally spaced. Instead, it is the entropy that is equally spaced~\cite{km2014}:
\begin{align}
S\simeq &4\pi^{3/2}\hbar \frac{\Gamma(\frac{n-1}{2(n-3)}+\frac12)}{\Gamma(\frac{n-1}{2(n-3)})}N=:S_N,\label{spectrum-flat}
\end{align}
where $N$ is a large integer.
This is independent of the extension parameter $L$ in the boundary condition (\ref{eq:robin bc}).
Some particular cases of the above are
\begin{eqnarray}
S_N=
\left\{ \begin{array}{ll}
4\pi^{3/2}\hbar N\frac{\Gamma(2)}{\Gamma(3/2)}=4\pi \hbar N~~(n=4)\\
4\pi^{3/2}\hbar N\frac{\Gamma(3/2)}{\Gamma(1)} =\pi^2 \hbar N~(n=5)\\
4\pi^{3/2}\hbar \frac{\Gamma(1)}{\Gamma(1/2)}N=2\pi\hbar N~~(n\to \infty).
\end{array} \right.
\end{eqnarray}
The relation (\ref{entropy-k=1}) with the result (\ref{spectrum-flat}) means that for large $N$ the mass spectrum behaves as
\begin{align}
\langle M\rangle_N\propto N^{(n-3)/(n-2)}. \label{spectrum-mass-flat}
\end{align}
\subsubsection{Asymptotically AdS black hole}
In the presence of negative $\Lambda$, the mass spectrum for toroidal black holes ($k=0$) under the Dirichlet or Neumann boundary condition can be obtained exactly as
\begin{align}
\langle M\rangle_N=\frac{(n-1)\hbar}{2l}\biggl(N+\frac12\biggl) \quad (N=0,1,2,\cdots), \label{EN:toroidal}
\end{align}
where even (odd) $N$ corresponds to the Neumann (Dirichlet) boundary condition~\footnote{The integer $N$ in Eq.~(3.12) in Ref.~\cite{km2014} with Dirichlet boundary conditions corresponds to half of the integer $N$ in Eqs.~(\ref{EN:toroidal}) and (\ref{spectrum-AdS}) in the present paper.}.
On the other hand, the spectrum for black holes with $k=\pm 1$ has only been obtained in the WKB approximation:
\begin{align}
\langle M\rangle_N \simeq \frac{(n-1)\hbar}{2l}N, \label{spectrum-AdS}
\end{align}
where $N$ is a large integer.
The spectrum (\ref{spectrum-AdS}) for large $N$ is independent of the extension parameter $L$ in the boundary condition (\ref{eq:robin bc}).
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.65\linewidth]{AdSmassspectrumk1}
\caption{\label{fig-AdSspectrum} Mass-horizon relation (\ref{mass-horizon}) for spherical AdS black holes ($k=1$). The mass spectrum is given by Eq.~(\ref{spectrum-flat}) in the region of $\ell_{\rm p}\ll r_{\rm h}\ll l$ and by Eq.~(\ref{spectrum-AdS}) in the region of $r_{\rm h}\gg l$.}
\end{center}
\end{figure}
At first glance, the two results (\ref{spectrum-mass-flat}) and (\ref{spectrum-AdS}) appear to be inconsistent in the limit of vanishing cosmological constant $\Lambda\to0$ ($l\to\infty$) for spherical AdS black holes ($k= 1$).
However, this apparent discrepancy can easily be understood by noting the presence of two different length scales, namely the Planck length $\ell_{\rm p}$ and the AdS radius $l$.
The mass-horizon relation (\ref{mass-horizon}) shows that a spherical AdS black hole is approximated by an asymptotically flat black hole if the horizon radius $r_{\rm h}$ is much smaller than the AdS radius $l$, while it is approximated by the toroidal AdS black hole ($k=0$) if $r_{\rm h}$ is much larger than $l$.
Thus, the spectrum of the spherical AdS black hole is given by Eq.~(\ref{spectrum-flat}) in the region of $\ell_{\rm p}\ll r_{\rm h}\ll l$ and by Eq.~(\ref{spectrum-AdS}) in the region of $r_{\rm h}\gg l$. (See Fig.~\ref{fig-AdSspectrum}.)
In the limit of zero cosmological constant ($l\to\infty$), the black-hole spacetime becomes asymptotically flat so that the region $r_{\rm h}\gg l$ disappears in Fig.~\ref{fig-AdSspectrum} and the mass spectrum (\ref{spectrum-flat}) is valid in the region of $r_{\rm h}\gg \ell_{\rm p}$\footnote{The authors thank Shunichiro Kinoshita for this interpretation of our result.}.
\subsection{Horizons of the quantum toroidal black hole}
Now let us focus on the toroidal black hole ($k=0$).
Classically, the horizon radius $r=r_{\rm h}$ with mass $M$ is given by
\begin{align}
r_{\rm h}=&\biggl(\frac{2\kappa_n^2l^2M}{(n-2)V_{n-2}^{(0)}}\biggl)^{1/(n-1)} =l_{\rm p}\biggl(\frac{2l^2M}{(n-2)V_{n-2}^{(0)}l_{\rm p}^2m_{\rm p}}\biggl)^{1/(n-1)}. \label{c-horizon}
\end{align}
How, then, does one define the location of the ``horizon'' for a quantum black hole?
\subsubsection{``Classical" horizon}
One option is simply to use Eq.~(\ref{c-horizon}) with $M$ replaced by the mass expectation value, namely
\begin{align}
r_{\rm C}:=l_{\rm p}\biggl(\frac{2l^2\langle M\rangle_N}{(n-2)V_{n-2}^{(0)}l_{\rm p}^2m_{\rm p}}\biggl)^{1/(n-1)}. \label{q-horizon1}
\end{align}
In the stationary state, there is no mass fluctuation, so that $r_{\rm C}$ is fixed without ambiguity.
For this reason, we call $r_{\rm C}$ the areal radius of the ``classical'' horizon.
The surface area of the classical horizon is then defined by
\begin{align}
A_{\rm C}:=V_{n-2}^{(0)}r_{\rm C}^{n-2}=A_{\rm p}\biggl(\frac{2l^2\langle M\rangle_N}{(n-2)V_{n-2}^{(0)}l_{\rm p}^2m_{\rm p}}\biggl)^{(n-2)/(n-1)}.\label{def-Ac}
\end{align}
However, in non-stationary states, mass fluctuations $\delta \langle M\rangle_N$ cause the fluctuations of the radius $\delta r_{\rm C}$ and the surface area $\delta A_{\rm C}$ of the classical horizon.
\subsubsection{Quantum horizon}
Another option to define the location of the horizon for a dynamical quantum black hole comes from the throat dynamics.
As explained in Sec.~\ref{sec:ClassicalThroatDynamics}, the classical throat radius $a(t)$ starts to grow from zero corresponding to the white-hole singularity, turns to decrease after reaching the maximum value at the bifurcation $(n-2)$-surface, and finally becomes zero again at the black-hole singularity. (See Fig.~\ref{SingleBH}(b).)
Because the maximum value of $a(t)$ at the bifurcation $(n-2)$-surface corresponds to the horizon radius classically, one may identify the location of the horizon of a quantum black hole by the maximum value of $\langle a\rangle_N(t)$, namely
\begin{align}
r_{\rm Q}:=\max_{t\in\mathbb{R}}\langle a\rangle_N(t).
\end{align}
We call $r_{\rm Q}$ the areal radius of the ``quantum'' horizon.
Note that such an option is not available for stationary states, since of course the expectation values do not change with time.
One can instead identify the quantum turning point in the semi-classical limit with the value of $a$ (far from the origin in the case of the half line) where the probability density is greatest, since for the classical system the probability of finding the particle is greatest near the point where the velocity vanishes.
However, using the above arguments, one may also choose to define the size of the black hole by the maximum value, not of the areal radius $a(t)$, but instead of the surface area $A(t):=V_{n-2}^{(0)}a^{n-2}$.
Then, the surface area of the quantum black hole is defined by
\begin{align}
A_{\rm Q}:=\max_{t\in\mathbb{R}}\langle A\rangle_N(t),
\end{align}
where $A_{\rm Q}\ne V_{n-2}^{(0)}r_{\rm Q}^{n-2}$ holds in general.
The fluctuations of $r_{\rm Q}$ and $A_{\rm Q}$ are, respectively, given by
\begin{align}
\delta r_{\rm Q}:=&\sqrt{|\langle a^2\rangle_N-\langle a\rangle_N^2|}\biggl|_{\langle a\rangle_N=r_{\rm Q}}, \\
\delta A_{\rm Q}:=&\sqrt{|\langle A^2\rangle_N-\langle A\rangle_N^2|}\biggl|_{\langle A\rangle_N=A_{\rm Q}}.
\end{align}
In terms of $x(:=a^{(n-1)/2})$, $\langle a\rangle_N$ and $\langle A\rangle_N$ are written as
\begin{align}
\langle a\rangle_N=&\langle x^{2/(n-1)}\rangle_N, \label{<a>0} \\
\langle A\rangle_N=&V_{n-2}^{(0)}\langle x^{2(n-2)/(n-1)}\rangle_N. \label{<A>0}
\end{align}
In fact, the ``quantum" horizon can in principle be defined using the maximum value of any power of $x$.
They should all agree in the semi-classical limit, if it exists.
\subsubsection{Euclidean volume of the horizon}
The surface area is proportional to the Bekenstein-Hawking entropy, and hence is a natural geometrical quantity characterizing a black hole.
However, because of the rational powers in the expressions (\ref{<a>0}) and (\ref{<A>0}), it is difficult to obtain $\langle a\rangle_N$ or $\langle A\rangle_N$ in a closed form.
In contrast, the Euclidean volume (or special relativistic volume), which we define as $V_{\rm E}:= V_{n-2}^{(0)} a^{n-1}/(n-1)$, has a computable form in terms of $x$:
\begin{align}
\langle V_{\rm E}\rangle_N=\frac{V_{n-2}^{(0)}}{n-1}\langle x^{2}\rangle_N(t).
\end{align}
In the black-hole spacetime, $V_{\rm E}$ does not represent the volume inside a spacelike hypersurface with constant $r$ because $r$ becomes a timelike coordinate inside the horizon\footnote{It is interesting to note that the Euclidean volume of a stationary black hole is related to the more general concept of a vector volume, defined in Ref.~\cite{ballik2013} by examining the rate of growth of an invariant volume of a spacetime region along a divergence-free vector field.}.
In order to obtain clear, analytic versions of the main results, we will use the Euclidean volume $V_{\rm E}$ to measure the size of a quantum black hole instead of the surface area $A$ in the present paper.
The Euclidean volume of the classical horizon and its fluctuation are given, respectively, by
\begin{align}
V_{\rm C}:=&\frac{V_{n-2}^{(0)}r_{\rm C}^{n-1}}{n-1}=\frac{2V_{\rm p}l^2\langle M\rangle_N}{(n-2)V_{n-2}^{(0)}l_{\rm p}^2m_{\rm p}},\label{VC-def} \\
\delta V_{\rm C}:=&\frac{2V_{\rm p}l^2\delta \langle M\rangle_N}{(n-2)V_{n-2}^{(0)}l_{\rm p}^2m_{\rm p}}=V_{\rm C}\frac{\delta \langle M\rangle_N}{\langle M\rangle_N},\label{deltaVC-def}
\end{align}
while the Euclidean volume of the quantum horizon and its fluctuation are
\begin{align}
V_{\rm Q}:=&\max_{t\in\mathbb{R}}\langle V_{\rm E}\rangle_N(t)=\frac{V_{n-2}^{(0)}}{n-1}\max_{t\in\mathbb{R}}\langle x^{2}\rangle_N(t),\\
\delta V_{\rm Q}:=&\sqrt{|\langle V_{\rm E}^2\rangle_N-\langle V_{\rm E}\rangle_N^2|}\biggl|_{\langle V_{\rm E}\rangle_N=V_{\rm Q}}.
\end{align}
\subsubsection{Semi-classical limit}
For a standard quantum harmonic oscillator in the stationary state on the whole line $x\in(-\infty,\infty)$, the energy expectation value $\langle E\rangle_N$ and the uncertainty relation between $x$ and its momentum conjugate $p$ are given by
\begin{align}
\langle E\rangle_N=&\hbar\omega\biggl(N+\frac12\biggl),\\
\delta\langle x\rangle_N \delta\langle p\rangle_N=&\hbar\biggl(N+\frac12\biggl),
\end{align}
where $\omega$ is the angular frequency and $N=0,1,2,\cdots$.
The classical limit is formally defined by $\hbar\to 0$ in which canonical pairs of physical quantities become commutable and the spectrum becomes continuous as the energy spacing $\Delta\langle E\rangle_N:=\langle E\rangle_{N+1}-\langle E\rangle_N=\hbar\omega$ converges to zero.
Independently, {and somewhat more practically}, the limit of large $N$ is also often called the semi-classical limit since the relative energy spacing reduces to zero, namely
\begin{align}
\lim_{N\to\infty}\frac{\Delta\langle E\rangle_N}{\langle E\rangle_N}\to 0.
\end{align}
However, this limit does not reproduce a {purely} classical behaviour, since the uncertainty relation still remains valid.
In fact, the uncertainty in the unit of the Planck constant $\delta\langle x\rangle_N\delta\langle p\rangle_N/\hbar$ is diverging for $N\to \infty$.
We illustrate this point using the quantity $\langle x^2\rangle_N$ because $\delta \langle E\rangle_N=0$ and $\langle x\rangle_N=\langle p\rangle_N=0$ are satisfied in the stationary state.
$\langle x^2\rangle_N$ and its fluctuation are given by
\begin{align}
\langle x^2\rangle_N=&\biggl(N+\frac12\biggl)\ell_0^2,\\
\delta \langle x^2\rangle_N:=&\sqrt{\langle x^4\rangle_N-\langle x^2\rangle_N^2}=\sqrt{\frac{2(N^2+N+1)}{(2N+1)^2}}\langle x^2\rangle_N.
\end{align}
Note that $\ell_0:=\sqrt{\hbar/m\omega}$ is the only length scale in the system, where $m$ is the mass of the oscillator.
These expressions show
\begin{align}
\frac{\Delta\langle x^2\rangle_N}{\ell_0^2}=1,\qquad \frac{\delta \langle x^2\rangle_N}{\ell_0^2}=\sqrt{\frac{N^2+N+1}{2}},
\end{align}
where as before $\Delta\langle x^2\rangle_N:= \langle x^2\rangle_{N-1}-\langle x^2\rangle_N$ denotes the spacing between levels. Hence, the quantum fluctuation $\delta \langle x^2\rangle_N$ increases as $N$ increases and cannot be negligible in comparison with the scale $\ell_0^2$ even for $N=0$.
Alternatively, one may use the following ratios to evaluate the fluctuations:
\begin{align}
\frac{\Delta\langle x^2\rangle_N}{\langle x^2\rangle_N}=&\frac{2}{2N+1},\\
\frac{\delta \langle x^2\rangle_N}{\langle x^2\rangle_N}=&\sqrt{\frac{2(N^2+N+1)}{(2N+1)^2}}.
\end{align}
While the relative spacing $\Delta\langle x^2\rangle_N/\langle x^2\rangle_N$ converges to zero for large $N$, the relative fluctuation $\delta\langle x^2\rangle_N/\langle x^2\rangle_N$ reduces to $1/\sqrt{2}\simeq 0.7071$ for $N\to\infty$. {Of course, for the simple harmonic oscillator on the whole line, the semi-classical behaviour is verified by observing, for example, the shape of the probability amplitude for the position operator $x$. In the large-$N$ limit the quantum probability of finding the particle in an finite segment of the $x$-axis coincides with the corresponding classical property obtained by calculating how much time the oscillating particle spends in that region. In both cases, the probability of finding the particle at the turning points, for example, is greatest, whereas the probability of finding it near the equilibrium point is small.}
In the context of quantum gravity, one expects classical behaviour for black holes that are sufficiently large.
For dynamical quantum black holes, one may therefore define the ``semi-classical limit" by requiring that the quantum fluctuations to be negligible--namely that $\delta \langle X\rangle/\langle X\rangle\ll 1$ for any physical quantity $X$ associated with the black hole.
In such a limit, both $A_{\rm Q}/A_{\rm C}\to 1$ and $V_{\rm Q}/V_{\rm C}\to 1$ are also expected to hold, so that the classical and quantum horizons coincide.
One of the main purposes of the present paper is to identify such a limit for the dynamical quantum black holes described by exact wave functions.\\
\section{{Exact time-dependent wave functions for toroidal black holes}}
\label{sec:wave}
In the previous section, we summarized our main results for stationary quantum black holes.
In this section, we will present exact time-dependent wave functions describing quantum black holes in non-stationary states.
We focus on the case of toroidal black holes ($k=0$) whose Schr\" odinger equation (\ref{Schro-eq}) reduces to that of the simple harmonic oscillator.
By the coordinate transformations $x=\sqrt{(n-1)^2\hbar \kappa_n^2/4(n-2)V_{n-2}^{(0)}}{\bar x}$, the Schr\" odinger equation (\ref{Schro-eq}) with $k=0$ is cast into the following standard form:
\begin{align}
\biggl(-\frac12\frac{\partial^2}{\partial{\bar x}^2}+\frac12\omega^2{\bar x}^2\biggl)\Psi=&i\frac{\partial\Psi}{\partial t},\label{tosolve}
\end{align}
where
\begin{align}
\omega:=\frac{n-1}{2l}.\label{def-omega}
\end{align}
As anticipated, this equation is precisely that of a quantum harmonic oscillator on the half line ${\bar x}\in [0,\infty)$. In a previous paper~\cite{km2014}, we obtained the exact stationary-state mass spectrum (\ref{EN:toroidal}) for the Dirichlet or Neumann boundary condition at ${\bar x}=0$.
We now take advantage of the fact that for this system exact time-dependent solutions of the Schr\" odinger equation are also available.
\subsection{Time-dependent wave function on the whole line: A review}
\label{sec:harmonic-review}
There is a six-parameter family of exact solutions to the Schr\" odinger equation (\ref{tosolve})~\cite{lsv2013}~\footnote{We have done coordinate transformations and reparametrization from the original expressions in Ref.~\cite{lsv2013}.}:
\begin{align}
\Psi({\bar x},t)=&\Psi_N({\bar x},t) \nonumber \\
:=&\frac{e^{i\left(\alpha(t){\bar x}^2+\delta(t){\bar x}+\kappa(t)\right)+i(2N+1)\gamma(t)}}{\sqrt{2^NN!\mu(t)\sqrt{\pi}}} e^{-(\beta(t){\bar x}+\varepsilon(t))^2/2}H_N(\beta(t){\bar x}+\varepsilon(t)). \label{six-solution}
\end{align}
In the above, $H_N$ are the Hermite polynomials with integer $N(=0,1,2,\cdots)$ and
\begin{align}
\mu(t)=&\mu_0\sqrt{{\bar\beta}_0^4\sin^2\omega t/\omega ^2+(2\alpha_0\sin \omega t+\cos \omega t)^2}, \label{six-sol1}\\
\alpha(t)=&\frac{\omega \alpha_0\cos 2\omega t+(\sin 2\omega t/\omega )({\bar\beta}_0^4+4\omega ^2\alpha_0^2-\omega ^2)/4}{{\bar\beta}_0^4\sin^2\omega t/\omega ^2+(2\alpha_0\sin \omega t+\cos \omega t)^2},\\
\beta(t)=&\frac{{\bar\beta}_0}{\sqrt{{\bar\beta}_0^4\sin^2\omega t/\omega ^2+(2\alpha_0\sin \omega t+\cos \omega t)^2}},\\
\gamma(t)=&\gamma_0-\frac12\arctan\biggl(\frac{{\bar\beta}_0^2\sin \omega t/\omega }{2\alpha_0\sin \omega t+\cos \omega t}\biggl),\\
\delta(t)=&\frac{{\bar\delta}_0(2\alpha_0\sin \omega t+\cos \omega t)+\varepsilon_0{\bar\beta}_0^3\sin \omega t/\omega }{{\bar\beta}_0^4\sin^2\omega t/\omega ^2+(2\alpha_0\sin \omega t+\cos \omega t)^2},\\
\varepsilon(t)=&\frac{\varepsilon_0(2\alpha_0\sin \omega t+\cos \omega t)-{\bar\beta}_0{\bar\delta}_0\sin \omega t/\omega }{\sqrt{{\bar\beta}_0^4\sin^2\omega t/\omega ^2+(2\alpha_0\sin \omega t+\cos \omega t)^2}},\\
\kappa(t)=&\kappa_0+\frac{\sin^2 \omega t}{\omega ^2}\frac{\varepsilon_0{\bar\beta}_0^2(\omega \alpha_0\varepsilon_0-{\bar\beta}_0{\bar\delta}_0)-\omega \alpha_0{\bar\delta}_0^2}{{\bar\beta}_0^4\sin^2\omega t/\omega ^2+(2\alpha_0\sin \omega t+\cos \omega t)^2} \nonumber \\
&+\frac14\frac{\sin 2\omega t}{\omega }\frac{\varepsilon_0^2{\bar\beta}_0^2-{\bar\delta}_0^2}{{\bar\beta}_0^4\sin^2\omega t/\omega ^2+(2\alpha_0\sin \omega t+\cos \omega t)^2},\label{six-sol2}
\end{align}
where $\mu_0,{\alpha}_0,{\bar \beta}_0,\gamma_0,{\bar\delta}_0,\kappa_0,\varepsilon_0$ are constants.
Among them, $\gamma_0$ and $\kappa_0$ are phase constants.
In order to verify that the solution (\ref{six-solution}) solves the Schr\" odinger equation (\ref{tosolve}), the following relations are useful:
\begin{align}
& {\dot \mu}=2\mu \alpha,\quad {\dot \alpha}=\frac12(\beta^4-4\alpha^2-\omega^2),\label{t-derivative1} \\
& {\dot \beta}=-2\beta\alpha,\quad {\dot \gamma}=-\frac12\beta^2,\quad {\dot \delta}=\beta^3\varepsilon-2\alpha\delta, \label{t-derivative4}\\
& {\dot \varepsilon}=-\beta \delta,\quad {\dot \kappa}=\frac12(\beta^2\varepsilon^2-\delta^2),\label{t-derivative7}
\end{align}
where a dot denotes differentiation with respect to $t$.
Using the orthogonality condition (\ref{Hermite-orthogonal}) of the Hermite polynomials, we compute the squared norm of $\Psi_N$ on the whole line:
\begin{align}
\langle \Psi_N|\Psi_N \rangle=\int_{-\infty}^\infty|\Psi_N({\bar x},t)|^2{\rm d} {\bar x}=\frac{1}{\mu(t) \beta(t)}=\frac{1}{\mu_0{\bar \beta}_0}.
\end{align}
Since normalization requires $\mu_0{\bar\beta}_0=1$, the number of independent parameters is six, of which two ($\kappa_0$ and $\gamma_0$) are pure phase. The physical properties of the solution are therefore characterized by four continuous parameters ${\alpha}_0$, ${\bar \beta}_0$, ${\bar\delta}_0$, and $\varepsilon_0$ in addition to the quantum number $N$.
Within this class of solutions, stationary states are realized as special cases for which ${\alpha}_0={\bar\delta}_0=\varepsilon_0=0$ and ${\bar \beta}_0=\sqrt{\omega}$. In this case we have
\begin{align}
\mu(t)=&\mu_0,\quad \alpha(t)=0,\quad \beta(t)=\sqrt{\omega},\\
\gamma(t)=&\gamma_0-\frac12\omega t, \quad \delta(t)=\varepsilon(t)=0,\quad \kappa(t)=\kappa_0.
\end{align}
We will use this time-dependent solution (\ref{six-solution}) to construct exact wave functions on the half line, but first we will review some properties of the solution on the whole line.
The details of computations are presented in Appendix~\ref{app:wholeline}.
\subsubsection{Energy expectation value and uncertainty relation}
The energy expectation value is given by
\begin{align}
\langle E \rangle_N=&\frac{\langle \Psi_N|{\hat H}\Psi_N \rangle}{\langle \Psi_N|\Psi_N \rangle}=\biggl(N+\frac12\biggl)\hbar\Omega+\frac12\hbar\eta, \label{Energy-wholeline}
\end{align}
where ${\hat H}=i\hbar\partial/\partial t$ is the Hamiltonian operator and
\begin{align}
\Omega:=&\frac{{\bar\beta}_0^4+4\alpha_0^2\omega^2+\omega^2}{2{\bar\beta}_0^2}, \label{Omega-def}\\
\eta:=&\frac{(2\alpha_0\varepsilon_0\omega-{\bar\beta}_0{\bar \delta}_0)^2+\varepsilon_0^2\omega^2}{{\bar\beta}_0^{2}}.\label{eta-def}
\end{align}
Hence, the expectation value of the energy is time independent and equally spaced, with $\Delta \langle E\rangle_N:=\langle E\rangle_{N-1}-\langle E\rangle_N=\hbar\Omega$.
On the other hand, the fluctuation of the energy $\delta \langle E\rangle_N:=\sqrt{\langle E^2\rangle_N-\langle E\rangle_N^2}$ is in general given by a complicated expression that is time dependent.
However, there is a subclass of states for which both the energy expectation value and the energy fluctuation are constant. These are the so-called ``shape-preserving states'', which as the name suggests are time-dependent states whose probability amplitudes move without deformation. They include as a special case the usual coherent state of the simple harmonic oscillator and will be discussed in full in Sec.~\ref{sec:exact2}.
It turns out that the energy expectation value (\ref{Energy-wholeline}) can in general be written as
\begin{align}
\langle E \rangle_N=&\biggl(N+\frac12\biggl)\hbar\Omega+\frac{\hbar}{2}\biggl(\omega^2\langle {\bar x}\rangle_N^2+\hbar^{-2}\langle p\rangle_N^2\biggl),
\end{align}
where $\langle {\bar x}\rangle_N$ and $\langle p\rangle_N$ are expectation values of the position ${\bar x}$ and the momentum ${\hat p}=-i\hbar\partial/\partial{\bar x}$, respectively, given by
\begin{align}
\langle {\bar x}\rangle_N=&-\frac{\varepsilon}{\beta} \nonumber \\
=&\frac{1}{\omega}\biggl\{\biggl({\bar \delta}_0-\frac{2\alpha_0\varepsilon_0\omega}{{\bar \beta}_0}\biggl)\sin\omega t-\frac{\varepsilon_0\omega}{{\bar \beta}_0}\cos\omega t\biggl\},\label{exp-x-full}\\
\langle p\rangle_N=&\hbar\biggl(\delta-\frac{2\alpha\varepsilon}{\beta}\biggl) \nonumber \\
=&\hbar\biggl\{\biggl({\bar \delta}_0-\frac{2\alpha_0\varepsilon_0\omega}{{\bar \beta}_0}\biggl)\cos\omega t+\frac{\varepsilon_0\omega}{{\bar \beta}_0}\sin\omega t\biggl\}.\label{exp-p-full}
\end{align}
Thus, the constant $\eta$ represents the sum of the kinetic and potential energies.
The fluctuations of the position and momentum are given by
\begin{align}
\delta \langle {\bar x}\rangle_N:=&\sqrt{\langle {\bar x}^2\rangle_N-\langle {\bar x}\rangle_N^2}=\sqrt{\frac{2N+1}{2\beta^2}} \nonumber \\
=&\sqrt{\frac{(2N+1)}{2{\bar\beta}_0^2}\biggl\{\frac{{\bar\beta}_0^4}{\omega^2}\sin^2\omega t+(2\alpha_0\sin \omega t+\cos \omega t)^2\biggl\}},\\
\delta \langle p\rangle_N:=&\sqrt{\langle p^2\rangle_N-\langle p\rangle_N^2}=\sqrt{\frac{\hbar^2(2N+1)}{2\beta^2}(4\alpha^2+\beta^{4})} \nonumber \\
=&\sqrt{\frac{\hbar^2(2N+1)}{2{\bar\beta}_0^2}\biggl\{{\bar\beta}_0^4\cos^2\omega t+\omega^2(\sin\omega t-2\alpha_0\cos\omega t)^2\biggl\}}.
\end{align}
Hence, the uncertainty relation is given by
\begin{align}
&\delta \langle {\bar x}\rangle_N\delta \langle p\rangle_N \nonumber \\
&~~=\frac{\hbar(2N+1)}{4{\bar\beta}_0^2\omega}\biggl\{(4\alpha_0^2\omega^2+{\bar\beta}_0^4+\omega^2)^2 -\biggl((4\alpha_0^2\omega^2+{\bar\beta}_0^4-\omega^2)\cos2\omega t-4\alpha_0\omega^2\sin2\omega t\biggl)^2\biggl\}^{1/2}, \label{uncertainty-full}
\end{align}
where the expression inside the square root is positive definite.
\subsubsection{Shape-preserving state}
\label{sec:ShapePreserving}
The shape-preserving states are realized for $\alpha_0=0$ and ${\bar\beta}_0=\sqrt{\omega}$, in which case one has:
\begin{align}
\mu(t)=&\mu_0,\quad \alpha(t)=0,\quad \beta(t)=\sqrt{\omega},\quad \gamma(t)=\gamma_0-\frac12\omega t,
\label{eq:SPmu}\\
\delta(t)=&{\bar\delta}_0\cos \omega t+\varepsilon_0\sqrt{\omega}\sin \omega t=\sqrt{\omega\zeta}\sin (\omega t+\theta_1),\label{eq:SPdelta}\\
\varepsilon(t)=&\varepsilon_0\cos \omega t-{\bar\delta}_0\sin \omega t/\sqrt{\omega}=\sqrt{\zeta}\sin (\omega t+\theta_2),\label{eq:SPepsilon}\\
\kappa(t)=&\kappa_0-\frac{\varepsilon_0{\bar\delta}_0\sin^2 \omega t}{\sqrt{\omega}} +\frac{(\varepsilon_0^2\omega-{\bar\delta}_0^2)\sin 2\omega t}{4\omega },
\label{eq:SPkappa}
\end{align}
where $\theta_1:=\arctan({\bar\delta}_0/\varepsilon_0\sqrt{\omega})$, $\theta_2:=\arctan(-\varepsilon_0\sqrt{\omega}/{\bar\delta}_0)$, and
\begin{align}
\zeta:=\varepsilon^2+\frac{\delta^2}{\beta^2}=\varepsilon_0^2+\frac{{\bar\delta}_0^2}{\omega}. \label{zeta-def}
\end{align}
$\Omega=\omega$ and $\eta=\omega\zeta$ hold in this shape-preserving state and $\langle E\rangle_N$, $\langle {\bar x}\rangle_N$, and $\langle p\rangle_N$ reduce to
\begin{align}
\langle E\rangle_N=&\biggl(N+\frac12\biggl)\hbar\omega+\frac12\hbar\omega\zeta,\label{energy-shapepreserving}\\
\langle {\bar x}\rangle_N=&-\frac{\varepsilon}{\sqrt{\omega}},\quad \langle p\rangle_N=\hbar\delta.
\end{align}
A characteristic property of the shape-preserving states is that the uncertainty relation (\ref{uncertainty-full}) becomes constant:
\begin{align}
\delta \langle {\bar x}\rangle_N\delta \langle p\rangle_N=\hbar\biggl(N+\frac12\biggl).
\end{align}
For shape-preserving states, the relative fluctuation of $\langle {\bar x}\rangle_N$ is given by
\begin{align}
\frac{\delta \langle {\bar x}\rangle_N}{\langle {\bar x}\rangle_N}=&\frac{\sqrt{\langle {\bar x}^2\rangle_N-\langle {\bar x}\rangle_N^2}}{\langle {\bar x}\rangle_N}=-\frac{1}{\varepsilon}\sqrt{\frac{2N+1}{2}}.
\end{align}
The absolute value of this quantity diverges for large $N$ as
\begin{align}
\lim_{N\to \infty}\biggl|\frac{\delta \langle {\bar x}\rangle_N}{\langle {\bar x}\rangle_N}\biggl|\simeq |\varepsilon|^{-1}N^{1/2}\to \infty.
\end{align}
Because the range of $\varepsilon(t)$ is $-\sqrt{\zeta}\le \varepsilon\le \sqrt{\zeta}$, we obtain
\begin{align}
\min_{t\in\mathbb{R}}\biggl|\frac{\delta \langle {\bar x}\rangle_N}{\langle {\bar x}\rangle_N}\biggl|=\sqrt{\frac{2N+1}{2\zeta}},\qquad \max_{t\in\mathbb{R}}\biggl|\frac{\delta \langle {\bar x}\rangle_N}{\langle {\bar x}\rangle_N}\biggl|=\infty
\end{align}
for a given value of $N$, where the minimum and maximum are realized when $|\langle {\bar x}\rangle_N|=\max_{t\in\mathbb{R}}|\langle {\bar x}\rangle_N|=\sqrt{\zeta/\omega}$ and $\langle {\bar x}\rangle_N=0$ hold, respectively.
The minimum of the relative fluctuation reduces to zero for $\zeta\to \infty$ as
\begin{align}
\lim_{\zeta\to \infty}\min_{t\in\mathbb{R}}\biggl|\frac{\delta \langle {\bar x}\rangle_N}{\langle {\bar x}\rangle_N}\biggl|\simeq \sqrt{\frac{N}{\zeta}}\to 0.
\end{align}
Actually, the limit $\zeta\to \infty$ is unphysical since it gives infinite energy; however, the minimum of the relative fluctuation becomes very small in the situation of $\zeta\gg N$.
Equation~(\ref{energy-shapepreserving}) shows that, for $\zeta\gg N$, the energy of the shape-preserving state is much larger than the energy in the stationary state.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.7\linewidth]{deltax2coherent}
\caption{\label{deltax2-coherent} $\delta \langle {\bar x}^2\rangle_N/\langle {\bar x}^2\rangle_N$ as a function of $\varepsilon^2$ for the shape-preserving state with $N=0,1,\cdots,5$.}
\end{center}
\end{figure}
For later reference, we also present the relative fluctuation of $\langle {\bar x}^2\rangle_N$:
\begin{align}
\frac{\delta \langle {\bar x}^2\rangle_N}{\langle {\bar x}^2\rangle_N}=&\frac{\sqrt{\langle {\bar x}^4\rangle_N-\langle {\bar x}^2\rangle_N^2}}{\langle {\bar x}^2\rangle_N} \nonumber \\
=&\frac{\sqrt{2(N^2+N+1)+8(2N+1)\varepsilon^2}}{(2N+1)+2\varepsilon^2}.\label{deltax2-shape}
\end{align}
This quantity has a finite limit for $N\to \infty$ as
\begin{align}
\lim_{N\to \infty}\frac{\delta \langle {\bar x}^2\rangle_N}{\langle {\bar x}^2\rangle_N}=\frac{1}{\sqrt{2}}.
\end{align}
Figure~\ref{deltax2-coherent} shows the shape of $\delta \langle {\bar x}^2\rangle_N/\langle {\bar x}^2\rangle_N$ as a function of $\varepsilon^2$ for $N=0,1,\cdots,5$.
It has a maximum at $\varepsilon^2=3N(N+1)/2(2N+1)$.
The relative fluctuation at $\varepsilon=\pm\sqrt{\zeta}$ is given by
\begin{align}
\frac{\delta \langle {\bar x}^2\rangle_N}{\langle {\bar x}^2\rangle_N}\biggl|_{\varepsilon^2=\zeta}=&\frac{\sqrt{2(N^2+N+1)+8(2N+1)\zeta}}{(2N+1)+2\zeta},
\end{align}
which reduces to zero for $\zeta\to \infty$ as
\begin{align}
\lim_{\zeta\to \infty}\frac{\delta \langle {\bar x}^2\rangle_N}{\langle {\bar x}^2\rangle_N}\biggl|_{t=t_{\rm max}}\simeq &\frac{\sqrt{2(2N+1)}}{\zeta^{1/2}}\to 0.
\end{align}
\subsection{Time-dependent wave function on the half line}
The wave function (\ref{six-solution}) is an exact solution of the Schr\" odinger equation (\ref{tosolve}) on the whole line ${\bar x}\in(-\infty,\infty)$.
By symmetrizing or anti-symmetrizing this solution about ${\bar x}=0$, one can construct exact wave functions on the half line ${\bar x}\in[0,\infty)$ satisfying Neumann or Dirichlet boundary conditions, respectively, at the origin.
This yields
\begin{align}
\Psi^{(\pm)}_{N}({\bar x},t):=&\frac12\biggl(\Psi_N({\bar x},t)\pm \Psi_N(-{\bar x},t)\biggl) \nonumber \\
=&\frac{e^{i(\alpha{\bar x}^2+\kappa)+i(2N+1)\gamma}}{2\sqrt{2^NN!\mu\sqrt{\pi}}}\biggl\{e^{i\delta{\bar x}} e^{-(\beta{\bar x}+\varepsilon)^2/2}H_N(\beta{\bar x}+\varepsilon)\pm e^{-i\delta{\bar x}}e^{-(\beta{\bar x}-\varepsilon)^2/2}H_N(-\beta{\bar x}+\varepsilon)\biggl\}, \label{SolHalfLine}
\end{align}
which satisfies Neumann (Dirichlet) boundary conditions at ${\bar x}=0$ for the upper (lower) sign.
The square of the wave function is an even function given by
\begin{align}
|\Psi^{(\pm)}_{N}|^2=&\frac{1}{2^{N+2}N!\mu\sqrt{\pi}}\biggl\{e^{-(\beta{\bar x}+\varepsilon)^2}H_N(\beta{\bar x}+\varepsilon)^2 \nonumber \\
&+e^{-(\beta{\bar x}-\varepsilon)^2}H_N(-\beta{\bar x}+\varepsilon)^2\pm 2\cos(2\delta{\bar x}) e^{-\beta^2{\bar x}^2-\varepsilon^2}H_N(\beta{\bar x}+\varepsilon)H_N(-\beta{\bar x}+\varepsilon) \biggl\}.
\label{SquareNormGeneral}
\end{align}
While the integral of the first two terms are easily evaluated to yield
\begin{align}
&\int_{0}^\infty e^{-(\beta{\bar x}+\varepsilon)^2}H_N(\beta{\bar x}+\varepsilon)^2 {\rm d} {\bar x} \nonumber \\
=&\int_{0}^\infty e^{-(\beta{\bar x}-\varepsilon)^2}H_N(-\beta{\bar x}+\varepsilon)^2 {\rm d} {\bar x}=\frac{\sqrt{\pi}2^{N-1}N!}{\beta},
\end{align}
it is difficult to compute the last term for general $N$.
Nevertheless, using the following integral and its derivatives with respect to $q$,
\begin{align}
\int_{0}^\infty \cos (2qy) e^{-y^2}{\rm d} y=&\frac12\int_{-\infty}^\infty \cos (2qy)e^{-y^2}{\rm d} y \nonumber \\
=&\frac{\sqrt{\pi}}{2}e^{-q^2},
\end{align}
one can compute the squared norm for each value of $N$.
Equation~(\ref{SolHalfLine}) provides two families of dynamical solutions to the harmonic oscillator Schr\" odinger equation on the half line, one satisfying Neumann boundary conditions at the origin, the other satisfying Dirichlet boundary conditions. Each family retains all the parameters of the solutions on the whole line. For the reasons given before, the physical parameters are $\varepsilon_0$, ${\bar \delta}_0$, $\alpha_0$ , ${\bar \beta}_0$ and the integer $N$. We have omitted $\mu_0$ since it is determined in terms of $\varepsilon_0$ and ${\bar \beta}_0$ via the normalization condition (cf. Eq.~(\ref{SquareNormGeneral})). We emphasize that the two families live in distinct Hilbert spaces in which the Hamiltonian is self-adjoint.
Given that there is such a large parameter space of solutions, it is difficult to analyze the dynamical behaviour in full generality. In the following, we will therefore focus on two different subclasses of solutions of particular interest. The subclasses retain two distinct pairs of parameters and are characterized by different dynamical behaviour.
\subsubsection{Wave function I}
The first class of wave functions on the half line is obtained from (\ref{SolHalfLine}) by setting ${\bar\delta}_0=\varepsilon_0=0$ (and hence $\delta(t)=\varepsilon(t)=0$). It is interesting to note that making this substitution directly into the wave function (\ref{six-solution}) also yields Neumann or Dirichlet boundary conditions, without the need to (anti-)symmetrize first~\footnote{If one chooses to (anti-)symmetrize first, then the anti-symmetric solutions vanish identically for even $N$, while the symmetric solutions vanish identically for odd $N$. The remaining solutions are precisely those given above.}. One obtains
\begin{align}
\Psi_{{\rm I}(N)}({\bar x},t):=&\frac{e^{i(\alpha(t){\bar x}^2+\kappa(t))+i(2N+1)\gamma(t)}}{\sqrt{2^NN!\mu(t)\sqrt{\pi}}}e^{-\beta(t)^2{\bar x}^2/2}H_N(\beta(t){\bar x}), \label{sol-first}
\end{align}
where
\begin{align}
\mu(t)=&\mu_0\sqrt{{\bar\beta}_0^4\sin^2\omega t/\omega ^2+(2\alpha_0\sin \omega t+\cos \omega t)^2}, \\
\alpha(t)=&\frac{\omega \alpha_0\cos 2\omega t+(\sin 2\omega t/\omega )({\bar\beta}_0^4+4\omega ^2\alpha_0^2-\omega ^2)/4}{{\bar\beta}_0^4\sin^2\omega t/\omega ^2+(2\alpha_0\sin \omega t+\cos \omega t)^2},\\
\beta(t)=&\frac{{\bar\beta}_0}{\sqrt{{\bar\beta}_0^4\sin^2\omega t/\omega ^2+(2\alpha_0\sin \omega t+\cos \omega t)^2}},\\
\gamma(t)=&\gamma_0-\frac12\arctan\biggl(\frac{{\bar\beta}_0^2\sin \omega t/\omega }{2\alpha_0\sin \omega t+\cos \omega t}\biggl),\\
\kappa(t)=&\kappa_0.
\end{align}
Neumann (Dirichlet) boundary conditions are satisfied for even (odd) $N$. In this case, the normalization can be carried out explicitly:
\begin{align}
\langle \Psi_{{\rm I}(N)}|\Psi_{{\rm I}(N)}\rangle=&\int_{0}^\infty|\Psi_{{\rm I}(N)}({\bar x},t)|^2{\rm d} {\bar x} \nonumber \\
=&\frac{1}{2\mu(t)\beta(t)}=\frac{1}{2\mu_0{\bar \beta}_0}. \label{s-norm-half}
\end{align}
As claimed previously, this class of solutions has two independent continuous physical parameters $\alpha_0$ and ${\bar \beta}_0$ in addition to the quantum number $N$. The parameters $\gamma_0$ and $\kappa_0$ are, as before, pure phase, while the parameter $\mu_0$ is determined by normalization of $\Psi_{{\rm I}(N)}$.
Stationary states on the half line are realized for all $N$ when $\alpha_0=0$ and ${\bar\beta}_0=\sqrt{\omega}$.
\subsubsection{Wave function II}
The second class of exact wave functions $\Psi=\Psi^{(\pm)}_{{\rm II}(N)}({\bar x},t)$ on the half line is given by setting $\alpha_0=0$ and $\beta_0=\sqrt{\omega}$ into the (anti-)symmetrized wave functions (\ref{SolHalfLine}). One thereby obtains subclasses of (anti-)symmetric states $\Psi^{(\pm)}_{{\rm II}(N)}$. These are the analogues from the state
on the half line of the shape-preserving solutions described in Sec.~\ref{sec:ShapePreserving}. They can also be obtained directly by (anti-)symmetrizing the shape-preserving solutions on the whole line.
The wave function $\Psi^{(\pm)}_{{\rm II}(N)}({\bar x},t)$ is given by Eq.~(\ref{SolHalfLine}) with the functions (\ref{eq:SPmu})--(\ref{eq:SPkappa}). It is useful to exhibit the first six squared norms:
\begin{align}
\int_0^\infty|\Psi^{(\pm)}_{{\rm II}(0)}|^2{\rm d} {\bar x}=&\frac{1}{4\mu\beta}\biggl\{1\pm e^{-\zeta}\biggl\},\\
\int_0^\infty|\Psi^{(\pm)}_{{\rm II}(1)}|^2{\rm d} {\bar x}=&\frac{1}{4\mu\beta}\biggl\{1\pm \biggl(2\zeta-1\biggl)e^{-\zeta}\biggl\},\\
\int_0^\infty|\Psi^{(\pm)}_{{\rm II}(2)}|^2{\rm d} {\bar x}=&\frac{1}{4\mu\beta}\biggl\{1\pm \biggl(2\zeta^2-4\zeta+1\biggl)e^{-\zeta}\biggl\},\\
\int_0^\infty|\Psi^{(\pm)}_{{\rm II}(3)}|^2{\rm d} {\bar x}=&\frac{1}{12\mu\beta}\biggl\{3\pm \biggl(4\zeta^3-18\zeta^2+18\zeta-3\biggl)e^{-\zeta}\biggl\},\\
\int_0^\infty|\Psi^{(\pm)}_{{\rm II}(4)}|^2{\rm d} {\bar x}=&\frac{1}{12\mu\beta}\biggl\{3\pm \biggl(2\zeta^4-16\zeta^3+36\zeta^2-24\zeta+3\biggl)e^{-\zeta}\biggl\},\\
\int_0^\infty|\Psi^{(\pm)}_{{\rm II}(5)}|^2{\rm d} {\bar x}=&\frac{1}{60\mu\beta}\biggl\{15\pm \biggl(4\zeta^5-50\zeta^4+200\zeta^3-300\zeta^2+150\zeta-15\biggl)e^{-\zeta}\biggl\},
\end{align}
where
\begin{align}
\zeta:=\varepsilon(t)^2+\frac{\delta(t)^2}{\beta(t)^2}=\varepsilon_0^2+\frac{{\bar\delta}_0^2}{{\bar\beta}_0^2}.
\label{eq:zeta}
\end{align}
It is interesting to note that the physical properties of these solutions depend only on the time-independent combination of $\varepsilon(t)$ and $\delta(t)$ given by $\zeta$, which measures the difference from the stationary state (${\bar \delta}_0=0$ and ${\varepsilon}_0=0$) or a vanishing wave function $\Psi^{(\pm)}_{{\rm II}(N)}\equiv 0$.
The latter is realized for $\zeta=0$ in the case of the upper sign with odd $N$ or the lower sign with even $N$.
The evolutions of the (unnormalized) squared norm $|\Psi^{(+)}_{{\rm II}(N)}|^2$ and $|\Psi^{(-)}_{{\rm II}(N)}|^2$ with $N=0,1,2$ are drawn in Figs.~\ref{fig-wave2(all-)} and \ref{fig-wave2(all+)}, respectively.
They show that, while the evolution is rather chaotic for $\zeta=1$, the wave packet is shape preserving when it is away from the origin for $\zeta=40$.
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=1.0\linewidth]{wave2minusall}
\caption{\label{fig-wave2(all-)} Three snapshots during the evolution of the unnormalized squared norm $|\Psi^{(+)}_{{\rm II}(N)}({\bar x},t)|^2$ with $\mu_0=1$, $\omega=1/100$, and ${\bar\delta}_0=0$.
The left row corresponds to $\varepsilon_0^2=1$ ($\zeta=1$) with (a1) $N=0$, (a2) $N=1$, and (a3) $N=2$.
The right row corresponds to $\varepsilon_0^2=40$ ($\zeta=40$) with (b1) $N=0$, (b2) $N=1$, and (b3) $N=2$.
}
\end{center}
\end{figure*}
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=1.0\linewidth]{wave2plusall}
\caption{\label{fig-wave2(all+)} Three snapshots during the evolution of the unnormalized squared norm $|\Psi^{(-)}_{{\rm II}(N)}({\bar x},t)|^2$ with $\mu_0=1$, $\omega=1/100$, and ${\bar\delta}_0=0$.
The left row corresponds to $\varepsilon_0^2=1$ ($\zeta=1$) with (a1) $N=0$, (a2) $N=1$, and (a3) $N=2$.
The right row corresponds to $\varepsilon_0^2=40$ ($\zeta=40$) with (b1) $N=0$, (b2) $N=1$, and (b3) $N=2$.
}
\end{center}
\end{figure*}
\section{Quantum toroidal black hole: Wave function I}
\label{sec:exact1}
In this section, we will study quantum toroidal black holes described by the wave function (\ref{sol-first}). Hereafter we assume ${\bar \beta}_0\ge 0$ (hence $\beta(t)\ge 0$) without loss of generality.
The details of computations are presented in Appendix~\ref{app:halfline}.
\subsection{Dynamical singularity resolution}
The solution (\ref{sol-first}) is a function in the Hilbert space ${\cal L}^2([0,\infty))$ with the Dirichlet or Neumann boundary condition at the origin and nonsingular for $t\in (-\infty,\infty)$.
Hence the classical initial singularity at $r=0$ (hence at $x={\bar x}=0$) is avoided in the quantum system.
The oscillatory evolution of the normalized squared norm $|\Psi_{{\rm I}(N)}({\bar x},t)|^2$ is exhibited in Fig.~\ref{fig-wave1(all)} for several values of the quantum number $N$.
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=1.0\linewidth]{wave1minusall}
\caption{\label{fig-wave1(all)} The evolution of the normalized squared norm $|\Psi_{{\rm I}(N)}({\bar x},t)|^2$ with $\alpha_0={\bar\beta}_0=1$ for (a) $N=0$, (b) $N=1$, (c) $N=2$, and (d) $N=3$.
$t_0$ is the time when $|\Psi_{{\rm I}(N)}({\bar x},t)|^2$ has the maximum value in the domain, and $T$ is the period of oscillation.
}
\end{center}
\end{figure*}
We now examine the time evolution of the expectation value of the Euclidean volume constructed from the areal radius $r(=a)$, given by $V_{\rm E}=V_{n-2}^{(0)}x^2/(n-1)$.
Classically, along the orbit of the wormhole throat in the maximally extended spacetime, $V_{\rm E}(t)$ starts from zero at the white-hole singularity at $x=0$, reaches the maximum value at the bifurcation $(n-2)$-surface corresponding to the event horizon, and then turns to decrease toward $x=0$ at the black-hole singularity.
The volume expectation value is given by
\begin{align}
\langle V_{\rm E}\rangle_N(t)=&\frac{(2N+1)(n-1)\hbar \kappa_n^2}{8(n-2)\beta(t)^2} \nonumber \\
=&V_{\rm p}\frac{(2N+1)(n-1)l}{4(n-2)V_{n-2}^{(0)}\ell_{\rm p}}\frac{\Omega}{\omega}\biggl(1+\sqrt{1-\frac{\omega^2}{\Omega^2}}\sin (2\omega t+\theta_0)\biggl),\label{V_E-0}
\end{align}
where $\Omega$ is defined by Eq.~(\ref{Omega-def}) and we use the following expression:
\begin{align}
\frac{1}{\beta(t)^2}=&\frac{\Omega}{\omega^2}\biggl(1+\sqrt{1-\frac{\omega^2}{\Omega^2}}\sin (2\omega t+\theta_0)\biggl),\label{beta-inverse} \\
\theta_0:=&\arctan\biggl\{\frac{1}{2\alpha_0}\biggl(1-\frac{{\bar\beta}_0^2\Omega}{\omega^2}\biggl)\biggl\}.
\end{align}
(See Appendix~\ref{app:area} for derivation of $\langle V_{\rm E}\rangle_N$.)
This shows that $\langle V_{\rm E}\rangle_N$ is positive definite and oscillating, and hence the classical spacelike singularity at $r=0$ is resolved quantum mechanically and replaced by a big-bounce.
The classical description of the regularized black hole is presented in Fig.~\ref{SingleBH-bounce}.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.3\linewidth]{SinglehorizonBHAdSbounce}
\caption{\label{SingleBH-bounce} A portion of the Penrose diagram for the maximally extended spacetime of the regularized toroidal black hole. A double line corresponds to AdS infinity, and $i^+$ and $i^-$ are future and past timelike infinities, respectively. Dashed lines represent big bounces which replace the spacelike curvature singularities in classical general relativity.}
\end{center}
\end{figure}
Note that the combination of the parameters $\Omega$ defined by Eq.~(\ref{Omega-def}) provides a measure of the difference between the above states and the corresponding stationary state with the same $N$.
The expression
\begin{align}
\frac{\Omega}{\omega}=\frac{{\bar \beta}_0^4+4\alpha_0^2\omega^2+\omega^2}{2{\bar \beta}_0^2\omega}
\end{align}
shows that $\Omega\ge \omega$ holds, with equality holding only in the stationary case ($\alpha_0=0$ and ${\bar\beta}_0^2=\omega$).
On the other hand, the high-amplitude limit $\Omega/\omega\to \infty$ is realized in two independent limits: ${\bar\beta}_0^2/\omega\to\infty$ or $\alpha_0^2\omega/{\bar \beta}_0^2\to \infty$.
In the next subsection, we will see that this high-amplitude limit $\Omega/\omega\to \infty$ corresponds to infinite mass spacing, and therefore the limit is unphysical.
The limit $\Omega/\omega\to \infty$ should therefore be interpreted as $\Omega\gg \omega$, as done in the following.
In terms of the Planck length $\ell_{\rm p}$ and the Planck volume $V_{\rm p}$, the maximum and minimum values of $\langle V_{\rm E}\rangle_N(t)$ are given by
\begin{align}
\max_{t\in\mathbb{R}}\langle V_{\rm E}\rangle_N(t)=&\frac{(2N+1)(n-1)l}{4(n-2)V_{n-2}^{(0)}\ell_{\rm p}}\frac{\Omega}{\omega}\biggl(1+\sqrt{1-\frac{\omega^2}{\Omega^2}}\biggl)V_{\rm p},\\
\min_{t\in\mathbb{R}}\langle V_{\rm E}\rangle_N(t)=&\frac{(2N+1)(n-1)l}{4(n-2)V_{n-2}^{(0)}\ell_{\rm p}}\frac{\Omega}{\omega}\biggl(1-\sqrt{1-\frac{\omega^2}{\Omega^2}}\biggl)V_{\rm p},
\end{align}
where we use $\omega:=(n-1)/2l$.
In the stationary case ($\Omega=\omega$), $\langle V_{\rm E}\rangle_N$ is constant:
\begin{align}
\lim_{\Omega/\omega\to 1}\langle V_{\rm E}\rangle_N(t)=\frac{(2N+1)(n-1)l}{4(n-2)V_{n-2}^{(0)}\ell_{\rm p}}V_{\rm p}.
\end{align}
In the high-amplitude limit $\Omega/\omega\to \infty$, on the other hand, we obtain
\begin{align}
\lim_{\Omega/\omega\to \infty}\max_{t\in\mathbb{R}}\langle V_{\rm E}\rangle_N(t)\simeq& \frac{(2N+1)(n-1)l}{2(n-2)V_{n-2}^{(0)}\ell_{\rm p}}\frac{\Omega}{\omega}\to \infty,\\
\lim_{\Omega/\omega\to \infty}\min_{t\in\mathbb{R}}\langle V_{\rm E}\rangle_N(t)\simeq& \frac{(2N+1)(n-1)l}{8(n-2)V_{n-2}^{(0)}\ell_{\rm p}}\frac{\omega}{\Omega}\to 0.
\end{align}
\subsection{Mass of the quantum black hole}
The physical meaning of the parameter $\Omega$ is made explicit by the following expression for the expectation value of the mass of the black hole:
\begin{align}
\langle M\rangle_N:=&\frac{\langle \Psi_{{\rm I}(N)}|{\hat H}\Psi_{{\rm I}(N)}\rangle}{\langle \Psi_{{\rm I}(N)}|\Psi_{{\rm I}(N)}\rangle}=\biggl(N+\frac12\biggl)\hbar \Omega. \label{mass-spectrum-dynamical}
\end{align}
(See Appendix.~\ref{app:mass} for derivation.)
Equation~(\ref{mass-spectrum-dynamical}) shows that $\Omega$, defined by Eq.~(\ref{Omega-def}), defines the mass step (i.e. spacing) between the neighboring states.
Since we have $\Omega\ge \omega$, with equality holding only for the stationary state, the mass step in the non-stationary state is larger than in the stationary state.
The expression (\ref{mass-spectrum-dynamical}) also confirms that the strict high-amplitude limit $\Omega/\omega\to \infty$ is unphysical.
However, we will still consider the situation $\Omega\gg \omega$ corresponding to a large black hole.
Although both of the limits $\Omega/\omega\to \infty$ and $N\to \infty$ correspond to the large-mass limit of a black hole, they are independent.
The mass spacing $\Delta\langle M\rangle_N$ is given by
\begin{align}
\Delta\langle M\rangle_N:=\langle M\rangle_{N+1}-\langle M\rangle_N=\hbar \Omega
\end{align}
and hence $N\to\infty$ is a semi-classical limit in the sense that the relative mass spacing reduces to zero:
\begin{align}
\lim_{N\to\infty}\frac{\Delta\langle M\rangle_N}{\langle M\rangle_N}=&\lim_{N\to\infty}\biggl(N+\frac12\biggl)^{-1}=0.
\end{align}
On the other hand, the mass fluctuation $\delta \langle M\rangle_N:=\sqrt{\langle M^2\rangle_N-\langle M\rangle_N^2}$ satisfies
\begin{align}
\frac{\delta \langle M\rangle_N}{\langle M\rangle_N}=\sqrt{\frac{2(N^2+N+1)}{(2N+1)^2}\biggl(1-\frac{\omega^2}{\Omega^2}\biggl)}. \label{mass-uncertainty-dynamical}
\end{align}
(See Appendix~\ref{app:mass} for the derivation of $\delta \langle M\rangle_N$.)
Clearly, there is no mass uncertainty in the stationary case ($\Omega=\omega$).
In the non-stationary case, by contrast, the mass uncertainty is non-zero.
In the two independent large-mass limits $\Omega/\omega\to \infty$ and $N\to 0$, we obtain
\begin{align}
\lim_{\Omega/\omega\to \infty}\frac{\delta \langle M\rangle_N}{\langle M\rangle_N}=&\sqrt{\frac{2(N^2+N+1)}{(2N+1)^2}}\biggl(>\frac{1}{\sqrt{2}}\biggl),\\
\lim_{N\to\infty}\frac{\delta \langle M\rangle_N}{\langle M\rangle_N}=&\sqrt{\frac12\biggl(1-\frac{\omega^2}{\Omega^2}\biggl)}\biggl(<\frac{1}{\sqrt{2}}\biggl).
\end{align}
The relative mass fluctuations therefore are non-negligible even for large-mass quantum black holes with
$\Omega \gg \omega$.
On the other hand, for large black holes for which $N\to \infty$, the mass fluctuations can be small if the state is very close to the stationary state.
If $\delta \langle M\rangle_N/\hbar\Omega\ge 1$, the mass fluctuation is greater than the spacing of the mass expectation value between states.
Its physical implications are still not clear.
From the expression
\begin{align}
\frac{\delta \langle M\rangle_N}{\hbar\Omega}=\sqrt{\frac{N^2+N+1}{2}\biggl(1-\frac{\omega^2}{\Omega^2}\biggl)},
\end{align}
we see that $\delta \langle M\rangle_N/\hbar\Omega<1$ is satisfied for $N=0$, independent $\Omega$.
Also, the above expression shows that the condition $\delta \langle M\rangle_N/\hbar\Omega \ge 1$ is satisfied for $N\ge N_{\rm c}$, where the critical value $N_{\rm c}$ is a positive solution of the following algebraic equation:
\begin{align}
N_{\rm c}^2+N_{\rm c}+1=2\biggl(1-\frac{\omega^2}{\Omega^2}\biggl)^{-1}.
\end{align}
If $\Omega$ is close to $\omega$, then $\delta \langle M\rangle_N/\hbar\Omega<1$ is satisfied even for large $N$.
However, if the state is non-stationary, then $\delta \langle M\rangle_N/\hbar\Omega>1$ is satisfied only for sufficiently large $N$.
\subsection{Euclidean volume of the quantum black hole}
\label{app:volumeBH}
From Eqs.~(\ref{VC-def}) and (\ref{mass-spectrum-dynamical}), the Euclidean volume of the classical horizon $V_{\rm C}$ is given by
\begin{align}
V_{\rm C}=&V_{\rm p}\frac{l^2(2N+1)\hbar\Omega}{(n-2)V_{n-2}^{(0)}l_{\rm p}^2m_{\rm p}}. \label{Vh}
\end{align}
By Eqs.~(\ref{deltaVC-def}) and (\ref{mass-uncertainty-dynamical}), the uncertainty $\delta V_{\rm C}$ satisfies
\begin{align}
\frac{\delta V_{\rm C}}{V_{\rm C}}=\sqrt{\frac{2(N^2+N+1)}{(2N+1)^2}\biggl(1-\frac{\omega^2}{\Omega^2}\biggl)}. \label{deltaVh}
\end{align}
On the other hand, the Euclidean volume of the quantum horizon $V_{\rm Q}$ and its uncertainty $\delta V_{\rm Q}$ are defined by $V_{\rm Q}:=\langle V_{\rm E}\rangle_N(t_{\rm max})$ and $\delta V_{\rm Q}:=\delta \langle V_{\rm E}\rangle_N(t_{\rm max})$, respectively, where $t_{\rm max}$ is the time when $\langle V_{\rm E}\rangle_N$ has the maximum value.
Equation~(\ref{V_E-0}) can be written as
\begin{align}
\langle V_{\rm E}\rangle_N(t)=&V_{\rm p}\frac{l^2(2N+1)\hbar\Omega}{2(n-2)V_{n-2}^{(0)}\ell_{\rm p}^2m_{\rm p}} \biggl(1+\sqrt{1-\frac{\omega^2}{\Omega^2}}\sin (2\omega t+\theta_0)\biggl)
\end{align}
and its uncertainty is given by
\begin{align}
\delta \langle V_{\rm E}\rangle_N(t)=&\sqrt{\frac{2(N^2+N+1)}{(2N+1)^2}}\langle V_{\rm E}\rangle(t) \label{deltaV}
\end{align}
(See in Appendix~\ref{app:area} for derivation.)
These expressions show
\begin{align}
V_{\rm Q}=&\frac12V_{\rm C}\biggl(1+\sqrt{1-\frac{\omega^2}{\Omega^2}}\biggl), \label{VQ} \\
\delta V_{\rm Q}=&V_{\rm Q}\sqrt{\frac{2(N^2+N+1)}{(2N+1)^2}} \nonumber \\
=&\frac12V_{\rm C}\sqrt{\frac{2(N^2+N+1)}{(2N+1)^2}}\biggl(1+\sqrt{1-\frac{\omega^2}{\Omega^2}}\biggl). \label{deltaVQ}
\end{align}
Equations~(\ref{deltaVh}), (\ref{VQ}), and (\ref{deltaVQ}) imply the following relations:
\begin{align}
\frac{V_{\rm Q}}{V_{\rm C}}=&\frac12\biggl(1+\sqrt{1-\frac{\omega^2}{\Omega^2}}\biggl),\\
\frac{\delta V_{\rm C}}{V_{\rm C}}=&\frac{\delta V_{\rm Q}}{V_{\rm Q}}\sqrt{1-\frac{\omega^2}{\Omega^2}}, \label{V-relations}
\end{align}
which are independent of $N$.
$\delta V_{\rm C}/V_{\rm C}$ and $V_{\rm Q}/V_{\rm C}$ are drawn in Fig.~\ref{VQVC-deltaVC} as functions of $\omega/\Omega$ for $N=0,1,\cdots,5$.
We have $1/2\le V_{\rm Q}/V_{\rm C}<1$ and hence the quantum horizon is located inside the classical horizon.
The Euclidean volume of the quantum black hole is half that of the classical black hole in the stationary state ($\omega/\Omega=1$).
As $\omega/\Omega$ decreases, the ratio $V_{\rm Q}/V_{\rm C}$ monotonically increases and $V_{\rm Q}/V_{\rm C}=\delta V_{\rm Q}/\delta V_{\rm C}=1$ is satisfied, namely the quantum horizon coincides with the classical horizon, in the high-amplitude limit $\omega/\Omega\to 0$ (and hence for large black holes with $\Omega\gg \omega$).
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.7\linewidth]{VQVCdeltaVC}
\caption{\label{VQVC-deltaVC} Dependence of $V_{\rm Q}/V_{\rm C}$ (thick curve) and $\delta V_{\rm C}/V_{\rm C}$ for $N=0,1,\cdots,5$ (thin curves) on $\omega/\Omega$.
}
\end{center}
\end{figure}
Equation~(\ref{deltaVQ}) shows that $\delta V_{\rm Q}/V_{\rm Q}$ is independent of $\omega/\Omega$ and is a monotonically decreasing function of $N$ satisfying
\begin{align}
\frac{1}{\sqrt{2}}<\frac{\delta V_{\rm Q}}{V_{\rm Q}} \le \sqrt{2},
\end{align}
where the lower and upper bounds correspond to $N\to\infty$ and $N=0$, respectively.
Hence the relative volume fluctuation of the quantum horizon is not negligible even for large black holes described both by $N\to \infty$ and by $\Omega\gg\omega$.
This implies that the quantum black hole described by wave function I does not have a semi-classical limit.
On the other hand, Eq.~(\ref{deltaVh}) shows that $\delta V_{\rm C}/V_{\rm C}$ depends on $\omega/\Omega$ and is also a monotonically decreasing function of $N$ satisfying
\begin{align}
\sqrt{\frac{1}{2}\biggl(1-\frac{\omega^2}{\Omega^2}\biggl)}<\frac{\delta V_{\rm C}}{V_{\rm C}} \le \sqrt{2\biggl(1-\frac{\omega^2}{\Omega^2}\biggl)}.
\end{align}
As $\omega/\Omega$ becomes smaller, the properties of the classical horizon are more similar to the quantum horizon.
In contrast, the fluctuation of the classical horizon gets smaller and finally becomes zero in the stationary limit $\omega/\Omega\to 1$.
In closing this section, we present physical quantities in terms of the surface area for comparison.
By Eqs.~(\ref{def-Ac}), (\ref{beta-inverse}), and (\ref{mass-spectrum-dynamical}) and the results in Appendix~\ref{app:area}, we obtain
\begin{align}
\frac{A_{\rm Q}}{A_{\rm C}}=&\frac{\int_{0}^{\infty} y^{2(n-2)/(n-1)} e^{-y^2}H_N(y)^2 {\rm d} y}{2^{N-1}N!\sqrt{\pi}}\biggl\{\frac{1}{2N+1}\biggl(1+\sqrt{1-\frac{\omega^2}{\Omega^2}}\biggl)\biggl\}^{(n-2)/(n-1)}, \label{A-relations1}\\
\frac{\delta A_{\rm C}}{A_{\rm C}}=&\biggl\{\frac{2(N^2+N+1)}{(2N+1)^2}\biggl(1-\frac{\omega^2}{\Omega^2}\biggl)\biggl\}^{(n-2)/2(n-1)},\label{A-relations2}\\
\frac{\delta A_{\rm Q}}{A_{\rm Q}}=&\biggl(\frac{2^{N-1}N!\sqrt{\pi}\int_{0}^{\infty} y^{4(n-2)/(n-1)} e^{-y^2}H_N(y)^2 {\rm d} y}{(\int_{0}^{\infty} y^{2(n-2)/(n-1)} e^{-y^2}H_N(y)^2 {\rm d} y)^2}-1\biggl)^{1/2}. \label{A-relations3}
\end{align}
\begin{table}[htb]
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|c|} \hline
& $A_{\rm Q}/A_{\rm C}|_{\omega/\Omega=1}$ & $A_{\rm Q}/A_{\rm C}|_{\omega/\Omega=0}$ & $\delta A_{\rm C}/A_{\rm C}|_{\omega/\Omega=1}$ & $\delta A_{\rm C}/A_{\rm C}|_{\omega/\Omega=0}$ & $\delta A_{\rm Q}/A_{\rm Q}$ \\ \hline \hline
$n=3$ & 0.6366 & 0.9003 & 0 & 0.8409 & 0.4834 \\
$n=4$ & 0.5798 & 0.9204 & 0 & 0.7937 & 0.5709 \\
$n=5$ & 0.5564 & 0.9358 & 0 & 0.7711 & 0.6090 \\
$n=6$ & 0.5436 & 0.9465 & 0 & 0.7579 & 0.6304 \\
$n=7$ & 0.5356 & 0.9543 & 0 & 0.7492 & 0.6441 \\
$n=8$ & 0.5301 & 0.9602 & 0 & 0.7430 & 0.6536 \\
$n=9$ & 0.5260 & 0.9647 & 0 & 0.7384 & 0.6607 \\
$n=10$ & 0.5229 & 0.9683 & 0 & 0.7349 & 0.666 \\
$n\to\infty$ & 0.5 & 1 & 0 & 0.7071 & 0.7071 \\
\hline
\end{tabular}
\caption{Values of $A_{\rm Q}/A_{\rm C}$, $\delta A_{\rm C}/A_{\rm C}$, and $\delta A_{\rm Q}/A_{\rm Q}$ for $\Psi=\Psi_{{\rm I}(N)}$ with $N=100$ are shown for $n=3,4,\cdots,10$ and $n\to\infty$. The values of $\omega/\Omega$ in $A_{\rm Q}/A_{\rm C}$ and $\delta A_{\rm C}/A_{\rm C}$ are $\omega/\Omega=1$ and $0$. In the limit $n\to\infty$, the values are identical to the ones in terms of the Euclidean volume with $N=100$, shown in Table~\ref{Table:largeN-V}.}
\label{Table:largeN}
\begin{tabular}{|l|c|c|c|c|c|c|} \hline
& $V_{\rm Q}/V_{\rm C}|_{\omega/\Omega=1}$ & $V_{\rm Q}/V_{\rm C}|_{\omega/\Omega=0}$ & $\delta V_{\rm C}/V_{\rm C}|_{\omega/\Omega=1}$ & $\delta V_{\rm C}/V_{\rm C}|_{\omega/\Omega=0}$ & $\delta V_{\rm Q}/V_{\rm Q}$ \\ \hline \hline
$N=0$ & 0.5 & 1 & 0 & 1.414 & 1.414 \\
$N=1$ & 0.5 & 1 & 0 & 0.8165 & 0.8165 \\
$N=2$ & 0.5 & 1 & 0 & 0.7483 & 0.7483 \\
$N=3$ & 0.5 & 1 & 0 & 0.7284 & 0.7284 \\
$N=4$ & 0.5 & 1 & 0 & 0.7201 & 0.7201 \\
$N=5$ & 0.5 & 1 & 0 & 0.7158 & 0.7158 \\
$N=10$ & 0.5 & 1 & 0 & 0.7095 & 0.7095 \\
$N=100$ & 0.5 & 1 & 0 & 0.7071 & 0.7071 \\
$N\to \infty$ & 0.5 & 1 & 0 & 0.7071 & 0.7071 \\
\hline
\end{tabular}
\caption{Values of $V_{\rm Q}/V_{\rm C}$, $\delta V_{\rm C}/V_{\rm C}$, and $\delta V_{\rm Q}/V_{\rm Q}$ are shown for $\Psi=\Psi_{{\rm I}(N)}$ for some values of $N$. The values of $\omega/\Omega$ in $V_{\rm Q}/V_{\rm C}$ and $\delta V_{\rm C}/V_{\rm C}$ are $\omega/\Omega=1$ and $0$.}
\label{Table:largeN-V}
\end{center}
\end{table}
The values of $A_{\rm Q}/A_{\rm C}$, $\delta A_{\rm C}/A_{\rm C}$, and $\delta A_{\rm Q}/A_{\rm Q}$ in the large-$N$ (actually $N=100$) case are shown in Table~\ref{Table:largeN}, where $\omega/\Omega$ is set to be $1$ and $0$ for $A_{\rm Q}/A_{\rm C}$ and $\delta A_{\rm C}/A_{\rm C}$.
For comparison, we present the same quantities using the Euclidean volume instead of area in Table~\ref{Table:largeN-V}.
\section{Quantum toroidal black hole: Wave function II}
\label{sec:exact2}
In this section, we will study quantum black holes described by the wave function $\Psi=\Psi^{(\pm)}_{{\rm II}(N)}$, which is given by Eq.~(\ref{SolHalfLine}) with $\alpha_0=0$ and ${\bar \beta}_0=\sqrt{\omega}$ and corresponds to generalized coherent states on the half line.
In the present section, the upper and lower signs in the expression of the expectation value $\langle X\rangle_{N}$ for a quantity $X$ are for $\Psi=\Psi^{(+)}_{{\rm II}(N)}$ and $\Psi=\Psi^{(-)}_{{\rm II}(N)}$, respectively.
\subsection{Dynamical singularity resolution}
Let us compute the expectation value of the Euclidean volume with the areal radius $r(=a)$:
\begin{align}
\langle V_{\rm E}\rangle_N(t):=&\frac{V_{n-2}^{(0)}}{n-1}\langle x^{2}\rangle_N =V_{\rm p}\frac{(n-1)l\omega\langle {\bar x}^2\rangle_N}{2(n-2)V_{n-2}^{(0)}\ell_{\rm p}}. \label{<V>E-wave2}
\end{align}
By construction, $\langle V_{\rm E}\rangle_N(t)$ is positive definite and hence the classical singularity at $r=0$ (hence ${\bar x}=0$) is resolved quantum mechanically.
The best friendly forms of the first six of the quantity $\omega\langle {\bar x}^2\rangle_N$ are
\begin{align}
\omega\langle {\bar x}^2\rangle_{0}=&\frac{2\varepsilon^2+1}{2}\mp \frac{\zeta}{e^\zeta\pm 1} \nonumber \\
\omega\langle {\bar x}^2\rangle_1=&\frac{2\varepsilon^2+3}{2}\mp \frac{4\varepsilon^2+\zeta(2\zeta-3)}{e^\zeta\pm(2\zeta-1)},\label{<X2>-II1}\\
\omega\langle {\bar x}^2\rangle_2=&\frac{2\varepsilon^2+5}{2}\mp \frac{4(2\zeta-3)\varepsilon^2+\zeta(2\zeta^2-8\zeta+7)}{e^\zeta\pm(2\zeta^2-4\zeta+1)},\\
\omega\langle {\bar x}^2\rangle_3=&\frac{2\varepsilon^2+7}{2} \mp \frac{24(\zeta-1)(\zeta-3)\varepsilon^2+\zeta(4\zeta^3-30\zeta^2+66\zeta-39)}{3e^\zeta\pm(4\zeta^3-18\zeta^2+18\zeta-3)},\\
\omega\langle {\bar x}^2\rangle_4=&\frac{2\varepsilon^2+9}{2}\mp \frac{8(2\zeta^3-15\zeta^2+30\zeta-15)\varepsilon^2+\zeta(2\zeta^4-24\zeta^3+96\zeta^2-144\zeta+63)}{3e^\zeta\pm(2\zeta^4-16\zeta^3+36\zeta^2-24\zeta+3)},\\
\omega\langle {\bar x}^2\rangle_5=&\frac{2\varepsilon^2+11}{2} \nonumber \\
&\mp \frac{20(2\zeta^4-24\zeta^3+90\zeta^2-120\zeta+45)\varepsilon^2+\zeta(4\zeta^5-70\zeta^4+440\zeta^3-1200\zeta^2+1350\zeta-465)}{15e^\zeta\pm(4\zeta^5-50\zeta^4+200\zeta^3-300\zeta^2+150\zeta-15)}.\label{<X2>-II2}
\end{align}
It is seen that the time dependence of $\langle V_{\rm E}\rangle_N$ is determined by $\varepsilon(t)^2$.
Because $\varepsilon(t)$ itself is oscillating within the range $-\sqrt{\zeta}\le \varepsilon\le\sqrt{\zeta}$ as shown in Eq.~(\ref{eq:SPepsilon}), the dynamics of $\langle V_{\rm E}\rangle_N$ is a simple oscillation.
The above expressions imply
\begin{align}
\lim_{\zeta\to\infty}\omega\langle {\bar x}^2\rangle_N\simeq &\varepsilon^2+\biggl(N+\frac12\biggl)+{\cal O}(\zeta^{N+1}e^{-\zeta}) \label{limit<x^2>}
\end{align}
and hence $\langle V_{\rm E}\rangle_N$ is almost equally spaced for large $\zeta$.
\subsection{Mass of quantum black hole}
The first six mass expectation values $\langle M\rangle_N$ are given by
\begin{align}
\langle M\rangle_0=&\frac12\hbar\omega+\frac12\hbar\omega\zeta\frac{e^{\zeta}\mp1}{e^{\zeta}\pm1},\label{<M>0-II}\\
\langle M\rangle_1=&\frac32\hbar\omega+\frac12\hbar\omega\zeta\frac{e^\zeta\mp(2\zeta-1)}{e^{\zeta}\pm(2\zeta-1)},\\
\langle M\rangle_2=&\frac52\hbar\omega+\frac12\hbar\omega\zeta\frac{e^\zeta\mp(2\zeta^2-4\zeta+1)}{e^{\zeta}\pm(2\zeta^2-4\zeta+1)},\\
\langle M\rangle_3=&\frac72\hbar\omega+\frac12\hbar\omega\zeta\frac{3e^\zeta\mp(4\zeta^3-18\zeta^2+18\zeta-3)}{3e^{\zeta}\pm(4\zeta^3-18\zeta^2+18\zeta-3)},\\
\langle M\rangle_4=&\frac92\hbar\omega+\frac12\hbar\omega\zeta\frac{3e^\zeta\mp(2\zeta^4-16\zeta^3+36\zeta^2-24\zeta+3)}{3e^{\zeta}\pm(2\zeta^4-16\zeta^3+36\zeta^2-24\zeta+3)},\\
\langle M\rangle_5=&\frac{11}{2}\hbar\omega+\frac12\hbar\omega\zeta\frac{15e^\zeta\mp(4\zeta^5-50\zeta^4+200\zeta^3-300\zeta^2+150\zeta-15)}{15e^{\zeta}\pm(4\zeta^5-50\zeta^4+200\zeta^3-300\zeta^2+150\zeta-15)}.\label{<M>5-II}
\end{align}
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=0.65\linewidth]{McoherentqallPRD}
\caption{\label{fig-Mcoherent} The first six mass expectation values $\langle M\rangle_N/\hbar\omega$ as a function of $\zeta$ for (a) $\Psi=\Psi^{(+)}_{{\rm II}(N)}$ and (b) $\Psi=\Psi^{(-)}_{{\rm II}(N)}$.
}
\end{center}
\end{figure*}
The forms of $\langle M\rangle_N$ are drawn in Fig.~\ref{fig-Mcoherent}.
It shows that $\langle M\rangle_N$ is not a monotonic function of $\zeta$ in general and $\langle M\rangle_N$ is almost equally spaced for large $\zeta$, shown as
\begin{align}
\lim_{\zeta\to\infty}\langle M\rangle_N\simeq&\biggl(N+\frac12\biggl)\hbar\omega+\frac12\hbar\omega\zeta+{\cal O}(\zeta^{-1}). \label{<M>largezeta}
\end{align}
For small $\zeta$, on the other hand, $\langle M\rangle_{N+1}$ can be smaller than $\langle M\rangle_N$ in some ranges of $\zeta$.
It is intriguing that this type of level crossing occurs in time-dependent states of quantum black holes.
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=0.65\linewidth]{MspacingallPRD}
\caption{\label{fig-Mspacing} The first five relative mass spacings $\Delta \langle M\rangle_N/\langle M\rangle_{N}$ as functions of $\zeta$ for (a) $\Psi=\Psi^{(+)}_{{\rm II}(N)}$ and (b) $\Psi=\Psi^{(-)}_{{\rm II}(N)}$.
}
\end{center}
\end{figure*}
It is noted that all of the above $\langle M\rangle_N$ have finite limits for $\zeta\to 0$, in spite of the fact that $\Psi^{(\pm)}_{II(N)}\equiv 0$ is realized in some cases.
The limits are
\begin{align}
\lim_{\zeta\to 0}\langle M\rangle_0=&\frac12\hbar\omega,\quad \lim_{\zeta\to 0}\langle M\rangle_1=\frac{11}{6}\hbar\omega,\\
\lim_{\zeta\to 0}\langle M\rangle_2=&\frac52\hbar\omega,\quad \lim_{\zeta\to 0}\langle M\rangle_3=\frac{51}{14}\hbar\omega,\\
\lim_{\zeta\to 0}\langle M\rangle_4=&\frac92\hbar\omega,\quad \lim_{\zeta\to 0}\langle M\rangle_5=\frac{123}{22}\hbar\omega
\end{align}
for $\Psi=\Psi^{(+)}_{{\rm II}(N)}$ (where the states with $N=1,3,5$ are unphysical for $\zeta\to 0$), and
\begin{align}
\lim_{\zeta\to 0}\langle M\rangle_0=&\frac32\hbar\omega,\quad \lim_{\zeta\to 0}\langle M\rangle_1=\frac32\hbar\omega, \\
\lim_{\zeta\to 0}\langle M\rangle_2=&\frac{27}{10}\hbar\omega,\quad \lim_{\zeta\to 0}\langle M\rangle_3=\frac72\hbar\omega,\\
\lim_{\zeta\to 0}\langle M\rangle_4=&\frac{83}{18}\hbar\omega,\quad \lim_{\zeta\to 0}\langle M\rangle_5=\frac{11}{2}\hbar\omega
\end{align}
for $\Psi=\Psi^{(-)}_{{\rm II}(N)}$ (where the states with $N=0,2,4$ are unphysical for $\zeta\to 0$).
The expressions (\ref{<M>0-II})--(\ref{<M>5-II}) and Eq.~(\ref{<M>largezeta}) imply that $N\to \infty$ and $\zeta\to \infty$ are independent large black-hole limits.
The relative mass spacing $\Delta \langle M\rangle_N/\langle M\rangle_{N}$ is drawn in Fig.~\ref{fig-Mspacing}, which shows that $\Delta \langle M\rangle_N/\langle M\rangle_{N}$ converges to zero for $N\to \infty$ and also for $\zeta\gg N$.
Indeed, the expressions (\ref{<M>0-II})--(\ref{<M>5-II}) imply
\begin{align}
\lim_{\zeta\to\infty}\frac{\Delta \langle M\rangle_N}{\langle M\rangle_N} \propto \frac{1}{\zeta}\to0.
\end{align}
Here it is emphasized that the limit $\zeta\to\infty$ in the following argument does not mean infinite $\zeta$ but sufficiently large $\zeta$ satisfying $N/\zeta\ll 1$.
We will now examine the behaviour of the quantum fluctuation in the limit of $N\to \infty$ or $\zeta\to \infty$.
The relative mass fluctuation is given by
\begin{align}
\frac{\delta\langle M\rangle_N}{\langle M\rangle_N}=&\frac{\sqrt{|\langle M^2\rangle_N-\langle M\rangle_N^2|}}{\langle M\rangle_N} =\sqrt{\biggl|\frac{\int_{0}^\infty \Psi^*{\hat H}^2\Psi{\rm d} {\bar x}}{(\int_{0}^\infty\Psi^*{\hat H}\Psi{\rm d} {\bar x})^2}-1\biggl|},
\end{align}
where $\Psi=\Psi^{(\pm)}_{{\rm II}(N)}$.
The simplest forms of the first six are given by
\begin{align}
\biggl(\frac{\delta\langle M\rangle_0}{\langle M\rangle_0}\biggl)^2=&\biggl|\frac{\zeta-1}{(\zeta+1)^2}-\frac{2\zeta h_0(\zeta)}{(\zeta+1)^2\{(\zeta+1)e^{\zeta}\mp (\zeta-1)\}^2}\biggl|,\\
\biggl(\frac{\delta\langle M\rangle_1}{\langle M\rangle_1}\biggl)^2=&\biggl|\frac{3(\zeta-3)}{(\zeta+3)^2}-\frac{2\zeta h_1(\zeta)}{(\zeta+3)^2\{(\zeta+3)e^{\zeta}\mp (\zeta-3)(2\zeta-1)\}^2}\biggl|,\\
\biggl(\frac{\delta\langle M\rangle_2}{\langle M\rangle_2}\biggl)^2=&\biggl|\frac{5(\zeta-5)}{(\zeta+5)^2}-\frac{2\zeta h_2(\zeta)}{(\zeta+5)^2\{(\zeta+5)e^{\zeta}\mp (\zeta-5)(2\zeta^2-4\zeta+1)\}^2}\biggl|,\\
\biggl(\frac{\delta\langle M\rangle_3}{\langle M\rangle_3}\biggl)^2=&\biggl|\frac{7(\zeta-7)}{(\zeta+7)^2}-\frac{2\zeta h_3(\zeta)}{(\zeta+7)^2\{3(\zeta+7)e^{\zeta}\mp (\zeta-7)(4\zeta^3-18\zeta^2+18\zeta-3)\}^2}\biggl|,\\
\biggl(\frac{\delta\langle M\rangle_4}{\langle M\rangle_4}\biggl)^2=&\biggl|\frac{9(\zeta-9)}{(\zeta+9)^2}-\frac{2\zeta h_4(\zeta)}{(\zeta+9)^2\{3(\zeta+9)e^{\zeta}\mp (\zeta-9)(2\zeta^4-16\zeta^3+36\zeta^2-24\zeta+3)\}^2}\biggl|,\\
\biggl(\frac{\delta\langle M\rangle_5}{\langle M\rangle_5}\biggl)^2=&\biggl|\frac{11(\zeta-11)}{(\zeta+11)^2} \nonumber \\
&-\frac{2\zeta h_5(\zeta)}{(\zeta+11)^2\{15(\zeta+11)e^{\zeta}\mp (\zeta-11)(4\zeta^5-50\zeta^4+200\zeta^3-300\zeta^2+150\zeta-15)\}^2}\biggl|,
\end{align}
where
\begin{align}
h_0:=&(\zeta^2+3)\mp (\zeta+1)(2\zeta^2+3\zeta-3)e^\zeta,\\
h_1:=&(2\zeta-1)(2\zeta^3-27\zeta^2+126\zeta-81)\mp (\zeta-1)(\zeta+3)(4\zeta^2+24\zeta-27)e^\zeta,\\
h_2:=&(2\zeta^2-4\zeta+1)(2\zeta^4-88\zeta^3+675\zeta^2-1200\zeta+375) \nonumber \\
&\mp (\zeta+5)(4\zeta^4+30\zeta^3-180\zeta^2+255\zeta-75) e^\zeta,\\
h_3:=&(4\zeta^3-18\zeta^2+18\zeta-3)(4\zeta^5-366\zeta^4+4338\zeta^3-14847\zeta^2+14994\zeta-3087) \nonumber \\
&\mp 3(\zeta+7)(8\zeta^5+72\zeta^4-858\zeta^3+2418\zeta^2-2205\zeta+441)e^\zeta,\\
h_4:=&(2\zeta^4-16\zeta^3+36\zeta^2-24\zeta+3) \nonumber \\
&\times(2\zeta^6-312\zeta^5+5322\zeta^4-29688\zeta^3+61479\zeta^2-42768\zeta+6561) \nonumber \\
&\mp 3(\zeta+9)(4\zeta^6+38\zeta^5-822\zeta^4+3972\zeta^3-7350\zeta^2+4833\zeta-729)e^\zeta,\\
h_5:=&(4\zeta^5-50\zeta^4+200\zeta^3-300\zeta^2+150\zeta-15) \nonumber \\
&\times(4\zeta^7-950\zeta^6+22092\zeta^5-182070\zeta^4+634350\zeta^3-927465\zeta^2+490050\zeta-59895) \nonumber \\
&\mp 15(\zeta+11)(8\zeta^7+72\zeta^6-2742\zeta^5+20670\zeta^4-64800\zeta^3+88320\zeta^2-45045\zeta+5445)e^\zeta.
\end{align}
The forms of $\delta\langle M\rangle_N/\langle M\rangle_N$ are drawn in Fig.~\ref{fig-deltaMcoherent}.
The second term in the above expressions of $\delta\langle M\rangle_N/\langle M\rangle_N$ rapidly converges to zero as $\zeta$ increases and therefore the fall-off rate of the fluctuation should be
\begin{align}
\lim_{\zeta\to\infty}\frac{\delta\langle M\rangle_N}{\langle M\rangle_N}\simeq& \frac{(2N+1)\{\zeta-(2N+1)\}}{(\zeta+2N+1)^2}+{\cal O}(\zeta^Ne^{-\zeta}) \to 0. \label{delta<M>largezeta}
\end{align}
Hence, the relative mass fluctuation of the black hole in the non-stationary state becomes very small in a large-mass limit $\zeta\gg N$.
In contrast, as seen in Fig.~\ref{fig-deltaMcoherent}, the relative mass fluctuation is not negligible in a different large-mass limit $N\to\infty$.
\begin{figure}[htbp]
\begin{center}
\includegraphics[width=0.7\linewidth]{deltaMqallPRD}
\caption{\label{fig-deltaMcoherent} The first six relative mass fluctuations $\delta\langle M\rangle_N/\langle M\rangle_N$ as functions of $\zeta$ for (a) $\Psi=\Psi^{(+)}_{{\rm II}(N)}$ and (b) $\Psi=\Psi^{(-)}_{{\rm II}(N)}$. }
\end{center}
\end{figure}
\subsection{Euclidean volume of quantum black hole}
The Euclidean volume $\langle V_{\rm E}\rangle_N$ with the areal radius $r=a(t)$ is given by Eq.~(\ref{<V>E-wave2}), and $\omega\langle {\bar x}^2\rangle_N$ has the form of Eq.~(\ref{limit<x^2>}) for large $\zeta$.
In this subsection, we study the Euclidean volume of the classical and quantum horizons in the case of large $\zeta$, in which we can obtain their analytic expressions.
For large $\zeta$, the volume of the quantum horizon $V_{\rm Q}$ behaves as
\begin{align}
\lim_{\zeta\to\infty} V_{\rm Q}=&\lim_{\zeta\to\infty} \max_{t\in\mathbb{R}}\langle V_{\rm E}\rangle_N(t) \nonumber \\
\simeq& \frac{(n-1)V_{\rm p}l(2\zeta+2N+1)}{4(n-2)V_{n-2}^{(0)}\ell_{\rm p}}\to \infty.\label{VQ2-large}
\end{align}
In order to see the volume fluctuation of the quantum horizon $\delta V_{\rm Q}/V_{\rm Q}$, we need to compute
\begin{align}
\frac{\delta\langle V_{\rm E}\rangle_N}{\langle V_{\rm E}\rangle_N}=&\frac{\sqrt{|\langle V_{\rm E}^2\rangle_N-\langle V_{\rm E}\rangle_N^2|}}{\langle V_{\rm E}\rangle_N} =\sqrt{\biggl|\frac{\int_{0}^\infty {\bar x}^4|\Psi|^2{\rm d} {\bar x}}{(\int_{0}^\infty {\bar x}^2|\Psi|^2{\rm d} {\bar x})^2}-1\biggl|},
\end{align}
where $\Psi=\Psi^{(\pm)}_{{\rm II}(N)}$.
Although it is difficult to obtain the expression for general $N$, the first six are written as
\begin{align}
\biggl(\frac{\delta\langle V_{\rm E}\rangle_0}{\langle V_{\rm E}\rangle_0}\biggl)^2=&\biggl|\frac{2(1+4\varepsilon^2)}{(1+2\varepsilon^2)^2}+j_{0(\pm)}(\zeta,\varepsilon^2)\biggl|,\\
\biggl(\frac{\delta\langle V_{\rm E}\rangle_1}{\langle V_{\rm E}\rangle_1}\biggl)^2=&\biggl|\frac{2(3+12\varepsilon^2)}{(3+2\varepsilon^2)^2}+j_{1(\pm)}(\zeta,\varepsilon^2)\biggl|,\\
\biggl(\frac{\delta\langle V_{\rm E}\rangle_2}{\langle V_{\rm E}\rangle_2}\biggl)^2=&\biggl|\frac{2(7+20\varepsilon^2)}{(5+2\varepsilon^2)^2}+j_{2(\pm)}(\zeta,\varepsilon^2)\biggl|,\\
\biggl(\frac{\delta\langle V_{\rm E}\rangle_3}{\langle V_{\rm E}\rangle_3}\biggl)^2=&\biggl|\frac{2(13+28\varepsilon^2)}{(7+2\varepsilon^2)^2}+j_{3(\pm)}(\zeta,\varepsilon^2)\biggl|,\\
\biggl(\frac{\delta\langle V_{\rm E}\rangle_4}{\langle V_{\rm E}\rangle_4}\biggl)^2=&\biggl|\frac{2(21+36\varepsilon^2)}{(9+2\varepsilon^2)^2}+j_{4(\pm)}(\zeta,\varepsilon^2)\biggl|,\\
\biggl(\frac{\delta\langle V_{\rm E}\rangle_5}{\langle V_{\rm E}\rangle_5}\biggl)^2=&\biggl|\frac{2(31+44\varepsilon^2)}{(11+2\varepsilon^2)^2}+j_{4(\pm)}(\zeta,\varepsilon^2)\biggl|,
\end{align}
where the expressions of the functions $j_{N(\pm)}(\zeta,\varepsilon^2)$ are complicated but converge to zero for large $\zeta$ as
\begin{align}
\lim_{\zeta\to \infty}j_{N(\pm)}\simeq &\frac{\zeta^{N+2}e^{-\zeta}}{\{(2N+1)+2\varepsilon^2\}^2}\to0.
\end{align}
Hence, the above expressions imply that the relative volume fluctuation should behave as
\begin{align}
\lim_{\zeta\to\infty}\biggl(\frac{\delta\langle V_{\rm E}\rangle_N}{\langle V_{\rm E}\rangle_N}\biggl)^2\simeq &\frac{2(N^2+N+1)+8(2N+1)\varepsilon^2}{\{(2N+1)+2\varepsilon^2\}^2}+{\cal O}(\zeta^{N+2}e^{-\zeta}).
\end{align}
The leading term is exactly the same as in the shape-preserving state on the whole line.
Thus, the relative volume fluctuation of the quantum horizon for large $\zeta$ should be given by
\begin{align}
\lim_{\zeta\to\infty}\frac{\delta V_{\rm Q}}{V_{\rm Q}}\simeq &\frac{\sqrt{2(N^2+N+1)+8(2N+1)\zeta}}{(2N+1)+2\zeta} \to 0.
\end{align}
On the other hand, by Eqs.~(\ref{VC-def}) and (\ref{<M>largezeta}), we obtain the Euclidean volume of the classical horizon $V_{\rm C}$ for large $\zeta$ as
\begin{align}
\lim_{\zeta\to\infty}V_{\rm C}\simeq&\frac{(n-1)(\zeta+2N+1)V_{\rm p}l}{2(n-2)V_{n-2}^{(0)}\ell_{\rm p}}\to \infty.\label{VC2-large}
\end{align}
Equations~(\ref{deltaVC-def}) and (\ref{delta<M>largezeta}) show
\begin{align}
\lim_{\zeta\to\infty}\frac{\delta V_{\rm C}}{V_{\rm C}}\simeq \frac{(2N+1)\{\zeta-(2N+1)\}}{(\zeta+2N+1)^2}\to 0.
\end{align}
Now we have shown that the relative volume fluctuations of the horizon are very small in a large-mass limit $\zeta\gg N$.
In addition, Eqs.~(\ref{VQ2-large}) and (\ref{VC2-large}) show that $V_{\rm Q}/V_{\rm C}$ is close to $1$, shown as
\begin{align}
\lim_{\zeta\to\infty} \frac{V_{\rm Q}}{V_{\rm C}}\simeq& \frac{2\zeta+2N+1}{2(\zeta+2N+1)}\simeq 1-\frac{N}{\zeta}.
\end{align}
These results show that classical configurations are realized in this limit.
\begin{figure*}[htbp]
\begin{center}
\includegraphics[width=1.0\linewidth]{deltaVallzeta11040PRD}
\caption{\label{fig-deltaV(all)zeta=1-10-40} Values of $\delta \langle V_{\rm E}\rangle_N/\langle V_{\rm E}\rangle_N$ in the range of $\min_{t\in\mathbb{R}}\omega\langle {\bar x}^2\rangle_N\le \omega\langle {\bar x}^2\rangle_N\le \max_{t\in\mathbb{R}}\omega\langle {\bar x}^2\rangle_N$ for $N=0,\cdots 5$.
The left row corresponds to $\Psi=\Psi^{(+)}_{{\rm II}(N)}$, with (a1) $\zeta=1$, (a2) $\zeta=10$, and (a3) $\zeta=40$.
The right row corresponds to $\Psi=\Psi^{(-)}_{{\rm II}(N)}$, with (b1) $\zeta=1$, (b2) $\zeta=10$, and (b3) $\zeta=40$.
}
\end{center}
\end{figure*}
On the other hand, in another large-mass limit $N\to \infty$, with $\zeta$ fixed, quantum fluctuations are not negligible.
Figure~\ref{fig-deltaMcoherent} shows that the relative volume fluctuation of the classical horizon is not negligible for $N\to \infty$ because $\delta V_{\rm C}/V_{\rm C}=\delta \langle M\rangle_N/\langle M\rangle_N$ holds.
In addition, the relative volume fluctuation of the quantum horizon $\delta V_{\rm Q}/V_{\rm Q}$ is not negligible in this limit, either.
This is seen in Fig.~\ref{fig-deltaV(all)zeta=1-10-40}, which shows that the relative volume fluctuation $\delta\langle V_{\rm E}\rangle_N/\langle V_{\rm E}\rangle_N$ is in the range of $\min_{t\in\mathbb{R}}\omega\langle {\bar x}^2\rangle_N\le \omega\langle {\bar x}^2\rangle_N\le \max_{t\in\mathbb{R}}\omega\langle {\bar x}^2\rangle_N$ for $\zeta=1,10,40$.
Since $\omega\langle {\bar x}^2\rangle_N$ is proportional to $\langle V_{\rm E}\rangle_N$ and its dynamics is simple oscillation as shown in Eqs.~(\ref{<X2>-II1})--(\ref{<X2>-II2}), $\delta V_{\rm Q}/V_{\rm Q}$ is given by the value of $\delta\langle V_{\rm E}\rangle_N/\langle V_{\rm E}\rangle_N$ at $\omega\langle {\bar x}^2\rangle_N=\max_{t\in\mathbb{R}}\omega\langle {\bar x}^2\rangle_N$.
Therefore, Fig.~\ref{fig-deltaV(all)zeta=1-10-40} shows that $\delta V_{\rm Q}/V_{\rm Q}$ does not become small for $N\to\infty$.
Lastly, Table~\ref{tab:V-zeta1} shows the values of $V_{\rm Q}/V_{\rm C}$ for $N=0,1,\cdots, 5$ with $\zeta=1$ and $\zeta=20$.
Table~\ref{tab:V-zeta1} implies that classical and quantum horizons do not coincide for $N\to\infty$.
\begin{table}[htb]
\begin{center}
\begin{tabular}{|l|c|c|c|c|c|c|} \hline
\multicolumn{1}{|l|}{} & \multicolumn{2}{c|}{$\zeta=1$} & \multicolumn{2}{c|}{$\zeta=20$}\\ \hline \hline
&$\Psi=\Psi^{(+)}_{{\rm II}(N)}$ & $\Psi=\Psi^{(-)}_{{\rm II}(N)}$ &$\Psi=\Psi^{(+)}_{{\rm II}(N)}$ & $\Psi=\Psi^{(-)}_{{\rm II}(N)}$ \\ \hline
$N=0$ & 0.8420 & 0.6580 & 0.9762 & 0.9762 \\ \hline
$N=1$ & 0.5109 & 0.8222 & 0.9348 & 0.9348 \\ \hline
$N=2$ & 0.7323 & 0.5069 & 0.9000 & 0.9000 \\ \hline
$N=3$ & 0.5643 & 0.5604 & 0.8704 & 0.8704 \\ \hline
$N=4$ & 0.5382 & 0.6574 & 0.8449 & 0.8448 \\ \hline
$N=5$ & 0.7238 & 0.5736 & 0.8228 & 0.8224 \\ \hline
\end{tabular}
\caption{\label{tab:V-zeta1} Values of $V_{\rm Q}/V_{\rm C}$ with $\zeta=1$ and $\zeta=20$ for $N=0,1,\cdots,5$.}
\end{center}
\end{table}
\section{Summary and future prospects}
\label{sec:summary}
In this paper we have studied the quantum dynamics of toroidal AdS black holes in the framework of throat quantization pioneered by Louko and M\"akel\"a~\cite{LM96}. The resulting Schr\"odinger equation is equivalent to that of a quantum harmonic oscillator on the half line.
While the classical dynamics represented by the Hamiltonian is equivalent to the dynamics of the wormhole throat in the maximally extended black-hole spacetime, the time $t$ is by construction the proper time at one AdS infinity in the spacetime.
Therefore, all the expectation values of physical/geometrical quantities are for an observer at rest at AdS infinity.
The Hamiltonian operator in the Hilbert space ${\cal L}^2([0,\infty))$ admits a one-parameter family of boundary conditions at the location of the classical singularity $x=0$ with extension parameter $L$.
For any value of $L$, the evolution of the wave function is unitary, and the classical curvature singularity at the center is resolved quantum mechanically and replaced by a big bounce. We were able to construct exact time-dependent wave functions that satisfy either Dirichlet ($L=0$) or Neumann ($L=\infty$) boundary conditions.
The existence of such exact time-dependent solutions for this class of black-hole spacetimes provides a useful framework for studying conceptual issues surrounding the quantum mechanics of black holes, such as mass uncertainty and the definition of the quantum horizon in the dynamical setting.
We presented two distinct classes of exact time-dependent quantum states. Both cases clearly exhibit dynamical singularity resolution: The probability amplitudes, $|\Psi|^2$, oscillate between the origin and a maximum ``near" the classical turning point. We evaluated analytically the mass fluctuations and examined the dynamics and fluctuations of two definitions for the horizon of a quantum black hole, the first obtained from the classical expression in terms of the mass, and the second in terms of a suitably defined Euclidean volume.
The first class, wave function I, is fully dynamical in that the shape of the probability amplitude changes with time throughout the evolution. It is generally spread out far from the origin and becomes more focused near the origin. As a result, the quantum fluctuations are large for most parameter values, and a semi-classical limit is not realized except in the stationary limit. The second class, wave function II, provides a generalization to the half line of coherent states of the harmonic oscillator on the whole line. The probability amplitudes are shape preserving far from the origin and, under the circumstances where the kinetic term energy dominates the potential energy (i.e., $\zeta\ll N$), quantum fluctuations become small, and the classical and quantum horizons coincide so that they are able to represent semi-classical black-hole states. As expected the semi-classical, shape-preserving nature of the solutions breaks down near the origin due to the boundary conditions that provide the mechanism for singularity resolution.
We have shown that the stationary states are fully quantum even in the large-mass limit $N\to\infty$, in the sense that the classical and quantum horizons do not coincide and quantum fluctuations do not become small.
On the other hand, quantum effects are very much suppressed when the states are highly dynamical, i.e., far from the stationary states. This suggests that semi-classical configurations are realized in this limit.
Since our results were obtained only within two subspaces of the full parameter space of the exact time-dependent wave functions, it is still an open question as to which of the properties we observed are generic.
In addition, it would be interesting to study time-dependent states for asymptotically flat black holes.
We will address these problems elsewhere.
\subsection*{Acknowledgments}
H.~M. thanks Shunichiro Kinoshita for useful comments after the 70th JSPS Annual Meeting at Waseda University.
H.~M. thanks the Department of Physics, University of Winnipeg, and the Winnipeg Institute for Theoretical Physics for their kind hospitality and support while part of this work was carried out.
G.~K. is grateful to the Natural Sciences and Engineering Research Council for support.
| 8aa7ebd85077a69016baa1946f70cffc06b3cf0a | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The theoretical understanding of quantum critical phenomena is one of the cornerstones of modern theoretical physics. It has been known for some time that such phenomena, both classical and quantum, are characterized by universal laws described by conformal field theory (CFT) \cite{DMS}. Conformal invariance is constraining enough as to fully determine the space-time dependence of two- and three-point functions of local fields, demonstrating that they must exhibit power-law decay, a defining feature of criticality. Conformal field theory is particularly powerful in two dimensions \cite{BPZ}, where it is associated with the presence of infinite-dimensional symmetry algebras. Beyond criticality, universal behaviours subsist at large observation scales when correlation lengths are much larger than microscopic distances. These are described by non-conformal quantum field theory (QFT). One of the deepest ideas in this context is that of the renormalization group (RG). RG flows describe how, in parameter space, a QFT model varies with the scale, connecting an ultraviolet (small length scale) to an infrared (large length large scale) fixed-point CFT. These give rise to flows between CFT models, and one might wonder what structure such RG flows may take.
In the eighties Alexander B. Zamolodchikov \cite{zamc} showed that the space of {\it unitary} CFT is partially ordered by unitary RG flows: an RG flow associated to a non-conformal unitary QFT may exist from an ultraviolet (UV) to an infrared (IR) fixed point only if the central charges of the corresponding CFTs are ordered,
\begin{equation}\label{ineq}
c_{\rm UV} > c_{\rm IR}.
\end{equation}
This implies irreversibility of (nontrivial) RG flows.
Specifically, he constructed for any 1+1 dimensional unitary QFT, a function $c(s)$ (called $c$-function, or scaling function) of the renormalization group parameter $s=2\log(mr)$, where $m$ is a characteristic mass scale and $r$ is the observation length scale, with the following properties: it is non-negative for all $s$, it is monotonically strictly decreasing along the RG flow (from the UV to the IR) and it takes constant values at critical points, given by the central charges of the corresponding CFTs. Therefore, the function ``flows" between its ultraviolet (or high-energy) $c_{\rm UV}=\lim_{s\rightarrow -\infty}c(s)$ and its infrared (or low-energy) $c_{\rm IR}=\lim_{s\rightarrow \infty}c(s)$ values, which are the central charges of the two CFTs that are found at high and low energies. This implies the strict inequality \eqref{ineq}.
This statement is known as the $c$-theorem and it constitutes one of the most fundamental results for 1+1 dimensional QFT. Several alternative proofs of the $c$-theorem exist which, for instance, employ finite-size field-theory methods \cite{cth2} and holographic arguments for the entanglement entropy \cite{Casinic}.
The function $c(s)$ may be interpreted as counting degrees of freedom at a given energy scale\footnote{This interpretation is particularly well illustrated for some classes of integrable models \cite{roaming, CastroAlvaredo:2000ag} where the $c$-function visits the vicinity of multiple critical points between its UV and its IR values (the function considered in \cite{roaming} is a different $c$-function than that defined in \cite{zamc}, which however takes the same values at critical points).}. It also has the interpretation as an off-critical Casimir energy, since the Casimir energy of CFT on a cylinder is proportional to the central charge $c$ \cite{BCN,Affleck}, and it characterizes the logarithmic growth of entanglement in one-dimensional quantum critical systems \cite{HolzheyLW94, Latorre1, Calabrese:2004eu}. The key role played by the central charge $c$ in the description of critical phenomena has been nicely reviewed in John Cardy's Boltzmann Medal lecture \cite{ubi}.
In higher dimensions, similar concepts arise. Starting with John Cardy's work \cite{higherd1}, the existence in four dimensional theories of an $a$-function with similar properties to Zamolodchikov's $c$-function has been recently proven \cite{atheorem}.
The existence of a similar monotonic function of the RG flow in three dimensions, known as an $F$-function, was conjectured in \cite{Ftheorem} and later proven \cite{Ftheorem2} by using concepts of holography. Similar holographic arguments were also employed in \cite{Casinia} to provide an alternative proof of the $a$-theorem. Entropic proofs exploit the properties of the bi-partite entanglement entropy in unitary QFT (e.g.~subadditivity), emphasizing its interpretation as counting the number of degrees of freedom.
In \cite{zamc} Zamolodchikov also provided a precise construction procedure for the function $c(s)$ which employs the following correlators:
\begin{eqnarray}
&&F(z\bar{z}):= z^4 \langle T(z, \bar{z}) T(0,0) \rangle=\bar{z}^4 \langle \bar{T}(z, \bar{z}) \bar{T}(0,0) \rangle, \label{f}\\
&&G(z\bar{z}):=z^3 \bar{z}\langle T(z, \bar{z})\Theta(0,0) \rangle=\bar{z}^3 {z}\langle \bar{T}(z, \bar{z})\Theta(0,0) \rangle, \label{gg}\\
&& H(z\bar{z}):= z^2 \bar{z}^2 \langle \Theta (z, \bar{z})\Theta(0,0) \rangle,\label{hhh}
\label{fgh}
\end{eqnarray}
\noindent in terms of the usual complex coordinates $z=x+iy, \bar{z}=x-iy$ (where $y$ is imaginary time). The operators above are nothing but the various components of the energy-momentum tensor $T_\mu^\nu$ in these variables, namely:
$T(z,\bar{z}):= T_{zz}(z,\bar{z})$,
$ \bar{T}(z,\bar{z}):=T_{\bar{z}\bar{z}}(z,\bar{z})$ and the trace
$T^\mu_\mu(z,\b z) = \Theta(z,\bar{z})=4 T_{z\bar{z}}(z,\bar{z}) = 4T_{\bar{z}z}(z,\bar{z})$. Employing conservation of the energy-momentum tensor $\bar{\partial}T(z,\bar{z})+\frac{1}{4}\partial \Theta(z,\bar{z})=0$ and $\partial\bar{T}(z,\bar{z})+\frac{1}{4}\bar\partial\Theta(z,\bar{z})=0$, where $\partial:=\frac{\partial}{\partial z}$ and $\bar{\partial}:=\frac{\partial}{\partial \bar{z}}$ it then follows that the function
\begin{equation}
c(s)=4\pi^2\left(2 F(r)-G(r)-\frac{3}{8}H(r)\right),
\label{tada}
\end{equation}
with $r=z\bar{z}$ and $s=2\log(mr)$, satisfies the equation:
\begin{equation}
\frac{dc}{ds}=-3\pi^2H(r).
\end{equation}
Since $\Theta$ is a hermitian operator, the right-hand side is non-negative by reflection positivity of unitary theories. Thus $c(s)$ is monotonically decreasing. It is well known that $H(r)$ vanishes at critical points (where the trace of the stress-energy tensor is vanishing). In addition, it is also known that at critical points, $\Theta$ vanishes and
\begin{equation}
4\pi^2 \langle T(0) T(z) \rangle=\frac{c}{2z^4},
\label{7}
\end{equation}
where $c$ is the central charge, and so $4\pi^2 F(z\bar{z})=\frac{c}{2}$ at critical points. Note that the unusual $4\pi^2$ factor is due to the normalization of the energy-momentum tensor chosen above, which differs from the standard normalization $T(z,\bar{z})=-2\pi T_{zz}$ (see e.g.~\cite{DMS}).
This means that $c(s)$ as defined above satisfies all three requirements of Zamolodchikov's $c$-theorem.
As powerful as this result is, one may wonder if a proof of existence of such an RG scaling function may be found under less stringent conditions. There is a wide class of QFTs in 1+1 dimensions which are non-unitary: in their usual CFT description in terms of Virasoro representations, negative-norm states exist. Yet in many cases the spectrum is real and bounded from below, and such QFTs describe {\em bona fide}, near-critical points of local quantum models with non-hermitian but positive-spectrum hamiltonian acting on a proper Hilbert space. A good example is the famous Lee-Yang model of CFT which describes the Lee-Yang edge singularity \cite{Fisher,LYCardy}. It has an integrable massive QFT counterpart which leads to a perfectly reasonable QFT \cite{Cardymuss, TCSA, Z} (a good discussion of non-unitarity and scattering theory in the context of massive integrable QFTs can be found in \cite{TW,TW2}).
Several manifestly non-hermitian discrete implementations of the Lee-Yang model exist, such as the quantum spin chain studied in \cite{gehlen1,meandreas}, which possesses an Ising critical point and a Lee-Yang critical line. A non-unitary QFT describing the near-critical region of this spin chain, including the Ising $\mapsto$ Lee-Yang flow, was studied in \cite{fonseca}.
May we also define a $c$-function for these theories? If so, how can we establish that it is indeed a $c$-function without appealing to unitarity of the QFT?
Non-unitary 1+1 dimensional CFTs abound (e.g. there are infinitely many examples within the minimal models of CFT) and much is known about their properties. A feature that has emerged in various contexts is that for such theories, the role of the central charge $c$ is taken up by an effective central charge $c_{\mathrm{eff}}:=c-24\Delta$ where $\Delta$ is the smallest conformal dimension of the spectrum of local fields. For unitary theories $\Delta=0$, corresponding to the identity field but for non-unitary models $\Delta$ and $c$ may both be negative in such a way as to give $c_{\mathrm{eff}}\geq 0$. For instance, the ground state energy of a non-unitary CFT on the cylinder is proportional to $c_{\mathrm{eff}}$ \cite{ceff} whereas for unitary theories it is proportional to $c$ \cite{BCN,Affleck}. Similarly, the von Neumann entanglement entropy of connected sub-system in a 1+1 dimensional CFT diverges logarithmically with the size of the sub-system with a coefficient which is proportional to $c$ in unitary theories \cite{HolzheyLW94, Latorre1, Calabrese:2004eu} and to $c_{\mathrm{eff}}$ in non-unitary ones \cite{BCDLR,BR,Saleur}. These results all highlight the important role played by $c_{\mathrm{eff}}$ at criticality for non-unitary CFTs.
Furthermore, there are many known examples of RG flows between such CFTs with the property
\begin{equation}
(c_{\mathrm{eff}})_{\rm UV}> (c_{\mathrm{eff}})_{\rm IR},
\label{order}
\end{equation}
hence generalising (\ref{ineq}) for non-unitary models.
{\begin{floatingfigure}[h!]{7cm}
\begin{center}
\includegraphics[width=7cm]{cfun.pdf}
\end{center}
\caption{Typical RG flows from and to the Lee-Yang model $\mathcal{M}_{2,5}$ with $c_{\mathrm{eff}}=0.4$.\vspace{1cm}}
\label{typical}
\end{floatingfigure}}
A particularly well-known family of non-unitary CFTs are the non-unitary minimal models, commonly denoted by $\mathcal{M}_{p,q}$ with $c_{\mathrm{eff}}=1-\frac{6}{pq}$, $q>p+1$ and $p, q$ coprime. RG flows between the non-unitary minimal models generated through perturbation by particular CFT fields have been investigated for a long time. Early examples of such flows were pointed out in \cite{Lassig:1991an, Ahn:1992qi}. They proved that perturbation by a field $\phi_{1,3}$ (the least relevant field of the theory), generates the family of flows
\begin{equation}
\mathcal{M}_{p,q} + \phi_{1,3} \rightarrow \mathcal{M}_{2p-q,p}.
\label{flow1}
\end{equation}
The proof relied on perturbation theory under the assumption that $p/(q-p)\gg 1$. Further examples of RG flows amongst non-unitary minimal models have been found through the use of the thermodynamic Bethe ansatz (TBA) approach \cite{TBA1} and its massless version \cite{massless}. In particular, the families
\begin{eqnarray}
&& \mathcal{M}_{p,2p-1} + \phi_{2,1} \rightarrow \mathcal{M}_{p-1,2p-1}\nonumber \\
&& \mathcal{M}_{p,2p+1} + \phi_{1,5} \rightarrow \mathcal{M}_{p,2p-1}
\label{flows2}
\end{eqnarray}
were proposed in \cite{Martins:1991hi} and further studied in \cite{ravanini:1994pt}. As an example, the solid line in Fig.~\ref{typical} shows the massless flow $ \mathcal{M}_{3,5} + \phi_{2,1} \rightarrow \mathcal{M}_{2,5}$
as presented in \cite{dorey:2000zb}. A more general set of flows was proposed and explored in \cite{dorey:2000zb} giving the following families:
\begin{eqnarray}
&& \mathcal{M}_{p,q} + \phi_{2,1} \rightarrow \mathcal{M}_{q-p,q} \quad \mathrm{for} \quad p<q<2p \nonumber \\
&& \mathcal{M}_{p,q} + \phi_{1,5} \rightarrow \mathcal{M}_{p,4p-q} \quad \mathrm{for} \quad 2p<q<3p \nonumber \\
&& \mathcal{M}_{p,q} + \phi_{1,5} \rightarrow \mathcal{M}_{4p-q,p}\quad \mathrm{for} \quad 3p<q<4p.
\label{flows3}
\end{eqnarray}
which contain some of the examples above but are more general.
Within the TBA approach it is common to define a scaling function $c_{\mathrm{eff}}^{\mathrm{TBA}}(s)$ which for unitary theories has the same features as Zamolodchikov's $c$-function (even though they are distinct functions). The TBA scaling function has been evaluated for some of the flows listed above. In all cases the property (\ref{order}) is observed (and can be easily checked by simply evaluating $c_{\mathrm{eff}}$ for the listed theories). Moreover, in many examples, such as the flows (\ref{flows2}), the TBA scaling function is monotonic, although this is not always the case \cite{dorey:2000zb}.
In addition to the massless flows (\ref{flow1})-(\ref{flows3}) between nontrival non-unitary CFTs, there are also known examples of massless flows connecting unitary to non-unitary minimal models.
An example is provided by the QFT studied in \cite{fonseca} which may be described as implementing the flow
\begin{equation}
\mathcal{M}_{3,4}+ \lambda_ 1\phi_{1,3}+i \lambda_2 \phi_{1,2} \mapsto \mathcal{M}_{2,5}
\end{equation}
with $\lambda_{1,3}\in \mathbb{R}$. Here $\mathcal{M}_{3,4}$ is the critical Ising model, $\mathcal{M}_{3,4}+ \lambda_ 1\phi_{1,3}$ is the massive Ising model, $i\lambda_2\phi_{1,2}$ represents an imaginary magnetic field, and the final theory (the IR point) is the critical Lee-Yang model. Finally, there are also examples of massive flows from a non-unitary minimal model to the trivial fixed point, $(c_{\mathrm{eff}})_{\mathrm{IR}}=0$. The dashed line in Fig.~\ref{typical} shows the massive flow
\begin{equation}
\mathcal{M}_{2,5}+i \lambda \phi_{1,2} \mapsto \{{\bf 1}\}
\end{equation}
with $\lambda\in{\mathbb{R}}$, from the minimal Lee-Yang model described by its massive integrable QFT counterpart \cite{Cardymuss,TCSA, Z}.
The examples above make it reasonable to expect that a generalization of Zamolodchikov's $c$-theorem will give rise to a similar-looking ``$c_{\mathrm{eff}}$-theorem''. In this letter we show that a $c_{\mathrm{eff}}$-function exists with all the properties of Zamolodchikov's $c$-function.
This is shown under the standard QFT properties of Poincar\'e invariance and locality, as well as {\it unbroken parity--time-reversal ($\mathcal{PT}$) symmetry}. At critical points, under certain additional natural assumptions which hold at least in rational models of CFT, this function equals the effective central charge \cite{ceff}. This thus shows that the space of $\mathcal{PT}$-symmetric CFTs is partially ordered by ($\mathcal{PT}$-symmetric) RG flows, and that this order is characterized by \eqref{order}.
Interestingly, there also exist examples of flows where the inequality (\ref{order}) is violated \cite{nonmo1,nonmo2,nonmo25,nonmo3,nonmo4}. We discuss how these may break some of the assumptions underlying our result in Section \ref{sectest}.
The paper is organized as follows. In Section \ref{secgen} we recall some general facts about QFT in order to emphasize which results hold without unitarity, and we express the precise meaning of $\mathcal{PT}$-symmetry we use in order to replace unitarity. In Section \ref{sechash} we derive some basic consequences including a reflection positivity statement, and we discuss aspects of locality. In Section \ref{secthm} we prove our main irreversibility theorem, following closely the steps of Zamolodchikov's original proof. Finally in Section \ref{sectest} we briefly discuss various examples, and we conclude in Section \ref{secconclu}.
\section{Generalities and main assumptions}\label{secgen}
In this section, we establish the precise context in which we work, stating the assumptions and their immediate consequences. Except where otherwise stated, we work in real time with Minkowskian metric. There are three assumptions. The first two, Poincar\'e invariance and locality, are standard assumptions in relativistic QFT. The last one is $\mathcal{PT}$-symmetry. If unbroken -- that is, if there is a basis of states that are $\mathcal{PT}$-invariant -- this symmetry guarantees the reality of the hamiltonian spectrum in non-hermitian quantum models \cite{wigner,BB}. In our proof of irreversibility, we need both reality of spectra and the more dynamical statement of $\mathcal{PT}$-invariance of the stress-energy tensor.
\subsection{Poincar\'e invariance and locality}
We consider a Hilbert space ${\cal H}$ on which we define a non-unitary quantum system. The Hilbert space has all appropriate structures, including a non-degenerate inner product which we denote with the usual Dirac bra-ket notation $\langle v|w\rangle$, and an associated hermitian structure ${}^\dag$, with as usual $|v\rangle^\dag = \langle v|$. The non-unitary quantum system is defined by its non-hermitian, diagonalizable hamiltonian $H$, with $H \neq H^\dagger$. Except for unitarity, other standard assumptions of relativistic QFT are made, which we review here for clarity.
We assume translation invariance, with associated momentum operator $P$ which satisfies
\begin{equation}
[H,P]=0.
\end{equation}
Space and time translations are effected as usual as
\begin{equation}
{\cal O}(x,t) = e^{iHt-iPx}{\cal O} e^{-iHt+iPx}
\end{equation}
where we identify ${\cal O} = {\cal O}(0,0)$.
As per the standard precepts of QFT, $H$ and $P$ are integrals of hamiltonian and momentum densities with locality properties:
\begin{equation}
H=\int dx \, h(x,t), \quad P=\int dx \, p(x,t)
\end{equation}
with
\begin{equation}\label{hh}
[h(x,t),h(y,t)] = [h(x,t),p(y,t)] = [p(x,t),p(y,t)]=0\quad \forall\;x\neq y,\quad \forall t
\end{equation}
A (homogeneous) local field ${\cal O}(x)={\cal O}(x,0)$ is an operator satisfying $[P,{\cal O}(x)]=i\partial_x{\cal O}(x)$, and such that $[h(x),{\cal O}(y)]=[p(x),{\cal O}(y)]=0$ for all $x\neq y$. The usual locality arguments are assumed: let ${\cal O}(x)$ be local; then if $\int \mathrm{d} x\,{\cal O}(x)=0$ then ${\cal O}(x) = \partial_x \t{\cal O}(x)$ for local $\t{\cal O}(x)$; and if $\partial_x {\cal O}(x) = 0$ then ${\cal O}(x)= a {\bf 1}$ for some constant $a\in{\mathbb{C}}$. The above equations then imply the existence of local currents $j(x,t)$ and $k(x,t)$ such that
\begin{equation}\label{cur}
\partial_t {h}(x,t)+\partial_x {j}(x,t)=0, \quad
\partial_t {p}(x,t)+\partial_x {k}(x,t)=0.
\end{equation}
As usual, these hold inside correlation functions except at the space-time positions where other local fields are inserted, in which case $\delta$-type contact terms arise, see e.g. \cite{DMS} or any QFT textbook.
We assume clustering of correlation functions, namely factorization of correlators of local fields at large space distances.
We further assume Lorentz invariance. The basic condition of Lorentz invariance is the equality between the energy current and the momentum density,
\begin{equation}
j=p.
\end{equation}
Then we can construct the boost operator
\begin{equation}
B = \int \mathrm{d} x\,xh(x),
\end{equation}
which satisfies the correct relation for the (two-dimensional) Poincar\'e group,
\begin{equation}\label{bh}
[B,H] = i\int \mathrm{d} x\,x \partial_th(x,t) = -i\int \mathrm{d} x\,x\partial_x p(x,t)
= i\int \mathrm{d} x\,p(x,t) = iP
\end{equation}
and
\begin{equation}\label{bp}
[B,P] = -i\int \mathrm{d} x\,x\partial_x h(x,t) = i H.
\end{equation}
These relations hold up to local densities at infinity. By the clustering property, such local densities at infinity do not contribute whenever the operators are appropriately exponentiated, for instance when acting on local observables by adjoint action, hence can be neglected.
By using \eqref{bh}, \eqref{bp} as well as arguments based on locality, it is possible to deduce the following relations $i[B,h] = -2p$, $i[B,k]=-2p$ and $i[B,p]=-h-k$, up to derivatives of local fields. In the standard form of the stress-energy tensor, these derivatives are assumed to be zero (see a discussion of the stress-energy tensor in \cite{DMS}). Since such fundamental relations are not expected to be based on unitarity, here we assume that they hold.
Let $z=x-t$ and $\bar{z}=x+t$. By the Poincar\'e algebra, in general we have
\begin{equation}
e^{i\alpha B}{\cal O}(z,\b z)e^{-i\alpha B} = (e^{i\alpha B}{\cal O} e^{-i\alpha B})
(e^{-\alpha} z,e^{\alpha} \b z).
\end{equation}
Following the usual construction, consider the operators
\begin{eqnarray}
{\tau}(z,\bar{z})&=&\frac{h(x,t)+k(x,t)+2p(x,t)}{4}\nonumber\\ \bar{\tau}(z,\bar{z})&=&\frac{h(x,t)+k(x,t)-2p(x,t)}{4}\nonumber\\ \theta(z,\bar{z})&=&k(x,t)-h(x,t).
\label{operators}
\end{eqnarray}
These have spin 2, -2 and 0 respectively,
\begin{equation}\label{btau}
i[B,\tau] = -2\tau,\quad i[B,\b \tau] = 2\b\tau,\quad
i[B,\theta]=0,
\end{equation}
from which we find the standard transformation properties,
\begin{equation}\label{Btrans}
e^{i\alpha B}\tau e^{-i\alpha B} = e^{-2\alpha}\tau,\quad
e^{i\alpha B}\b\tau e^{-i\alpha B} = e^{-2\alpha}\b\tau,\quad
e^{i\alpha B}\theta e^{-i\alpha B} = \theta.
\end{equation}
These operators satisfy
\begin{equation}\label{curb}
\partial \bar{\tau}+\frac{1}{4}\bar{\partial}\theta=0 \quad\mbox{and} \quad
\bar{\partial}\tau +\frac{1}{4}\partial \theta=0
\end{equation}
where $\partial = \partial/\partial z$ and $\b\partial = \partial/\partial\b z$.
\subsection{$\mathcal{PT}$ symmetry and real spectra}
The core of our derivation of the $c_\mathrm{eff}$ theorem is the reality of the hamiltonian and momentum spectra, as well as $\mathcal{PT}$-invariance of the stress-energy tensor and of the ground state. For a $\mathcal{PT}$-invariant diagonalizable hamiltonian, the statement that all its eigenstates have real eigenvalues is equivalent to the statement that one can choose a basis of $\mathcal{PT}$-invariant eigenstates (that is, to the statement that $\mathcal{PT}$-symmetry is not spontaneously broken)\footnote{It is well-known that unbroken $\mathcal{PT}$-symmetry is not a necessary condition for the energy spectrum to be real. This can be guaranteed by the condition of pseudo-hermiticity as discussed at length in Ali Mostafazadeh's work \cite{Mo1,Mo2,Mo3}. However we would like to emphasize that our current derivation does require unbroken $\mathcal{PT}$-symmetry since, together with the reality of the spectrum, we require QFT correlators to be $\mathcal{PT}$-invariant (see section 4).}.
The simple idea that unbroken $\mathcal{PT}$-symmetry guarantees the reality of the energy spectrum even if a hamiltonian is non-hermitian is at the heart of an active area of research. The ideas underpinning this research go back to the early days of quantum mechanics \cite{wigner} when it was realized that hermiticity of the hamiltonian is not strictly required in order to define a meaningful quantum mechanical model. The area has been popularized more recently thanks to a great extent to the pioneering work \cite{BB} in which a family of non-hermitian quantum mechanical hamitonians was shown to exhibit real energy spectrum. Various reviews of the field can be found in \cite{Carla,Bender,Mosta}. More recent developments are covered in the special issue \cite{spePT}.
Parity invariance is the transformation $x\rightarrow -x$, equivalently $z\leftrightarrow -\bar{z}$. Time reversal is the operation of complex conjugation of coefficients of states and operators as written in a chosen basis (the choice of the basis defines the time reversal operation). The combination is the operation $\mathcal{PT}$, which we will see (by a slight abuse of notation) as acting on operators and on vectors and co-vectors. We understand this operation as an anti-linear involution of the operator algebra, which preserves the inner product up to complex conjugation, $\mathcal{PT}(\langle v|)\mathcal{PT}(|w\rangle) = \langle v|w\rangle^*$, and which preserves the momentum operator,
\begin{equation}\label{PTP}
\mathcal{PT}(P) = P.
\end{equation}
Our basic dynamical assumption is that {\em the stress-energy tensor is $\mathcal{PT}$-invariant}. That is, we assume
\begin{equation}\label{PTass}
\mathcal{PT}(h(0,0)) = h(0,0),\quad
\mathcal{PT}(p(0,0)) = p(0,0),\quad
\mathcal{PT}(k(0,0)) = k(0,0).
\end{equation}
Thanks to \eqref{PTP} and anti-linearity this implies $\mathcal{PT}(h(x,0)) = h(-x,0)$, hence $\mathcal{PT}(H) = H$, and therefore
\begin{equation}\label{PTh}
\mathcal{PT}(h(x,t)) = h(-x,-t),\quad
\mathcal{PT}(p(x,t)) = p(-x,-t),\quad
\mathcal{PT}(k(x,t)) = k(-x,-t).
\end{equation}
Consider the simultaneous right-eigenvalue equations for the hamiltonian and momentum operator, $H|R_n\rangle = E_n|R_n\rangle$ and $P|R_n\rangle = p_n|R_n\rangle$. We further assume that {\em all eigenvalues $E_n$ and $p_n$ are real}, or equivalently \cite{wigner,BB} that {\em all eigenstates $|R_n\rangle$ are $\mathcal{PT}$-invariant} (unbroken $\mathcal{PT}$-symmetry).
We also assume, as usual in QFT, the set $\{E_n\}$ to be bounded from below, and the lowest-energy state $|R_0\rangle$ to be unique; by appropriate identity-shift of $H$ and $P$, we choose it to have zero energy and momentum, $E_0=p_0=0$. Note that by Lorentz invariance in QFT, we must have $E_n \geq |p_n|$.
Denoting by $\langle L_n|$ the simultaneous left-eigenvectors for the hamiltonian and momentum operator, we therefore have
\begin{equation} \label{Heq}
H \neq H^\dagger, \quad H |R_n\rangle= E_n |R_n\rangle \quad \mathrm{and} \quad \langle L_n|H= E_n \langle L_n|,\quad E_n\in{\mathbb{R}}
\end{equation}
and
\begin{equation}\label{Peq}
P|R_n\rangle = p_n|R_n\rangle \quad \mathrm{and} \quad \langle L_n|P= p_n \langle L_n|,\quad p_n\in{\mathbb{R}}.
\end{equation}
It follows that $H^\dagger |L_n\rangle= E_n |L_n\rangle$ and $\langle R_n|H^\dagger= E_n \langle R_n|$. The vector $|R_n\rangle$ and co-vector $\langle L_n|$ are the right and left eigenvectors, respectively, of the hamiltonian $H$. For $H$ non-hermitian we have in general that $|R_n \rangle^\dagger \neq \langle L_n|$, and the vectors $|R_n\rangle$ and $| R_m\rangle$ are not orthogonal under the Hilbert space structure, $\langle R_m|R_n\rangle\neq \delta_{nm}$. However, we may construct our basis in such a way as to have
\begin{equation}
\langle L_n| R_m\rangle =\delta_{nm}.
\end{equation}
In the context of non-hermitian quantum mechanics this is termed a biorthogonal basis \cite{bi0, bi1,bi2}. In particular, we have
\begin{equation} \label{decomp}
{\bf{1}}=\sum_n |R_n \rangle \langle L_n|.
\end{equation}
We choose the vectors $|R_m\rangle$ and $|L_m\rangle$ as such. Note that both are $\mathcal{PT}$-invariant.
That is,
\begin{equation}
\mathcal{PT}(|R_n\rangle) = |R_n\rangle,\quad
\mathcal{PT}(\langle L_n|) = \langle L_n|.
\end{equation}
\section{The hash operation and reflection positivity}\label{sechash}
In unitary models, the hermitian conjugation of the Hilbert space guarantees a very important property, at the basis of the ordinary $c$-theorem: reflection positivity. For non-hermitian hamiltonians, this does not hold anymore. However, in the theory of $\mathcal{PT}$-symmetric quantum mechanics, it is known that when $\mathcal{PT}$ symmetry is not broken, a similarity transformation exists which relates the hamiltonian to a hermitian counterpart\footnote{Such a similarity transformation also exists under the condition of pseudo-hermiticity \cite{Mo1,Mo2,Mo3}, distinct from $\mathcal{PT}$-symmetry and less stringent in some respects. However, as noted in footnote 2, we do require $\mathcal{PT}$-symmetry in the present context.} \cite{Carla,Bender,Mosta,spePT}.
With this new hermitian structure, the usual results of quantum mechanics hold. In the present context, it is important to specify how this new hermitian structure interacts with locality and Poincar\'e invariance.
\subsection{Hash operation}
Instead of exhibiting the explicit similarity transformation, we instead define the ${}^\#$ operation on ${\rm End}({\cal H})$, which plays the role of the new hermitian conjugation. Let ${\cal O}\in {\rm End}({\cal H})$ be an operator on the Hilbert space. Then ${\cal O}^\#$ is defined by the relation
\begin{equation} \label{hash}
\langle L_m| {\mathcal{O}}^{\#} |R_n \rangle:= \langle L_n| {\mathcal{O}} |R_m \rangle^*=
\langle R_m| {\mathcal{O}}^\dagger |L_n \rangle
\end{equation}
(where we recall that ${}^*$ is complex conjugation). It is a simple matter to see that the hash operation is an antilinear involution, and further that
\begin{equation}\label{o1o2}
({\cal O}_1 {\cal O}_2)^\# = {\cal O}_2^\# {\cal O}_1^\#.
\end{equation}
The latter is derived as follows:
\begin{eqnarray}
\langle L_m| (\mathcal{O}_1\mathcal{O}_2)^{\#}
|R_n \rangle
&=&
\langle L_n| \mathcal{O}_1 \mathcal{O}_2 |R_m \rangle^*
=
\lt(\sum_k \langle L_n| \mathcal{O}_1|R_k\rangle \langle L_k|
\mathcal{O}_2 |R_m \rangle\rt)^* \nonumber\\
&=&
\sum_k \langle L_k| \mathcal{O}_1^\#|R_n\rangle \langle L_m|
\mathcal{O}_2^\# |R_k \rangle
=
\langle L_m| \mathcal{O}_2^\#
\mathcal{O}_1^\# |R_n \rangle.
\end{eqnarray}
Also, thanks to reality of spectra, we find that $\langle L_m| {H}^{\#} |R_n \rangle= \langle R_m| {H}^\dagger |L_n \rangle=E_n \delta_{m,n}=\langle L_m| {H} |R_n \rangle$ and similarly for $P$, and thus
\begin{equation}\label{hphash}
{H}^{\#}=H\quad\mathrm{and}\quad P^\# = P.
\end{equation}
In particular,
\begin{equation}\label{oxt}
{\cal O}(x,t)^\# = (e^{iHt-iPx}{\cal O}(0,0)e^{-iHt+iPx})^\#
= e^{iHt-iPx}{\cal O}(0,0)^\# e^{-iHt+iPx}
= {\cal O}^\#(x,t).
\end{equation}
The hash operation may also be defined as an antilinear operation from $\mathcal H$ to its dual $\mathcal H^*$, and vice versa. Consistency with the above definition on operators implies
\begin{equation}
|R_n\rangle^\# = \langle L_n|,\quad
\langle L_n|^\# = |R_n\rangle.
\end{equation}
From this and the fact that the basis states $|R_n\rangle$ and $\langle L_n|$ are $\mathcal{PT}$-invariant, we see that $^\#$, as acting on $\mathcal H$ and as acting $\mathcal H^*$, commutes with $\mathcal{PT}$. Therefore, it does so as well as acting on ${\rm End}(\mathcal{H})$:
\begin{equation}\label{PThash}
\mathcal{PT}({\cal O}^\#) = \mathcal{PT}({\cal O})^\#.
\end{equation}
\subsection{Reflection positivity}
Let us now consider a generic ground-state two-point function $\langle L_0| {{\cal O}}_1(x_1,t_1) {\cal O}_2(x_2,t_2)|R_0\rangle$ of some operators ${\cal O}_1,\, {\cal O}_2$.
Using the decomposition of the identity \eqref{decomp} as well as \eqref{Peq} and \eqref{Heq}, we find that it is a function of time and position differences only:
\begin{equation}
\langle L_0| {{\cal O}}_1(x_1,t_1) {\cal O}_2(x_2,t_2)|R_0\rangle=\sum_n e^{-i(t_1-t_2)E_n + i(x_1-x_2)p_n} \langle L_0| {{\cal O}}_1| R_n\rangle \langle L_n| {\cal O}_2|R_0\rangle.
\end{equation}
Using \eqref{hash}, \eqref{o1o2} and \eqref{oxt}, we also see that
\begin{equation}
\langle L_0| {{\cal O}}_1(x_1,t_1) {\cal O}_2(x_2,t_2)|R_0\rangle=\langle L_0| {{\cal O}}_2^{\#}(x_2,t_2) {\cal O}_1^{\#}(x_1,t_1)|R_0\rangle^* .
\label{rel1}
\end{equation}
Finally, consider the time-dependent two-point function of ${\mathcal{O}}^{\#}$ and $\mathcal{O}$ in imaginary time $t=-iy,\;y\in{\mathbb{R}}$, in the ground state of the theory,
\begin{equation}
\langle L_0|{\mathcal{O}}^{\#} (x,-iy) \mathcal{O}(x,0) |R_0\rangle=\langle L_0|e^{-y H}{\mathcal{O}}^{\#} (x,0)e^{y H} \mathcal{O}(x,0) |R_0\rangle=\langle L_0|{\mathcal{O}}^{\#}(x,0)e^{y H} \mathcal{O}(x,0) |R_0\rangle.
\end{equation}
Using \eqref{decomp} we write
\begin{equation}
\langle L_0|{\mathcal{O}}^{\#}(x,-iy) \mathcal{O}(x,0) |R_0\rangle=\sum_n e^{y E_n} \langle L_0| {\mathcal{O}}^{\#}(x,0) |R_n \rangle \langle L_n| {\mathcal{O}}(x,0) |R_0 \rangle.
\end{equation}
Thanks to \eqref{hash}, this sum is {\it positive-(semi)definite}:
\begin{equation}\label{pos}
\langle L_0|{\mathcal{O}}^{\#}(x,-iy) \mathcal{O}(x,0) |R_0\rangle \geq0
\end{equation}
(equality occurs for some value of $y$, if and only if all correlation functions involving ${\cal O}(x,t)$ vanish). This is reflection positivity.
\subsection{Hash-locality and conserved currents}
Since $H$ and $P$ are invariant under the hash operation, it is clear that the relations defining locality (see the discussion below \eqref{hh}) may be hashed, keeping invariant the position $x$. We may therefore distinguish two classes of locality, both associated to the same real space parameterized by $x$: {\em locality} and {\em hash-locality}, the first with respect to $h(x)$ and $p(x)$, the second to $h^\#(x)$ and $p^\#(x)$. Hashing \eqref{cur}, we have
\begin{equation}\label{curhash}
\partial_t {h}^{\#}(x,t)+\partial_x {j}^{\#}(x,t)=0, \quad
\partial_t {p}^{\#}(x,t)+\partial_x {k}^{\#}(x,t)=0,
\end{equation}
where all fields are hash-local. Relations \eqref{curhash} hold in correlation functions except at space-time positions $(x,t)$ at which other hash-local fields are inserted, where additional standard contributions in the form of $\delta$-function contact terms will arise. Passing to the $\tau^\#$, $\b\tau^\#$ and $\theta^\#$ fields, these equations are
\begin{equation}\label{hashcurb}
\partial \bar{\tau}^\#+\frac{1}{4}\bar{\partial} \theta^\#=0\quad\mbox{and} \quad
\bar{\partial}\tau^\# +\frac{1}{4}\partial \theta^\#=0.
\end{equation}
It will be important below to consider mixed correlation functions, $\langle L_0|{\cal O}_1^\#(x,t){\cal O}_2(0,0)|R_0\rangle$ for local ${\cal O}_2$ and hash-local ${\cal O}_1^\#$. Clearly, local fields are not necessarily hash-local, and can be hash-supported on extended regions. Yet, since conservations equations hold if the fields are time-separated, relations \eqref{curhash} still hold in correlation functions with insertion of local fields at times different from $t$.
\subsection{Spin}
By hashing \eqref{btau}, it is clear that the hash-local fields $\tau^\#$, $\b\tau^\#$ and $\theta^\#$ have hash-spin 2, $-2$ and 0. However, in order to determine the space-time dependence of two-point functions involving a mixture of local and hash-local fields, we need to have a common notion of spin for both types of locality. Thanks to $H^\#=H$ and $P^\#=P$, the Poincar\'e algebra implies $[H,B-B^\#]=[P,B-B^\#]=0$. It is in fact natural to assume hash-invariance of all Poincar\'e generators, imposing $B^\#=B$ and thus a single notion of spin. In order to justify this assumption, we provide general arguments that indeed lead to $B^\#=B$ for the boost operator.
Let us parametrize the energy and momentum eigenvalues of a right eigenstate $|R_n\rangle$ by $m_n\cosh\theta_n$ and $m_n\sinh\theta_n$ respectively\footnote{Note that although this parametrization may seem reminiscent of two-dimensional QFT where $m_n$ typically represents the mass and $\theta_n$ the rapidity of a one particle excitation, here $m_n$ and $\theta_n$ are introduced as generic parameters where $m_n$ is just the minimum energy (corresponding to the zero-momentum state) and $\sinh \theta_n$, $\cosh\theta_n$ simply arise from a Lorentz boost of the associated minimal energy state.}, for some $m_n > 0$ and $\theta_n\in{\mathbb{R}}$. This can always be done by positivity of the $H$ spectrum and reality of the $P$ spectrum if $E_n>|p_n|$; in the case $E_n=|p_n|$ (``massless states"), an alternative parametrization can be used, without affecting the argument\footnote{Massless states can be parameterized using the exponential function $me^\theta$ instead, for an arbitrarily chosen $m$. For such states, a boost cannot bring the momentum to zero, and cannot change its sign. Instead, the two values $\pm m$ of momentum can be used as anchors, one positive and one negative.}. Its boost $e^{i\alpha B}|R_n\rangle$ by the rapidity $\alpha\in{\mathbb{R}}$ also is a right eigenstate, with energy and momentum eigenvalues $m_n\cosh(\theta_n+\alpha)$ and $m_n\sinh(\theta_n+\alpha)$ respectively. Therefore, we may always bring its momentum to 0. The same holds for the boost of the left eigenstate $\langle L_n|e^{-i\alpha B^\#}$.
Let us change the labeling, and use the symbols $|R_n,0\rangle$ in order to label a basis of independent center-of-momentum right eigenstates and $\langle L_n,0|$ the corresponding orthogonal left eigenstates, $\langle L_n,0|R_m,0\rangle = \delta_{m,n}$ and $P|R_n,0\rangle = 0$, $\langle L_n,0|P = 0$.
Let us define the states $|R_n,\alpha\rangle := e^{i\alpha B}|R_n,0\rangle$ and $\langle L_n,\alpha|:=\langle L_n,0|e^{-i\alpha B^\#}$ for all $n$ and all $\alpha\in{\mathbb{R}}$. By the above discussion the set of states $\{|R_n,\alpha\rangle\}$ span the Hilbert space, and $\{\langle L_n,\alpha|\}$ spans its dual. Let us now consider the Fourier transforms
\begin{equation}
|R_n,\omega\rangle\ket = \int d\alpha\,e^{-i\omega\alpha}|R_n,\alpha\rangle,\quad
\langle\bra L_n,\omega| = \int d\alpha\,e^{i\omega\alpha}\langle L_n,\alpha|
\end{equation}
for $\omega\in{\mathbb{R}}$. Note that, since real boosts $\alpha\in{\mathbb{R}}$ are used, it is important that $\omega$ be real for existence of the integrals\footnote{Overlaps $\langle L_n,\alpha|R_m,\beta\rangle$ may be chosen of the form $f(\alpha)\delta(\alpha-\beta)\delta_{m,n}$. The argument presented here requires $f(\alpha)$ to grow at most polynomially at large $|\alpha|$, and the conclusion is that $f(\alpha)$ is in fact independent of $\alpha$.}. Since the Fourier transform is invertible, these states span the Hilbert space and its dual respectively, or at least dense subsets of these. Since $B|R_n,\alpha\rangle = -i\partial_\alpha|R_n,\alpha\rangle$ and $\langle L_n,\alpha|B^\# = i\partial_\alpha\langle L_n,\alpha|$, they right- and left-diagonalize $B$ and $B^\#$ respectively,
\begin{equation}
B|R_n,\omega\rangle\ket = \omega |R_n,\omega\rangle\ket,\quad
\langle\bra L_n,\omega| B^\# = \omega \langle\bra L_n,\omega|.
\end{equation}
Therefore,
\begin{equation}
\langle\bra L_n,\omega|B|R_m,\nu\rangle\ket = \omega
\langle\bra L_n,\omega|R_m,\nu\rangle\ket =
\langle\bra L_n,\omega|B^\#|R_m,\nu\rangle\ket
\end{equation}
for all $n,m$ and $\omega,\nu$. Since the vectors and covectors span (dense subsets of) the Hilbert space and its dual, we conclude that $B=B^\#$.
As a consequence,
\begin{equation}\label{Btranshash}
e^{i\alpha B}\tau^\# e^{-i\alpha B} = e^{-2\alpha}\tau^\#,\quad
e^{i\alpha B}\b\tau^\# e^{-i\alpha B} = e^{-2\alpha}\b\tau^\#,\quad
e^{i\alpha B}\theta^\# e^{-i\alpha B} = \theta^\#.
\end{equation}
\section{Irreversibility theorem}\label{secthm}
The irreversibility theorem is composed of two parts. The first part is the proof that a certain function $c_\mathrm{eff}(s)$ monotonically decreases along the RG flow (where $s$ is the log of the RG distance scale). This follows very closely Zamolodchikov's proof \cite{zamc}. The function is defined in terms of two-point functions involving stress-energy tensor components and their hash counterparts. Reflection positivity, $\mathcal{PT}$-symmetry, and the fact that the Poincar\'e group is hash-invariant ($H^\#=H$, $P^\#=P$ and $B^\#=B$) are used in a fundamental way. The function strictly decreases from its initial (UV) to its final (IR) point, unless the trace of the stress-energy tensor $\theta(x,t)$ vanishes, in which case the function is constant.
The second part aims at identifying $c_\mathrm{eff}(s)$ at fixed points. This is based on an independent argument, using the critical specific free energy. It is known that, in many models of CFT including all rational models, the specific free energy is proportional to the CFT effective central charge $c_\mathrm{eff} = c-24\Delta$, where $c$ is the central charge and $\Delta$ is the lowest dimension of the spectrum of local fields (in non-rational, non-compact models the choice of $\Delta$ may be more delicate). Thus, we identify
\begin{equation}
c(\infty) = c_\mathrm{IR}-24\Delta_\mathrm{IR}\qquad\mbox{and}\qquad c(-\infty) = c_\mathrm{UV}-24\Delta_\mathrm{UV}.
\end{equation}
\subsection{A monotonic function}
We first define the operation of taking the ``real part'', or hash-invariant part, of an operator:
\begin{equation}
A^{\mathsf R} = \frc{A+A^\#}2.
\end{equation}
Clearly $(A^{\mathsf R})^\# = A^{\mathsf R}$.
In the spirit of Zamolodchikov's original proof of the $c$-theorem, we then define the correlators:
\begin{eqnarray}
&&f(z\bar{z}):= z^4 \langle L_0| \tau^{\mathsf R}(z, \bar{z}) \tau^{\mathsf R}(0,0)|R_0 \rangle,\qquad \,\, \,\,\,\bar{f}(z\bar{z}):= \bar{z}^4 \langle L_0| \bar{\tau}^{\mathsf R}(z, \bar{z}) \bar{\tau}^{\mathsf R}(0,0)|R_0 \rangle,\nonumber\\
&&g_1(z\bar{z}):=z^3 \bar{z}\langle L_0| \tau^{\mathsf R}(z, \bar{z})\theta^{\mathsf R}(0,0) |R_0\rangle, \qquad \bar{g}_1(z\bar{z}):=\bar{z}^3 {z}\langle L_0| \bar{\tau}^{\mathsf R}(z, \bar{z})\theta^{\mathsf R}(0,0) |R_0\rangle, \nonumber\\
&& g_2(z\bar{z}):=z^3 \bar{z}\langle L_0| \theta^{\mathsf R}(z, \bar{z})\tau^{\mathsf R}(0,0) |R_0\rangle, \qquad
\bar{g}_2(z\bar{z}):=\bar{z}^3 {z}\langle L_0| {\theta}^{\mathsf R}(z, \bar{z})\bar{\tau}^{\mathsf R}(0,0) |R_0\rangle, \nonumber\\
&& q(z\bar{z}):= z^2 \bar{z}^2 \langle L_0| \theta^{\mathsf R} (z, \bar{z})\theta^{\mathsf R}(0,0) |R_0 \rangle.
\label{fgh2}
\end{eqnarray}
The fact that all these functions depend on $z\b z$ is a consequence of \eqref{Btrans} and \eqref{Btranshash}. Here and below we consider space-like distances $z\b z>0$. Thanks to \eqref{curb} and \eqref{hashcurb}, there are various relations between the derivatives of these correlators, which hold for all $z\neq \b z$:
\begin{equation}
\bar{z}\bar{\partial} f+\frac{z}{4} \partial g_2=\frac{3}{4} g_2 \quad \mathrm{and} \quad {z}{\partial} \bar{f}+\frac{\bar{z}}{4} \bar\partial \bar{g}_2=\frac{3}{4} \bar{g}_2,
\end{equation}
as well as
\begin{equation}
\bar{z}\bar{\partial} g_1+\frac{z}{4} \partial q=g_1+\frac{1}{2}q \quad \mathrm{and} \quad {z}{\partial} \bar{g}_1+\frac{\bar{z}}{4} \bar{\partial} q=\bar{g}_1+\frac{1}{2}q.
\end{equation}
These are identical to the relations found by Zamolodchikov if we identify $g_{1}=\bar{g}_1=g_2=\bar{g}_2$ and $f=\bar{f}$. However, in general these identifications do not all hold in the present case.
We may combine the relations above to write:
\begin{equation}
\bar{z}\bar{\partial} f+{z}{\partial} \bar{f}+\frac{z}{4} (\partial g_2-3\partial \bar{g}_1)+\frac{\bar{z}}{4} (\bar\partial \bar{g}_2-3\bar\partial {g}_1)-\frac{3}{16} (z\partial+\bar{z}\bar{\partial}) q=\frac{3}{4}(g_2-g_1+\bar{g}_2-\bar{g}_1-q).
\end{equation}
We may now use $\mathcal{PT}$ symmetry to show that $g_1=g_2$ and $\bar{g}_1=\bar{g}_2$. By translation invariance, hash-invariance of $\tau^{\mathsf R}$ and $\theta^{\mathsf R}$, and \eqref{rel1}, it follows that
\begin{equation}
\langle L_0| \tau^{\mathsf R}(z,\b z)\theta^{\mathsf R}(0,0) |R_0\rangle
=\langle L_0| \tau^{\mathsf R}(0,0)\theta^{\mathsf R}(-z, -\bar{z}) |R_0\rangle=\langle L_0| \theta^{\mathsf R}(-z,-\bar{z})\tau^{\mathsf R}(0, 0) |R_0\rangle^*.
\end{equation}
Using $\mathcal{PT}$-symmetry \eqref{PTh} and \eqref{PThash}, we have
\begin{eqnarray}
\langle L_0| \theta^{\mathsf R}(-z,-\bar{z})\tau^{\mathsf R}(0, 0) |R_0\rangle^*
&=&
\mathcal{PT}(\langle L_0|)\,\mathcal{PT}(\theta^{\mathsf R}(-z,-\b{z}))\,\mathcal{PT}(\tau^{\mathsf R}(0, 0))\,\mathcal{PT}(|R_0\rangle) \nonumber\\
&=&
\langle L_0|\theta^{\mathsf R}(z,\b{z})\tau^{\mathsf R}(0, 0)|R_0\rangle.
\end{eqnarray}
This implies $g_1=g_2$. The same reasoning may be applied to $\bar{g}_1$ to show that $\bar{g}_1=\bar{g}_2$. Using this, the equation above simplifies to:
\begin{equation}
\bar{z}\bar{\partial} f+{z}{\partial} \bar{f}+\frac{z}{4} (\partial g_1-3\partial \bar{g}_1)+\frac{\bar{z}}{4} (\bar\partial \bar{g}_1-3\bar\partial {g}_1)-\frac{3}{16} (z\partial+\bar{z}\bar{\partial}) q=-\frac{3}{4}q.
\end{equation}
We may now change variables to polar coordinates by writing $z=r e^{\theta}$, $\bar{z}=r e^{-\theta}$, with $r>0$. We then have that $2 z \partial = r\partial_r + \partial_\theta $ and $2\bar{z}\bar\partial=r\partial_r- \partial_\theta$, and:
\begin{equation}
\frac{1}{2} r\partial_r\left(f+\bar{f}-\frac{1}{2}(g_1+\bar{g}_1)-\frac{3}{8} q\right)- \frac{1}{2} \partial_\theta \left(\bar{f}-f +\bar{g}_1-g_1\right)=-\frac{3}{4}q.
\end{equation}
Given that all functions involved are functions of $z\bar{z}=r^2$ only, it follows that the $\theta$-derivative must be zero, thus the equation simplifies to:
\begin{equation}
r\partial_r\left(f+\bar{f}-\frac{1}{2}(g_1+\bar{g}_1)-\frac{3}{8} q\right)=-\frac{3}{2}q,
\end{equation}
or, introducing the standard RG parameter $s=2\log(mr)$ (where $m$ is an energy scale) we can write:
\begin{equation}
\frac{d}{ds}\left(f+\bar{f}-\frac{1}{2}(g_1+\bar{g}_1)-\frac{3}{8} q\right)=-\frac{3}{4}q.
\end{equation}
This now takes almost exactly the same form as Zamolodchikov's $c$-theorem.
We may now define a function
\begin{equation}
c_{\mathrm{eff}}(s):= 4\pi^2\left( f(r)+\bar{f}(r)-\frac{1}{2}(g_1(r)+\bar{g}_1(r))-\frac{3}{8}q(r)\right),
\label{defcef}
\end{equation}
which satisfies the equation
\begin{equation}\label{dc}
\frac{dc_{\mathrm{eff}}}{ds}=-3\pi^2 q(r).
\end{equation}
The factor $(2\pi)^2$ in (\ref{defcef}) is introduced in order to reestablish the standard conformal normalization of the fields $\tau, \bar{\tau}$ and $\theta$ which, as shown in the paragraph before equation (\ref{7}), usually involves an extra factor $-2\pi$.
Consider the function $q(r)$. Its values for all $r>0$ may be obtained from the analytic continuation of the correlator $\langle L_0|\theta^{\mathsf R}(z,\b z)\theta^{\mathsf R}(0,0)|R_0\rangle$ to purely imaginary times $t=-iy$ (recall that $z=x-t$ and $\b z=x+t$). The inequality $q(r) \geq 0$ follows from reflection positivity \eqref{pos}. Thus, if $c_{\mathrm{eff}}(s)$ is complex, then its imaginary part is in fact independent of $s$, and its real part is strictly monotonically decreasing except when $q(r)$ is zero. By \eqref{pos}, it is clear that $q(r)$ vanishes for some $r$ if and only if it does so for all $r>0$, and this happens if and only if $\langle L_n|\theta|R_0\rangle=0$ for all $n$. In this case all vacuum correlation functions involving the trace of the stress-energy tensor vanish, and thus we may set it to zero. That is, either the real part of $c_\mathrm{eff}(s)$ is strictly monotonically decreasing as $s$ increases from $-\infty$ to $\infty$, or it is constant for all $s$ and $\theta(x,t)=0$.
As usual, we assume that the limits $\lim_{s\to\pm\infty}c_\mathrm{eff}(s)$ exist, and that these correspond to the UV and IR quantum critical points, at which scale invariance holds and thus the trace of the energy-momentum tensor vanishes. At these points, $g_1=\b g_1=0$. By analytic continuation to imaginary times, one shows by similar arguments as above that $f$ and $\b f$ are real and non-negative. Therefore, $c_\mathrm{eff}(\pm\infty)\geq 0$, and combining with \eqref{dc}, this implies that $c_\mathrm{eff}(s)\geq 0$
Thus we have established the existence of a function $c_{\mathrm{eff}}(s)$ of the renormalization group parameter $s$ which is non-negative and monotonically decreasing along renormalization group flows, with
\begin{equation}
c_\mathrm{eff}(-\infty) \geq c_\mathrm{eff}(\infty)
\end{equation}
(there is equality if and only if $\theta(x,t)=0$, in which case $c_\mathrm{eff}(s)$ is independent of $s$). This is irreversibility of the RG flow.
The only missing bit of the puzzle is the determination of the values $c_{\mathrm{eff}}(\pm\infty)$, at the IR and UV quantum critical points. From the definitions above these are given by the correlator $4\pi^2(f(r_*)+\bar{f}(r_*))$ at $r_*=0$ or $r_*=\infty$. In contrast to the unitary case, however, we do not immediately know these values as we do not have explicit expressions for the operators $\tau, \tau^\#$ within the standard CFT framework. We thus need to resort to a different strategy.
\medskip
\noindent {\em Remark.} In fact, even for unitary theories, the result above is slightly more general, as it does not assume that $f=\bar{f}$ and $g_1=\bar{g}_1$. Indeed, there exist CFTs (even unitary ones) where the left and right central charges may be different. In such cases $f\neq \bar{f}$ and $g_1\neq \bar{g}_1$. As expected even for those theories there is an irreversibility theorem, which, following the reasoning here, would be a ``$c+\bar{c}$-theorem" where the value of the scaling function at critical points would be $\frac{c+\bar{c}}{2}$. Having $c\neq \bar{c}$ means that the theory is not separately parity and time-reversal invariant, even if it is $\mathcal{PT}$-symmetric\footnote{Note that modular invariance implies that, even in cases where $c\neq \bar{c}$, the difference $c-\bar{c}$ is constrained to take values in $12\mathbb{Z}$ \cite{DMS}.}.
\medskip
\noindent {\em Remark.} Different choices of the functions $f, \b f, g_i, \b g_i$ and $q$ are possible. For instance, we could have used $f(z\b z) = z^4 \langle L_0|\tau^\#(z,\b z)\tau(0,0)|R_0\rangle$, etc., with the same result. These may simply be different monotonic functions along the RG flow. For the argument presented in the next section, the choice used here is more convenient.
\subsection{Connection with the CFT effective central charge}
In order to evaluate the values of $c_\mathrm{eff}(\pm\infty)$, we need to calculate the function $4\pi^2(f(r)+\bar{f}(r))$ in a CFT. The statement of scale invariance is the vanishing of the trace of the stress-energy tensor,
\begin{equation}
\theta(x,t)=0.
\end{equation}
From this alone, it is possible to conclude that chiral factorization occurs, and that $f(r)$ and $\b f(r)$ are constant. In order to emphasize that these arguments do not depend on unitarity, we repeat them briefly here.
First, equations \eqref{curb} and the similar relations for hash-fields imply that in any correlation function, $\tau$ and $\tau^\#$ are solely functions of $z$, and $\b\tau$ and $\b\tau^\#$ functions of $\b z$ (this is true, as usual, except at the space-time positions of other local or hash-local field insertions). Second, suppose a state $\langle\cdots\rangle$ is space-time translation invariant and clustering at large distance. Then,
\begin{eqnarray}
&& \langle \tau(x,0) \b\tau(x',0)\rangle
= \lim_{t\to\infty} \langle\tau(x,t) \b\tau(x',t)\rangle \nonumber\\
&& = \lim_{t\to\infty} \langle \tau(x-t,0) \b\tau(x'+t,0)\rangle
= \langle \tau(0,0)\rangle\,\langle \b\tau(0,0)\rangle.
\label{chiral}
\end{eqnarray}
This is (a part of) chiral factorization. Third, using the fact that $\langle L_0|\tau^{\mathsf R}(z,\b z)\tau^{\mathsf R}(0)|R_0\rangle$ (resp. $\langle L_0|\b \tau^{\mathsf R}(z,\b z)\b \tau^{\mathsf R}(0)|R_0\rangle$) only depends on $z$ (resp. $\b z$) for all $z\neq\b z$, and that the ground state is Lorentz invariant, the unique one-parameter solutions to \eqref{Btrans} are
\begin{equation}\label{gsres}
\langle L_0|\tau^{\mathsf R}(z,\b z)\tau^{\mathsf R}(0)|R_0\rangle = A z^{-4},
\quad
\langle L_0|\b \tau^{\mathsf R}(z,\b z)\tau^{\mathsf R}(0)|R_0\rangle = \b A\b z^{-4}
\end{equation}
for some constants $A$, $\b A$. This shows that $f(r)$ and $\b f(r)$ are indeed constants.
In the following, we will use the standard notation $\tau(z) = \tau(z,\b z)$ and $\b\tau(\b z) = \b \tau(z,\b z)$ (similarly for hashed fields) in order to emphasize chirality. We will also understand the variables $z$ and $\b z$ as complex variables (complex conjugate of each other), and use Euclidean, imaginary-time fields. Therefore $\tau(z)$ is holomorphic and $\b \tau(z)$ is anti-holomorphic, except at positions of other fields insertions, where singularities may occur.
Consider the partition function of a CFT at finite temperature $T=\beta^{-1}$ in a system of length $\ell$,
\begin{equation}
Z=\sum_n e^{-\beta E_n(\ell)}.
\end{equation}
It was shown in \cite{BCN,Affleck}, and then generalized to non-unitary models \cite{ceff} that, at least for minimal models of CFT, the specific free energy is given by:
\begin{equation} \label{Z}
\lim_{\ell \rightarrow \infty} \ell^{-1} \log Z=f_0\beta +\frac{\pi(c_{\mathrm{eff}}+\bar{c}_{\mathrm{eff}})}{12 \beta}
\end{equation}
where $f_0$ is an energy per unit length, and
\begin{equation}\label{ceffcft}
c_\mathrm{eff} = c-24\Delta,\quad \b c_\mathrm{eff} = \b c-24\b\Delta,
\end{equation}
with $c$ (resp. $\b c$) the holomorphic (resp. antiholomorphic) central charge and $\Delta$ (resp. $\b\Delta$) the lowest holomorphic (resp. antiholomorphic) dimension of the CFT.
In fact, in \cite{ceff} it was assumed that $c_{\mathrm{eff}}=\bar{c}_{\mathrm{eff}}$ as this holds for most CFTs, but the above is a simple generalization. More generally, relation \eqref{Z} is expected to hold simply based on scale invariance, and defines the quantity $c_{\mathrm{eff}}+\bar{c}_{\mathrm{eff}}$, which in non-compact models might or might not be determined by \eqref{ceffcft}.
For a non-unitary CFT we can write
\begin{equation}
Z=\sum_n \langle L_n | e^{-\beta H}| R_n\rangle.
\end{equation}
Differentiating with respect to $\beta$ twice we find
\begin{equation}
\frac{\partial^2 Z}{\partial \beta^2}=\langle H^2 \rangle_\beta^\mathrm{c}
\end{equation}
where $\langle AB\rangle_\beta^\mathrm{c} = \langle AB\rangle_\beta- \langle A\rangle_\beta\langle B\rangle_\beta$ and
\begin{equation}
\langle \mathcal{O} \rangle_\beta= Z^{-1}{\sum_n\langle L_n|\mathcal{O}e^{-\beta H}|R_n\rangle}.
\end{equation}
At criticality, we have $h = \tau + \b \tau$ and $p=\tau-\b\tau$, and thanks to chiral factorization \eqref{chiral},
\begin{eqnarray}
&& \langle (\tau(z)+ \bar{\tau}(\bar{z}))(\tau(z')+\bar{\tau}(\bar{z}')) \rangle_\beta-\langle \tau(z)+ \bar{\tau}(\bar{z})\rangle_\beta\, \langle \tau(z')+\bar{\tau}(\bar{z}')\rangle_\beta\nonumber\\
&&= \langle \tau(z) \tau(z')\rangle_\beta - \langle \tau\rangle^2_\beta + \langle \bar{\tau}(\bar{z}) \bar{\tau}(\bar{z}')\rangle_\beta - \langle \b\tau\rangle_\beta^2.
\end{eqnarray}
Therefore, defining $H_\pm = (H\pm P)/2$,
\begin{equation}
\langle H^2\rangle_\beta^\mathrm{c}
= \langle H_+^2\rangle_\beta^\mathrm{c}
+ \langle H_-^2\rangle_\beta^\mathrm{c}.
\end{equation}
Using the fact that ${H}^{\#}=H$ and $P^\#=P$ we may also write
\begin{equation}
\frac{\partial^2 Z}{\partial \beta^2}=
\langle H_+^{\mathsf R} H_+^{\mathsf R} \rangle_\beta^\mathrm{c}+
\langle H_-^{\mathsf R} H_-^{\mathsf R} \rangle_\beta^\mathrm{c}
\end{equation}
and so
\begin{eqnarray}
\frac{\partial^2}{\partial \beta^2} \left(\lim_{\ell \rightarrow \infty} \ell^{-1}\log Z\right)&=&\int dx \, \left(\langle {\tau}^{\mathsf R}(x,t+i\epsilon) \tau^{\mathsf R} (0,t)\rangle_\beta^\mathrm{c}+\langle \b\tau^{\mathsf R} (x,t+i\epsilon) \b\tau^{\mathsf R} (0,t)\rangle_\beta^\mathrm{c}\right)\nonumber\\
&=&\frac{\pi(c_{\mathrm{eff}}+\bar{c}_{\mathrm{eff}})}{6\beta^3}
\end{eqnarray}
where we shift the time variable slightly to ensure that operators are time-ordered.
In unitary CFT we may compute the correlators involved by identifying the CFT at finite temperature with a CFT on a cylinder of radius $\beta^{-1}$, and then employing the transformation properties of the energy-momentum tensor under a conformal map from the plane to the cylinder. However, once more, we have not identified correlation functions of our fields $\tau, \tau^\#, \bar{\tau}, \bar{\tau}^\#$ with those of the Virasoro-generated holomorphic and anti-holomorphic energy-momentum tensor of the standard formulation of CFT, hence we have not shown their transformation properites under conformal maps. We may instead use a more general QFT result, namely the Kubo-Martin-Schwinger relation \cite{kubo,MS}
\begin{equation} \label{kms}
\langle {\tau}^{{\mathsf R}}(z) \tau^{\mathsf R} ({z}')\rangle_\beta=\langle {\tau}^{\mathsf R} (z') \tau^{{\mathsf R}}(z-i\beta)\rangle_\beta.
\end{equation}
In order to go further, we need to argue that
\begin{equation}\label{comm}
\langle {\tau}^{\mathsf R} (z') \tau^{{\mathsf R}}(z)\rangle_\beta
=\langle \tau^{{\mathsf R}}(z){\tau}^{\mathsf R} (z')\rangle_\beta.
\end{equation}
Consider the function $F(z,\beta) = \langle \tau^{\mathsf R}(z)\tau^{\mathsf R}(0)\rangle_\beta$. It is analytic in a neighborhood of the line ${\rm Im}(z)=0$, except possibly at ${\rm Im}(z)=0$. On the one hand, for $z\approx 0$, by scaling one can use the ground-state result \eqref{gsres}, and thus there must be an isolated pole of order 4 at the origin. On the other hand, one can argue that the position of any other singularity, which might appear for instance due to the non-locality of $\tau^\#$, cannot depend on the temperature, as it is a property of the operators, not of the state. Therefore, by scaling again, no other singularity should exist on the real line. By imaginary-time ordering, for $z\in{\mathbb{R}}\setminus\{0\}$ we have $F(z+i0) = \langle \tau^{\mathsf R}(z)\tau^{\mathsf R}(0)\rangle_\beta$ and $F(z-i0) = \langle \tau^{\mathsf R}(0)\tau^{\mathsf R}(z)\rangle_\beta$. Since $F(z)$ is analytic, hence continuous, on ${\mathbb{R}}\setminus\{0\}$, this implies \eqref{comm} on $z-z'\in{\mathbb{R}}\setminus\{0\}$, and therefore for all $z,z'$ by analytic continuation. Note that relation \eqref{comm} points to the equivalence, at least from the viewpoint of the stress-energy tensor and at criticality, of locality and hash-locality.
Combining \eqref{kms} and \eqref{comm} we obtain the statement of periodicity for the analytic function $F(z,\beta) = \langle \tau^{\mathsf R}(z)\tau^{\mathsf R}(0)\rangle_\beta$:
\begin{equation}\label{per}
F(z,\beta) = F(z+i\beta,\beta).
\end{equation}
The function is in fact expected to be analytic within the full strip ${\rm Im}(z)\in(-\beta/2,\beta/2]$ except for the pole at $z=0$. By general results in one-dimensional models at nonzero temperature, the two point function vanishes exponentially at large distances. The unique family of solutions to the periodicity \eqref{per}, the requirement of a singularity \eqref{gsres} at $z=0$, and exponential vanishing at larges distances is
\begin{equation}
\langle {\tau}^{{\mathsf R}}(z) \tau^{{\mathsf R}}({z}')\rangle_\beta=\frac{\pi^4 A}{\beta^4 \sinh^4\frac{\pi (z-z')}{\beta}},
\end{equation}
where $A$ is a constant to be determined. Similarly:
\begin{equation}
\langle \bar{\tau}^{{\mathsf R}}(\bar{z}) \bar{\tau}^{{\mathsf R}}(\bar{z}')\rangle_\beta=\frac{\pi^4 \bar{A}}{\beta^4 \sinh^4\frac{\pi (\bar{z}-\bar{z'})}{\beta}}.
\end{equation}
Performing the integral
\begin{equation}
\frac{\pi^4}{\beta^4}\int dx \, \left(\frac{A}{\sinh^4\frac{\pi (x+i \epsilon)}{\beta}}+\frac{\bar{A}}{\sinh^4\frac{\pi (x-i\epsilon)}{\beta}}\right)=\frac{4\pi^3}{3\beta^3}(A+\bar{A}),
\end{equation}
and requiring that
\begin{equation}
\frac{4\pi^3}{3\beta^3}(A+\bar{A})=\frac{\pi(c_{\mathrm{eff}}+\bar{c}_{\mathrm{eff}})}{6\beta^3},
\end{equation}
we find
\begin{equation}
A+\b A=\frac{c_{\mathrm{eff}}+\bar{c}_{\mathrm{eff}}}{8\pi^2}.
\end{equation}
As argued earlier, we know that the function $c_{\mathrm{eff}}(s)$ defined in (\ref{defcef}) has the value $4\pi^2(f(r_{*})+\bar{f}(r_{*}))$ at conformal critical points characterized by a length scale $r_*$. The correlators above show that this values is nothing but $\frac{c_{\mathrm{eff}}+\bar{c}_{\mathrm{eff}}}{2}$. This completes our proof.
\medskip
\noindent {\em Remark.} The factor $8\pi^2$ arises once more from the fact that our normalization of the operators (\ref{operators}) is not the standard ``conformal normalization" that is used when defining the energy-momentum tensor. In the latter context one normally works with operators $\varepsilon:=-2\pi \tau $ and $\bar{\varepsilon}:=-2\pi \bar{\tau}$ and similarly for the hashed operators. Employing such operators we have then
that as $\beta\rightarrow \infty$ we recover the results for CFT on the plane. In such a case the conformal correlators become:
\begin{equation}
\langle L_0| {\varepsilon}^{{\mathsf R}}(z) \varepsilon^{\mathsf R}(0)|R_0 \rangle=\frac{c_{\mathrm{eff}}}{2 z^4} \qquad \mathrm{and} \qquad \langle L_0| \bar{\varepsilon}^{{\mathsf R}}(\bar{z}) \bar{\varepsilon}^{\mathsf R}(0)|R_0 \rangle=\frac{\bar{c}_{\mathrm{eff}}}{2 \bar{z}^4}.
\end{equation}
\section{Testing the $c_{\rm{eff}}$-theorem: some examples}\label{sectest}
In order to illustrate the $c_\mathrm{eff}$-theorem, we now discuss a few interesting examples of RG flows where the conditions of the theorem are met, and some where they are not.
Firstly, we would like to discuss a QFT studied in \cite{fonseca} whose lagrangian density is
\begin{equation}
\mathcal{L}_{FZ}=\psi\bar\partial{\psi}+\bar{\psi}\partial \bar{\psi}+ im \bar{\psi}\psi + ih \sigma,
\label{fon}
\end{equation}
where $\psi$, $\bar{\psi}$ are the chiral components of the Majorana free Fermion field, $\sigma$ is the corresponding spin field and $m,h\in \mathbb{R}$ with $m>0$. It was shown in \cite{fonseca} that the theory (\ref{fon}) displays an RG flow from the critical Ising to the critical Lee-Yang model, provided the ratio $\eta=m/|h|^{\frac{8}{15}}$ is fixed to a particular value. This describes the near-critical, universal region of a spin chain which was found earlier \cite{gehlen1} to be in the Ising criticality class at one point, and in the Lee-Yang class along a curve emanating from that point defined by an algebraic relation between the two coupling constants involved (corresponding to $m$ and $h$ in \eqref{fon}). Both in the QFT \eqref{fon} and in the quantum chain, this critical curve separates a $\mathcal{PT}$-broken phase, where some energy eigenvalues occur in complex conjugated pairs, from an unbroken phase, where all eigenvalues are real. Therefore, at the critical curve and more generally in the $\mathcal{PT}$-unbroken phase, the theory (\ref{fon}) is the type of non-unitary model where we expect the $c_{\rm{eff}}$-theorem to hold. The explicit $\mathcal{PT}$-symmetry of the lagrangian in this case is:
\begin{equation}
\sigma \mapsto -\sigma,\quad \psi\mapsto i\psi,\quad \b\psi \mapsto i\b\psi, \quad x\rightarrow -x,\quad i \rightarrow -i,
\end{equation}
which guarantees $\mathcal{PT}$-symmetry of the stress-energy tensor as per \eqref{PTass}. The flow \eqref{fon} on the critical curve satisfies the condition \eqref{order}, because $(c_\mathrm{eff})_{\rm UV}=0.5$ and $(c_\mathrm{eff})_{\rm IR}=0.4$, thus confirming $c_\mathrm{eff}$-theorem.
Further support for the existence of irreversible flows between non-unitary minimal models is the fact that, in some cases at least, it is possible to argue that they exhibit $\mathcal{PT}$-symmetry themselves. This may be shown by employing an effective Landau-Ginzburg description \cite{LGZ}. For the Lee-Yang minimal model \cite{Fisher}, the corresponding lagrangian density is
\begin{equation}\label{lyl}
\mathcal{L}_{\mathrm{LY}}=\frac{1}{2} (\partial_\mu \phi)^2+ im \phi^3
\end{equation}
A natural realization of $\mathcal{PT}$-symmetry in this case is the transformation
\begin{equation}
\phi \mapsto -\phi, \quad i\mapsto -i \quad \mathrm{and} \quad x\mapsto -x,
\label{ptre}
\end{equation}
under which the lagrangian is obviously invariant (this is a feature that is also well known from the study of the quantum-mechanical counterpart of this model \cite{BB}).
More generally, it is known that unitary minimal models can be described by Landau-Ginzburg lagrangians with potentials which are even polynomials $V(\phi)=V(-\phi)$ with real coefficients \cite{LGZ}, where $\mathcal{PT}$-symmetry in the sense of (\ref{ptre}) is also present. Unfortunately there is no known generic Landau-Ginzburg description of the non-unitary minimal models, as discussed in detail in \cite{amoroso}, even if the Lee-Yang case is well understood \cite{Fisher, LYCardy}. However, the presence of $\mathcal{PT}$-symmetry in all theories where the potential $V(\phi)$ either involves even powers of $\phi$ with real coefficients and/or odd powers of $\phi$ with imaginary coefficients, has been noted to hold quite generally, even in higher dimensions \cite{Codello}. Note that $\mathcal{PT}$-symmetry has been argued to be sufficient to ensure the stability of \eqref{lyl}, guaranteeing the spectrum to be real and bounded from below \cite{BBJ}.
Secondly, there are known RG flows where the condition (\ref{order}) is violated. Some of these examples have been discussed in \cite{nonmo1,nonmo2,nonmo25,nonmo3,nonmo4}. These examples would deserve more attention as they can only be reconciled with our result if some of the properties required for a $c_{\mathrm{eff}}$-theorem are not met. In such theories a decreasing monotonic function flowing between the UV and IR fixed points cannot exist. This could be explained in two possible ways: either $\mathcal{PT}$-symmetry is absent or broken, or the values of $c_{\rm{eff}}(s)$ at critical points do not coincide with the effective central charge as defined in (\ref{ceffcft}). The latter point is relevant because some of these theories have non-compact target space. This is associated with non-compact CFTs for which the result (\ref{Z}) is not always guaranteed to hold. However, in some cases it is hard to determine which of these two conditions is broken, in particular it is not easy to determine if $\mathcal{PT}$-symmetry is present or not.
An example where the situation is simpler is the sine-Gordon model with purely imaginary coupling \cite{nonmo1, nonmo2, nonmo25}. This describes a non-unitary RG flow between two critical points, both with $c=1$. The theory has lagrangian density of the form:
\begin{equation} \label{ish}
\mathcal{L}_{\mathrm{SG}}=\frac{1}{2} (\partial_\mu \phi)^2+ i\mu \cos\beta \phi,
\end{equation}
where $\beta, \mu \in \mathbb{R}$ are coupling constants and $\phi$ is a scalar field. Using $p=\beta^2/(8\pi-\beta^2)$, it is natural to restrict to $p\geq 2$ \cite{nonmo1, nonmo2, nonmo25}.
If we employ the $\mathcal{PT}$ transformation (\ref{ptre}) it is clear that the cosine term in the lagrangian is not invariant. However, there are other possible realizations of $\mathcal{PT}$-symmetry, such as
\begin{equation}\label{ptsG}
\phi \mapsto \frc{\pi}\beta-\phi, \quad i\mapsto -i \quad \mathrm{and} \quad x\mapsto -x,
\end{equation}
which preserves the lagrangian. Thus, the theory possesses dynamical $\mathcal{PT}$-symmetry as per \eqref{PTass}, yet the strict inequalty \eqref{order} is violated and thus the $c_\mathrm{eff}$-theorem does not hold. We speculate that $\mathcal{PT}$-symmetry is in fact broken in this case: some of the energy eigenvalues are complex.
This speculation is based on the following two observations. First, in \cite{nonmo2,nonmo25}, it was observed that, when a TBA analysis of the massless scattering matrix proposed to describe this model is performed, the wrong UV value of the central charge is obtained. For $p> 3$ this was assumed to be due to technical difficulties, but seen to be more fundamental for $2\leq p \leq 3$. At least in this region, the real-energy massless states associated to this $S$-matrix might not form the complete set of states necessary for a TBA analysis. Indeed, in \cite{nonmo25}, it was conjectured that a certain pole of the scattering matrix, for $2\leq p < 3$, should be associated to a ``monstron" particle whose mass $M_m=e^{i\pi(3-p)/4}M$ (where $M$ is the ``intercept scale" of the massless spectrum) has positive imaginary part. This leads to exponentially growing amplitudes. The monstron particle provides energy eigenvalues with nonzero imaginary parts, thus breaking of $\mathcal{PT}$-symmetry and making the $c_\mathrm{eff}$-theorem inapplicable. Beyond this range of $p$, for instance for $3<p<7$, a simple analytic continuation suggests that the monstron's mass gets a negative imaginary part. This would give rise to decaying amplitudes which supposedly have a vanishing influence on the TBA analysis. However, one may speculate that it still breaks $\mathcal{PT}$-symmetry, and thus again makes the $c_\mathrm{eff}$-theorem inapplicable. In fact, by $\mathcal{PT}$-invariance of the hamiltonian, an ``anti-monstron" should also be present with complex conjugated mass. The transformation \eqref{ptsG} maps minima of the potential to maxima and {\em vice versa}. Thus, in a scattering theory whose particle spectrum is built with respect to a given supremum, only the monstron is visible; yet for $\mathcal{PT}$-symmetry to be applicable, one also needs to consider the other suprema, and thus the anti-monstron.
Second, quantum-group restrictions of the theory \eqref{ish} are known to correctly reproduce massless flows between unitary minimal models, lending support to the massless scattering matrix proposed. Something similar happens in the context of quantum-group invariant open XXZ chains, where boundary terms break hermiticity. These chains in general can be expected to possess states with complex eigenvalues, representing gain and loss processes. In this context, it is known that at ``roots of unity", diagonalizability does not hold and Jordan blocks appear (so that biorthogonality is broken), which quantum-group restrictions heal \cite{AB,PS,JK,Korff} giving rise to minimal models. A possible scenario is that such Jordan blocks arise as conjugate pairs collapse into a single real eigenvalue (such collapses are called ``exceptional points"), indeed suggesting the presence, at generic parameters, of eigenvalues with nonzero imaginary parts.
Finally, we note that there are known scaling functions, distinct from $c_\mathrm{eff}(s)$, which flow between two conformal critical points satisfying the property (\ref{order}) but which do so in a non-monotonic fashion. All the examples we are aware of arise in the context of the thermodynamic Bethe ansatz approach, where a natural scaling function $c^{\mathrm{TBA}}_\mathrm{eff}(s)$ can be defined which in unitary theories is known to encapsulate the same information as Zamolodchikov's $c$-function. Examples of such scaling functions have been presented in Fig.~1 and briefly discussed in the introduction. Examples of non-monotonic scaling functions have been presented in \cite{dorey:2000zb}. Since these scaling functions are different from our $c_{\rm{eff}}$-function and they satisfy (\ref{order}) it is clear that they do not provide counterexamples to our theorem. Obviously there are infinitely many continuous functions flowing between two points that may be constructed, both monotonic and non-monotonic. Our claim is that the function defined by (\ref{defcef}) exists and is monotonic under certain conditions. The existence of other non-monotonic functions does not challenge this claim.
\section{Conclusions and Outlook}\label{secconclu}
In this letter we have shown that a function with all the properties of a $c$-function may be constructed for non-unitary 1+1-dimensional QFTs under certain conditions, including crucially {\it unbroken $\mathcal{PT}$-symmetry} as expressed in \eqref{PTP}, \eqref{PTass}, and the ensuing positivity of the spectrum. Thus, the requirement of unitarity, in the sense of a hermitian hamiltonian, is not necessary for the irreversibility of RG flows to hold. Besides unbroken $\mathcal{PT}$-symmetry, the properties we require (e.g. locality, Poincar\'e invariance etc.) are in fact very natural and routinely assumed to hold in QFT. Much of this paper is an effort to extricate some fundamental QFT concepts from unitarity and to show which properties are strictly necessary for the existence of a monotonic RG function and which are not. The resulting scaling function is monotonically decreasing along RG flows and it is constant at critical points where, at least in rational models of CFT, it takes the value $\frac{c_{\mathrm{eff}}+\bar{c}_{\mathrm{eff}}}{2}$. In parity symmetric critical points this is the usual effective central charge introduced in \cite{ceff}.
There exist massive perturbations of non-unitary minimal models of CFT. In these cases, the $c_\mathrm{eff}$-theorem provides an alternative understanding of the positivity of $c_\mathrm{eff}$. It also guarantees that there cannot exist limit cycles with varying $c_\mathrm{eff}$ in non-unitary RG flows with unbroken $\mathcal{PT}$-symmetry.
The proof relies on somewhat abstract considerations of the non-hermitian stress-energy tensor components $\tau, \bar{\tau}$ and $\theta$ acting on a {\em bona fide} Hilbert space. For practical purposes, a very important open question is how to explicitly construct these operators and their ${\,}^\#$ versions within the standard formulation of non-unitary CFTs (e.g. the non-unitary minimal series). In this standard formulation, the hamiltonian is hermitian, and one instead constructs the stress-energy tensor components in terms of generators of the Virasoro algebra of (possibly) negative central charge. These act on Verma modules, with an orthogonal basis (generated by the action of Virasoro operators on states created by primary fields) but with negative-norm states. Can we relate $\tau, \bar{\tau}$ and $\theta$ to such operators? Can we relate the states $|R_0\rangle$ and $\langle L_0|$ to the lowest-energy state $|\Delta\rangle, \langle \bar{\Delta}|$ of standard CFT formulations? These are questions we would like to address in the future.
We briefly discussed various examples where the assumptions and statement of the $c_\mathrm{eff}$-theorem can be checked, and an example where the statement does not hold, explaining how in this case $\mathcal{PT}$-symmetry appears to be broken. It would be important to provide more details on these examples, as well as other fully-worked cases, including a better understanding of the precise conditions of the $c_{\rm{eff}}$-theorem that are violated in those discussed in \cite{nonmo1,nonmo2,nonmo25,nonmo3,nonmo4}.
Another interesting question is how the entropic arguments used in \cite{Casinic} to prove Zamolodchikov's $c$-theorem can be generalized to give an alternative proof of the $c_{\rm{eff}}$-theorem.
\paragraph{Acknowledgement:} We would like to thank Patrick Dorey, Andreas Fring, Hubert Saleur, Germ\'an Sierra and Gerard Watts for useful discussions. We thank Hubert Saleur for bringing references \cite{nonmo2,nonmo3,nonmo4} to our attention and for sharing some private notes with us. We also thank Andreas Fring for bringing the special issue \cite{spePT} and reference \cite{bi0} to our attention.
Olalla Castro-Alvaredo and Benjamin Doyon are grateful to EPSRC for providing funding through the standard proposal ``Entanglement Measures, Twist Fields, and Partition Functions in Quantum Field Theory" under reference numbers EP/P006108/1 and EP/P006132/1. They also thank the Physics Department of the University of Bologna, INFN and Elisa Ercolessi for hospitality and financial support during an extended visit in November 2016.
Francesco Ravanini thanks INFN, in particular the Commission 4--Theory, for partial financial support through the grant GAST.
| d2cfe371f6ca7dff64c4e4a96084045bd6c742f7 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The goal of face alignment is to localize a set of predefined facial landmarks (eye corners, mouth corners etc.) in an image of a face. Face alignment is an important component of many computer vision applications, such as face verification \cite{DeepFace}, facial emotion recognition \cite{FacialAffect}, human-computer interaction \cite{HCI} and facial motion capture \cite{Animation}.
Most of the face alignment methods introduced in the recent years are based on shape indexed features \cite{SDM, LBF, ESR, CFSS, MDM}. In these approaches image features, such as SIFT \cite{SDM, CFSS} or learned features \cite{LBF, MDM}, are extracted from image patches extracted around each of the landmarks. The features are then used to iteratively refine the estimates of landmark locations. While those approaches can be successfully applied to face alignment in many photos, their performance on the most challenging datasets~\cite{300-W} leaves room for improvement~\cite{ESR, SDM, CFSS, MDM}. We believe that this is due to the fact that for the most difficult images the features extracted at disjoint patches do not provide enough information and can lead the method into a local minimum.
In this work, we address the above shortcoming by proposing a novel face alignment method which we dub Deep Alignment Network (DAN). It is based on a multi-stage neural network where each stage refines the landmark positions estimated at the previous stage, iteratively improving the landmark locations. The input to each stage of our algorithm (except the first stage) are a face image normalized to a canonical pose and an image learned from the dense layer of the previous stage. To make use of the entire face image during the process of face alignment, we additionally input at each stage a {\it landmark heatmap}, which is a key element of our system.
A landmark heatmap is an image with high intensity values around landmark locations where intensity decreases with the distance from the nearest landmark. The convolutional neural network can use the heatmaps to infer the current estimates of landmark locations in the image and thus refine them. An example of a landmark heatmap can be seen in Figure~\ref{fig:diagram1} which shows an outline of our method. By using landmark heatmaps, our DAN algorithm is able to reduce the failure rate on the 300W public test set by a large margin of 72\% with respect to the state of the art.
\begin{figure*}[tb]
\centering
\includegraphics[width=0.9\textwidth]{./scheme/scheme_general_v3.png}
\caption{A diagram showing an outline of the proposed method. Each stage of the neural network refines the landmark location estimates produced by the previous stage, starting with an initial estimate $S_0$. The connection layers form a link between the consecutive stages of the network by producing the landmark heatmaps $H_t$, feature images $F_t$ and a transform $T_t$ which is used to warp the input image to a canonical pose. By introducing landmark heatmaps and feature images we can transmit crucial information, including the landmark location estimates, between the stages of our method. }
\label{fig:diagram1}
\end{figure*}
To summarize, the three main contributions of this work are the following:
\begin{enumerate}
\item We introduce landmark heatmaps which transfer the information about current landmark location estimates between the stages of our method. This improvement allows our method to make use of the entire image of a face, instead of local patches, and avoid falling into local minima.
\item The resulting robust face alignment method we propose in this paper reduces the failure rate by 60\% on the 300W private test set \cite{300-W} and 72\% on the 300-W public test set \cite{300-W} compared to the state of the art.
\item Finally, we publish both the source code of our implementation of the proposed method and the models used in the experiments.
\end{enumerate}
The remainder of the paper is organized in the following manner. In section \ref{related} we give an overview of the related work. In section \ref{DAN} we provide a detailed description of the proposed method. Finally, in section \ref{experiments} we perform an evaluation of DAN and compare it to the state of the art.
\section{Related work} \label{related}
Face alignment has a long history, starting with the early Active Appearance Models \cite{AAM, AAM-revisited}, moving to Constrained Local Models \cite{CLM1, CLM2} and recently shifting to methods based on Cascaded Shape Regression (CSR) \cite{SDM, ESR, LBF, cGPRT, KRFWS, MIX, Trees} and deep learning \cite{Zhou2013, Fan2016, MDM, RAR, Bulat16}.
In CSR based methods, the face alignment begins with an initial estimate of the landmark locations which is then refined in an iterative manner. The initial shape $S_0$ is typically an average face shape placed in the bounding box returned by the face detector\cite{ESR, SDM, LBF, MDM}. Each CSR iteration is characterized by the following equation:
\begin{equation}
S_{t+1}=S_t + r_t(\phi(I, S_{t})),
\end{equation}
where $S_t$ is the estimate of landmark locations at iteration $t$, $r_t$ is a regression function which returns the update to $S_t$ given a feature $\phi$ extracted from image $I$ at the landmark locations.
The main differences between the variety of CSR based methods introduced in the literature lie in the choice of the feature extraction method $\phi$ and the regression method $r_t$. For instance, Supervised Descent Method (SDM) \cite{SDM} uses SIFT \cite{SIFT} features and a simple linear regressor. LBF \cite{LBF} takes advantage of sparse features generated from binary trees and intensity differences of individual pixels. LBF uses Support Vector Regression \cite{SVR} for regression which, combined with the sparse features, leads to a very efficient method running at up to 3000 fps.
Coarse to Fine Shape Searching (CFSS) \cite{CFSS}, similarly to SDM, uses SIFT features extracted at landmark locations. However the regression step of CSR is replaced with a search over the space of possible face shapes which goes from coarse to fine over several iterations. This reduces the probability of falling into a local minimum and thus improves convergence.
MIX \cite{MIX} also uses SIFT for feature extraction, while regression is performed using a mixture of experts, where each expert is specialized in a certain part of the space of face shapes. Moreover MIX, warps the input image before each iteration so that the current estimate of the face shape matches a predefined canonical face shape.
Mnemonic Descent Method (MDM) \cite{MDM} fuses the feature extraction and regression steps of CSR into a single Recurrent Neural Network that is trained end-to-end. MDM also introduces memory into the process which allows information to be passed between CSR iterations.
While all of the above mentioned methods perform face alignment based only on local patches, there are some methods \cite{Zhou2013,Fan2016} that estimate initial landmark positions using the entire face image and use local patches for refinement. In contrast, DAN localizes the landmarks based on the entire face image at all of its stages.
The use of heatmaps for face alignment related tasks precedes the proposed method. One method that uses heatmaps is \cite{Bulat16}, where a neural network outputs predictions in the form of a heatmap. In contrast, the proposed method uses heatmaps solely as a means for transferring information between stages.
The development of novel methods contributes greatly in advancing face alignment. However it cannot be overlooked that the publication of several large scale datasets of annotated face images \cite{300-W, Reannotation} also had a crucial role in both improving the state of the art and the comparability of face alignment methods.
\section{Deep Alignment Network} \label{DAN}
In this section, we describe our method, which we call the Deep Alignment Network (DAN). DAN is inspired by the Cascade Shape Regression (CSR) framework, just like CSR our method starts with an initial estimate of the face shape $S_0$ which is refined over several iterations. However, in DAN we substitute each CSR iteration with a single stage of a deep neural network which performs both feature extraction and regression. The major difference between DAN and approaches based on CSR is that DAN extracts features from the entire face image rather than the patches around landmark locations. This is achieved by introducing additional input to each stage, namely a landmark heatmap which indicates the current estimates of the landmark positions within the global face image and transmits this information between the stages of our algorithm. An outline of the proposed method is shown in Figure \ref{fig:diagram1}.
Therefore, each stage of DAN takes three inputs: the input image $I$ which has been warped so that the current landmark estimates are aligned with the canonical shape $S_0$, a landmark heatmap $H_t$ and a feature image $F_t$ which is generated from a dense layer connected to the penultimate layer of the previous stage $t-1$. The first stage only takes the input image as the initial landmarks are always assumed to be the average face shape $S_0$ located in the middle of the image.
A single stage of DAN consists of a feed-forward neural network which performs landmark location estimation and \textit{connection layers} that generate the input for the next stage. The details of the feed-forward network are described in subsection \ref{sec:feed-forward}. The connection layers consist of the Transform Estimation layer, the Image Transform layer, Landmark Transform layer, Heatmap Generation layer and Feature Generation layer. The structure of the connection layers is shown in Figure \ref{fig:diagramconnection}.
\begin{figure}[tb]
\centering
\includegraphics[width=0.9\linewidth]{./scheme/scheme_connection.png}
\caption{A diagram showing an outline of the connection layers. The landmark locations estimated by the current stage $S_{t}$ are first used to estimate the normalizing transform $T_{t+1}$ and its inverse $T^{-1}_{t+1}$. $T_{t+1}$ is subsequently used to transform the input image $I$ and $S_{t}$. The transformed shape $T_{t+1}(S_t)$ is then used to generate the landmark heatmap $H_{t+1}$. The feature image $F_{t+1}$ is generated using the fc1 dense layer of the current stage $t$.}
\label{fig:diagramconnection}
\end{figure}
The transform estimation layer generates the transform $T_{t+1}$, where $t$ is the number of the stage. The transformation is used to warp the input image $I$ and the current landmark estimates $S_{t}$ so that $S_{t}$ is close to the canonical shape $S_0$. The transformed landmarks $T_{t+1}(S_{t})$ are passed to the heatmap generation layer. The inverse transform $T^{-1}_{t+1}$ is used to map the output landmarks of the consecutive stage back into the original coordinate system.
The details of the Transform Estimation, Image Transform and Landmark Transforms layer are described in subsection \ref{sec:normalization}. The Heatmap Generation and Feature Image layers are described in sections \ref{sec:heatmap}, \ref{sec:feature}. Section \ref{sec:training} details the training procedure.
\begin{table}
\centering
\caption{ Structure of the feed-forward part of a Deep Alignment Network stage. The kernels are described as height $\times$ width $\times$ depth, stride.} \label{tab:feed-forward}
\begin{tabular}{| c | c | c | c |}
\hline
Name & Shape-in & Shape-out & Kernel \\
\hline
conv1a & 112$\times$112$\times$1 & 112$\times$112$\times$64 & 3$\times$3$\times$1,1\\
\hline
conv1b & 112$\times$112$\times$64 & 112$\times$112$\times$64 & 3$\times$3$\times$64,1\\
\hline
pool1 & 112$\times$112$\times$64 & 56$\times$56$\times$64 & 2$\times$2$\times$1,2\\
\hline
conv2a & 56$\times$56$\times$64 & 56$\times$56$\times$128 & 3$\times$3$\times$64,1\\
\hline
conv2b & 56$\times$56$\times$128 & 56$\times$56$\times$128 & 3$\times$3$\times$128,1\\
\hline
pool2 & 56$\times$56$\times$128 & 28$\times$28$\times$128 & 2$\times$2$\times$1,2\\
\hline
conv3a & 28$\times$28$\times$128 & 28$\times$28$\times$256 & 3$\times$3$\times$128,1\\
\hline
conv3b & 28$\times$28$\times$256 & 28$\times$28$\times$256 & 3$\times$3$\times$256,1\\
\hline
pool3 & 28$\times$28$\times$256 & 14$\times$14$\times$256 & 2$\times$2$\times$1,2\\
\hline
conv4a & 14$\times$14$\times$256 & 14$\times$14$\times$512 & 3$\times$3$\times$256,1\\
\hline
conv4b & 14$\times$14$\times$512 & 14$\times$14$\times$512 & 3$\times$3$\times$512,1\\
\hline
pool4 & 14$\times$14$\times$512 & 7$\times$7$\times$512 & 2$\times$2$\times$1,2\\
\hline
fc1 & 7$\times$7$\times$512 & 1$\times$1$\times$256 & -\\
\hline
fc2 & 1$\times$1$\times$256 & 1$\times$1$\times$136 & -\\
\hline
\end{tabular}
\end{table}
\subsection{Feed-forward neural network} \label{sec:feed-forward}
The structure of the feed-forward part of each stage is shown in Table \ref{tab:feed-forward}. With the exception of max pooling layers and the output layer, every layer takes advantage of batch normalization and uses Rectified Linear Units (ReLU) for activations. A dropout \cite{dropout} layer is added before the first fully connected layer. The last layer outputs the update $\Delta S_t$ to the current estimate of the landmark positions.
The overall shape of the feed-forwad network was inspired by the network used in \cite{VGG} for the ImageNet ILSVRC 2014 competition.
\begin{figure}
\centering
\includegraphics[width=0.24\linewidth]{./afterStage1/0004a.png}
\includegraphics[width=0.24\linewidth]{./afterStage1/0004b.png}
\includegraphics[width=0.24\linewidth]{./afterStage1/0004c.png}
\includegraphics[width=0.24\linewidth]{./afterStage1/0004d.png}
\includegraphics[width=0.24\linewidth]{./afterStage1/0087a.png}
\includegraphics[width=0.24\linewidth]{./afterStage1/0087b.png}
\includegraphics[width=0.24\linewidth]{./afterStage1/0087c.png}
\includegraphics[width=0.24\linewidth]{./afterStage1/0087d.png}
\includegraphics[width=0.24\linewidth]{./afterStage1/0123a.png}
\includegraphics[width=0.24\linewidth]{./afterStage1/0123b.png}
\includegraphics[width=0.24\linewidth]{./afterStage1/0123c.png}
\includegraphics[width=0.24\linewidth]{./afterStage1/0123d.png}
\caption{Selected images from the IBUG dataset and intermediate results after 1 stage of DAN. The columns show: the input image $I$, the input image normalized to canonical shape using transform $T_2$, the landmark heatmap showing $T_2(S_1)$, the corresponding feature image.}
\label{fig:afterStage1}
\end{figure}
\subsection{Normalization to canonical shape} \label{sec:normalization}
In DAN the input image $I$ is transformed for each stage so that the current estimates of the landmarks are aligned with the canonical face shape $S_0$. This normalization step allows the further stages of DAN to be invariant to a given family of transforms. This in turn simplifies the alignment task and improves accuracy.
The Transform Estimation layer of our network is responsible for estimating the parameters of transform $T_{t+1}$ at the output of stage $t$. As input the layer takes the output of the current stage $S_{t}$. Once $T_{t+1}$ is estimated the Image Transform and the Landmark Transform layers transform the image $I$ and landmarks $S_{t}$ to the canonical pose. The image is transformed using bilinear interpolation. Note that for the first stage of DAN the normalization step is not necessary since the input shape is always the average face shape $S_0$, which is also the canonical face shape.
Since the input image is transformed, the output of every stage has to be transformed back to match the original image, the output of any DAN stage is thus:
\begin{equation}
S_t = T^{-1}_t(T_t(S_{t-1}) + \Delta S_t),
\end{equation}
where $\Delta S_t$ is the output of the last layer of stage $t$ and $T^{-1}_t$ is the inverse of transform $T_t$
A similar normalization step has been previously proposed in \cite{MIX} with the use of affine transforms. In our implementation we chose to use similarity transforms as they do not cause non-uniform scaling and skewing of the output image. Figure \ref{fig:afterStage1} shows examples of images before and after the transformation.
\subsection{Landmark heatmap} \label{sec:heatmap}
The landmark heatmap is an image where the intensity is highest in the locations of landmarks and it decreases with the distance to the closest landmark. Thanks to the use of landmark heatmaps the Convolutional Neural Network can infer the landmark locations estimated by the previous stage. In consequence DAN can perform face alignment based on entire facial images.
At the input to a DAN stage the landmark heatmap is created based on the landmark estimates produced by the previous stage and transformed to the canonical pose: $T_t(S_{t-1})$. The heatmap is generated using the following equation:
\begin{equation}
H(x,y)=\frac{1}{1 + \min_{s_i \in T_t(S_{t-1})} ||(x, y) - s_i||},
\end{equation}
where $H$ is the heatmap image and $s_i$ is the i-th landmark of $T_t(S_{t-1})$. In our implementation the heatmap values are only calculated in a circle of radius 16 around each landmark to improve performance. Note that similarly to normalization, this step is not necessary at the input of the first stage, since the input shape is always assumed to be $S_0$, which would result in an identical heatmap for any input.
An example of a face image and a corresponding landmark heatmap is shown in Figure \ref{fig:afterStage1}.
\subsection{Feature image layer} \label{sec:feature}
The feature image layer $F_t$ is an image created from a dense layer connected to the fc1 layer (see Table \ref{tab:feed-forward}) of the previous stage $t-1$. Such a connection allows any information learned by the preceding stage to be transferred to the consecutive stage. This naturally complements the heatmap which transfers the knowledge about landmark locations learned by the previous stage.
The feature image layer is a dense layer which has 3136 units with ReLU activations. The output of this dense layer is reshaped to a 56$\times$56 2D layer and upscaled to 112$\times$112, which is the input shape of DAN stages. We use the smaller 56$\times$56 image rather than 112$\times$112 since it showed similar results in our experiments, with considerably less parameters. Figure \ref{fig:afterStage1} shows an example of a feature image.
\subsection{Training procedure} \label{sec:training}
The stages of DAN are trained sequentially. The first stage is trained by itself until the validation error stops improving. Subsequently the connection layers and the second stage are added and trained. This procedure is repeated until further stages stop reducing the validation error.
While many face alignment methods \cite{ESR, SDM, LBF} learn a model that minimizes the Sum of Squared Errors of the landmark locations, DAN minimizes the landmark location error normalized by the distance between the pupils:
\begin{equation} \label{interpupil}
\min_{\Delta S_t} \frac{||T^{-1}_t(T_t(S_{t-1}) + \Delta S_t) - S^*||}{d_{ipd}},
\end{equation}
where $S^*$ is a vector of ground truth landmark locations, $T_t$ is the transform that normalizes the input image and shape for stage $t$ and $d_{ipd}$ is the distance between the pupils of $S^*$. The use of this error is motivated by the fact that it is a far more common \cite{SDM,LBF,CFSS} benchmark for face alignment methods than the Sum of Squared Errors.
Thanks to the fact that all of the layers used in DAN are differentiable DAN can also be trained end-to-end. In order to evaluate end-to-end training in DAN we have experimented with several approaches. Pre-training the first stage for several epochs followed by training of the entire network yielded similar accuracy to the proposed approach but the training was significantly longer. Training the entire network from scratch yielded results significantly inferior to the proposed approach.
While we did not manage to obtain improved results with end-to-end training we believe that it is possible with a better training strategy. We leave the creation of such a strategy for future work.
\section{Experiments} \label{experiments}
In this section we perform an extensive evaluation of the proposed method on several public datasets as well as the test set of the Menpo challenge \cite{Menpo} to which we have submitted our method. The following paragraphs detail the datasets, error measures and implementation. Section \ref{sec:comparison} compares our method with the state of the art, while section \ref{sec:menpo} shows our results in the Menpo challenge. Section \ref{sec:further} discusses the influence of the number of stages on the performance of DAN.
\paragraph*{Datasets}
In order to evaluate our method we perform experiments on the data released for the 300W competition \cite{300-W} and the recently introduced Menpo challenge \cite{Menpo} dataset.
The 300W competition data is a compilation of images from five datasets: LFPW \cite{LFPW}, HELEN \cite{HELEN}, AFW \cite{AFW}, IBUG \cite{300-W} and 300W private test set \cite{300-W}. The last dataset was originally used for evaluating competition entries and at that time was private to the organizers of the competition, hence the name. Each image in the dataset is annotated with 68 landmarks \cite{Reannotation} and accompanied by a bounding box generated by a face detector. We follow the most established approach \cite{LBF,CFSS,KRFWS,RAR} and divide the 300-W competition data into training and testing parts. The training part consists of the AFW dataset as well as training subsets of LFPW and HELEN, which results in a total of 3148 images. The test data consists of the remaining datasets: IBUG, 300W private test set, test sets of LFPW, HELEN. In order to facilitate comparison with previous methods we split this test data into four subsets:
\begin{itemize}
\item the \textit{common subset} which consists of the test subsets of LFPW and HELEN (554 images),
\item the \textit{challenging subset} which consists of the IBUG dataset (135 images),
\item the \textit{300W public test set} which consists of the test subsets of LFPW and HELEN as well as the IBUG dataset (689 images),
\item the \textit{300W private test set} (600 images).
\end{itemize}
The annotation for the images in the 300W public test set were originally published for the 300W competition as part of its training set. We use them for testing as it became a common practice to do so in the recent years \cite{LBF, CFSS, RAR, cGPRT, KRFWS}.
The Menpo challenge dataset consists of semi-frontal and profile face image datasets. In our experiments we only use the semi-frontal dataset. The dataset consists of training and testing subsets containing 6679 and 5335 images respectively. The training subset consists of images from the FDDB \cite{FDDB} and AFLW \cite{AFLW} datasets. The image were annotated with the same set of 68 landmarks as the 300W competition data but no face detector bounding boxes. The annotations of the test subset have not been released.
\paragraph*{Error measures}
Several measures of face alignment error for an individual face image have been recently introduced:
\begin{itemize}
\item the mean distance between the localized landmarks and the ground truth landmarks divided by the inter-ocular distance (the distance between the outer eye corners) \cite{CFSS, LBF, RAR},
\item the mean distance between the localized landmarks and the ground truth landmarks divided by the inter-pupil distance (the distance between the eye centers) \cite{MDM, 300-W},
\item the mean distance between the localized landmarks and the ground truth landmarks divided by the diagonal of the bounding box \cite{Menpo}.
\end{itemize}
In our work, we report our results using all of the above measures.
For evaluating our method on the test datasets we use three metrics: the mean error, the area under the cumulative error distribution curve (AUC$_\alpha$) and the failure rate.
Similarly to \cite{MIX, MDM}, we calculate AUC$_\alpha$ as the area under the cumulative distribution curve calculated up to a threshold $\alpha$, then divided by that threshold. As a result the range of the AUC$_\alpha$ values is always 0 to 1. Following \cite{MDM}, we consider each image with an inter-ocular normalized error of 0.08 or greater as failure and use the same threshold for AUC$_{0.08}$. In all the experiments we test on the full set of 68 landmarks.
\paragraph*{Implementation}
We train two models, DAN which is trained on the training subset of the 300W competition data and DAN-Menpo which is trained on both the above mentioned dataset and the Menpo challenge training set. Data augmentation is performed by mirroring around the Y axis as well as random translation, rotation and scaling, all sampled from normal distributions. During data augmentation a total of 10 images are created from each input image in the training set.
Both models (DAN and DAN-Menpo) consist of two stages. Training is performed using Theano 0.9.0 \cite{Theano} and Lasagne 0.2 \cite{Lasagne}. For optimization we use Adam stochastic optimization \cite{Adam} with an initial step size of 0.001 and mini batch size of 64. For validation we use a random subset of 100 images from the training set.
The Python implementation runs at 73 fps for images processed in parallel and at 45 fps for images processed sequentially on a GeForce GTX 1070 GPU. We believe that the processing speed can be further improved by optimizing the implementation of some of our custom layers, most notably the Image Transform layer.
To enable reproducible research, we release the source code of our implementation as well as the models used in the experiments\footnote{\url{https://github.com/MarekKowalski/DeepAlignmentNetwork}}. The published implementation also contains an example of face tracking with the proposed method.
\subsection{Comparison with state-of-the-art} \label{sec:comparison}
We compare the DAN model with state-of-the-art methods on all of the test sets of the 300W competition data. We also show results for the DAN-Menpo model but do not perform comparison since at the moment there are no published methods that use this dataset for training. For each test set we initialize our method using the face detector bounding boxes provided with the datasets.
Tables \ref{tab:public1} and \ref{tab:public2} show the mean error, AUC$_{0.08}$ and the failure rate of the proposed method and other methods on the \textit{300W public test set}. Table \ref{tab:private} shows the mean error, the AUC$_{0.08}$ and failure rate on the \textit{300W private test set}.
All of the experiments performed on the two most difficult test subsets (the \textit{challenging subset} and the \textit{300W private test set}) show state-of-the-art results, including:
\begin{itemize}
\item a failure rate reduction of 60\% on the \textit{300W private test set},
\item a failure rate reduction of 72\% on the \textit{300W public test set},
\item a 9\% improvement of the mean error on the \textit{challenging subset}.
\end{itemize}
This shows that the proposed DAN is particularly suited for handling difficult face images with a high degree of occlusion and variation in pose and illumination.
\begin{table}[tb]
\caption{Mean error of face alignment methods on the 300W public test set and its subsets. All values are shown as percentage of the normalization metric.} \label{tab:public1}
\begin{tabularx}{\linewidth}{ >{\centering\arraybackslash}X c c c c }
\Xhline{4\arrayrulewidth}
Method & \makecell{Common \\ subset} & \makecell{Challenging \\ subset} & Full set \\
\hline
\multicolumn{4}{c}{inter-pupil normalization} \\
\hline
ESR \cite{ESR} & 5.28 & 17.00 & 7.58\\
SDM \cite{SDM} & 5.60 & 15.40 & 7.52\\
LBF \cite{LBF} & 4.95 & 11.98 & 6.32\\
cGPRT \cite{cGPRT} & - & - & 5.71\\
CFSS \cite{CFSS} & 4.73 & 9.98 & 5.76\\
Kowalski et al. \cite{KRFWS} & 4.62 & 9.48 & 5.57\\
RAR \cite{RAR} & \textbf{4.12} & 8.35 & 4.94\\
\hline
\textbf{DAN} & 4.42& \textbf{7.57} & 5.03 & \\
\textbf{DAN-Menpo} & 4.29 & \textbf{7.05} & \textbf{4.83} & \\
\hline
\multicolumn{4}{c}{inter-ocular normalization} \\
\hline
MDM \cite{MDM} & - & - & 4.05\\
Kowalski et al. \cite{KRFWS} & 3.34 & 6.56 & 3.97\\
\hline
\textbf{DAN} & \textbf{3.19} & \textbf{5.24} & \textbf{3.59} & \\
\textbf{DAN-Menpo} & \textbf{3.09} & \textbf{4.88} & \textbf{3.44} & \\
\hline
\multicolumn{4}{c}{bounding box diagonal normalization} \\
\hline
\textbf{DAN} & \textbf{1.35} & \textbf{2.00} & \textbf{1.48} & \\
\textbf{DAN-Menpo} & \textbf{1.31} & \textbf{1.87} & \textbf{1.42} & \\
\Xhline{4\arrayrulewidth}
\end{tabularx}
\end{table}
\begin{table}[tb]
\caption{AUC and failure rate of face alignment methods on the 300W public test set. } \label{tab:public2}
\begin{tabularx}{\linewidth}{ >{\centering\arraybackslash}X c c c c }
\Xhline{4\arrayrulewidth}
Method & AUC$_{0.08}$ & Failure (\%) \\
\hline
\multicolumn{3}{c}{inter-ocular normalization} \\
\hline
ESR \cite{ESR} & 43.12 & 10.45\\
SDM \cite{SDM} & 42.94 & 10.89\\
CFSS \cite{CFSS} & 49.87 & 5.08\\
MDM \cite{MDM} & 52.12 & 4.21\\
\hline
\textbf{DAN} & \textbf{55.33} & \textbf{1.16} & \\
\textbf{DAN-Menpo} & \textbf{57.07} & \textbf{0.58} & \\
\Xhline{4\arrayrulewidth}
\end{tabularx}
\end{table}
\begin{table}[tb]
\caption{Results of face alignment methods on the 300W private test set. Mean error is shown as percentage of the inter-ocular distance.} \label{tab:private}
\begin{tabularx}{\linewidth}{ >{\centering\arraybackslash}X c c c c }
\Xhline{4\arrayrulewidth}
Method & Mean error & AUC$_{0.08}$ & Failure (\%) \\
\hline
\multicolumn{4}{c}{inter-ocular normalization} \\
\hline
ESR \cite{ESR} & - & 32.35 & 17.00\\
CFSS \cite{CFSS} & - & 39.81 & 12.30\\
MDM \cite{MDM} & 5.05 & 45.32 & 6.80\\
\hline
\textbf{DAN} & \textbf{4.30} & \textbf{47.00} & \textbf{2.67} & \\
\textbf{DAN-Menpo} & \textbf{3.97} & \textbf{50.84} & \textbf{1.83} & \\
\Xhline{4\arrayrulewidth}
\end{tabularx}
\end{table}
\begin{figure}[!htb]
\centering
\includegraphics[width=0.325\linewidth]{./ibug_imgs/121.jpg}
\includegraphics[width=0.325\linewidth]{./ibug_imgs/035.jpg}
\includegraphics[width=0.325\linewidth]{./ibug_imgs/011.jpg}
\includegraphics[width=0.325\linewidth]{./ibug_imgs/101.jpg}
\includegraphics[width=0.325\linewidth]{./ibug_imgs/045.jpg}
\includegraphics[width=0.325\linewidth]{./ibug_imgs/034.jpg}
\includegraphics[width=0.325\linewidth]{./ibug_imgs/041.jpg}
\includegraphics[width=0.325\linewidth]{./ibug_imgs/122.jpg}
\includegraphics[width=0.325\linewidth]{./ibug_imgs/127.jpg}
\caption{ The 9 \textit{worst} results on the challenging subset (IBUG dataset) in terms of inter-ocular error produced by the DAN model. Only the first 7 images have an error of more than 0.08 inter-ocular distance and can be considered failures. }
\label{fig:ibug_worst}
\end{figure}
\begin{figure*}[!htb]
\centering
\includegraphics[width=0.16\textwidth]{./menpo_examples/10670.png}
\includegraphics[width=0.16\textwidth]{./menpo_examples/11251.png}
\includegraphics[width=0.16\textwidth]{./menpo_examples/11914.png}
\includegraphics[width=0.16\textwidth]{./menpo_examples/12268.png}
\includegraphics[width=0.16\textwidth]{./menpo_examples/3398.png}
\includegraphics[width=0.16\textwidth]{./menpo_examples/5876.png}
\includegraphics[width=0.16\textwidth]{./menpo_examples/3456.png}
\includegraphics[width=0.16\textwidth]{./menpo_examples/6310.png}
\includegraphics[width=0.16\textwidth]{./menpo_examples/7218.png}
\includegraphics[width=0.16\textwidth]{./menpo_examples/7220.png}
\includegraphics[width=0.16\textwidth]{./menpo_examples/9385.png}
\includegraphics[width=0.16\textwidth]{./menpo_examples/9794.png}
\caption{ Results of our submission to the Menpo challenge on some of the difficult images of the Menpo test set. The blue squares denote the initialization bounding boxes. The images were cropped to better visualize the results, in the original images the bounding boxes are always located in the center. } \label{fig:menpo_examples}
\end{figure*}
\subsection{Results on the Menpo challenge test set} \label{sec:menpo}
In order to evaluate the proposed method on the Menpo challenge test dataset we have submitted our results to the challenge and received the error scores from the challenge organizers. The Menpo test data differs from the other datasets we used in that it does not include any bounding boxes which could be used to initialize face alignment. For that reason we have decided to use a two step face alignment procedure, where the first step serves as an initialization for the second step.
The first step performs face alignment using a square initialization bounding box placed in the middle of the image with a size set to a percentage of image height. The second step takes the result of the first step, transforms the landmarks and the image to the canonical face shape and creates a bounding box around the transformed landmarks. The transformed image and bounding box are used as input to face alignment. An inverse transform is later applied to get landmark coordinates for the original image.
In order to determine the optimal size of the bounding boxes in the first step we ran DAN on a small subset of the Menpo test set for several bounding box sizes. The optimal size was determined using a method that would estimate the face alignment error of a given set of landmarks and an image. Said method extracts HOG \cite{HOG} features at each of the landmarks and uses a linear model to estimate the error. The method was trained on the 300W training set using ridge regression. The chosen bounding box size was 46\% of the image height.
Figure \ref{fig:Menpo} and Table \ref{tab:Menpo} show the CED curve, mean error, $AUC_{0.03}$ and failure rate for the DAN-Menpo model on the Menpo test set. In all cases the errors are calculated using the diagonal of the bounding box normalization, used by the challenge organizers. For the AUC and the failure rate we have chosen a threshold of 0.03 of the bounding box diagonal as it is approximately equivalent to 0.08 of the interocular distance used in the previous chapter.
Figure \ref{fig:menpo_examples} shows examples of images from the Menpo test set and corresponding results produced by our method. Note that even though DAN was trained primarily on semi-frontal images it can handle fully profile images as well.
\begin{figure}[!htb]
\centering
\includegraphics[width=\linewidth]{./ced/ced.png}
\caption{The Cumulative Error Distribution curve for the DAN-Menpo model on the Menpo test set. The Point-to-Point error is shown as percentage of the bounding box diagonal.}
\label{fig:Menpo}
\end{figure}
\begin{table}[tb]
\caption{Results of the proposed method on the semi-frontal subset of the Menpo test set. Mean error is shown as percentage of the bounding box diagonal.} \label{tab:Menpo}
\begin{tabularx}{\linewidth}{ >{\centering\arraybackslash}X c c c c }
\Xhline{4\arrayrulewidth}
Method & Mean error & AUC$_{0.03}$ & Failure (\%) \\
\hline
\multicolumn{4}{c}{bounding box diagonal normalization} \\
\hline
\textbf{DAN-Menpo} & \textbf{1.38} & \textbf{56.20} & \textbf{1.74} & \\
\Xhline{4\arrayrulewidth}
\end{tabularx}
\end{table}
\subsection{Further evaluation} \label{sec:further}
In this subsection we evaluate several DAN models with a varying number of stages on the 300W private test set. All of the models were trained identically to the DAN model from section \ref{sec:comparison}.
Table \ref{tab:stage_number} shows the results of our evaluation. The addition of the second stage increases the AUC$_{0.08}$ by 20\% while the mean error and failure rate are reduced by 14\% and 56\% respectively. The addition of a third stage does not bring significant benefit in any of the metrics.
\begin{table}[tb]
\caption{Results of the proposed method with a varying number of stages on the 300W private test set. Mean error is shown as percentage of the inter-ocular distance.} \label{tab:stage_number}
\begin{tabularx}{\linewidth}{ >{\centering\arraybackslash}X c c c c }
\Xhline{4\arrayrulewidth}
\# of stages & Mean error & AUC$_{0.08}$ & Failure (\%) \\
\hline
\multicolumn{4}{c}{inter-ocular normalization} \\
\hline
1 & 5.02 & 39.04 & 6.17 & \\
2 & 4.30 & 47.00 & 2.67 & \\
3 & 4.32 & 47.08 & 2.67 & \\
\Xhline{4\arrayrulewidth}
\end{tabularx}
\end{table}
\section{Conclusions}
In this paper, we introduced the Deep Alignment Network - a robust face alignment method based on convolutional neural networks. Contrary to the recently proposed face alignment methods, DAN performs face alignment based on entire face images, which makes it highly robust to large variations in both initialization and head pose. Using entire face images instead of local patches extracted around landmarks is possible thanks to the use of novel landmark heatmaps which transmit the information about landmark locations between DAN stages. Extensive evaluation performed on two challenging, publicly available datasets shows that the proposed method improves the state-of-the-art failure rate by a significant margin of over 70\%.
Future research includes investigation of new strategies for training DAN in an end-to-end manner. We also plan to introduce learning into the estimation of the transform $T_t$ that normalizes the shapes and images between stages.
\section{Acknowledgements}
The work presented in this article was supported by The National Centre for Research and Development grant number DOB-BIO7/18/02/2015. The results obtained in this work were a basis for developing the software for the grant sponsor, which is different from the software published with this paper.
We thank NVIDIA for donating a Titan X Pascal GPU which was used to train the proposed neural network.
{\small
\bibliographystyle{ieee}
| b223c78e2bef75b9996445a2e3b660726b3b7486 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The radiation fields, winds and SNe of massive stars destroy and
disperse molecular material. This eventually ends star formation in
clusters, though before that happens, massive stars may trigger
further star formation \citep[e.g.][]{koenig12}. The removal of mass
from a cluster affects the cluster dynamics and plays a key role in
cluster dissolution \citep{portegiesZwart10}. Stellar feedback also
sustains turbulence in the ISM \citep[e.g.][]{dobbs11} and powers
galactic fountains and winds \citep{veilleux05}. The mass, momentum,
energy and ionizing radiation fluxes escaping a cluster into a galaxy,
and beyond, depend on how stellar and supernova energy output
regulates cluster gas. However, the coupling of stellar winds, SNe and
ionizing radiation to clumpy, inhomogeneous molecular clouds
surrounding a massive stellar cluster is not well constrained.
Detections of diffuse X-ray emission from many young (pre-SN) massive
star forming regions support the conjecture that winds play important
roles in clusters. Cold molecular material sometimes confines X-ray
emitting gas, but around other clusters, hot gas appears to shape and
remove cold clouds \citep[e.g.][]{townsley14}. Direct evidence for large
scale outflows from stellar clusters is provided by observations of
stellar bowshocks in and near clusters
\citep[e.g.][]{winston12}. ``Leakage'' of the hot gas is also implied by
the much lower mass of hot cluster gas than expected for the cluster
ages and the mass-loss rates of stars \citep{townsley03}, and by
energy budget considerations \citep{rosen14}. This leakage reduces the
thermal pressure within the hot bubble enough that H{\sc ii} gas
pressure may drive the dynamics \citep{harper09}, at least for
molecular clouds of $\sim 10^{5}$\,M$_\odot$
\citep{walch12,dale12}. Simulations of momentum-driven or isothermal
winds \citep{dale15,offner15} give a lower limit to their impact.
Pre-SN feedback enhances the impact of SNe \citep{fierlinger16},
and, in whole galaxy models, clears dense gas from star forming
regions, reducing the star formation rate
\citep[e.g.][]{agertz13}. However, whole galaxy models remain very
sensitive to assumptions in the feedback scheme.
In our previous work \citep{rogers13,rogers14,wareing17} we have
examined the effect that winds and SNe have on surrounding molecular
material. In \citet[][hereafter Paper II]{wareing17}, we explored the
effect of mechanical feedback from a single 15\,M$_\odot$ star and a
single 40\,M$_\odot$ star into a sheet-like molecular cloud formed by
the action of the thermal instability under the influence of magnetic
fields, that in projection appears remarkably filamentary. That cloud
was one of three cases studied in \citet[][hereafter Paper
I]{wareing16} which explored the formation of dense, cold, molecular
clouds from quiescent diffuse thermally unstable clouds, under the
influence of the thermal instability. There we considered the
hydrodynamic case, the case of equal magnetic and thermal pressures
and the case of dominating magnetic pressure (10$\times$ greater than
thermal pressure). Paper II explored feedback in the case of equal
pressures. Here, we now explore the hydrodynamic case of feedback into
a roughly spherical clumpy molecular cloud. We have presented a review
of the relevant literature in Paper II and refer the interested reader
back to that work.
In the next section, we present our numerical method and define the
initial conditions used in our model, as well as the basis for
realistic input of mechanical energy from each star. In Section
\ref{results} we present and discuss the resulting simulations,
organised by results for the wind phase, the early SN phase and the
late SN phase, rather than by each star. In Section \ref{analysis} we
present the global evolution with time of energy, density, phase,
mass-weighted temperature-density and pressure-density distribution,
total mass and mixing behaviour in each of our simulations. In Section
\ref{discussion} we compare our results to previous works and relevant
observations. We summarise and conclude the work in Section
\ref{conclusions}.
\section{Numerical Methods and initial conditions}\label{numerical}
\subsection{Numerical methods}
We present 3D, hydrodynamical (HD) simulations of stellar feedback
with self-gravity using the established astrophysical code MG
\citep{falle91} as recently used in Papers I and II. The code employs
an upwind, conservative shock-capturing scheme and is able to employ
multiple processors through parallelisation with the message passing
interface (MPI) library. MG uses piece-wise linear cell interpolation
to solve the Eulerian equations of hydrodynamics. The Riemann problem
is solved at cell interfaces to obtain the conserved fluxes for the
time update. Integration in time proceeds according to a second-order
accurate Godunov method \citep{godunov59}. A Kurganov Tadmor Riemann
solver is again used in this work \citep{kurg00}. Self-gravity is
computed using a full-approximation multigrid to solve the Poisson
equation. We reduce the Magneto-HD code to HD by setting all the
magnetic field components to zero. For full details of the AMR method,
please see Papers I and II. The simulations presented below employed 8
levels of AMR. The physical size of the computational volume and the
physical resolutions are as detailed below. Continued care has been
taken with the implementation of realistic heating and cooling in the
same way as used in Papers I and II.
\subsection{Initial conditions}
\begin{figure*}
\centering
\includegraphics[width=150mm]{figure1.pdf}
\caption{Initial condition. Snapshot of a clumpy cloud after 27.1\,Myrs
of evolution used as the initial condition in this work. Shown are
logarithm of mass density on planes (a) x=-0.025, (b) y=-0.2351, (c) z=-0.11
and projected column density along the x-axis in (d).
Length is scaled in units of 50\,pc. Raw data: doi.org/10.5518/201.}
\label{figure-initial}
\end{figure*}
As well as creating filamentary molecular clouds through the effect of
background magnetic fields and the influence of the thermal
instability \citep{parker53,field65}, Paper I also explored the
formation of clumpy molecular clouds in the zero-field case. In that
work, we examined the evolution of diffuse clouds varying $\beta$, the
ratio of thermal pressure to magnetic pressure. We examined three
cases, $\beta$ = 0.1, $\beta$ = 1.0 and $\beta$ = $\infty$, equivalent
to the hydrodynamic case of zero magnetic field and of further
interest here. In Paper I, the initial condition consisted of a
stationary cloud of radius 50\,pc with a number density of hydrogen
throughout the cloud of n$_H$ = 1.1 cm$^{-3}$ giving the cloud a total
mass of $\sim$16,500\,M$_\odot$. In the cloud, 10\% density variations
about the uniform initial density were introduced. The pressure was
set according to the local density and thermal equilibrium between
heating and cooling prescriptions at P$_{eq}$/k = $4700\pm300$
K\,cm$^{-3}$, resulting in an initial temperature T$_{eq}$ =
$4300\pm700$\,K (an unstable part of the equilibrium curve - for more
details see Paper I). An advecting scalar, $\alpha_{cloud}$, was set equal
to 1 in the cloud material. The pressure of the lower density (n$_H$ = 0.1
cm$^{-3}$) surroundings was set equal to that of the unperturbed
cloud, with $\alpha_{cloud}$ set equal to zero.
No magnetic field was present in the simulation.
The computational volume consists of a 150\,pc$^3$ box with free-flow
boundary conditions (non-periodic for self-gravity) and AMR level G0
containing 4$^3$ cells. Eight levels of AMR mean an effective
resolution of 512$^3$ cells on level G7, although by the time of the
initial condition G2 with 16$^3$ cells is the finest fully populated
level (rather than the MG default of G1). Extra tests have shown we
have resolved the action of the thermal instability. As previously
noted in Papers I and II, such a large number of AMR levels is
employed in order to efficiently compute the self-gravity on the
coarsest levels and also fully resolve the structures formed in the
molecular cloud. The finest physical resolution is 0.293\,pc. The total
number of cells across all eight AMR levels is $13.9\times10^6$ grid
cells, ten times less than that required
by a fixed grid code with the equivalent 512$^3$ cells. The
large number of AMR levels is computationally costly. Each 3D HD
simulation with stellar feedback and self-gravity presented here with
8 AMR levels took approximately 60,000 CPUhours ($\sim$10-12 48-hour
cycles on 128 cores of the high performance computing facility
at Leeds), so 240,000 CPUhours for the four runs presented. Supporting
investigations included a parameter exploration of the initial
condition (see Paper I) and a high resolution re-run of the purely
hydrodynamic case which also explored clump-collapse to form
pre-stellar cores. This cost more than the total CPUhours used by the
simulations presented herein. Confidence in these simulations also
comes from the results presented in Paper II. Each further level of
AMR introduces a computational-cost-multiplier of between 6 and 8 for
this model, given the spherical nature of the cloud, thus making
complete higher resolution simulations prohibitively expensive at this
time. Tests in Paper II showed the resolution used is appropriate for
these simulations, though it is close to the cooling length at times.
The influence of the thermal instability causes the cloud to evolve
into multiple clumps. Specifically,
molecular clumps form throughout the diffuse cloud.
Self-gravity accelerates the contraction of the cloud.
Besides exploring the importance of the thermal
instability in molecular cloud evolution, the secondary aim of Paper I
was to provide a more realistic initial condition for this work, by
including more accurate heating and cooling, the effect of thermal
instability, self-gravity and magnetic fields, as compared to our
previous feedback studies \citep{rogers13,rogers14} where
the clump structure was based upon the simulations of
\cite{vazquez08} of turbulent and clumpy molecular clouds and
contained 3240\,M$_{\odot}$ of material in a 4\,pc radius.
\begin{figure*}
\centering
\includegraphics[width=150mm]{figure2.pdf}
\caption{Stellar evolution tracks \citep{vink00,vink01,ekstrom12} for
a 60\,M$_\odot$ star, showing mass-loss rate and wind velocity on the upper graphs, and energy
injection rate and total injected mass on the lower graphs. Raw data: doi.org/10.5518/201.}
\label{60evolution}
\end{figure*}
In this work, we take as our initial condition a repeat of a
cloud simulation without magnetic field following the method of Paper
I. Different random seeds result in a simulation that is qualitatively
the same as the result in Paper I, but quantitatively
different. After 27.1\,Myrs of evolution, densities in the
condensations formed in this new simulation have reached
100\,cm$^{-3}$ - the density threshold often used for injection of
stars in similar simulation work \citep[e.g.,][]{fogerty16}. In this
work, we have decided to inject stars at this time following the work
of such other authors and also because of our own high-resolution
simulations.
We show snapshots of this initial condition
in Figure \ref{figure-initial} and refer the interested reader to
Paper I for a full description of the evolutionary process that led to
the formation of this cloud. It is important to note though that this
time-scale of 27.1\,Myrs should not be considered as the `age' of the
parent molecular cloud - this is the length of time required to go
from a diffuse cloud with an average density of n$_H$ = 1.1 cm$^{-3}$
to a structured molecular cloud where feedback can be introduced.
Cold (less than 100\,K) condensations in the cloud have only existed
for a few Myrs, in reasonable agreement with observed ages of
molecular clouds.
In this paper, we consider four scenarios, each employing this initial
condition, in order to examine the effect of stellar feedback in this cloud.
We use the same method as Paper II, but repeat ourselves here for
clarity in this work.
\begin{figure*}
\centering
\includegraphics[width=150mm]{figure3.pdf}
\caption{Stellar evolution tracks \citep{vink00,vink01,ekstrom12} for
a 120\,M$_\odot$ star, showing
mass-loss rate and wind velocity on the upper graphs, and energy
injection rate and total injected mass on the lower graphs.
Raw data: doi.org/10.5518/201.}
\label{120evolution}
\end{figure*}
\subsubsection{Scenario 1 - a 15\,M$_{\odot}$ star}
In this scenario we introduce a 15\,M$_\odot$ star at the position
($x$,$y$,$z$) = (-0.025, -0.2351, -0.11) where the coordinates are
given in scaled code units and the grid extends to $3\times3\times3$
centred on (0, 0, 0). This is the location of the highest
density condensation in the cloud, closest to the centre of the
volume. We remove enough mass present in a spherical volume with a 5-cell
radius centred at this point to form a 15\,M$_\odot$ star, assuming
100\% conversion of cloud material to star. For this first
investigation, the cloud mass in this spherical region is removed at the
switch on of the stellar wind, t$_{wind}$=0, under the assumption
that this material has formed the star. If the mass is left in the
injection region, the stellar
wind rapidly and unrealistically cools and hence the feedback effects
are significantly underestimated. An advected scalar, $\alpha_{wind}$,
previously set to zero throughout the grid, is set to 1 in the wind
injection region in order to track the movement and mixing of the wind
material.
For the stellar evolution, a 15\,M$_{\odot}$ non-rotating Geneva
stellar evolution model calculated by \cite{ekstrom12} is used in
order to provide a realistic mass-loss rate over the lifetime of the
star, as per the method used in Paper II. Detailed plots of this
evolution can be found in Paper II. In this case, after 12\,Myrs
of stellar evolution, the cloud has also collapsed under the influence
of gravity. Densities in the centre of the cloud have reached levels
where the resolution is insufficient so the simulation was stopped at
this point, close to the end of the star's lifetime of 12.5\,Myrs.
During the wind phase, the total mass and total energy injected by the
star are 1.75\,M$_\odot$ and $1.05\times10^{49}$\,erg respectively.
The location of the star remains constant throughout this simulation.
The centre of cloud-collapse is not exactly at the location of the
star though, so in this case of a 15\,M$_\odot$ star, a moving source
is strictly required to accurately model this case. In future work, we
plan to convert the mass into a `star' particle following the method
in MG of \cite{vanloo13} and \cite{vanloo15}. The star can then move
through the computational volume whilst feeding back through winds and
SNe and remain consistent with self-gravity in the simulation.
\subsubsection{Scenario 2 - a 40\,M$_{\odot}$ star}
In this scenario we introduce a 40\,M$_\odot$ star using the same
method and at the same position as in Scenario 1, removing enough
mass to form a 40\,M$_{\odot}$ star from the spherical injection
region. A 40\,M$_{\odot}$
non-rotating stellar evolution model calculated by \cite{ekstrom12} is
used in order to provide a realistic mass-loss rate over the lifetime
of the star. Detailed plots of this evolution can be found in Paper
II. The total mass and total energy injected by the star prior to
supernova explosion are 27.2\,M$_\odot$ and $2.50\times10^{50}$\,erg
respectively. After 4.97\,Myrs, the star explodes as a SN, injecting
10\,M$_\odot$ of stellar material and 10$^{51}$ ergs of energy into
the same wind injection volume. The SN mass and energy is injected
over 500\,yrs, roughly consistent with the time taken for a remnant to
reach the size of the injection volume. An advected scalar,
$\alpha_{SN}$, previously set to zero throughout the grid, is set to 1
in the supernova injection region in order to track the movement and
mixing of the SN material. At this time the wind scalar
$\alpha_{wind}$ is set to zero. The fraction of cloud material in any
given cell is $\alpha_{cloud} - \alpha_{wind} - \alpha_{SN}$.
Gravity plays less of a role during this shorter stellar wind phase
as compared to the 15\,M$_\odot$ star case and the early SN
phase, partly due to the comparatively powerful dynamics and partly
due to the shorter timescale, but we continue to include it
for consistency within the cloud and to explore the evolution
post-SN when it again plays more of a role.
\subsubsection{Scenario 3 - a 60\,M$_{\odot}$ star}
In this scenario we introduce a 60\,M$_\odot$ star using the same
method and at the same position as in Scenario 1, removing enough
mass to form a 60\,M$_{\odot}$ star from the spherical injection
region. A 60\,M$_{\odot}$ non-rotating
stellar evolution model calculated by \cite{ekstrom12} is used in
order to provide a realistic mass-loss rate over the lifetime of the star.
The calculated mass-loss rate and wind velocity are shown in
Fig.~\ref{60evolution}. Also shown in Fig.~\ref{60evolution} are the energy
injection rate and total injected mass.
The total mass and total energy injected by the star prior to supernova
explosion are 47.5\,M$_\odot$ and $8.9\times10^{50}$\,erg respectively,
almost equivalent to the energy introduced in the SN event and considerably
more material. After
3.97\,Myrs, the star explodes as a SN in the same manner as in
Scenario 2.
\subsubsection{Scenario 4 - a 120\,M$_{\odot}$ star}
In this scenario we introduce a 120\,M$_\odot$ star using the same
method and at the same position as in Scenario 1. In this case, we simply
remove all the mass in a spherical region twice the size of the injection
region in order to form the 120\,M$_\odot$ star. A 120\,M$_{\odot}$ non-rotating
stellar evolution model calculated by \cite{ekstrom12} is used in
order to provide a realistic mass-loss rate over the lifetime of the star.
The calculated mass-loss rate and wind velocity are shown in
Fig.~\ref{120evolution}. Also shown in Fig.~\ref{120evolution} are the energy
injection rate and total injected mass. The total mass and total
energy injected by the star prior to supernova explosion are
89.1\,M$_\odot$ and $2.44\times10^{51}$\,erg respectively,
considerably more than the amount of material and energy
introduced in a SN event. After
3.01\,Myrs, the star explodes as a SN in the same manner as in
Scenario 2.
\begin{figure*}
\centering
\includegraphics[width=160mm]{figure4.pdf}
\caption{Cloud-wind interaction during the lifetime of a 15\,M$_\odot$ star.
Shown is the logarithm of mass density at various times and on various
planes through the computational volume. Length is scaled in units of 50\,pc.
Raw data: doi.org/10.5518/201.}
\label{wnd-15M}
\end{figure*}
\section{Results}\label{results}
In this section we present our results. We present both 2D slices
through the computational volume created within MG and 3D contour and
volume visualisations, created using the VisIt software \citep{visit}.
Raw data for the all figures in this paper is available
from the University of Leeds Repository at doi.org/10.5518/201.
\subsection{The wind phase}\label{windphase}
\begin{figure*}
\centering
\includegraphics[width=160mm]{figure5.pdf}
\caption{Cloud-wind interaction during the lifetime of a 40\,M$_\odot$ star.
Shown is the logarithm of mass density at various times and on various
planes through the computational volume. Length is scaled in units of 50\,pc.
Raw data: doi.org/10.5518/201.}
\label{wnd-40M}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=160mm]{figure6.pdf}
\caption{Cloud-wind interaction during the lifetime of a 60\,M$_\odot$ star.
Shown is the logarithm of mass density at various times and on various
planes through the computational volume. Length is scaled in units of 50\,pc.
Raw data: doi.org/10.5518/201.}
\label{wnd-60M}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=160mm]{figure7.pdf}
\caption{Cloud-wind interaction during the lifetime of a 120\,M$_\odot$ star.
Shown is the logarithm of mass density at various times and on various
planes through the computational volume. Length is scaled in units of 50\,pc.
Raw data: doi.org/10.5518/201.}
\label{wnd-120M}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=160mm]{figure8.pdf}
\caption{Column density, naive emission and 3D visualisation of the cloud-wind
structure just before each star goes SN. Naive emission is calculated as the
radiative cooling energy source for each cell at this time, projected along the
axes as defined in the figure. Length is scaled in units of 50\,pc. Each 3D
visualisation shows the isosurface of the wind scalar (in colour) in half the
computational volume ($x\leq0$) in order to show the core of the wind-blown
bubble and the distribution of high density clumps (black isosurfaces) in the
entire computational volume. Raw data: doi.org/10.5518/201.}
\label{wnd-proj}
\end{figure*}
In Fig.~\ref{wnd-15M} we show density slices through the computational
volume during the evolution of the 15\,M$_\odot$ star up to the end of its
life. The low mass-loss rate and corresponding low energy injection rate have
minimal effect on the cloud structure, generating only a wind cavity that
spreads away from the star into the cloud around the clumps for the first
few Myrs, but is eventually confined by the gravitational contraction of the
cloud itself. All three planes through the location of the star reveal similar
structure for all timepoints. This structure is a result of the weak wind that
is able to affect the inter-clump material, but is not able to
affect the high-density clumps themselves. Examining the extent of the
stellar wind material we see that very little wind material gets more than
10\,pc away from the star.
During the red super giant (RSG) phase of evolution, from
t$_{wind}$=11.2\,Myr until its SN explosion, the slow dense wind
deposits considerable amounts of material into the cavity formed by
the earlier wind and completely refills it. We do not show this stage
of evolution here, as the cloud itself has collapsed and dominates the
end stages of the stellar bubble's evolution. Neither do we trigger a
supernova in this simulation, as the cloud collapses to a high density
which is unrealistic given our resolution.
In Fig.~\ref{wnd-40M} we show the logarithm of density on all three
planes through the location of the 40\,M$_\odot$ star at various times
through the star's lifespan. By only 0.707\,Myrs into the main
sequence evolution of the star, the impact on the molecular cloud is
significant and clearly different from the 15\,M$_{\odot}$ star case,
as shown in the first row of Fig.~\ref{wnd-40M}. The stellar wind is
expanding away from the star, streaming past the clumps and already
forming clear channels through the cloud. The wind is also beginning
to ablate material from the clumps into the wind flow along the
channels. By 1.53\,Myrs, as shown in the second row of
Fig.~\ref{wnd-40M}, the wind has continued to expand, establishing a
distinct reverse shock at a radius of approximately 5\,pc. The
ablated clumps are losing further material and being pushed away from
the star - the force of the stellar wind on the clumps is clearly
overcoming the gravitational contraction of the cloud and the wind is
now blowing the cloud apart. After 3.01\,Myrs, the wind has reached
the edge of the cloud material and can be seen in the
$x$-plane expanding into the low-density surrounding medium. The wind
has not yet escaped the cloud on the other planes, probably
due to a combination of the centre-offset position of the star and the
clump distribution in the cloud. A small number of low-density
channels have now also become dominant in the structure, but by
4.42\,Myrs, at the end of the Main Sequence evolution of the star,
these channels have become less distinct on the x-plane, but more
distinct on the $y$-plane. This structure, with its channels and
clumps connected together is reminiscent of our previous work
\citep{rogers13}.
The star now enters the LBV phase. The wind mass-loss rate increases
by two orders of magnitude to $\sim10^{-4}$\,M$_\odot$\,yr$^{-1}$, and
the terminal wind speed reduces to $\sim100$\,km\,s$^{-1}$. This slow
dense wind forms a high density environment around the location of the
star, which eventually contains $\sim20$\,M$_\odot$ of LBV wind
material. This phase lasts approximately 200\,kyrs and is followed by
the WR phase of stellar evolution, where a variable, faster, less
dense but more powerful wind sweeps up the LBV wind over the course of
the final 400 kyrs of the star's life. The structure formed is presented in
the next section.
In Fig.~\ref{wnd-60M} we show the logarithm of density on all three
planes through the location of the 60\,M$_\odot$ star at various times
through the star's lifespan. After only 1.5\,Myrs, the impact on the
molecular cloud is already almost equivalent to the total impact of the 40\,M$_\odot$
star. The wind has carved channels to allow the flow of wind material
out of the cloud and is now breaking out into the surrounding medium.
After 3.3\,Myrs, at the end of the Main Sequence evolution, the structure
is more extreme, with several deep channels carved through the cloud
and considerable amounts of wind material flowing out of the
cloud. The reverse shock is clear and very close to spherical, almost
isolated from the parent cloud, at a radius of approximately 20\,pc
from the star. Many clumps towards the lower left of the star have been
ablated and are mass-loading the wind flowing away from the star.
However, towards the upper left of the star, many of the clumps are
still distinct, intact and still comparatively close to the star, especially
compared to the now close-to empty region towards the lower left of the
x-plane. Clearly the distribution of clumps and the position of the star
play a key role in defining the wind-blown structure around the star.
After 2.74\,Myrs, the star enters the LBV phase and again the mass-loss rate increases
and terminal wind speed reduces. As we saw also for the 40\,M$_\odot$
scenario, this slow dense wind forms a high density
environment around the location of the star. We show the nature of the
environment after this LBV phase along the third row of Fig.~\ref{wnd-60M}.
After this, during the WR phase of stellar evolution, the variable, faster,
less dense but more powerful wind is able to sweep the LBV and Main Sequence material
out into the cloud creating the final environment into which the supernova
will go off. By this time, the reverse shock is now up to 35\,pc from the star
and there are no traces of the high-density environment formed during the
LBV phase. The large reverse shock structure dominates the cloud, outside
which multiple channels will allow SN material to be
transported rapidly out of the cloud. We note that some clumps
have again survived this more extreme stellar evolution and remain
relatively close to the star.
In Fig.~\ref{wnd-120M} we show the logarithm of density on all three
planes through the location of the 120\,M$_\odot$ star at various
times through the star's lifespan. The 120\,M$_\odot$ star evolves
through the Main Sequence period in 2.24\,Myrs. The first two rows of
Fig.~\ref{wnd-120M} show the extreme effect this star has on its
environment. By the end of the Main Sequence, half the cloud has been
blown away, as shown in the $x$-plane. The reverse shock, previously
relatively small, now approaches half the radius of the original
cloud. It is almost spherical - the molecular cloud material has been
carried out by the force of the stellar wind. In the $y$- and
$z$-planes, channels through the cloud are clear. In the $z$-plane,
some molecular cloud material is still reasonably close to the star,
just outside the reverse shock. During the following LBV phase,
multiple non-spherical shells form around the star, as shown in the
third row of Fig~\ref{wnd-120M}, caused by the oscillatory changes in
the mass-loss rate and wind speed (see Fig~\ref{120evolution}). The
structure of the cloud is quite different depending upon which plane
is considered. Low-density empty voids exist to the left in the
$x$-plane and clumpy structure surrounds the shells formed during the
LBV phase in the $y$-plane.
After the LBV phase, the variable, faster, less dense WR wind is able to
evacuate the interior of the cloud in all directions, generating a strong
reverse shock that in the empty void extends to more than 50\,pc from
the star. We show the final structure of the cloud at the end of the star's
life, after 3.01\,Myrs, across the fourth row of Fig.~\ref{wnd-120M}. Much
of the cloud has been swept away. Compared to the previous cases, the
least amount of clumpy cloud material remains around the star - something
we will study in more detail in the following Analysis section.
In Fig.~\ref{wnd-proj} we show column density, naive emission (calculated
from the radiative cooling energy source term) and 3D visualisations of the
cloud-wind structure at the end of the life of the three most massive stars. In the
40\,M$_\odot$ star case, the blown-out side of the cloud is very clear in the
column density plot. The radiative cooling plot highlights the `emission' from
the isolated LBV/WR shell very clearly, as discussed in more detail in the next
sub-section. In the case of the 60\,M$_\odot$ star, the column density
projected along the $y$-axis highlights the fact that whilst the wind has blown
out a larger section of the cloud than in the previous case (as shown in
Fig.~\ref{wnd-60M}), the line-of-sight has strong effect on how obvious this
blow-out is to the observer. The naive emission from the radiative cooling
energy source term highlights the much greater extent of the more
powerful WR wind in this case, reaching throughout the cloud. Column
density shown collapsed along the $z$-direction for the 120\,M$_\odot$ star
does show very clearly how much of the cloud has been dispersed from the
original structure, as compared to how much is left in the top half of the cloud.
Distinct in this plot and the plot for the 40\,M$_\odot$ star are the radially aligned
`spokes' of the cloud after sculpting by the stellar wind.
The hot, fast WR wind has triggered 'emission' all over the structure in the
case of the 120\,M$_\odot$ star. Of course the `emission' assumes optically thin
conditions and true estimates of emission require radiative transfer to be
simulated alongside the hydrodynamic evolution of the cloud-wind interaction.
In the third column of Fig.~\ref{wnd-proj} we show a complex 3D isosurface
of the wind scalar on half the volume (indicated by the colour scale for the
domain x$\leq$0) and a second isosurface of the high density clumpy cloud
structure (indicated by the black isosurfaces at
$\rho$=25\,cm$^{-3}$). In all three star cases, the yellow core highlights the
inner core of the wind-blown bubble as well as the increasing extent of the
wind-carved channels with greater stellar mass, reaching off the domain in
many places for the case of the 120\,M$_\odot$ star. Noticeable also from
these 3D visualisations is the widening distribution of the high density structure
(indicated by the black clumps), dispersed by the increasingly powerful stellar
winds with increasing stellar mass.
\subsection{The early supernova phase}\label{snphase}
\begin{figure*}
\centering
\includegraphics[width=165mm]{figure9.pdf}
\caption{Early SN-cloud-wind interaction for a 40\,M$_\odot$ star.
Shown is the logarithm of mass density at various times on the
plane at $x$ = -0.025. Length is scaled in units of 50\,pc.
Raw data: doi.org/10.5518/201.}
\label{sn-40M}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=165mm]{figure10.pdf}
\caption{Early SN-cloud-wind interaction for a 60\,M$_\odot$ star.
Shown is the logarithm of mass density at various times on the
plane at $x$ = -0.025. Length is scaled in units of 50\,pc.
Raw data: doi.org/10.5518/201.}
\label{sn-60M}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=165mm]{figure11.pdf}
\caption{Early SN-cloud-wind interaction for a 120\,M$_\odot$ star.
Shown is the logarithm of mass density at various times on the
plane at $x$ = -0.025. Length is scaled in units of 50\,pc.
Raw data: doi.org/10.5518/201.}
\label{sn-120M}
\end{figure*}
The environment into which the SN mass and energy from the
40\,M$_\odot$ star are injected is shown in Fig.~\ref{sn-40M}(a). The
non-spherical high density shell structure formed during the WR stage
of evolution is centred on the location of the star, distinct from,
but strongly influenced by, the surrounding molecular cloud structure.
In 4700\,yrs the supernova remnant (SNR) has propagated far enough to
fill the LBV/WR shell and after 7073 years the SNR has overrun the
shell and is now expanding into the star's wind-blown bubble, as shown
in panel (b) and (c) of Fig.~\ref{sn-40M}. After 10,000 yrs, shown in
panel (d), the forward shock of the SNR can clearly be seen expanding
out into the bubble. Meanwhile, internal shocks reflecting off the inner shell
are refilling the SNR. Clump remnants that had survived the wind phase
are compressed by the SNR over the next 10,000 years (panels (e) and
(f)) and then accelerated outwards, towards the edge of the wind
bubble. By 42,000 years, in the final panel of Fig.~\ref{sn-40M},
the internal shocks running back towards the location of the star from the
edge of the wind bubble are acting to generate even more complex
structure inside the SNR. The SNR now continues to expand, encompassing
the wind bubble. The forward shock of the SNR reaches the edge of the
grid after 69,000 years, at the point where the SNR is closest to the edge
of the grid in this panel.
In our previous work, studying the evolution of 40\,M$_\odot$ star in
a sheet-like cloud, we found the stellar wind generated a tunnel
through the cloud, through which the SNR was then able to escape the
parent cloud in a matter of only 30,000 years. Here, the SNR is able
to progress through the wind-blown bubble equally as quickly, going
beyond the extent of the original parent cloud where the wind-blown
bubble had swept up and then beyond the molecular cloud itself. In
other directions though, the SNR has been rapidly deccelerated as it
enters the parent molecular cloud - only in panel (e) of
Fig.~\ref{sn-40M}, 13,000 years after the SN event, do we begin to
clearly see the SNR progressing into clumpy cloud material. So in this
case, as in the magnetically-influenced case of the 40\,M$_\odot$ star
in Paper II, the wind-blown environment is the key to the evolution of
the SNR. Progress in the hydrodynamic case is dramatically slowed by
the denser clumps and the inter-clump material, as compared to
progress through the low-density wind cavity. We now examine the same
intra-cavity phase for the 60\,M$_\odot$ and 120\,M$_\odot$ stars.
In Fig.~\ref{sn-60M} we show the logarithm of density on planes
through the location of the 60\,M$_\odot$ star at times after the SN
event corresponding approximately to those shown in
Fig.~\ref{sn-40M}. Immediately clear in panel (a) is the larger size
of the wind-blown bubble and also the lack of any LBV/WR shell. The
late-stage wind has blown out through the entire cavity formed in the
preceding phase. The reverse shock of the WR wind is visible above and
to the right of the star's location, located at a stand-off against
the remaining clumpy cloud material. In the other directions, the
shock of the WR wind is not visible in this figure due to the scaling
chosen to show the supernova evolution, but can be seen in the final
row of Fig.~\ref{wnd-60M}. The supernova explodes into this
environment. For the first 4000 years the SNR expands into WR wind,
unhindered by material from the preceding phases, or cloud material,
as shown in panel (b). By 7000 years the SNR has passed the reverse
shock of the WR wind in all directions and is now interacting with the
remaining cloud material close to the star. Shocks are beginning to
propagate down the channels between the network of clumps, whilst
stand-off bow shocks are also forming ahead of the radial spokes of
the network. Meanwhile, in the opposite direction, the SNR continues
to expand into shocked LBV/WR wind material. This scenario of free
expansion in one direction with interaction in the other now describes
the next ~20,000 years of SNR evolution as shown in the middle row of
Fig.~\ref{sn-60M}. Compared to the case of the 40\,M$_\odot$ star,
the wind structure is much larger and the channels off the grid have
allowed the forward shock to leave the grid much earlier, specifically
only 15,000 years after the SN event, propagating down the widest
wind-blown channel shown on the left of these panels. By 50,000 years,
the interactions of the SNR with this structure have setup multiple
bow shocks which interact and form an even more complex
structure. Channels allowing the early propagation of the SNR off the
grid have now been filled with multiple wide bow-shock structures, as
has the inside of the remnant. It is of particular interest here to
note the survival of the spoke-like network of clumpy material to the
left of the star. As the preceding wind rearranged the individual
clumps into aligned spokes, the cold cloud material aligned in the
spokes is self-shielded against the passage of the SNR (and the
preceding WR wind). Thus, whilst the head of each spoke is ablated,
the length of the structures at 10-20\,pc means much of the structure
survives beyond this early stage of the SNR - clearly until at least
80,000 years beyond the SN event, as shown in panel (i). We will
explore the further evolution of these structures in the next
sub-section and analyse the amount of cold cloud component remaining
in the following Section. Some slight numerical artifacts can be seen
in the interior of the SNR at times, caused by derefinements of the AMR
scheme in the smooth but ballistic and expanding flow. These have no
effect on the passage of strong internal shocks, reflected off the
cloud-wind structure, as can be seen across the third row of the figure.
In Fig.~\ref{sn-120M} we show the logarithm of density on planes
through the location of the 120\,M$_\odot$ star at times after the SN
event corresponding approximately to those shown in Figs.~\ref{sn-60M}
and ~\ref{sn-120M}. Clear again in panel (a) is the greater effect
upon the cloud of the stellar wind - half the cloud has been blown
away. Again there is no LBV/WR shell - the reverse shock of the WR
wind is visible to the right of the star's location, located at a
greater stand-off distance against the remaining clumpy cloud material
than in the case of the 60\,M$_\odot$ star. In the other directions,
the reverse shock of the WR wind is almost off the grid and far
outside the original extent of the parent cloud
(r$_{cloud}\approx1$). Panel (b) shows the SNR expanding into the
undisturbed WR wind cavity for the first 5,000 years or so. After this
time, the SNR passes the WR reverse shock boundary, creating internal
structure at the edge of the SNR. It also begins to interact with the
remaining `spokes' of cold cloud material, ablating the inner end of
the radially-aligned spoke structure.
Over the next 10,000 years (second row of Fig~\ref{sn-120M}), the
forward shock of the SNR continues to expand into shocked WR wind
material in one direction, whilst interacting with remaining cloud
material in the other. This makes it easy to see how the SNR evolves
differently under almost unhindered expansion versus expansion into
multiple cold molecular clumps. This also highlights, as with the
previous cases, that it is possible to observe {\it both} scenarios in
one supernova event. The forward shock leaves the grid after
approximately 15,000 years, similarly to the 60\,M$_\odot$ star case.
Across the third row of Fig.~\ref{sn-120M} it is
possible to see the same SNR interaction behaviour as in the previous
case of the 40\,M$_\odot$ star, in that the SNR expands along the free
channels, only in this case the left side of the cloud is so dispersed
that it is more like remnants of cloud material in a low-density
shocked wind medium, than channels through shocked medium. For this
reason the bow-shocks forming ahead of the remaining material against
the SNR are not able to intersect to the extent that a reverse shock
propagates back towards the explosion site. Indeed, the evolution of
the bow-shocks over time shows the relatively rapid cooling of these
structures, as they fall back onto the clumps they were initially
detached from. A large void in the centre of the SNR now
exists. Evolving rapidly into this void, launched off the remaining
cloud material to the right, is an internal shock. This can clearly be
seen moving across the grid with a negative $y$-velocity in the bottom
row of Fig.~\ref{sn-120M}. By 50,000 years, this structure has passed
the original location of the SN event. This creates the
observationally interesting scenario of an offset `centre' of the SNR,
compared to the original location of the star (and the current
location of any remaining stellar remnant). This particular simulation
demonstrates the case of the stellar remnant being close to the edge
of the SNR within a comparatively short 50,000 years since the SN
event.
\subsection{The late supernova phase}\label{latesnphase}
\begin{figure*}
\centering
\includegraphics[width=165mm]{figure12.pdf}
\caption{Late SN-cloud-wind interaction for a 40\,M$_\odot$ star.
Shown is the logarithm of mass density and temperature at various
times on the plane at $y$ = -0.2351. Length is scaled in units of 50\,pc.
Raw data: doi.org/10.5518/201.}
\label{snlate-40M}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=160mm]{figure13.pdf}
\caption{Late SN-cloud-wind interaction for a 60\,M$_\odot$ star.
Shown is the logarithm of mass density at various
times on the plane at $y$ = -0.2351. Length is scaled in units of 50\,pc.
Raw data: doi.org/10.5518/201.}
\label{snlate-60M}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=160mm]{figure14.pdf}
\caption{Late SN-cloud-wind interaction for a 120\,M$_\odot$ star.
Shown is the logarithm of mass density at various
times on the plane at $y$ = -0.2351. Length is scaled in units of 50\,pc.
Raw data: doi.org/10.5518/201.}
\label{snlate-120M}
\end{figure*}
In Fig.~\ref{snlate-40M} we show the late post-SN evolution of the
40\,M$_\odot$ star simulation on $y$-planes through the location of
the star. We show $y$-planes in this figure (as opposed to the
$x$-planes shown for the early SN stage), as the SNR leaves the
computational volume later on in the $y$-plane (due to the position of
the star and structure of the cloud) and hence boundary effects come
into play later on this plane. After 98,000 years, as shown in panel
(a), the SNR can be seen escaping the molecular cloud, having now
swept up the wind material into the thin shell, punctured by the
spokes of remaining cloud material. Over the next few hundred thousand
years, internal shocks bounce back from the edge of the SNR, and the
ablation of the remaining cloud material increases due to the passage
of these shocks within the internal void of the SNR. The SNR has also
expanded and cooled far enough to reach the radiative phase of
evolution. In panel (b), the forward shock is becoming subject to
Rayleigh-Taylor instabilities and beginning to break up into
individual clumps, in the same way as noted previously in Paper II.
The remnants of the molecular cloud survive this
phase reasonably well.
By 1.51\,Myrs, the forward shock of the SNR has reached the edge of
the computational volume and the simulation after this point should
not be taken too literally, since boundary effects have come into
play. Nevertheless, the later evolution can be used as a guide to the
sort of behaviour one might expect to see. We are now particularly
interested in the fate of the remaining cloud material. Considering
the second row of panels, (d) at 2.1\,Myrs, (e) at 3.28\,Myrs and (f)
at 4.46\,Myrs post-SN, it is clear that the molecular cloud has been
dispersed as all that is left are a few disparate high-density clumps.
It's clear by looking at the temperature plots at equivalent times, as
shown in the third row of the figure, that these are still cold
($\leq$100\,K), even after the passage of the SNR and are the final
remaining parts of the molecular cloud. The structure of the cloud is
destroyed, but parts of the cold cloud material remain, much as
occurred in the case of the 40\,M$_\odot$ star in a
magnetically-collimated sheet-like cloud in Paper II. As also noted in
Paper II, self-gravity in the dispersed clumps is now likely to
dominate their evolution, possibly leading to further star formation.
It is highly likely that the disruption of the molecular cloud is
caused by the SN event and not influenced by any numerical effects, as
the outward motion of the clumps can be traced back to before the SNR
reached the edge of the computational volume. We will consider in the
following cases whether this can be confirmed further, by examining
the rate of cloud destruction in each case, but to be more confident
of this result, future computationally-costly simulations with a
larger volume are necessary. Further questions over the fate of the
remaining molecular material and whether the SN triggers any further
formation of cold material by disrupting the thermal stability of the
wind/cloud material are addressed in Section \ref{analysis}. Some
indications are given by the decreasing lower limit of the temperature
range going from panels (h) to (i), indicating the presence of
ever-colder material on this plane. Note also the `turbulent' nature
of the interior of the SNR, caused by the interaction of multiple
shocks. Even after 2\,Myrs, the temperature inside the SNR is still
above 10$^6$\,K. The SNR cools noticeably over the next
2.5\,Myrs. With a cooling and dispersing remnant, gravity plays an
increasingly important role. We have simulated on to 6\,Myrs post-SN,
but are unable to confidently assess the balance in the simulation
between the evolution caused by the SNR, by gravity and by the
influence of any boundary effects.
In Fig.~\ref{snlate-60M} we show the late post-SN evolution of the
60\,M$_\odot$ star simulation. Whilst structural similarities to the
40\,M$_\odot$ case are apparent, it should be noted that the SNR is
larger and has left the cloud at a much earlier time. Clearly the
major reason is the enhanced disruption of the cloud by the wind
during the preceding evolution. While the SNR transitions into the
radiative phase at around the same time, comparing panel (b) in both
this and Fig.\ref{snlate-40M}, it is clear that the rest of the
remnant is much more dispersed. The same effects come into play as
the star evolves into Myrs post-SN, but in this case it should be
noted that the clumpy remains of the molecular cloud are fewer and
further out from the star. We show the SNR at 5.64\,Myrs post-SN to
display the final turbulent state of the SNR. We analyse the evolution
of the cold molecular material in the computational volume as a whole
in the next Section.
In Fig.~\ref{snlate-120M} we show the late post-SN evolution of the
120\,M$_\odot$ star simulation. The effect of the preceding wind phase
- to allow the SNR to have completely blown through the wind-cloud
structure in the $y$-plane after only 100,000 years - is shown in
panel (a). A slight ``echo'' of the edge of the cloud remains, but
otherwise turbulent conditions dominate the interior of the SNR. This
arc-like echo of the cloud breaks up after 300,000 years of SNR
evolution. After 1.5\,Myrs, the last few high-density cloud components
are leaving the grid. After this, a low density SNR void is surrounded
by average ISM conditions (densities ranging from 10$^{-2}$ to 10$^0$
particles per cm$^3$). There are no cloud components left on this
plane - the cloud has been completely dispersed by the SNR.
\section{Analysis}\label{analysis}
\subsection{Energy}
In Fig.~\ref{energy} we show how the distribution of energy between
kinetic, thermal and hot thermal (i.e. material above 10,000K) in the
cloud, wind and SN material combined, varies
over time following the introduction of stellar feedback, with
reference to a simulation of the cloud evolution without feedback for
each star. The thermal and kinetic energy fractions together sum to
1.0, i.e. we are ignoring gravitational energy in this figure. Note
that the hot thermal profile indicates the fraction of the thermal
energy which is ``hot''. In the second plot for each star, we show how
the total energy in the cloud varies over time, with reference to the
simulations of the cloud without feedback. The third plot for the 40,
60 and 120\,M$_\odot$ star cases shows the post-SN ``radiated and
escaped energy", i.e. energy that is radiated away or leaves the grid
dynamically - we do not differentiate between these two means of
energy loss in the plot.
\begin{figure*}
\centering
\includegraphics[width=170mm]{figure15.pdf}
\caption{Time profiles of energy fractions, totals and escaped amounts
of energy in the cloud, wind and SN material combined. The thermal and kinetic energy fractions together
sum to 1.0 i.e. we are ignoring the change in gravitational energy. The
`hot thermal' profile indicates the fraction of the thermal energy which
is hot (above $10^{4}$\,K), and so also has a range between 0.0 and
1.0. Lines with markers indicate the energy behaviour in the cloud with no
stellar feedback. On the third row of the 40, 60 and 120 M$_{\odot}$
star plots, the ``radiated and escaped energy" is the total energy
subtracted from the injected energy of the SN, 10$^{51}$ erg. Note
the different timescales in these plots. Raw data: doi.org/10.5518/201.}
\label{energy}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=135mm]{figure16.pdf}
\caption{Time profiles of the energy fractions of the stellar wind
material as traced by the advected scalar $\alpha_{wind}$. The thermal
and kinetic energy fractions together sum to 1.0 i.e. we are ignoring the
change in gravitational energy. The
`hot thermal' profile indicates the fraction of the thermal energy which
is hot (above $10^{4}$\,K), and so also has a range between 0.0 and
1.0. Raw data: doi.org/10.5518/201.}
\label{wndenergy}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=135mm]{figure17.pdf}
\caption{Time profiles of the energy fractions of the supernova
ejecta, as traced by the advected scalar $\alpha_{SN}$The thermal
and kinetic energy fractions together sum to 1.0 i.e. we are ignoring the
change in gravitational energy. The
`hot thermal' profile indicates the fraction of the thermal energy which
is hot (above $10^{4}$\,K), and so also has a range between 0.0 and
1.0. Raw data: doi.org/10.5518/201.}
\label{snenergy}
\end{figure*}
For the first 10\,Myrs of the 15\,M$_{\odot}$ star case, the energy
fractions do not change considerably during the wind phase, as
compared to the reference case. This was expected from the minimal
dynamic and structural impact the star has on the cloud. In fact the
stellar wind introduces so little energy, that the overall energy in
the cloud reduces over the first 10\,Myr of the star's life, as
compared to the `no feedback' reference case. This likely arises due
to relatively efficient radiative cooling of the predominantly thermal
energy injected by the stellar wind, perhaps due to weakly compressing
and heating neighbouring cloud material. From 10\,Myr post star
formation (37\,Myrs in the figure), the collapse of the cloud
dominates the evolution. The stellar wind has little effect on this
and after 37\,Myrs, these results contain no meaningful information.
In the 40\,M$_{\odot}$ star case, the stellar wind has a clear effect
on the energy fractions during the wind phase. The stellar wind
supplies and creates large amounts of kinetic and hot thermal energy,
raising both fractions above the close-to-zero reference case. Given
that the wind blows a hot bubble, it's likely that radiative losses
are responsible for the drop in total energy approaching the end of
the star's life, before the introduction of 10$^{51}$\,erg of thermal
energy and 10\,M$_{\odot}$ of material in the SN event. The
introduction of this thermalised SN kinetic energy is most obvious in
the total energy plot, as the spike just after 32\,Myrs. Since the SN
energy is injected thermally, both the thermal energy and the `hot
thermal' phase rise sharply at this time. The third plot for the
40\,M$_{\odot}$ star case shows that the SN energy is rapidly
transported and radiated out of the simulation. Note that the forward
shock of the SNR first reaches the edge of the computational volume
approximately 69,000 years after the SN event. At approximately
1\,Myr post-SN, the energy balances are most different from the
reference case without feedback (apart from the first few thousand
years of SN evolution, which we discuss in more detail later in this
sub-section). After this, as the SNR escapes from the cloud/wind
structure into the surroundings and then off
the grid, the kinetic energy fraction falls and the thermal energy
fraction rises. This behaviour continues, albeit at a decreasing rate,
until the end of the simulation. The actual amount of thermal energy
drops until approximately 2.5\,Myrs post-SN and then plateaus, whereas
the kinetic energy simply continues to drop during the simulation
post-SN. The plateauing of thermal energy corresponds to when the hot
thermal fraction of energy remaining on the grid has fallen to close
to zero - much of the material on the grid has now dropped to
temperatures below 10,000K and densities in the range 0.01-1
cm$^{-3}$, with random motions at a range of velocities and scales
(except for the last remaining gas in the hot, central void of the
SNR, but this is at low density). This is fairly characteristic of the
warm, neutral medium, formed in this case from evolution of an
isolated single star, approximately 2.5\,Myrs post-SN.
\begin{figure*}
\centering
\includegraphics[width=170mm]{figure18.pdf}
\caption{Maximum densities reached in the stellar feedback simulations
(solid lines) shown with the maximum density in a continued simulation
of the same initial conditions without feedback (lines with markers) for reference.
Note that without feedback, the cloud collapses to unrealistic densities for
the resolution of the simulation. The SN occurs at $t = 32.1$, 31.1
and 30.1\,Myr for the $M = 40$, 60 and 120\,M$_{\odot}$ cases.
Raw data: doi.org/10.5518/201.}
\label{maxrho}
\end{figure*}
In the 60\,M$_{\odot}$ star case, the wind effect upon the thermal and
kinetic energy fraction is an amplified version of the 40\,M$_{\odot}$
star case. During the Main Sequence wind phase, the amounts of both
thermal and kinetic energy on the grid increase smoothly for
2.5\,Myrs, then decrease smoothly, with the amount of kinetic energy
increasing and decreasing more than thermal energy. In the final
stages of the star's evolution, the total energy in the simulation is
increased by a factor of 5 by the stellar wind, with a sharp increase
in amounts of both thermal and kinetic energy. This corresponds
closely with the increased energy injection rate of the stellar wind,
as shown in Fig.~\ref{60evolution}. The kinetic energy increases by a
factor of 10, leading to the apparent drop in the thermal energy
fraction, although there is no drop in the actual amount of
energy. Just before the supernova, ten times more energy is present in
the simulation than in the reference case, or the 40\,M$_{\odot}$ star
case (which is close to the reference case pre-supernova). One must
bear in mind that a substantial amount of the total wind energy up to
this point has already been transported off the grid, while some has
also been radiated away.
The SN event increases the amount of energy on the grid by a factor of
10. The post-SN evolution of the energy fraction and the total energy
in the simulation evolves in a similar way to the 40\,M$_{\odot}$ star
case, with an initial peak in kinetic energy (and trough in thermal
energy) as the SNR blows out of the cloud. In this case though, the
forward shock leaves the grid along the wind-carved channels much
more quickly, first reaching the edge of the computational volume after
only 15,000 years. The initial peaks are followed by similar
evolution between the thermal, hot thermal and kinetic energy
fractions when compared to the 40\,M$_\odot$ star case.
The thermal energy plateau of warm, neutral medium is
reached on the same timescale. Whilst the detailed evolution of the
cloud with feedback from a 60\,M$_{\odot}$ star is visibly very
different to the 40\,M$_{\odot}$ star case, as shown in the previous
Section, the evolution of the energy fraction, total energy and
radiated/escaped energy is not considerably different.
The 120\,M$_{\odot}$ star case is similar to the case of the
60\,M$_{\odot}$ star. The wind introduces considerably more kinetic
energy though, driving shock heating and a greater fraction of hot
thermal energy. The late stages of stellar evolution are equally
complex and track the variations in energy injection rate, but the
post-SN evolution is remarkably similar. Most interestingly, the
supernova from the 120\,M$_{\odot}$ star very rapidly loses thermal
energy post-SN and the kinetic energy peaks at a higher fraction than
both the 40 and 60\,M$_{\odot}$ star cases. Again, the forward shock
leaves the grid in around 15,000 years, along the low density channels
carved by the wind through the cloud and surroundings. Although the post-SN
material in the 120\,M$_{\odot}$ case does briefly plateau in the
warm, neutral phase between 2.5 to 3.5\,Myrs post-SN, the thermal
energy on the grid then continues to decrease. The grid is dominated
by the hot, low-density void of the inner SNR, which is still cooling,
leading to this effect. Any warm, netural medium would be outside the
grid. This highlights the range upon which each star acts, increasing
with the mass of the star (and hence strength of the wind). A larger
computational volume is clearly necessary to capture the
characteristics of the surrounding medium that the 120\,M$_{\odot}$
star creates post-SN.
In Fig.~\ref{wndenergy} we show the evolution of the energy fraction
of stellar wind material for each star. The wind material injected by
the 15\,M$_{\odot}$ star is almost completely dominated by hot thermal
energy, and has very little kinetic energy, until the end of the
star's life (indicating that the wind is ``bottled up''). The wind
material from the 40\,M$_{\odot}$ star has a greater fraction of
kinetic energy, which increases steeply close to the supernova event
as the star enters its WR phase. Post-SN, the energy of the wind
material follows a similar trend to that of the total energy in the
simulation, as shown in the last figure. The same trends are seen in
the 60 and 120 \,M$_{\odot}$ star cases. In Fig.~\ref{snenergy} we
show the evolution of the energy fractions of the supernova material
for each star. Clearest in panel (a), showing the first few tens of
thousands of years of the SNR evolution, is the initial conversion of
thermal energy into the kinetic energy of expansion of the ejecta
(causing the thermal energy to drop to only 10\% of the total ejecta
energy), followed by a rapid increase to 35\% as the expanding ejecta
passes through a reverse shock and is rethermalized to an 85\%
fraction. Over the next 4\,Myr, the thermal energy fraction drops to
60\% of the total (but since material is lost off the grid one cannot
draw any further conclusions).
\subsection{Maximum density}
In Fig.~\ref{maxrho} we show the maximum densities reached in the
simulations, compared to the maximum density at the same time in the
reference case without stellar feedback. In the 15\,M$_{\odot}$ star
case, the maximum density during the wind phase follows the reference
case. The maximum density is more than 10$^6$ H\,cm$^{-3}$ during the
cloud collapse; realistic in the gravitational collapse of a molecular
cloud, but beyond the accuracy of the resolution used in this
simulation. In the other three cases, the stellar winds are able to
create considerably higher densities than in the reference case during
the wind phase, enough for the possibility of a second generation of
star formation with densities over 10$^3$ H\,cm$^{-3}$. Variations are
not due to the stellar wind - it is relatively steady with smooth
variation over the majority of the star's lifetime - but instead must
be due to the compressional effect the stellar wind has on the cloud,
rearranging the structure and compressing ``spokes''.
By the time of the supernova, the maximum density in all three cases
is back to similar levels to the reference case, showing that whilst
the final wind stages inject a considerable amount of mass, densities
elsewhere reduce as the cause of increased density, compression by the
wind, has reduced. The SN instantly raises the density in the
injection location, but the spikes seen in the plots are instead
associated with compression of surrounding LBV/WR material and then
cloud material in the 40\,M$_{\odot}$ star case, and the cloud
material in the 60 and 120\,M$_{\odot}$ star cases. Later peaks in the
maximum density plots, showing against the overall decreasing trend of
maximum density, may be associated with the formation of new cold
material after SN affected material returns to the thermally unstable
phase. Localised star formation, according to our star formation
rules, would occur in all these cases. Only in the case of the
60\,M$_{\odot}$ star is a late onset increase in maximum density
apparent (post 34.5\,Myrs). This must be associated with increasing
amounts of cold molecular material, even in the warm, neutral
medium. This is investigated further in the next sub-section.
\subsection{Phases}
\begin{figure*}
\centering
\includegraphics[width=170mm]{figure19.pdf}
\caption{Variation of phase fractions with time for the feedback simulations.
The reference case without feedback is indicated by lines with markers in the
phase fraction plots. Raw data: doi.org/10.5518/201.}
\label{phases}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=150mm]{figure20.pdf}
\caption{Mass-weighted temperature-density and pressure-density distributions at the
end of the star's life for each of the simulations. Over-plotted dashed
lines indicate the approximate thermal equilibrium between heating and cooling
for the prescriptions used. Raw data: doi.org/10.5518/201.}
\label{wndphases}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=160mm]{figure21.pdf}
\caption{Mass-weighted temperature-density and pressure-density distributions for the simulations
presented here during the early evolution of the supernova in each case. Over-plotted dashed
lines indicate the approximate thermal equilibrium between heating and
cooling for the prescriptions used. Raw data: doi.org/10.5518/201.}
\label{snphases}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=150mm]{figure22.pdf}
\caption{Mass-weighted temperature-density and pressure-density distributions for the 60\,M$_{\odot}$
star case in the late stages of the supernova event. The evolution of the 40\,M$_{\odot}$
and 120\,M$_{\odot}$ star cases are very similar. Over-plotted dashed lines indicate
the approximate thermal equilibrium between heating and cooling for
the prescriptions used. Raw data: doi.org/10.5518/201.}
\label{snlatephases}
\end{figure*}
In Fig.~\ref{phases} we show the phase fraction and total cloud mass
for all four feedback simulations. At the injection of a star,
27\,Myrs into the cloud's evolution, the cloud is in thermal
equilibrium (i.e. on the equilibrium curve between heating and cooling
processes) with 50\% of its mass in the warm phase (5,000\,K $\leq$ T
$\leq$ 10,000\,K), $\approx 25$\% of its mass in the cold phase (T
$\leq$ 160\,K) and $\approx 25$\% of its mass in the unstable
equilibrium part of the pressure-density distribution (160\,K $\leq$ T
$\leq$ 5,000\,K). No material is in the hot thermal phase (T $\geq$
10,000\,K).
Once mechanical feedback starts the cloud material is shock heated out
of thermal equilibrium and requires a cooling time to radiate away its
excess energy. The low density shocked stellar wind gas has a
comparatively long cooling timescale, while the denser and cooler
swept-up gas has a much shorter cooling timescale. Later on, the hot
SN ejecta and the gas it sweeps up have comparatively long cooling
timescales.
Fig.~\ref{phases}(a) shows that during the wind phase of the
15\,M$_{\odot}$ star case before the cloud collapses (beyond
$t=37$\,Myrs), the fraction of cold material decreases during the
first 0.5\,Myr, with a corresponding increase in unstable material,
then slowly increases very much like the reference case with no
feedback - the molecular cloud continues to form around the stellar
wind bubble. The introduction of stellar feedback has little effect
upon the amount of warm thermally stable material, until close to the
collapse of the cloud. The presence of a larger quantity of unstable
material as compared to the reference case enables the condensation of
slightly more cold material, allowing the fraction of cold material to
increase above the reference case with feedback. One of the very few
effects of the 15\,M$_{\odot}$ star then, is to slightly increase the
amount of cold molecular material in the cloud.
In panel (b) of Fig.~\ref{phases} we show the phase evolution of the
material in the simulation of mechanical feedback from the
40\,M$_{\odot}$ star. The effect upon the cloud is greater than that
of the 15\,M$_{\odot}$ star. Whilst the warm fraction remains
reasonably constant for the first 2\,Myrs, the stellar wind rapidly
heats cold cloud material, moving it back into the unstable phase
(from where it all originally began in the diffuse initial condition
preceding the formation of the molecular cloud at t=0). The unstable
phase fraction increases and the cold phase fraction decreases. Over
the rest of the star's life though, both the cold and unstable phase
fractions continue to grow, at the expense of the warm phase, which
drops to below 20\% by the time of the supernova, compared to 50\% in
the reference case at the same time. Clearly the star reduces the
amount of cold material formed in the cloud (to 30\%, rather than more
than 40\% in reference case), but is not able to stop the cloud
continuing to form cold material. The SN then transports hundreds and
eventually thousands of solar masses of SN, wind and then cloud
material out of the cloud. The SN rapidly heats nearly 10\% of the
cloud material into the hot thermal phase (above 10,000\,K), but the
majority of this material radiatively cools quickly and the fraction
of hot material drops back close to zero before the SNR leaves the
cloud (from about 0.5\,Myrs after the explosion). The major effect of
the supernova is to transition nearly 70\% of the material on the grid
into the thermally unstable phase, mostly from the warm phase. The
amount of cold material also decreases very rapidly post-SN, returns
briefly to nearly 30\% after 0.5\,Myrs post explosion - the point at
which the SN forward shock has progressed through the cloud - and then
decreases as the remnant evolves. These changes can be understood as
the sweeping up and ejection of material from the cloud by the
SN, whilst the cold cloud material reacts much more slowly. Some of
this behaviour is also due to the (SN) shock heating and subsequent
cooling of dense material, as found by \citet{rogers13}. The fraction
of cold dense material remains at around 20\% throughout the post-SN
evolution, with a slight increase at late-time, matched by an increase
in warm phase material, with the unstable phase feeding into
both. Cold molecular components are still present in the simulation,
albeit at a 20\% fraction rather than the 50\% fraction seen without
feedback. They have not been destroyed by the supernova, even if the
structure of the cloud has been disrupted. It should be noted though
that this result is at the limit of applicability of these
simulations, as numerical issues are present due to the forward SN
shock having long since passed off the grid.
In the case of the 60\,M$_{\odot}$ star, we observe many similarities
to the case of the 40\,M$_{\odot}$ star. The 120\,M$_{\odot}$ star
also shows similar changes, though they are more extreme, during the
stellar wind phase, with less than half the cold phase material
present by the time of supernova, as compared to the reference case,
and less than all the other cases. An appreciable fraction of hot
thermal phase material is also present, around 5\% by the time of
supernova. After the SN, the first Myr is also similar to the
preceding cases, but after that the evolution is markedly different to
the preceding cases. There is no leveling off of the cold phase
fraction, nevermind an increase. By the end of the simulation, less
than 5\% of the material in the computational volume is in the cold
phase. The volume is completely dominated by warm phase material -
nearly 80\% of the total material present - with the remainder in the
unstable phase. This is most likely because of the combined strength
of the stellar wind and subsequent supernova which completely destroy
the parent molecular cloud. Any remaining cold molecular material has
been blown out of the computational volume. We can therefore state
that it takes more than a 60\,M$_{\odot}$ star to completely disperse
a molecular cloud by mechanical feedback under these conditions, but
that a 120\,M$_{\odot}$ star is very efficient at this dispersal.
In Fig.~\ref{wndphases} we plot the mass-weighted temperature-density
(in the left column) and pressure-density (in the right column)
distributions for the four cases considered in this
work, using 200 bins in log density and 200 bins in log
temperature/pressure. We also over-plot the approximate thermal
equilibrium curve for the heating and cooling prescriptions used in
this work. In the first row, we show the distributions 10.97\,Myrs
through the wind phase for the case of the 15\,M$_{\odot}$ star,
shortly before complete cloud collapse. Most of the material remains
in thermal equilibrium, tracing the equilibrium curve. The two
distinct stable phases (warm and cold) are, as expected, in
approximate pressure equilibrium with each other. A low density,
higher temperature stellar wind bubble (as shown in the bottom row of
Fig.~\ref{wnd-15M}) is indicated by the branch stretching away from
the equilibrium curve horizontally to the left at the same P/k
$\sim10^{3.8}$\,K\,cm$^{-3}$ in the pressure-density distribution.
In the second row of Fig.~\ref{wndphases}, we show the distributions
for the case of the 40\,M$_{\odot}$ star at the end of the wind
phase. The distributions have the same components on the equilibrium
curve that are present in the case of the 15\,M$_{\odot}$ star, but
there are also other considerable regions. We can identify branches
from gas in the wind bubble and the LBV/WR shell. The lower pressure
horizontal branch at P/k$ \sim10^{3}$\,K\,cm$^{-3}$ is the wind
bubble, at notably lower pressure and temperature than the previous
case. The LBV/WR shell that formed during the last stages of stellar
evolution is responsible for the branch and arc of material out of
equilibrium at higher pressure. Specifically, the diagonal line that
spans from (log(P),\,log($\rho$)) = (6,-1) down toward the tunnel
branch at (log(P),\,log($\rho$)) = (3,-3) is the wind injection region
and region of undisturbed wind material up to the reverse shock. The
reverse shock is indicated by the jump back up in pressure to the
over-pressured (with respect to the cloud and other wind bubble
tunnel) LBV/WR shell that is itself indicated by the horizontal branch
at P/k\ $\sim10^{5}$\,K\,cm$^{-3}$ that arcs back to the equilibrium
curve, broadening as it cools and increases in density. The greater
power of the wind has also driven a broadening of the distributions,
with heated material above the equilibrium curve radiatively cooling
towards the equilibrium curve. Cooling by expansion has led to low
density, cool gas below the equilibrium curve. The
distributions in this case are very similar to the case of the
40\,M$_{\odot}$ star in Paper II, despite the marked difference in
structures; a bubble in a clumpy cloud here, a tunnel through a
filamentary sheet there. Common to both cases and responsible for much
structure in these distributions is the LBV/WR shell
structure. Looking at the distributions, it is impossible
to discern the considerably different overall wind-cloud structure
responsible for each. It may be possible to conclude that lower
pressure wind branches are indicative of larger, less confined
structures.
The third and fourth rows of Fig.~\ref{wndphases} show the
distributions for the cases of the 60 and 120\,M$_{\odot}$ stars and
are remarkably similar. Both contain isolated diagonal lines that are
again indicative of the wind injection regions and regions of
undisturbed WR wind material up to the reverse shock. In both cases,
the jump back up in pressure at the reverse shock is at, or just off,
the lower density axis boundary of each figure, as can in fact be seen
at low density in Figs.~\ref{wnd-60M} and \ref{wnd-120M}. The greater
power of the winds and merger of the LBV/WR material with the whole
wind bubble has created a single, broader distribution, pushing more
and more material out of equilibrium, until in the case of the
120\,M$_{\odot}$ star, most of the material is out of equilibrium.
In Fig.~\ref{snphases} we show the distributions 30\,kyrs after the SN
for the three simulations with supernovae. Towards the upper left of
each distribution is the low density, hot phase, which consists of SN
shock heated gas. It is not yet in pressure equilibrium with the warm
phase, but evolves towards this equilibrium as the simulation
progresses beyond that shown here. The distribution at low density is
reasonably wide in pressure and temperature and widens with increasing
stellar mass. The reason for this is the presence of both shock heated
and adiabatically cooled gas, as noted in a similar analysis of such
distributions influenced by feedback simulated by \cite{walch15b}. The
SN has also shock-heated an appreciable amount of the cold dense
material, spreading the distribution upwards at high density in all
three plots.
In Fig.~\ref{snlatephases} we show the mass-weighted
temperature-density and pressure-density distributions at
late times after the SN for the case of the 60\,M$_{\odot}$ star. We
can see the broad distribution is cooling (shifting downwards) in both
plots. The material is also moving back towards thermal equilibrium,
and by 4.5\,Myrs post-SN essentially all the material is back in
thermal equilibrium. The evolution of the 40\,M$_{\odot}$ and
120\,M$_{\odot}$ star cases is very similar, albeit with lower amounts
of high density material and lower maximum densities in each case
(i.e. less material on the right of equivalent plots to those shown in
Fig.~\ref{snlatephases}).
\subsection{Total cloud mass}
\begin{figure}
\centering
\includegraphics[width=80mm]{figure23.pdf}
\caption{Variation of total mass in the computational volume with time
for the feedback simulations. The reference case without feedback is
not shown, but almost precisely overlies the 15M$_{\odot}$ star case.
Raw data: doi.org/10.5518/201.}
\label{cloudmass}
\end{figure}
\begin{figure}
\centering
\includegraphics[width=80mm]{figure24.pdf}
\caption{Variation of gravitationally bound mass (i.e. kinetic energy plus
gravitational energy is less than zero) in the computational volume with time
for the feedback simulations. Raw data: doi.org/10.5518/201.}
\label{boundmass}
\end{figure}
\begin{figure*}
\centering
\includegraphics[width=160mm]{figure25.pdf}
\caption{Levels of mixing of the wind material with the cloud material
throughout the stellar feedback simulations. A value of 100\% indicates
no mixing of the wind material with the cloud material.
Raw data: doi.org/10.5518/201.}
\label{windmixing}
\end{figure*}
\begin{figure*}
\centering
\includegraphics[width=160mm]{figure26.pdf}
\caption{Levels of mixing of the supernova material with the cloud/wind material
throughout the stellar feedback simulations. A value of 100\% indicates
no mixing of the SN material with the cloud/wind material.
Raw data: doi.org/10.5518/201.}
\label{snmixing}
\end{figure*}
In Fig.~\ref{cloudmass} we show the variation of total mass in the computational
volume with time for the case of each star. The reference case is not shown, as it overlaps
the 15\,M$_{\odot}$ star case, highlighting again the little influence
of the 15\,M$_{\odot}$ star. In the 40\,M$_{\odot}$ star case, the
total mass remains approximately the same until the supernova (at
32.1\,Myrs), when the SN event drives mass out of the computational
volume. The wind phase
of the 60\,M$_{\odot}$ star drives a few hundred solar masses of
material out of the computational volume, before the supernova (at 31.1\,Myrs) expels
considerably more mass. In both the 40 and 60\,M$_{\odot}$ star cases,
the majority of mass remains on the grid post-SN until beyond
the end of the simulations. Most of it is out of equilibrium in the
thermally unstable temperature range, as revealed by the preceding
plots in this sub-section. Even by the time of supernova (at
30.1\,Myrs), the wind from the 120\,M$_{\odot}$ star has already
driven nearly 2,000\,M$_{\odot}$ of material out of volume - more than
10\% of the cloud's original 16,500\,M$_{\odot}$. Most of this
expulsion occurred during the stellar evolution off the Main Sequence,
during the final Myr or so of the star's life. The expulsion of mass
continues during the SN phase, until 6\,Myrs post-SN, only 20\% of the
initial mass remains in the simulation. As such, the cloud has been
effectively dispersed and a larger simulation is required to track the
final properties of this material. This would suggest that whilst the
star has destroyed the cloud and replaced it with a tenuous warm-phase
SNR surrounded by warm medium, some cold material may have survived
the violent evolution of the 120\,M$_{\odot}$ star, but is now
entirely separate from its origins in the parent molecular cloud.
In Fig.~\ref{boundmass} we show the variation of mass gravitationally
bound to the potential of the cloud in the computational volume with time
for the case of each star.
That is, we show the amount of mass within the gravitational potential
well of the original cloud that does not have enough kinetic energy to
escape the gravitational potential of the cloud.
Here is the only place the 15\,M$_{\odot}$ star has an appreciable
effect over the reference case, although by the end of the star's evolution,
the amount of gravitationally bound mass has returned to the same level as the
case without
feedback. In the 40\,M$_{\odot}$ star case, the main sequence evolution
of the star provides enough energy for nearly half the mass of the cloud to
overcome the gravitational energy and either leave the domain, or have
enough kinetic energy to overcome the gravitational potential, but the late stages of evolution increase
the amount of mass associated with the cloud. The effect of the supernova
is to disperse the remainder of the cloud, although 2\,Myrs after the SN, the
amount of gravitationally bound mass is increasing again, indicating that the
increasing levels of cold material noted previously are now bound to the
cloud itself. The 60\,M$_{\odot}$ star shows a similar, but more disruptive
pattern, with the cold molecular material formed more than 2\,Myrs post-SN
now associated with the cloud. There is considerably less material in this
case. The 120\,M$_{\odot}$ star is able to
completely disperse the cloud material during the star's lifetime ($\sim$3\,Myrs),
confirming the conclusions drawn above. No increase post-SN is noted.
\subsection{Mixing}
In Fig.~\ref{windmixing} we show the mixing of the stellar wind
material with cloud material. In the 15\,M$_{\odot}$ star case, wind
material is immediately well mixed and remains that way, since it is
all at fractions of no higher than 20\% with respect to the cloud
material. Whilst an equivalent amount of material to the star's mass
was taken out of the injection region to make the star, some cloud
material remained in the injection region and the wind material was
introduced as a source term adding to this, hence the wind material
mixed very quickly. There is no brief period when the
injection region is almost entirely wind material i.e. there is no
transition observed from ``no mixing" (at 100\%) down to well
mixed. Therefore, the initial few thousand years of stellar wind
may be quite effectively quenched, but this is a transitory period
and outweighed by the benefit of using a source term rather than
an imposed wind structure in those cells, which tests and previous
experience \citet[e.g.][and other works therein]{wareing07} has
shown can impose grid-based structure upon the resulting wind.
The wind from the main sequence phase of the 40\,M$_{\odot}$
star is almost as well mixed as that of the 15\,M$_{\odot}$ star. In
this case, the brief initial period of mixing down from almost 100\%
in the injection region is indicated by a very small initial
peak, due to the stronger stellar wind.
However, the final 0.5\,Myrs of stellar wind injection shows the
isolated and unmixed nature of the LBV/WR shell, and also how much
material is in the shell, given the height of the peak. At
32.08\,Myrs, the SN explodes and in short time mixes the wind material
down to concentrations of below 20\% wind material. Both the 60 and
120\,M$_{\odot}$ star cases show peaks of unmixed material in the
final stages of stellar evolution, but with decreasing levels, down to
the 120\,M$_{\odot}$ star case where the majority of wind material
remains efficiently well-mixed throughout the star's lifetime, albeit
with larger fractions of other levels of mixing than in any of the
other cases. Peaks and troughs in the final stages of stellar
evolution can be attributed to the different phases of the stellar
wind and transitory shells that arise during these stages. The effect
of the supernova on the mixing of wind material is similar and is very
efficient in all cases.
In Fig.~\ref{snmixing} we show the mixing of supernova material with
cloud and wind material in the three supernova cases presented in this
work. In panel (a) we focus upon the early mixing of the supernova
ejecta, in the first few hundred thousand years following the
explosion of the 40\,M$_{\odot}$ star. Initially all the SN material
is unmixed (80-100\%), remaining so for a few thousand years and then
is slowly mixed down into the more mixed brackets in this
figure. Within 0.1\,Myrs though, the SN material is efficiently mixed
down to 0-20\% fractions within the cloud/wind material. At later
times, the SN material continues to remain well mixed. Mixing is more
efficient, taking less time to completely reach the 0-20\% brackets
for the more massive stars.
For these simulations, as for those in Paper II, we have not
quantitatively examined the nature of the material leaving the
grid. In the three more massive cases, the SN drives material out of
the cloud, as investigations of the phase fractions of material have
shown. Further investigations may shed more light on mass-loading and
the entrainment of material leaving the grid.
\section{Discussion}\label{discussion}
\subsection{Comparisons to previous work}
Comparing the present work with our previous efforts in the
hydrodynamic case \citep{rogers13,rogers14} is not straightforward due
to the large differences in initial conditions. However, there are
some clear similarities. In both, the later wind energy and SN energy
is transferred to the wider surroundings through a multitude of porous
channels. In Rogers \& Pittard this was as a result of the imposed
initial condition and was associated with a much higher density
filamentary cloud in a smaller volume. Here, the early phases of the
stellar wind result in similar channels and gaps, even though the
initial condition is that of a clumpy cloud with no filamentary
network. This finding, of clear gaps and channels carved out through
parental molecular clouds, would appear to be a common evolutionary
conclusion when mechanical feedback from stars is involved. Only when
a magnetic field is present, effectively focussing the outflow from a
massive star into a single channel as we found in Paper II, are the
results morphologically and quantitatively different. Nevertheless, it
is clear that winds are on the whole quite capable of breaking out of
their confining clouds. Such βleakageβ of the hot gas is consistent
with the much lower mass of hot cluster gas than expected for the
cluster ages and the mass-loss rates of stars \citep{townsley03}
and by energy budget considerations \citep{rosen14}.
In common with Rogers \& Pittard, the densest molecular regions are
surprisingly resistant to ablation by the stellar/cluster wind. In
both works this is partially due to shielding by other dense regions
closer to the star/cluster. Pressure gradients within the flow appear
to rearrange the molecular cloud structure into radially-aligned
spokes or filaments, so that the head of the spoke closest to the star
shields the remainder of the structure. Further simulations are
required to investigate whether this behaviour extends to clumpy
clouds of higher mass and density.
In common with all our previous work, SNe are able to transport large
amounts of energy directly out of the parent cloud, with existing
channels and gaps allowing weaker coupling to the remaing dense
material. Such channels are further carved out by the SN. The key
factor is the shaping effect of the pre-SN stellar winds, which make
the cloud highly porous \citep[][and the current paper]{rogers13}, or
as in Paper II where they are focussed by magnetic fields to open up
large-scale channels directing SN energy out of the cloud. In all
cases, the winds appear to be better at removing molecular material
from the cloud/cluster environment, despite typically injecting less
energy than the SN(e). The molecular material in all cases was found
to be almost completely removed from the original cloud after 6\,Myr
post-SN. These works together demonstrate the complexity of the
interaction of a stellar wind with an inhomogeneous environment. The
results are far removed from simple spherically symmetric models and
compare nicely with the simulations of \cite{geen16} who argue that
strong pre-supernova (radiative) feedback is required to allow
supernova blastwaves to propagate efficiently into the interstellar
medium.
A key common finding between this work and Paper II is that lower mass
stars (15\,M$_{\odot}$ or less) have little global effect on the
parent cloud. Here we find that the weak wind is unable to overcome
the gravitational collapse of the cloud. In Paper II, the local small
bipolar wind bubble is eventually refilled by the slow wind from the
late stages of the star's evolution, as we also note here. The effect
in both cases is as if the stellar wind had not been present at all.
\cite{harper09} considered stellar wind feedback into an inhomogeneous
environment and postulated that the non-uniform surrounding medium
causes gaps in the swept-up shell surrounding the wind-blown bubble
where some of the high-pressure gas in the bubble interior can leak
out. We have simulated a scenario very similar to this, with a stellar
wind expanding into a non-uniform, inhomogeneous clumpy cloud. For the
intermediate mass cases (stars of mass 40\,M$_{\odot}$ and
60\,M$_{\odot}$) we find exactly that - the wind shell is
non-spherical and has gaps or channels through it which allow gas to
vent or blow out of the parent cloud. In the case of the
15\,M$_{\odot}$ star, this effect is far less pronounced - the bubble
which forms remains close to the star (within 10 or so parsecs) and no
gaps or channels are seen around or outside the bubble. At the other
extreme, the wind-bubble around the 120\,M$_{\odot}$ star destroys
half the cloud by the end of the wind phase, and channels are present
through the remaining cloud. In this case, dramatic structural changes
occur within the cloud near the end of the star's life due to the
increasingly powerful wind blowing at this time.
The effect of stellar winds on molecular clouds has also been
considered by \cite{dale15} and \cite{offner15}. These simulations
used either momentum driven or isothermal winds and so give a lower
limit to their impact. With this is mind, \cite{offner15} concluded
that stellar mass-loss rates for individual stars must be greater than
10$^{-7}$\,M$_\odot$\,yr$^{-1}$ in order to reproduce shell
properties. As we noted in Paper II, the mass-loss rates for the
15\,M$_{\odot}$ star are less than
$3 \times 10^{-7}$\,M$_\odot$\,yr$^{-1}$ this limit for the entire MS
evolution of the star (see figure 2 of Paper II). When the mass-loss
rate exceeds this limit the star is in its RSG phase, so the wind is
slow and has little power. On the other hand, the higher mass stars
always have mass-loss rates above this limit (as shown in
Figs.~\ref{60evolution} and \ref{120evolution} of the current paper
and figure 3 of Paper II). Therefore our findings are not in
disagreement with \cite{offner15}, although we have not tested stars
with masses between 15\,M$_{\odot}$ and 40\,M$_{\odot}$. The wind from
the 15\,M$_{\odot}$ star has little effect on the parent cloud,
restricted to a 10\,pc radius around the star. In contrast, the winds
from higher mass stars strongly affect the parent molecular cloud.
Numerous recent works have explored the effects of supernovae on the
multi-phase ISM and parent molecular clouds
\citep[e.g.][]{gatto15,walch15,walch15b,giri16,kortgenphd,kortgen16}.
However, there is still much disagreement on the ability of supernovae
to drive outflows. For instance, \cite{giri16} found that strong
outflows were only generated when SNe were randomly positioned and had
the opportunity to inject energy into relatively low density
environments. SNe placed at density peaks instead radiate away too
much energy to drive any noticeable outflow. However, they did not
account for the effect of the SN-preceding stellar wind. In contrast,
\cite{simpson16}, who accounted for the production of cosmic rays in
supernova events, showed that outflows can be driven from SNe placed
at density peaks. They found their outflows have similar mass loading
as obtained from random placement of SNe with no cosmic rays.
\cite{kortgen16} find that single SNe disperse $\sim$10\,pc-sized
regions but do not disrupt entire clouds, which instead requires
clustered, short-interval SNe to form large hot bubbles. This is in
contrast to our findings, but in mitigation our simulations focus on
feedback in a cloud of lower mass than the \cite{walch15} and
\cite{kortgen16} studies, and also include the effect of the stellar
wind, so it is not unexpected that our conclusions are somewhat
different. Compared to these other works, our simulations clearly show
that when stars are placed at high density locations, their stellar
winds can create gaps and carve channels out of the clumps and the
parent cloud, allowing wind material and SN material to escape easily
from the cloud, mass-loaded with material stripped from clumps in the
parent cloud (\cite{fierlinger16} also find that pre-SN feedback
enhances the impact of SNe). Thus the prior influence of winds should
be included in simulations involving SN feedback.
\subsection{Comparison to observations}
\cite{lopez11} support the scenario presented by \cite{harper09},
concluding that leakage through gaps in the swept-up wind-blown bubble
may be occurring within 30 Doradus. They claim that the lower X-ray
gas pressure relative to the direct radiation pressure suggests that
the hot gas is only partially confined and is hence leaking out of
``pores'' in the H{\sc ii} shell. They also conclude that the
significant radiation pressure near the central star cluster indicates
that direct stellar radiation pressure dominated the interior dynamics
at earlier times, but this claim has proved far more controversial,
and other works are in favour of the thermal pressure of hot X-ray
emitting plasma shaping the large-scale structure and dynamics in 30
Doradus. For example, \cite{pellegrini11} find that the dynamics and
large-scale structure of 30 Doradus are set by a confined system of
X-ray bubbles in rough pressure equilibrium with each other and with
the confining molecular gas. In both the 60 and 120\,M$_{\odot}$ cases
presented herein, we find a structure of wind and cloud material in
rough pressure equilibrium, as shown in Fig.~\ref{wndphases}. The
existence of low-density channels carrying hot wind material is also
clear from Figs.~\ref{wnd-40M} and \ref{wnd-120M}. The more powerful
the wind, the larger the cavity, the quicker the channels form and the
wider they are, implying that a cluster of stellar winds will rapidly
form multiple wide channels through a low-mass parent molecular cloud
such as studied here. However, to determine whether hot gas thermal
pressure or direct radiation pressure is dominant requires simulations
with both mechanical and radiative feedback included.
\cite{lopez14} further assessed the role of stellar feedback at
intermediate scales of 10-100\,pc. They studied a sample of 32 H{\sc
ii} regions (with ages from 3-10\,Myr) in the Large and Small
Magellanic Clouds and found that warm ionized gas dominates the
dynamics whilst the hot gas pressures are comparatively weak. They
emphasize that younger, smaller H{\sc ii} regions should be studied to
elucidate the roles of hot gas and direct pressure in the early
evolution of these regions. They also conclude that the hot gas is
only partially confined in all of their sources, and suggest that hot
gas leakage is common. Our simulations are by design influenced only
by hot gas pressure, but nevertheless create the kind of structures
observed by \cite{lopez14} and others.
It is also interesting to note the existence of structures within
H{\sc ii} regions. For example, in RCW 120 arcs of dust emission are
clearly seen within the H{\sc ii} region bubble in mid-IR {\it
Spitzer} data. \citep{mackey16} interpret this as the outer edge of
the wind-blown-bubble. An alternative viewpoint is that dust arcs are
waves induced by photo-evaporation flows inside H{\sc ii} bubbles
\citep{ochsendorf14}. Our simulations, although not able to define
the extent of the H{\sc ii} region, do show complex structure within
the wind bubble itself, for example bow shocks and LBV/WR shells,
which could perhaps also lead to such observed structures. Such arcs
could also be related to the aligned radial structures formed in the
simulation of the 60\,M$_{\odot}$ star. Photoionisation of the complex
structures formed by the 60 and 120\,M$_{\odot}$ stars is bound to
show one-sided H{\sc ii} regions too, as the stellar wind has blown
out the molecular cloud in unequal measure around the star, as shown
in Figure \ref{wnd-proj}.
Some H{\sc ii} regions (e.g. RCW 79, RCW 82, RCW 120) show a central
hole in the 24\,$\mu$m emission \citep[e.g.][]{martins10}, which may
also be evidence of stellar winds. However, these holes could also have been
produced by radiation pressure. Simulations by
\citet{freyer03,freyer06} show that stellar winds can have important
dynamical effects even when the ratio of the injected wind to ionizing
photon energy is as low as 0.01, but it is not known if this remains
true at still lower ratios.
Two ``smoking guns'' which reveal strong evidence for the potential
impact of stellar winds are bowshocks around stars on the periphery of
stellar clusters \citep[e.g.][]{winston12}, and diffuse X-ray emission within
and around young pre-SN clusters
\citep[e.g.][]{gudel08,townsley14}. However, it is clear that further
observational and theoretical work is needed for a better
understanding of the dynamics of H{\sc ii} regions and
wind-blown-bubbles/cluster outflows.
\section{Summary and conclusions}\label{conclusions}
In this work we have explored the effects of mechanical stellar wind
and supernova (SN) feedback on realistic molecular clouds without
magnetic fields. Our initial condition has been based on the work of
\cite{wareing16} in which a diffuse atomic cloud was allowed to form
structure through the action of the thermal instability, under the
influence of gravity, but without injected turbulence. The resulting
structure is best described as a clumpy near-spherical cloud of
approximately 100\,pc-diameter surrounded by a diffuse atomic cloud,
when we introduce mechanical stellar feedback. The clumps are roughly
equally distributed over the inner region of the cloud (r$\leq$50\,pc)
when the densities in some clumps reach 100 cm$^{-3}$. A single
massive star was then introduced at the highest density location
closest to the centre of the cloud. We considered four cases:
formation of a 15\,M$_{\odot}$ star, a 40\,M$_{\odot}$ star, a
60\,M$_{\odot}$ star and a 120\,M$_{\odot}$ star. Their stellar winds
(based on realistic Geneva non-rotating stellar evolution models)
subsequently affect the parent cloud, and at the end of life of the
three highest mass stars a SN explosion is modelled. We do not model a
SN explosion in the case of the 15\,M$_{\odot}$ star as the cloud
collapses under gravity before the end of the star's life and the
simulation loses meaning at the resolution available.
In the 15\,M$_{\odot}$ star case the stellar wind has very little
effect on the molecular cloud, and forms only a small cavity around
the star which reduces in size as the cloud collapses and the star
enters the final stages of stellar evolution. The stellar wind is
unable to support the wind-blown bubble against the surrounding
gravitational collapse of the cloud.
In the more massive star cases, the stellar wind has a significant
effect on the molecular cloud, carving channels and gaps through the
cloud which are reminiscent of the interaction seen in our previous
work \citep{rogers13,rogers14}. Each star's wind halts the
gravitational collapse of the cloud. The density, temperature,
pressure and velocity of material in the cloud environment all span
many orders of magnitude. The hottest gas typically occurs at the
reverse shock of the stellar wind, and cools as it expands away from
the cloud and mixes in with denser surrounding material. A multitude
of weaker shocks form around the remaining dense clumps within the
cloud and material is ablated from these into the stellar wind flow,
as they also rearrange and form radially aligned spokes in the cloud.
This change of structure allows self-shielding to occur and for cold
molecular cloud material to survive for the lifetime of the star, both
in these shielded spokes and entrained into the wind escaping the
cloud along the channels.
In the most massive (120\,M$_{\odot}$) star case, nearly half the
molecular cloud is dispersed and destroyed during the final stages of
stellar evolution, before the SN occurs. The channels through the
wind bubble allow the forward shock of the SN to leave the wind
structure very quickly through these 'leaky gaps'. In the cases of both
the 60\,M$_\odot$ and 120\,M$_\odot$ stars, along these wind-carved
channels the forward shock leaves the computational volume after only
15,000 years, at an average speed of over 4,500 km\,s$^{-1}$, efficiently
transporting kinetic and thermal energy into surroundings beyond the
parent molecular cloud. The SN subsequently continues
this dispersal and destruction, overrunning the structures formed
during the wind phase. In all three cases, by 2.5\,Myrs after the SN,
material with the characteristics of the warm neutral medium surrounds
a low-density inner cavity (the central region of the SNR) at the
original location of the isolated star.
It can be concluded from this work that stellar winds from the lower
mass end of the range of stars undergoing core-collapse have little
effect on their parent cloud. In such cases, the cloud is likely to
evolve subject to gravity, radiation pressure, and external
influences, until the star explodes. In contrast, higher mass stars
are able to disperse and destroy the cold molecular material of the
parent cloud even before SNe occur. Massive stars of intermediate mass
carve channels and gaps through their parent clouds, allowing SN
material to remain fairly hot and energetic as it escapes \citep[see
also][]{rogers13,rogers14,wareing17}.
In this work, as in \citet{wareing17}, we have taken a single initial
condition, realistically formed from the action of the thermal
instability, but none-the-less in isolation and at the lower end of
molecular cloud masses. Molecular cloud masses in the Milky Way reach
10$^{5-6}$\,M$_{\odot}$, considerably more than the
16,500\,M$_{\odot}$ of material in the cloud investigated here. The
nature and distribution of this material is key to how the stellar
winds and SN affect the molecular cloud. In future work we will
examine feedback into higher mass clouds and also account for
radiative feedback effects.
\section*{Acknowledgments}
This work was supported by the Science \& Technology Facilities
Council [Research Grant ST/L000628/1]. The calculations for this paper
were performed on the DiRAC Facility jointly funded by STFC, the Large
Facilities Capital Fund of BIS and the University of Leeds and on
other HPC facilities at the University of Leeds. These facilities are
hosted and enabled through the ARC HPC resources and support team at
the University of Leeds (A. Real, M.Dixon, M. Wallis, M. Callaghan \&
J. Leng), to whom we extend our grateful thanks. We acknowledge
useful discussions with T. W. Hartquist and S. Van Loo and extend
further thanks to S. Van Loo for the provision of analysis routines to
produce PDFs. We also express thanks to the Reviewer, Prof. Vincent
Icke, and Editors for their comments on the first draft of this manuscript.
Raw data for the all figures in this paper is available
from the University of Leeds Repository at http://doi.org/10.5518/201.
VisIt is supported by the Department of
Energy with funding from the Advanced Simulation and Computing Program
and the Scientific Discovery through Advanced Computing Program.
| 7ff24db917bc13733343307a034b9df371c4e2ad | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:intro}
The explosive growth of mobile traffic demand is the driving force
behind the development of signal processing and communication technologies to
significantly upgrade the high-end experiences of
communication such as high throughput, high reliability, and ubiquitous access.
It is widely believed that interference-limited techniques such as
coordinated multipoint transmission (CoMP) \cite{GH10,JYH12}, which treat interference as noise
cannot meet the edge throughput requirements for 5G \cite{Anetal14}. Dirty-paper coding (DPC) \cite{WeingartenIT06}, under which the interference is successively mitigated, can improve both the edge and sum throughput but is difficult to implement in practice and remains only as a theoretical concept due to its high computational complexity. Nonorthogonal multiple access (NOMA)
has been recently recognized as an essential enabling technology for 5G systems \cite{DingMag16} due to its potential
to improve the edge throughput \cite{Saietal13}.
In NOMA, a base station (BS) transmits a signal superposition to all users. The users are paired so that
in each pair there is one with a better channel condition and another with a poorer channel condition \cite{Benjebbour14}.
The messages intended for each pair of users are sequentially decoded as follows. First,
the message for the user with the poorer channel condition is decoded by
both users. The message for the user with the better channel condition is then successively decoded
by this user after canceling the interference from the other user \cite{DingTVT15}. Thus, while the throughput
at the users with poorer channel condition remains the same as that in interference-limited techniques, the throughput
at the users with better channel condition is clearly improved, leading to a higher system throughput.
Multi-input multi-output (MIMO) is widely known for its enormous potential in improving the capacity of wireless communication systems without requiring extra bandwidth or power. NOMA for MIMO communication
(MIMO-NOMA) in single-cell systems for achieving higher throughput has been investigated in \cite{DingTWC16,DSP16}, and an extension to multi-cell cases has been considered in \cite{ShinComL16}. In multi-cell systems, the effects of inter-cell interference are acute and unpredictable, limiting the quality-of-service (QoS) for cell-edge users. It is therefore challenging to realize the benefit that NOMA may bring to multi-cell systems.
\subsection{Related Works}
In this subsection we discuss the state-of-the-art of signal processing techniques
for NOMA downlink transmission. NOMA was mostly studied for single-cell
multi-input single-output (MISO) systems known as MISO-NOMA, where the multiple-antenna BS
broadcasts signal superpositions to single-antenna users. Under the assumption on low QoS requirement for
the near user (with a good channel condition) and high QoS requirement for the far user (with a poor channel condition) in
a two-user MISO-NOMA, \cite{Choi15} proposed a heuristic computational procedure with neither convergence nor
optimality guaranteed for the beamforming power minimization. Under
similar users' QoS requirements in a $2K$-user MISO-NOMA, it used a particular zero-forcing beamformer
to cancel the inter-pair interference, so the problem of $2K$-user MISO-NOMA beamforming is decomposed
into $K$ independent subproblems of two-user MISO-NOMA beamforming. A closed-form solution
for minimization of beamforming power in two-user MISO-NOMA subject to natural users' QoS requirements
was obtained in \cite{ChenCL16,ChenTSP16,ChenAC16}. In \cite{HDRK16}, users performed
successive interference cancellation (SIC) based on the channel gain differences.
Its proposed algorithm for beamforming is of high computational complexity.
Regarding MIMO-NOMA, \cite{DingTWC16} and \cite{DSP16} derived the outage probability experienced
by users in zero-forcing postcoding or signal alignment. Power allocation for achieving the ergodic capacity of
two-user MIMO NOMA was considered in \cite{SunWCL15} and \cite{ChoiTWC16}.
User-pairing to enhance the throughput of users of poor channel condition was proposed in \cite{LuiCL16}.
\cite{ShinComL16} proposed two interference alignment based coordinated beamforming
for a two-cell MIMO-NOMA, where the interference at all cell-center users and edge-center users is canceled.
\subsection{Motivation and Contributions }
The paper considers the problem of designing linear precoders/beamformers at the BSs for MIMO-NOMA multi-cell systems
to maximize their sum throughput while simultaneously meeting the users' QoS requirements.
In general, such a design problem is very complicated as
the objective function is nonlinear and nonsmooth, and the QoS constraints are highly nonconvex, for which
even finding a feasible point is already challenging.
The main contributions of the paper are three-fold:
\begin{itemize}
\item For MIMO-NOMA, two path-following optimization algorithms are proposed for computation, which
at least converge to a locally optimal solution.
At each iteration, the first algorithm invokes a convex quadratic program while the second
algorithm invokes a semi-definite program (SDP). Both these convex programs are of moderate dimension so
their computation is very efficient.
\item Another path-following algorithm tailored for MISO-NOMA is proposed, which explores much simpler structures of
the throughput functions in MISO systems for more efficient computation.
\item The provided numerical results show the essential performance improvements of NOMA based systems compared to
the conventional systems. The capability of NOMA to improve both edge and sum throughput is revealed.
\end{itemize}
\subsection{ Paper Organization and Notation}
The rest of the paper is organized as follows. Section~\ref{sec:sys_model} presents the system model and formulates the problem. A convex quadratic programming based path-following algorithm for the MIMO-NOMA problem is developed in Section~\ref{sec:CQBI}, while another
SDP based path-following algorithm is developed in Section~\ref{SDPBI}. Section~\ref{sec:TL} devotes to computation for MISO-NOMA problem. Numerical results are provided in Section~\ref{sec:simulation}, and Section~\ref{sec:conclusion} concludes the paper.
\textit{Notation.} Bold-faced upper-case letters are used for matrices, bold-faced lower-case letters
are used for vectors, and lower-case letters are used for for scalars.
$\mathbf{I}_n$ is the identity matrix of size $n\times n$.
$\mathbf{X}^{H}$, $\mathbf{X}^{T}$, and $\mathbf{X}^{*}$ are the Hermitian transpose, normal transpose, and conjugate of a matrix $\mathbf{X}$, respectively. The inner product $\langle{\mathbf{X}, \mathbf{Y}}\rangle$ of matrices $\mathbf{X}$ and $\mathbf{Y}$
is defined as $\textrm{trace}(\mathbf{X}^H \mathbf{Y})$. Denote by $\langle \mathbf{A}\rangle$ the trace of a matrix $\mathbf{A}$, and by $|\mathbf{A}|$ its determinant. $\|\cdot\|$ stands for matrix's Frobenius norm
or vector's Euclidean norm.
For Hermitian symmetric matrices $\mathbf{A}\succeq \mathbf{0}$ ($\mathbf{A}\succ \mathbf{0}$, resp.) means that
$\mathbf{A}$ is a positive semidefinite (positive definite, resp.) matrix. Accordingly
$\mathbf{A} \succeq \mathbf{B}$ ($\mathbf{A} \succ \mathbf{B}$, resp.) means
$\mathbf{A}-\mathbf{B}\succeq \mathbf{0}$ ($\mathbf{A}-\mathbf{B}\succ \mathbf{0}$, resp.).
$\mathbb{C}$ is the set of all complex numbers, and $\emptyset$ is an empty set. $\Re\{x\}$ denotes the real part of a complex number $x$. $\mathbf{x}\sim\mathcal{CN}(\boldsymbol{\eta},\boldsymbol{Z})$ means that $\mathbf{x}$ is a random vector following a complex circular Gaussian distribution with mean $\boldsymbol{\eta}$ and covariance matrix $\boldsymbol{Z}$. $\nabla_{\mathbf{x}}f(\mathbf{x})$ is the gradient of function $f(\cdot)$ with respect to its variable $\mathbf{x}$. $\mathbb{E}\{\cdot\}$ denotes the expectation operator.
\section{System Model and Problem Formulation}\label{sec:sys_model}
This section presents a system model for signal superposition in MIMO-NOMA multi-cell systems
and formulates an optimization problem for the precoder design. Relations
to CoMP and DPC are also briefly clarified.
\subsection{Signal Processing Model}
\begin{figure}[t]
\centering
\includegraphics[width=0.6\textwidth,trim={0.5cm 0.0cm -4.5cm -0.0cm}]{Systemmodel.eps}
\caption{An illustration of the cell of interest in a NOMA system serving $2K$ users.}
\label{fig:SM:1}
\end{figure}
Consider a downlink system consisting of $N$ cells, where
the BS of each cell is equipped with $N_t$ antennas to serve $2K$ users
(UEs) within its cell as illustrated in Fig.~\ref{fig:SM:1}. Each UE is equipped with $N_r$ antennas. In each cell, there are $K$ near UEs
(cell-center UEs), which are located inside the circular area with radius $r_n$ and the BS at its center,
and $K$ far UEs (cell-edge UEs), which are located within the ring area with inner radius $r_n$ and
outer radius $r_o$. $K$ far UEs in each cell are not only in poorer
channel conditions than other $K$ near UEs but also are under
more intensified inter-cell interference from adjacent cells.
Upon denoting $\mathcal{I} \triangleq \{1, 2, \cdots, N\}$ and $\mathcal{J} \triangleq \{1, 2, \cdots, 2K\}$,
the $j$-th UE in the $i$-th cell is referred to as UE $(i,j)\in \mathcal{S}\triangleq \mathcal{I}\times\mathcal{J}$.
The cell-center UEs are UE $(i,j)$, $j\in\mathcal{K}_1\triangleq\{1,\cdots, K\}$ while the cell-edge UEs are
UE $(i,j)$, $j\in\mathcal{K}_2\triangleq\{K+1,\cdots, 2K\}$. Thus the set of cell-center UEs and the set of cell-edge UEs
are $\mathcal{S}_1\triangleq \mathcal{I}\times \mathcal{K}_1$ and $\mathcal{S}_2\triangleq \mathcal{I}\times \mathcal{K}_2$,
respectively. In NOMA each
cell-center UE $(i,j)\in\mathcal{S}_1$ is randomly paired with cell-edge UE $(i,p(j))\in \mathcal{S}_2$ of the same cell to create
a virtual cluster.\footnote{Using more sophisticated user-pairing strategies may improve the performance of MIMO-NOMA networks (see e.g. \cite{DingTVT15,LuiCL16}) but it is beyond the scope of this paper.} The signal superpositions are precoded at the BSs prior to being transmitted to the UEs. Specifically, the message intended for UE $(i,j)$ is
$\mathbf{s}_{i,j} \in \mathbb{C}^L$ with $\mathbb{E}\{\mathbf{s}_{i,j}(\mathbf{s}_{i,j})^H\}=\mathbf{I}_L$, which
is precoded by
matrix $\mathbf{V}_{i,j}\in \mathbb{C}^{N_t\times L}$, where $L$ is the number of concurrent data streams and $L \leq \min\{N_t,N_r\}$.
For notational convenience, let us define
$\mathbf{V} \triangleq [\mathbf{V}_{i,j}]_{(i,j)\in\mathcal{S}}$. The received signals at UE $(i,j)$ and UE $(i,p(j))$ are expressed as
\begin{IEEEeqnarray}{rCl}
\mathbf{y}_{i,j} &=& \mathbf{H}_{i,i,j}\mathbf{V}_{i,j}\mathbf{s}_{i,j}+\mathbf{H}_{i,i,j}\mathbf{V}_{i,p(j)}\mathbf{s}_{i,p(j)} \nonumber\\
& & + \displaystyle\sum_{(s,l)\in {\cal S}\setminus \{(i,j), (i,p(j)) \}} \mathbf{H}_{s,i,j}\mathbf{V}_{s,l} \mathbf{s}_{s,l} + \mathbf{n}_{i,j},\label{eq:yij}
\end{IEEEeqnarray}
and
\begin{IEEEeqnarray}{rCl}
&& \mathbf{y}_{i,p(j)} = \mathbf{H}_{i,i,p(j)}\mathbf{V}_{i,p(j)}\mathbf{s}_{i,p(j)}+\mathbf{H}_{i,i,p(j)}\mathbf{V}_{i,j}\mathbf{s}_{i,j}\nonumber\\
&&\qquad + \displaystyle\sum_{(s,l)\in {\cal S}\setminus \{(i,j), (i,p(j)) \}} \mathbf{H}_{s,i,p(j)}\mathbf{V}_{s,l} \mathbf{s}_{s,l} + \mathbf{n}_{i,p(j)}\quad\label{eq:yipj}
\end{IEEEeqnarray}
where $\mathbf{H}_{s,i,j} \in\mathbb{C}^{N_r \times N_t}$ is the MIMO channel from the BS $s\in\mathcal{I}$ to UE $(i,j)\in\mathcal{S}$. The entries of the additive noise $\mathbf{n}_{i,j}\in\mathbb{C}^{N_r}$ are independent and identically
distributed (i.i.d.) noise samples with zero mean and variance $\sigma^2$.
The covariances of $\mathbf{y}_{i,j}$ and $\mathbf{y}_{i,p(j)}$ are thus
\begin{equation}\label{eqLMij}
{\mathbf{\mathcal{ M}}}_{i,j}(\mathbf{V})= \displaystyle\sum_{(s,l)\in{\cal S}}\mathbf{H}_{s,i,j}\mathbf{V}_{s,l}\mathbf{V}_{s,l}^H \mathbf{H}_{s,i,j}^H + \sigma^2\mathbf{I}_{N_r},
\end{equation}
and
\begin{equation}\label{eqLMipj}
{\mathbf{\mathcal{ M}}}_{i,p(j)}(\mathbf{V})= \displaystyle\sum_{(s,l)\in{\cal S}}\mathbf{H}_{s,i,p(j)}\mathbf{V}_{s,l}\mathbf{V}_{s,l}^H \mathbf{H}_{s,i,p(j)}^H + \sigma^2\mathbf{I}_{N_r}.
\end{equation}
The purpose of the paper is to design complex-valued precoding matrices $\mathbf{V}_{i,j}$ to maximize the overall
spectral efficiency under a given pairing for NOMA. The MIMO channel states $\mathbf{H}_{s,i,j}$
are assumed unchanged during message transmission but may change independently from one message to another
and are perfectly known at all nodes \cite{Choi15,HDRK16,DSP16}.
\subsection{Problem Formulation}
In NOMA, the message $\mathbf{s}_{i,p(j)}$ intended for the cell-edge UE $(i,p(j))$ is decoded by both UE $(i, p(j))$ and
UE $(i,j)$. Then message $\mathbf{s}_{i,j}$ intended for the cell-center UE $(i,j)$ is decoded by itself only.
The cell-edge UE $(i,p(j))$ decodes its own message $\mathbf{s}_{i,p(j)}$ with the achievable rate
\begin{IEEEeqnarray}{rCl}\label{eq:Rijj}
&& {\cal R}_{i,p(j)}^{p(j)}(\mathbf{V}) = \ln\Bigl| \mathbf{I}_L + \nonumber\\
&&\quad (\mathbf{V}_{i,p(j)})^H\mathbf{H}_{i,i,p(j)}^H
{\mathbf{\mathcal{ M}}}_{i,p(j)}^{p(j)}(\mathbf{V})^{-1}\mathbf{H}_{i,i,p(j)}\mathbf{V}_{i, p(j)} \Bigl| \qquad
\end{IEEEeqnarray}
where ${\mathbf{\mathcal{ M}}}_{i,p(j)}^{p(j)}(\mathbf{V})$ is defined by
\begin{IEEEeqnarray}{rCl}\label{eq:Mij}
&&{\mathbf{\mathcal{ M}}}_{i,p(j)}^{p(j)}(\mathbf{V}) \nonumber\\
&&\ \triangleq {\mathbf{\mathcal{ M}}}_{i,p(j)}(\mathbf{V}) - \mathbf{H}_{i,i,p(j)}\mathbf{V}_{i,p(j)}(\mathbf{V}_{i,p(j)})^H\mathbf{H}_{i,i,p(j)}^H \nonumber\\
&&\ =\displaystyle\sum_{(s,l)\in{\cal S}\setminus\{(i,p(j)) \}}\mathbf{H}_{s,i,p(j)}\mathbf{V}_{s,l}\mathbf{V}_{s,l}^H \mathbf{H}_{s,i,p(j)}^H + \sigma^2\mathbf{I}_{N_r}.\quad\,
\end{IEEEeqnarray}
On the other hand, the cell-center UE $(i,j)$ decodes the message $\mathbf{s}_{i,p(j)}$ with the achievable rate
\begin{IEEEeqnarray}{rCl}\label{eq:Rijpj}
&&{\cal R}_{i,p(j)}^j(\mathbf{V}) = \nonumber\\
&&\ \; \ln\Bigl|\mathbf{I}_L + (\mathbf{V}_{i,p(j)})^H\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V})^{-1}\mathbf{H}_{i,i,j}\mathbf{V}_{i,p(j)}\Bigl|\quad
\end{IEEEeqnarray}
where
\begin{IEEEeqnarray}{rCl}\label{eq:Mijpj}
&&{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}) \triangleq {\mathbf{\mathcal{ M}}}_{i,j}(\mathbf{V}) - \mathbf{H}_{i,i,j}\mathbf{V}_{i,p(j)}(\mathbf{V}_{i,p(j)})^H\mathbf{H}_{i,i,j}^H\quad\nonumber\\
&&\qquad=\displaystyle\ds\sum_{(s,l)\in{\cal S}\setminus\{(i,p(j))\}}\mathbf{H}_{s,i,j}\mathbf{V}_{s,l}\mathbf{V}_{s,l}^H \mathbf{H}_{s,i,j}^H + \sigma^2\mathbf{I}_{N_r}.
\end{IEEEeqnarray}
Hence, the throughput by decoding the message $\mathbf{s}_{i,p(j)}$ by UEs $(i,p(j))$ and $(i,j)$ is
\begin{equation}\label{noma1}
{\cal R}_{i,p(j)}=\min \Bigl\{{\cal R}^{j}_{i,p(j)}(\mathbf{V}), {\cal R}^{p(j)}_{i,p(j)}(\mathbf{V})\Bigr\}.
\end{equation}
Next, the throughput by decoding the message $\mathbf{s}_{i,j}$
by the cell-center UE $(i,j)$ after decoding the message $\mathbf{s}_{i,p(j)}$ is
\begin{equation}\label{noma2}
{\cal R}_{i,j}(\mathbf{V}) = \ln\Bigl| \mathbf{I}_L + (\mathbf{V}_{i,j})^H\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{p}(\mathbf{V})^{-1}\mathbf{H}_{i,i,j}\mathbf{V}_{i,j} \Bigl|
\end{equation}
where
\begin{IEEEeqnarray}{rCl}\label{eq:Mipjpj}
&& {\mathbf{\mathcal{ M}}}_{i,j}^{p}(\mathbf{V}) \triangleq {\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}) - \mathbf{H}_{i,i,j}\mathbf{V}_{i,j}(\mathbf{V}_{i,j})^H\mathbf{H}_{i,i,j}^H \nonumber\qquad\quad\\
&& =\displaystyle\ds\sum_{(s,l)\in{\cal S}\setminus\{(i,p(j)),(i,j)\}}\mathbf{H}_{s,i,j}\mathbf{V}_{s,l}\mathbf{V}_{s,l}^H \mathbf{H}_{s,i,j}^H + \sigma^2\mathbf{I}_{N_r}.
\end{IEEEeqnarray}
Our goal is to maximize the total sum throughput of the system under QoS for each individual UE and
power budget at each BS, which is mathematically formulated as
\begin{IEEEeqnarray}{rCl}\label{Rmtgor}
&&\underset{\mathbf{V}}{{\mathrm{maximize}}}\; \mathcal{P}(\mathbf{V})\triangleq\sum_{i=1}^N\sum_{j=1}^K\Bigl({\cal R}_{i,j}(\mathbf{V})+{\cal R}_{i,p(j)}(\mathbf{V})\Bigr)\qquad \IEEEyessubnumber\label{Rmtgora}\\
&&{\mathrm{subject\ to}}\; {\cal R}_{i,j}(\mathbf{V}) \geq r_{i,j},\ \forall i \in\mathcal{I},\ \forall j\in\mathcal{K}_1, \IEEEyessubnumber\label{Rmtgorb}\\
&&\qquad\qquad\; {\cal R}_{i,p(j)}(\mathbf{V}) \geq r_{i,p(j)},\, \forall i \in\mathcal{I},\, \forall j\in\mathcal{K}_2, \IEEEyessubnumber\label{Rmtgorc}\\
&&\qquad\qquad\; \sum_{j\in {\cal J}} \Bigl\langle \mathbf{V}_{i,j}\mathbf{V}_{i,j}^H \Bigl\rangle \le P^{\max}_{i},\ \forall i \in\mathcal{I}\IEEEyessubnumber\label{Rmtgord}
\end{IEEEeqnarray}
where $P^{\max}_i$ in \eqref{Rmtgord} is the transmit power budget of BS $i$. The QoS constraints
\eqref{Rmtgorc} and \eqref{Rmtgord} set a minimum throughput requirement $r_{i,j}$ at the UE $(i,j)$ and $r_{i,p(j)}$ at the UE $(i,p(j))$.
\subsection{Relations to CoMP and DPC}
In CoMP \cite{GH10,JYH12}, the problem of maximizing the sum throughput under QoS constraints is formulated as \cite{TTN16}
\begin{IEEEeqnarray}{rCl}\label{CoMPobj}
&&\underset{\mathbf{V}}{{\mathrm{maximize}}}\;\ \mathcal{P}^{\mathrm{CoMP}}(\mathbf{V})\triangleq\sum_{i=1}^N\sum_{j=1}^{2K}{\cal R}_{i,j}^{'}(\mathbf{V})\IEEEyessubnumber\label{CoMPa}\\
&&{\mathrm{subject\ to}}\;\ {\cal R}_{i,j}^{'}(\mathbf{V}) \geq r_{i,j},\ \forall i \in\mathcal{I},\ \forall j\in\mathcal{J}, \IEEEyessubnumber\label{CoMPb}\qquad\\
&&\qquad\qquad\quad \sum_{j\in {\cal J}} \Bigl\langle \mathbf{V}_{i,j}\mathbf{V}_{i,j}^H \Bigl\rangle \le P^{\max}_{i},\ \forall i \in\mathcal{I}\IEEEyessubnumber\label{CoMPc}
\end{IEEEeqnarray}
where ${\cal R}_{i,j}^{'}(\mathbf{V})$ is given by
\begin{equation}\label{eq:Rij:CoMP}
{\cal R}_{i,j}^{'}(\mathbf{V}) = \ln\Bigl|\mathbf{I}_L + (\mathbf{V}_{i,j})^H\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{j}(\mathbf{V})^{-1}\mathbf{H}_{i,i,j}\mathbf{V}_{i,j}\Bigl|
\end{equation}
with
\begin{equation}
\begin{array}{lll}
{\mathbf{\mathcal{ M}}}_{i,j}^{j}(\mathbf{V}) \triangleq \displaystyle\ds\sum_{(s,l)\in{\cal S}\setminus\{(i,j)\}}\mathbf{H}_{s,i,j}\mathbf{V}_{s,l}\mathbf{V}_{s,l}^H \mathbf{H}_{s,i,j}^H + \sigma^2\mathbf{I}_{N_r}.
\end{array}
\end{equation}
Compared ${\cal R}_{i,j}^{'}(\mathbf{V})$ defined by $(\ref{eq:Rij:CoMP})$ to
${\cal R}_{i,p(j)}(\mathbf{V})$ and ${\cal R}_{i,j}(\mathbf{V})$ defined by (\ref{noma1}) and (\ref{noma2}) one can see that
\begin{IEEEeqnarray}{rCl}
{\cal R}_{i,p(j)}^{'}(\mathbf{V})&=&{\cal R}^{p(j)}_{i,p(j)}(\mathbf{V})\label{com1}\\
&\geq&\min \Bigl\{{\cal R}^{j}_{i,p(j)}(\mathbf{V}), {\cal R}^{p(j)}_{i,p(j)}(\mathbf{V})\Bigr\}\label{com2}\\
&=& {\cal R}_{i,p(j)},\ \forall j\in\mathcal{K}_2,\label{com3}
\end{IEEEeqnarray}
and
\begin{equation}
{\cal R}_{i,j}^{'}(\mathbf{V})\leq {\cal R}_{i,j}(\mathbf{V}),\ \forall j\in\mathcal{K}_1,\label{com4}
\end{equation}
i.e., under the same precoder $\mathbf{V}$, the throughput at cell-edge UEs is higher with CoMP while
that at cell-center UEs is higher with NOMA. Thus, under the same precoder, NOMA does not need to perform better
than CoMP in terms of the total throughput. The preference of NOMA now critically depends on its performance at its
optimal precoder, which is sought in the next sections.
On the other hand, DPC at the BS $i$ with encoding order from UE $(i,2K)$ to UE $(i,1)$
enables UE $(i,j)$ view the massages intended for UEs $(i,j')$, $j'>j$ as non-causually known and
thus cancel them from its received signal.
Hence, the throughput by decoding the message $\mathbf{s}_{i,j}$ for UE $(i,j)$ is defined by
\begin{equation}\label{DPC1}
{\cal R}_{i,j}^{''}(\mathbf{V}) = \ln\Bigl|\mathbf{I}_L + (\mathbf{V}_{i,j})^H\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{i}(\mathbf{V})^{-1}\mathbf{H}_{i,i,j}\mathbf{V}_{i,j} \Bigl|
\end{equation}
where
\begin{IEEEeqnarray}{rCl}\label{eq:DPC2}
{\mathbf{\mathcal{ M}}}_{i,j}^{i}(\mathbf{V}) &\triangleq& \displaystyle\ds\sum_{s \neq i}^{N}\sum_{l = 1}^{2K}\mathbf{H}_{s,i,j}\mathbf{V}_{s,l}\mathbf{V}_{s,l}^H \mathbf{H}_{s,i,j}^H \nonumber\\
&+&\; \sum_{k=1}^{j-1}\mathbf{H}_{i,i,j}\mathbf{V}_{i,k}\mathbf{V}_{i,k}^H\mathbf{H}_{i,i,j}^H + \sigma^2\mathbf{I}_{N_r}.\quad
\end{IEEEeqnarray}
The problem of maximizing the sum throughput under QoS constraints is formulated as
\begin{IEEEeqnarray}{rCl}\label{DPCobj}
&&\underset{\mathbf{V}}{{\mathrm{maximize}}}\;\ \mathcal{P}^{\mathrm{DPC}}(\mathbf{V})\triangleq\sum_{i=1}^N\sum_{j=1}^{2K}{\cal R}_{i,j}^{''}(\mathbf{V})\IEEEyessubnumber\label{DPCa}\\
&&{\mathrm{subject\ to}}\;\ {\cal R}_{i,j}^{''}(\mathbf{V}) \geq r_{i,j},\ \forall i \in\mathcal{I},\ \forall j\in\mathcal{J}, \IEEEyessubnumber\label{DPCb}\qquad\\
&&\qquad\qquad\quad \sum_{j\in {\cal J}} \Bigl\langle \mathbf{V}_{i,j}\mathbf{V}_{i,j}^H \Bigl\rangle \le P^{\max}_{i},\ \forall i \in\mathcal{I}.\IEEEyessubnumber\label{DPCc}
\end{IEEEeqnarray}
Apparently, under DPC the intra-cell interference from UEs of better channel
condition is canceled in decoding the messages intended for UEs of poorer channel condition resulting in much better
edge throughput compared with CoMP. NOMA is advantageous over DPC in terms of ease of implementation.
By increasing the QoSs' requirement in
(\ref{Rmtgorb}), (\ref{CoMPb}) and (\ref{DPCb}), we will see that the NOMA's sum throughput is fully superior over CoMP's one and catches up DPC's one.
\section{Convex Quadratic Based Iterations}\label{sec:CQBI}
Note that the QoS constraints \eqref{Rmtgorb} and \eqref{Rmtgorc} in \eqref{Rmtgor},
and \eqref{CoMPb} in \eqref{CoMPobj} are set beforehand, which are dependent on the UEs' throughput requirements but
independent on their channel condition. In maximizing the sum throughput objective \eqref{CoMPa}, most of CoMP techniques
(see e.g. \cite{CACC08}) are unable to address the QoS constraints \eqref{CoMPb}. Without setting such QoS constraints,
the UEs of poor channel condition are easily disconnected from the service because
it is well known that almost all of transmit power will be
allocated to a very few UEs of the best channel conditions in maximizing the sum throughput, causing almost zero throughput at other UEs. The weighted sum throughput maximization is only an ad hoc way to
balance the UEs' throughput. Alternatively, the throughput satisfaction can be effectively handled via maximizing the
users' worst throughput but the latter involves optimization of a nonsmooth objective function, for which these techniques are powerless. Both QoS constrained sum throughput maximization problem \eqref{CoMPobj} and UEs' worst throughput maximization
problem could be addressed very recently in \cite{TTN16}.
Although the optimization problem \eqref{Rmtgor} is different from \eqref{CoMPobj} and \eqref{DPCobj}, all functions appearing in the former
have a similar structure to that appearing in the latter. Therefore, a systematic approach to solve the latter is expected
to be applicable to the former. In this section, we adopt the approach of \cite{TTN16} to address \eqref{Rmtgor}. Unlike
\cite{Cetal15}, which aims at expressing the nonsmooth function ${\cal R}_{i,p(j)}(\mathbf{V})$ in \eqref{noma1} and then
the nonsmooth objective function in \eqref{Rmtgorb} as d.c. (\underline{d}ifference of two \underline{c}onvex functions)
by using the universality of d.c. functions \cite{Tuybook} and leads to d.c. iterations \cite{KTN12} of high computational complexity, we will see now that each below iteration invokes only a simple convex quadratic
program of low computational complexity.
Suppose that $\mathbf{V}^{(\kappa)}\triangleq [\mathbf{V}^{(\kappa)}_{i,j}]_{(i,j)\in\mathcal{S}}$ is a feasible
point found at the $(\kappa-1)$-th iteration. Define the following
quadratic functions in $\mathbf{V}$:
\begin{IEEEeqnarray}{rCl}
&&{\cal R}^{j,(\kappa)}_{i,p(j)}(\mathbf{V})
\triangleq a_{i,p(j)}^{j,(\kappa)}+2\Re\Bigl\{\bigl\langle{\cal A}_{i,p(j)}^{j,(\kappa)},\mathbf{V}_{i,p(j)}\bigl\rangle
\Bigl\}\nonumber\\
&&\quad -\,\Bigl\langle{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}-{\mathbf{\mathcal{ M}}}_{i,j}(\mathbf{V}^{(\kappa)})^{-1},{\mathbf{\mathcal{ M}}}_{i,j}(\mathbf{V})\Bigl\rangle,
\nonumber\\
&&{\cal R}^{p(j),(\kappa)}_{i,p(j)}(\mathbf{V}) \triangleq
a_{i,p(j)}^{p(j),(\kappa)}
+2\Re\Bigl\{\bigl\langle {\cal A}_{i, p(j)}^{p(j),(\kappa)},\mathbf{V}_{i, p(j)}\bigl\rangle
\Bigl\}\nonumber\\
&&\quad-\,\Bigl\langle {\mathbf{\mathcal{ M}}}_{i,p(j)}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}-{\mathbf{\mathcal{ M}}}_{i,p(j)}(\mathbf{V}^{(\kappa)})^{-1},
{\mathbf{\mathcal{ M}}}_{i,p(j)}(\mathbf{V})\Bigl\rangle,\label{eq:qdt} \nonumber \\
&&{\cal R}^{(\kappa)}_{i,j}(\mathbf{V}) \triangleq a_{i,j}^{(\kappa)}
+2\Re\Bigl\{\bigl\langle{\cal A}_{i,j}^{(\kappa)},\mathbf{V}_{i,j}\bigl\rangle\Bigl\}\nonumber \\
&&\quad-\,\Bigl\langle {\mathbf{\mathcal{ M}}}_{i,j}^{p}(\mathbf{V}^{(\kappa)})^{-1}-{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1},
{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V})\Bigl\rangle
\end{IEEEeqnarray}
where $a_{i,p(j)}^{j,(\kappa)}$, $a_{i,p(j)}^{p(j),(\kappa)}$, $a_{i,j}^{(\kappa)}$ are given as
\begin{IEEEeqnarray}{rCl}
a_{i,p(j)}^{j,(\kappa)} &\triangleq& {\cal R}^{j}_{i,p(j)}(\mathbf{V}^{(\kappa)}) - \Bigl\langle (\mathbf{V}_{i,p(j)}^{(\kappa)})^H \mathbf{H}_{i,i,j}^H\nonumber\\
&&\quad \times{\mathbf{\mathcal{ M}}}_{i,j}(\mathbf{V}^{(\kappa)})^{-1}\mathbf{H}_{i,i,j}\mathbf{V}_{i,p(j)}^{(\kappa)}\Bigl\rangle\, < 0,\nonumber\\
a_{i,p(j)}^{p(j),(\kappa)} &\triangleq& {\cal R}^{p(j)}_{i,p(j)}(\mathbf{V}^{(\kappa)}) - \Bigl\langle(\mathbf{V}_{i,p(j)}^{(\kappa)})^H\mathbf{H}_{i,i,p(j)}^H \nonumber\\
&&\quad \times {\mathbf{\mathcal{ M}}}_{i,p(j)}(\mathbf{V}^{(\kappa)})^{-1}\mathbf{H}_{i,i,p(j)}\mathbf{V}_{i,p(j)}^{(\kappa)}\Bigl\rangle\, < 0,\nonumber\quad\\
a_{i,j}^{(\kappa)} &\triangleq& {\cal R}_{i,j}(\mathbf{V}^{(\kappa)}) - \Bigl\langle (\mathbf{V}_{i,j}^{(\kappa)})^H\mathbf{H}_{i,i,j}^H \nonumber \\
&&\quad \times{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}
\mathbf{H}_{i,i,j}\mathbf{V}_{i,j}^{(\kappa)}\Bigl\rangle\, < 0,
\end{IEEEeqnarray}
and ${\cal A}_{i,p(j)}^{j,(\kappa)}$, ${\cal A}_{i,p(j)}^{p(j),(\kappa)}$, ${\cal A}_{i,j}^{(\kappa)}$ are given as
\begin{IEEEeqnarray}{rCl}
{\cal A}_{i,p(j)}^{j,(\kappa)}&\triangleq& \mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}(\mathbf{V}^{(\kappa)})^{-1}\mathbf{H}_{i,i,j}\mathbf{V}_{i,p(j)}^{(\kappa)},\nonumber\\
{\cal A}_{i,p(j)}^{p(j),(\kappa)}&\triangleq&\mathbf{H}_{i,i,p(j)}^H{\mathbf{\mathcal{ M}}}_{i,p(j)}(\mathbf{V}^{(\kappa)})^{-1}
\mathbf{H}_{i,i,p(j)}\mathbf{V}_{i,p(j)}^{(\kappa)},\nonumber \qquad\\
{\cal A}_{i,j}^{(\kappa)}&\triangleq&\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}
\mathbf{H}_{i,i,j}\mathbf{V}_{i,j}^{(\kappa)}.
\end{IEEEeqnarray}
Note that all functions in (\ref{eq:qdt}) are concave due
to \eqref{eq:Mij}, \eqref{eq:Mijpj}, and \eqref{eq:Mipjpj}:
\begin{IEEEeqnarray}{rCl}
{\mathbf{\mathcal{ M}}}_{i,p(j)}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}-{\mathbf{\mathcal{ M}}}_{i,p(j)}(\mathbf{V}^{(\kappa)})^{-1}&\succeq& \mathbf{0},\nonumber \\
{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}-{\mathbf{\mathcal{ M}}}_{i,j}(\mathbf{V}^{(\kappa)})^{-1}
&\succeq& \mathbf{0},\nonumber \\
{\mathbf{\mathcal{ M}}}_{i,j}^{p}(\mathbf{V}^{(\kappa)})^{-1}-{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1} &\succeq& \mathbf{0}.\nonumber
\end{IEEEeqnarray}
The following result shows that the complicated function defined by \eqref{noma1} and \eqref{noma2}
is lower bounded by concave quadratic functions.
\par\medskip\noindent
\begin{myth}\label{basicth} For ${\cal R}^{(\kappa)}_{i,p(j)}(\mathbf{V}) \triangleq \min\Bigl\{{\cal R}^{j,(\kappa)}_{i,p(j)}(\mathbf{V}),$ $
{\cal R}^{p(j),(\kappa)}_{i,p(j)}(\mathbf{V})\Bigl\}$ it is true that
\begin{IEEEeqnarray}{rCl}\label{main}
{\cal R}_{i,p(j)}(\mathbf{V}^{(\kappa)})&=&{\cal R}^{(\kappa)}_{i,p(j)}(\mathbf{V}^{(\kappa)})\quad\mbox{and}\quad \nonumber \\
{\cal R}_{i,p(j)}(\mathbf{V})&\geq& {\cal R}^{(\kappa)}_{i,p(j)}(\mathbf{V}), \ \forall\, \mathbf{V},
\end{IEEEeqnarray}
and
\begin{IEEEeqnarray}{rCl}\label{mainbound}
{\cal R}_{i,j}(\mathbf{V}^{(\kappa)}) &=& {\cal R}^{(\kappa)}_{i,j}(\mathbf{V}^{(\kappa)})\quad\mbox{and}\quad \nonumber \\
{\cal R}_{i,j}(\mathbf{V}) &\geq& {\cal R}^{(\kappa)}_{i,j}(\mathbf{V}), \ \forall\, \mathbf{V}.
\end{IEEEeqnarray}
\end{myth}
\begin{IEEEproof}
See \cite[Appendix B]{TTN16}.
\end{IEEEproof}\par\medskip\noindent
Based on these results, at the $\kappa$-th iteration, the following convex program, which is
an inner approximation for the nonconvex optimization problem \eqref{Rmtgor}, is solved to generate the next feasible point
$\mathbf{V}^{(\kappa+1)}$:
\begin{IEEEeqnarray}{rCl}\label{RmtgFWA}
&& \displaystyle\underset{\mathbf{V}}{\max} \quad \mathcal{P}^{(\kappa)}(\mathbf{V}) \triangleq\sum_{i=1}^N\sum_{j=1}^K \Bigl({\cal R}_{i,j}^{(\kappa)}(\mathbf{V})+{\cal R}_{i,p(j)}^{(\kappa)}(\mathbf{V}) \Bigr)\IEEEyessubnumber\label{RmtgFWA:a}\qquad \\
&& {\mathrm{subject\ to}} \ \
{\cal R}_{i,j}^{(\kappa)}(\mathbf{V})\geq r_{i,j},\ \forall i\in\mathcal{I},\ \forall j\in\mathcal{K}_1,\IEEEyessubnumber\label{RmtgFWA:b}\\
&&\qquad\qquad\quad {\cal R}^{(\kappa)}_{i,p(j)}(\mathbf{V})\geq r_{i,p(j)}, \ \forall i\in\mathcal{I},\ \forall j\in\mathcal{K}_2,\IEEEyessubnumber\label{RmtgFWA:c}\\
&&\qquad\qquad\quad \eqref{Rmtgord}.\IEEEyessubnumber\label{RmtgFWA:d}
\end{IEEEeqnarray}
A pseudo-code of this quadratic programing (QP)-based path-following
procedure is given by Algorithm~\ref{alg_SCALE_FW}.
Note that $\mathbf{V}^{(\kappa)}$ is also feasible for \eqref{RmtgFWA} with $\mathcal{P}(\mathbf{V}^{(\kappa)})=
\mathcal{P}^{(\kappa)}(\mathbf{V}^{(\kappa)})$ by the equalities in \eqref{main} and \eqref{mainbound}.
It is then true
that $\mathcal{P}^{(\kappa)}(\mathbf{V}^{(\kappa+1)})>\mathcal{P}^{(\kappa)}(\mathbf{V}^{(\kappa)})=\mathcal{P}(\mathbf{V}^{(\kappa)})$ whenever
$\mathbf{V}^{(\kappa+1)}\neq \mathbf{V}^{(\kappa)}$. Together with $\mathcal{P}(\mathbf{V}^{(\kappa+1)})\geq
\mathcal{P}^{(\kappa)}(\mathbf{V}^{(\kappa)})$ and according to the inequalities in \eqref{main} and \eqref{mainbound}, we have
$\mathcal{P}(\mathbf{V}^{(\kappa+1)})>\mathcal{P}(\mathbf{V}^{(\kappa)})$, i.e., the optimal solution
$\mathbf{V}^{(\kappa+1)}$ of the convex quadratic problem \eqref{RmtgFWA} is a better point for the
nonconvex nonsmooth optimization problem \eqref{Rmtgor} than $\mathbf{V}^{(\kappa)}$.
Therefore, once initialized from an feasible point $\mathbf{V}^{(0)})$, the sequence $\{\mathbf{V}^{(\kappa)}\}$ obtained by solving \eqref{RmtgFWA} is of improved feasible points for \eqref{Rmtgor}. By following the same arguments as those in \cite[Proposition 1]{TTN16}, we can prove that
Algorithm~\ref{alg_SCALE_FW} converges to a Karush-Kuh-Tucker (KKT) point of \eqref{Rmtgor}.
\begin{algorithm}[t]
\begin{algorithmic}[1]
\protect\caption{QP-based path-following algorithm for the STM \eqref{Rmtgor} in MIMO-NOMA}
\label{alg_SCALE_FW}
\global\long\def\textbf{Initialization:}{\textbf{Initialization:}}
\REQUIRE Set $\kappa:=0$ and solve \eqref{kiter1} to generate an initial feasible point $\mathbf{V}^{(0)}$ for constraints
\eqref{Rmtgorb}-\eqref{Rmtgord}.
\REPEAT
\STATE Solve the convex quadratic program \eqref{RmtgFWA} to obtain the optimal solution: $\mathbf{V}^{\star}$.
\STATE Update\ \ $\mathbf{V}^{(\kappa+1)}:= \mathbf{V}^{\star}.$
\STATE Set $\kappa:=\kappa+1.$
\UNTIL Convergence\\
\end{algorithmic} \end{algorithm}
\textit{Generation of the initial points}: A feasible point for constraints \eqref{Rmtgorb}-\eqref{Rmtgord}
for initializing Algorithm~\ref{alg_SCALE_FW} is found via the following problem of QoS feasibility
\begin{IEEEeqnarray}{rCl}\label{qos}
&&\displaystyle\max_{\mathbf{V}} \underset{(i,j)\in\mathcal{S}}{\min}\min \Biggl\{\frac{{\cal R}_{i,j}(\mathbf{V})}{r_{i,j}},
\frac{{\cal R}_{i,p(j)}(\mathbf{V})}{r_{i,p(j)}}\Biggr\}\IEEEyessubnumber\label{qosa}\\
&&{\mathrm{subject\ to}}\quad \eqref{Rmtgord}.\IEEEyessubnumber\label{qosb}
\end{IEEEeqnarray}
Initialized from a feasible point $\mathbf{V}^{(0)}$ for the convex constraint
\eqref{Rmtgord}, the following iterations are invoked
\begin{IEEEeqnarray}{rCl}\label{kiter1}
&&\max_{\mathbf{V}} \underset{(i,j)\in\mathcal{S}}{\min}\min \Biggl\{\frac{{\cal R}^{(\kappa)}_{i,j}(\mathbf{V})}{r_{i,j}},
\frac{{\cal R}^{(\kappa)}_{i,p(j)}(\mathbf{V})}{r_{i,p(j)}}\Biggr\} \IEEEyessubnumber\\
&& {\mathrm{subject\ to}}\quad \eqref{Rmtgord}\IEEEyessubnumber
\end{IEEEeqnarray}
till reaching a value more than or equal to $1$ in satisfying \eqref{Rmtgorb}-\eqref{Rmtgord}.
\textit{Complexity Analysis}:
Problem \eqref{RmtgFWA} is convex quadratic with
$m_{\text{QP}}=N(1+3K)$ quadratic constraints and $n=KN(2N_tL+1)$ real decision variables. Its
computational complexity is $\mathcal{O}(n^2m_{\text{QP}}^{2.5}+m_{\text{QP}}^{3.5})$.
\section{Semi-Definite Programming Based Iterations}\label{SDPBI}
To further improve the convergence speed of solving \eqref{Rmtgor}, we need explore more partial convex structures of
functions \eqref{noma1} and \eqref{noma2}. In this section, we propose a novel
SDP-based path-following algorithm for \eqref{Rmtgor}. To this end, we will use the following matrix inequalities:
\begin{eqnarray}
\mathbf{V}^H\mathbf{X}^{-1}\mathbf{V} &\succeq& \bar{\boldsymbol{V}}^H\bar{\boldsymbol{X}}^{-1}\mathbf{V}+
\mathbf{V}^H\bar{\boldsymbol{X}}^{-1}\bar{\boldsymbol{V}} \nonumber\\
&& -\; \bar{\boldsymbol{V}}^H\bar{\boldsymbol{X}}^{-1}\mathbf{X}\bar{\boldsymbol{X}}^{-1}\bar{\boldsymbol{V}}, \nonumber\\
&&\forall\ \mathbf{V}, \bar{\boldsymbol{V}}, \mathbf{X}\succ \mathbf{0} , \bar{\boldsymbol{X}}\succ \mathbf{0} , \label{in1}
\end{eqnarray}
and
\begin{eqnarray}
\ln|\mathbf{X}|&\geq& \ln|\bar{\boldsymbol{X}}|-\bigl\langle \bar{\boldsymbol{X}}, \mathbf{X}^{-1}-\bar{\boldsymbol{X}}^{-1}\bigl\rangle, \nonumber\\
&&\forall\ \mathbf{X}\succ \mathbf{0} , \bar{\boldsymbol{X}}\succ \mathbf{0},\label{in2}
\end{eqnarray}
whose proofs are given by Appendix A and Appendix B.
Let us treat the rate function ${\cal R}_{i,p(j)}^j(\mathbf{V})$ from (\ref{eq:Rijpj}) first.
Applying (\ref{in1}) yields
\begin{IEEEeqnarray}{rCl}
\mathbf{V}_{i,p(j)}^H\mathbf{H}_{i,i,j}^H
{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V})^{-1}\mathbf{H}_{i,i,j}\mathbf{V}_{i, p(j)} \succeq \widetilde{{\cal Q}}^{j,(\kappa)}_{i,p(j)}(\mathbf{V}) \label{lmi1} \qquad
\end{IEEEeqnarray}
for
\begin{IEEEeqnarray}{rCl}
\widetilde{\mathbf{{\cal Q}}}^{j,(\kappa)}_{i,p(j)}(\mathbf{V})&& \triangleq(\mathbf{V}^{(\kappa)}_{i,p(j)})^H\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}
\mathbf{H}_{i,i,j}\mathbf{V}_{i,p(j)} \nonumber\\
&&+\,(\mathbf{V}_{i,p(j)})^H\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}
\mathbf{H}_{i,i,j}\mathbf{V}^{(\kappa)}_{i,p(j)}\nonumber\\
&&-\,(\mathbf{V}^{(\kappa)}_{i,p(j)})^H\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}
{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}) \nonumber \\
&&\times {\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}
\mathbf{H}_{i,i,j}\mathbf{V}^{(\kappa)}_{i,p(j)},\label{map1}
\end{IEEEeqnarray}
which also satisfies
\begin{IEEEeqnarray}{rCl}\label{match1}
&&\widetilde{\mathbf{{\cal Q}}}^{j,(\kappa)}_{i,p(j)}(\mathbf{V}^{(\kappa)})= \nonumber\\
&&\qquad \; (\mathbf{V}_{i,p(j)}^{(\kappa)})^H\mathbf{H}_{i,i,j}^H
{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}\mathbf{H}_{i,i,j}\mathbf{V}_{i, p(j)}^{(\kappa)}.\qquad
\end{IEEEeqnarray}
For
${\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V})$ defined from (\ref{eq:Mijpj}), applying (\ref{in1}) again yields
\begin{eqnarray}
{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}) & \succeq & {\cal L}_{i,j}^{p(j),(\kappa)}(\mathbf{V})+\sigma^2\mathbf{I}_{N_r}\label{lmi1a}
\end{eqnarray}
over the trust region
\begin{equation}\label{tr1}
{\cal L}_{i,j}^{p(j),(\kappa)}(\mathbf{V})\succeq \mathbf{0},
\end{equation}
for the linear mapping
\begin{eqnarray}
{\cal L}_{i,j}^{p(j),(\kappa)}(\mathbf{V})\triangleq \displaystyle\sum_{(s,l)\in{\cal S}\setminus\{(i,p(j))\}}\mathbf{H}_{s,i,j}
\Bigl[\mathbf{V}_{s,l}(\mathbf{V}_{s,l}^{(\kappa)})^H \nonumber\\
+\; \mathbf{V}_{s,l}^{(\kappa)}\mathbf{V}_{s,l}^H
-\mathbf{V}_{s,l}^{(\kappa)}(\mathbf{V}_{s,l}^{(\kappa)})^H\Bigl] \mathbf{H}_{s,i,j}^H.\label{linear1}
\end{eqnarray}
It follows from (\ref{lmi1}) that
\begin{equation}\label{lmi1.1}
\widetilde{{\cal Q}}^{j,(\kappa)}_{i,p(j)}(\mathbf{V})\preceq {\cal Q}^{j,(\kappa)}_{i,p(j)}(\mathbf{V})
\end{equation}
for
\begin{IEEEeqnarray}{rCl}
&&{\cal Q}^{j,(\kappa)}_{i,p(j)}(\mathbf{V})\triangleq \mathbf{V}_{i,p(j)}^H\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}
\mathbf{H}_{i,i,j}\mathbf{V}^{(\kappa)}_{i,p(j)}\nonumber\\
&&\qquad+\, (\mathbf{V}^{(\kappa)}_{i,p(j)})^H\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}
\mathbf{H}_{i,i,j}\mathbf{V}_{i,p(j)} \nonumber\\
&&\qquad-\,(\mathbf{V}^{(\kappa)}_{i,p(j)})^H\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}
\Bigl[{\cal L}_{i,j}^{p(j),(\kappa)}(\mathbf{V})\nonumber \qquad \\
&&\qquad +\, \sigma^2\mathbf{I}_{N_r}\Bigr] {\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}
\mathbf{H}_{i,i,j}\mathbf{V}^{(\kappa)}_{i,p(j)}.\label{map1:1}
\end{IEEEeqnarray}
Therefore,
\begin{IEEEeqnarray}{rCl}
&& {\cal R}_{i,p(j)}^j(\mathbf{V}) \nonumber\\
&&= \ln\Bigl|\mathbf{I}_L + (\mathbf{V}_{i,p(j)})^H\mathbf{H}_{i,i,j}^H
{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V})^{-1}\mathbf{H}_{i,i,j}\mathbf{V}_{i, p(j)}\Bigl|\nonumber \quad\\
&&\geq\ln\Bigl|\mathbf{I}_L +\widetilde{{\cal Q}}^{j,(\kappa)}_{i,p(j)}(\mathbf{V}) \Bigl|\label{lmi2}\\
&&\geq{\cal R}_{i,p(j)}^j(\mathbf{V}^{(\kappa)})+L \nonumber\\
&&\quad -\; \Bigl\langle \mathbf{I}_L +\widetilde{{\cal Q}}^{j,(\kappa)}_{i,p(j)}(\mathbf{V}^{(\kappa)}),\Bigl(\mathbf{I}_L+ \widetilde{{\cal Q}}^{j,(\kappa)}_{i,p(j)}(\mathbf{V})\Bigl)^{-1} \Bigr\rangle\label{l1.0} \\
&&\geq\widetilde{{\cal R}}_{i,p(j)}^{j, (\kappa)}(\mathbf{V})\label{l1}
\end{IEEEeqnarray}
for
\begin{IEEEeqnarray}{rCl}
&&\widetilde{{\cal R}}_{i,p(j)}^{j, (\kappa)}(\mathbf{V})
\triangleq{\cal R}_{i,p(j)}^j(\mathbf{V}^{(\kappa)})+L \nonumber\\
&&\qquad -\; \Bigl\langle \mathbf{I}_L +(\mathbf{V}_{i,p(j)}^{(\kappa)})^H\mathbf{H}_{i,i,j}^H
{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}\nonumber \qquad\\
&&\qquad\qquad \times \mathbf{H}_{i,i,j}\mathbf{V}_{i, p(j)}^{(\kappa)}
,\Bigl(\mathbf{I}_L+ {\cal Q}^{j,(\kappa)}_{i,p(j)}(\mathbf{V})\Bigl)^{-1} \Bigr\rangle,\label{l1.1}
\end{IEEEeqnarray}
which is a concave function. Inequality (\ref{lmi2}) follows from (\ref{lmi1}) and the condition
\[
\ln|\mathbf{I}_L+\mathbf{X}|\geq \ln|\mathbf{I}_L+\mathbf{Y}|,\ \forall\ \mathbf{X}\succeq\mathbf{Y}\succeq \mathbf{0}.
\]
Inequality (\ref{l1.0}) follows by applying (\ref{in2}) and using equality
\[
\ln\bigl|\mathbf{I}_L +\widetilde{{\cal Q}}^{j,(\kappa)}_{i,p(j)}(\mathbf{V}^{(\kappa)}) \bigl|=
{\cal R}_{i,p(j)}^j(\mathbf{V}^{(\kappa)}).
\]
Inequality (\ref{l1}) follows from inequality (\ref{lmi1}), equality (\ref{match1}), and the condition
\[
\Bigl\langle \mathbf{M},\mathbf{X}^{-1}\Bigl\rangle \geq \Bigl\langle \mathbf{M},\mathbf{Y}^{-1}\Bigl\rangle, \ \forall\ \mathbf{M}\succeq \mathbf{0},
\mathbf{Y}\succeq \mathbf{X}\succ \mathbf{0}.
\]
Analogously, the rate function ${\cal R}_{i,p(j)}^{p(j)}(\mathbf{V})$ and throughput function ${\cal R}_{i,j}(\mathbf{V})$
defined from (\ref{eq:Rijj}) and (\ref{noma2}) are lower bounded by
\begin{IEEEeqnarray}{rCl}\label{eq:Ripjpj}
{\cal R}_{i,p(j)}^{p(j)}(\mathbf{V})&\geq& \widetilde{{\cal R}}_{i,p(j)}^{p(j), (\kappa)}(\mathbf{V})\label{l2}
\end{IEEEeqnarray}
over the trust region
\begin{equation}\label{tr2}
{\cal L}_{i,p(j)}^{p(j),(\kappa)}(\mathbf{V})\succeq \mathbf{0},
\end{equation}
and
\begin{IEEEeqnarray}{rCl}\label{eq:Rij}
{\cal R}_{i,j}(\mathbf{V})&\geq&\widetilde{{\cal R}}_{i,j}^{(\kappa)}(\mathbf{V})
\end{IEEEeqnarray}
over the trust region
\begin{equation}\label{tr3}
{\cal L}_{i,j}^{p,(\kappa)}(\mathbf{V})\succeq \mathbf{0},
\end{equation}
for linear mappings
\begin{IEEEeqnarray}{rCl}\label{linear2}
&&{\cal L}_{i,p(j)}^{p(j),(\kappa)}(\mathbf{V}) \triangleq \displaystyle\sum_{(s,l)\in{\cal S}\setminus\{(i,p(j)) \}}\mathbf{H}_{s,i,p(j)}\Bigl[\mathbf{V}_{s,l}(\mathbf{V}_{s,l}^{(\kappa)})^H \nonumber\\
&&\qquad\qquad\quad +\; \mathbf{V}_{s,l}^{(\kappa)}\mathbf{V}_{s,l}^H-
\mathbf{V}_{s,l}^{(\kappa)}(\mathbf{V}_{s,l}^{(\kappa)})^H\Bigl]\mathbf{H}_{s,i,p(j)}^H
\end{IEEEeqnarray}
and
\begin{IEEEeqnarray}{rCl}\label{linear3}
&&{\cal L}_{i,j}^{p,(\kappa)}(\mathbf{V}) \triangleq \displaystyle\ds\sum_{(s,l)\in{\cal S}\setminus\{(i,p(j)),(i,j)\}}\mathbf{H}_{s,i,j}\Bigl[\mathbf{V}_{s,l}(\mathbf{V}_{s,l}^{(\kappa)})^H \nonumber\\
&&\qquad\qquad\quad +\; \mathbf{V}_{s,l}^{(\kappa)}\mathbf{V}_{s,l}^H-\mathbf{V}_{s,l}^{(\kappa)}(\mathbf{V}_{s,l}^{(\kappa)})^H\Bigl] \mathbf{H}_{s,i,j}^H,
\end{IEEEeqnarray}
and for concave functions
\begin{IEEEeqnarray}{rCl}
&&\widetilde{{\cal R}}_{i,p(j)}^{p(j), (\kappa)}(\mathbf{V})
\triangleq {\cal R}_{i,p(j)}^{p(j)}(\mathbf{V}^{(\kappa)})
+ L \nonumber\\
&&\qquad -\Bigl\langle \mathbf{I}_L +(\mathbf{V}^{(\kappa)}_{i,p(j)})^H\mathbf{H}_{i,i,p(j)}^H
{\mathbf{\mathcal{ M}}}_{i,p(j)}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}\nonumber\\
&&\qquad \times \mathbf{H}_{i,i,p(j)}\mathbf{V}_{i, p(j)}^{(\kappa)}
,\Bigl(\mathbf{I}_L +{\cal Q}_{i,p(j)}^{p(j), (\kappa)}(\mathbf{V})\Bigl)^{-1}
\Bigr\rangle\label{l2:2}
\end{IEEEeqnarray}
and
\begin{IEEEeqnarray}{rCl}
&&\widetilde{{\cal R}}_{i,j}^{(\kappa)}(\mathbf{V}) \triangleq
{\cal R}_{i,j}^{(\kappa)}(\mathbf{V}^{(\kappa)})
+ L \nonumber\\
&&\qquad - \; \Bigl\langle \mathbf{I}_L +(\mathbf{V}^{(\kappa)}_{i,j})^H\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{p}(\mathbf{V}^{(\kappa)})^{-1}
\mathbf{H}_{i,i,j}\mathbf{V}_{i,j}^{(\kappa)}, \nonumber\\
&& \qquad\qquad \Bigl(\mathbf{I}_L +{\cal Q}^{(\kappa)}_{i,j}(\mathbf{V})\Bigl)^{-1}\Bigr\rangle\label{l3},
\end{IEEEeqnarray}
with
\begin{IEEEeqnarray}{rCl}
&&{\cal Q}_{i,p(j)}^{p(j), (\kappa)}(\mathbf{V})\triangleq \nonumber\\
&&\quad (\mathbf{V}^{(\kappa)}_{i,p(j)})^H\mathbf{H}_{i,i,p(j)}^H
{\mathbf{\mathcal{ M}}}_{i,p(j)}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}\mathbf{H}_{i,i,p(j)}\mathbf{V}_{i, p(j)} \nonumber\\
&&\quad +\, (\mathbf{V}_{i,p(j)})^H\mathbf{H}_{i,i,p(j)}^H
{\mathbf{\mathcal{ M}}}_{i,p(j)}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}\mathbf{H}_{i,i,p(j)}\mathbf{V}^{(\kappa)}_{i, p(j)} \nonumber\\
&&\quad -\, (\mathbf{V}^{(\kappa)}_{i,p(j)})^H\mathbf{H}_{i,i,p(j)}^H
{\mathbf{\mathcal{ M}}}_{i,p(j)}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}
\Bigl[{\cal L}_{i,p(j)}^{p(j),(\kappa)}(\mathbf{V}) \nonumber \quad \\
&&\quad +\; \sigma^2\mathbf{I}_{N_r}\Bigr]
{\mathbf{\mathcal{ M}}}_{i,p(j)}^{p(j)}(\mathbf{V}^{(\kappa)})^{-1}
\mathbf{H}_{i,i,p(j)}\mathbf{V}^{(\kappa)}_{i, p(j)} \label{map2}\quad\
\end{IEEEeqnarray}
and
\begin{IEEEeqnarray}{rCl}
{\cal Q}^{(\kappa)}_{i,j}(\mathbf{V})&\triangleq& (\mathbf{V}^{(\kappa)}_{i,j})^H\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{p}(\mathbf{V}^{(\kappa)})^{-1}
\mathbf{H}_{i,i,j}\mathbf{V}_{i,j} \nonumber\\
&+&\;(\mathbf{V}_{i,j})^H\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{p}(\mathbf{V}^{(\kappa)})^{-1}
\mathbf{H}_{i,i,j}\mathbf{V}^{(\kappa)}_{i,j}\nonumber\\
&-&\,(\mathbf{V}^{(\kappa)}_{i,j})^H\mathbf{H}_{i,i,j}^H{\mathbf{\mathcal{ M}}}_{i,j}^{p}(\mathbf{V}^{(\kappa)})^{-1}
\Bigl[{\cal L}_{i,j}^{p,(\kappa)}(\mathbf{V}) \nonumber\\
&+&\; \sigma^2\mathbf{I}_{N_r}\Bigr]{\mathbf{\mathcal{ M}}}_{i,j}^{p}(\mathbf{V}^{(\kappa)})^{-1}
\mathbf{H}_{i,i,j}\mathbf{V}^{(\kappa)}_{i,j}.\label{map3}
\end{IEEEeqnarray}
We also define
\begin{equation}\label{l4}
\widetilde{{\cal R}}^{(\kappa)}_{i,p(j)}(\mathbf{V})\triangleq \min\Bigl\{\widetilde{{\cal R}}^{j,(\kappa)}_{i,p(j)}(\mathbf{V}),
\widetilde{{\cal R}}^{p(j),(\kappa)}_{i,p(j)}(\mathbf{V})\Bigl\}.
\end{equation}
In summary, at the $\kappa$-th iteration, the following SDP, which is an inner approximation of \eqref{Rmtgor}, is solved
to generate the next feasible point $\mathbf{V}^{(\kappa+1)}$:
\begin{IEEEeqnarray}{rCl}\label{sdp1}
&& \displaystyle\underset{\mathbf{V}}{\max} \ \, \mathcal{\widetilde{P}}^{(\kappa)}(\mathbf{V}) \triangleq\sum_{i=1}^N\sum_{j=1}^K \Bigl(\widetilde{{\cal R}}_{i,j}^{(\kappa)}(\mathbf{V}) + \widetilde{{\cal R}}^{(\kappa)}_{i,p(j)}(\mathbf{V}) \Bigr)\IEEEyessubnumber\label{sdp1:a}\qquad\\
&& {\mathrm{subject\ to}} \
\widetilde{{\cal R}}_{i,j}^{(\kappa)}(\mathbf{V})\geq r_{i,j},\ \forall i\in\mathcal{I},\ \forall j\in\mathcal{K}_1,\IEEEyessubnumber\label{sdp1:b}\\
&& \qquad\qquad\quad \widetilde{{\cal R}}^{(\kappa)}_{i,p(j)}(\mathbf{V}) \geq r_{i,p(j)}, \; \forall i\in\mathcal{I},\; \forall j\in\mathcal{K}_2,\IEEEyessubnumber\label{sdp1:c}\\
&&\qquad\qquad\quad \eqref{Rmtgord}, \eqref{tr1}, \eqref{tr2}, \eqref{tr3}.\IEEEyessubnumber\label{sdp1:d}
\end{IEEEeqnarray}
The proposed Algorithm~\ref{alg_SDP_FW}
generates a sequence $\{\mathbf{V}^{(\kappa)}\}$ of improved points of \eqref{Rmtgor}, which also converges to a KKT point.
\begin{algorithm}[t]
\begin{algorithmic}[1]
\protect\caption{SDP-based path-following algorithm for the STM \eqref{Rmtgor} in MIMO-NOMA}
\label{alg_SDP_FW}
\global\long\def\textbf{Initialization:}{\textbf{Initialization:}}
\REQUIRE Set $\kappa:=0$ and solve \eqref{kiter2} to generate an initial feasible point $\mathbf{V}^{(0)}$ for constraints
\eqref{Rmtgorb}-\eqref{Rmtgord}.
\REPEAT
\STATE Solve the semi-definite program \eqref{sdp1} to obtain the optimal solution: $\mathbf{V}^{\star}$.
\STATE Update\ \ $\mathbf{V}^{(\kappa+1)}:= \mathbf{V}^{\star}.$
\STATE Set $\kappa:=\kappa+1.$
\UNTIL Convergence\\
\end{algorithmic} \end{algorithm}
A feasible point $\mathbf{V}^{(0)}$ for the constraints \eqref{Rmtgorb}-\eqref{Rmtgord} to initialize Algorithm \ref{alg_SDP_FW}
can be found by invoking the iterations
\begin{IEEEeqnarray}{rCl}\label{kiter2}
&&\max_{\mathbf{V}} \underset{(i,j)\in\mathcal{S}}{\min}\min \Biggl\{\frac{\widetilde{{\cal R}}^{(\kappa)}_{i,j}(\mathbf{V})}{r_{i,j}},
\frac{\widetilde{{\cal R}}^{(\kappa)}_{i,p(j)}(\mathbf{V})}{r_{i,p(j)}}\Biggl\}\ \IEEEyessubnumber\\
&&{\mathrm{subject\ to}}\quad \eqref{Rmtgord},
\eqref{tr1}, \eqref{tr2}, \eqref{tr3} \IEEEyessubnumber
\end{IEEEeqnarray}
to reach a value more than or equal to $1$ in satisfying \eqref{Rmtgorb}-\eqref{Rmtgord}.
\textit{Complexity analysis:} The SDP \eqref{sdp1} involves $N(1+3K)$ quadratic constraints, $3NK$ semi-definite
constraints with $N_r$ rows and $n=KN(2N_tL+1)$ real decision variables. For $m_{\text{SDP}}\triangleq N(1+3K)+3NKN_r$,
its computational complexity is $\mathcal{O}(n^2m_{\text{SDP}}^{2.5}+m_{\text{SDP}}^{3.5})$, which is seen higher than that of
the convex quadratic problem \eqref{RmtgFWA}.
\section{Tailored algorithm for MISO-NOMA}\label{sec:TL}
In this case, all channels are row vectors ($\mathbf{H}_{s,i,j}\in\mathbb{C}^{1\times N_t}$)
and $s_{i,j}\in\mathbb{C}$ ($L=1$).
As observed first time in \cite{WES06}, for
\begin{equation}\label{varch}
\bar{\mathbf{V}}_{i,j}=e^{-\jmath {\sf arg}(\mathbf{H}_{i,i,j}\mathbf{V}_{i,j})}\mathbf{V}_{i,j}
\end{equation}
one has
$|\mathbf{H}_{i,i,j}\mathbf{V}_{i,j}|=\mathbf{H}_{i,i,j}\bar{\mathbf{V}}_{i,j}=\Re\{ \mathbf{H}_{i,i,j}\bar{\mathbf{V}}_{i,j}\}\geq 0$ and
$|\mathbf{H}_{i',i,j'}\mathbf{V}_{i,j}|=|\mathbf{H}_{i',i,j'}\bar{\mathbf{V}}_{i,j}|$ for $(i',j')\neq (i,j)$. Therefore,
without loss of generality we can replace
\[
\mathbf{H}_{i,i,j}\mathbf{V}_{i,j}\mathbf{V}_{i,j}^H\mathbf{H}_{i,i,j}^H = \bigl|\mathbf{H}_{i,i,j}\mathbf{V}_{i,j}\bigr|^2, \ j=1,\dots, 2K
\]
by
\[
\bigl(\Re\{\mathbf{H}_{i,i,j}\mathbf{V}_{i,j}\}\bigr)^2, \ j=1,\dots, 2K
\]
with
\begin{equation}\label{positivereal}
\Re\{\mathbf{H}_{i,i,j}\mathbf{V}_{i,j}\}\geq 0, \ j=1,\dots, 2K
\end{equation}
(including $p(j)$ for $j=K+1,...,2K$).
Accordingly, write
\[
{\mathbf{\mathcal{ M}}}_{i,j}(\mathbf{V})= \displaystyle\sum_{(s,l)\in{\cal S}}\bigl|\mathbf{H}_{s,i,j}\mathbf{V}_{s,l}\bigr|^2+ \sigma^2.
\]
Then the message $\mathbf{s}_{i,p(j)}$ intended for cell-edge UE $(i,p(j))$ is decoded by the cell-center UE $(i,j)$
with the achievable rate
\begin{equation}\label{t1}
r_{i,p(j)}^j(\mathbf{V}) = \ln\Biggl(1 + \frac{\bigl|\mathbf{H}_{i,i,j}\mathbf{V}_{i,p(j)}\bigr|^2}{{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V})}\Biggr),
\end{equation}
and is decoded by the cell-edge UE $(i,p(j))$ itself with the achievable rate
\begin{equation}\label{t2}
r_{i,p(j)}^{p(j)}(\mathbf{V}) = \ln\Biggl( 1 +
\frac{\bigl(\Re\{ \mathbf{H}_{i,i,p(j)}\mathbf{V}_{i, p(j)}\}\bigr)^2}{{\mathbf{\mathcal{ M}}}_{i,p(j)}^{p(j)}(\mathbf{V})} \Biggl)
\end{equation}
where
\begin{IEEEeqnarray}{rCl}
{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}) & \triangleq& {\mathbf{\mathcal{ M}}}_{i,j}(\mathbf{V}) - \bigl|\mathbf{H}_{i,i,j}\mathbf{V}_{i,p(j)}\bigr|^2 \nonumber\\
&=&\displaystyle\ds\sum_{(s,l)\in{\cal S}\setminus\{(i,p(j))\}}
\bigl|\mathbf{H}_{s,i,j}\mathbf{V}_{s,l}\bigl|^2 + \sigma^2, \nonumber
\end{IEEEeqnarray}
and
\begin{IEEEeqnarray}{rCl}
{\mathbf{\mathcal{ M}}}_{i,p(j)}^{p(j)}(\mathbf{V})& \triangleq& {\mathbf{\mathcal{ M}}}_{i,p(j)}(\mathbf{V}) - \bigl|\mathbf{H}_{i,i,p(j)}\mathbf{V}_{i,p(j)}\bigl|^2 \nonumber\\
&=&\displaystyle\sum_{(s,l)\in{\cal S}\setminus\{(i,p(j)) \}}
\bigl|\mathbf{H}_{s,i,p(j)}\mathbf{V}_{s,l}\bigl|^2 + \sigma^2.\nonumber
\end{IEEEeqnarray}
Also, the message $\mathbf{s}_{i,j}$ intended for the cell-center UE $(i,j)$ is successively decoded by UE $(i,j)$ itself
with the throughput
\begin{equation}\label{t3}
r_{i,j}(\mathbf{V}) = \ln\Biggr( 1 + \frac{\bigl(\Re\{\mathbf{H}_{i,i,j}\mathbf{V}_{i,j}\}\bigr)^2}{ {\mathbf{\mathcal{ M}}}_{i,j}^{p}(\mathbf{V})} \Biggr)
\end{equation}
where
\begin{IEEEeqnarray}{rCl}
{\mathbf{\mathcal{ M}}}_{i,j}^{p}(\mathbf{V}) & \triangleq & {\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}) - \bigl|\mathbf{H}_{i,i,j}\mathbf{V}_{i,j}\bigl|^2 \nonumber\\
&=&\displaystyle\ds\sum_{(s,l)\in{\cal S}\setminus\{(i,p(j)),(i,j)\}}\bigl|\mathbf{H}_{s,i,j}\mathbf{V}_{s,l}\bigl|^2 + \sigma^2. \nonumber\nonumber
\end{IEEEeqnarray}
For $r_{i,p(j)}(\mathbf{V}) \triangleq \min\bigl\{r^j_{i,p(j)}(\mathbf{V}), r^{p(j)}_{i,p(j)}(\mathbf{V})\bigl\}$, the problem (\ref{Rmtgor}) in this case is
\begin{IEEEeqnarray}{rCl}\label{beam1}
&&\underset{\mathbf{V}}{{\mathrm{maximize}}}\ \mathcal{P}(\mathbf{V})\triangleq\sum_{i=1}^N\sum_{j=1}^K\Bigr(r_{i,j}(\mathbf{V})+ r_{i,p(j)}(\mathbf{V})
\Bigl)\quad \IEEEyessubnumber\label{beam1a} \quad\\
&& {\mathrm{subject\ to}} \
r_{i,j}(\mathbf{V}) \geq r_{i,j}, \ \forall i\in\mathcal{I},\ \forall j\in\mathcal{K}_1, \IEEEyessubnumber\label{beam1b} \\
&&\qquad\qquad\ r^{p(j)}_{i,p(j)}(\mathbf{V}) \geq r_{i,p(j)}, \ \forall i\in\mathcal{I},\ \forall j\in\mathcal{K}_2, \IEEEyessubnumber\label{beam1c}\\
&&\qquad\qquad\ r^{j}_{i,p(j)}(\mathbf{V}) \geq r_{i,p(j)}, \ \forall i\in\mathcal{I},\ \forall j\in\mathcal{K}_2, \IEEEyessubnumber\label{beam1d}\\
&&\qquad\qquad\ \sum_{j\in {\cal J}} \|\mathbf{V}_{i,j}\|^2 \le P^{\max}_{i}, \forall i \in\mathcal{I}. \IEEEyessubnumber\label{beam1e}
\end{IEEEeqnarray}
Due to the above transforms (\ref{t2}) and (\ref{t3}) under condition (\ref{positivereal}), the nonconvex
constraints (\ref{beam1b}) and (\ref{beam1c}) are expressed by the second-order cone (SOC) constraints
\begin{IEEEeqnarray}{rCl}
\Re\bigl\{\mathbf{H}_{i,i,j}\mathbf{V}_{i,j}\bigr\} &\geq& \sqrt{e^{r_{i,j}}-1}\sqrt{{\mathbf{\mathcal{ M}}}_{i,j}^{p}(\mathbf{V})}, \nonumber\\
&& \ \forall i\in\mathcal{I},\ \forall j\in\mathcal{K}_1, \label{beam1be} \\
\Re\bigl\{ \mathbf{H}_{i,i,p(j)}\mathbf{V}_{i, p(j)} \bigl\} &\geq&
\sqrt{e^{r_{i,p(j)}}-1}\sqrt{{\mathbf{\mathcal{ M}}}_{i,p(j)}^{p(j)}(\mathbf{V})}, \qquad \nonumber\\
&& \ \forall i\in\mathcal{I},\ \forall j\in\mathcal{K}_2\label{beam1ce}
\end{IEEEeqnarray}
but the constraint (\ref{beam1d}) remains to be nonconvex.
To approximate functions in (\ref{beam1a}) we use the inequality
\begin{equation}\label{zf8}
\ln(1+z)\geq a(\bar{z})-b(\bar{z})\frac{1}{z},\quad
\forall\ z>0, \bar{z}>0
\end{equation}
with
\begin{equation}\label{zf9}
0<a(\bar{z}) \triangleq \ln(1+\bar{z})+\displaystyle\frac{\bar{z}}{\bar{z}+1},\
0<b(\bar{z}) \triangleq \frac{\bar{z}^2}{\bar{z}+1},
\end{equation}
whose proof is provided by Appendix~C.
Applying (\ref{zf8}) for
$\bar{z}=z^{j, (\kappa)}_{i,p(j)}\triangleq |\mathbf{H}_{i,i,j}\mathbf{V}^{(\kappa)}_{i,p(j)}|^2/{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V}^{(\kappa)})$ and
$z= |\mathbf{H}_{i,i,j}\mathbf{V}_{i,p(j)}|^2/{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V})$ yields
\begin{IEEEeqnarray}{rCl}
&&r^{j}_{i,p(j)}(\mathbf{V})\geq a\bigl(z_{i,p(j)}^{j,(\kappa)}\bigl)-b\bigl(z_{i,p(j)}^{j,(\kappa)}\bigl)\displaystyle\frac{{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V})}{\bigl|\mathbf{H}_{i,i,j}\mathbf{V}_{i,p(j)}\bigr|^2} \nonumber\\
&&\qquad\qquad\ \geq r^{j, (\kappa)}_{i,p(j)}(\mathbf{V}) \nonumber\\
&&\qquad\qquad\ \triangleq a\bigl(z_{i,p(j)}^{j,(\kappa)}\bigl)-
b\bigl(z_{i,p(j)}^{j,(\kappa)}\bigl)
\displaystyle\frac{{\mathbf{\mathcal{ M}}}_{i,j}^{p(j)}(\mathbf{V})}{\varphi_{i,p(j)}^{j,(\kappa)}(\mathbf{V})} \qquad
\end{IEEEeqnarray}
over the trust region
\begin{IEEEeqnarray}{rCl}\label{trb1}
\varphi_{i,p(j)}^{j,(\kappa)}(\mathbf{V}) \triangleq
2\Re\bigl\{\mathbf{H}_{i,i,j}\mathbf{V}^{(\kappa)}_{i,p(j)}
(\mathbf{H}_{i,i,j}\mathbf{V}_{i,p(j)})^*\bigr\} && \nonumber\\
-\ \bigl|\mathbf{H}_{i,i,j}\mathbf{V}^{(\kappa)}_{i,p(j)}\bigl|^2 &&\; > 0.\qquad
\end{IEEEeqnarray}
Analogously,
\begin{IEEEeqnarray}{rCl}
r^{p(j)}_{i,p(j)}(\mathbf{V})&\geq&r^{p(j), (\kappa)}_{i,p(j)}(\mathbf{V}) \nonumber\\
&\triangleq&a\bigl(z_{i,p(j)}^{p(j),(\kappa)}\bigl) -
b\bigl(z_{i,p(j)}^{p(j),(\kappa)}\bigl)\displaystyle\frac{{\mathbf{\mathcal{ M}}}_{i,p(j)}^{p(j)}(\mathbf{V})}
{\varphi_{i,p(j)}^{p(j),(\kappa)}(\mathbf{V})}\qquad
\end{IEEEeqnarray}
with
\begin{IEEEeqnarray}{rCl}
\varphi_{i,p(j)}^{p(j),(\kappa)}(\mathbf{V}) \triangleq&& \Re\bigl\{\mathbf{H}_{i,i,p(j)}\mathbf{V}^{(\kappa)}_{i,p(j)}\bigl\}
\Bigl(2\Re\bigl\{\mathbf{H}_{i,i,p(j)}\mathbf{V}_{i,p(j)}\bigl\} \nonumber \\
&& -\ \Re\bigl\{\mathbf{H}_{i,i,p(j)}\mathbf{V}^{(\kappa)}_{i,p(j)}\bigl\}\Bigl) \nonumber
\end{IEEEeqnarray}
over the trust region
\begin{equation}\label{trb2}
2\Re\bigl\{\mathbf{H}_{i,i,p(j)}\mathbf{V}_{i,p(j)}\bigl\}-\Re\bigl\{\mathbf{H}_{i,i,p(j)}\mathbf{V}^{(\kappa)}_{i,p(j)}\bigl\}\; > 0,
\end{equation}
and
\begin{IEEEeqnarray}{rCl}
r_{i,j}(\mathbf{V})&\geq& r^{(\kappa)}_{i,p(j)}(\mathbf{V}) \nonumber\\
&\triangleq&a(z_{i,j}^{(\kappa)})-
b(z_{i,j}^{(\kappa)})\displaystyle\frac{{\mathbf{\mathcal{ M}}}_{i,j}^{p}(\mathbf{V})}{ \varphi_{i,j}^{(\kappa)}(\mathbf{V})}
\end{IEEEeqnarray}
with
\begin{IEEEeqnarray}{rCl}
\varphi_{i,j}^{(\kappa)}(\mathbf{V}) \triangleq&&\ \Re\bigl\{\mathbf{H}_{i,i,j}\mathbf{V}^{(\kappa)}_{i,j}\bigl\}
\Bigl(2\Re\bigl\{\mathbf{H}_{i,i,j}\mathbf{V}_{i,j}\bigl\} \nonumber\\
&&-\ \Re\bigl\{\mathbf{H}_{i,i,j}\mathbf{V}^{(\kappa)}_{i,j}\bigl\}\Bigl) \nonumber
\end{IEEEeqnarray}
over the trust region
\begin{equation}\label{trb3}
2\Re\bigl\{\mathbf{H}_{i,i,j}\mathbf{V}_{i,j}\bigl\}-\Re\bigl\{\mathbf{H}_{i,i,j}\mathbf{V}^{(\kappa)}_{i,j}\bigl\} > 0.
\end{equation}
In Algorithm \ref{alg_Tailored_FW}, we propose an QP-based path-following algorithm to solve problem (\ref{beam1}). At
the $\kappa$-th iteration it solves the following SOC program to generate the next feasible point $\mathbf{V}^{(\kappa+1)}$:
\begin{IEEEeqnarray}{rCl}\label{beamk}
&& \underset{\mathbf{V}}{\max} \quad \mathcal{P}^{(\kappa)}(\mathbf{V}) \triangleq\sum_{i=1}^N\sum_{j=1}^K \Bigl(r_{i,j}^{(\kappa)}(\mathbf{V})+r_{i,p(j)}^{(\kappa)}(\mathbf{V}) \Bigl) \IEEEyessubnumber \qquad\\
&& {\mathrm{subject\ to}}\ r^{j,(\kappa)}_{i,p(j)}(\mathbf{V})\geq r_{i,p(j)}, \ \forall i\in\mathcal{I},\ \forall j\in\mathcal{K}_2, \IEEEyessubnumber \\
&&\qquad\qquad\ \eqref{beam1e}, (\ref{beam1be}), (\ref{beam1ce}), (\ref{trb1}), (\ref{trb2}), (\ref{trb3}) \IEEEyessubnumber
\end{IEEEeqnarray}
where $r_{i,p(j)}^{(\kappa)}(\mathbf{V}) \triangleq \min\Bigl\{r^{j, (\kappa)}_{i,p(j)}(\mathbf{V}), r^{p(j), (\kappa)}_{i,p(j)}(\mathbf{V})\Bigr\}$.
\begin{algorithm}[t]
\caption{Tailored QP-based path-following algorithm for the
STM \eqref{beam1} in MISO-NOMA}\label{alg_Tailored_FW}
\begin{algorithmic}
\STATE \textbf{Initialization}: Initialize a feasible point $\mathbf{V}^{(0)}$ for constraints in (\ref{beam1}).
\STATE \textbf{$\kappa$-th iteration}: Solve the convex quadratic program (\ref{beamk}) to find the optimal solution $\mathbf{V}^{\star}$. If $\bigr|\bigl({\cal P}(\mathbf{V}^{\star}) - \mathcal{P}(\mathbf{V}^{(\kappa)})\bigr) \big/ \mathcal{P}(\mathbf{V}^{(\kappa)}) \bigl| \le \epsilon$, terminate. Otherwise, set $\kappa :=\kappa+1, \mathbf{V}^{(\kappa)}:=\mathbf{V}^{\star}$ and continue.
\end{algorithmic}
\end{algorithm}
To find a feasible point for constraints in (\ref{beam1}) for initializing Algorithm \ref{alg_Tailored_FW},
initialized by a feasible point $\mathbf{V}^{(0)}$ for
the convex constraints (\ref{beam1e}), (\ref{beam1be}), and (\ref{beam1ce}), the following SOC based iterations
\begin{IEEEeqnarray}{rCl}\label{beami}
&&\max_{\mathbf{V}} \min_{(i,j)\in\mathcal{S}_2}\ \Biggl\{\frac{r^{j,(\kappa)}_{i,p(j)}
(\mathbf{V})}{r_{i,p(j)}}\Biggr\} \quad \IEEEyessubnumber\\
&&{\mathrm{subject\ to}}\quad
\eqref{beam1e}, (\ref{beam1be}), (\ref{beam1ce}), (\ref{trb1})\IEEEyessubnumber
\end{IEEEeqnarray}
are invoked for reaching a value more or equal to $1$ in satisfying constraints in (\ref{beam1}).
\textit{Complexity analysis:} The SOC program \eqref{beamk} involves $m_{\text{SOC}}=N(1+6K)$ quadratic or
SOC constraints and $n=KN(2N_tL+1)$ real decision variables. Its computational complexity is $\mathcal{O}(n^2m_{\text{SOC}}^{2.5}+m_{\text{SOC}}^{3.5})$, which is seen higher than that of
the convex quadratic problem \eqref{RmtgFWA}.
\section{Numerical Results}\label{sec:simulation}
In this section we use numerical examples to evaluate the performance of the proposed algorithms.
A system topology shown in Fig. \ref{fig:SM:RatePerUE} is set up. There are
$N=3$ macro cells and $4$ UEs per cell with two cell-center UEs and two cell-edge UEs, which are
located near to the boundaries with the two adjacent cells. Unless stated otherwise, $N_t = 4$ and $N_r = 2$
are set for MIMO-NOMA, for which $L=N_r$ is set. Thus, the precoder-matrices $\mathbf{V}_{i,j}$ are of dimension $N_t\times N_r$.
The channel matrix between a BS and a UE at a distance $d$ (in kilometres) is generated as $\mathbf{H}=\sqrt{10^{-\sigma_{\mathsf{PL}}/10}}\tilde{\mathbf{H}}$ \cite{ETSI}. Here, $\sigma_{\mathsf{PL}}$ is the path loss (PL) in dB and the entries of $\tilde{\mathbf{H}}$ are independent and identically distributed complex Gaussian variables with zero mean and unit variance. Without loss of generality, the requirement thresholds for
all UEs are set as $r_{i,j} = r_{i,p(j)}\equiv \bar{\mathsf{R}}$ and the same
power budget $P^{\max}_i = P^{\max},\, \forall i\in\mathcal{I}$ is given to all BSs.
For the ease of reference, the other parameters given in Table \ref{parameter} including $\bar{\mathsf{R}}$
are used. The error tolerance in the proposed Algorithms is set to $\epsilon=10^{-3}$. The numerical results are obtained using the parser YALMIP \cite{L04}. The achieved sum throughput results are divided by $\ln(2)$ to arrive at the unit of bps/channel-use.
Each simulation is run $100$ times and the result are averaged to arrive at the final figures.
\begin{figure}[t]
\centering
\includegraphics[width=0.45\textwidth,trim={0cm 0.0cm -0cm -0cm}]{Simulation1.eps}
\caption{MIMO-NOMA multi-cell system.}
\label{fig:SM:RatePerUE}
\end{figure}
\begin{table}[t]
\caption{SIMULATION PARAMETERS}
\label{parameter}
\centering
{\setlength{\tabcolsep}{0.695em}
\setlength{\extrarowheight}{0.6em}
\begin{tabular}{l|l}
\hline
Parameters & Value \\
\hline\hline
Carrier frequency/ Bandwidth & 2 [GHz]/ 20 [MHz] \\
Noise power density & -174 [dBm/Hz] \\
Path loss from the BS to a UE, $\sigma_{\mathsf{PL}}$ & 128.1 + 37.6$\log_{10}(d)$ [dB]\\
Shadowing standard deviation & 8 [dB] \\
Radius of each cell, $r_o$ & 500 [m]\\
Coverage of near UEs, $r_n$ & 150 [m]\\
Distance between BS and nearest UE & $>$ 10 [m]\\
Threshold $\bar{\mathsf{R}}$ & 1 [bps/Hz] \\
\hline
\end{tabular}}
\end{table}
\subsection{Algorithms' Convergence}
\begin{figure}
\begin{center}
\begin{subfigure}[MIMO-NOMA networks.]{
\includegraphics[width=0.49\textwidth]{ConvergenceRate_MIMO.eps}}
\label{fig:a}
\end{subfigure}
\begin{subfigure}[MISO-NOMA networks.]{
\includegraphics[width=0.49\textwidth]{ConvergenceRate_MISO.eps}}
\label{fig:b}
\end{subfigure}
\caption{Convergence pattern for $P^{\max} = 30$ dBm.}\label{fig:Convergencebehavior:Iteration}
\end{center}
\end{figure}
Fig.~\ref{fig:Convergencebehavior:Iteration} shows the typical convergence behavior of the proposed algorithms for a given set of channel realizations that are randomly generated for the two cases.
According to Fig.~\ref{fig:Convergencebehavior:Iteration}(a), both Algorithm~\ref{alg_SCALE_FW} and \ref{alg_SDP_FW}
for MIMO-NOMA reach the almost optimal value of sum throughput in $18$ and $13$ iterations.
As expected, Algorithm~\ref{alg_SDP_FW} converges faster than Algorithm~\ref{alg_SCALE_FW}.
On the other hand, according to Fig.~\ref{fig:Convergencebehavior:Iteration}(b), Algorithm~\ref{alg_Tailored_FW} for NOMA-MISO
converges very fast reaching the optimal value in 6 iterations.
\subsection{Numerical Results for MIMO-NOMA}
\begin{figure}
\begin{center}
\begin{subfigure}[Sum throughput versus $P^{\max}$.]{
\includegraphics[width=0.48\textwidth]{NOMA_MIMO_Pmax.eps}}
\label{fig:SRMMIMO:Pmax}
\end{subfigure}
\begin{subfigure}[Sum throughput versus $\bar{\mathsf{R}}$ with $P^{\max} = 30$ dBm.]{
\includegraphics[width=0.48\textwidth]{NOMA_MIMO_Rbar.eps}}
\label{fig:SRMMIMO:Rbar}
\end{subfigure}
\caption{Sum throughput of MIMO-NOMA, (a) versus the transmit power at the BSs and (b) versus the throughput threshold ($N = 3, K = 2, N_t = 4$, and $N_r = 2$).}\label{fig:SRMMIMO:PmaxRbar}
\end{center}
\end{figure}
\begin{table}[t]
\centering
\caption{Achieved User Throughput (bps/Hz) in Multi-user MIMO Multi-cell Systems for $\bar{\mathsf{R}} = 0$ bps/Hz}
\label{Table:RateperUENoQoS}
{\setlength{\tabcolsep}{0.50em}
\setlength{\extrarowheight}{0.65em}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\multicolumn{1}{|c|}{\multirow{2}{*}{}} & \multicolumn{3}{c|}{Throughput per UE} & \multicolumn{3}{c|}{ST per cell} \\ \cline{2-7}
\multicolumn{1}{|c|}{} & NOMA & CoMP & DPC & \multicolumn{1}{c|}{NOMA} & \multicolumn{1}{c|}{CoMP} & \multicolumn{1}{c|}{DPC} \\ \hline
UE (1,1) & 6.36 & 8.27 & 9.43 & \multirow{4}{*}{18.48} & \multirow{4}{*}{19.11} & \multirow{4}{*}{22.54} \\ \cline{1-4}
UE (1,2) & 7.13 & 9.39 & 7.86 & & & \\ \cline{1-4}
UE (1,3) & 1.39 & 1.07 & 3.27 & & & \\ \cline{1-4}
UE (1,4) & 3.60 & 0.38 & 1.98 & & & \\ \hline
UE (2,1) & 5.96 & 9.70 & 10.76 & \multirow{4}{*}{18.50} & \multirow{4}{*}{23.35} & \multirow{4}{*}{24.01} \\ \cline{1-4}
UE (2,2) & 7.02 & 7.87 & 8.93 & & & \\ \cline{1-4}
UE (2,3) & 3.59 & 2.19 & 2.69 & & & \\ \cline{1-4}
UE (2,4) & 1.93 & 3.59 & 1.63 & & & \\ \hline
UE (3,1) & 5.04 & 7.62 & 10.06 & \multirow{4}{*}{21.27} & \multirow{4}{*}{19.25} & \multirow{4}{*}{22.62} \\ \cline{1-4}
UE (3,2) & 7.30 & 8.44 & 8.24 & & & \\ \cline{1-4}
UE (3,3) & 4.40 & 0.63 & 2.86 & & & \\ \cline{1-4}
UE (3,4) & 4.53 & 2.56 & 1.46 & & & \\ \hline
\multicolumn{4}{|c|}{Total ST} & \underline{\textbf{58.25}} & \underline{\textbf{61.71}} & \underline{\textbf{69.17}} \\ \hline
\end{tabular}}
\end{table}
For the ease of reference, the result achieved by \eqref{CoMPobj} is labeled by ``Conventional MIMO-CoMP'' whereas
that achieved by \eqref{Rmtgor} is labeled by ``Proposed MIMO-NOMA.'' Fig.~\ref{fig:SRMMIMO:PmaxRbar}(a) plots
the sum throughput versus the power budget $P^{\max}$ under setting $\bar{\sf R} = \{0,1\} $ bps/Hz.
For $\bar{\sf R}=0$ bps/Hz, i.e. there is no UEs' QoS requirement imposed, CoMP slightly outperforms
MIMO-NOMA by achieving throughput concentrated at the cell-center UEs of good conditions.
Table~\ref{Table:RateperUENoQoS} details the UEs's throughput distribution for $P^{\max}=30$ dBm.
The high ratio $9.39/0.38=24.7$ between the best UE throughput and the worst UE throughput (BWR) implies that CoMP would
perform wobbly in the QoS maximization problem (\ref{qos}) (with $r_{i,j}\equiv 1$), which also expresses the
system ability to offer the uniform service to UEs. BWR for MIMO-NOMA is $5.1=7.13/1.39$ so it is expected to outperform
CoMP in maximizing (\ref{qos}). The low throughput $1.46$ bps/Hz
at UE $(3,4)$ by DPC is a result of a strong interference from an adjacent cell, which cannot be mitigated by DPC.
For $\bar{\sf R}=1$ bps/Hz, MIMO-NOMA of course offers a higher sum throughput than
CoMP, where the BSs are seen spending a nearly full power budget
$P^{\max}$ in gaining the sum throughput. Increasing $P^{\max}$ also leads to a remarkable gain
in sum throughput by NOMA compared with CoMP. The sum throughput by the former also catches up
that by DPC. The gain of MIMO-NOMA is a result of canceling interference from intra-cluster interference, as shown in \eqref{eq:Mipjpj}. The cell-center UEs in CoMP experience intra-cluster interference that becomes stronger when transmit power increases.
The plot of sum throughput versus QoS requirement threshold $\bar{\mathsf{R}}\in[0.6,\,2.4]$ bps/Hz is shown by Fig.~\ref{fig:SRMMIMO:PmaxRbar}(b) for $P^{\max} = 30$ dBm. The sum throughput are nearly flat for $\bar{\mathsf{R}} \leq 1.2$ bps/Hz and are degraded after that.
The BSs in CoMP must allocate much more power to serve cell-edge UEs when QoS threshold increases. As a result,
the system sum throughput is dropped quickly.
In contrast, the sum throughput of MIMO-NOMA is still slightly sensitive to QoS requirement threshold
because BSs can tune the power allocation in meeting cell-edge UEs' QoS requirements
whenever the cell-center UEs' QoS requirement is easily met.
\begin{figure}
\begin{center}
\begin{subfigure}[Sum throughput versus $K$ with $N_t = 4$.]{
\includegraphics[width=0.48\textwidth]{NOMA_MIMO_K.eps}}
\label{fig:SRMMIMO:K}
\end{subfigure}
\begin{subfigure}[Sum throughput versus $N_t$ with $K = 2$.]{
\includegraphics[width=0.48\textwidth]{NOMA_MIMO_Nt.eps}}
\label{fig:SRMMIMO:N}
\end{subfigure}
\caption{Sum throughput of the MIMO-NOMA, (a) versus the number of clusters per cell and (b) versus the number of transit antennas at the BS ($N = 3, N_r = 2$, and $P^{\max} = 30$ dBm).}\label{fig:SRMMIMO:KN}
\end{center}
\end{figure}
Fig.~\ref{fig:SRMMIMO:KN} shows the impact of the number of UEs per cell and the number of transmit antennas at the BS on the performance of the system. Fig.~\ref{fig:SRMMIMO:KN}(a) shows that MIMO-NOMA can deliver an
acceptable sum throughput for large $K$. Again, MIMO-NOMA outperforms CoMP in all $K$.
The sum throughput of the both systems decreases from a certain value of $K$
where there is not much degree-of-freedom (DoF) for leveraging multi-user diversity. Interestingly, MIMO-NOMA achieves its best
sum throughput for $K=3$ (6 UEs) while CoMP is peaked at $K=2$ (4 UEs).
Of course, these numbers are not magic and can be changed in other settings. Fig.~\ref{fig:SRMMIMO:KN}(b)
plots the sum throughput vs the number $N_t$ of antennas at the BSs.
\begin{table}[t]
\centering
\caption{Achieved User Throughput (bps/Hz) in Multi-user MIMO Multi-cell Systems}
\label{Table:RateperUE}
{\setlength{\tabcolsep}{0.45em}
\setlength{\extrarowheight}{0.65em}
\begin{tabular}{|c|c|c|c|c|c|c|}
\hline
\multirow{2}{*}{} & \multicolumn{2}{c|}{Throughput per UE} & \multicolumn{2}{c|}{ST per cell} & \multicolumn{2}{c|}{Total ST} \\ \cline{2-7}
& CoMP & NOMA & CoMP & NOMA & CoMP &NOMA \\ \hline
UE (1,1) &8.37&7.83& \multirow{4}{*}{20.68} & \multirow{4}{*}{27.16} & \multirow{12}{*}{\underline{\textbf{59.63}}} & \multirow{12}{*}{\underline{\textbf{77.23}}} \\ \cline{1-3}
UE (1,2) &8.74&7.62& & & & \\ \cline{1-3}
UE (1,3) &2.32&5.12& & & & \\ \cline{1-3}
UE (1,4) &1.25&6.59& & & & \\ \cline{1-5}
UE (2,1) &7.81&7.56& \multirow{4}{*}{18.76} & \multirow{4}{*}{24.34} & & \\ \cline{1-3}
UE (2,2) &6.31&6.24& & & & \\ \cline{1-3}
UE (2,3) &3.13&5.78& & & & \\ \cline{1-3}
UE (2,4) &1.51&4.76& & & & \\ \cline{1-5}
UE (3,1) &9.01&8.86& \multirow{4}{*}{20.19} & \multirow{4}{*}{25.73} & & \\ \cline{1-3}
UE (3,2) &6.25&6.92& & & & \\ \cline{1-3}
UE (3,3) &1.89&6.33& & & & \\ \cline{1-3}
UE (3,4) &3.04&3.62& & & & \\ \hline
\end{tabular}}
\end{table}
Table~\ref{Table:RateperUE} details the throughput at UEs under setting $N_t=6$ and $P^{\max}=30$ dBm.
Under the same QoS requirement threshold for UEs, the CoMP's throughput is mostly contributed
by the cell-center UEs, i.e. CoMP still tends to punish the cell-edge UEs, who are in poor channel condition.
Raising the QoS requirement for the cell-edge UEs to counter this discrimination would lead to the risk of
CoMP service feasibility. In contrast, in maximizing the system's throughput, MIMO-NOMA offers
much fairer and balanced services without contrasting the QoS requirement thresholds so it is very suitable for new
quality-of-experience (QoE) services for cell-edge UEs.
\begin{figure}
\begin{center}
\begin{subfigure}[CDF versus sum throughput for $P^{\max} = 15$ dBm.]{
\includegraphics[width=0.48\textwidth]{CDF_MIMO15dBm.eps}}
\label{fig:CDF:b}
\end{subfigure}
\begin{subfigure}[CDF versus sum throughput for $P^{\max} = 30$ dBm.]{
\includegraphics[width=0.48\textwidth]{CDF_MIMO30dBm.eps}}
\label{fig:CDF:a}
\end{subfigure}
\caption{CDF of the sum throughput (a) for $P^{\max} = 30$ dBm and (b) for $P^{\max} = 15$ dBm with $N = 2, K = 4, N_t = 5$, and $ N_r = 4$.}\label{fig:CDFMIMO}
\end{center}
\end{figure}
By Fig.~\ref{fig:CDFMIMO},
the system's throughput achieved by the proposed precoder design is compared to that achieved by
signal alignment MIMO-NOMA (SA MIMO-NOMA) \cite{DSP16} and interfering channel alignment CoMP MIMO-NOMA (ICA-CoMP MIMO-NOMA) \cite{ShinComL16} in the two-cell scenario with $K = 4$ users, $N_t = 5$, while $ N_r = 4$ is set to make the signal
and channel alignment feasible. In SA MIMO-NOMA, the inter-cluster interference is canceled
by detection vectors based SA technique at the UEs and zero-forcing (ZF) based precoder matrix at the BSs. In ICA-CoMP MIMO-NOMA,
a receive beamformer is constructed at the cell-edge UEs to align the interfering channels, and then a transmit beamformer based on the null space at the BS is designed to ensure zero inter-cell and inter-cluster interference. For the cell-center UEs, a ZF decoder is designed to cancel the inter-cluster interference only. To ensure a fair comparison, it is additionally
assumed that the cell-center UEs do not experience inter-cell interference as
they are far away from the neighboring cell in practice \cite{ShinComL16}.
The two different settings of $P^{\max}$ = (15 dBm, 30 dBm), according to 3GPP TR 36.942 v.9.0.1 with a 46-dBm
maximum transmit power for the 20 MHz bandwidth are under consideration.
Fig.~\ref{fig:CDFMIMO} plots the cumulative distribution function (CDF) of the sum throughput. As expected, MIMO-NOMA outperforms both ICA-CoMP MIMO-NOMA and SA MIMO-NOMA. Specifically,
by controlling the interference to the cell-edge UEs more efficiently, MIMO-NOMA reaches
1.5 bps/Hz and 5.9 bps/Hz higher than the ICA-CoMP MIMO-NOMA and SA MIMO-NOMA, respectively, in about $60\%$ of the simulated trials with $P^{\max}$ = 15 dBm (see Fig.~\ref{fig:CDFMIMO}(a)). With $P^{\max}$ = 30 dBm, it is even more essential (see Fig.~\ref{fig:CDFMIMO}(b)).
\subsection{Numerical Results for MISO-NOMA}
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth,trim={0cm 0.0cm -0cm -0cm}]{NOMA_MISO_Nt.eps}
\caption{Average sum throughput of the MISO-NOMA versus the number of transit antennas at the BS ($N = 3, K = 2$, and $P^{\max} = 30$ {dBm}).}
\label{fig:SRMMISO:N}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.48\textwidth,trim={0cm 0.0cm -0cm -0cm}]{CDF_MISO.eps}
\caption{CDF of the sum throughput for two different settings of the transmit power ($N = 3, K = 2, N_t = 6$, and $ N_r = 1$).}
\label{fig:SRMMISO:CDF}
\end{figure}
The performance of MISO-NOMA achieved by \eqref{beam1} is compared to that achieved
by the two-stage beamforming in \cite{Choi15}. The key idea of the two-stage beamforming
is that ZF beamforming is employed at the BS first to cancel the inter-pair interference and
then SIC is used for each pair of UEs. The optimal solution for two-stage beamforming can be easily found by using Algorithm~\ref{alg_Tailored_FW}.
Fig.~\ref{fig:SRMMISO:N} plots the total sum throughput vs. the number of transmit antennas at the BS
under setting $N=3$, $K=2$ and $P^{\max} = 30$ dBm. Note that two-stage beamforming requires $N_t\geq 2(K-1) + 2=4$.
Unsurprisingly, MISO-NOMA achieves a better sum throughput than the two-stage beamforming.
Two-stage beamforming achieves closer performance to MISO-NOMA as the number $N_t$
of transmit antennas increases, providing more degrees of freedom to leverage multi-user diversity.
Fig.~\ref{fig:SRMMISO:CDF} plots the CDF of the sum throughput at $N_t = 6$.
The performance gap is narrower at a higher power budget.
\subsection{Comparison for Different Cluster Sizes}
\begin{figure}
\begin{center}
\begin{subfigure}[Simulation setup considered in Fig.~\ref{fig:SRM3UE}(b) and Fig.~\ref{fig:SRM3UE}(c) with $r_n = 100 $ m, $r_m = 250 $ m, and $r_o = 500 $ m.]{
\includegraphics[width=0.35\textwidth]{Simulationlayout_3UE.eps}}
\label{fig:SRM3UE:a}
\end{subfigure}
\begin{subfigure}[Average sum throughput of the MIMO-NOMA versus $\bar{\mathsf{R}}$ for $N_r = 2$.]{
\includegraphics[width=0.47\textwidth]{SRM_clustersizeMIMO.eps}}
\label{fig:SRM3UE:b}
\end{subfigure}
\begin{subfigure}[Average sum throughput of the MISO-NOMA versus $\bar{\mathsf{R}}$ for $N_r = 1$.]{
\includegraphics[width=0.47\textwidth]{SRM_clustersizeMISO.eps}}
\label{fig:SRM3UE:c}
\end{subfigure}
\caption{Average sum throughput (b) for the MIMO-NOMA, and (c) for the MISO-NOMA with different cluster sizes ($N_t = 12$ and $P^{\max} = 30$ dBm).}\label{fig:SRM3UE}
\end{center}
\end{figure}
As the last numerical example, we investigate the system performance
in a single-cell scenario of more than two UEs grouped to create a virtual cluster for NOMA. There are 6 UEs in total,
which are randomly placed in three different areas, as shown by Fig.~\ref{fig:SRM3UE}(a). Two cell-center UEs
are located inside the disc of radius $r_n$ = 100 m, two cell-middle UEs
are located inside the ring of inner radius $100$ m and outer radius $250$ m, and
two cell-edge UEs are located inside the ring of inner radius $250$ m and outer radius $500$ m.
Different cluster sizes are considered: two UEs per cluster, three UEs per cluster, and six UEs per cluster.
For two-UE-per-cluster, a cell-center UE is randomly paired with a cell-middle UE, while the other
is randomly paired with a cell-edge UE. The unpaired cell-middle UE and cell-edge are then paired to create the third cluster.
There are two scenarios in grouping for three-UE-per-cluster.
\begin{itemize}
\item scenario-1 (\textit{more distinct channel conditions}): each cluster consists of a cell-center UE, a cell-middle UE
and a cell-edge UE;
\item scenario-2 (\textit{less distinct channel conditions}): each cluster consists of a cell-center UE.
\end{itemize}
The order of decoding messages for UEs in the same cluster of size three is as follows:
the message for a cell-edge UE is decoded by all UEs, the message for a cell-middle UE is decoded by itself and the third
UE by canceling the previously decoded message for the cell-edge UE from the intra-cluster,
and the message for the third UE is decoded by itself only by canceling all previously decoded messages from the
intra-cluster interference. Analogously, the messages for UEs $(1,1)$, $(1,2)$, $(1,3)$, $(1,4)$, $(1,5)$ and $(1,6)$
are successively decoded and canceled from the intra-cell interference in the case of six-UE-per-cluster, i.e. all UEs
of the same cell have NOMA.
The proposed algorithms are easily adapted for solution of the corresponding sum throughput maximization problems.
Fig.~\ref{fig:SRM3UE}(b) and Fig.~\ref{fig:SRM3UE}(c) plot the sum throughput
achieved by different clustering versus throughput threshold for MIMO-NOMA and MISO-NOMA
under setting $N_t = 12$ and $P^{\max} = 30$ dBm.
In general, the sum throughput achieved by the three-UE-per-cluster and six-UE-per-cluster based schemes are
dropped less than that achieved by two-UE-per-cluster based one when the threshold
$\bar{\mathsf{R}}$ raises. Specifically, in Fig.~\ref{fig:SRM3UE}(b), the sum throughput
of the three-UE scenario 2 and six-UE are worse than the two-UE for $\bar{\mathsf{R}} < 2.4$ bps/Hz and vice versa. The BS will allocate a much higher transmit power to the UE of the worst channel condition
in the three-UE and six-UE schemes than in the two-UE scheme to meet the QoS constraints. In other words, the cell-edge UEs' throughput is significantly improved in larger cluster sizes.
Notably, the sum throughput by three-UE scenario 1 catches up that by DPC
for larger $\bar{\mathsf{R}}$ in both Fig.~\ref{fig:SRM3UE}(b) and Fig.~\ref{fig:SRM3UE}(c).
In addition, six-UE scheme cannot provide a good sum throughput since large UEs per cluster may have error propagation in SIC leading to drastically reduce NOMA performance. Another interesting observation
is that the sum throughput of the three-UE scenario 1 outperforms that of the other schemes. Recalling that NOMA is more efficient by exploiting their channel condition differences, i.e. 3-UE scenario 1 with more distinct channel conditions. Consequently, a larger cluster size is recommended for more distinct channel conditions while a smaller cluster size is recommended for less distinct channel conditions.
\section{Conclusions}\label{sec:conclusion}
We have addressed the problem of sum throughput maximization in NOMA based systems
by proposing new path-following optimization algorithms. Numerical examples with realistic parameters have confirmed
their fast convergence to an optimal solution. They reveal that NOMA not only helps
increase the cell-edge UEs' throughput substantially but also achieves much higher total sum throughput.
The appropriate size of UE cluster with more distinctive channel gains has also been shown to achieve
remarkable gains in NOMA systems.
\section*{Appendix A: Proof for inequality (\ref{in1})}
By \cite[Appendix B]{RTKN14}
\begin{IEEEeqnarray}{rCl}
\bigl(\alpha \mathbf{V}_1+\beta\mathbf{V}_2\bigr)^H\bigl(\alpha\mathbf{X}_1+\beta\mathbf{X}_2\bigl)^{-1}\bigl(\alpha \mathbf{V}_1+\beta\mathbf{V}_2\bigl)
\preceq&& \nonumber\qquad\\
\alpha \mathbf{V}_1^H\mathbf{X}_1^{-1}\mathbf{V}_1+
\beta\mathbf{V}_2^H\mathbf{X}_2^{-1}\mathbf{V}_2&&
\end{IEEEeqnarray}
for all $\alpha\geq 0$, $\beta\geq 0$, $\alpha+\beta=1$ and $\mathbf{V}_1$, $\mathbf{V}_2$, $\mathbf{X}_1\succ \mathbf{0} $,
$\mathbf{X}_2\succ \mathbf{0} $. This means for all $\mathbf{x}$, function
\begin{IEEEeqnarray}{rCl}
f(\mathbf{V},\mathbf{X})=\mathbf{x}^H\mathbf{V}^H\mathbf{X}^{-1}\mathbf{V}\mathbf{x}
\end{IEEEeqnarray}
is convex. Then, for all $\mathbf{V}$, $\bar{\mathbf{V}}$, $\mathbf{X}\succ \mathbf{0} $,
$\bar{\mathbf{X}}\succ \mathbf{0} $ it is true that \cite{Tuybook}
\begin{IEEEeqnarray}{rCl}
f(\mathbf{V},\mathbf{X})&\geq& f(\bar{\mathbf{V}},\bar{\mathbf{X}}) + \bigl\langle \nabla f(\bar{\mathbf{V}},\bar{\mathbf{X}}),
(\mathbf{V},\mathbf{X})-(\bar{\mathbf{V}},\bar{\mathbf{X}})\bigl\rangle \nonumber\\
&=&\mathbf{x}^H\Bigl[\bar{\boldsymbol{V}}^H\bar{\boldsymbol{X}}^{-1}\mathbf{V}+
\mathbf{V}^H\bar{\boldsymbol{X}}^{-1}\bar{\boldsymbol{V}} \nonumber\\
&&\qquad\quad -\; \bar{\boldsymbol{V}}^H\bar{\boldsymbol{X}}^{-1}\mathbf{X}\bar{\boldsymbol{X}}^{-1}\bar{\boldsymbol{V}} \Bigl]\mathbf{x},
\end{IEEEeqnarray}
i.e.
\begin{IEEEeqnarray}{rCl}
\mathbf{x}^H\mathbf{V}^H\mathbf{X}^{-1}\mathbf{V}\mathbf{x}\geq \mathbf{x}^H\Bigr[\bar{\boldsymbol{V}}^H\bar{\boldsymbol{X}}^{-1}\mathbf{V}+
\mathbf{V}^H\bar{\boldsymbol{X}}^{-1}\bar{\boldsymbol{V}} \nonumber\\ -
\bar{\boldsymbol{V}}^H\bar{\boldsymbol{X}}^{-1}\mathbf{X}\bar{\boldsymbol{X}}^{-1}\bar{\boldsymbol{V}} \Bigl]\mathbf{x},\
\forall\mathbf{x}\
\end{IEEEeqnarray}
proving (\ref{in1}).
\section*{Appendix B: Proof for inequality (\ref{in2})}
Since function $\ln|\mathbf{X}|$ is concave on $\mathbf{X}\succ \mathbf{0} $, it is true that
\cite{Tuybook}
\begin{IEEEeqnarray}{rCl}
-\ln|\mathbf{A}|\geq -\ln|\mathbf{B}| - \bigl\langle \mathbf{B}^{-1},\mathbf{A}-\mathbf{B}\bigr\rangle, \nonumber\\
\forall\ \mathbf{A}\succ \mathbf{0}, \mathbf{B}\succ \mathbf{0}
\end{IEEEeqnarray}
or equivalently
\begin{IEEEeqnarray}{rCl}
\ln|\mathbf{A}^{-1}|\geq \ln|\mathbf{B}^{-1}| - \bigl\langle \mathbf{B}^{-1},\mathbf{A}-\mathbf{B} \bigl\rangle,\;\nonumber\\
\forall\ \mathbf{A}\succ \mathbf{0}, \mathbf{B}\succ \mathbf{0}. \label{proof28}
\end{IEEEeqnarray}
Then (\ref{in2}) follows by substituting $\mathbf{X}=\mathbf{A}^{-1}$ and $\bar{\mathbf{X}}=\mathbf{B}^{-1}$ into \eqref{proof28}.
\section*{Appendix C: Proof for inequality (\ref{zf8})}
By \cite[Th. 6]{Taetal16} function $\ln (1+x^{-1})$ is convex on $x>0$ so for all $x>0$ and $\bar{x}>0$, it is true that
\cite{Tuybook}
\begin{equation}\label{proof61}
\ln\bigl(1+x^{-1}\bigl) \geq \ln \bigl(1+\bar{x}^{-1}\bigl) + \bigl[(1+\bar{x})^{-1}-\bar{x}^{-1}\bigr](x-\bar{x}).
\end{equation}
Inequality (\ref{zf8}) then follows by substituting $z=x^{-1}$ and $\bar{z}=\bar{x}^{-1}$ into \eqref{proof61}.
\bibliographystyle{IEEEtran}
\balance
| c9b381e27432eee3b388cb9f3bda1902a517dc77 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Nematic liquid crystal phases are states of matter that possess long-range orientational order but are translationally invariant \cite{DeGennesBook}.
Historically, they were discovered in systems of rod-like molecules with a
$D_{\infty h}$ symmetry and had a revolutionary impact on the display industry.
However, it is generally accepted that the classification of nematic phases coincides
with three-dimensional ($3D$) point groups. Since the early $1970$s there
have been steady and tremendous efforts in the search for new nematic phases beyond uniaxial order.
Indeed, $D_{2h}$ biaxial nematics were proposed \cite{Freiser70} and their properties were discussed \cite{Alben73, Straley74} just shortly after theories of the uniaxial ones were established \cite{DeGennesBook}.
There is now strong evidence of their existence
\cite{Madsen04, Acharya04, Severing04} and they are believed to be
promising candidates for the next generation of liquid crystal displays (LCDs) \cite{BiaxialBook15}.
Another remarkable example of unconventional nematic phases whose existence has been established is the twist-bent liquid crystals formed from bent-core constituents with $C_{2v}$ symmetry \cite{Madsen04, Acharya04}.
They exhibit intriguing optical \cite{Merkel04, Neupane06, Gortz09, Chen14} and elastic \cite{Gleeson14, Kaur16} properties and rich transition sequences \cite{Lubensky02, Mettout05, Takezoe06, LongaPajakWydro09}, and are still subject to present studies.
Moreover, there is also great interest in more complex polyhedral
nematic phases classified by polyhedral groups. Attention has been focused on
the search for those phases experimentally in various chemical and colloidal systems \cite{Wiant08, Qazi10, Aoshima12, Lekkerkerker12} and numerically from packing shapes \cite{Blaak04, John04, John08, Duncan09, Duncan11, Marechal12, Wilson12, Glotzer12, Glotzer16a, Glotzer16b}, classifying the associated order parameter tensors \cite{Fel95, Michel01,HajiAkbariGlotzer15} and topological defects \cite{Mermin79, Michel80}, constructing theories supporting such orders and investigating their phase diagrams \cite{NelsonToner81, SteinhardtNelson1981, Lubensky02, Romano06, Romano08, TrojanowskiLonga12}, and examining the consequential macroscopic properties \cite{Stallinga94, Brand05}.
Nevertheless, in spite of considerable progress, we may have unveiled only a small corner of the rich landscape of the polyhedral phases. There are still many open questions on, e.g., their transition sequences and the nature of those phase transitions, the interactions of the associated topological defects and their influence on the thermodynamical, optical and mechanical properties of the system.
From a theoretical point of view, the difficulty is closely tied to the
complexity of those symmetries and their subgroup structure. These demand tensor
order parameters of high rank and lead to rich patterns of phase transitions where dynamics of topological defects also plays a crucial role.
Traditional Landau schemes and the associated lattice models are explicitly
based on order parameters, and, hence, may become extremely involved and difficult to handle when dealing with complicated symmetries.
They are also not convenient in accessing topological defects.
Furthermore, the full classification of the explicit form of those order
parameter tensors has been attained only recently \cite{Liu16, Nissinen16}.
However, lattice gauge theory, adopted from high-energy physics \cite{Kogut79},
has opened up new avenues to address these issues.
The application of this method to nematic orders dates back to the seminal works
of Lammert, Rokshar, and Toner in the mid-$1990$s \cite{LammertRoksharToner93, LammertRoksharToner95, TonerLammertRokshar95b}.
The authors utilized a $Z_2$ gauge theory to promote Heisenberg vectors to
directors, and formulated their model in terms of vectors and $Z_2$ gauge
fields, instead of $Q_{ab}$ tensors. They successfully capture the important
statistical physics of uniaxial nematics, especially the first-order
nematic-isotropic (NI) transition,
and show the power of lattice gauge theory in controlling dynamics of topological defeccts.
Variants of the Lammert-Rokshar-Toner model have also been applied to strongly correlated electron systems, for instance, in studies of charge fractionalization of superconductors \cite{SenthilFisher00, SenthilFisher01, DemlerSenthil05} and spin nematics \cite{ZaanenNussinov02, PodolskyDemler05}.
The works mentioned above have focussed exclusively on $Z_2$
symmetries and uniaxial orders. Only recently has the gauge-theoretical description been extended to accommodate general point-group symmetries, in the studies of $2+1 d$ quantum melting \cite{Liu15, Beekman17} and $3D$ thermal nematics \cite{Liu16, Nissinen16, Liu17} by ourselves and collaborators.
Its advantages have been proven both mathematically and practically, especially when dealing with $3D$ point groups which are in general non-Abelian.
First of all, it has been shown in solid mathematical terms that the gauge model fits all nematic orders into a uniform and efficient framework, regardless of their symmetry \cite{Liu16}.
This is in stark contrast to traditional order-parameter methods which are typically specific only for a single symmetry and often suffer from the complexity arising from high-rank ordering tensors.
Moreover, the formulation of the gauge model requires no prior knowledge of the underlying order parameter which is an essential input for traditional methods.
Instead, it acts as a machinery and generates a full classification of nematic ordering tensors which, to our knowledge, has never been done before \cite{Nissinen16}, though remarkable results of a more narrow scope have been obtained previously by other means \cite{Fel95, Michel01,HajiAkbariGlotzer15}.
Furthermore, in virtue of its generality, the gauge-theoretical method can also provide a global view over symmetries, which allows us to explore universal properties of different nematic orders \cite{Liu16}.
These include the insight of a relation between thermal fluctuations and symmetries, and the finding of a vestigial chiral phase that is reminiscent the chiral liquid reported in a recent experiment \cite{Dressel14}.
Last but not least, the gauge model is also naturally compatible with anisotropic interactions. By allowing anisotropy, it has mathematically predicted and numerically verified rich patterns of biaxial-uniaxial transitions and new types of biaxial-biaxial$^{\star}$ transitions \cite{Liu17}, enriching our understanding of biaxial orders.
In earlier works, we have focused on building the connection between generalized $3D$ nematics and non-Abelian gauge theories, and on exploring the topology of their phase diagrams.
In the present paper we study the nature of the NI transition for polyhedral orders by means of Monte Carlo simulations and a renormalization group analysis. This is not only important for physical properties of the system near the phase transition, but also relates to a fundamental question in statistical physics. That is, whether breaking $O(3)$ in different manners can give rise to new universality classes.
Moreover, it is worth pointing out that the gauge-theory model allows us to
easily exclude irrelevant symmetries and focus on the most important degrees of
freedom. Meanwhile, the model remains flexible enough to incorporate competing orders and disorders.
The symmetries we are interested in are the $7$ polyhedral subgroups of $O(3)$, $\{T, T_d, T_h, O, O_h, I, I_h\}$, requiring orientational tensors of rank higher than $2$.
Nematic phases of these symmetries are sometimes dubbed as octupolar or tetrahedral ($T_d$), cubic ($O, O_h$) or icosahedral ($I, I_h$) phases in literature. However, for convenience we will refer to them as generalized nematic phases when discussing general symmetries and $G-$nematics when discussing a specific instance of the symmetry $G$.
This convention was by no means invented by us, but already used in the textbook Ref.~\onlinecite{DeGennesBook}.
These polyhedral nematic phases have not been clearly identified in experiments.
Nevertheless, this does not mean that they are only of academical interest.
Indeed, modern technologies in nano and colloid science are able to synthesize and
manipulate mesoscopic particles with the desired symmetry to a high degree
\cite{Glotzer07, Glotzer10, Kraft09, Kraft11, Mark13, Huang15}, hence providing essential building blocks for the realization of polyhedral phases.
Moreover, it is also worth noting that these phases may emerge from systems of lower-symmetry constituents with suitable interactions or geometrical constraints, such as the proposed tetrahedral $T_d$ phase from $C_{2v}$-shaped molecules \cite{Lubensky02, TrojanowskiLonga12} and the cubic $O_h$ phase from $D_{\infty h}$-components \cite{Blaak99, Marechal12, Romano16}.
This paper is organized as follows. In Section \ref{sec:model}, we define the necessary degrees of freedom, and review the realization of generalized nematics in the language of lattice gauge theories.
Section \ref{sec:results} is devoted to Monte Carlo simulations. We first discuss the results of the chiral tetrahedral $T$ nematics in detail, then present those for other polyhedral symmetries with a discussion on their general features.
We compare our results with those from a renormalization scenario and other lattice models in Section \ref{sec:other_methods}.
Finally, we conclude and provide an outlook in Section \ref{sec:done}.
\section{Gauge-theory description of nematic phases} \label{sec:model}
\subsection{Degrees of freedom} \label{subsec:dof}
Instead of directly using physical order parameters, the fundamental degrees
of freedom are nonphysical matter fields and gauge fields in the gauge theoretical description.
The matter fields are $O(3)$ rotors describing all possible rotations in $3D$ real space. They can be parameterized by local orthonormal triads as
\begin{equation}
R = (\mb{l}, \mb{m}, \mb{n} )^{\mr{T}} \in O(3),
\end{equation}
where $\mb{n}^{\alpha} = \{\mb{l}, \mb{m}, \mb{n}\}$ with $\alpha = 1, 2, 3$ are the three axes of a local triad.
In concrete terms, they are defined by rotations that let $R$ coincide with the fixed ``laboratory'' axes $\mb{e}^{a} = \{\mb{e}_1, \mb{e}_2, \mb{e}_3 )$, and are parametrized by three Euler angles with respect to $\mb{e}^{a}$.
The three axes of $R$ satisfy the relation
\begin{equation}
\sigma = \mr{det}(R) = \mb{l} \cdot (\mb{m} \times \mb{n}) = \pm 1. \label{eq:handedness}
\end{equation}
This also defines the chirality or handedness, denoted by a pseudo-scalar $\sigma$, of the triad.
For $\sigma = 1$, $R$ describes rotations in $SO(3)$ and is usually referred to
as a proper rotation.
For convenience later on, we also define triads formed by pseudo-vectors
$\tmb{n}^{\alpha} = \{\tmb{l}, \tmb{m}, \tmb{n}\}$,
\begin{equation}
\widetilde{R} = (\tmb{l}, \tmb{m}, \tmb{n} )^{\mr{T}} \in SO(3), \label{SO3_triad}
\end{equation}
with $\tmb{l} \cdot (\tmb{m} \times \tmb{n}) \equiv 1$, describing rotations of a rigid body.
Correspondingly, a rotation in $O(3)$ can be decomposed as
\begin{equation}
R = \sigma \widetilde{R}.
\end{equation}
The gauge fields are defined as a connection between two neighboring triads.
They are also rotations, but in contrast to the global $O(3)$ rotations, they describe local rotations with respect to the axes of a triad.
The introduction of gauge fields makes it is possible to compare two triads
locally at different locations. The symmetry of the gauge fields is a point
group by construction.
In the simplest situation, when homogenous distributions of order parameters are preferred, it coincides with the
symmetry of the ``mesogens'' of a liquid crystal.
In terms of the terminology of Ref. \onlinecite{Mettout06}, this symmetry is
chosen to be the symmetry of the effective building blocks of the system, and
can in turn represent the symmetry of the state in the fully symmetry-broken
phase. In other words, the scheme works at a coarse-grained level by construction.
\subsection{The model} \label{subsec:model}
\begin{figure}
\centering
\subfigure[]{
\includegraphics[width=0.45\textwidth]{cube}\label{fig:cube}
}
\vfill
\subfigure[]{
\includegraphics[width=0.45\textwidth]{lattice}\label{fig:lattice}
}
\caption{ The correspondence between the shape of mesogens and the configurations of local triads. Here, a mesogen with the chiral cubic $O$ symmetry is considered as an example.
Fig.~\ref{fig:cube}: The orientation of a cube, representing physical order parameter fields, is encoded into $24$
triad configurations, generated by a gauge symmetry $O$ acting on local axes of a triad.
Fig.~\ref{fig:lattice}: A lattice of cubes maps on to a lattice with triad matter
fields residing on the sites and point-group symmetric gauge fields defined on the links.
}
\end{figure}
Having established the necessary information on the degrees of freedom, we now introduce the gauge-theoretic model.
It is defined on an auxiliary cubic lattice, which is permitted by the continuous translational symmetry of nematic liquid crystals. It hence applies to both continuous and discrete point groups.
The Hamiltonian takes the following form,
\begin{equation}
H = -\sum_{\corr{i,j}} \mr{Tr}(R_i^{\mr{T}} \mathbb{J} U_{ij} R_j). \label{eq:model}
\end{equation}
$R_i$ is the $O(3)$ triad at a lattice site $i$, $U_{ij}$ is the gauge field of a point-group symmetry $G$ mediating the interaction between two nearest-neighboring sites and lives on the link $\corr{i,j}$.
$\mathbb{J}$ is a coupling matrix and can act as a tuning parameter. It is constrained by the symmetry of nematic mesogens, \textit{i.e.} the gauge symmetry $G$, in such a way that it has to be invariant under the transformation $\Lambda \mathbb{J} \Lambda^{\mr{T}} = \mathbb{J}$, $\forall \Lambda \in G$.
It follows that $\mathbb{J}$ is isotropic and takes the form $\mathbb{J} = J \mathbbm{1}$ for all polyhedral groups, where $J$ is positive for ferromagnetic (alignment) coupling.
However, anisotropies of $\mbb{J}$ are possible for nematics with axial symmetries, and are responsible to the generalized biaxial-uniaxial and biaxial-biaxial$^{\star}$ transitions \cite{Liu17}.
The Hamiltonian Eq. \eqref{eq:model} is invariant under gauge transformations
\begin{equation}
R_i \rightarrow \Lambda_i R_i,\quad U_{ij} \rightarrow \Lambda_i U_{ij} \Lambda_j^{\mr{T}},\quad \forall \Lambda_i \in G. \label{eq:gauge_trans}
\end{equation}
This invariance identifies the orientation of a triad defined by $R_i$ with that defined by $\Lambda_i R_{i}$, and thus encodes the symmetry of the mesogens under consideration.
To be more concrete, we define a local triad $n^{\prime \beta}_i = U^{\beta\gamma}_{ij} n^{\gamma}_j$ at a site $i$, and rewrite the Hamiltonian Eq. \eqref{eq:model} in the following form,
\begin{equation}
H = -\sum_{\corr{i,j}} J^{\alpha\beta} n^{\alpha}_i \cdot n^{\prime \beta}_j\label{eq:triad_form},
\end{equation}
where Greek letters in superscripts are associated with the axes of local triads, and $J^{\alpha\beta} = J \delta_{\alpha, \beta}$ for polyhedral symmetries.
By doing so, the triad $n^{\gamma}_j$ has been brought to the same local gauge, namely the same body-fixed coordinate, as the site $i$ by parallel transporting so that the orientation of the two triads can be compared.
Then, considering the gauge transformations in Eq. \eqref{eq:gauge_trans}
running over $G$ at a site $j$, but letting other sites unchanged, $\Lambda_{i
\neq j} \equiv \mathbbm{1}$, this generates a set of $n^{\prime \beta}_j$ which
consists of all the equivalent definitions of the orientation of the underlying mesogen of the symmetry $G$ at the site $j$.
Let us take a chiral cube with the symmetry $O$ as an example. The orientation of the cube maps to $24$ configurations of a local triad, corresponding to the $24$ transformations of the group $O$.
When all these configurations are considered and identified, we are effectively describing the orientation of a cube via that of a set of local triads. The symmetry of the underlying mesogens is thus realized by the gauge symmetry. Note that the choice of $\Lambda_{i \neq j} \equiv \mathbbm{1}$ in the above example is purely for simplifying the example.
Gauge transformations Eq. \eqref{eq:gauge_trans} can be performed independently to all the sites.
Consequently, in the low-energy limit, the orientational interaction of physical mesogens (order parameter fields) is hence effectively encoded in the gauge model Eq. \eqref{eq:model} of nonphysical degrees of freedom, as depicted by Fig. \ref{fig:lattice}.
This procedure can also be shown in explicit mathematical terms by integrating out the gauge fields Eq. \eqref{eq:model}, and we refer to our earlier publications Refs. \onlinecite{Liu16} and \onlinecite{Nissinen16} for detailed proofs.
However, though maybe less intuitive, it is advantageous to work with gauge degrees of freedom.
As the symmetry of the order parameter fields is directly implemented by the gauge symmetry, the gauge model applies to all point group symmetries by simply choosing a desired $G$.
\subsection{Discussion on the phases}
It is well known that gauge symmetries cannot break spontaneously \cite{Elitzur75}. As a consequence, the fully symmetry-broken phase of the gauge model, Eq. \eqref{eq:model}, features a ground state manifold $O(3)/G$ which is just the order-parameter manifold of a $G-$nematic phase.
This phase is usually referred to as the Higgs phase in the language of gauge theories, and corresponds to the aligned state of mesogens in the current context, i.e., a nematic phase of the symmetry $G$.
On the other hand, the disordered $O(3)$ liquid phase is realized by the confinement phase of the gauge model Eq. \eqref{eq:model}.
There are three comments to be made on the above statements to avoid confusion.
First, the Higgs phase just mentioned corresponds to a situation where $O(3)$
has been completely broken to the local symmetry $G$. However, aside from this,
there can also be an intermediate Higgs phase that breaks $O(3)$ to a larger point group $G^{\prime}$ satisfying $G \subset G^{\prime} \subset O(3)$, featuring vestigial order.
This could happen when fluctuations in some sectors of the degrees of freedom
are more pronounced than in others.
For instance, in case $G$ is a finite axial group, the fully ordered Higgs phase is a biaxial nematic phase of the symmetry $G$.
When fluctuations in the plane perpendicular to the so-called primary axis are sufficiently strong or weak, upon changing temperature, the system may experience an intermediate uniaxial and/or biaxial$^{\star}$ phase, respectively, before entering the disordered isotropic liquid phase, as discussed in detail in Ref. \onlinecite{Liu17}.
As another example, an intermediate Higgs phase may also appear as a chiral
liquid phase. Possible realizations of this phase are systems formed from
mesogens of a chiral polyhedral symmetry, $G \in \{T, O, I\}$. For these
symmetries, fluctuations in orientations are much more pronounced than those in the chirality.
Thus a phase that breaks real-space inversion and mirror symmetry but is invariant under $SO(3)$ rotations can emerge between the nematic phase and the $O(3)$ liquid phase. We will encounter this situation again in the next section and a systematic and detailed discussion can be found in Ref. \onlinecite{Liu16}.
Second, the distinction between the Higgs phase and the confinement phase is only a property when $G$ is a nontrivial subgroup of $O(3)$. In the limit $G = O(3)$, these two phases are continuously connected and indistinguishable \cite{FradkinShenker79}, consistent with the fact that there is no symmetry breaking for $O(3)/O(3)$.
Last but no least, as mentioned earlier, we will focus on homogeneous distributions of order parameter fields, as is realized by the gauge model Eq. \eqref{eq:model}.
However, inhomogeneous distributions may also lead to interesting phenomena. One example is the chiral-$T$ phase with a helical structure of $T_d$ molecules, owing to an explicit chiral elastic term in Landau free energy, as discussed in Ref.~\onlinecite{Fel95}. We can also introduce a gauge invariant chiral term to Eq.~\eqref{eq:model} to incorporate such helical structure for general symmetries, but will leave it for future study.
What is relevant to the current paper is that, as such a chiral term is independent to the additional quartic terms of high rank tensors, it is unlikely that they can change the nature of fluctuation-induced first-order phase transitions for the $T$ and $T_d$ symmetry which will be discussed in Sec. \ref{subsec:MF_FG}.
\subsection{Topological defects}
Before closing the section, let us briefly comment on the dynamics of gauge fields in the model Eq. \eqref{eq:model} and its relation to topological defects of liquid crystals.
From the point of view of gauge theories, the model Eq. \eqref{eq:model} consists of a single Higgs term, in which form the dynamics of gauge fields $U_{ij}$ arises purely from the interaction with the matter fields $R_i$.
In general, however, the gauge fields can have their own dynamics, which in the simplest case is described by a plaquette term in the following form,
\begin{equation}
H_{\mr{YM}} = - \sum_{\square} K_{C_{\mu}} (U_{\square}) \mr{Tr}[U_{\square}]. \label{eq:K_term}
\end{equation}
This generalizes the defect suppression term in Refs.~\onlinecite{LammertRoksharToner93, LammertRoksharToner95}, and essentially is
an analog of the Yang-Mills theory as the $U_{ij}$'s are in general non-Abelian \cite{Kogut79}. However, comparing to usual lattice Yang-Mills theories in high energy physics, here we are interested in discrete symmetries.
$U_{\square} \in G$ denotes the orientated product of the gauge fields around a minimal plaquette of the lattice, $U_{\square} = \prod_{\corr{ij} \in \square} U_{ij}$. A plaquette of $U_{\square} \neq \mathbbm{1}$ is frustrated and represents a gauge flux.
The coupling strength $K_{C_{\mu}}$ depends on the trace of $U_{\square}$, so it is a function of the conjugation class $C_{\mu}$ of $G$, which means that gauge fluxes in the same conjugation class are physically equivalent.
A gauge flux has the effect that after a triad travels around it in a closed circuit, the triad is rotated by $U_{\square}$, just like circling a disclination. Furthermore, the classification of gauge fluxes coincides with the Volterra classification of disclinations \cite{Kleman08}. Even though this classification, as well as the Volterra classification, is in general not identical to the homotopy classification of topologically stable defects \cite{Mermin79}, it includes all the elementary topologically stable defects and can be used to construct the full homotopy classification.
As we can easily tune $K_{C_{\mu}}$ to suppress or prompt certain types of defects, the interaction in $H_{\mr{YM}}$ provides a possible route to study the influence of topological defects on thermodynamical properties of nematic liquid crystals.
As is known from the study of lattice gauge theory \cite{FradkinShenker79}, this may qualitatively change both the topology of the phase diagrams and the nature of underlying phase transitions.
Refs.~\onlinecite{LammertRoksharToner93, LammertRoksharToner95, TonerLammertRokshar95b} also showed remarkable examples in this context, where the first-order uniaxial-isotropic transition is split into two continuous ones when the defect suppression is sufficiently large.
For general symmetries, we expect rich physics to explore when treating the elastic term ~\eqref{eq:model} and non-Abelian defect term ~\eqref{eq:K_term} at equal footing.
Nevertheless, this is beyond the scope of the current paper, and deserves a separate systematic study. Thus, for simplicity, we will set $K_{C_{\mu}} \equiv 0$ and focus on Eq. \eqref{eq:model} in the following. Physically, this means none of the topological defects are assigned a particular core energy.
\setlength{\tabcolsep}{2pt}
\renewcommand{\arraystretch}{1.8}
\begin{table}
\centering
\caption{{\bf Generators of $3D$ polyhedral point groups.}
The first column specifies the symmetries. The second column shows a set of generators which produce the entire group elements of the underlying symmetry. The third column gives the order of the symmetries. Note that, there are multiple ways to choose the generator set, but they are all equivalent (for more information see Refs. \onlinecite{BookButler12} and \onlinecite{BookBishop93}). A representation of the generators listed below is catalogued in Table \ref{table:generators}.
}
\label{table:groups}
\begin{tabular}{ | c | c | c | }
\hline
\hline
\parbox{2.1 cm}{\centering{\bf Symmetry}}
& \parbox{3.5 cm}{\centering{\bf Generators}}
& \parbox{1.3 cm}{\centering{\bf Order}}
\\ \hline
{\centering{$T$}}
& $c_2(\mathbf{n})$, $c_3(\mathbf{l}+\mathbf{m}+\mathbf{n})$
& 12
\\ \hline
{\centering{$T_d$}}
& $c_2(\mathbf{n})$, $c_3(\mathbf{l}+\mathbf{m}+\mathbf{n})$ , $\sigma_d$
& 24
\\ \hline
{\centering{$T_h$}}
& $c_2(\mathbf{n})$, $c_3(\mathbf{l}+\mathbf{m}+\mathbf{n})$ , $\sigma_h$
& 24
\\ \hline
{\centering{$O$}}
& $c_2(\mathbf{m} + \mathbf{n})$,
$c_3(\mathbf{l}+\mathbf{m}+\mathbf{n})$, $c_4({\mathbf{n}})$
& 24
\\ \hline
{\centering{$O_h$}}
& \parbox{3.5 cm}{\centering $c_2(\mathbf{m} + \mathbf{n})$,
$c_3(\mathbf{l}+\mathbf{m}+\mathbf{n})$, $c_4({\mathbf{n}})$, $\sigma_h$}
& 48
\\ \hline
{\centering{$I$}}
& $c_2(\mathbf{n})$, $c_3(\mathbf{l}+\mathbf{m}+\mathbf{n})$, $c_5(\mathbf{l} + \tau\mathbf{n})$
& 60
\\ \hline
{\centering{$I_h$}}
& \parbox{3.5 cm} {\centering $c_2(\mathbf{n})$, $c_3(\mathbf{l}+\mathbf{m}+\mathbf{n})$, $c_5(\mathbf{l} + \tau\mathbf{n})$, $\sigma_h$}
& 120
\\ \hline
\hline
\end{tabular}
\end{table}
\setlength{\tabcolsep}{2pt}
\renewcommand{\arraystretch}{1.8}
\begin{table}
\centering
\caption{{\bf Definitions of the generators.}
Here we specify the representation of the generators for the $3D$ polyhedral point groups used in our simulations. $c_N(\mathbf{p})$ denotes a $N$-fold rotation about a vector $\mathbf{p}$ defined by the local axes $\{\mathbf{l}, \mathbf{m}, \mathbf{n}\}$. $\tau = (\sqrt{5}+1)/2$ is the golden ratio, which is involved in case of icosahedral groups $I$ and $I_h$. $\sigma_h$ defines a reflection about the $(\mathbf{l}, \mathbf{m})$ plane, while $\sigma_d$ indicates a reflection about the plane $(\mathbf{l} + \mathbf{m}, \mathbf{n})$.
}
\label{table:generators}
\begin{tabular}{ | c | c | }
\hline
\hline
\parbox{2.9 cm}{\centering{\bf Generator}}
& \parbox{5 cm}{\centering{\bf Representation}}
\\ \hline
{\centering $c_{2}(\mathbf{n})$}
& $\begin{pmatrix}
-1 & 0 & 0 \\
0 & -1 & 0 \\
0 & 0 & 1
\end{pmatrix}$
\\ \hline
{\centering $c_{2}(\mathbf{m}+\mathbf{n})$}
& $\begin{pmatrix}
-1 & 0 & 0 \\
0 & 0 & 1 \\
0 & 1 & 0
\end{pmatrix}$
\\ \hline
{\centering $c_{3}(\mathbf{l}+\mathbf{m}+\mathbf{n})$}
& $\begin{pmatrix}
0 & 1 & 0 \\
0 & 0 & 1 \\
1 & 0 & 0
\end{pmatrix}$
\\ \hline
{\centering $c_{4}(\mathbf{n})$}
& $\begin{pmatrix}
0 & -1 & 0 \\
1 & 0 & 0 \\
0 & 0 & 1
\end{pmatrix}$
\\ \hline
{\centering $c_{5}(\mathbf{l} + \tau \mathbf{n})$}
& $\begin{pmatrix}
1/2 & -\tau/2 & 1/(2\tau) \\
\tau/2 & 1/(2\tau) & -1/2 \\
1/(2\tau) & 1/2 & \tau/2
\end{pmatrix}$
\\ \hline
{\centering $\sigma_h$}
& $\begin{pmatrix}
1 & 0 & 0 \\
0 & 1 & 0 \\
0 & 0 & -1
\end{pmatrix}$
\\ \hline
{\centering $\sigma_d$}
& $\begin{pmatrix}
0 & -1 & 0 \\
-1 & 0 & 0 \\
0 & 0 & 1
\end{pmatrix}$
\\ \hline
\hline
\end{tabular}
\end{table}
\section{Numerical results} \label{sec:results}
As discussed in the last section, the gauge model Eq. \eqref{eq:model} can act
as an efficient and flexible framework for studying nematic order with arbitrary point-group symmetries. Moreover, it is readily accessible by Monte Carlo simulations.
In light of this, we examined the nature of the nematic-isotropic transition for all $3D$ polyhedral groups, \textit{i.e.}, $\{T, T_d, T_h, O, O_h, I, I_h\}$.
For the convenience of the reader, generators of these symmetries are provided in Tables \ref{table:groups} and \ref{table:generators}, while more information can be found in textbooks, e.g., Refs. \onlinecite{BookButler12} and \onlinecite{BookBishop93}. Sch\"{o}nflies notation is used thorough out the manuscript.
We use the Metropolis algorithm and run simulations on cubic lattices with volume $V = 8^3, 16^3, 24^3$.
The simulations include three steps. In the first step, the transition
temperatures are located as precise as possible by examining the peak position of the heat capacity and the nematic susceptibility (see below).
Procedures of cooling random initial states and heating uniform states are compared.
In the next step, we perform extensive simulations near the transition, histogramming the distribution of the observables of interest. The typical amount of \textit{independent} samples used are of the order of $10^5$ to $10^7$.
In the last step, the histograms are further improved by Ferrenberg-Swendsen reweighting \cite{Ferrenberg88, BookNewman99}, and the transition temperatures are estimated from the shape of the histograms \cite{Lee90, BookLandau13}.
In the rest of the section, we first present the results for tetrahedral $T$ nematics and discuss the general features of this phase transition in detail. Then we discuss other symmetries.
\subsection{$SO(3)/T$ transition} \label{subsec:SO3_T}
\begin{figure}\centering
\includegraphics[width=0.45\textwidth]{Nematicity_T}
\caption{ Nematicity of the tetrahedral $T$ symmetry for a broad range of temperatures in units of the Higgs coupling $J$, computed by Eq. \eqref{eq:SO3}. Error bars are smaller than the data points. These, together with the susceptibility and heat capacity (not shown), give us information of the nontrivial region where massive simulations are performed to locate and determine the order of the transition.} \label{fig:Nematicity}
\end{figure}
The tetrahedral group $T$ consists of $12$ proper rotations leaving a tetrahedron invariant. Therefore, aside from the orientational order, a $T$-symmetric nematic phase also takes an intrinsic chiral order, breaks inversion and any kinds of mirror symmetries of real space.
Note that the $T$-nematic phase we discuss here is different from the $T$-phase discussed by Fel in Ref. \onlinecite{Fel95}.
In the later case the $T$ symmetry arises from the helical structure of $T_d$-symmetric mesogens and is associated with a different order parameter (see below and also Sec. \ref{subsec:MF_FG} for further discussion).
It turns out that for nematics formed from constituents with a $T$-symmetry and
a flexible handedness, as well as from those with an $O$- or $I$-symmetry,
fluctuations in the orientation sector are in general much more pronounced than those in the chirality sector \cite{Liu16}.
Consequently, the system develops orientational order and chiral order sequentially.
Furthermore, by comparing numerical results with a mean-field analysis, it is shown that the two phase transitions are well separated.
This implies that the relevant degrees of freedom associated with the NI transition lie in the $SO(3)$ sector in Eq. \eqref{eq:model}.
In mathematical terms, we can rewrite the gauge model Eq. \eqref{eq:model} as
\begin{equation}
H = - J \sum_{\corr{i,j}} \sigma_i \sigma_j \mr{Tr}(\widetilde{R}_i^{\mr{T}} U_{ij} \widetilde{R}_j),
\end{equation}
by taking out the $Z_2$ center of $O(3)$, where $\sigma_i$ denotes the handedness fields and $\widetilde{R}_i$ are $SO(3)$ triads, defined in Eq. \eqref{eq:handedness} and Eq. \eqref{SO3_triad}.
The handedness fields are frozen during the NI transition, thus simplifying the problem to a phase transition breaking $SO(3)$, governed by a Hamiltonian
\begin{equation}
H^{\prime} = - J \sum_{\corr{i,j}} \mr{Tr}(\widetilde{R}_i^{\mr{T}} U_{ij} \widetilde{R}_j). \label{eq:SO3}
\end{equation}
\begin{figure}
\centering
\subfigure[]{
\includegraphics[width=0.45\textwidth]{Histo_E_T}\label{fig:Histo_E_T}
}
\vfill
\subfigure[]{
\includegraphics[width=0.45\textwidth]{Histo_Nem_T}\label{fig:Histo_Nem_T}
}
\caption{ Histograms of the energy density, Fig. \ref{fig:Histo_E_T}, and of the nematicity, Fig, \ref{fig:Histo_Nem_T}, for tetrahedral $T$ nematics at the phase transition at different lattice sizes. Transition temperatures, $T_c$, are estimated by finding where the histogram peaks are of equal height. The data shown have been improved by Ferrenberg-Swendsen reweighting in the vicinity of the simulated temperatures, which are at the precision of four decimal digits.}\label{fig:Histo_T}
\end{figure}
The orientational order parameter of $T$-nematic phases is a rank-$3$ tensor taking the following form (note the difference to the $T_d$ order parameter, see Sec. \ref{subsec:MF_FG}),
\begin{align} \label{eq:op_T}
\overline{\mbb{O}}^{T} &= \corr{\mbb{O}^{T}_i}_V \nonumber \\
&= \sum_{\mr{cyc}}\corr{\tmb{l} \otimes \tmb{m} \otimes \tmb{n} - \frac{1}{6}\varepsilon_{abc} \tmb{e}_a \otimes \tmb{e}_b \otimes \tmb{e}_c }_V,
\end{align}
where $\mbb{O}^{T}_i$ denotes the local ordering tensor at a coarse-grained lattice site $i$, $\corr{...}_V$ denotes the average over the volume, and $\sum_{\mr{cyc}}$ is the sum running over cyclic permutations of local axes $\{\tmb{l}, \tmb{m}, \tmb{n}\}$ \cite{Nissinen16}.
The Levi-Civita tensor is introduced to make the ordering tensor traceless, so $\overline{\mbb{O}}^{T}$ becomes a zero tensor in the liquid phase. This term is only needed when working with $SO(3)$ triads where the handedness of each local triad is fixed. If the handedness is allowed to fluctuate (\textit{i.e}, the case of an $O(3)$ triad) summing over the two kinds of chirality cancels this term.
In case of homogenous distributions, instead of using the tensor form Eq. \eqref{eq:op_T}, we can characterize a nematic order of symmetry $G$ by its magnitude defined as
\begin{equation}
q = \sqrt{\mr{Tr}(\mbb{O}^G \cdot \mbb{O}^G)}.
\end{equation}
This quantity is called the \textit{nematicity} and generalizes the concept of magnetization \cite{Nissinen16}.
Consequently, we can further define the susceptibility of $q$ in the standard way,
\begin{equation}
\chi_{q} = \beta V \left( \corr{q^2} - \corr{q}^2 \right),
\end{equation}
and detect the NI transition by the peak of $\chi_{q}$, where $\beta$ is the inverse temperature.
As confirmed in our simulations, the peak of $\chi_{q}$ coincides with that of the heat capacity defined in the standard way, indicating that the nematicity $q$ is indeed a valid scalar order parameter.
We have measured the $SO(3)/T$ NI transition with Eq. \eqref{eq:SO3} by monitoring several quantities, including the energy, the nematicity, histograms of the two, the heat capacity and the susceptibility. As many of them reveal the same information, we present only those which are necessary for the following discussions.
In Fig. \ref{fig:Nematicity}, we show the behavior of the nematicity for a broad range of temperatures, and Figs. \ref{fig:Histo_E_T} and \ref{fig:Histo_Nem_T} show the histograms of the energy density and the nematicity at phase transition, $P(E)$ and $P(q)$, respectively.
As shown in the energy histogram Fig. \ref{fig:Histo_E_T}, a double-peak
behavior emerges at sufficiently large lattice sizes, and appears more pronounced when the lattice size increases, indicating the occurrence two stable, co-existing phases, which is a hallmark of a first-order phase transition. The distance between the valley and peak of a histogram, measuring the difficulty for the system to tunnel between two phases, also increases with the lattice size, as expected.
The physical meaning of the two peaks is revealed by the nematicity histogram
from the same simulations, shown in Fig. \ref{fig:Histo_Nem_T}. With increasing
lattice size, aside from the behavior that the peaks become more pronounced, the
first peak notably moves to the left. We expect it eventually goes to zero in the thermodynamical limit, indicating a disordered liquid phase. The other peak corresponds to the nematic phase.
\subsection{Other symmetries} \label{subsec:others}
\begin{figure}
\centering
\includegraphics[width=0.45\textwidth]{Histo_E_all}
\caption{ Histograms of the energy density for symmetries $\{T_d, T_h, O, I\}$ at the NI transition, for lattice size $V = 24^3$. $T_c$'s are estimated in the same way as for the $SO(3)/T$ case.
Depending on symmetries, the simulations are performed with the original gauge theory Eq. \eqref{eq:model} or its $SO(3)$ sector Eq. \eqref{eq:SO3}, indicated in the figure. The data of the $SO(3)/O$ and $SO(3)/I$ transition also represent those of the $O(3)/O_h$ and $O(3)/I_h$ transition, respectively.
The two-peaks behavior reveals the first-order nature for all these transitions. The histograms of the nematicity (not shown) exhibit similar features.
The same binning size is used for the first three symmetries, while a smaller
binning size is used for the $SO(3)/I$ transition. However, the heights of these
histograms are not comparable, since even though these symmetries are studied
via a common framework here, they correspond to different physical models.
} \label{fig:Histo_others}
\end{figure}
We performed the same procedure as in the case of the $SO(3)/T$ NI transition for all other polyhedral symmetries, \textit{i.e.}, symmetries of $\{T_d, T_h, O, O_h, I, I_h\}$. Similar to the $T$-nematics, the breaking of chiral and of rotational symmetry in case of the cubic $O$ and the icosahedral $I$ symmetry are separated, owing to huge fluctuations in orientation \cite{Liu16}.
Thus, they are also simulated in terms of Eq. \eqref{eq:SO3}, resulting in a $SO(3)/O$ and a $SO(3)/I$ transition, respectively.
On the other hand, the nonchiral symmetries $\{T_d, T_h, O_h, I_h\}$ are studied via the original model Eq. \eqref{eq:model}, corresponding to a direct breaking of $O(3)$.
We find first-order behavior for all these transitions. Fig. \ref{fig:Histo_others} shows the energy histogram of symmetries $\{T_h, T_d, O, I\}$.
The results of the $SO(3)/O$ and $SO(3)/I$ transition also represent those of the $O(3)/O_h$ and $O(3)/I_h$ transition, which have very similar behavior as the former two, with slightly higher $T_c$'s.
This may be understood by the fact that the $Z_2$ center in the latter two cases can be factorized as a trivial $Z_2/Z_2$ theory, leading to the same order parameter manifold as the $SO(3)/O$ and $SO(3)/I$ cases, respectively.
Indeed, $O$-nematics and $O_h$ nematics, as well as $I$- and $I_h$-nematics,
share the same orientational order parameter, only distinguished by a pseudo-scalar chiral order parameter \cite{Nissinen16}.
We also studied the behavior of the nematicity and its histogram for these symmetries. Although the curves corresponding to different symmetries are well separated (since the phase transitions occur at different temperature scales) in Fig. \ref{fig:Histo_others},
the histograms of the nematicities overlap closely for these symmetries,
especially in the disorder region where all the disordering peaks are located at some small nematicity value close to $0$.
Moreover, they show very similar features as seen in Figs. \ref{fig:Nematicity} and \ref{fig:Histo_T} for the $SO(3)/T$ transition case.
They are therefore not presented.
One notable feature of Fig. \ref{fig:Histo_others} is that the peaks of the
histogram shift to lower energy scales as symmetries increase (the energy density is normalized via $E_G = - 9V$ which is the energy when all mesogens uniformly align up),
indicating a decrease in the corresponding transition temperatures. This can be understood as
a consequence of more pronounced orientational fluctuations for high symmetries, which in turn results in an increasing difficulty to stabilize the order.
Note that this feature is manifest when using a common metric, the Higgs coupling strength $J$, for all the symmetries. This metric is not a direct physical measure in the sense that it describes the interaction strength between the auxiliary gauge fields and matter fields, rather than that of physical order parameter fields.
Although the latter one depends in principle on the Higgs coupling, the derivation of the relation is in general nontrivial. Nevertheless, this does not prevent us to obtain general insights on the nature of the phase transitions.
Clearly, Fig. \ref{fig:Histo_others} reveals the generic first-order nature of
the NI transition for all these symmetries. It is not clear yet how the strength
of these first-order transitions depends on their respective symmetries. However, this depends on microscopic details of a system and a universal conclusion might not exist.
\section{Relations and comparisons to other methods} \label{sec:other_methods}
We have numerically reached the conclusion of a generic first order NI transition from a particular framework. Moreover, this is consistent with existing results from other methods, including a general perspective from mean field theories, renormalization group (RG) analyses \cite{Radzihovsky01} and other lattice models \cite{Romano06, Romano08}, as we elaborate below.
\subsection{Mean-field theories and RG} \label{subsec:MF_FG}
A significant difference between nematic order and spin or vector order is that in general the former requires a tensor order parameter due to nontrivial internal symmetries. In case of the $D_{\infty h}$ uniaxial nematics, the order parameter is a rank-$2$ tensor, $Q_{ab}$, which gives rise to a third order term, $Q_{ab} Q_{bc} Q_{ca}$, in the Landau-de Gennes theory and makes the NI transition discontinuous.
For polyhedral symmetries $\{T_h, O, O_h, I, I_h\}$ the nematic order parameters are also even rank tensors \cite{Nissinen16}, which take the following form for nonchiral groups $\{T_h, O_h, I_h\}$,
\begin{align}
\mathbb{O}^{T_h}
&= \mathbf{l}^{\otimes 2} \otimes \mathbf{m}^{\otimes 2} + \mathbf{m}^{\otimes 2} \otimes \mathbf{n}^{\otimes 2} + \mathbf{n}^{\otimes 2} \otimes \mathbf{l}^{\otimes 2}
\nonumber \\
& \quad -\frac{2}{5} \delta_{ab} \delta_{cd}\bigotimes_{\substack{ \mu = a,b,c, d}} \mathbf{e}_{\mu}
+ \frac{1}{10} \bigg( \delta_{ac} \delta_{bd} \bigotimes_{\substack{ \mu = a,c,b, d}} \mathbf{e}_{\mu}
\nonumber \\
& \quad + \delta_{ad} \delta_{bc}\bigotimes_{\substack{ \mu = a,d,b, c}} \mathbf{e}_{\mu} \bigg), \label{eq:op_Th}\\
\mathbb{O}^{O_h} &= \mathbf{l}^{\otimes 4} + \mathbf{m}^{\otimes 4} + \mathbf{n}^{\otimes 4}
-\frac{1}{5}\sum_{\text{pairs}} \delta_{ab} \delta_{cd} \bigotimes_{\substack{ \mu = a,b,c, d}} \mathbf{e}_{\mu}, \label{eq:op_Oh} \\
\mathbb{O}^{I_h} &= \sum_{\rm{cyc}} \bigg[ \mathbf{l}^{\otimes 6} +\sum_{ \{+,- \}}
\Big(\frac{1}{2} \mathbf{l} \pm \frac{\tau}{2} \mathbf{m} \pm \frac{1}{2\tau} \mathbf{n} \Big)^{\otimes 6} \bigg]
\nonumber \\
& \quad -\frac{1}{7} \sum_{\text{pairs}} \delta_{ab} \delta_{cd} \delta_{ef} \bigotimes_{\substack{ \mu = a,b,c, \\ d,e,f}} \mathbf{e}_{\mu}, \label{eq:op_Ih}
\end{align}
and $\mathbb{O}^{O} = \{\mathbb{O}^{O_h}, \sigma\}$ and $\mathbb{O}^{I} = \{\mathbb{O}^{I_h}, \sigma\}$ for chiral groups $\{O, I\}$, where $\sum_{\text{cyc}}$ runs over the cyclic permutations of $\{\mb{l}, \mb{m}, \mb{n}\}$, $\sum_{\text{pairs}}$ sums over all nonequivalent pairings of the indices of the Kronecker delta functions, and $\otimes$ denotes the tensor product.
Similar to the $D_{\infty h}$ uniaxial case, even in naive mean-field theories for these symmetries, there are third order terms of the form
$\mbb{O}_{abcd} \mbb{O}_{cdef} \mbb{O}_{efab} $
or $\mbb{O}_{abcdef} \mbb{O}_{defghk} \mbb{O}_{ghkabc}$,
giving rise to a first order phase transition.
On the other hand, the tetrahedral $T_d$ order requires a rank-$3$ order parameter tensor, $\mbb{O}^{T_d}$, where
\begin{align} \label{eq:op_Td}
\mathbb{O}^{T_d} = \sum_{\mathrm{cyc}} (\mathbf{l} \otimes \mathbf{m} + \mathbf{m} \otimes \mathbf{l})\otimes \mathbf{n}.
\end{align}
This forbids the appearance of the third order term in a naive mean-field theory which has the following free energy density,
\begin{align} \label{eq:MF_Td}
f_{T_d} = \frac{1}{2} (\partial_i \mbb{O}^{T_d}_{abc} \partial_i \mbb{O}^{T_d}_{abc}) + \frac{r}{2} \mbb{O}^{T_d}_{abc} \mbb{O}^{T_d}_{abc} + u (\mbb{O}^{T_d}_{abc} \mbb{O}^{T_d}_{abc})^2,
\end{align}
and predicts a continuous NI transition, where $r$ and $u$ are phenomenological coefficients.
A further RG study by Radzihovsky and Lubensky in Ref. \onlinecite{Radzihovsky01} shows that Eq. \eqref{eq:MF_Td} has a second order phase transitions, falling into the $O(7)$ universality class, as the symmetric traceless tensor $\mbb{O}^{T_d}$ can be mapped to a $7$ dimensional vector.
However, this transition is unstable against fluctuations. Following the general symmetry principles there exists another fourth order term,
$\mbb{O}^{T_d}_{abc} \mbb{O}^{T_d}_{ade} \mbb{O}^{T_d}_{bdf} \mbb{O}^{T_d}_{cef}$,
representing fluctuations, which qualitatively modifies the nature of the NI transition and makes it first order.
The tetrahedral $T$ order faces a similar situation as the $T_d$ one. Its order parameter $\mbb{O}^T$ defined in Eq. \eqref{eq:op_T} is also an odd rank and symmetric traceless tensor (The kernel of $\mbb{O}^T$ is not symmetric, but becomes so after carrying out the trace). Therefore, the paradigm of Ref. \onlinecite{Radzihovsky01} for the $T_d$ case equally applies to the $SO(3)/T$ transition, with a different tensor representation for the $O(7)$ vector. Consequently, the second order NI transition predicted by the MF theory is converted to a first order one, agreeing with our results from the gauge model.
With this new perspective, let us look back to the symmetries $\{T_h, O, O_h, I, I_h\}$. Although the first-order nature of their NI transitions can already be concluded from a naive mean-field treatment, we should not omit other quartic couplings for completeness.
$\mbb{O}^{O_h}$ and $\mbb{O}^{I_h}$ are also symmetric and traceless tensors, and lead to $5$ and $7$ quartic terms
(the number of independent terms may be reduced by one), respectively.
It is more tricky in the $T_h$ case. $\mbb{O}^{T_h}$ is only partially symmetric (it is only invariant under switching the first or the last two indices), hence even the second and the third order coupling are not unique.
Of course not all of those couplings are necessary to appear in a particular
system. However, this implies that it requires very precise fine tuning to
promote the NI transition for polyhedral symmetries to second order.
\subsection{Other lattice models} \label{subsec:other_lattice}
Next we compare our results with those from other lattice models. Remarkable examples are Refs. \onlinecite{Romano06} and \onlinecite{Romano08} for the $T_d$ and $O_h$ order. (Other examples have been mainly focused on uniaxial and biaxial orders \cite{BiaxialBook15}.)
They are typically constructed by an interaction potential between two rigid molecules or mesogens of a certain symmetry.
The orientation of a mesogen is described by $M$ (nonorthogonal) unit vectors, spanned from a common local origin and organized in a way that explicitly has the desired symmetry.
The potential, $V_{ij}$, is then defined in terms of Legendre polynomials of the inner product of those vectors at the lowest nontrivial order,
\begin{align} \label{eq:pair_potential}
V_{ij} &= V(\mb{v}^{(m)}_i, \mb{v}^{(m^{\prime})}_j) \nonumber \\
& \sim -c_l \sum^M_{m,m^{\prime} = 1} P_l(\mb{v}^{(m)}_i \cdot \mb{v}^{(m^{\prime})}_j),
\end{align}
where the coefficient $c_l$ is positive for ferromagnetic coupling, $P_l(...)$ denotes the Legendre polynomial at the order $l$, and $\mb{v}^{(m)}_i$ are local unit vectors at the lattice site $i$.
For the cubic $O_h$ order, the $\mb{v}^{(m)}_i$ coincide with our local triads
$\mb{n}^{\alpha}_i$, and the Legendre polynomials are trivial for $l < 4$
\cite{Romano06}. On the other hand, in case of the tetrahedral $T_d$ order,
the $\mb{v}^{(m)}_i$ are the $4$ three-fold axes of a regular tetrahedron and are
nonorthogonal \cite{Fel95}, whereas the Legendre polynomials become nontrivial only
from $P_3$ onwards \cite{Romano08}.
Following the same principle and given the expression Eq. \eqref{eq:op_Ih}, one expects that this requires $15$ vectors, which are the $2$-fold axes of a regular icosahedron \cite{Fel95}, and sixth order Legendre polynomials.
It is not hard to show that the interaction potential Eq. \eqref{eq:pair_potential} can be understood as a counterpart of the Lebwohl-Lasher model \cite{Lebwohl72} for general symmetries, and is equivalent to the inner product between two order parameter tensors,
\begin{equation} \label{eq:effective_potential}
V_{ij} \sim \mr{Tr}(\mbb{O}^G_i \cdot \mbb{O}^G_j).
\end{equation}
As shown in Refs. \onlinecite{Liu16, Nissinen16}, this is exactly the leading order of the effective Hamiltonian of the gauge model Eq. \eqref{eq:model} after tracing out the gauge fields.
Therefore, it is not a surprise that our results of the $T_d$ and $O_h$ NI transition agree with those in Refs. \onlinecite{Romano06, Romano08}, and the agreements to other polyhedral symmetries, $\{T, T_h, O, I, I_h\}$, can also be expected.
However, lattice models for these symmetries in the type of the potential Eq.
\eqref{eq:pair_potential} are yet to be developed to the best of our knowledge.
Unlike the gauge model Eq. \eqref{eq:model}, which can readily be applied for all point-group symmetries, the potential Eq. \eqref{eq:pair_potential} is symmetry-dependent and involves large amounts of vectors and high-order Legendre polynomials, whose high complexity in actual use can be anticipated.
For instance, in case of the $I_h$ order, it involves $225$ Legendre polynomials of order $6$.
However, the potential Eq. \eqref{eq:pair_potential} has advantages in a more straightforward connection with microscopic interactions of liquid crystal mesogens, as it is built directly on physical order parameter fields.
Moreover, it is interesting to see how this method applies to the $T$ and $T_h$
symmetries, where the role of mirrors may be manifest, as well as the relation and difference of the resultant lattice models to those of the $T_d$ and $O_h$ case, which are the symmetry they halve, respectively.
\section{Summary} \label{sec:done}
Rotational symmetry breaking is ubiquitous and plays an important role in
condensed matter physics and statistical physics. One of its intriguing features
is that there is a multitude of ways to break a symmetry into its subgroups, leading to a large array of exotic phases.
In this work, we have examined the nature of phase transitions breaking the rotational group $O(3)$ to polyhedral point groups.
Such phases are prime candidates in the search for unconventional nematic liquid crystals, in particular in the field of nano and colloidal science.
We found that the transitions from the nematic phase to the isotropic liquid phase are generically first order for all polyhedral symmetries.
Furthermore, the polyhedral NI transitions are robust in the sense that they require fine tuning of a high precision in order to achieve a second order phase transition. This feature is inherited from the complexity of the group structure of polyhedral symmetries.
Moreover, along the lines of the discussion in Sec. \ref{subsec:MF_FG}, we
anticipate the NI transition of generalized uni- and bi-axial nematics, which
breaks $O(3)$ to axial point groups $\{C_n, C_{nv}, C_{nh}, S_{2n}, D_{n},
D_{nh}, D_{nd} \}$, to be generically of first order as well.
As discussed in detail in Ref. \cite{Nissinen16}, the order parameter of
axial symmetries in general has the structure $\mbb{O}^G = \{\mbb{A}^G,
\mbb{B}^G, \sigma\}$, where $\mbb{A}^G = \mbb{A}^G[\mb{n}]$ defines the order of
the primary axis chosen to be $\mb{n}$, $\mbb{B}^G = \mbb{B}^G[\mb{l}, \mb{m}]$
or $\mbb{B}^G[\mb{l}, \mb{m}, \mb{n}]$ defines the order in the perpendicular
plane and is required for finite axial symmetries, and $\sigma$ defines the
chiral order as seen in Sec. \ref{subsec:SO3_T} and is only relevant for the proper axial groups $\{C_n, D_n\}$.
For symmetries $\{ C_{nh}, S_{2n}, D_{n}, D_{nh}, D_{nd} \}$, $\mbb{A}^G$ is a
rank-two tensor and coincides with the $Q_{ab}$ director. Hence, following the
Landau-de Gennes theory, it is immediately clear that regardless of the in-plane structure, the NI transition for these symmetries will be generically first order.
For symmetries $C_n$ and $C_{nv}$, the primary order parameter $\mbb{A}^G$ is a vector, and continuous phase transitions seem to be preferred. However, when $n > 1$ but finite, the direct NI transition will be also first order, owing to the existence of an even and/or high rank $\mbb{B}^G$ tensor, as in the cases of polyhedral symmetries.
Even at $n = 1$, where both $\mbb{A}$ and $\mbb{B}$ are vectorial, the order of the phase transition will depend on their coupling.
Therefore, even though there are diverse patterns to break the $O(3)$ symmetry, second-order transitions and corresponding universality classes may be quite rare.
The familiar Heisenberg universality class related to the breaking of $O(3)$ to $O(2) \cong C_{\infty v}$ is a special case.
Our results add new insights to the physics of exotic orientational phases
and hopefully facilitate the understanding of future experiments.
Finally, we would like to note that in the present work only a single symmetry is considered in the realization of each polyhedral nematic.
Nevertheless, as has been discussed by many authors for $T_d$ and $O_h$ symmetries \cite{Lubensky02, TrojanowskiLonga12, Blaak99, Marechal12, Romano16}, polyhedral phases may emerge from systems formed from less-symmetric constituents.
Although it is hard to imagine a second-order NI transition from this, it would be interesting to explore the general pattern of symmetry emergence in liquid crystal systems.
Given the compatibility with competing orders and the potential power on
controlling topological defects, we expect that the gauge-theory scenario will be suitable to achieve this aim without losing simplicity.
\textbf{Acknowledgments}
We would like to thank Leo Radzihovsky and Henk Bl\"{o}te for stimulating discussions, and Jaakko Nissinen for useful discussions and related collaborations.
This work is supported by FP7/ERC starting grant No. 306897. Our simulations made use of the ALPSCore library \cite{Gaenko17}.
\bibliographystyle{apsrev4-1}
| b440fb185e92d51c990c1c8a33d2de55277581e2 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\par
\setcounter{num}{1}
\setcounter{equation}{0}
Inverse relations play an important role in combinatorics \cite{Riordan}.
The {\em binomial inversion formula}, which states that for sequences $\{a_n\}$ and $\{b_n\}$ ($n=0,1, 2, \ldots$),
\begin{equation}
\label{b1}
a_n=\sum_{k=0}^{n} {n \choose k}(-1)^k b_k
\;\;\;\mbox{\rm if and only if}\;\;\;
b_n=\sum_{k=0}^{n} {n \choose k}(-1)^k a_k,
\end{equation}
is a typical inverse relation of interest in \cite{Choi,Donaghey,Prodinger,Sun,Wang,Wojcik}. Specifically,
(\ref{b1}) motivated Sun \cite{Sun} to investigate the following sequences.
\begin{defn}
{\rm
Let $\{a_n\}$ ($n=0,1, 2, \ldots$) be a sequence such that
\begin{equation}
\label{first}
(-1)^{s-1} \, a_n=\sum_{k=0}^{n} {n \choose k}(-1)^k a_k; \;\;s=1 \;\;\mbox{or}\;\; s=2.
\end{equation}
We refer to $\{a_n\}$ as an {\em invariant sequence} (when $s=1$) or an
{\em inverse invariant sequence} (when $s=2$) {\em of the first kind}.
}
\end{defn}
Several examples of invariant sequences of the first kind can be found in \cite{Sun}, including
$$ \{\frac{1}{2^n}\}, \; \{ nF_{n-1}\},
\; \{L_n\},\; \{(-1)^nB_n\} \;\; (n\geq 0),
$$
where $F_{-1}=0$ and $\{F_n\}$, $\{L_n\}$, and $\{B_n\}$ are the Fibonacci sequence, Lucas sequence,
and Bernoulli numbers \cite{Comtet}, respectively.
In this paper, we will establish (see Lemma \ref{bf2proof}) the {\em modified binomial inversion formula} such that
\begin{equation}
\label{b2}
a_n=\sum_{k=n}^{\infty} {k \choose n}(-1)^k b_k
\;\;\;\mbox{\rm if and only if}\;\;\;
b_n=\sum_{k=n}^{\infty} {k \choose n}(-1)^k a_k.
\end{equation}
Motivated by (\ref{b2}), we will introduce and consider the following sequences.
\begin{defn}
{\rm
Let $\{a_n\}$ ($n=0, 1, 2, \ldots$) be a sequence such that
\begin{equation}
\label{second}
(-1)^{s-1} a_n = \sum_{k=n}^{\infty} {k \choose n}(-1)^k a_k; \;\; s=1 \;\;\mbox{or}\;\; s=2.
\end{equation}
We refer to $\{a_n\}$ as an {\em invariant sequence} (when $s=1$) or an {\em inverse invariant sequence}
(when $s=2$) {\em of the second kind}.
}
\end{defn}
Naturally arising are the questions of existence, identification, and construction of (inverse) invariant sequences
of the second kind, as well as the problem of characterizing such sequences and examining their
relationship to their counterparts of the first kind.
Invariant sequences, which are also called self-inverse sequences in \cite{Wang}, have indeed been studied by several authors
\cite {Choi,Donaghey,Sun,Wang}. They are naturally connected to involutory (also known as involution or self-invertible) matrices
\cite{Horn} and to Riordan involutions \cite{Cheon}. Involutory matrices find use in numerical methods for
differential equations \cite{Andrew, Horn}.
They are also useful in cryptography, information theory, and computer security by providing convenient encryption and decryption methods \cite{Acharya}.
Motivated by Shapiro's open questions \cite{Shapiroq}, Riordan involutions have been intensely investigated as a combinatorial
concept \cite{Cameron,Cheon}.
In this paper, we investigate invariant sequences by means of the eigenspaces of $PD$ and $P^TD$, where $P$ is the Pascal matrix and $D$
an infinite diagonal matrix with alternating diagonal entries in $\{1, \,-1\}$ (see Sections 2, 3). In fact, $PD$ and $P^TD$ are
involutory matrices and $PD$ is a Riordan involution.
Our investigation follows the ideas and connections of invariant sequences to the eigenspaces of $PD$ and $P^TD$
developed in Choi et al. \cite{Choi}. This will allow us to associate (inverse) invariant sequences of the first
and second kinds, as well as identify and construct such sequences (Section 4).
\section{Notation and preliminaries}
\par
\setcounter{num}{2}
\setcounter{equation}{0}
The following notation and conventions are used throughout the manuscript.
\begin{itemize}
\item
The infinite matrices in this paper have infinite numbers of rows $i$ and columns
$j$, with $i,j\in\{0,1, 2, \ldots\}$.
\item
$\mathbf{E}_\lambda (A)$ denotes the eigenspace of a (finite or infinite)
matrix $A$ corresponding to its eigenvalue $\lambda$.
\item
For a matrix $A$ with columns $A_j$ ($j=0,1, 2, \ldots$) and with
${\mathbf 0}_j$ denoting the vector of zeros in $\mathbb{R}^j$,
${A\hspace{0.01cm}^\downarrow}$ denotes the matrix whose $j$th column
is
${\displaystyle
\left[\begin{matrix}
\mathbf 0_j\\
A_j
\end{matrix} \right]
}$ where $\mathbf 0_0$ is vacuous.
\item
For a matrix $A$, its (possibly infinite) row and column index sets are $J$ and $K$,
respectively. For $J_0 \subseteq J, K_0 \subseteq K$, let
$A(J_0|K_0)$ denote the matrix obtained from $A$ by deleting rows in $J_0$ and
columns in $K_0$, and let $A[J_0|K_0]$ denote the matrix
$A(\bar{J_0}|\bar{K_0})$, where $\bar{J_0}=J\setminus J_0, \bar{K_0}=K\setminus K_0$.
For brevity, write $A(\cdot|K_0)$ and $A(J_0|\cdot)$ in place of $A(\emptyset|K_0)$
and $A(J_0|\emptyset)$, respectively.
Further, for $m,n\in\{0,1, 2, \ldots\}$, we let $A_{m,n}=A[\{0,1, 2, \ldots, m\}|\{0,1, 2, \ldots, n\}]$;
$A_{m,m}$ is abbreviated by $A_m$.
\item
The binomial coefficient (``$i$ choose $j$") is denoted by $i\choose j$
with the convention that it equals $0$ when $i<j$ or $j<0$.
\item
$P=\left[\begin{matrix}
{i}\choose{j}
\end{matrix} \right]$
$(i,j=0, 1, 2, \ldots)$ denotes the (infinite) {\em Pascal matrix}.
\item
$D={\rm diag}(1, -1, 1, -1, \ldots)$.
\item
Infinite real sequences $\{x_n\}$ are identified with the infinite
dimensional real vector space $\mathbb{R}^\infty$ consisting
of column vectors $\mathbf{x}=[x_0, x_1, x_2, \ldots ]^T$.
\end{itemize}
Notice that as a consequence of the binomial inversion formula (\ref{b1}), we have that
$$ P^{-1}=DPD=
\left[ (-1)^{i-j}
\begin{matrix}
{i}\choose{j}
\end{matrix} \right] \;\;
(i,j=0, 1, 2, \ldots).$$
Thus $(PD)^{-1}=PD$ and (\ref{b1}) can be converted \cite{Choi} into a vector equation for
$\mathbf{x}=[a_0,a_1, a_2, \ldots]^T$ and $\mathbf{y}=[b_0,b_1, b_2, \ldots]^T \in \mathbb{R}^\infty$,
as follows:
\begin{equation}
\label{vbf}
PD\mathbf{x}=\mathbf{y}
\;\;\; \mbox{\rm if and only if} \;\;\;
PD\mathbf{y}=\mathbf{x}.
\end{equation}
\begin{lem}
\label{bf2proof}
Let $P$ and $D$ be the Pascal matrix and the diagonal matrix defined above,
and let $\mathbf{x}, \mathbf{y} \in \mathbb{R}^\infty$. Then
\begin{equation}
\label{vbf2}
P^TD\mathbf{x}=\mathbf{y}
\;\;\;\mbox{\rm if and only if}\;\;\;
P^TD\mathbf{y}=\mathbf{x},
\end{equation}
and the modified binomial inversion formula {\rm (\ref{b2})} holds.
\end{lem}
\begin{proof}
As $P^{-1}=DPD$, we have $(P^TD)^{-1}=D(P^T)^{-1}=P^TD$. As a consequence, (\ref{vbf2}) holds.
Letting $\mathbf{x}=[a_0, a_1, a_2, \ldots]^T$ and $\mathbf{y}=[b_0, b_1, b_2, \ldots]^T$ implies (\ref{b2}).
\end{proof}
Let $\mathbf{F}=[F_0, F_1, F_2, \ldots ]^T$ and $\mathbf{L}=[L_0, L_1, L_2, \ldots]^T$
denote the vectors in $\mathbb{R}^{\infty}$ whose entries are the members of the Fibonacci and Lucas
sequences, respectively; that is
\[ F_0=0, \; F_1=1, \; F_n=F_{n-1}+F_{n-2} \;\; (n\geq 2), \]
\[ L_0=2, \; L_1=1, \; L_n=L_{n-1}+L_{n-2} \;\; (n\geq 2). \]
The generating functions of $\mathbf{F}$ and $\mathbf{L}$
are $h_1(x)={x \over {1-x-x^2}}$ and $h_2(x)={{2-x}\over {1-x-x^2}}$, respectively \cite{Brualdi, Comtet}.
The following fact is known, however we include a proof for completeness.
\begin{lem}
\label{wellknown}
$PD\,\mathbf{F} = - \mathbf{F} \;\;\mbox{and}\;\; PD\,\mathbf{L} = \mathbf{L}.$
\end{lem}
\begin{proof}
Let $g(x)={1 \over {1-x}}$ and $f(x)={-x \over {1-x}}$.
For $j=0, 1, 2, \ldots$, the generating function of the $j$th column of $PD$ is $g(x)f(x)^j$ \cite{Shapiro}.
Thus the generating function of $PD\mathbf F$ is
\begin{equation*}
\begin{split}
[g(x), g(x)f(x), g(x)f(x)^2, \ldots][F_0, F_1, F_2, \ldots]^T&=g(x)(F_0+F_1f(x)+F_2f(x)^2+\cdots)\\
&=g(x)h_1(f(x))=-h_1(x),
\end{split}
\end{equation*}
which implies that $PD\,\mathbf{F} = - \mathbf{F}$. The proof of $PD\,\mathbf{L} = \mathbf{L}$ is similar.
\end{proof}
That is, $-1$ and $1$ are eigenvalues of $PD$ and consequently of $P^TD$.
In fact, these are the only eigenvalues of $PD$ and $P^TD$; see \cite{Choi}.
The corresponding eigenspaces are infinite dimensional. Indeed, if we
consider the Pascal-type matrices ${P\hspace{0.01cm}^\downarrow}$ and
${Q\hspace{0.01cm}^\downarrow}$ constructed via the Pascal matrix $P$ and
the matrix
\[ Q=P+\left[
\begin{array}{cc}
1 & {\mathbf 0}^T\\
{\mathbf 0}& P
\end{array}\right],
\]
then, as shown in \cite{Choi}, the columns of
\[
\left[\begin{matrix}
{\mathbf 0}^T\\
{P\hspace{0.01cm}^\downarrow}
\end{matrix} \right]\quad
\mbox{and} \quad
{Q\hspace{0.01cm}^\downarrow}
\]
form bases for $\mathbf{E}_{-1}(PD)$ and $\mathbf{E}_{1}(PD)$, respectively.
The following observation follows directly from the definitions and properties mentioned above.
\begin{obs}
\label{obs1}
The entries of $\mathbf{x} \in \mathbb{R}^\infty$ form
\begin{itemize}
\item
an invariant sequence of the first kind if and only if
$\mathbf{x} \in \mathbf{E}_{1} (PD)$;
\item
an inverse invariant sequence of the first kind if and only if
$\mathbf{x} \in \mathbf{E}_{-1} (PD)$;
\item
an invariant sequence of the second kind if and only if
$\mathbf{x} \in \mathbf{E}_{1} (P^TD)$;
\item
an inverse invariant sequence of the second kind if and only if
$\mathbf{x} \in \mathbf{E}_{-1} (P^TD)$.
\end{itemize}
\end{obs}
Based on Observation \ref{obs1}, our goal is to study the
eigenspaces $\mathbf{E}_\lambda (PD)$ and $\mathbf{E}_\lambda (P^TD)$
($\lambda \in \{ 1, -1\}$) and discover their relationships. Our approach
entails showing the existence of an infinite invertible matrix $N$ such that
\begin{equation}
\label{goal1}
N(P^TD)N^{-1}=(P_1^TD_1)\bigoplus (P_1^TD_1)\bigoplus \cdots \bigoplus (P_1^TD_1)\bigoplus \cdots
\end{equation}
and
\begin{equation}
\label{goal2}
D(N^{-1})^TD\,(PD)\,DN^TD=(P_1D_1)\bigoplus (P_1D_1)\bigoplus \cdots \bigoplus (P_1D_1)\bigoplus \cdots,
\end{equation}
which are infinite direct sums of copies of $P_1^TD_1$ and $P_1D_1$, respectively. This result will be
applied to characterize $\mathbf{E}_\lambda(PD)$ and $\mathbf{E}_\lambda(P^TD)$. Extending the work in
\cite{Choi}, we will also show that the columns of $P^{T\downarrow}$ and ${Q}^{T\downarrow}(0|0)$ form
bases for $\mathbf{E}_\lambda(P^TD)$. This will indeed allow us to investigate the relationships between
invariant sequences of the first and second kinds.
\section{The Eigenspaces of $P^TD$ and $PD$}
\par
\setcounter{num}{3}
\setcounter{equation}{0}
Let $A=[a_{ij}]$ $(i, j=0, 1, 2, \dots)$ be the matrix defined by
\[
a_{ij}=\left \{ \begin{array}{lr}
(-1)^{j-i}, & {\rm if}~ i \leq j, \\
0, & {\rm if}~ i>j,
\end{array} \right.
\]
and let $J(a)$ denote the infinite Jordan block of the form
\[
\left[ \begin{array}{ccccccc}
~~ a & 1 & & & & & \\
~~ & a & 1 & & & O& \\
& & a & 1 & & & \\
& & & \ddots & \ddots & & \\
& & & & a & 1 &\\
& & O & & & a &\ddots\\
& & & & & & \ddots~~
\end{array} \right].
\]
It readily follows that $A^{-1}=J(1)$.
In the next two lemmas, we will construct an infinite matrix $N$ and its
inverse $M=N^{-1}$, which will give rise to similarity transformations of
$P^TD$ and $PD$ into direct sums as in (\ref{goal1}) and (\ref{goal2}).
\begin{lem}
\label{lemma1}
Let $m$ be a positive integer and let $N_{(m)}=H_{(m)} H_{(m-1)}\cdots H_{(1)}$,
where $H_{(k)}=I_{2k-2}\bigoplus F$ {\rm ($k=1, 2, \ldots, m$)}. Then
$$\lim_{m \to \infty}N_{(m)}=N=[n_{ij}^{\infty}]$$
is the infinite matrix defined by
$n_{00}^{\infty}=1$, $n_{0j}^{\infty}=0$ and $n_{i0}^{\infty}=0$ for $i, j=1, 2, \ldots$, and
\[
n_{ij}^{\infty}=\left \{ \begin{array}{lr}
(-1)^{j-i} {{\lfloor {{i-1}\over 2}\rfloor+j-i}\choose {\lfloor {{i-1}\over 2}\rfloor}},& {\rm if}~~ 1\leq i \leq j, \\
0, & {\rm if}~~ i>j\geq 1.
\end{array} \right.
\]
\end{lem}
\begin{proof}
Let $m$ be a positive integer and let $N_{(m)}=H_{(m)}H_{(m-1)}\cdots H_{(1)}=[n_{ij}^{m}].$
We will prove that
\[
n_{ij}^{m}=\left \{ \begin{array}{lr}
(-1)^{j-i} {{k+j-i}\choose {k}},& {\rm if}~~ i \leq j, \\
0, & {\rm if}~~ i>j
\end{array} \right.
\]
by induction on $m$, where
\[
k=\left \{ \begin{array}{lr}
\lfloor {{i-1}\over 2}\rfloor,& {\rm if}~i=1, 2, \ldots, 2m;~~ j=1, 2, \ldots, \\
m-1, & {\rm if}~i=2m+1, 2m+2, \ldots;~j=1, 2, \ldots
\end{array} \right.
\]
The claim is clear for $m=1$. For $m=2$, by the construction of $H_{(1)}$ and $H_{(2)}$, we have
\[
n_{ij}^{2}=\left \{ \begin{array}{lr}
(-1)^{j-i} {{k+j-i}\choose {k}}, & {\rm if}~~ i \leq j, \\
0, & {\rm if}~~ i>j,
\end{array} \right.
\]
where
\[
k=\left \{ \begin{array}{lr}
\lfloor {{i-1}\over 2}\rfloor, & {\rm if}~i=1, 2, 3, 4;~~ j=1, 2, \ldots, \\
1, & {\rm if}~i=5, 6, \ldots;~j=1, 2, \ldots,
\end{array} \right.
\]
since for each $i=3, 4, \ldots$ and each $j=1, 2, \ldots$,
$$
n_{ij}^{2}=\sum_{l=i}^{\infty}(-1)^{l-i}(-1)^{j-l} {{j-l}\choose{0}} =(-1)^{j-i}\sum_{l=i}^{j} {{j-l}\choose{0}}
=(-1)^{j-i} {{1+j-i}\choose{1}}.
$$
Let now $m \geq 3$. Then by the construction of $H_{(m)}$,
we have $n_{ij}^{m}=n_{ij}^{m-1}$ for all $i=1, 2, \ldots, 2m-2$
and all $j=1, 2, \ldots$
By the induction hypothesis,
\[
n_{ij}^{m}=\left \{ \begin{array}{lr}
(-1)^{j-i} {{k+j-i}\choose {k}},& {\rm if}~~ i \leq j, \\
0, & {\rm if}~~ i>j,
\end{array} \right.
\]
where
\begin{eqnarray}
k=\left \{ \begin{array}{lr}
\lfloor {{i-1}\over 2}\rfloor, & {\rm if}~i=1, 2, \ldots, 2m;~~ j=1, 2, \ldots, \\
m-1, & {\rm if}~i=2m+1, 2m+2, \ldots;~j=1, 2, \ldots,\\
\end{array} \right.
\end{eqnarray}
because
\begin{equation*}
\begin{split}
n_{ij}^{m}&=\sum_{l=i}^{\infty}(-1)^{l-i}(-1)^{j-l} {{m-2+j-l} \choose {m-2}}\\
&=(-1)^{j-i}\sum_{l=i}^{j}{{m-2+j-l} \choose {m-2}}=(-1)^{j-i} {{m-1+j-i} \choose {m-1}}
\end{split}
\end{equation*}
for each $i=2m-1, 2m, \ldots$ and each $j=1,2, \ldots$
Thus, by $(2.1)$, we have that $\lim_{m \to \infty}N_{(m)}=N=[n_{ij}^{\infty}]$ given by
\[
n_{ij}^{\infty}=\left \{ \begin{array}{lr}
(-1)^{j-i} {{\lfloor {{i-1}\over 2}\rfloor+j-i}\choose {\lfloor {{i-1}\over 2}\rfloor}},& {\rm if}~~ i \leq j, \\
0, & {\rm if}~~ i>j
\end{array} \right.
\]
for each $i, j=1, 2, \ldots$ Clearly, we have
$n_{00}^{\infty}=1$, $n_{0j}^{\infty}=0$ and $n_{i0}^{\infty}=0$ for $i, j=1, 2, \ldots$
by the construction of $H_{(l)}$ ($l=1,2, \ldots$),
and the proof is complete.
\end{proof}
The {\em difference sequence} $\Delta \mathbf{a}=[\Delta a_0, \Delta a_1 , \Delta a_2 , \ldots]^T$
of a sequence $\mathbf{a}=[a_0, a_1 , a_2 , \ldots ]^T$ is defined by $\Delta a_i=a_{i+1}-a_i$ for each $i=0, 1, 2, \ldots$
Let $\Delta^k \mathbf{a}=[\Delta^k a_0, \Delta^k a_1 , \Delta^k a_2 , \ldots ]^T$ $(k=0, 1, 2, \ldots)$ be the
$k$th difference sequence defined inductively by
$\Delta^k \mathbf{a}=\Delta(\Delta^{k-1} \mathbf{a})$, where $\Delta^0 \mathbf{a}=\mathbf{a}$.
The infinite matrix
\[ \left[
\begin{array}{cccc}
a_0 & a_1 & a_2 & \cdots~\\
\Delta a_0 & \Delta a_1 & \Delta a_2 & \cdots~ \\
\Delta^2 a_0 & \Delta^2 a_1 & \Delta^2 a_2 & \cdots~ \\
\vdots & \vdots & \vdots & \ddots
\end{array}\right]
\]
is called the {\em difference matrix} of $\mathbf{a}$. It is well known \cite{Brualdi}
that for each $n=0,1,2, \ldots$,
$$ a_n= a_0 {{n}\choose{0}} + \Delta a_0 {{n}\choose{1}} + \Delta^2 a_0 {{n}\choose{2}}+
\cdots + \Delta^n a_0 {{n}\choose{n}},$$
which is used in the proof of the following lemma.
\begin{lem}
\label{lemma2}
Let $M=[m_{ij}^{\infty}]$ be the matrix with $m_{00}^{\infty}=1$, $m_{0j}^{\infty}=0$ and $m_{i0}^{\infty}=0$ for $i, j=1, 2, \ldots$, and
\[
m_{ij}^{\infty}=\left \{ \begin{array}{lr}
{{\lfloor {j\over 2}\rfloor}\choose{j-i}}, & {\rm if}~~ i \leq j, \\
0, & {\rm if}~~ i>j
\end{array} \right. (i, j=1, 2, \ldots).
\]
Then $M=N^{-1}$, where $N$ is the limit matrix in {\em Lemma \ref{lemma1}}.
\end{lem}
\begin{proof}
Let $n_{ij}^{\infty}$ denote the $(i, j)$ entry of $N$ ($i, j=0, 1, 2, \ldots$). We would like to show that $$\sum_{l=0}^{\infty}n_{il}^{\infty}m_{lj}^{\infty}=\delta_{ij},$$ the Kronecker delta.
For $i=0$ or $j=0$, there is nothing to show, so let $i\geq 1$ and $j\geq 1$. If $i=j$ resp. $i>j$,
then clearly
$$\sum_{l=0}^{\infty}n_{il}^{\infty}m_{lj}^{\infty}=\sum_{l=i}^{i}n_{il}^{\infty}m_{li}^{\infty}={{\lfloor {i-1\over 2}\rfloor}\choose{\lfloor {i-1\over 2}\rfloor}}{{\lfloor {i\over 2}\rfloor}\choose{0}}=1
\;\;\mbox{resp.}\;\;
\sum_{l=0}^{\infty}n_{il}^{\infty}m_{lj}^{\infty}=0.$$
So it is enough to show that if $j=i+r$ with $r > 0$, then $\sum_{l=0}^{\infty}n_{il}^{\infty}m_{lj}^{\infty}=0$.
Let $k={\lfloor {i-1\over 2}\rfloor}$ and consider the sequence
$\mathbf{z}=\left[{{k+r}\choose{k}}, - {{k+r-1} \choose {k}}, {{k+r-2}\choose{k}}, \ldots, (-1)^{r-1} {{k+1}\choose{k}}, (-1)^r {{k}\choose{k}}, 0, \ldots\right]^T$.
We can construct the difference matrix $A=[a_{ij}]$ having $\mathbf{z}$
as its first column as follows:
\[
A=\left[
\begin{array}{cccc|ccc|c}
~ {k+r}\choose{k} & {k+r-1}\choose{k-1} & \cdots & {r}\choose{0} & & & &\cdots~\\
~- {{k+r-1}\choose{k}}&-{{k+r-2}\choose{k-1}}&\cdots & -{{r-1}\choose{0}} & & & & \cdots~\\
~{k+r-2}\choose{k} & {k+r-3}\choose{k-1} & \cdots & {r-2}\choose{0}& & & & \cdots~\\
~\vdots& \vdots& \cdots & \vdots& & B & & \cdots~\\
~(-1)^{r-2} {{k+2}\choose{k}} & (-1)^{r-2} {{k+1}\choose{k-1}} & \cdots & (-1)^{r-2} {{2}\choose{0}} & & & &\cdots~\\
~(-1)^{r-1} {{k+1}\choose{k}} & (-1)^{r-1} {{k}\choose{k-1}} & \cdots & (-1)^{r-1} {{1}\choose{0}} & & & & \cdots~\\
~(-1)^r {{k}\choose{k}} & (-1)^r {{k-1}\choose{k-1}} & \cdots & (-1)^r {{0}\choose{0}}& & & & \cdots~\\\cline{1-8}
~0 & 0 & 0 & 0 & 0 &\cdots & 0 & \cdots~\\
~\vdots & \vdots & \vdots & \vdots &\vdots & \vdots & \vdots & \ddots~
\end{array}
\right],
\]
where $B$ is the $(r+1)\times (r+1)$ matrix given by
\[
B=\left[
\begin{array}{c|c|c|c|c|c|c|c|c}
~ 0 & 0 & 0 & 0 & \cdots & 0 & 0 & 0 & * ~\\\cline{8-8}
~ 0 & 0 & 0 & 0 & \cdots & 0 & 0 & * & * ~\\\cline{7-7}
~ \vdots&\vdots&\vdots&\vdots& \iddots & 0 & * & * & * ~\\\cline{6-6}
~ 0 & 0 & 0 & 0 & \iddots & * & * & * & * ~\\\cline{5-5}
~ 0 & 0 & 0 & 0 &\iddots & \vdots & \vdots & \vdots & \vdots ~\\\cline{4-4}
~ 0 & 0 & 0 & * & \cdots & * & * & * & * ~\\\cline{3-3}
~ 0 & 0 & * & * & \cdots & * & * & * & * ~\\\cline{2-2}
~0 & * & * & * & \cdots & * & * & * & * ~\\\cline{1-1}
* & * & * & * & \cdots & * & * & * & * ~
\end{array}
\right].
\]
Thus $a_{0j}=0$ for $j=k+1, \ldots, k+r$. Since $k+1 \leq \lfloor {i+r\over 2}\rfloor \leq \lfloor {i-1+r+r\over 2}\rfloor =k+r$, we obtain
\begin{equation*}
\begin{split}
\sum_{l=0}^{\infty}n_{il}^{\infty}m_{lj}^{\infty}&=\sum_{l=i}^{j}n_{il}^{\infty}m_{lj}^{\infty}=\sum_{l=i}^{i+r}(-1)^{l-i} {{k+l-i}\choose{k}} {{t}\choose{i+r-l}}\\
&=\sum_{s=0}^{t}(-1)^{r+s} {{k+r-s}\choose{k}} {{t}\choose{s}}=(-1)^r a_{0t}=0,
\end{split}
\end{equation*}
where $k=\lfloor {i-1\over 2}\rfloor$ and $t=\lfloor {i+r\over 2}\rfloor$, completing the proof.
\end{proof}
Let $M_{(t)}=U_{(1)} U_{(2)}\cdots U_{(t)}$, where $U_{(j)}=I_{2j-2}\bigoplus J(1)$ ($j=1, 2, \ldots, t$).
Then, by Lemmas \ref{lemma1} and \ref{lemma2}, we have $\lim_{t \to \infty}M_{(t)}=M$, which is the matrix in Lemma \ref{lemma2}.
In fact we have
\[
N= \left[ \begin{array}{ccccccccc}
~ 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots~\\
~0 & 1 &-1 & 1 & -1 & 1 & -1 & 1 & \cdots~\\
~0 & 0 & 1 & -1 & 1 & -1 & 1 & -1 & \cdots~\\
~0 & 0 & 0 & 1 & -2 & 3 & -4 & 5 & \cdots~\\
~0 & 0 & 0 & 0 & 1 & -2 & 3 & -4 & \cdots~\\
~0 & 0 & 0 & 0 & 0 & 1 & -3 & 6 & \cdots~\\
~0 & 0 & 0 & 0 & 0 & 0 & 1 & -3 & \cdots~\\
~0 & 0 & 0 & 0 & 0 &0 & 0 & 1 & \cdots~\\
~\vdots & \vdots & \vdots & \vdots &\vdots & \vdots & \vdots & \vdots & \ddots~
\end{array} \right]
\]
and
\[
M=N^{-1} = \left[ \begin{array}{ccccccccc}
~ 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots~\\
~0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & \cdots~\\
~0 & 0 & 1 & 1 & 1 & 0 & 0 & 0 & \cdots~\\
~0 & 0 & 0 & 1 & 2 & 1 & 1 & 0 & \cdots~\\
~0 & 0 & 0 & 0 & 1 & 2 & 3 & 1 & \cdots~\\
~0 & 0 & 0 & 0 & 0 & 1 & 3 & 3 & \cdots~\\
~0 & 0 & 0 & 0 & 0 & 0 & 1 & 3 & \cdots~\\
~0 & 0 & 0 & 0 & 0 & 0 & 0 & 1 & \cdots~\\
\vdots& \vdots&\vdots &\vdots&\vdots& \vdots &\vdots & \vdots& \ddots~
\end{array} \right].
\]
We can now state and prove the similarity transformations of $PD$ and $P^TD$ claimed
in (\ref{goal1}) and (\ref{goal2}).
\begin{thm}
\label{similarity}
Let $P=\left[ \begin{matrix}
{i}\choose{j}
\end{matrix} \right] (i, j=0, 1, \ldots)$ and $D={\rm diag}(1, -1, 1, -1, \ldots)$. Then,
\begin{itemize}
\item [\rm (a)] $N(P^TD)M=(P_1^TD_1)\bigoplus (P_1^TD_1)\bigoplus \cdots \bigoplus (P_1^TD_1)\bigoplus \cdots$,
\item [\rm (b)] $(DM^TD)\, (PD)\, (DN^TD)=(P_1D_1)\bigoplus (P_1D_1)\bigoplus \cdots \bigoplus (P_1D_1)\bigoplus \cdots$,
\end{itemize}
where
$P_1$ resp. $D_1$ are the leading $2\times 2$ submatrices of $P$ resp. $D$,
and $N$ resp. $M$ are the matrices in {\rm Lemmas} $3.1$ resp. $3.2$.
\end{thm}
\begin{proof}
(a) For each $j=1, 2, \ldots$, let $H_{(j)}=I_{2j-2}\bigoplus F$ and $U_{(j)}=I_{2j-2}\bigoplus J(1)$.
Let $m$ be an arbitrary positive integer with $n=2m+1$. First, we will show by
induction on $m$ that
\begin{equation*}
\left(H_{(m)}H_{(m-1)}\cdots H_{(1)}\right)P^TD \left(U_{(1)}U_{(2)}\cdots U_{(m)}\right)=Z \bigoplus (P^TD),
\end{equation*}
where $Z$ is an $n \times n$ matrix such that
$Z=\overbrace{(P^T_1D_1)\bigoplus (P^T_1D_1)\bigoplus \cdots \bigoplus (P^T_1D_1)}^{m}$.
When $m=1$, the first row of $H_{(1)}P^TDU_{(1)}$ is clearly
$[1, -1, 0, 0, \ldots]$.
Since for each $i$ and $j$ with $i, k=1, 2, \ldots$,
\[
(H_{(1)}P^TD)_{ik}=(-1)^k\biggl({{k}\choose{i}}-{{k}\choose{i+1}}+\cdots+(-1)^{k}{{k}\choose{k}}\biggr)
=(-1)^k{{k-1}\choose{i-1}},
\]
the second row of $H_{(1)}P^TDU_{(1)}$ is $[0, -1, 1, -1, 1, \ldots]U_{(1)}=[0, -1, 0, 0, 0, \ldots]$.
For each $i$ and $j$ with $i, j \geq 2$, we have
\[
(H_{(1)}P^TDU_{(1)})_{ij}=(-1)^{j-1}{{j-2}\choose{i-1}}+(-1)^j{{j-1}\choose{i-1}}=(-1)^{j-2}{{j-2}\choose{i-2}},
\]
and so $H_{(1)}P^TDU_{(1)}=(P^T_1D_1)\bigoplus (P^TD)$.
By induction on $m$, it follows
\begin{equation*}
N_{(m)}P^TDU_{(m)}=\overbrace{(P^T_1D_1)\bigoplus (P^T_1D_1)\bigoplus \cdots \bigoplus (P^T_1D_1)}^{m}\bigoplus (P^TD),
\end{equation*}
where $N_{(m)}=H_{(m)}H_{(m-1)}\cdots H_{(1)}$ and $M_{(m)}=U_{(1)}U_{(2)}\cdots U_{(m)}$. Since $m$ was an arbitrary positive integer,
\[
NP^TDM=({P^T_1}D_1)\bigoplus ({P^T_1}D_1)\bigoplus \cdots \bigoplus ({P^T_1}D_1) \bigoplus \cdots,
\]
where $N=\lim_{m \to \infty}N_{(m)}$ and $M=\lim_{m \to \infty}M_{(m)}$.\\
(b) It follows directly from (a) that
$$(DM^TD)PD(DN^TD)=(P_1D_1)\bigoplus (P_1D_1)\bigoplus \cdots \bigoplus (P_1D_1)\bigoplus \cdots,$$
completing the proof.
\end{proof}
Let $\mathbf e_i$ denote the $i$th column of the identity matrix $I$ ($i=0, 1, \ldots$).
Then $\mathbf{E}_{\lambda} (P^TD)$ and $\mathbf{E}_{\lambda} (PD)$ ($\lambda \in \{ 1, -1\}$)
can be characterized via Theorem \ref{similarity} as follows:
\begin{thm}
\label{bases}
Let $N$ resp. $M=N^{-1}$ be the matrices in {\rm Lemma \ref{lemma1} resp.
Lemma \ref{lemma2}}. Then the following hold:
\begin{itemize}
\item [\rm (a)] $\{M \mathbf e_0, M\mathbf e_2, M\mathbf e_4, \ldots\}$ is a basis for $\mathbf{E}_1(P^TD)$.
\item [\rm (b)] $\{M(\mathbf e_0+2 \mathbf e_1), M(\mathbf e_2+2\mathbf e_3), \ldots\}$
is a basis for $\mathbf{E}_{-1}(P^TD)$.
\item [\rm (c)] $\{DN^TD(2\mathbf e_0+\mathbf e_1), DN^TD(2\mathbf e_2+\mathbf e_3) \ldots\}$
is a basis for $\mathbf{E}_1(PD)$.
\item [\rm (d)] $\{DN^TD \mathbf e_1, DN^TD\mathbf e_3, DN^TD\mathbf e_5, \ldots\}$
is a basis for $\mathbf{E}_{-1}(PD)$.
\end{itemize}
\end{thm}
\begin{proof}
(a) Let $\mathbf B_{(1)}=\{\mathbf e_0, \mathbf e_2, \mathbf e_4, \ldots\}$
and $\mathbf x=[x_0, x_1, x_2, \ldots]^T \in \mathbf{E}_1(NP^TDM)$. Then,
by Theorem \ref{similarity}, we have
$$(NP^TDM-I)\mathbf x=\biggl( \bigoplus_{i=1}^{\infty} \left[ \begin{array}{cc}
0 & -1 \\
0 & -2
\end{array} \right] \biggr) \mathbf x=\mathbf 0,$$
which implies that $x_{i}=t_i, x_{i+1}=0$ for each $i=0, 2, 4, \ldots$ and $t_i \in \mathbb R$.
So $\mathbf B_{(1)}$ spans $\mathbf{E}_1(NP^TDM)$ and since
$\mathbf e_0, \mathbf e_2, \mathbf e_4, \ldots$ are linearly independent,
$\mathbf B_{(1)}$ is a basis for $\mathbf{E}_1(NP^TDM)$.
Therefore $\{M \mathbf e_0, M\mathbf e_2, M\mathbf e_4, \ldots\}$ is a basis for $\mathbf{E}_1(P^TD)$.\\
(b) Let $\mathbf B_{(-1)}=\{\mathbf e_0+2 \mathbf e_1, \mathbf e_2+ 2\mathbf e_3, \ldots\}$
and $\mathbf y=[y_0, y_1, y_2, \ldots]^T \in \mathbf{E}_{-1}(NP^TDM)$. Then,
by Theorem \ref{similarity}, we have
\[
(NP^TDM+I)\mathbf y=\biggl( \bigoplus_{i=1}^{\infty} \left[ \begin{array}{cc}
2 & -1 \\
0 & 0
\end{array}\right] \biggr) \mathbf y=\mathbf 0,
\]
which implies that $\left[ \begin{array}{c}
y_{i} \\
y_{i+1}
\end{array}\right] =s_i \left[ \begin{array}{c}
1 \\
2
\end{array}\right]$
for each $i=0, 2, 4, \ldots$ and $s_i \in \mathbb R$.
So $\mathbf B_{(-1)}$ spans $\mathbf{E}_1(NP^TDM)$ and since
$\mathbf e_0+2 \mathbf e_1, \mathbf e_2+ 2\mathbf e_3, \ldots$ are
linearly independent, $\mathbf B_{(-1)}$ is a basis for $\mathbf{E}_{-1}(NP^TDM)$.
Therefore $\{M(\mathbf e_0+2 \mathbf e_1), M(\mathbf e_2+2\mathbf e_3), \ldots\}$
is a basis for $\mathbf{E}_{-1}(P^TD)$. \\
Clauses (c) and (d) can be proven similarly.
\end{proof}
Consider now
\[
{P^T\hspace{0.01cm}^\downarrow}=\left[\begin{array}{ccccccccc}
~ 1 & 0 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots~\\
~0 & 1 & 0 & 0 & 0 & 0 & 0 & 0 & \cdots~\\
~0 & 1 & 1 & 0 & 0 & 0 & 0 & 0 & \cdots~\\
~0 & 0 & 2 & 1 & 0 & 0 & 0 & 0 & \cdots~\\
~0 & 0 & 1 & 3 & 1 & 0 & 0 & 0 & \cdots~\\
~0 & 0 & 0 & 3 & 4 & 1 & 0 & 0 & \cdots~\\
~0 & 0 & 0 & 1 & 6 & 5 & 1 & 0 & \cdots~\\
~0 & 0 & 0 & 0 & 4 &10 & 6 & 1 & \cdots~\\
~\vdots & \vdots & \vdots & \vdots &\vdots & \vdots & \vdots & \vdots & \ddots~
\end{array}\right].
\]
The following is a matrix expression of Theorem \ref{bases}; the last two
clauses appeared in \cite{Choi}.
\begin{cor}
\label{columns}
Let $Q=P+\left[\begin{array}{c|c}
1 & {\mathbf 0}^T\\\cline{1-2}
{\mathbf 0}& P
\end{array} \right]$
and $D={\rm diag}((-1)^0, (-1)^1, \ldots)$, where $P=\left[\begin{matrix}
{i}\choose{j}
\end{matrix} \right] (i, j=0, 1, 2, \ldots)$.
Then the following hold:
\begin{itemize}
\item [\rm (a)] The columns of $P^{T\downarrow}$ form a basis for $\mathbf{E}_1(P^TD)$.
\item [\rm (b)] The columns of $Q^{T\downarrow}(0|0)$ form a basis for $\mathbf{E}_{-1}(P^TD)$.
\item [\rm (c)] The columns of
$\left[\begin{array}{c}
{\mathbf 0}^T\\
{P\hspace{0.01cm}^\downarrow}
\end{array} \right]$
form a basis for $\mathbf{E}_{-1}(PD)$.
\item [\rm (d)] The columns of
${Q\hspace{0.01cm}^\downarrow}$ form a basis for $\mathbf{E}_1(PD)$.
\end{itemize}
\end{cor}
\begin{proof}
(a) Let $(i, j)$ be a pair of integers with $i, j\geq 0$.
The $i$th component of $M{\mathbf e}_{2j}$ equals
${\displaystyle j \choose {2j-i}}$ when $i \leq 2j$, and equals
$0$, otherwise; thus the $j$th column of $P^{T\downarrow}$ is
$$M{\mathbf e}_{2j}=(\overbrace{0, \ldots, 0}^j,{j \choose 0}, {j \choose 1}, \ldots, {j \choose j}, 0, 0, \ldots)^T.$$
\noindent
(b) The $i$th component of
$M({\mathbf e}_{2j}+2{\mathbf e}_{2j+1})$ equals
$$
{j \choose {2j-i}}+2{j \choose {2j-i+1}}={{j+1} \choose {2j-i+1}}+{j \choose {2j-i+1}},
$$
when $i\leq 2j+1$, and equals $0$ otherwise; thus the $j$th column of
$$ P^{T\downarrow}(0|0)+\left[\begin{array}{c|c}
1 & {\mathbf 0}^T\\\cline{1-2}
{\mathbf 0}& P^T
\end{array} \right]^{T\downarrow}(0|0)=Q^{T\downarrow}(0|0) $$ is
\begin{equation*}
\begin{split}
M({\mathbf e}_{2j}+2{\mathbf e}_{2j+1})&=[\overbrace{0, \ldots, 0}^j,{j+1 \choose 0}, {j+1 \choose 1}, \ldots, {j+1 \choose j+1}, 0, 0, \ldots]^T\\
&+[\overbrace{0, \ldots, 0}^{j+1},{j \choose 0}, {j \choose 1}, \ldots, {j \choose j}, 0, 0, \ldots]^T.
\end{split}
\end{equation*}
Using the fact that for each $i$ and $j$ with $i \geq j\geq 1$,
$$(DN^TD)_{ij}=(-1)^i(-1)^{i-j} {{\lfloor {{j-1}\over 2}\rfloor+i-j}\choose
{\lfloor {{j-1}\over 2}\rfloor}}(-1)^j,$$
clauses (c) and (d) can be proven similarly.
\end{proof}
\section{Invariant sequences of two kinds: Relations and examples}
\par
\setcounter{num}{4}
\setcounter{equation}{0}
We begin by some basic examples of (inverse) invariant sequences.
\begin{ex}
\label{fibomat}
{\rm
It follows from Corollary \ref{columns} (c) and (d) that
the Fibonacci sequence $\mathbf F$ is an invariant sequence of the first kind
and the Lucas sequence $\mathbf L$ is an inverse invariant sequence of the first kind.
The $i$th row of $P^{T\downarrow}$ ($i=0,1,\ldots$) is
\begin{equation}
\label{fibo}
[\overbrace{0, \ldots, 0}^{\lceil {{i}\over 2}\rceil}, {{\lceil {{i}\over 2}\rceil}\choose {\lfloor {{i}\over 2}\rfloor}},\ldots, {{i-2} \choose {2}}, {{i-1} \choose {1}}, {{i} \choose {0}}, 0, 0, \ldots],
\end{equation}
from which we can get that $J(0)\mathbf F$ is an invariant sequence of the
second kind and $J(0)\mathbf L$ is an inverse invariant sequence of the second kind.
Recall that $J(0)$ is the infinite Jordan block with $0$ in the main diagonal.
}
\end{ex}
In the following theorem, we provide a general mechanism for transforming
invariant into inverse invariant sequences, and vice versa.
\begin{thm}
\label{transform}
Let $J(\lambda)$ denote the infinite Jordan block with $\lambda$ in the main diagonal.
Then the following hold:
\begin{itemize}
\item [\rm (a)] If $\mathbf x$ is an invariant sequence of the second kind, then $(J(1)+J(0))^T{\mathbf x}$
is an inverse invariant sequence of the second kind.
\item [\rm (b)] If $\mathbf x$ is an inverse invariant sequence of the second kind,
then $J(2)^{-1}J(0) {\mathbf x}$ is an invariant sequence of the second kind.
\item [\rm (c)] If $\mathbf x$ is an invariant sequence of the first kind, then
$\left(-J(0)J(-2)^{-1}\right)^T {\mathbf x}$ is an inverse invariant sequence of the first kind.
\item [\rm (d)] If $\mathbf x$ is an inverse invariant sequence of the first kind,
then $(J(-1)+J(0)){\mathbf x}$ is an invariant sequence of the first kind.
\end{itemize}
\end{thm}
\begin{proof}
(a) If $\mathbf x$ is an invariant sequence of the second kind, then by
Corollary \ref{columns} (a), there exists $\mathbf b \in \mathbb{R}^\infty$
such that $P^{T\downarrow}\mathbf b =\mathbf x$. Let $L=[l_{ij}]$ ($i, j=0, 1, 2,\ldots$)
be the infinite lower triangular matrix defined by
$$ l_{ij}=\left \{ \begin{array}{lr}
(-1)^{i+j}, & {\rm if}~ i \geq j, \\
0, & {\rm if}~ i<j.
\end{array} \right.$$
Then $LP^{T\downarrow}(0|0)=P^{T\downarrow}$
because for each $i$ and $j$ with $i, j \geq 0$, the $(i, j)$ entry of $LP^{T\downarrow}(0|0)$ equals
$$ (-1)^{i+j}{j+1 \choose 0}+(-1)^{i+j+1}{j+1 \choose 1}+\cdots+(-1)^{i+j+i-j}{j+1 \choose {i-j}}={j \choose i-j}$$
when $i \geq j$, and equals $0$ otherwise, which coincides with the $(i, j)$ entry of $P^{T\downarrow}$.
Thus,
\begin{equation}
\label{2.2}
LQ^{T\downarrow}(0|0)=P^{T\downarrow}+L\left[\begin{array}{c}
{\mathbf 0}^T\\
P^{T\downarrow}
\end{array} \right].
\end{equation}
From the fact that $L^{-1}=J(1)^T$ and by Corollary \ref{columns} (b), it follows that
\begin{equation*}
Q^{T\downarrow}(0|0)\mathbf b=J(1)^TP^{T\downarrow}\mathbf b+\left[\begin{array}{c}
{\mathbf 0}^T\\
P^{T\downarrow}
\end{array} \right]\mathbf b=J(1)^T{\mathbf x}+[0, {\mathbf x}^T]^T=(J(1)+J(0))^T{\mathbf x}
\end{equation*}
is an inverse invariant sequence of the second kind.\\
(b) Let $\mathbf x$ be an inverse invariant sequence of the second kind.
From (\ref{2.2}) and $J(0)J(1)^T=J(1)$, it readily follows that
$$ J(2)^{-1}J(0)Q^{T\downarrow}(0|0)=P^{T\downarrow}$$
and so by Corollary \ref{columns} (b), we have that
$J(2)^{-1}J(0){\mathbf x}$ is an invariant sequence of the second kind. \\
(c) Since $Q^{\downarrow}=P^{\downarrow}+\left[\begin{array}{c|c}
1 & {\mathbf 0}^T\\\cline{1-2}
0& {\mathbf 0}^T\\\cline{1-2}
{\mathbf 0}& P^{\downarrow}
\end{array} \right]$, we have
\begin{equation*}
J(0)^TQ^{\downarrow}=J(0)^TP^{\downarrow}+\left[\begin{array}{c|c}
0& {\mathbf 0}^T\\\cline{1-2}
1 & {\mathbf 0}^T\\\cline{1-2}
0& {\mathbf 0}^T\\\cline{1-2}
{\mathbf 0}& P^{\downarrow}
\end{array} \right].
\end{equation*}
So $\Omega \left(
\left[\begin{array}{c}
{\mathbf 0}^T\\
Q^{\downarrow}
\end{array} \right]-
\left[\begin{array}{c}
{\mathbf 0}^T\\
P^{\downarrow}
\end{array} \right]\right)=\left[\begin{array}{c}
{\mathbf 0}^T\\
P^{\downarrow}
\end{array} \right]$, because
${i\choose i}+{i\choose i+1}+\cdots+ {i\choose i+k}={{i+1} \choose {i+k+1}}$ for $i, k=0, 1,\ldots$,
where $\Omega$ is the infinite $(0,1)$-matrix with $1$'s everywhere on and below its main diagonal.
Since $-J(-1)^T=\Omega^{-1}$, we get
$(I-J(-1)^T)^{-1}\left[\begin{array}{c}
{\mathbf 0}^T\\
Q^{\downarrow}
\end{array} \right]=\left[\begin{array}{c}
{\mathbf 0}^T\\
P^{\downarrow}
\end{array} \right]$, which implies that $\left(-J(0)J(-2)^{-1}\right)^T {\mathbf x}$ is an inverse
invariant sequence of the first kind. \\
Clauses (d) easily follows from (c) similarly.
\end{proof}
Let $\tau_1={{1+\sqrt{5}}\over 2}$ and $\tau_2={{1-\sqrt{5}}\over 2}$.
It is well known that $F_n={1 \over \sqrt{5}}\tau_1^{n}-{1 \over \sqrt{5}}\tau_2^{n}$
and $L_n=\tau_1^{n}+\tau_2^{n}$ where $F_n$ and $L_n$ are the $n$th terms
of ${\mathbf F}$ and ${\mathbf L}$, respectively ($n=0, 1, 2, \ldots$) \cite{Brualdi}.
Since $J(0){\mathbf F}$ is an invariant sequence of the second kind by (\ref{fibo}),
it follows from Theorem \ref{transform} (a) that
$$J(1)^TJ(0){\mathbf F}+J(0)^TJ(0){\mathbf F}=J(1)^TJ(0){\mathbf F}+{\mathbf F}$$
is an inverse invariant sequence of the second kind. In fact, the $n$th term ($n=0, 1, 2, \ldots$)
of $J(1)^TJ(0){\mathbf F}+{\mathbf F}$ is
$$ {1 \over \sqrt{5}}\tau_1^{n+2}-{1 \over \sqrt{5}}\tau_2^{n+2}+
{1 \over \sqrt{5}}\tau_1^{n}-{1 \over \sqrt{5}}\tau_2^{n}=L_{n+1}, $$
which implies that $J(1)^TJ(0){\mathbf F}+{\mathbf F}=J(0){\mathbf L}$.
On the other hand, since for each $i$ and $j$ with $i, j=0, 1, 2,\ldots$,
$$J(2)^{-1}=\left \{ \begin{array}{lr}
(-1)^{j-i}({1 \over 2})^{j-i+1}, & {\rm if}~ i \leq j, \\
0, & {\rm if}~ i>j,
\end{array} \right.$$
and the $n$-th term of $J(2)^{-1}J(0)^2{\mathbf L}$ is
$$
\sum_{k=0}^{\infty}(-1)^k(1/2)^{k+1}(\tau_1^{k+2+n}+\tau_2^{k+2+n})={1 \over \sqrt{5}}\tau_1^{n+1}-{1 \over \sqrt{5}}\tau_2^{n+1}=F_{n+1},$$
we have $J(2)^{-1}J(0)^2{\mathbf L}=J(0){\mathbf F}$, which is an invariant sequence of the second
kind by Theorem \ref{transform} (b), as stated in Example \ref{fibomat}.
It directly follows from Theorem \ref{transform} (d) that $(J(0)+J(-1)){\mathbf F}$,
namely $\,\mathbf L$, is an invariant sequence of the first kind.
Since for each $i$ and $j$ with $i, j=0, 1, 2,\ldots$,
$$-J(-2)^{-1}=\left \{ \begin{array}{lr}
({1 \over 2})^{j-i+1}, & {\rm if}~ i \leq j, \\
0, & {\rm if}~ i>j,
\end{array} \right.$$
we obtain $\left(-J(-2)^{-1}\right)^T [0,{\mathbf L}^T]^T=\mathbf F$, which
is an inverse invariant sequence of the first kind by Theorem \ref{transform} (c),
because for $n=0, 1, 2, \ldots$,
$$\sum_{k=0}^{n-1}(1/2)^{n-k}L_k=(1/2)^{n}~\sum_{k=0}^{n-1}((2\tau_1)^{k}+(2\tau_2)^{k})
={1 \over \sqrt{5}}\tau^{n}-{1 \over \sqrt{5}}\tau_2^{n}=F_n.$$
The sequence $\mathbf B=(B_0, B_1, \ldots)^T$ defined by
$B_0=1$ and $\sum_{k=0}^{n} {{n+1} \choose k}B_k=0~(n \geq 1)$ comprises the Bernoulli numbers and
$D \mathbf B$ is an invariant sequence of the first kind \cite{Sun},
which also follows directly from the fact that $PDD \mathbf B=D \mathbf B$.
A new inverse invariant sequence of the first kind from the Bernoulli numbers $\mathbf B$ is
provided next. See Table $1$ for explicit members of these sequences.
\begin{cor}
Let $\mathbf B=(B_0, B_1, \ldots)^T$ be the Bernoulli numbers.
Then the sequence $\mathbf K=(K_0, K_1, K_2, \ldots)^T$ defined by
\begin{equation*}
K_0=0,~ K_n=\sum_{k=0}^{n-1}(1/2)^{n-k}(-1)^kB_k \;~(n=1, 2, \ldots)
\end{equation*}
is an inverse invariant sequence of the first kind.
\end{cor}
\begin{proof}
It follows from Theorem \ref{transform} (c) that
$$ (-J(-2)^{-1})^TJ(0)^TD \mathbf B=(-J(-2)^{-1})^T[0,D{\mathbf B}^T]^T $$
is an inverse invariant sequence of the first kind.
Notice now that the sequence ${\mathbf K}$ in the statement is indeed equal to
$(-J(-2)^{-1})^T[0,D{\mathbf B}^T]^T. $
\end{proof}
\begin{table}[!ht]
\caption[]{\footnotesize (Inverse) invariant sequences of the first kind associated with the Bernoulli numbers.} \label{mytable}
\begin{tabular}{cccccccccccccccccc}
\hline
n&0&1&2&3&4&5&6&7&8&9&10&11&12&$\cdots$\\
\hline
$\mathbf B$&1&$-{1\over2}$&${1\over6}$&0&$-{1\over 30}$&0&${1\over 42}$&0&$-{1\over 30}$&0&${5\over 66}$&0&$-{691 \over 2730}$&$\cdots$\\
\hline
$D\mathbf B$&1&${1\over2}$&${1\over6}$&0&$-{1\over 30}$&0&${1\over 42}$&0&$-{1\over 30}$&0&${5\over 66}$&0&$-{691 \over 2730}$&$\cdots$\\
\hline
$\mathbf K$&0&${1 \over 2}$&${1\over2}$&${1 \over 3}$&${1 \over 6}$&${1 \over 15}$&${1 \over 30}$&${1 \over 35}$&${1 \over 70}$&$-{1 \over 105}$&$-{1 \over 210}$&${41 \over 1155}$&${41 \over 2310}$&$\cdots$\\
\hline
\end{tabular}
\end{table}
By Theorem \ref{transform} (d), we get $D \mathbf B=(J(0)+J(-1)){\mathbf K}$, since the first component of $(J(0)+J(-1)){\mathbf K}$
is clearly $(-1)^0B_0$, and for $i=1, 2, \ldots$, $i$th component of $(J(0)+J(-1)){\mathbf K}$ is
\begin{equation*}
-\sum_{k=0}^{i-1}(1/2)^{i-k}(-1)^kB_k+2\sum_{k=0}^{i}(1/2)^{i+1-k}(-1)^kB_k=(-1)^iB_i.
\end{equation*}
By Corollary \ref{columns} and Theorem \ref{transform}, we can directly get more (inverse) invariant sequences of the
first and second kind as follows:
\begin{cor}
\label{newcor1}
For a positive integer $n$, let
$\mathbf \Phi_n={\mathcal S}_{l_n}{\mathcal S}_{l_{n-1}} \ldots {\mathcal S}_{l_3}S_{l_2}{\mathcal S}_{l_1}$,
where
\begin{equation*}
{\mathcal S}_{l_i}=\begin{cases}
P^{T\downarrow},&\text{if $l_i=1$},\\
J(1)^T+J(0)^T, & \text{if $l_i=-1$},
\end{cases}
\end{equation*}
and let $\tilde{\mathbf \Phi}_n=\tilde{\mathcal S}_{l_n}\tilde{\mathcal S}_{l_{n-1}} \ldots \tilde{\mathcal S}_{l_3}\tilde{\mathcal S}_{l_2}\tilde{\mathcal S}_{l_1}$ where
\begin{equation*}
\tilde{\mathcal S}_{l_i}=\begin{cases}
J(2)^{-1}J(0),&\text{if $l_i=1$},\\
Q^{T\downarrow}(0|0), & \text{if $l_i=-1$}
\end{cases}
\end{equation*}
for $i=1, 2, \ldots, n$. Then for $\mathbf x \in \mathbb{R}^\infty$, we have the following:
\begin{itemize}
\item [\rm (a)] If $n$ is odd (even) and $l_i=(-1)^{i+1}$ for $i=1, 2, \ldots, n$, then $\mathbf \Phi_n{\mathbf x}$
is an invariant (inverse invariant) sequence of the second kind.
\item [\rm (b)] If $n$ is odd (even) and $l_i=(-1)^i$ for $i=1, 2, \ldots, n$, then $\tilde{\mathbf \Phi}_n{\mathbf x}$
is an inverse invariant (invariant) sequence of the second kind.
\end{itemize}
\end{cor}
\begin{ex}
{\rm
It follows from (\ref{fibo}) and Corollary \ref{newcor1} (a) that
for $\mathbf x=[x_0, x_1, x_2, \ldots]^T \in \mathbb{R}^\infty$, $\mathbf \Phi_2 \mathbf x=\mathbf y$
is an inverse invariant sequence of the second kind, where $\mathbf y=[y_0, y_1, y_2, \ldots]^T$ with $$y_i=\sum_{t=\lfloor {i \over 2}\rfloor}^{i+1}\left({\binom{t}{i-1-t}}+{\binom{t}{i+1-t}}\right)x_t \;\;\; (i=0, 1, \ldots). $$
For example, let $\mathbf x=[0, 0, 0, 0, 0, 0, 0, 1, 0, 0, \ldots]^T$.
This results to the inverse invariant sequence of the second kind $\mathbf y=[y_0, y_1, y_2, \ldots]^T$,
where
$$ y_i={\binom{7}{i-1-7}}+{\binom{7}{i+1-7}} \;\;\; (i=0, 1, 2, \ldots). $$
That is, $y_0=\cdots=y_5=0$, $y_{16}=y_{17}=\cdots=0$, and by direct calculation, one can compute the nonzero components $y_6, y_7, \ldots, y_{15}$;
e.g., $y_{10}={\binom{7}{10-1-7}}+{\binom{7}{10+1-7}}=56$.
}
\end{ex}
\begin{cor}
\label{newcor2}
For a positive integer $n$, let $\mathbf \Psi_n={\mathcal T}_{l_n}{\mathcal T}_{l_{n-1}} \ldots {\mathcal T}_{l_3}{\mathcal T}_{l_2}{\mathcal T}_{l_1}$,
where
\begin{equation*}
{\mathcal T}_{l_i}=\begin{cases}
Q^{\downarrow},&\text{if $l_i=1$},\\
\left(-J(-2)^{-1}\right)^TJ(0)^T, & \text{if $l_i=-1$}
\end{cases}
\end{equation*}
and let $\tilde{\mathbf \Psi}_n=\tilde{\mathcal T}_{l_n}\tilde{\mathcal T}_{l_{n-1}} \ldots \tilde{\mathcal T}_{l_3}\tilde{\mathcal T}_{l_2}\tilde{\mathcal T}_{l_1}$, where
\begin{equation*}
\tilde{\mathcal T}_{l_i}=\begin{cases}
J(0)+J(-1),&\text{if $l_i=1$},\\
\left[\begin{array}{c}
{\mathbf 0}^T\\
{P\hspace{0.01cm}^\downarrow}
\end{array} \right], & \text{if $l_i=-1$}
\end{cases}
\end{equation*}
for $i=1, 2, \ldots, n$. Then for $\mathbf x \in \mathbb{R}^\infty$, we have the following:
\begin{itemize}
\item [\rm (a)] If $n$ is odd (even) and $l_i=(-1)^{i+1}$ for $i=1, 2, \ldots, n$,
then $\mathbf \Psi_n{\mathbf x}$ is an invariant (inverse invariant) sequence of the first kind.
\item [\rm (b)] If $n$ is odd (even) and $l_i=(-1)^i$ for $i=1, 2, \ldots, n$,
then $\tilde{\mathbf \Psi}_n{\mathbf x}$ is an inverse invariant (invariant) sequence of the first kind.
\end{itemize}
\end{cor}
\begin{ex}
{\rm
For $i=1, 2, \ldots$, the $i$th row of $\left[\begin{array}{c}
{\mathbf 0}^T\\
{P\hspace{0.01cm}^\downarrow}
\end{array} \right]$ is
$[{\binom{i-1}{0}},{\binom{i-2}{1}}, \ldots, {{\lceil {{i-1}\over 2}\rceil}\choose
{\lfloor {{i-1}\over 2}\rfloor}}, 0, 0, \ldots].$
It follows that
for $\mathbf x=[x_0, x_1, x_2, \ldots]^T \in \mathbb{R}^\infty$, $\tilde{\mathbf \Psi}_2 \mathbf x=\mathbf y$
is an invariant sequence of the first kind, where by Corollary \ref{newcor2} (b), $\mathbf y=(y_0, y_1, y_2, \ldots)^T$ satisfies
$$y_i=\sum_{t=0}^{\lfloor {i \over 2}\rfloor}\left({\binom{i-1-t}{t-1}}+{\binom{i-t}{t}}\right)x_t.$$
For example, let $\mathbf x=[0, 0, 0, 0, 0, 0, 0, 1, 0, 0, \ldots]^T$. This results to the invariant sequence of
the first kind $\mathbf y=[y_0, y_1, y_2, \ldots]^T$, where
$$y_i={\binom{i-1-7}{6}}+{\binom{i-7}{7}} \;\;\;(i=1, 2, \ldots).$$
Thus $y_0=y_1=\cdots=y_{13}=0$ and one can find by direct calculation, the nonzero components $y_{14}, y_{15}, \ldots$; e.g., $y_{14}={\binom{6}{6}}+{\binom{7}{7}}=2$.
}
\end{ex}
In the next two theorems, we obtain direct relationships among (inverse) invariant sequences
of the first kind and (inverse) invariant sequences of the second kind.
\begin{thm}
\label{new2}
Let $\mathbf{x}$ and $\mathbf{y}$ be, respectively, either
\\
{\small
{\rm (i)} an invariant sequence of the first kind and an inverse invariant
sequence of the second kind,
\\
\hspace*{.25in} or
\\
{\rm (ii)} an inverse invariant sequence of the first kind and an invariant
sequence of the second kind.}\\
Then $\mathbf{x}^TD\mathbf{y}=0$.
\end{thm}
\begin{proof}
From $PD\mathbf{x}=\lambda\mathbf{x}$ and $P^TD\mathbf{y}=-\lambda\mathbf{y}$, we get $\mathbf{x}^TDP^TD\mathbf{y}=\lambda\mathbf{x}^TD\mathbf{y}$, which
implies that $2\lambda\mathbf{x}^TD\mathbf{y}=0$ for $\lambda \in \{ 1, -1\}$.
\end{proof}
The following lemma is a useful tool for proving the final theorem.
\begin{lem}
\label{new3}
For every positive integer $n$,
\begin{itemize}
\item [\rm (a)] the columns of $(P+D)^n$ are invariant sequences of the first kind, and the columns of $(P-D)^n$ are
inverse invariant sequences of the first kind.
\item [\rm (b)] the columns of $(P^T+D)^n$ are invariant sequences of the second kind, and the columns of $(P^T-D)^n$ are
inverse invariant sequences of the second kind.
\end{itemize}
\end{lem}
\begin{proof}
Let $n$ be a positive integer. Then $$PD(P+D)^n=PD(P+D)(P+D)^{n-1}=(P+D)^n$$ and $$P^TD(P^T-D)^n=P^TD(P^T-D)(P^T-D)^{n-1}=-(P^T-D)^n,$$
which respectively imply that each column of $(P+D)^n$ is an invariant sequence of the first kind,
and each column of $(P^T-D)^n$ is an inverse invariant sequence of the second kind. The other assertions of the theorem
follow similarly.
\end{proof}
\vskip 0.5mm
The following theorem is a form of converse of Theorem \ref{new2}.
\begin{thm}
\label{new4}
Let ${\mathbf x_i}\in \mathbb{R}^\infty~(i=0, 1, 2, \ldots)$, ${\mathbf y} \in \mathbb{R}^\infty\setminus\{{\mathbf 0}\}$,
and let $X=[\mathbf x_0, \mathbf x_1, \mathbf x_2, \ldots]$. Then for each $i=0, 1, 2, \ldots$,
\begin{itemize}
\item [\rm (a)] if ${\mathbf x_i}^TD{\mathbf y}=0$ and $X=P+D~(X=P-D)$, then ${\mathbf y}$ is an inverse invariant (invariant)
sequence of the second kind, and ${\mathbf x_i}$ is an invariant (inverse invariant) sequence of the first kind.
\item [\rm (b)] if ${\mathbf x_i}^TD{\mathbf y}=0$ and $X=P^T+D~(X=P^T-D)$, then ${\mathbf y}$ is an inverse invariant (invariant) sequence
of the first kind and ${\mathbf x_i}$ is an inverse invariant (invariant) sequence of the second kind.
\end{itemize}
\end{thm}
\begin{proof}
If ${\mathbf x_i}^TD{\mathbf y}=0$ and $X=P+D$, then since $X^TD \mathbf{y}=(P^T+D)D\mathbf{y}=\mathbf 0$, we get $P^TD\mathbf{y}=-\mathbf{y}$ and by Lemma \ref{new3}, $PD \mathbf{x}_i=\mathbf{x}_i$. So ${\mathbf y}$ is an inverse invariant sequence of the second kind and ${\mathbf x_i}$ is an invariant sequence of the first kind.
This proves the first case of part (a). For the case of $X=P-D$ and part (b), the results can be shown similarly.
\end{proof}
\vskip 0.5cm
\section*{References}
| 21f65ec4d2252233156832d3d9969c7b8411f348 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Introduction}
The quantum anomalous Hall effect (QAHE) allows the resistance quantization and dissipationless edge states seen in the quantum Hall effect \cite{vonKlitzing1980}, but without the need for an applied magnetic field. The topologically-protected edge states of the QAHE are also chiral and spin-polarised \cite{Bernevig2006,Yu2010}, acting as a natural spin filter. Additional applications are starting to be explored, for example in spintronics as a pure spin current source or detector \cite{Gotte2016}, and in topological quantum computing \cite{HasanKane2010,Qi2010}.
In experimental devices, the QAHE is remarkably robust, and has been observed in chromium- and vanadium-doped (Bi,~Sb)$_2$Te$_3$ across a range of film thicknesses, and grown on multiple substrates \cite{Chang2013,Checkelsky2014,Grauer2015,Kou2014,Bestwick2015}. Quantization has been improved by reducing film thickness and doping with vanadium to suppress dissipative channels \cite{Chang2015a}, but whether these variables affect the crystal structure and electronic band structure has largely been overlooked experimentally. To date, the \emph{a}-axis parameter has only been determined for Cr-doped (Bi,~Sb)$_2$Te$_3$ grown on SrTiO$_3$~(1~1~1) with or without a Te capping layer \cite{Park2015}, with no systematic study of the role of thickness or doping level. This is despite the potential for uniaxial or biaxial strain in the films to either drive these materials into the topologically trivial (non-QAH) regime by altering the band structure \cite{Aramberri2016}, or affect the fabrication and performance of ferromagnetic/non-ferromagnetic topological insulator heterostructure devices due to lattice mismatch \cite{Yasuda2016,He2016,Mogi2017}.
Here, we conduct a comprehensive study of the effect of thickness, vanadium-doping and substrate choice on the crystal structure of MBE-grown (Bi,~Sb)$_{2-x}$V$_x$Te$_3$ thin films, and compare to existing results in the literature. We use high-resolution X-ray diffractometry (HRXRD) to determine the in- and out-of-plane lattice parameters, orientation relationships and epitaxial quality of the film, substrate, and Te capping layer. The HRXRD data is supported by information about the local structure and doping, provided by scanning transmission electron microscopy (STEM) and energy-dispersive X-ray spectroscopy (EDX).
\section*{Results}
\subsection*{Crystallinity and elemental composition}
We first demonstrate the quality of the films with symmetrical high-resolution X-ray diffraction measurements, probing planes parallel to the substrate. Figure \ref{epitaxy}a shows a 2$\theta$/$\omega$ scan of a 4~QL (quintuple layer, each QL~$\approx$1~nm) film of (Bi,~Sb)$_{1.89}$V$_{0.11}$Te$_3$ on a SrTiO$_3$~(1~1~1) substrate, with a 10~nm Te capping layer. We observe (Bi,~Sb)$_{2-x}$V$_{x}$Te$_3$ peaks at (0 0 3$n$), as expected for its space group $R\bar{3}m$, and tellurium peaks at ($m$ 0 0), echoing previous results on Cr-doped (Bi,~Sb)$_2$Te$_3$ \cite{Chang2015a,Park2015}.
\begin{figure}[ht]
\centering
\includegraphics[width=12.75cm]{FMTI_XRD_Set2_Fig1_v1}
\caption{High-resolution X-ray diffractometry of vanadium-doped (Bi,~Sb)$_2$Te$_3$ thin films on SrTiO$_3$~(1~1~1). \textbf{a} Indexed 2$\theta$/$\omega$ scan of a 4 QL film (quintuple layer; 1 QL $\approx$ 1 nm) with 10~nm tellurium capping layer. (Bi,~Sb)$_{2-x}$V$_x$Te$_3$ peaks are unnamed. \textbf{b-d} Rocking curves of (0 0 6), Te (1 0 0) and (0 0 15) peaks (shown in \textbf{a} in blue, red and dark blue, respectively). \textbf{e} 2$\theta$/$\omega$ scans of 4 (black), 10 (red) and 20 QL (blue) films (10 and 20 QL data offset for clarity). Grey dashed lines are guides to the eye, showing the contraction of the lattice as thickness increases.}
\label{epitaxy}
\end{figure}
Both layers grow epitaxially, as shown by the rocking curves in Figure \ref{epitaxy}b-\ref{epitaxy}d (rocking curves taken on the (Bi,~Sb)$_{2-x}$V$_{x}$Te$_3$ (0~0~6), Te (1 0 0) and (Bi,~Sb)$_{2-x}$V$_{x}$Te$_3$ (0 0 15) peaks, respectively). The (Bi,~Sb)$_{1.89}$V$_{0.11}$Te$_3$ (Figure \ref{epitaxy}b and \ref{epitaxy}d) has a smaller full-width at half-maximum (FWHM, Gaussian fit) than tellurium: $\sim$0.11$^{\circ}$ as opposed to 0.65$^{\circ}$. The (Bi,~Sb)$_{2-x}$V$_x$Te$_3$ (0~0~6) curve shows dual peaks $\sim$0.15$^{\circ}$ from the expected value of $\omega$, consistent with the observation of twinned crystal domains in Cr-doped films \cite{Richardella2015}.
These results hold for all samples grown on SrTiO$_3$~(1~1~1), even with changes to doping and thickness. Figure \ref{epitaxy}e shows 2$\theta$/$\omega$ scans of samples with $x$~=~0.06-0.07 and thicknesses of 4, 10 and 20~QLs (black, red and blue, respectively). The plots are offset for clarity. The changes in doping and thickness do not affect the relative intensities of the peaks. The (Bi, Sb)$_{2-x}$V$_{x}$Te$_3$ \emph{c}-axis appears to shorten as the thickness increases; this is clearly seen in the 2$\theta$ position of the (0 0 24) peak as it changes relative to the Te~(3 0 0) peak, which in turn stays constant with respect to SrTiO$_3$ (1 1 1). However, averages of Gaussian fits to 3 reflections show that 4 and 10~QL films have approximately equal thickness (30.51 $\pm$ 0.05 \AA and 30.54 $\pm$ 0.04 \AA respectively, compared to \emph{c}~=~30.44 $\pm$ 0.02 \AA for the 20~QL film).
HRXRD measurements of 10~QL films grown on Al$_2$O$_3$~(0~0~0~1) and Si~(1~1~1) show very similar epitaxial growth and out-of-plane lattice parameters (See Supplementary Figure S1). The (Bi,~Sb)$_{2-x}$V$_{x}$Te$_3$ lattice constant \emph{c}~=~30.47~$\pm$~0.08~{\AA} and 30.45~$\pm$~0.07~\AA, respectively, a slight decrease compared to 30.54~$\pm$~0.03~{\AA} on SrTiO$_3$~(1~1~1) (10~QL film). On an Al$_2$O$_3$~(0~0~0~1) substrate, rocking curves on (Bi,~Sb)$_{2-x}$V$_{x}$Te$_3$ (0 0 15) and Te~(1~0~0) have a FWHM of 0.12$^{\circ}$ and 0.77$^{\circ}$, respectively, similar to the SrTiO$_3$~(1~1~1) samples. Growth appears to be more disordered on Si~(1~1~1), where the (Bi, Sb)$_{2-x}$V$_{x}$Te$_3$ (0 0 15) rocking curve has a larger FWHM of $\sim$0.36$^{\circ}$.
We now investigate the local crystallinity and interfaces using scanning transmission electron microscopy (STEM). The High Angle Annular Dark Field (HAADF) signal, generating the images shown in Figure~\ref{TEM}, is proportional to the local thickness and atomic number. For a relatively homogeneous thickness, such as a TEM lamella, the brightness is proportional to the average Z for a given pixel such that brighter atomic columns correspond to heavier atoms.
\begin{figure}[ht]
\centering
\includegraphics[width=12.75cm]{FMTI_XRD_Set2_Fig1B_v1}
\caption{Scanning transmission electron microscopy (STEM) and elemental analysis of (Bi,~Sb)$_{2-x}$V$_{x}$Te$_3$ on SrTiO$_3$. \textbf{a} STEM cross-sectional view of the film. \textbf{b} HRSTEM view of the substrate-film interface. \textbf{c} HAADF reference image and EDX elemental maps.}
\label{TEM}
\end{figure}
We observe regular growth of the QLs, with homogeneous thickness of $\sim$1~nm for each layer (Figure~\ref{TEM}a and \ref{TEM}b). The Te capping layer is crystalline and appears to grow at an angle of $\sim$7$^{\circ}$ to the underlying quintuple layers. However, detailed HRXRD indicates that this is not the case (see Figure~\ref{inplane} and discussion). While above the first QL we observe good epitaxial growth, the interface between STO and the QLs is relatively irregular, with the first QL being discontinuous and heavy atomic species coexisting with the lighter substrate. The SrTiO$_3$ substrate also displays a change in contrast over a thickness of $\sim$1.5~nm at the substrate surface, where a lower HAADF signal suggests that only light elements are present.
STEM-EDX (energy-dispersive X-ray spectroscopy) elemental maps are reported in Figure~\ref{TEM}c (non-negative matrix factorisation (NMF) of the EDX data is shown in Supplementary Figure S2). The Pt signal originates from the protective layer deposited during focused ion beam (FIB) sample preparation. The QL region contains strong signals from Bi, Sb, V and Te, which are all homogeneous throughout the device thickness. Tellurium extends above the active region into the capping layer. As expected, Sr, Ti and O signals dominate the substrate region. However, Ti and O extend further towards the (Bi,~Sb)$_{2-x}$V$_{x}$Te$_3$ compared to Sr, indicating that the interfacial region that appears darker in the STEM-HAADF images has a lower Sr concentration than the bulk SrTiO$_3$.
\begin{figure}[ht]
\centering
\includegraphics[width=12.75cm]{FMTI_XRD_Set2_Fig2_v1}
\caption{In-plane orientation of layers. \textbf{a-c} Reciprocal space maps of the (1 0 20) and Te (2 1 1) peaks of a 10 QL, x~=~0.07 film, at azimuthal angles $\phi$~=~60$^{\circ}$, 120$^{\circ}$ and 180$^{\circ}$ respectively. Grey dashed lines indicate fitted peak value from data in Figure \ref{thickness}\textbf{b}. Film peaks measured with respect to SrTiO$_3$~(1~1~2). \textbf{d} Azimuthal angle scans of substrate (grey, STO (1 1 2): 2$\theta$~=~57.84$^{\circ}$, offset~=~19.50$^{\circ}$), topological insulator (blue, (1 0 20): 2$\theta$~=~66.10$^{\circ}$, offset~=~23.34$^{\circ}$) and capping layer (red, Te (2 1 1): 2$\theta$~=~65.56$^{\circ}$, offset~=~24.78$^{\circ}$) peaks, offset for clarity. The substrate and topological insulator peaks have the expected 3- and 6-fold symmetries, whereas Te (2 1 1) has 6-fold rather than 2-fold symmetry.}
\label{inplane}
\end{figure}
Having confirmed that the films are epitaxial and homogeneously doped, we use reciprocal space mapping of asymmetrical peaks (those whose corresponding atomic planes are not parallel to the substrate) to find the in-plane orientations of the substrate, ferromagnetic topological insulator, and capping layer. Figure \ref{inplane}a-\ref{inplane}c shows reciprocal space maps of a 10~QL, x~=~0.07 sample, where $S_x$ is the reciprocal of the in-plane d-spacing, and $S_z$ is the reciprocal of the out-of-plane d-spacing. The three maps show equivalent areas of reciprocal space at three different azimuthal angles. The maps include the (Bi,~Sb)$_{2-x}$V$_{x}$Te$_3$~(1 0 20) and Te~(2 1 1) peaks (right- and left- hand side of the panels, respectively). Figure \ref{inplane}a also shows (Bi,~Sb)$_{2-x}$V$_{x}$Te$_3$~(1 0 19) directly below (1 0 20).
Whilst the higher-intensity (1 0 20) peak has the expected six-fold symmetry of a (Bi,~Sb)$_2$Te$_3$-based compound, the appearance of the Te~(2 1 1) peak at all three angles indicates that the capping layer is polycrystalline. Azimuthal angle scans confirm these findings (Figure \ref{inplane}d), and indicate the epitaxial relationship (Bi, Sb)$_{2-x}$V$_{x}$Te$_3$~[0~1~0]~//~SrTiO$_3$~[1~1~$\bar{2}$].
Note that one of the Te peaks was not captured in this scan due to slight sample misalignment, and that the broader peaks visible on the thin-film scans are due to sample geometry, confirmed by repeating the measurement away from the peaks.
We also observe a Te~(2 1 0) peak at $\Delta\phi \approx$~53$^{\circ}$ from Te~(2 1 1), confirming that the $a$- and $c$-axes of the tellurium cap are in-plane (see Supplementary Figure~S3), and an epitaxial relationship of (Bi,~Sb)$_{2-x}$V$_{x}$Te$_3$~[1~0~0]~//~Te~[$\bar{1}$~2~2]. Since the crystallographic axes are in-plane, we conclude that the $\sim$7.4$^{\circ}$ tilt observed with STEM is not a tilt of the unit cell itself, but rather an alignment of Te atoms revealed by an off-axis cut.
\subsection*{Thickness, doping and substrate dependence of lattice parameters}
To determine whether the unit cell of (Bi, Sb)$_{2-x}$V$_{x}$Te$_3$ changes with doping level or film thickness, we repeat the reciprocal-space measurements detailed above. A longer counting time is used in order to precisely determine the in- and out-of-plane lattice parameters. Figure \ref{doping} is a comparison of 4~QL films with $x$~=~0.06 and 0.11, and Figure \ref{thickness} shows reciprocal space maps for 4, 10 and 20~QL films on SrTiO$_3$, and 10~QL films on Al$_2$O$_3$~(0~0~0~1) and Si~(1~1~1). Table \ref{table} summarises the calculated lattice parameters, along with previous HRXRD results from the literature for comparison.
\begin{figure}[ht]
\centering
\includegraphics[width=8.5cm]{FMTI_XRD_Set2_Fig3_v1}
\caption{Effect of vanadium doping on the unit cell. Reciprocal space maps of the (1 0 20) and Te (2 1 1) (right and left peaks in panels, respectively) in films with \textbf{a} x~=~0.06 and \textbf{b} x~=~0.11, where (Bi,~Sb)$_{2-x}$V$_x$Te$_3$. \textbf{c} 2$\theta$/$\omega$ scans (offset for clarity) of the same films.}
\label{doping}
\end{figure}
Reciprocal-space maps (Figure \ref{doping}a and b, x~=~0.06 and 0.11, respectively) and 2$\theta$/$\omega$ scans (Figure \ref{doping}c, $x$~=~0.11 data (grey) offset for clarity) indicate that the $c$-axis lengthens as the doping is increased.
From the asymmetrical peaks (Figure \ref{doping}a and \ref{doping}b), $c$~=~30.55~$\pm$~0.42~{\AA} for $x$~=~0.06 and 30.50~$\pm$~0.42~{\AA} for $x$~=~0.11, whereas from the symmetrical peak measurements (Figure\ \ref{doping}c), $c$~=~30.51~$\pm$~0.05~{\AA} and 30.62~$\pm$~0.06~{\AA}. For both doping levels, \emph{c} is still close to the bulk value of 30.60~\AA.
The in-plane lattice parameter is also close to the bulk value ($a$~=~4.30~\AA). For the $x$~=~0.06 and 0.11 films, the $a$-axis parameter is 4.28~$\pm$~0.14~{\AA} or 4.26~$\pm$~0.15~\AA, respectively, and matches to the effective in-plane spacing of the substrate (3.565~\AA) with a 30$^{\circ}$ rotation ($a^{TI}\cos 30^{\circ}$~=~3.71~\AA). The slight discrepancy may be due to the Sr-deficient interfacial layer shown in Figure~\ref{TEM}, which appears to have a larger lattice than the bulk of the substrate.
\begin{figure}[ht]
\centering
\includegraphics[width=12.75cm]{FMTI_XRD_Set2_Fig4_v2}
\caption{Unit cell parameters as a function of thickness and substrate choice. Reciprocal space maps showing the (Bi,~Sb)$_{2-x}$V$_x$Te$_3$ (1 0 20) and Te (2 1 1) peaks, from films grown on SrTiO$_3$ (1~1~1) (\textbf{a-c}: 4, 10 and 20~QL, respectively), Al$_2$O$_3$ (0~0~0~1) (\textbf{d}: 10~QL), and Si (1~1~1) (\textbf{e}: 10~QL). Film peaks measured with respect to SrTiO$_3$~(1~1~2), Al$_3$O$_3$~(0 1 -1 8) or Si~(3~3~1). Panel \textbf{a} shows the same dataset as Figure \ref{doping}\textbf{a}.}
\label{thickness}
\end{figure}
The in-plane parameter also remains constant as a function of thickness (Figure \ref{thickness}a-\ref{thickness}c): the 10 and 20~QL samples ($x$~=~0.07, Figure \ref{thickness}b and \ref{thickness}c) have $a$~=~4.28~$\pm$~0.16~{\AA} and 4.30~$\pm$~0.16~\AA. Out-of-plane, the thickest (20~QL) film has a smaller $c$, at 30.34~$\pm$~0.41~\AA. The 10 and 4~QL films have similar unit cells, with $c$~=~30.54~$\pm$~0.43~{\AA} and 30.55~$\pm$~0.42~\AA.
Figure~\ref{thickness}b, \ref{thickness}d and \ref{thickness}e shows data taken for 10~QL films grown on SrTiO$_3$ (1 1 1), Al$_2$O$_3$ (0~0~0~1) and Si (1 1 1), respectively. There is very little difference between the in-plane parameters on the three substrates: $a$~=~4.28~$\pm$~0.16, 4.29~$\pm$~0.04 and 4.30~$\pm$~0.20~\AA, respectively. Out-of-plane, the film grown on Si~(1~1~1) appears to have a shorter unit cell, 30.22~$\pm$~0.41~{\AA} as opposed to 30.49~$\pm$~0.32~{\AA} on Al$_2$O$_3$ and 30.54~$\pm$~0.43~{\AA} on SrTiO$_3$. The Te~(2~1~1) peak was not observed for the Si~(1~1~1) sample; based on the relative peak breadths and intensities of symmetric 2$\theta$/$\omega$ scans, the Te capping layer was too thin (2--5~nm) for the peak to be detected (see Supplementary Figure~S1). The peaks measured on Al$_2$O$_3$ (0~0~0~1) and Si (1~1~1) were all less intense and broader than for the equivalent film grown on SrTiO$_3$~(1~1~1), which indicates less well-defined crystallographic orientation in-plane. Whilst the thickness of the films contributes to the breadth and low intensity of the peaks, Scherrer fits to the symmetrical data show that the nominally 10~QL films on SrTiO$_3$ (1 1 1), Al$_2$O$_3$ (0~0~0~1) and Si (1~1~1) are 9.9~$\pm$~1.9~nm, 10.8~$\pm$~0.1~nm and 7.2~$\pm$~0.8~nm, respectively, which does not correlate to the observed difference in intensity.
\section*{Discussion}
Our results, and those from previous HRXRD studies of ferromagnetic (Bi,~Sb)$_2$Te$_3$, are presented in Table~\ref{table}. Figure \ref{schematic} is a schematic summarizing the results for 10~QL of (Bi,~Sb)$_{2-x}$V$_x$Te$_3$ on SrTiO$_3$, with a Te capping layer.
Surprisingly, measurements of the in-plane lattice constant of these quantum anomalous Hall insulators have only been made for Cr-doped films grown on SrTiO$_3$ (see Table~\ref{table}), even though the QAHE has also been observed in (Bi, Sb)$_{2-x}$Cr$_x$Te$_3$ on InP~(1~1~1) \cite{Checkelsky2012,Checkelsky2014}, Si~(1~1~1) \cite{Grauer2015,Peixoto2016} and GaAs~(1~1~1) \cite{Kou2014,Kou2015}, and grown on Al$_2$O$_3$~(0~0~0~1) \cite{Chang2015a}. Vanadium-doped films have not previously been characterised in-plane.
\begin{figure}[ht]
\centering
\includegraphics[width=8.5cm]{FMTI_XRD_Set2_Fig5_v3}
\caption{Schematic of SrTiO$_3$/(Bi,~Sb)$_{2-x}$V$_x$Te$_3$/Te. \textbf{a} View along the (Bi,~Sb)$_{2-x}$V$_x$Te$_3$ ($\bar{1}$~1~0) direction (Te atoms = dark blue, Sb = dark green, Bi = light green, V = red, Sr = grey, Ti = dark red, O = dark grey). Parameters used are those of the 10~QL film. \textbf{b} Unit cells and experimental lattice parameters of the substrate and films, demonstrating in-plane lattice matching.}
\label{schematic}
\end{figure}
Although we do not observe significant differences in the (Bi,~Sb)$_{2-x}$V$_x$Te$_3$ unit cell between various doping levels, thicknesses or substrates, there are small trends in \emph{c}, whereas \emph{a} remains unchanged across all the samples measured. The latter indicates that (Bi, Sb)$_{2}$Te$_3$ does not strongly match to the substrate, perhaps due to weak, Van der Waals bonding between quintuple layers and the substrate. Doping with vanadium or chromium slightly lengthens the unit cell along the \emph{c}-axis, though asymmetrical measurements show no difference between our V-doped samples.
Where data is available for comparison, 20~QL films have a shorter unit cell than 4--10~QL films. This could be due to changing conditions during growth, or perhaps a gradient in the doping after around 10~QL (up to at least 10~QL, V-doping is homogeneous, as shown by EDX). Finally, the unit cells of films grown on Al$_2$O$_3$~(0~0~0~1) and SrTiO$_3$~(1~1~1) are almost identical, but the \emph{c}-axis on Si~(1~1~1) is consistently smaller. This may be because Si~(1~1~1) is not as close a match to the (Bi,~Sb)$_{2-x}$V$_x$Te$_3$ lattice as Al$_2$O$_3$~(0~0~0~1) and SrTiO$_3$~(1~1~1)---rocking curves on the latter samples also show far less disorder (see Supplementary Figure~S1). Although small changes in the unit cell within the parameters used here indicate that strain engineering is not feasible, the lack of biaxial strain in particular is important for growth of doped/undoped heterostructures such as those in Reference~\cite{Mogi2017}.
\begin{table}[ht] \centering \begin{threeparttable}
\begin{tabular}{llclll} \toprule
Substrate & Thickness & Doping & c (\AA, $2\theta$/$\omega$) & c (\AA, RSM) & a (\AA, RSM)\\ \midrule
Bulk & & & & 30.60 & 4.30 \\[10pt]
STO (1~1~1) & 4 QL & V$_{0.06}$ & 30.51 $\pm$ 0.05 & 30.55 $\pm$ 0.42 & 4.28 $\pm$ 0.14 \\
& & V$_{0.11}$ & 30.62 $\pm$ 0.06 & 30.50 $\pm$ 0.42 & 4.26 $\pm$ 0.15 \\[5pt]
& 5 QL\tnote{a} & Cr$_{0.15}$\tnote{*} & 30.62 $\pm$ 0.06 & & 4.28 \\
& & Cr$_{0.15}$ & 30.66 $\pm$ 0.19 & & 4.28 \\[5pt]
& 10 QL & V$_{0.07}$ & 30.54 $\pm$ 0.03 & 30.54 $\pm$ 0.43 & 4.28 $\pm$ 0.16 \\[5pt]
& 20 QL & V$_{0.07}$ & 30.44 $\pm$ 0.01 & 30.34 $\pm$ 0.41 & 4.30 $\pm$ 0.16 \\[10pt]
Al$_2$O$_3$ (0~0~0~1) & 10 QL & V$_{0.07}$ & 30.47 $\pm$ 0.08 & 30.49 $\pm$ 0.32 & 4.29 $\pm$ 0.13 \\[5pt]
& 20 QL\tnote{b} & V$_{0.15}$ & 30.39 & & \\
& & Cr$_{0.16}$ & 30.15 & & \\
& & None & 30.34 & & \\[10pt]
Si (1~1~1) & 10 QL & V$_{0.07}$ & 30.45 $\pm$ 0.07 & 30.22 $\pm$ 0.41 & 4.30 $\pm$ 0.20 \\[10pt]
InP (1~1~1) & 8 QL\tnote{c} & Cr$_{0.22}$ & 30.26 & & \\ \bottomrule
\end{tabular}
\begin{tablenotes}[para,flushleft]
\item[*] Uncapped
\item[a] J. Park \emph{et al.} \cite{Park2015}
\item[b] C.-Z. Chang \emph{et al.} \cite{Chang2015a}
\item[c] J. Checkelsky \emph{et al.} \cite{Checkelsky2014}
\end{tablenotes}
\caption{\label{table} Unit cell parameters of Cr- and V-doped (Bi,~Sb)$_2$Te$_3$ calculated from HRXRD measurements of samples with various substrates, thicknesses and doping levels. Parameters from this study are calculated from three (0~0~\emph{n}) reflections or a single (1~0~20) reflection.}
\end{threeparttable} \end{table}
We also find that the crystal structure of the V-doped films is unchanged compared to Cr-doped (Bi,~Sb)$_2$Te$_3$ \cite{Park2015}, and that the Te capping layer is epitaxial and grows in three equivalent orientations on (Bi,~Sb)$_2$Te$_3$. For all our samples, we observe the epitaxial relationship (Bi,~Sb)$_2$Te$_3$~[1~0~0]~//~Te~[$\bar{1}$~2~2]. The slight out-of-plane disorder in the Te capping layer (observed as a larger FWHM of the rocking curve than (Bi,~Sb)$_{2-x}$V$_x$Te$_3$) is probably due to dislocations at monocrystalline domain boundaries or island formation on the top of the topological insulator during growth (as observed in STEM measurements).
In summary, we have investigated the crystal structure of the ferromagnetic topological insulator (Bi,~Sb)$_{2-x}$V$_x$Te$_3$ as a function of doping level, thickness and substrate, using high-resolution X-ray diffractometry supported by STEM and EDX. Focusing on the range commonly used in devices, we find that the unit cell is largely unaffected by vanadium doping changes of $\sim$2~at.\%, and remains unchanged over a thickness range of 4--10 quintuple layers (4--10~nm). Substrate choice does not affect the in-plane lattice parameter (\emph{a}), however, out-of-plane the \emph{c}-axis is weakly reduced in films grown on less closely lattice-matched substrates. These results are consistent with previous studies of ferromagnetic topological insulators. We also confirm the previous results of Park \emph{et al.}\cite{Park2015} regarding the Te capping layer growth and orientation, which grows epitaxially in three equivalent orientations on (Bi,~Sb)$_{2}$Te$_3$. Since the in-plane lattice parameter remains constant over this experimentally-relevant range, heterostructures of doped and undoped (Bi,~Sb)$_{2}$Te$_3$ (for example, References \citen{He2016} and \citen{Mogi2017}) are a more promising route to applications than devices which rely on inducing strain to tailor the electronic band structure (as in Reference \citen{Aramberri2016}).
\section*{Methods}
Vanadium-doped bismuth antimony telluride films were grown on SrTiO$_3$~(1~1~1), Al$_2$O$_3$~(0~0~0~1) and Si~(1~1~1) substrates by molecular beam epitaxy, and capped with 10~nm of tellurium (a detailed description is given in Chang et al.~\cite{Chang2015a}).
We performed high-resolution X-ray diffractometry on 4, 10 and 20 quintuple-layer films in a Panalytical Empyrean (Series 2) $\theta$-$\theta$ diffractometer. Optics were optimised for intensity over resolution due to the film thickness. We used a Ge(2 2 0) hybrid monochromator (for Cu$_{\kappa_{\alpha}}$), and a 1/2$^{\circ}$ divergence slit on the incident beam. Diffracted-beam optics comprised either a Xe proportional counter and 1~mm beam tunnel (for symmetrical 2$\theta$/$\omega$ measurements and rocking curves), or a PIXcel$^{3D}$ (Medipix2) detector in 1D frame grab mode (for asymmetrical reciprocal space measurements). Typically, a step time of at least 10s was required to detect the thin-film peaks, whereas the data shown in Figure \ref{thickness} has a step time of 20 minutes. Reciprocal-space maps are measured relative to the STO~(1~1~2), Al$_2$O$_3$~(0~1~-1~8) or Si~(3~3~1) peak.
A cross-sectional sample for TEM analysis was prepared using a dual beam FIB/SEM (FEI Helios Nanolab), from a 10 quintuple-layer film on SrTiO$_3$~(1~1~1). STEM-EDX analysis was carried out in a FEI Osiris, operated at 200~kV, equipped with a set of four EDX detectors in a cross configuration (Super-X by Bruker). Elemental maps were denoised using Principal Component Analysis (PCA) routines integrated in Hyperspy, an open source toolkit for EM data analysis, and the maps shown in Figure \ref{TEM}c are 100x40~px, with a px size 0.5~nm~x~0.5~nm and Gaussian blur 0.5~px for display. The High Resolution STEM images (HRSTEM) were acquired on a probe-corrected FEI Titan with an acceleration voltage of 300~kV.
\footnotesize
\bibliographystyle{clr4}
| 380619ac95c0036bcf0fe9da481df18deacdfe72 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The notion of weak containment was introduced by Kechris in the book \cite{K10} in the context of cost theory of Borel eqiuvalence relations. It turned out to be a very fruitfull concept. It is known that this pre-order has a maximal element(see \cite{GTW06} and \cite{K10}). It was proved by Abert and Weiss \cite{AbW13} that Bernoulli actions are minimal in the class of free actions. In the work \cite{S16} Seward introduced the notion of weak containment for joinings and used it in the context of so-called Rokhlin entropy.
Another notion we will work with is the notion of sofic action. In the groundbreaking paper \cite{B10} Lewis Bowen have defined so-called sofic entropy, the notion realiant on mimicking of a given acion over so-called sofic approximation(which is a finite approximation of a group).
An action which could be approximated in such a way is called sofic action.
It was already known that weak containent has something to do with soficity. It is a folklore fact that an action weakly contained in the sofic action is sofic itself. Carderi in \cite{C15} connected this notion to the soficity of action relative to given sofic approximation, namely he showed that action is approximable by given ultralimit sofic approximation iff it is weakly contained in the ultraproduct action of this sofic approximation. In the work \cite{BTD17} Bowen an Tucker-Drob have studied the space of so-called stable weak equivalence classes.
In this work we will firstly show that product of two sofic actions is also sofic (theorem \ref{theor: product is sofic}). We then show that there is a (non-unique) maximal element in the class of sofic actions of given sofic group with respect to the weak containment order (theorem \ref{theor: universal sofic action}) we will call it a {\em universal sofic action}. We then define a {\em maximal sofic approximation} to be any sofic approximation which approximates some (and hence any) universal action.
In the theorem \ref{theor: maximal sofic approximation} we show that any sofic action will be aprroximable by it in a nice fashion (without need to taking subsequences). Afterwards we recall the notion of doubly-quenched convergence introduced by Austin in \cite{Au16}. In the theorem \ref{theor: all sofic are strongly sofic} we prove using result of Austin that any sofic action admits a doubly-quenched approximation, the situation called by Hayes ``strong soficity'', with respect to maximal sofic approximation. In works \cite{H16a}, \cite{H16b} he showed that strong soficity of an action with respect to given sofic approximation has a lot of interesting consequences.
In the section \ref{sec: applications} we show how our results are combining with theorems of Hayes from \cite{H16a}, \cite{H16b}. It is a an important problem to understand whether for algebraic actions topological sofic entropy equals to the measure-theoretic one (then we endow this action with the Haar measure). In the corolary \ref{cor: algebraic} we use the result from \cite{H16a} to show that for maximal sofic approximations this equality holds whenever the measure-theoretic entropy is not $-\infty$. We also show in corolary \ref{cor: pinsker} that a product formula for Pinsker factors holds for maximal sofic approximations .
We study the class $\mathcal{S}$ of sofic groups all of whose actions are soficin the section \ref{sec: class s}. By a simple argument and using well-known result that all the treable actions are sofic (see \cite{EL10},\cite{Pa11}), we show that all the treeable groups (groups admitting an essentially free treeable action) belong to this class. We also show that this class is closed under taking subgroups.
{\em Acknowledgements. } I would like to thank Ben Hayes for his comments. Research is supported by the Russian Science Foundation grant β14-21-00035.
\section{Measures and Kantorovich distance}
Let $G$ be a countable group. We will fix an arbitrary enumeration of its elements $(\gamma_i)_{i \in N}$.
For a standard Borel space $X$ we will denote $\Me(X)$ the set of all the Borel probability measures on $X$.
For a probability space $(X,\mu)$ let us denote $\mathrm{MALG}(X,\mu)$ the space of all the measurable subsets endowed with the symmetric difference distance.
Let $(X, r)$ be a compact metric space. We denote $l$ --- the {\em Kantorovich distance} on $\Me(X)$ defined by
\[
l(\mu_1,\mu_2) = \inf_\xi \int_{X \times X} r(x_1,x_2) d\xi(x_1,x_2),
\]
there infimum is taken along all the couplings $\xi$ of measures $\mu_1$ and $\mu_2$.
It is known that the Kantorovich distance defines the weak* topology on the space $\Me(X)$.
Let $f$ be a map between two metric compacta $X$ and $Y$. If $f$ does not increase distance then the induced map on measures does not increase the Kantorovich distance.
A convention on metrics.
\begin{enumerate}
\item If metric for the space is explicitly stated then this is it.
\item If the space has the form $\Me(X)$ for $X$ a compact set with specified metric then we endow it with the Kantorovich distance.
\item If the space has the form $X^V$ there $V$ is a finite set and metric for $X$ is specified somehow then we will define metric on $X^V$ by
\[
d_{X^V}(t_1,t_2) = \sup_{v \in V} d_X(t_1(v),t_2(v))
\]
for $t_1,t_2 \in X^V$.
\item If the space has the form $X^G$ and for $X$ the metric is specified then we define the metric by
\[
d_{X^G} = \sup_{i \in \mathbb{N}} \frac1i d_X (t_1(\gamma_i),t_2(\gamma_i))
\]
for $t_1,t_2 \in X^G$.
\item All the finite sets we will endow with the discrete metric ($1$ for distinctive element and $0$ otherwise).
\end{enumerate}
\section{Obseravables}
All the actions of countable groups on probability spaces will be measure preserving if otherwise is not stated.
Let us fix a measure preserving action of countable group $G$ on the standard probability space $(X,\mu)$. An {\em observable} is any measrable map $f : X \to A $ where $A$ is a finite space. We denote $\mathrm{Obs}(X,\mu,A)$ the space of observables from $X$ to $A$.
We will endow it with the metric: for $f_1, f_2 \in \mathrm{Obs}(X,\mu,A)$ it will be defined as $\mu(\lbrace x \vert f_1(x) \neq f_2(x)\rbrace)$.
For observable $f$ we will define new map $f^G: X \to A^G$ by $(f^G(x))(g)= f(g x)$ for $x \in X$ and $g \in G$.
Consider a map $\mathrm{distr}: \mathrm{Obs}(X, \mu,A) \to \Me(A^G)$ defined as $f \mapsto f^{G}(\mu)$. We will also denote the action if needed: $\mathrm{distr}_T (f)$.
\begin{lem}\label{lem: distr continuous}
The map $\mathrm{distr}$ is continuous.
\end{lem}
\begin{proof}
It is enough to prove that measure of any cylinder set depends continuously on $f$. We can see that for any cylinder subset $C$ of $A^G$ the map $\mathrm{Obs}(X,\mu,A) \to \mathrm{MALG}(X,\mu) $ defined by $f \mapsto {(f^G)}^{-1}(C)$ is continuous. This implies the desired.
\end{proof}
Let $V$ be any set. Let $A$, $B$ be any finite sets. Let $\pi$ be a map from $B$ to $A$. We will denote $\pi^V: B^V \to A^V$ to be the map defined naturally by applying $\pi$ coordinate-wise. We will define a {\em tower of observables} to be the sequence of observables $f_i : X \to A_i$ together with the maps $\pi_{i,j}: A_i \to A_j$ such that for any $i<j$ holds $(\pi_{i,j} \circ f_i)(x) = f_j(x)$ for $\mu$-a.e. $x \in X$ and that for any $i<j<k$ we have $\pi_{j,i} \circ \pi_{k,j} = \pi_{k,i}$. We will say that the tower is {\em sufficient} if for any observable $f: X \to B$ and $\varepsilon>0$ there is an index $i$ and a map $\pi': A_i \to B$ such that observable $\pi' \circ f_i$ is $\varepsilon$-close to $f$. We note that it is easy to construct a sufficient tower of observables on the standard probability space.
We will say that p.m.p. action of $G$ on $(X_1,\mu_1)$ is {\em weakly contained} in the p.m.p. action of $G$ on $(X_2,\mu_2)$ if for any observable $f_1: X_1 \to A $ and any $\varepsilon>0$ there is an observable $f_2 : X_2 \to A$ such that $\mathrm{distr}(f_2)$ is $\varepsilon$-close to $\mathrm{distr}(f_2)$.
\section{Sofic groups and sofic actions}
For a finite set $V$ we will denote $\mathrm{Sym}(V)$ the group of all its permutations. We will endow it with the normalised Hamming distance:
\[
d_H(g_1,g_2) = \frac{ \lvert \lbrace v \in V \;,\; g_1(v) \neq g_2(v)\rbrace\rvert }{\lvert V\rvert}
\]
Let $G$ be a countable group. A {\em sofic approximation} $\mathfrak{S}$ is a sequence of finite sets $(V_i)$ together with a sequence of maps $(\sigma_i)$, $\sigma_i : G \to \mathrm{Sym}(V_i)$ (we will use the notation $\sigma_i : g \mapsto \sigma_i^g$) satisfying two properties:
\begin{enumerate}
\item $\lim_{i \to \infty} d_H(\sigma_i^{g_1},\sigma_i^{g_2}) = 1$ for any $g_1 \neq g_2$ from $G$,
\item $\lim_{i \to \infty} d_H(\sigma_i^{g_1} \circ \sigma_i^{g_2}, \sigma_i^{g_1 g_2}) = 0$ for any $g_1,g_2$ from $G$.
\end{enumerate}
A group is said to be {\em sofic} if it has at least one sofic approximation.
Let $\sigma$ be a map from $G$ to $\mathrm{Sym}(V)$ for some finite set $V$. Let $X$ be any set. We define a map $\theta_{v, \sigma}: X^V \to X^G$ by equality
\[
(\theta_{v,\sigma}(\tau))(g) = \tau(\sigma^g(v)).
\]
We will denote $\Theta_{v,\sigma}$ the map from $\Me(X^V)$ to $\Me(X^G)$ defined by equality
\[
\Theta_{\sigma}(\eta) = \frac1{\lvert V \rvert}\sum_{v \in V} \theta_{v,\sigma}(\eta).
\]
We will omit $\sigma$ in the notations above if it is clear from the context.
Let $\mathfrak{S} = ((V_i),(\sigma_i))$ be a sofic approximation. Let $G$ be acting on the the standard probability space $(X,\mu)$. We will say that this action is $\mathfrak{S}$-weakly approximable if for any observable $f: X \to A $ there is a growing sequence $(i_j)$ of natural numbers and a sequence $(\tau_j)$, $\tau_j \in A^{V_{i_j}}$ such that
\[
\lim_{j \to \infty} \Theta(\delta_{\tau_{i_j}}) = \mathrm{distr}(f).
\]
It is not hard to see that
We will say that this action is $\mathfrak{S}$-approximable if for any observable $f: X \to A $ there is a a sequence $(\tau_i)$, $\tau_i \in A^{V_i}$ such that
\[
\lim_{i \to \infty} \Theta(\delta_{\tau_i}) = \mathrm{distr}(f).
\]
We will say that action is {\em sofic} if it is $\mathfrak{S}$-weakly aproximable for some sofic approximation $\mathfrak{S}$ of the group $G$.
For fixed sofic group $G$ we will say that the map $\sigma: G \to \mathrm{Sym}(V)$ is $k$-good if
\begin{enumerate}
\item $ d_H(\sigma_i^{\gamma_i},\sigma_i^{\gamma_j}) > 1 - 1/k$ for any $i < j \leq k $,
\item $ d_H(\sigma_i^{\gamma_i} \circ \sigma_i^{\gamma_j}, \sigma_i^{\gamma_i \gamma_j}) < 1/k$ for any $i,j \leq k$.
\end{enumerate}
It is not hard to see that we can reformulate now the requirements from the definition of sofic approximation in the following way: for any natural $k$ there is such $i_0$ that for any $i>i_0$ we have that $\sigma_i$ is $k$-good.
It is easy to see that an action $G \curvearrowright (X,\mu)$ is sofic iff for any $f:X \to A$ and for any natural $k$ there is a $k$-good map $G \to \mathrm{Sym}(V)$ for some finite $V$ and an element $\tau \in A^V$ with $\Theta_{\sigma}(\delta_{\tau})$ being $1/k$-close to $\mathrm{distr}(f)$.
Let $(f_i)$, $f_i : X \to A_i$ be a sufficient tower of observables on $(X,\mu)$. Then an action $T$ of $G$ on $(X,\mu)$ is sofic iff for any natural $k$ we have that there is $\tau \in A_k^V$ and $\sigma: G \to \mathrm{Sym}(V)$ --- such a $k$-good map(for some finite $V$) that $\Theta_{\sigma}(\delta_\tau)$ is $1/k$-close to $\mathrm{distr}(f_k)$.
\begin{lem}\label{lem: sofic approximation for sofic action}
Any sofic action $T$ of sofic group $G$ on a standard probability space $(X,\mu)$ is $\mathfrak{S}$-approximable for some sofic approximation $\mathfrak{S}$.
\end{lem}
\begin{proof}
We fix a sufficient tower of observables $(f_i)$, $f_i : X \to A_i$ and take $\sigma_i: G \to \mathrm{Sym}(V_i)$ such that it is $i$-good and that there is such a $\tau \in A_i^{V_i}$ that $\Theta_{\sigma_i}(\delta_\tau)$ is $1/i$-close to $\mathrm{distr}(f_i)$. It is not hard to see now that for any natural $j<i$ we have that $\Theta_{\sigma_i}(\delta_{\pi_{i,j}^{V_i}(\tau)})$ is $1/i$-close to $\mathrm{distr}(f_j)$. Hence our action is approximable with respect to constructed sofic approximation.
\end{proof}
\begin{theor}\label{theor: product is sofic}
Direct product of two sofic actions of sofic group on standard probability spaces is a sofic action.
\end{theor}
\begin{proof}
Let $T',T''$, be two sofic actions of $G$ on standard probability spaces $(X',\mu')$,$(X',\mu')$ respectively. Let $(f_i')$, $f'_i : X' \to A_i'$ be a sufficient tower of observables for $(X',\mu')$ and $(f''_i)$, $f''_i : X' \to A_i'$ --- on $(X'',\mu'')$. Let us consider a tower of observables $(f'_i \otimes f''_i)$ on $(X' \times X'', \mu' \otimes \mu'')$ defined by
\[
f'_i \otimes f''_i : (x',x'') \mapsto (f'_i(x'),f''_i(x'')).
\]
It is not hard to see that this tower of observables is sufficient.
Suppose $T'$ is $\mathfrak{S}'$-approximable and $T''$ is $\mathfrak{S}''$-approximable for some sofic approximations $\mathfrak{S}',\mathfrak{S}''$. For a natural $j$ let $(\tau'_i), \tau'_i \in {A'_j}^{V_i}$ be such a sequence that
\[
\lim_{i \to \infty} \Theta_{\sigma'_i}(\delta_{\tau'_i}) = \mathrm{distr}(f'_j).
\]
Analogously we take $(\tau''_i), \tau''_i \in {A''_j}^{V_i}$ be such a sequence that
\[
\lim_{i \to \infty} \Theta_{\sigma''_i}(\delta_{\tau''_i}) = \mathrm{distr}(f''_j).
\]
We now consider the sequence $(\tau'_i \otimes \tau''_i)$, $\tau'_i \otimes \tau''_i : V'_i \times V''_i \to A'_j \times A''_j$ defined by
\[
(\tau'_i \otimes \tau''_i)(v',v'') = (\tau'_i(v'), \tau''_i(v'')).
\]
We define a sofic approximation $\mathfrak{S}' \times \mathfrak{S}''$ in such a way that $\sigma'_i \times \sigma''_i: G \to \mathrm{Sym}(V'_i \times V''_i)$ and
\[
(\sigma'_i \times \sigma''_i)^g(v',v'') = ({\sigma'}^g_i(v'),{\sigma''}^g_i(v''))
\]
for any $g \in G$, $v' \in V'_i$ and $v'' \in V''_i$.
It is a sofic approximation indeed and we can check that
\[
\lim_{i \to \infty} \Theta_{\sigma'_i \times \sigma''_i}(\delta_{\tau'_i} \otimes \delta_{\tau''_i}) = \lim_{i \to \infty} \Theta_{\sigma'_i \times \sigma''_i}(\delta_{\tau'_i \otimes \tau''_i}) = \mathrm{distr}(f'_i \otimes f''_i).
\]
This implies that $T' \times T'$ is $\mathfrak{S}' \times \mathfrak{S}''$-approximable and hence sofic.
\end{proof}
\begin{lem}
Inverse limit of sofic actions is sofic.
\end{lem}
\begin{proof}
Let $T$ be an action of $G$ on a standard probability space $(X,\mu)$, let $(\sA_i)$ be a an increasing sequence of $G$-invariant subalgebras whose join is the whole algebra of measurable sets on $X$. Suppose that all the corresponding factors are sofic.
Fix any observable $f: X \to A$ and a natural $k$. We can find such a close observable $f': X \to A$ that it is $\sA_i$-measurable for some $i$ and that $\mathrm{distr}(f')$ is $1/(2k)$-close to $\mathrm{distr}(f)$ (by lemma \ref{lem: distr continuous}).
Since factor corresponding to $\sA_i$ is sofic, we can find a $k$-good map $\sigma : G \to \mathrm{Sym}(V)$ for some finite $V$ and $\tau \in A^V$ such that $\Theta_{\sigma}(\delta_{\tau})$ is $1/(2k)$-close to $\mathrm{distr}(f')$. We have now that $\Theta_{\sigma}(\delta_{\tau})$ is $1/k$-close to $\mathrm{distr}(f)$. This implies that $T$ is a sofic action.
\end{proof}
\begin{lem}\label{lem: countable products}
Countable product of sofic actions is sofic.
\end{lem}
\begin{proof}
It's a direct consequence of two previous lemmata.
\end{proof}
We are now ready to prove that for any countable group there is a universal sofic action which weakly contains any other.
\begin{theor}\label{theor: universal sofic action}
For any sofic group $G$ there is such a sofic action $T$ that any other sofic action is weakly contained in it.
\end{theor}
\begin{proof}
The proof is completely analogous to that of the fact stating the existence of weakly maximal action among all the actions of the given group (see \cite{GTW06}, \cite{K10} and \cite{BuK16}. Of course we will need the previous lemma.
For every natural $i$ let $A_i$ be a set of size $i$. We now consider the set $D_i$ of all such measures $\mu$ on $A_i^G$ that there is a sofic action $T$ on some standard probability space $(X,\mu)$ and an observable $f: X \to A_i$ that $\mu= \mathrm{distr}_T(f)$. We now pick a countable dense set of measures from this set and for each measure take a corresponding action. We now proceed like this for all the natural $i$ which gives us a countable collection of sofic actions. Their product will be sofic by the previous lemma. It is easy to check that any sofic approximation is weakly contained in this product.
\end{proof}
We will call any such action from the previous theorem a {\em universal sofic action}. Sofic approximation $\mathfrak{S}$ is called {\em maximal} if any universal sofic action is $\mathfrak{S}$-approximable.
We now show that maximal sofic approximation approximates any sofic action.
\begin{theor}\label{theor: maximal sofic approximation}
If $\mathfrak{S}$ is a maximal sofic approximation for sofic group $G$ then any sofic action is $\mathfrak{S}$-approximable.
\end{theor}
\begin{proof}
Any sofic action is weakly contained in a universal sofic approximation which is $\mathfrak{S}$-approximable. Hence any action is $\mathfrak{S}$-approximable.
\end{proof}
We note that our definition of $\mathfrak{S}$-approximable action is rather strong and does not allow for taking subsequences to approximate.
We will say that an action $T$ of group $G$ on $(X,\mu)$ is $\mathfrak{S}$-{\em strongly sofic} if for any observable $f: X \to A$ there is a sequence $(\eta_i)$, $\eta_i \in \Me(A^{V_i})$ such that
\[
\lim_{i \to \infty}\Theta(\eta_i \otimes \eta_i) = \mathrm{distr}(f) \otimes \mathrm{distr}(f)
\]
and for any $\varepsilon>0$ the sets
\[
W_{i,\varepsilon} = \lbrace \xi \in (A \times A)^V_i \quad\vert\quad l(\Theta(\delta_{\xi}, \mathrm{distr}(f) \otimes \mathrm{distr}(f)) <\varepsilon\rbrace
\]
satisfy
\[
\lim_{i \to \infty} \eta_i \otimes \eta_i(W_{i,\varepsilon}) = 1.
\]
This kind of convergence condition is called {\em doubly-quenched convergence} and it was introduced by Austin in \cite{Au16}. We will need the following theorem from his work. For action $T$ and a natural $n$ we will denote $T^n$ the product-action of the same group.
\begin{theor}[Austin, \cite{Au16}]
Let $\mathfrak{S}$ be a sofic approximation and $T$ be an action. If for any natural $n$ we have that $T^n$ is $\mathfrak{S}$-approximable then $T$ is strongly sofic with respect to $\mathfrak{S}$.
\end{theor}
\begin{theor}\label{theor: all sofic are strongly sofic}
Let $\mathfrak{S}$ be a maximal sofic approximation for a sofic group $G$. Then any sofic action of $G$ is strongly sofic with respect to $\mathfrak{S}$.
\end{theor}
\begin{proof}
We note that if action $T$ is sofic then $T^n$ is sofic for any natural $n$ by the theorem \ref{theor: product is sofic}. It implies that $T^n$ is $\mathfrak{S}$-approximable for any natural $n$. Hence it is $\mathfrak{S}$-stronly sofic by the theorem of Austin.
\end{proof}
\section{Class $\mathcal{S}$}\label{sec: class s}
The class $\mathcal{S}$ was defined in by Paunescu in \cite{Pa11} to be the class of all the sofic groups such that all the m.p.t. actions of these groups are sofic. He also discovered some properties of this class.
Let $T$ be an m.p.t. action of group $G$ on the standard probability space $(X,\mu)$. Consider correspondent orbit equivalence relation $\sim$. Consider any collection $(\psi_i)$ of partial Borel maps (i.e. defined on a Borel subset) from $X$ to itself such that their graphs are subsets of an equivalence relation $\sim$. This sequence will be called {\em graphing} if induced graph is such that for almos every $x \in X$ and every $y \sim x$ we have that $x$ and $y$ are connected via the graph induced by the collection $(\psi_i)$. This graphing is called {\em treeing} if this graph is acyclic. An action relation is called {\em treeable} if its equivalence relations admits a treeing.
A countable group group is called {\em treeable} if it admits an essentially free treeable action.
\begin{theor}
Class of treeable groups is subset of $\mathcal{S}$.
\end{theor}
\begin{proof}
Take any actions $T$. Take any essentially free treeable action $T'$. Their direct product will be essentially free and treeable. Hence it is sofic (see \cite{EL10} and \cite {Pa11}). So $T$ is a factor of sofic action, hence it is sofic.
\end{proof}
\begin{theor}
Class $\mathcal{S}$ is closed under taking subgroups.
\end{theor}
\begin{proof}
Let $H$ be a subgroup of $G \in \mathcal{S}$. Let $T$ be an action of $H$. Consider a co-induced action $\tilde{T}$ of $G$. It is sofic. We note that induced action $\bar{T}$ of $H$ will be sofic and it is isomorphic to the product of $\lvert G : H\rvert$ copies of $T$. Hence $T$ is a factor of sofic action and hence it is sofic.
\end{proof}
\section{Applications}\label{sec: applications}
Let $X$ be a metrizable compact topological group and let $T$ be an action of countable group $G$ on $X$ by continuous automorphisms. Such an action is called an {\em algebraic action}. It could be endowed with the Haar measure, it is easy to see that this measure is presserved by this action. We refer the reader to \cite{H16a} for the definition of sofic topological and measure entropy.
\begin{cor}\label{cor: algebraic}
Let us fix a maximal sofic approximation. Then for any algebraic action we will have that measure sofic entropy of this action endowed with the Haar measure is either $-\infty$ or equals to the topological entropy. If acting group belongs to $\mathcal{S}$ then we have equality for maximal sofic approximations without additional conditions.
\end{cor}
\begin{proof}
If measure entropy is not $-\infty$ then we have that algebraic action endowed with the Haar measure is sofic (we note that if acting group belongs to $\mathcal{S}$ then this is always the case).
Hence it is strongly sofic with respect to our approximation (theorem \ref{theor: all sofic are strongly sofic}). By the theorem 1.1 of \cite{H16a} its measure entropy equals to the topological entropy.
\end{proof}
We refer the reader to \cite{H16a} for the definitiojn of Pinsker factor for sofic entropy in the presence.
\begin{cor}\label{cor: pinsker}
Let $T_i$ act on standard probability space $(X_i,\mu_i)$ for $i=1,2$. Let $\mathfrak{S}$ be a maximal sofic approximation and suppose that $T_i$ are sofic. Then Pinsker factors for sofic entropy in the presence for considered actions obey the prodict formula: Pinsker subalgebra for product action equals to the join of Pinsker algebras for $T_i$, $i=1,2$.
\end{cor}
\begin{proof}
Our actions are strongly sofic with respect to $\mathfrak{S}$ by \ref{theor: all sofic are strongly sofic} and the statement follows by the theorem 1.1 from \cite{H16b}.
\end{proof}
| 2e1cf6d4cba2a63d338cd9cf24cd545ffd8b0820 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The need for search often arises from a person's need to achieve a goal, or a task such as booking travels, buying a house, etc., which would lead to search processes that are often lengthy, iterative, and are characterized by distinct stages and shifting goals. ~\cite{jones2008beyond}. Thus, identifying and representing these tasks properly is highly important for devising search systems that can help end users complete their tasks. It has previously been shown that these task representations can be used to provide users with better query suggestions \cite{hassan2014supporting}, offer improved personalization \cite{mehrotra2015terms,white2013enhancing}, provide better recommendations \cite{zhang2015task}, help in satisfaction prediction \cite{wang2014modeling} and search result re-ranking. Moreover, accurate representations of tasks could also be highly useful in aptly placing the user in the task-subtask space to contextually target the user in terms of better recommendations and advertisements, developing task specific ranking of documents, and developing task based evaluation metrics to model user satisfaction. Given the wide range of applications these tasks representations can be used for, significant amount of research has been devoted to task extraction and representation \cite{lucchese2013discovering,hua2013identifying,kotov2011modeling,jones2008beyond,li2014identifying}.
Task extraction is quite a challenging problem as search engines can be used to achieve very different tasks, and each task can be defined at different levels of granularity. A major limitation in existing task-extraction methods lies in their treatment of search tasks as flat structure-less clusters which inherently lack insights about the presence or demarcation of subtasks associated with individual search tasks. In reality, often search tasks tend to be hierarchical in nature. For example, a search task like planning a wedding involves subtasks like searching for dresses, browsing different hairstyles, looking for invitation card templates, finding planners, among others. Each of these subtasks (1) could themselves be composed of multiple subtasks, and (2) would warrant issuing different queries by users to accomplish them. Hence, in order to obtain more accurate representations of tasks, new methodologies for constructing hierarchies of tasks are needed.
As part of the proposed research, we consider the challenge of extracting hierarchies of search tasks and their associated subtasks from a search log given just the log data without the need of any manual annotation of any sort. In a recent poster we showed that Bayesian nonparametrics have the potential to extract a hierarchical representation of tasks~\cite{mehrotra2015towards}; we extend this model further to form more accurate representations of tasks.
We present an efficient Bayesian nonparametric model for discovering hierarchies and propose a tree based nonparametric model to discover this rich hierarchical structure of tasks/subtasks embedded in search logs. Most existing hierarchical clustering techniques result in binary tree structures with each node decomposed into two child nodes. Given that a complex task could be composed of an arbitrary number of subtasks, these techniques cannot directly be used to construct accurate representations of tasks. In contrast, our model is capable of identifying task structures that can be composed of an arbitrary number of children. We make use of a number of evaluation methodologies to evaluate the efficacy of the proposed task extraction methodology, including quantitative and qualitative analyses along with crowdsourced judgment studies specifically catered to evaluating the quality of the extracted task hierarchies. We contend that the techniques presented expand the scope for better recommendations and search personalization and opens up new avenues for recommendations specifically targeting users based on the tasks they involve in.
\section{Related Work}
Web search logs provide explicit clues about the information seeking behavior of users and have been extensively studied to improve search experiences of users. We cover several areas of related work and discuss how our work relates to and extends prior work.
\subsection{Task Extraction}
There has been a large body of work focused on the problem of segmenting and organizing query logs into semantically coherent structures. Many such methods use the idea of a \textit{timeout} cutoff between queries, where two consecutive queries are considered as two different sessions or tasks if the time interval between them exceeds a certain threshold \cite{catledge1995characterizing,he2002combining,silverstein1999analysis}. Often a 30-minute timeout is used to segment sessions.
However, experimental results of these methods indicate that the timeouts are of limited utility in predicting whether two queries belong to the same task, and unsuitable for identifying session boundaries.
More recent studies suggest that users often seek to complete multiple search tasks within a single search session \cite{mehrotra2016characterizing,lucchese2011identifying} with over 50\% of search sessions having more than 2 tasks \cite{mehrotra2016characterizing}. At the same time, certain tasks require significantly more effort, time and sessions to complete with almost 60\% of complex information gathering tasks continued across sessions \cite{agichtein2012search,ma2008exploring}. There have been attempts to extract in-session tasks \cite{jones2008beyond,lucchese2011identifying,spink2005multitasking}, and cross-session tasks \cite{kotov2011modeling,wang2013learning} from query sequences based on classification and clustering methods, as well as supporting users in accomplishing these tasks \cite{hassan2014supporting}. Prior work on identifying search-tasks focuses on task extraction from search sessions with the objective of segmenting a search session into disjoint sets of queries where each set represents a different task \cite{lucchese2013discovering,hua2013identifying}.
Kotov et al. \cite{kotov2011modeling} and Agichtein et al. \cite{agichtein2012search} studied the problem of cross-session task extraction via binary same-task classification, and found different types of tasks demonstrate different life spans. While such task extraction methods are good at linking a new query to an on-going task, often these query links form long chains which result in a task cluster containing queries from many potentially different tasks. With the realization that sessions are not enough to represent tasks, recent work has started exploring cross-section task extraction, which often results in complex non-homogeneous clusters of queries solving a number of related yet different tasks. Unfortunately, pairwise predictions alone cannot generate the partition of tasks efficiently and even with post-processing, the final task partitions obtained are not expressive enough to demarcate subtasks \cite{liao2012evaluating}. Finally, authors in \cite{li2014identifying} model query temporal patterns using a special class of point process called Hawkes processes, and combine topic model with Hawkes processes for simultaneously identifying and labeling search tasks.
Jones et al. \cite{jones2008beyond} was the first work to consider the fact that there may be multiple subtasks associated with a user's information need and that these subtasks could be interleaved across different sessions. However, their method only focuses on the queries submitted by a single user and attempts to segment them based on whether they fall under the same information need. Hence, they only consider solving the task boundary identification and same task identification problem and cannot be used directly for task extraction. Our work alleviates the same user assumption and considers queries across different users for task extraction. Finally, in a recent poster \cite{mehrotra2015towards}, we proposed the idea of extracting task hierarchies and presented a basic tree extraction algorithm. Our current work extends the preliminary model in a number of dimensions including novel model of query affinities and task coherence based pruning strategy, which we observe gives substantial improvement in results. Unlike past work, we also present detailed derivation and evaluation of the extracted hierarchy and application on task extraction.
\subsection{Supporting Complex Search Tasks}
There has been a significant amount of work on task continuation assistance \cite{morris2008searchbar,agichtein2012search}, building task tours and trails \cite{o2010tweetmotif,singla2010studying}, query suggestions \cite{baeza2005query,jones2006generating,mei2008query}, predicting next search action \cite{cao2009towards} and notes taking when accomplishing complex tasks \cite{donato2010you}. The quality of most of these methods depends on forming accurate representations of tasks, which is the problem we are addressing in this paper.
\subsection{Hierarchical Models}
Rich hierarchies are common in data across many domains, hence quite a few hierarchical clustering techniques have been proposed. The traditional methods for hierarchically clustering data are bottom-up agglomerative algorithms. Probabilistic methods of learning hierarchies have also been proposed \cite{blundell2013bayesian,liu2012automatic} along with hierarchical clustering based methods \cite{heller2005bayesian,chuang2002towards}. Most algorithms for hierarchical clustering construct binary tree representations of data, where leaf nodes correspond to data points and internal nodes correspond to clusters. There are several limitations to existing hierarchy construction algorithms. The algorithms provide no guide to choosing the correct number of clusters or the level at which to prune the tree. It is often difficult to know which distance metric to choose. Additionally and more importantly, restriction of the hypothesis space to binary trees alone is undesirable in many situations - indeed, a task can have any number of subtasks, not necessarily two. Past work has also considered constructing task-specific taxonomies from document collections \cite{yang2012constructing}, browsing hierarchy construction \cite{yang2015browsing}, generating hierarchical summaries \cite{lawrie2003generating}. While most of these techniques work in supervised settings on document collections, our work instead focused on short text queries and offers an unsupervised method of constructing task hierarchies.
Finally, Bayesian Rose Trees and their extensions have been proposed \cite{segal2002probabilistic,blundell2012bayesian,blundell2013bayesian} to model arbitrary branching trees. These algorithms naively cast relationships between objects as binary (0-1) associations while the query-query relationships in general are much richer in content and structure.
We consider a number of such existing methods as baselines and the various advantages of the proposed approach is highlighted in the evaluation section wherein the proposed approach in addition to being more expressive, performs better than state-of-the-art task extraction and hierarchical methods.
\begin{table}[t!]
\centering
\resizebox{!}{!}
{
\begin{tabular}{ll}
\textbf{Symbol} & \textbf{Description}\\
$n_T$ & number of children of tree T\\
$ab|c$ & partition of set $\lbrace a,b,c\rbrace$ into disjoint sets $\lbrace a,b\rbrace$,$\lbrace c\rbrace$\\
ch(T) & children of T\\
$\phi(T)$ & partition of tree T\\
$p(D_m|T_m)$ & likelihood of data $D_m$ given the tree $T_m$\\
$\pi_{T_m}$ & mixing proportions of partition of tree $T$\\
$f(D_m)$ & marginal probability of the data $D_m$\\
$\mathbb{H}(T)$ & set of all partitions of queries $Q = leaves(T)$\\
$f(Q)$ & task affinity function for set of queries Q\\
$r_{q_i,q_j}^k$ & the k-th inter-query affinity between $q_i$ \& $q_j$
\end{tabular}
}
\caption{Table of symbols}
\label{symbols}
\end{table}
\section{Defining Search Tasks}
Jones et al.~\cite{jones2008beyond} was one of the first papers to point out the importance of task representations, where they defined a search task as:
\theoremstyle{definition}
\begin{definition}{}
A \textit{search task} is an atomic information need resulting in one or more queries.
\end{definition}
Ahmed et al.~\cite{hassan2014supporting} later extended this definition to a more generic one, which can also capture task structures that could possibly consist of related subtasks, each of which could be complex tasks themselves or may finally split down into simpler tasks or atomic informational needs. Following Ahmed \textit{et al.}~\cite{hassan2014supporting}, a complex search task can then be defined as:
\theoremstyle{definition}
\begin{definition}{}
A \textit{complex search task} is a multi-aspect or a multi-step information need consisting of a set of related subtasks, each of which might recursively be complex.
\end{definition}\label{ComplexTask}
The definition of complex tasks is much more generic, and captures all possible search tasks, that can be either complex or atomic (non-complex). Throughout this paper we adopt the definition provided in Definition~\ref{ComplexTask}.2 as the definition for a search task.
Hence, by definition a search task has a hierarchical nature, where each task can consist of an arbitrary number of, possibly complex subtasks. An effective task extraction system should be capable of accurately identifying and representing such hierarchical structures.
\begin{table*}[t!]
\centering
{
\begin{tabular}{c|c}
\hline
\multicolumn {2}{c}{\textbf{Query-Term Based Affinity ($r^1$)}} \\
\hline
cosine & cosine similarity between the term sets of the queries\\
\hline
edit & norm edit distance between query strings\\
\hline
Jac & Jaccard coeff between the term sets of the queries\\
\hline
Term & proportion of common terms between the queries\\
\hline
\multicolumn {2}{c}{\textbf{URL Based Affinity ($r^2$)}} \\
\hline
Min-edit-U & Minimum edit distance between all URL pairs from the queries\\
\hline
Avg-edit-U & Average edit distance between all URL pairs from the queries\\
\hline
Jac-U-min & Minimum Jaccard coefficient between all URL pairs from the queries\\
\hline
Jac-U-avg & Average Jaccard coefficient between all URL pairs from the queries\\
\hline
\multicolumn {2}{c}{\textbf{Session/User Based Affinity ($r^3$)}} \\
\hline
Same-U & if the two queries belong to the same user\\
\hline
Same-S & if the two queries belong to the same session\\
\hline
\multicolumn {2}{c}{\textbf{Embedding Based Affinity ($r^4$)}} \\
\hline
Embedding & cosine distance between embedding vectors of the two queries\\
\hline
\end{tabular}
}
\caption{Query-Query Affinities.\label{tab:affinity}}
\end{table*}
\section{Constructing Task Hierarchies}
While hierarchical clustering are widely used for clustering, they construct binary trees which may not be the best model to describe data's intrinsic structure in many applications, for example, the task-subtask structure in our case. To remedy this, multi-branch trees are developed. Currently there are few algorithms which generate multi-branch hierarchies. Blundel \textit{et al.}~\cite{blundell2012bayesian,blundell2013bayesian} adopt a simple, deterministic, agglomerative approach called BRTs (Bayesian Rose Trees) for constructing multi-branch hierarchies. In this work, we adapt BRT as a basic algorithm and extend it for constructing task hierarchies. We next describe the major steps of BRT approach.
\begin{figure}[!]
\centering
\includegraphics[width=0.45\textwidth]{joining.png}
\vspace{-2mm}
\caption{The different ways of merging trees which allows us to obtain tree structures which best explain the task-subtask structure.}\label{fig4}
\vspace{-2mm}
\end{figure}
\subsection{Bayesian Rose Trees}
BRTs~\cite{blundell2012bayesian,blundell2013bayesian} are based on a greedy probabilistic agglomerative approach to construct multi-branch hierarchies. In the beginning,
each data point is regarded as a tree on its own: $T_i = \{x_i\}$ where $x_i$ is the feature vector of i-th data. For each step, the algorithm selects two trees $T_i$ and $T_j$ and merges them into a new tree $T_m$. Unlike binary hierarchical clustering, BRT uses three possible merging operations, as shown in Figure \ref{fig4}:
\begin{itemize}
\item \textbf{Join}: $T_m = {T_i, T_j}$, such that the tree $T_m$ has two children now
\item \textbf{Absorb}: $T_m = {children(T_i) \cup T_j}$, i.e., the children of one tree gets absorbed into the other tree forming an absorbed tree with $>$2 children
\item \textbf{Collapse}: $T_m = {children(T_i) \cup children(T_j)}$, all the children of both the subtrees get combined together at the same level.
\end{itemize}
Specifically, in each step, the algorithm greedily finds two trees $T_i$ and $T_j$ to merge which maximize the ratio of probability:
\begin{equation}
\frac{p(D_{m}|T_m)}{p(D_{i}|T_i)p(D_{j}|T_j)}
\end{equation}
where $p(D_m|T_m)$ is the likelihood of data $D_m$ given the tree $T_m$, $D_m$ is all the leaf data of $T_m$, and $D_m = D_i \cup D_j$
. The probability $p(D_m|T_m)$ is recursively defined on the children of $T_m$:
\begin{equation}
p(D_m | T_m) = \pi_{T_m} f(D_m) + (1 - \pi_{T_m}) \prod_{T_i \in ch(T_m)} p(D_i|T_i)
\end{equation}
where $f(D_m)$ is the marginal probability of the data $D_m$ and $\pi_{T_m}$ is the "\textit{mixing proportion}". Intuitively, $\pi_{T_m}$
is the prior probability that all the data in $T_m$ is kept in one cluster instead of partitioned into sub-trees. In BRT\cite{blundell2012bayesian}, $\pi_{T_m}$ is defined as:
\begin{equation}
\pi_{T_m} = 1 - (1 - \gamma)^{n_{T_m}-1}
\end{equation}
where $n_{T_m}$ is the number of children of $T_m$, and $0 \geq \gamma \leq 1$ is the hyperparameter to control the model. A larger $\gamma$ leads to coarser partitions and a smaller $\gamma$ leads to finer partitions. Table \ref{symbols} provides an overview of notations \& symbols used throughout the paper.
\subsection{Building Task Hierarchies}
We next describe our task hierarchy construction approach built on top of Bayesian Rose Trees. A tree node in our setting is comprised of a group of queries which potentially compose a search task, i.e. these are the set of queries that people tend to issue in order to achieve the task represented in the tree node.
We define the task-subtask hierarchy recursively: T is a task if either T contains all the queries at its node (an atomic search task) or if T splits into children trees as $T = \{T_1,T_2,...,T_{n_T}\}$ where each of the children trees ($T_i$) are disjoint set of queries corresponding to the $n_T$ subtasks associated with task $T$. This allows us to consider trees as a nested collection of sets of queries defining our task-subtask hierarchical relation.
To form nested hierarchies, we first need to model the query data. This corresponds to defining the marginal distribution of the data $f(D_m)$ as defined in Equation 2. The marginal distribution of the query data ($f(D_m)$) helps us encapsulate insights about task level interdependencies among queries, which aid in constructing better task representations. The original BRT approach \cite{blundell2012bayesian} assumes that the data can be modeled by a set of binary features that follow the Bernoulli distribution. In other words, features (that represent the relationship/similarities between data points) are not weighted and can only be binary. Binary (0/1) relationships are too simplistic to model inter-query relationships; as a result, this major assumption fails to capture the semantic relationships between queries and is not suited for modeling query-task relations. To this end, we propose a novel query affinity model and to alleviate the binary feature assumption imposed by BRT, we propose a conjugate model of query affinities, which we describe next.
\subsection{Conjugate Model of Query Affinities}
\label{taskaffinity}
A tree node in our setting is comprised of a group of queries which \textit{potentially} belong to the same search task. The likelihood of a tree should encapsulate information about the different relationships which exists between queries. Our goal here is to make use of the rich information associated with queries and their result set available to compute the likelihood of a set of queries to belong to the same task. In order to do so, we propose a query affinity model which makes use of a number of different inter-query affinities to determine the tree likelihood function.
We next describe the technique used to compute four broad categories of inter-query affinity and later describe the Gamma-Poisson conjugate model which makes use of these affinities to compute the marginal distribution of the data. \\
\noindent\textbf{Query-term based Affinity ($r^1$):}\\
Search queries catering to the same or similar informational needs tend to have similar query terms. We make use of this insight and capture query level affinities between a pair of queries. We make use of cosine similarity between the query term sets, the normalized edit distances between queries and the Jaccard Coefficient between query term sets.\\% We use the summation of all the pairwise similarities between two queries as defined in Query-term part of Table \ref{tab:affinity} as the Query-term based affinity.
\noindent\textbf{URL-based Affinity ($r^2$):}\\
Users tackling similar tasks tend to issue queries (possibly different) which return similar URLs, thus encoding the URL level similarity between pairs of queries into the query affinity model helps in capturing another task-specific similarity between queries. Any query pair having high URL level similarity increase the possibility of the query pair originating from similar informational needs. We capture a number of URL-based signals including minimum and average edit distances between URL domains and jaccard coefficient between URLs.\\% We use the summation of all the pairwise similarities between two queries as defined in URL-based part of \ref{tab:affinity} as the URL-based affinity.\\
\noindent\textbf{User/Session based Affinity ($r^3$):}\\
It is often the case that users issue related queries within a session so as to satisfy their informational need. We leverage this insight by making use of session level information (as a 0/1 binary feature) and user-level information (as a 0/1 binary feature) in our affinity model to identify queries issued in the same session and by the same user accordingly.\\% We sum both user and session information between two queries and assign it as the User/Session based affinity.\\
\noindent\textbf{Query Embedding based Affinity ($r^4$):}\\
Word embeddings capture lexico-semantic regularities in language, such that words with similar syntactic and semantic properties are found to be close to each other in the embedding space. We leverage this insight and propose a query-query affinity metric based on such embeddings. We train a skip-gram word embeddings model where a query term is used as an input to a log-linear classifier with continuous projection layer and words within a certain window before and after the words are predicted. To obtain a query's vector representation, we average the vector representations of each of its query terms and compute the cosine similarity between two queries' vector representations to quantify the embedding based affinity ($r^4$).
Table \ref{tab:affinity} summarizes all features considered to compute these affinities. Our goal is to capture information from all four affinities when defining the likelihood of the tree. We assume that the global affinity among a group of queries can be decomposed into a product of independent terms, each of which represent one of the four affinities from the query-group. For each query group $Q$, we take the normalized sum of the affinities from all pairs of queries in the group $Q$ to form each of the affinity component ($r^k$, k={1,2,3,4}).
Poisson models have been shown as effective query generation models for information retrieval tasks \cite{mei2007study}. While these affinities could be used with a lot of distributions, in the interest of computational efficiency and to avoid approximate solutions, our model will use a hierarchical Gamma-Poisson distribution to encode the query-query affinities. We incorporate the gamma-Poisson conjugate distribution in our model under the assumptions that the query affinities are discretized and for a group of queries $Q$, the affinities can be decomposed to a product of independent terms, each of which represents contributions from the four different affinity types. Finally, for a tree ($T_m$) consisting of the data ($D_m$), i.e. the set of queries $Q$, we define the marginal likelihood as:
\begin{equation}
f(D_m) = f(Q) = \prod_{k=1}^{k=4} p \bigg( \sum_{i \in 1 \cdots |Q|}\sum_{j \in 1 \cdots |Q|}r^k_{q_i,q_j} | \alpha_k, \beta_k\bigg)
\end{equation}
where $\alpha_k$ \& $\beta_k$ are respectively the shape parameter \& the rate parameter of the four different affinities. Making use of the Poisson-Gamma conjugacy, the probability term in the above product can be written as:
\begin{equation}
p(r|\alpha,\beta) = \int_\lambda p(r|\lambda) p(\lambda|\alpha,\beta)d\lambda
\end{equation}
\begin{equation}
= \Bigg\{ \frac{\Gamma (\alpha +r)}{r!\,\Gamma(\alpha)}\Bigg( \frac{\beta}{\beta +1}\Bigg)^\alpha \Bigg(\frac{1}{\beta +1}\Bigg)^r \Bigg\}
\end{equation}
where $\lambda$ is the Poisson mean rate parameter which gets eliminated from computations because of the Gamma-Poisson conjugacy and where $r$, $\alpha$ \& $\beta$ get replaced by affinity class specific values.
\subsection{Task Coherence based Pruning}
\label{sec:pruning}
The search task extraction algorithm described above provides us a way of constructing a task hierarchy wherein as we go down the tree, nodes comprising of complex multi-aspect tasks split up to provide finer tasks which ideally should model user's fine grained information needs. One key problem with the hierarchy construction algorithm is the continuous splitting of nodes which results in singleton queries occupying the leave nodes. While splitting of nodes which represent complex tasks is important, the nodes representing simple search task queries corresponding to atomic informational needs should not be further split into children nodes. Our goal in this section is to provide a way of quantifying the task complexity of a particular node so as to prevent splitting up nodes representing atomic search task into further subsets of query nodes.
\subsubsection{Identifying Atomic Tasks
We wish to identify nodes capturing search subtasks which represent atomic informational need. In order to do so, we introduce the notion of \textit{Task Coherence}:
\theoremstyle{definition}
\begin{definition}{}
\textit{Task Coherence} is a measure indicating the atomicity of the information need associated with the task. It is captured by the semantic closeness of the queries associated with the task.
\end{definition}\label{TaskCoherence}
By measuring Task Coherence, we intend to capture the semantic variability of queries within this task in an attempt to identify how complex or atomic a task is. For example, a tree node corresponding to a complex task like planning a vacation would involve queries from varied informational needs including flights, hotels, getaways, etc; while a tree node corresponding to a finer task representing an atomic informational need like finding discount coupons would involve less varied queries - all of which would be about discount coupons. Traditional research in topic modelling has looked into automatic evaluation of topic coherence \cite{newman2010automatic} via Pointwise Mutual Information. We leverage the same insights to capture task coherence.
\subsubsection{Pointwise Mutual Information}
PMI has been studied variously in the context of collocation extraction \cite{pecina2010lexical} and is one measure of the statistical independence of observing two words in close proximity. We wish to compute PMI scores for each node of the tree. A tree node in our discussion so far has been represented by a collection of search queries. We split queries into terms and obtain a set of terms corresponding to each node, and calculate a node's PMI scores using the node's set of query terms.
More specifically, the PMI of a given pair of query terms ($w_1$ \& $w_2$) is given by:
\begin{equation}
PMI(w_1,w_2) = log \frac{p(w_1,w_2)}{p(w_1)p(w_2)}
\end{equation}
where the probabilities are determined from the empirical statistics of some full standard collection. We employ the AOL log query set for this and treat two query terms as co-occurring if both terms occur in the same session. For a given task node ($Q$), we measure task coherence as the average of PMI scores for all pairs of the search terms associated with the task node:
\begin{equation}
PMI-Score(Q) = \frac{1}{|w|} \sum^{|w|}_{i=1} \sum^{|w|}_{j=1} PMI(w_i,w_j)
\end{equation}
where $|w|$ represents the total number of unique search terms associated with task node $Q$. The node's PMI-Score is used as the final measure of task coherence for the task represented via the corresponding node.
\subsubsection{Tree Pruning}
We use the task coherence score associated with each node of the task hierarchy constructed, and prune lower level nodes of the tree to avoid aggressive node splitting. The overall motivation here is to avoid splitting nodes which represent simple search tasks associated with atomic informational needs. We scan through all levels of the search task hierarchy obtained by the algorithm described above and for each node compute its task coherence score. If the task coherence score exceeds a specific threshold, it implies that all the queries in this particular node are aimed at solving the same or very similar informational need and hence, we prune off the sub-tree rooted at this particular node and ignore all further splits of this node.
\subsection{Algorithmic Overview}
We summarize the overall algorithm to construct the hierarchy by outlining the steps. The problem is treated as one of greedy model selection: each tree T is a different model, and we wish to find the model that best explains the search log data in terms of task-subtask structure.\\
\noindent\textbf{Step 1: Forrest Initialization}:\\
The tree is built in a bottom-up greedy agglomerative fashion, starting from a forest consisting of n (=$|Q|$) trivial trees, each corresponding to exactly one vertex. The algorithm maintains a forest F of trees, the likelihood $p(i) = p(D_{i}|T_i)$ of each tree $T_i \in F$ and the different query affinities. Each iteration then merges two of the trees in the forest. At each iteration, each vertex in the network is a leaf of exactly one tree in the forest. At each iteration a pair of trees in the forest F is chosen to be merged, resulting in forest $F*$.\\
\noindent\textbf{Step 2: Merging Trees}:\\
At each iteration, the best potential merge, say of trees X and Y resulting in tree I, is picked off the heap. Binary trees do not fit into representing search tasks since a task is likely to be composed of more than two subtasks. As a result, following \cite{blundell2013bayesian} we consider three possible mergers of two trees $T_i$ and $T_j$ into $T_m$. $T_m$ may be formed by joining $T_i$ and $T_j$ together using a new node, giving $T_m = \lbrace T_i, T_j\rbrace$. Alternatively $T_m$ may be formed by absorbing $T_i$ as a child of $T_j$, yielding $T_m = \lbrace T_j\rbrace \bigcup ch(T_i)$, or vice-versa, $T_m = \lbrace T_i\rbrace \bigcup ch(T_j)$. We explain the different possible merge operations in Figure \ref{fig4}. We obtain arbitrary shaped sub-trees (without restricting to binary tress) which are better at representing the varied task-subtask structures as observed in search logs with the structures themselves learnt from log data. Such expressive nature of our approach differentiates it from traditional agglomerative clustering approaches which necessarily result in binary trees.\\
\noindent\textbf{Step 3: Model Selection}:\\
Which pair of trees to merge, and how to merge these trees, is determined by considering which pair and type of merger yields the largest Bayes factor improvement over the current model. If the trees $T_i$ and $T_j$ are merged to form the tree M, then the Bayes factor score is:
\begin{equation}
SCORE(M;I,J) = \frac{p(D_{M}|F*)}{p(D_{M}|F)}
\end{equation}
\begin{equation}
\hspace{4mm} = \frac{p(D_{M}|M)}{p(D_{i}|T_i)p(D_{j}|T_j)}
\end{equation}
where $p(D_{i}|T_i)$ and $p(D_{j}|T_j)$ are given by the dynamic programming equation mentioned above. After a successful merge, the statistics associated with the new tree are updated. Finally, potential mergers of the new tree with other trees in the forest are considered and added onto the heap.
The algorithm finishes when no further merging results in improvement in the Bayes Factor score. Note that the Bayes factor score is based on data local to the merge - i.e., by considering the probability of the connectivity data only among the leaves of the newly merged tree. This permits efficient local computations and makes the assumption that local community structure should depend only on the local connectivity structure.\\
\noindent\textbf{Step 4: Tree Pruning}:\\
After constructing the entire hierarchy, we perform the post-hoc tree pruning procedure described in Section \ref{sec:pruning} wherein we identify atomic task nodes via their task coherence estimates and prune all child nodes of the identified atomic nodes.
\section{Experimental Evaluation}
We perform a number of experiments to evaluate the proposed task-subtask extraction method. First, we compare its performance with existing state-of-the-art task extraction systems on a manually labelled ground-truth dataset and report superior performance (\ref{exp1}). Second, we perform a detailed crowd-sourced evaluation of extracted tasks and additionally validate the hierarchy using human labeled judgments (\ref{exp2}). Third, we show a direct application of the extracted tasks by using the task hierarchy constructed for term prediction (\ref{exp3}).\\% Finally, we present a qualitative analysis of a sample hierarchy extracted (\ref{qualitative}).\\
\noindent\textbf{Parameter Setting:}\\
Unless stated otherwise, we made use of the best performing hyperparameters for the baselines as reported by the authors. The query affinities in the proposed approach were computed from the specific query collection used in the dataset used for each of the three experiments reported below. While hyperparmeter optimization is beyond the scope of this work, we experimented with a range of the shape and inverse scale hyperparameters ($\alpha$, $\beta$) used for the Poison Gamma conjugate model and used the ones which performed best on the validation set for the search task identification results reported in the next section. Additionally, for the tree pruning threshold, we empirically found that a threshold of 0.8 gave the best performance on our toy hierarchies, and was used for all future experiments.
\subsection{Search Task Identification}
\label{exp1}
\noindent To justify the effectiveness of the proposed model in identifying search tasks in query logs, we employ a commonly used AOL data subset with search tasks annotated which is a standard test dataset for evaluating task extraction systems.
We used the task extraction dataset as provided by Lucchese \textit{et al.}\cite{lucchese2011identifying}. The dataset comprises of a sample of 1000 user sessions for which human assessors were asked to manually identify the optimal task-based query sessions, thus producing a ground-truth that can be used for evaluating automatic task-based session discovery methods. For further details on the dataset and the dataset access links, readers are directed to Lucchese \textit{et al.}\cite{lucchese2011identifying}.
We compare our performance with a number of search task identification approaches:
\begin{itemize}
\item \textbf{Bestlink-SVM} \cite{wang2013learning}: This method identified search task using a semi-supervised clustering model based on the latent structural SVM framework.
\item \textbf{QC-HTC/QC-WCC} \cite{lucchese2011identifying}: This series of methods viewed search task identification as the problem of best approximating the manually annotated tasks, and proposed both clustering and heuristic algorithms to solve the problem
\item \textbf{LDA-Hawkes} \cite{li2014identifying}: a probabilistic method for identifying and labeling search tasks that model query temporal patterns using a special class of point process called Hawkes processes, and combine topic model with Hawkes processes for simultaneously identifying and labeling search tasks.
\item \textbf{LDA Time-Window(TW)}: This model assumes queries belong to the same search task only if they lie in a fixed or flexible time window, and uses LDA to cluster queries into topics based on the query co-occurrences within the same time window. We tested time windows of various sizes and report results on the best performing window size.
\end{itemize}
\subsubsection{Metrics}
A commonly used evaluation metric for search task extraction is the pairwise F-measure computed based on pairwise precision/recall \cite{jones2008beyond,kotov2011modeling} defined as,
\begin{equation}
p_{pair} = \frac{\Sigma_{i\le j} \delta(y(q_i),y(q_j))\delta(\hat{y}(q_i),\hat{y}(q_j))}{\delta(\hat{y}(q_i),\hat{y}(q_j))}
\end{equation}
\begin{equation}
r_{pair} = \frac{\Sigma_{i\le j} \delta(y(q_i),y(q_j))\delta(\hat{y}(q_i),\hat{y}(q_j))}{\delta(y(q_i),y(q_j))}
\end{equation}
where $p_{pair}$ evaluates how many pairs of queries predicted in the same task, i.e., $\delta(\hat{y}(q_i),\hat{y}(q_j) = 1$, are actually annotated as in the same task, i.e., $\delta(y(q_i),y(q_j)) = 1$ and $r_{pair}$ evaluates how many pairs annotated as in the same task are recovered by the algorithm.
Thus, globally F-measure evaluates the extent to which a task contains only queries of a particular annotated task and all queries of that task. Given $p_{pair}$ and $r_{pair}$, the F-measure is computed as:$F_1 = \frac{2\times p_{pair}\times r_{pair}}{p_{pair} + r_{pair}}$.
\subsubsection{Results \& Discussion}
Figure \ref{fig5} compares the proposed model with alternative probabilistic models and state-of-the-art task identification approaches by F1 score. To make fair comparisons, we consider the last level of the pruned tree constructed as task clusters when computing pairwise precision/recall values. It is important to note that the labelled dataset has only flat tasks extracted on a per user basis; as a result, this dataset is not ideal for making fair comparisons of the proposed hierarchy extraction method with baselines. Nevertheless, the proposed approach manages to outperform existing task extraction baselines while having much greater expressive powers and providing the subdivision of tasks into subtasks. LDA-TW performs the worst since its assumptions on query relationship within the same search task are too strong. The advantage over QC-HTC and QC-WCC demonstrates that appropriate usage of query affinity information can even better reflect the semantic relationship between queries, rather than exploiting it in some collaborative knowledge.
\begin{figure}[t!]
\centering
\includegraphics[width=0.5\textwidth]{Fscore4.pdf}
\caption{F1 score results on AOL tagged dataset}\label{fig5}
\end{figure}
\begin{table*}[t]
\centering
\resizebox{0.8\textwidth}{!}
{
\begin{tabular}{c|c|c|c|c|c}
& \multicolumn {4}{c}{\textbf{Task Relatedness}} \\
\hline
& \textbf{Proposed} & \textbf{LDA-TW} & \textbf{QC-WCC} & \textbf{LDA-Hawkes} & \textbf{QC-HTC}\\
\hline
Task Related & \textbf{72\%*} & 47\% & 60\% & 67\% & 61\%\\
\hline
Somewhat Related & 20\% & 14\% & 15\% & 13\% & 5\% \\
\hline
Unrelated & \textbf{10\%} & 23\% & 25\% & 20\% & 34\%\\
\hline
\end{tabular}
}
\caption{Performance on Task Relatedness. The results highlighted with * signify statistically significant difference between the proposed approach and best performing baseline using $\chi^2$ test with $p\leq 0.05$.\label{tab:relate}}
\end{table*}
\begin{table}[t!]
\centering
\resizebox{0.5\textwidth}{!}
{
\begin{tabular}{c|c|c|c|c|c}
& \multicolumn {4}{c}{\textbf{Subtask Validity}} \\
\hline
& \textbf{Proposed} & \textbf{Jones} & \textbf{BHCD} & \textbf{BAC}\\%\multicolumn {3}{c}{Purity} & \multicolumn {3}{c}{NMI Scores} & \multicolumn {3}{c|}{PMI Scores}\\
\hline
Valid & \textbf{81\%*} & 69\% & 51\% & 49\%\\
\hline
Somewhat Valid & 8\% & 19\% & 17\% & 21\% \\
\hline
Not Valid & 11\% & 12\% & 32\% & 30\%\\
\hline
& \multicolumn {4}{c}{\textbf{Subtask Usefulness}} \\
\hline
Useful & \textbf{67\%*} & 52\% & 41\% & 43\%\\
\hline
Somewhat Useful & 8\% & 17\% & 19\% & 20\%\\
\hline
Not Useful & 25\% & 31\% & 40\% & 37\%\\
\hline
\end{tabular}
}
\caption{Performance on Subtask Validity and Subtask Usefulness. Results highlighted with * signify statistically significant difference between the proposed framework and best performing baseline using $\chi^2$ test with $p\leq 0.05$.\label{tab:valid}}
\end{table}
\subsection{Evaluating the Hierarchy}
\label{exp2}
While there are no gold standard datasets for evaluating hierarchies of tasks, we performed crowd-sourced assessments to assess the performance of our hierarchy extraction method. We separately evaluated the coherence and quality of the extracted hierarchies via two different set of judgements obtained via crowdsourcing.\\
\noindent\textbf{\textit{Evaluation Setup}}\\
For the judgment study, we make use of the AOL search logs and randomly sampled entire query history of frequent users who had more than 1000 search queries. The AOL log is a very large and long-term collection consisting of about 20 million of Web queries issued by more than 657000 users over 3 months. We run the task extraction algorithms on the entire set of queries of the sampled users and collect judgments to assess the quality of the tasks extracted. Judgments were provided by over 40 judges who were recruited from the Amazon Mechanical Turk crowdsourcing service. We restricted annotators to those based in the US because the logs came from searchers based in the US. We also used hidden quality control questions to filter out poor-quality judges. The judges were provided with detailed guidelines describing the notion of search tasks and subtasks and were provided with several examples to help them better understand the judgement task.\\
\noindent\textbf{\textit{Evaluating Task Coherence}}\\
In the first study, we evaluated the quality of the tasks extracted by the task extraction algorithms. In an ideal task extraction system, all the queries belonging to the same task cluster should ideally belong to the same task and hence have better task coherence. To this end, we evaluate the task coherence property of the tasks extracted by the different algorithms. For each of the baselines and the proposed algorithm, we select a task at random from the set of tasks extracted and randomly pick up two queries from the selected task. We then ask the human judges the following question:\\
\noindent\textbf{RQ1: Task Relatedness:} Are the given pairs of queries related to the same task? The possible options include (i) Task Related, (ii) Somewhat Task Related and (iii) Unrelated.\\
The task relatedness score provides an estimate of how coherent tasks are. Indeed, a task cluster containing queries from different tasks would score less in Task Relatedness score since if the task cluster is impure, there is a greater chance that the 2 randomly picked queries belong to different tasks and hence get judged Unrelated.\\% In addition to evaluating purity of the task clusters, we also evaluate the quality of the hierarchy constructed, as described below.
\noindent\textbf{\textit{Evaluating the hierarchy}}\\
While there are no gold standard dataset to evaluate hierarchies, in our second crowd-sourced judgment study, we evaluate the quality of the hierarchy extracted. A valid task-subtask hierarchy would have the parent task representing a higher level task with its children tasks representing more focused subtasks, each of which help the user achieve the overall task identified by the parent task.
We evaluate the correctness of the hierarchy by validating parent-child task-subtask relationships. More specifically, we randomly select a parent node from the hierarchy and then randomly select a child node from the set of its immediate child nodes. Given such parent-child node pairs, we randomly pick 5 queries from the parent node and randomly pick 2 queries from the child node. We then show the human judges these parent and child queries and ask the following questions:\\
\noindent\textbf{RQ2: Subtask Validity:} Consider the set of queries representing the search task and the pair of queries representing the subtask. How valid is this subtask given the overall task?\\
The possible judge options include (i) Valid Subtask, (ii) Somewhat valid and (iii) Invalid. Answering this question helps us in analyzing the correctness of the parent-child task-subtask pairs.\\
\noindent\textbf{RQ3: Subtask Usefulness:} Consider the set of queries representing the search task and the pair of queries representing the subtask. Is the subtask useful in completing the overall search task?
The possible judge options include (i) Useful, (ii) Somewhat Useful and (iii) Not Useful. This helps us in evaluating the usefulness of task-subtask pairs by finding the proportion of subtasks which help users in completing the overall task described by the parent node. Overall, the RQ2 and RQ3 help in evaluating the correctness and usefulness of the hierarchy extracted.\\
\noindent\textbf{\textit{Baselines}}\\
Since RQ1 evaluates task coherence without any notion of task-subtask structure, we compare against the top performing baselines from the task extraction setup described in section \ref{exp1}. On the other hand, RQ2 \& RQ3 help in answering questions about the quality of hierarchy constructed. To make fair comparisons while evaluating the hierarchies, we introduce additional hierarchy extraction baselines:
\begin{itemize}
\item \textbf{Jones Hierarchies} \cite{jones2008beyond}: A supervised learning approach for task boundary detection and same task identification. We train the classifier using the supervised Lucchese AOL task dataset and use it to extract tasks on the current dataset used in the judgment study.
\item \textbf{BHCD} \cite{blundell2013bayesian}: A state-of-the-art bayesian hierarchical community detection algorithm based on stochastic blockmodels and makes use of Beta-Bernoulli conjugate priors to define a network. We build a network of queries and apply BHCD algorithm to extract hierarchies of query communities.
\item \textbf{Bayesian Agglomerative Clustering (BAC)} \cite{heller2005bayesian}: A standard agglomerative hierarchical clustering model based on Dirichlet process mixtures.
\end{itemize}
\noindent\textbf{\textit{Results \& Discussion}}\\
For the first judgment study, each HIT is composed of 20 query pairs per approach being judged for task relatedness. We had three judges work on every HIT. Overall, per method we obtained judgments for 60 query pairs to evaluate the performance on task-relatedness. From among the three judges judging each query-pair, we followed majority voting mechanism to finalize the label for the instance. Table \ref{tab:relate} presents the proportions of query pairs judged as related. About 72\% of query pairs were judged task-related for the proposed approach with LDA-Hawkes performing second best with 67\%. Task relatedness measures how pure the task clusters obtained are, a higher score indicates that the queries belonging to the same task are indeed used for solving the same search task. The overall results indicate that the tasks extracted by the proposed task-subtask extraction algorithm are indeed better than those extracted by the baselines.
For the second judgment study used for evaluating the quality of the hierarchy, we show 10 pairs of parent-child questions in each HIT and ask the human annotators to judge the subtask validity and usefulness. Overall, per method we evaluate 300 such judgments resulting in over 1200 judgments and used maximum voting criterion from among the 3 judges to decide the final label for each instance. Table \ref{tab:valid} compares the performance of the proposed hierarchy extraction method against other hierarchical baselines. The identified subtask was found useful in 67\% cases with the best performing baseline being useful in 52\% of judged instances. This highlights that the extracted hierarchy is indeed composed of better subtasks which are found to be useful in completing the overall task depicted by the parent task. It is interesting to note that for BHCD and BAC baselines, most often the subtasks were found to be invalid and not useful.
Since the same parent-child task-subtask was judged for validity and usefulness, it is expected that the proportion of task-subtasks judged useful would be less than the ones judged valid. Indeed, as can be seen from the Table \ref{tab:valid}, the relative proportions of tasks-subtasks found useful is much less than those found valid.
\begin{figure}[t]
\centering
\includegraphics[width=0.5\textwidth]{TermPrediction3.pdf}
\caption{Term Prediction performance}\label{fig6}
\vspace{-4mm}
\end{figure}
\subsection{Term Prediction}
\label{exp3}
\noindent In addition to task extraction and user study based evaluation, we chose to follow an indirect evaluation approach based on \textit{Query Term Prediction} wherein given an initial set of queries, we predict future query terms the user may issue later in the session
This is in line with our goal of supporting users tackling complex search tasks since a task identification system which is capable of identifying "\emph{good}" search tasks will indeed perform better in predicting the set of future query terms.
To evaluate the performance of the proposed task extraction method, we primarily work with the TREC Session Track 2014 \cite{carterette2013overview} and AOL log data and constructed a new dataset consisting of user sessions from AOL logs concerned with Session Track queries. The session track data consists of over 1200 sessions while AOL logs consists of 20M search queries issued by over 657K users. We find the intersection of queries between the Session Track data and AOL logs to identify user sessions in AOL data trying to achieve similar task objectives. The Session Track dataset consists of 60 different \textbf{\textit{topics}}.
For each of these 60 topics, we separately find user sessions from the entire AOL logs which contain query overlaps with these topics. For each topic, we iterate through the entire AOL logs and select any user session which contains query overlap with the current topic. As a result, we obtain a total of 14030 user sessions which contain around 6.4M queries.
Given the initial queries from a user session and a set of tasks extracted from Session Track data, we leverage queries from the identified task to predict future query terms. For each Session Track topic, we construct a task hierarchy and use the constructed task hierarchy to predict future query terms in the associated user sessions. More specifically, for each topic, we split each user session into two parts: (i) task matching and (ii) held-out evaluation part. We use queries from the task matching part of user sessions to obtain the right node in the task hierarchy from which we then recommend query terms. We pick the tree node which has the highest cosine similarity score based on all the query terms under consideration. We evaluate based on the absolute recall scores - the average number of recommended query terms which match with the query terms in the held-out evaluation part of user sessions.
We baseline against the top performing task extraction baselines from Section \ref{exp1} as well as the top performing hierarchical algorithms from Section \ref{exp2}. To make fair comparisons, we consider nodes at the bottom most level of the pruned tree for task matching and term recommendation.
Figure \ref{fig6} compares the performance on term prediction against the considered baselines. We plot the average number of query terms predicted against the proportion of user session data used. The proposed method is able to better predict future query terms than a standard task extraction baseline as well as a very recent hierarchy construction algorithm.
\section{Conclusion}
Search task hierarchies provide us with a more naturalistic view of considering complex tasks and representing the embedded task-subtask relationships. In this paper we first motivated the need for considering hierarchies of search tasks \& subtasks and presented a novel bayesian nonparametric approach which extracts such hierarchies. We introduced a conjugate query affinity model to capture query affinities to help in task extraction. Finally, we propose the idea of Task Coherence and use it to identify atomic tasks. Our experiments demonstrated the benefits of considering search task hierarchies. Importantly, we were able to demonstrate competitive performance while at the same time outputting a richer and more expressive model of search tasks. This expands the scope for better task recommendation, better search personalization and opens up new avenues for recommendations specifically targeting users based on the tasks they are involved in.
\bibliographystyle{ACM-Reference-Format}
| 00b58095b3f148c9229723aa706f53c8fa770223 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{sec:Intro}
In the past decades many theoretical methods have been developed in the attempt to predict and rationalise molecular electronic structures.
Coupled cluster approaches are amongst the most widespread reference methods and are based on the exponential \textit{Ansatz} for the molecule's wavefunction \cite{Bishop1991,Bartlett2007}.
The unfavourable scaling of these methods, however, makes them unsuitable for large scale calculations or predictive assessments for a large number of molecular systems.
In contrast to these well-established methods, many-body perturbation theory (MBPT) has emerged as a computational alternative for molecules \cite{Blase2011,Faber2011,Ke2011,Ren2012,Marom2012,Caruso2012a,Bruneval2013,Pham2013,Korbel2014,GW100}.
The computational scheme first proposed by Hedin \cite{Hedin1965}, and summarised in \autoref{Fig:Hedin} (a-b), involves several quantities: starting with an initial guess for the Green's function ($G$) one determines the system's irreducible polarizability ($\chi_0$) and the screened interaction $W$.
These quantities alone fully specify the self-energy operator $\Sigma$ in the first iteration of the scheme, which then produces an updated Green's function, as shown in panel (b).
For later iterations the inclusion of vertex effects is in principle required through the vertex function $\Lambda$.
The main difficulty in including the vertex is that it involves additional diagrams at each step, since the vertex is formally generated by the functional derivative of the self-energy with respect to the Green's function.
Exploiting exact functional relations \cite{Schindlmayr1998a} can simplify the calculations only to a certain extent, and no purely numerical scheme can be implemented for the original method of Hedin.
It is understandable then that the $GW$ approximation (GWA) completely neglects the vertex, thus iteratively applying the scheme in \autoref{Fig:Hedin}-b.
This approximation is justified when exchange and correlation effects are completely absent in the reference state and only for the first iteration of the scheme, \textit{i.e.} when $\Sigma_{\mathrm{xc}}=0$, as for the Hartree case \cite{Morris2007}.
However, it has been widely used also on Hartree-Fock (HF) or density functional theory (DFT) starting points \cite{Klimes2014b,Bruneval2013,Wang2003,Sommer2012,Miyake2013,Cazzaniga2012}.
Actual GWA calculations vary significantly in details. The most common approximation is to create the starting orbitals using
standard density functional theory and perform a single shot $G_0W_0$ calculation. This works well for extended systems,
but for small molecules this approximation tends
to yield underbound highest molecular orbitals (HOMO). One way to cure this problem is to perform the calculations
self-consistently.
An update of the quasi-particle (QP) energies alone produces what is now often called the ev-$G_nW_0$ or the ev-$G_nW_n$ schemes, if the updated eigenvalues are included only in the Green's function or also in the screened interaction, respectively.
Fully self-consistent $GW$ calculations \cite{Caruso2012a,Caruso2013a,Kutepov2016} require the calculation of the interacting Green's function $G(\omega) = (\omega -T-V_{\rm{ext}}-\Sigma)^{-1}$, with $T$ the kinetic energy operator and $V_{\rm{ext}}$ the external potential generated by the atomic nuclei and
$\Sigma$ being the self-energy.
The resulting quasi-particle equations involve an energy dependent non-Hermitian self-energy.
To make this problem more manageable, quasi-particle self-consistent $GW$ approaches \cite{Faleev2004,VanSchilfgaarde2006} have been devised that disregard the non-Hermitian component of the self-energy and approximate its energy dependence \cite{Shishkin2007a}.
Recent studies on selected organic molecules have shown that self-consistency significantly improves the quasi-particle energies\cite{Caruso2013a,Gallandi2016,Kaplan2016}.
However, the HOMO is now on average too negative; in other words self-consistent approaches tend to over-bind the electrons and yield a too large ionization potential (IP). As we will discuss below the inclusion of the vertex cures this shortcoming already in the first iteration of the scheme.
\begin {figure}
\centering
\includegraphics [width=12.5cm] {Hedin.jpg}
\caption{Many-Body perturbation theory (MBPT) computational approaches. (a) Hedin's pentagon, (b) $GW$ approximation. (c) Computational scheme adopted in this work. See text for a definition of the symbols.}
\label{Fig:Hedin}
\end{figure}
Apart from the overestimation of the binding energies, self-consistency is also computationally fairly expensive, and other simple
means to improve the QP energies have been considered as well.
One is the admixture of exact exchange in the reference mean-field ground state calculation. From a fundamental point of view,
this should certainly improve the occupied orbitals, since for the standard local and semi-local density functionals that are commonly used as starting point
the exchange-correction potential does not properly decay like one over the distance from the nuclei~\cite{Handy1969}. Also the HF one electron energies
are closer to the final $GW$ QP energies, so that an iterative solution of the $GW$ equations should not be necessary.
An elegant mean to achieve the correct $1/r$ decay of the nuclear-electron potential
is using long-range hybrid functionals \cite{Bruneval2013,Gallandi2016}, but standard hybrid functionals\cite{Kaplan2016}
as well as Hartree-Fock starting points have been used as well.
In the present work, we have also decided to use the HF reference point to initiate the $GW$ calculations. The reasons for this choice are summarized below.
(i) From a practical point of view, this choice obviously allows for a rather straightforward comparison with wavefunction based methods
that usually rely on the HF reference point. Accordingly, we will use CCSD(T) reference data taken from the literature \cite{Bruneval2013}.
(ii) There is evidence that HF theory is a reasonably accurate starting point for small molecules:
it is well known that, because of the Brillouin condition in HF theory, single excited configurations only contribute in second order in the perturbative expansion of the exact wavefunction, whereas for Br\"uckner orbitals their contribution is exactly zero to all orders.
From a diagrammatic standpoint this translates into an identical cancellation of single excited determinants in lowest order for the ground state energy evaluated using either the HF or the Br\"uckner orbitals \cite{Kobe1971}.
Since in localised systems the contribution of correlated higher order excitations is small \cite{Bhaskaran-Nair2016}, one can reasonably expect that the HF determinant has a large overlap with the exact many-body wavefunction, similarly to the Br\"uckner orbitals.
(iii) The main point of the present work is to investigate how important vertex corrections are. For the HF starting point, the vertex corrections
can be evaluated exactly, as they are give by the functional derivative of the exchange potential with respect to the Green's function. Hence we can unambiguously study the effect of vertex corrections.
This brings us back to a point already alluded to above. $G_0W_0$@HF, as well as self-consistent $GW$ place the HOMO at too negative binding energies, or in other words, the IP is too large \cite{Bruneval2013}. The reason for this is fairly simple. In the GWA--- which implies that the random phase approximation is used
to calculate the polarizability
---particle-hole ladder diagrams are not included.
The resulting $W$ underscreens the interaction between electrons because the energy to create an electron-hole pair is too large.
This largely explains why the IP is overestimated.
The second issue affecting the GWA stems from the so-called self-screening error \cite{Nelson2007a,Fernandez2009a,Aryasetiawan2012}.
In the $GW$ self-energy, one state can be occupied by two particles, since direct and exchange interactions are not treated on the same footing \cite{Romaniello2009}.
This situation is reminiscent of the self-interaction problem that appears at the mean-field level in the Hartree theory. The inclusion of the exact exchange in the Hartree-Fock theory completely cures this shortcoming on the mean-field level but neglects correlation. Including
the functional derivative of the exchange with respect to the Green's function in Hedin's equations also eliminates the problem on the level of correlations.
To set the stage, we now summarize our computational approach. The initial step is a HF calculation for a set of molecules.
We then calculate the standard $G_0W_0$@HF QP energies for this set, and find as expected that this approximation overestimates the IP.
We then present post-$GW$ calculations for the same set of molecules.
In these calculations, the vertex corrections are included and given by the functional derivative of the exchange potential $\Sigma_\mathrm{x}$
with respect to the HF Green's function $G(\omega) =(\omega -T -V_{\rm{ext}}-V_H -\Sigma_\mathrm{x})^{-1}$
(here $V_H$ is the Hartree potential).
The vertex is included either only in the polarizability ($G_0W_0^{\rm{tc-tc}}$) or in the polarizability and in the self-energy ($G_0W_0\Gamma$). In present work, only the diagonal components of the self-energy are calculated, and changes of the orbitals are concomitantly not considered,
nor do we perform the calculations self-consistently, since this would in principle require a consistent update of the vertex as already emphasized above.
The inclusion of the vertex in the polarizability is equivalent to calculating the polarizabilty using time-dependent Hartree-Fock \cite{RevModPhys.36.844}, and the computational cost is also similar with both methods scaling like $O(N^6)$.
There are also formal similarities to the Bethe-Salpeter equation (BSE) performed on top of a $GW$ reference state. The latter method is commonly adopted for solid state systems \cite{Albrecht1998,Rohlfing2000}, but has also been used to study the optical properties of molecular systems \cite{Bruneval2015,Jacquemin2015}.
To the best of our knowledge, vertex corrections have been routinely considered only for extended systems \cite{Bobbert1994,Delsole1994,Bruneval2005,Morris2007,Maebashi2011,Gruneis2014a,Kutepov2016}. For finite systems, calculations have been performed
for atoms \cite{Shirley1993}, for very simple molecules (within the Tamm-Dancoff approximation) \cite{Kuwahara2016},
or with the SOSEX \cite{Ren2015,Gallandi2016}, which approximates the vertex only in second order.
A local vertex $\Lambda_{\rm{LDA}}$ \cite{Hung2016} has also been used for a set of aromatic molecules, however, this simple two-point vertex behaves quite differently than the four-point many-body vertex used here.
Among other things, the local approximation breaks fundamental self-energy symmetries and, albeit these effects are small for extended solids \cite{Delsole1994}, they might be more relevant for very inhomogeneous systems such as molecules.
The work is structured in the following manner. In \autoref{sec:Theo} we derive simple equations for the vertex corrected $GW$,
reducing the set of equations to three equations that are reminiscent of the equation of motion. Contrary to Hedin's equation we use a four-point notation.
In \autoref{sec:2bleCounting} we report a diagrammatic analysis of the self-energy thus generated.
The computational details are given in \autoref{sec:Comp}, and our results are finally presented and discussed in \autoref{sec:Res}, where we apply the $GW\Gamma$ method perturbatively to a set of molecules.
\section{Theory}
\label{sec:Theo}
\subsection{Hedin's equations}
\label{sec:Hedin}
In principle, there are at least two complementary routes to improve on the MBPT approaches: either a self-consistent evaluation of the standard GWA or the inclusion of sub-leading diagrammatic contributions.
Practical implementations have been directed primarely towards self-consistency (see \autoref{sec:Intro} for a discussion), whereas vertex corrections remain largely unaddressed, because of the high implementational as well as computational complexity.
To present our computational approach we will rely on a four point notation, for instance used by Starke and Kresse \cite{Starke2012}, although in the present work we have decided to rearrange the indices to more easily connect with the available standard literature.
Also the specific succession of indices adopted here is easier to memorise.
Using a four index notation Hedin's equations can be written as
\begin{eqnarray}
G(1,2) & =& G_0(1,2) + G_0(1,3) \Sigma_{\mathrm{xc}}(3,4) G(4,2) \label{equ:h1} \\
\Sigma_{\mathrm{xc}}(1,2) &= & \mathrm{i} \,G(5,6) \Gamma(1,5,3,4) W(3,4,2,6) \label{equ:h2} \\
\Pi_0(1,2,3,4) & = &- G(1,5) G(6,2) \Gamma(5,6,3,4) \label{equ:h3} \\
W(1,2,3,4) & =& V(1,2,3,4) + \label{equ:h4} \\
& + &V(1,2,5,6) \Pi_0(5,6,7,8) W(7,8,3,4) \nonumber \\
\Gamma(1,2,3,4)& = &\delta(1,3)\delta(2,4) + \label{equ:h5} \\
&-& \mathrm{i} \, I(1,2,5,6) G(5,7) G(8,6) \Gamma(7,8,3,4) \nonumber
\end{eqnarray}
In these equations, integrals over repeated indices are assumed.
We note that if two four-point quantities are multiplied, for instance, $\Pi_0 W$ in Eq. (\ref{equ:h4}), then the integral is always over the two intermediate indices, in the equation below 5 and 6, and the order is easy to memorise:
\begin{equation}
\label{equ:matrixmul}
A(1,2,3,4) = \int \Pi_0(1,2,5,6) W(5,6,3,4) \, d5 \, d6.
\end{equation}
The advantage of using a four point notation is the following.
In Hedin's equations a numeral, for instance $1$, corresponds necessarily to a space-time and spin point $1=({\bf r}_1, t_1, \sigma_1)$;
Hedin's original equations do not apply in orbital space or momentum space.
In the four point notation above, however, one can perform a unitary transformation of the position coordinate to reciprocal space or to any set of orthogonal orbitals without changing the equations.
The slight disadvantage of the present ordering of the indices is that the first index and the fourth index in each four point object transform like covariant coordinates (transformed by say $U$), whereas the second and third indices transform like contra-variant coordinates (transformed by $U^\dagger$).
Before continuing we briefly reiterate the meaning of the individual objects.
The four-point Coulomb interaction in space-time coordinates is $V(1,2,3,4)= \delta(1,2) \delta(3,4) \delta(t_1, t_4) v(\mathbf{r}_1,\mathbf{r}_4)$ generalising the usual two-point counterpart $v(\mathbf{r}_1,\mathbf{r}_4)=\frac{1}{|\mathbf{r}_1-\mathbf{r}_4|}$.
$G_0(1,2)$ is the Hartree Green's function, generated by the corresponding Hartree self-energy: $\Sigma_0(1,2)=$ $V_H(1,2)=$ $\mathrm{i} \delta(1,2) \delta(t_1,t_3) [v(\mathbf{r}_1,\mathbf{r}_3) G(3,3^+)]$ in space-time coordinates, $\Sigma_{\mathrm{xc}}(1,2)$ is the self-energy including exchange as well as correlation terms.
$\Pi_0$ is the irreducible polarization (propagator); it can not be divided into two individual polarization propagators by cutting a single Coulomb line $V$.
$W$ is the screened interaction summing the Coulomb interactions to infinite order.
Finally $\Gamma$ is the four-point vertex, which is completely specified by the kernel $I$:
\begin{equation}
\label{eq:kern}
I(1,2,3,4) = \frac{\delta \, \Sigma_{\mathrm{xc}}(1,2) }{\delta\, G(3,4)}.
\end{equation}
There is an alternative way to write these equations that avoids the vertex altogether.
We first introduce an auxiliary independent two-particle propagator $L_0(1,2,3,4)= -G(1,3)G(4,2)$.
This allows to rewrite Eqs. (\ref{equ:h3}) and (\ref{equ:h5}) as
\begin{eqnarray}
\Pi_0(1,2,3,4) & = & L_0(1,2,5,6) \Gamma(5,6,3,4) \nonumber \\
\ \Gamma(1,2,3,4)& = &\delta(1,3)\delta(2,4) + \nonumber \\
&+& \mathrm{i} \, I(1,2,5,6) L_0(5,6,7,8) \Gamma(7,8,3,4). \nonumber
\end{eqnarray}
We can now combine both equations to a single equation
\begin{equation}
\label{equ:pi0}
\Pi_0 = L_0 + L_0 (\mathrm{i} \, I) \Pi_0 \quad\leftrightarrow \quad\Pi_0^{-1} = L_0^{-1} - \mathrm{i} \, I.
\end{equation}
Here we have suppressed the indices and the matrix multiplications need to be done according to the rule stated above (Eq. \ref{equ:matrixmul}).
For the inversion, it is understood that the first two and last two indices of each four point quantity are combined to a super index and the matrices are inverted using the two super indices as row and column indices.
Likewise, we can introduce the full polarization propagator
\begin{equation}
\label{equ:pi}
\Pi = \Pi_0 + \Pi_0 V \Pi \quad\leftrightarrow \quad\Pi^{-1} = \Pi_0^{-1} - V.
\end{equation}
It is a simple matter to see that these two equations can be further combined into a single equation, reading
\begin{equation}
\label{equ:pifinal}
\Pi = L_0 + L_0 (V+ \mathrm{i} \, I) \Pi \leftrightarrow\Pi^{-1} = L_0^{-1} - \mathrm{i} \, I- V.
\end{equation}
The proof is most easily done inspecting the right hand inverted Dyson like equations in the previous three equations.
Eq. (\ref{equ:pifinal}) is precisely the polarization propagator as calculated by means of the ``Bethe-Salpeter equation'' (BSE) in most solid state codes.
In fact, in the preceding lines, we have backtracked the calculations performed by Starke and Kresse, where Hedin's equations were derived from the equation of motion and the BSE.
We can now proceed with $W$ and the self-energy.
By expanding the right hand side of Eq. (\ref{equ:h4}) to infinite order in $V$ one can identify $\Pi$ and rewrite Eq. (\ref{equ:h4}) as
\begin{equation}
\label{eq:W}
W = V + V \Pi V.
\end{equation}
The final issue is to rewrite Eq. (\ref{equ:h2}), specifically the four point term $\Gamma W = \Gamma(1,2,3,4) W(3,4,5,6) $ such that it involves only the already calculated polarization propagators.
Using Eq. (\ref{eq:W}) for $W$ and inserting the vertex from Eq. (\ref{equ:h5}), we obtain:
\begin{equation}
\begin{split}
\Gamma W & = ( 1 - (\mathrm{i} I) L_0 -(\mathrm{i} I) L_0 (\mathrm{i} I) L_0 + ...) (V + V \Pi V ) \nonumber \\
& = ( 1 + (\mathrm{i} I) \Pi_0) (V + V \Pi V ) \\
& = V + V \Pi V + (\mathrm{i} I) (\Pi_0 +\Pi_0 V \Pi ) V \\
& = V + V \Pi V + (\mathrm{i} I) \Pi V = V + (V +\mathrm{i} I) \Pi V.
\end{split}
\end{equation}
From the first to the second line we have used Eq. (\ref{equ:pi0}), and from the third to the fourth line we have used Eq. (\ref{equ:pi}).
In summary, we can rewrite Hedin's equation in the much more compact form
\begin{eqnarray}
\Pi & = & L_0 + L_0 (V+ \mathrm{i} \, I) \Pi \label{eq:PiBSE} \\
\Sigma_{\mathrm{xc}} & =& \mathrm{i} \, G \Big( V + (V +\mathrm{i} I) \Pi V \Big ) \label{equ:GWGamma} \\
G & =& G_0 + G_0 \Sigma_{\mathrm{xc}} G, \label{equ:greens}
\end{eqnarray}
where the contraction of the two-point Green's function and a four-point quantity is here defined as
\begin{equation}
\Sigma_\mathrm{x}(1,3) =\mathrm{i} (GV)(1,3) = \int G(2,4) V (1,2,3,4) d 2 \, d 4.
\end{equation}
In these equations the somewhat arbitrary distinction between the Coulomb kernel $V$ and the remaining interaction kernel $\mathrm{i} I$ has been dropped.
These equations are essentially equivalent to the compact equations given by Starke and Kresse but have been derived here from Hedin's equations instead of the more fundamental equation of motion.
We note that the same equation for the self-energy can be also found in Ref. \cite{Romaniello2012} (albeit without derivation) and is seemingly common knowledge in quantum field theoretical publications \cite{Held2011,Ayral2013}.
A few comments are in place. As already emphasized, $\Pi$ in Eq. (\ref{eq:PiBSE}) is the polarization propagator (or polarizability) that many solid state BSE codes calculate. It describes how test charges are screened by the electronic system, hence it is often referred to as
test-charge test-charge (tc-tc) polarizability. The self-energy Eq. (\ref{equ:GWGamma}) describes
the effects of the many-body system on an added or removed electron, and Eq. (\ref{equ:greens}) is the related Green's function.
The first term on the right hand side of Eq. (\ref{equ:GWGamma}) is just the exchange potential ($GV$), whereas the second term describes, in principle, all correlation effects. If $iI$ is neglected in
the self-energy (\ref{equ:GWGamma}) but included in (\ref{eq:PiBSE}) the approximation is commonly
referred to as $GW^{\rm{tc-tc}}$.
Although this set of equations is in principle exact, approximations need to be made for the interaction kernel $I$ specified in \autoref{eq:kern}.
In our case, we perform calculations perturbatively on the Hartree-Fock reference state, \textit{i.e.} we start the Hedin scheme with the sum of the Hartree ($\Sigma_0$) and exchange self-energy ($\Sigma_\mathrm{x}$): $\Sigma= \Sigma_0 + \Sigma_\mathrm{x}$.
Then the interaction kernel is simply the functional derivative of the the exchange self-energy $\Sigma_\mathrm{x}$ and thus explicitly given by
\[
\mathrm{i} I(1,2,3,4) = - \delta(2,4) \delta(1,3) \delta(t_1, t_2) v(\mathbf{r}_1,\mathbf{r}_2),
\]
where $v( \mathbf{r}_1,\mathbf{r}_2)$ is the bare Coulomb interaction between particles. The polarization propagator is then just the polarizability of time-dependent Hartree-Fock \cite{RevModPhys.36.844}.
The approximation employed in this study is also related to the random phase approximation with exchange (RPAx) first proposed by Szabo and Ostlund \cite{Szabo1977}.
There has been a recent revival in interest for the RPAx \cite{Hesselmann2010,Bleiziffer2012,Furche2013,Nguyen2014,Mussard2015,Colonna2016}, however, previous studies have focused on the evaluation of correlation energies, rather than the self-energy considered here.
Finally, we comment on the use of the $GW$ reference state instead of the Hartree-Fock starting point employed here. For the $GW$ reference state, the interaction kernel is given by the functional derivative of the $GW$ self-energy with respect to the Green's function. This makes the interaction kernel more complicated. Specifically, (i) $W$ is frequency dependent, and (ii) since $W$ itself depends on the Green's function, the derivatives of $W$ with respect to $G$ should be accounted for. The resulting diagrams are sometimes refereed to as ``butterfly'' diagrams.
A common approximation is to neglect the dependence of $W$ on the Green's functions, and to approximate the frequency dependence of $W$ by an instantaneous interactions. The most common approximation for the interaction kernel is
\begin{equation}
\label{eq:BSEkern}
\mathrm{i} I(1,2,3,4) = - \delta(2,4) \delta(1,3) \delta(t_1, t_2) w(\mathbf{r}_1,\mathbf{r}_2),
\end{equation}
with $w(\mathbf{r}_1,\mathbf{r}_2)$ being the random phase approximation (RPA) screened Coulomb interaction at $\omega=0$.
These approximations are also commonly employed in solid state BSE codes \cite{Albrecht1998,Rohlfing2000,Fuchs2008,Sander2015,Ljungberg2015}, and our present $GW\Gamma$ code can be adopted to this case. We will report on such calculations in forthcoming publications.
\subsection{Bethe-Salpeter equation}
\label{sec:BSE}
As already emphasized above, we treat all the interactions as instantaneous in the present case. For the Hartree-Fock reference point this is exact, but
it constitutes an approximation, if one were to start from the self-energy of the $GW$ approximation.
In full generality, the BSE can be Fourier transformed into the frequency domain and its kernel becomes a function of three independent frequency variables $I=I(\omega,\omega',\tilde{\omega})$, with the screened interaction carrying a frequency dependence of $\omega'-\tilde{\omega}$ \cite{Romaniello2009a}.
Using an instantaneous interaction obviously yields a kernel $I$ that does not depend on the frequency,
since the Fourier transformation of a $\delta$-function in time is a constant in frequency space.
More detailed discussions can be found in literature \cite{Hanke1979,Hanke1980,Strinati1988}.
Then, in an orbital representation the full kernel $V +\mathrm{i} I$ is given by the matrix elements of the matrices $\mathbf{A'}$ and $\mathbf{B}$:
\begin{align}
A'_{iajb}=& \langle aj|V|ib \rangle - \langle aj|V|bi \rangle, \label{eq:RR}\\
B_{iajb}=& \langle ab|V|ij \rangle - \langle ab|V|ji \rangle.\label{eq:RA}
\end{align}
The BSE reduces to a generalised eigenvalue problem \cite{Fuchs2008}, where the eigenvalues $\Omega_{\lambda}$ correspond to the optical transition energies, and the matrix $\mathbf{A}$ is formed from $\mathbf{A'}$ adding the independent particle energy differences $\Delta E_{ia}=\epsilon_a - \epsilon_i$ to the diagonal elements:
\begin{align}
\label{eq:BSE}
\begin{pmatrix}
\mathbf{A} & \mathbf{B}\\
\mathbf{B}^* & \mathbf{A}^*
\end{pmatrix}
\left |
\begin{matrix}
\mathbf{X}_\lambda\\
\mathbf{Y}_\lambda
\end{matrix}
\right \rangle =
\Omega_{\lambda}
\left |
\begin{matrix}
\mathbf{X}_\lambda\\
\mathbf{Y}_\lambda
\end{matrix}
\right \rangle.
\end{align}
At the $\Gamma-$point the matrix elements of $\mathbf{A}$ and $\mathbf{B}$ can be chosen to be real-valued and since the matrix $\mathbf{A}$ is Hermitian and $\mathbf{B}$ is symmetric \cite{Onida2002}, the generalised eigenvalue problem (EVP) above can be recast as an Hermitian EVP, for which standard solvers are available.
A concise expression for the resulting polarisation propagator's spectral representation reads \cite{Ring1980,Maggio2016}:
\begin{align}
\label{eq:PiSR1}
\mathbf{\Pi} (\omega) = \mathbf{Z} \left ( \omega - \mathbf{\Omega} \right )^{-1} \mathbf{\Delta}\mathbf{Z}^\dagger,
\end{align}
with the following definitions for the matrices:
\begin{align}
\mathbf{Z}=
\begin{pmatrix}
\mathbf{X} & \mathbf{Y}^* \\
\mathbf{Y} & \mathbf{X}^*
\end{pmatrix},
\mathbf{\Delta}=
\begin{pmatrix}
\mathbf{1} & \mathbf{0} \\
\mathbf{0} & \mathbf{-1}
\end{pmatrix},
\mathbf{\Omega}=
\begin{pmatrix}
\mathrm{diag}\{\Omega_\lambda \} & \mathbf{0} \\
\mathbf{0} & \mathrm{diag}\{-\Omega_\lambda \}
\end{pmatrix}.
\end{align}
The matrix $\mathbf{Z}$ contains all the individual eigenvectors $\mathbf{X}_{\lambda}$ and $\mathbf{Y}_\lambda$, similarly the matrix $\mathbf{\Omega}$ includes resonant and antiresonant transition energies; the transition energies for the resonant and antiresonant branch are identical (apart from a sign) as a consequence of the downfolding of the original BSE into an Hermitian EVP.
For definitiveness we choose $\Omega_\lambda >0$.
Since the matrix $\omega - \mathbf{\Omega}$ is diagonal in the orbital representation, it can be commuted across; the resulting product $\mathbf{Z} \mathbf{\Delta} \mathbf{Z}^\dagger$ can be decomposed into the difference of two matrices:
\begin{align*}
\mathbf{Z \Delta Z}^\dagger =
\begin{pmatrix}
\mathbf{XX}^* & \mathbf{XY}^* \\
\mathbf{YX}^* & \mathbf{YY}^*
\end{pmatrix} -
\begin{pmatrix}
\mathbf{Y^*Y} & \mathbf{Y^*X} \\
\mathbf{X^*Y} & \mathbf{X^*X}
\end{pmatrix}.
\end{align*}
The resonant and antiresonant contributions can be more easily singled out by re-writing the matrices above as an external product of the eigenstates of \autoref{eq:BSE}.
The polarisation propagator then reads:
\begin{align}
\label{eq:PiSR2}
\Pi(\omega) = \sum_{\lambda} \frac{ \left |\begin{matrix} \mathbf{X}_\lambda \\ \mathbf{Y}_\lambda \end{matrix} \right \rangle \langle \mathbf{X}_\lambda^* \mathbf{Y}_\lambda^*| }{\omega - \Omega_\lambda} -
\frac{\left |\begin{matrix} \mathbf{Y}_\lambda^* \\ \mathbf{X}_\lambda^* \end{matrix} \right \rangle \langle \mathbf{Y}_\lambda \mathbf{X}_\lambda| }{\omega + \Omega_\lambda}.
\end{align}
This expression can now be inserted into \autoref{equ:GWGamma} to give the greater and lesser components of the correlation self-energy, corresponding to the propagation of a particle ($p$) and a hole ($h$), respectively:
\begin{align}
\label{eq:SigmaR}
\langle n | \Sigma_p (\omega) | n \rangle & =\sum_\lambda \langle n |\mathbf{V} |\mathbf{X}_\lambda + \mathbf{Y}_\lambda \rangle \langle \mathbf{AX}_\lambda + \mathbf{BY}_\lambda| n \rangle \left ( f(\epsilon_n) -1 \right ) \delta \left ( \epsilon_n + \Omega_\lambda - \omega \right ) \\
\label{eq:SigmaA}
\langle n | \Sigma_h (\omega)| n \rangle & =\sum_\lambda \langle n |\mathbf{V} |\mathbf{X}_\lambda + \mathbf{Y}_\lambda \rangle \langle \mathbf{AY}_\lambda + \mathbf{BX}_\lambda| n \rangle f(\epsilon_n) \delta \left ( \epsilon_n - \Omega_\lambda - \omega \right ).
\end{align}
$f(\epsilon_n)$ in the previous equations is the occupancy for the energy level $\epsilon_n$ and the summation goes over the set of particle-hole excitations $\lambda$.
The composite index $\lambda$ is constructed by considering all possible combinations of the single particle indices $(i,a)$, with $i$ and $a$ belonging to the occupied and unoccupied orbital manifold; in the following $n$ represents a generic orbital.
The total correlation part of the self-energy is then obtained by taking the Hilbert transform of the two components:
\begin{align}
\label{eq:Sigmac}
\langle n | \Sigma_c (\omega) | n \rangle & = \int \frac{d\omega'}{2} \frac{\langle n | \Sigma_p (\omega') - \Sigma_h (\omega')| n\rangle }{\omega - \omega' +\mathrm{i} \eta ~\mathrm{sgn}(\omega - \mu)}.
\end{align}
Expressions in Eqs. (\ref{eq:SigmaR}) and (\ref{eq:SigmaA}) reduce to the usual RPA limit \cite{Tiago2006}, if the exchange term in $\mathbf{A'}$ and $\mathbf{B}$ is omitted.
\subsection{Diagrammatic analysis}
\label{sec:2bleCounting}
We start our diagrammatic analysis with the lowest order expressions: these can be obtained by replacing the fully interacting polarisation propagator $\Pi$ in \autoref{eq:PiBSE} with the non-interacting counterpart $L_0$.
The corresponding diagrams are shown in \autoref{Fig:SE_2ndO}.
In the top row of \autoref{Fig:SE_2ndO} we are reporting the correlation contributions in \autoref{equ:GWGamma} that stem from the term labelled with $V$, in the bottom row the terms generated by the kernel $\mathrm{i} I$ are shown.
The two columns portray the different orientations of the intermediate state $n$ with respect to the incoming state $n'$ (which in the diagram is represented by the external legs which have not been amputated for clarity): in the left column $n$ and $n'$ belong to the same manifold (they are both either occupied or unoccupied), in the right column $n$ and $n'$ belong to different manifolds.
It is obvious, by inspecting the diagrams, that the two contributions (in the top and bottom row, called respectively bubble and exchange in the following) are non equivalent in lowest order and that by cutting a particle-hole pair the diagram does not break up into disjoint pieces.
This lower order contribution results in the MP2 self-energy, evaluated for instance in Ref. \cite{Gruneis2014a} for extended systems.
\begin {figure}
\centering
\includegraphics [width=7.5cm] {SE2ndOrder.jpg}
\caption{Self-energy diagrams for the lowest order approximation; these are obtained by replacing $\Pi$ with $L_0$ in \autoref{equ:GWGamma}. Wavy lines represent the bare Coulomb interaction. The direction of time is fixed by the thick arrow and it is understood in the following diagrams.}
\label{Fig:SE_2ndO}
\end{figure}
If we now switch on the interaction between particles and holes in the polarisation propagator we obtain additional classes of diagrams, as shown in \autoref{Fig:SE_3rdO} in third order in the interaction, \textit{i.e.} setting $\Pi=L_0 (V+\mathrm{i} I) L_0$.
Two of the third order diagrams belong to the same class as their second order counterparts, these are shown in the left column of \autoref{Fig:SE_3rdO}.
Additionally, the "mixed contributions" shown in the right column represent the insertion of an exchange interaction in the bubble term (top) and the insertion of a bubble in the exchange diagram (bottom).
The polarisation propagator, however, is obtained by a resummation to infinite order of the corresponding diagrams \textit{via} the solution of the BSE.
These are represented in \autoref{Fig:SE_infO}, where we are showing both orientations for the intermediate particle for completeness.
Starting from the top left diagram in \autoref{Fig:SE_3rdO} it is easy to identify an infinite resummation of this class with the usual RPA approximation, where the so-called bubble diagrams are summed up; this translates into a screened interaction (denoted by a double wavy line) showing up in the corresponding diagram in \autoref{Fig:SE_infO}.
Likewise, an infinite number of bubble insertions in the exchange diagram of \autoref{Fig:SE_3rdO} (bottom right) returns an additional screened line.
The infinite resummation of the exchange contributions (top right and bottom left in \autoref{Fig:SE_3rdO}), on the other hand, takes into account the interaction between particles and holes in the polarisation propagator and this is represented in \autoref{Fig:SE_infO} by a shaded triangle.
We should now point out that in the final self-energy diagrams, the resummation to infinite order of the polarisation propagator generates a frequency dependent, dynamical screened interaction vertex (pictorially defined at the bottom of \autoref{Fig:SE_infO}).
Considering the resulting diagrams it is easy to see that they result in proper, irreducible self-energy insertions: the presence of bubble and exchange contributions in the kernel assures that two topologically distinct classes are obtained (shown in the top and middle row of \autoref{Fig:SE_infO}) and that for either class there is no external polarisation insertion.
As it was mentioned in the introduction, the iteration to self-consistency modifies the diagrammatic structure presented up to this point (see \autoref{Fig:Hedin}-a). This is one reason why we do not consider self-consistency in the present work. As explained at the end of Sec. \ref{sec:Hedin}, by adopting some further
approximations, it is however possible to maintain the simple algebraic structure and extend the present
approach to $GW\Gamma$
calculations on top of a $GW$ reference state.
\begin {figure}
\centering
\includegraphics [width=7.5cm] {SE3rdOrder.jpg}
\caption{Self-energy diagrams for the third order contributions, only diagrams corresponding to states $n, n'$ belonging to the same manifold are shown. Wavy lines represent the bare Coulomb interaction.}
\label{Fig:SE_3rdO}
\end{figure}
\begin {figure}
\centering
\includegraphics [width=9.0cm] {SEtriangleAndVertex.jpg}
\caption{(Top) Self-energy diagrams for the infinite order resummation. A pictorial definition for the renormalised interaction lines (shaded triangles and double wavy lines) is given in the bottom line; these terms contain an infinite number of particle-hole interaction lines summed up by solving the BSE.}
\label{Fig:SE_infO}
\end{figure}
Finally, we mention that the exclusion of the exchange diagrams in the self-energy can be obtained by formally replacing the electron in the intermediate state $n$ with a classical test-charge.
In what follows we will refer to this level of theory as to the $GW^{\rm{tc-tc}}$ approximation.
This approximation neglects the $\mathrm{i} I$ term in \autoref{equ:GWGamma}, but includes it in \autoref{eq:PiBSE}.
\section{Computational procedure}
\label{sec:Comp}
Calculations were performed on a subset of molecules studied in Ref. \cite{Bruneval2013}. For consistency we adopted the same molecular geometries as obtained in Ref.\cite{Curtiss1998} at the MP2/6-31G(d) level of theory.
The initial step was always a self-consistent Hartree-Fock calculation. Since the solution of the Bethe-Salpeter equation is very demanding if many virtual orbitals are included, we used natural orbitals to represent the unoccupied states.
In our procedure we roughly followed the work of Gr\"uneis \textit{et al.} \cite{Gruneis2014a}, but instead of using the MP2 density matrix, we calculated the RPA density matrix using the cubically scaling $GW$ code \cite{Liu2016}.
We then diagonalized the virtual-virtual sub-block and sorted the natural orbitals by their occupancy.
In this step, the number of orbitals is reduced by a factor four compared to the plane wave basis set size specified by the cutoff $E_c$ and reported in \autoref{tab:G0W0}.
We note that the occupied ($N_{\rm{occ}}$) as well as few unoccupied states are not modified but kept fixed at the level of the HF canonical orbitals.
The number of unoccupied orbitals that are kept fixed at their Hartree-Fock level is set to 0.8$\times N_{\rm{occ}}$.
By comparing calculations using all canonical orbitals and natural orbitals for selected systems, we found that the reduction of the unoccupied states introduces errors that are smaller than about 10~meV.
After determining the natural orbitals, the Hartree-Fock Hamiltonian is again diagonalized using the natural orbitals as basis \footnote{in VASP this is performed by selecting \texttt{ALGO=SUBROT}}.
Using this basis, the Bethe-Salpeter equation is solved for the polarization propagator, and the diagonal elements of the self-energy are calculated in the orbital basis.
The self-energy is evaluated on the real frequency axis sampled with a uniform grid of 0.1 eV spacing.
The complex broadening is set to twice the frequency spacing and the resulting self-energy $\langle n |\Sigma_c(\omega) |n \rangle$ (shifted by the kinetic, ionic, Hartree and Fock exchange energy contributions) is piecewise linearly interpolated in each interval to determine the quasi-particle energies as the intersection with the bisector line as shown in \autoref{Fig:SE_PH3_CO}.
The quasi-particle shifts ($\delta(\varepsilon^{\rm{QP}})$ in the following) are then obtained by subtracting the corresponding mean-field single particle energies.
To minimise the impact of image charges due to the finite simulation box, the Hartree-Fock calculations were repeated increasing the simulation cell up to 25 {\AA} in each linear dimension and then correcting for the residual local potential present at the edge of the cell.
The quasi-particle calculations were performed in a smaller simulation box (8 {\AA} in size).
The volume dependence of the quasi-particle energies has been carefully analysed in our previous publication \cite{GW100vasp} and it was found to be neglible for the systems considered here.
The quality of this approximation can be assessed for the case of $G_0W_0$ for the HF reference state by comparison with the localised basis set calculations in Ref. \cite{Bruneval2013} and reported in \autoref{tab:G0W0}.
The final quasi-particle energies were obtained by adding to the cell size converged (HF) single particle energies the quasi-particle shifts and the basis set corrections, according to the expression:
\begin{align}
\label{eq:eQP}
\left . \varepsilon_n^{\rm{QP}} \right |_\infty = \left . \varepsilon_n^{\rm{HF}}\right |_N + \delta(\varepsilon_n^{\rm{QP}} ) + C_N(\varepsilon_n^{\rm{QP}}).
\end{align}
The basis set correction $C_N$ is defined as: $C_N(\varepsilon)=\left . \varepsilon \right |_\infty- \left . \varepsilon \right |_N$, where the subscripts indicate the number of natural orbitals ($N$) or the extrapolated value ($\infty$).
In the next section it will be shown how the interplay of quasi-particle energy shift and the basis set correction affects the final quasi-particle levels.
One of the advantages of plane wave basis sets is the full control of the basis set completeness by specifying their kinetic energy cutoff $E_c$. \footnote{in VASP this specified by the flag \texttt{ENCUT}}
To estimate the basis set convergence, we increased $E_c$ by a factor 1.3 and 1.5 beyond the value specified in \autoref{tab:G0W0} and extrapolated assuming that the QP shifts converge like one over the basis set size \cite{Klimes2014b}.
\begin{figure}
\centering
\includegraphics [width=7.5cm] {SEHOMO.png}
\caption{Self-energy for the HOMO level of $\rm{PH_3} $ (blue lines) and CO (red lines) evaluated at the $G_0W_0$ (dashed lines), $G_0W_0^{\rm{tc-tc}} $ (solid lines) and $G_0W_0\Gamma$ (dashed-dotted lines) level. The solid black line is the bisector line of equation $y=\omega$. }
\label{Fig:SE_PH3_CO}
\end{figure}
\section{Results and discussion}
\label{sec:Res}
In this section, we apply the computational scheme presented above to a subset of the molecules considered in Refs. \cite{Bruneval2013,Curtiss1998}.
As a first check on the accuracy of the computational procedure we compare in \autoref{tab:G0W0} the HOMO quasi-particle energies at the $G_0W_0$ level of theory with those obtained by Bruneval and Marques with Gaussian type orbitals (GTOs) using cc-pVQZ basis sets \cite{Bruneval2013}.
We immediately point out that we expect differences of the order of 100 meV between PW and GTO $GW$ results. First, pseudopotentials can introduce errors and second the GTO results were not extrapolated to the complete basis set limit.
Concerning the pseudpotential error, we know from previous work that flourine containing compounds are affected most strongly with errors of about $\approx$100~meV, owing to the strong degree of localisation for 2$p$ electrons. With the non-normconserving PAW potential used here, we somewhat underestimate the degree of localization for flourine, as explained elsewhere \cite{GW100vasp}. Concerning the basis set extrapolation of the GTO results,
its convergence was studied in Ref. \cite{Bruneval2013} up to the "correlation-consistent" pV5Z basis set for the specific case of CO; the HOMO energy difference between this basis set and the quadruple zeta employed for the calculations of all the remaining molecules is about $\sim$70 meV.
Since the error decays like $1/C_n^3$, where $C_n$ is the cardinal number (\textit{i.e.} four for quadruple basis sets), the quadruple zeta basis set results reported in Ref. \cite{Bruneval2013} are expected to have errors of the order of $2 \times 70~$meV$\approx 140~$meV.
The difference between our plane-wave (PW) results and the GTO ones is typically comparable with this order of magnitude, with the exception of two outliers: CS and P$_{\text{2}}$.
For CS we observe that in going from the HF to the $G_0W_0$ level there is a rearrangement of the orbital energies, with the HOMO-1 and the HOMO exchanging their positions.
We can therefore speculate that the value reported in the literature refers to a lower lying quasi-particle level that corresponds to the HOMO orbital at the mean-field level.
For this reason we have included also this quasi-particle energy in \autoref{tab:G0W0} for comparison.
On the other hand, for P$_{\text{2}}$, we notice how the finite box effects are particularly strong in this case, with a downward shift exceeding 1 eV for the HF HOMO level as the cell edge is increased from 8 to 25 \AA.
It is, however difficult to draw a solid conclusion on the origin of the observed mismatch, also considering that in Ref. \cite{Bruneval2013} P$_{\text{2}}$ fails to follow any of the trends observed for the remaining molecules.
\begin{table*}
\caption{ $G_0W_0$ quasi-particle energies for the highest occupied orbital (HOMO) evaluated on a Hartree-Fock reference state.
In the second column the smallest plane-wave cutoff $E_c$ (and the corresponding number of plane waves $N_{\rm{pw}}$) used in the extrapolation is given.
The last column reports the negative of the experimental ionisation potential (IP), vertical values are in italics.
Plane wave extrapolated results include also finite box corrections as specified in \autoref{sec:Comp}.
All values in eV.}
\label{tab:G0W0}
\begin{tabular}{l c l c c c c c }
\hline \hline
molecule & $E_c $ & ($N_{\rm{pw}}$) & $G_0W_0 $ & $G_0W_0$ \cite{Bruneval2013} & $\Delta$ & CCSD(T) \cite{Bruneval2013} & Expt. \\
& & & PW & GTO & PW $-$ GTO & GTO & \\
& & & & cc-pVQZ & & cc-pVQZ & \\
\hline
H$_{\text{2}}$ & 300.0 & (8620) & -16.72 & --- & --- & -16.39 \cite{Koval2014} & -15.43 \cite{McCormack1989} \\
Li$_{\text{2}}$ & 112.1 & (1940) & -5.29 & -5.36 & 0.07 & -5.17 & -4.73 \cite{Dugourd1992} \\
N$_{\text{2}}$ & 420.9 & (10060) & -16.56 & -16.48 & -0.08 & -15.49 & -15.58 \cite{Trickl1989} \\
P$_{\text{2}}$ & 255.0 & (9180) & -11.31 & -10.57 & -0.74 & -10.76 & -10.62 \cite{Bulgin1976} \\
Cl$_{\text{2}}$ & 262.5 & (9780) & -12.08 & -12.01 & -0.07 & -11.62 & {\em -11.49} \cite{Dyke1984} \\
HF & 487.7 & (12540) & -16.29 & -16.39 & 0.10 & -16.09 & {\em -16.12} \cite{Banna1975} \\
LiH & 300.0 & (8620) & -8.26 & -8.2 & -0.06 & -7.94 & -7.90 \cite{NIST2015} \\
NaCl & 262.5 & (7080) & -9.51 & -9.36 & -0.15 & -9.13 & {\em -9.80} \cite{Potts1977} \\
ClF & 487.7 & (12540) & -13.49 & -13.32 & -0.17 & -12.82 & {\em -12.77} \cite{Dyke1984} \\
CH$_{\text{4}}$ & 414.0 & (9800) & -14.95 & -14.86 & -0.09 & -14.4 & {\em -13.6} \cite{Bieri1980} \\
C$_{\text{2}}$H$_{\text{4}}$ & 414.0 & (9800) & -10.91 & -10.85 & -0.06 & -10.69 & {\em -10.68} \cite{Bieri1980} \\
C$_{\text{2}}$H$_{\text{2}}$ & 414.0 & (9800) & -11.73 & -11.65 & -0.08 & -11.42 & {\em -11.49} \cite{Bieri1980} \\
CH$_{\text{3}}$Cl & 414.0 & (9800) & -11.90 & -11.74 & -0.16 & -11.41 & {\em -11.29} \cite{Kimura1981} \\
CH$_{\text{3}}$OH & 434.4 & (10460) & -11.71 & -11.69 & -0.02 & -11.08 & {\em -10.96} \cite{Vorobev1989} \\
CH$_{\text{3}}$SH & 414.0 & (9800) & -9.93 & -9.81 & -0.12 & -9.49 & {\em -9.44} \cite{Cradock1972} \\
SiH$_{\text{4}}$ & 300.0 & (8620) & -13.40 & -13.31 & -0.09 & -12.82 & {\em -12.3} \cite{Roberge1978} \\
CO & 434.4 & (10460) & -15.03 & -14.97 & -0.06 & -14.05 & {\em -14.01} \cite{Potts1974} \\
CO$_{\text{2}}$ & 434.4 & (10460) & -14.35 & -14.38 & 0.03 & -13.78 & {\em -13.77} \cite{Eland1977} \\
SiO & 434.4 & (10460) & -12.05 & -11.98 & -0.07 & -11.55 & -11.3 \cite{Nakasgawa1981} \\
SO$_{\text{2}}$ & 434.4 & (20380) & -13.20 & -13.12 & -0.08 & -12.41 & {\em -12.50} \cite{Kimura1981} \\
CS & 414.0 & (9800) & -12.63 & -13.08 & 0.45 & -11.45 & -11.33 \cite{King1972} \\
CS (*) & & & -13.19 & -13.08 & -0.11 & -11.45 & -11.33 \cite{King1972} \\
H$_{\text{2}}$O & 434.4 & (10460) & -13.10 & -13.04 & -0.06 & -12.64 & {\em -12.62} \cite{Kimura1981} \\
H$_{\text{2}}$O$_{\text{2}}$ & 434.4 & (10460) & -12.12 & -12.13 & 0.01 & -11.49 & {\em -11.70} \cite{Ashmore1977} \\
H$_{\text{2}}$S & 300.0 & (6040) & -10.79 & -10.67 & -0.12 & -10.43 & {\em -10.50} \cite{Bieri1982} \\
HClO & 434.4 & (10460) & -11.96 & -11.83 & -0.13 & -11.3 & -11.12 \cite{Colbourne1978} \\
HCN & 420.9 & (10060) & -13.92 & -13.86 & -0.06 & -13.64 & {\em -13.61} \cite{Kreile1982} \\
NH$_{\text{3}}$ & 420.9 & (10060) & -11.45 & -11.38 & -0.07 & -10.92 & {\em -10.82} \cite{Baumgaertel1989} \\
N$_{\text{2}}$H$_{\text{4}}$ & 420.9 & (10060) & -10.86 & -10.78 & -0.08 & -10.24 & {\em -8.98} \cite{Vovna1975} \\
PH$_{\text{3}}$ & 300.0 & (6040) & -10.89 & -10.79 & -0.10 & -10.49 & {\em -10.59} \cite{Cowley1982} \\
\hline \hline
\end{tabular}
(*) quasi-particle value corresponds to Hartree-Fock HOMO level.
\end{table*}
We then consider the impact of the vertex on the self-energy and on the quasi-particle energy for the HOMO level. This level is defined as the highest occupied orbital in the (post-)$GW$ approximation.
This coincides with the HF HOMO level for all molecular systems except CS and N$_{\text{2}}$, for which it corresponds to the HOMO-1 level in the HF approximation.
Furthermore, the inclusion of the ladder diagrams only in the polarisation propagator produces the so-called $G_0W_0$ with test charge-test charge interactions (labelled $G_0W_0^{\rm{tc-tc}}$ in the following).
In terms of the diagrams involved, the latter approximation can be obtained from the more general $G_0W_0\Gamma$ by omitting the exchange diagrams in the self-energy (\textit{e.g.} those in the bottom row in Figs. \ref{Fig:SE_2ndO} - \ref{Fig:SE_infO} above).
For this level of theory we observe the same level alignment between the quasi-particle HOMO level and HF reference as in the $G_0W_0\Gamma$ calculations, with the sole exception of SiO.
For this molecule there is a level crossing as in the case of CS and N$_{\text{2}}$.
The inclusion of the ladder diagrams in the polarizability is reasonable from a physical standpoint: it introduces an interaction between the "virtual" particle and hole that are generated as a response of the many-body system to the introduction of a test charge (excitonic effects).
In turn, ladder diagrams increase the screening in the many-body system, since an interacting electron-hole pair requires less energy to be generated than a non-interacting pair.
This effect will be present in both the $G_0W_0\Gamma$ and the $G_0W_0^{\rm{tc-tc}}$, and it shows on the resulting self-energy as a shift of its resonances to higher (lower) energies below (above) the Fermi level (conventionally located mid-gap between the HOMO and LUMO levels).
\autoref{Fig:SE_PH3_CO} contrasts the HOMO self-energy for all approximations considered in two representative systems (PH$_\text{3}$ and CO): in both cases, there is a reduction of the frequency interval devoid of resonances when the polarizability includes ladder diagrams.
Consistently the intersection with the bisector line $y=\omega$ occurs in the $G_0W_0^{\rm{tc-tc}}$ approximation at higher (less negative) energies, thus reducing the underscreening error of $G_0W_0$, as shown in \autoref{tab:GWGamma}.
Not surprisingly, the most sizable changes occur for molecules whose highest occupied orbital (HOMO) has a non-bonding or anti-bonding character, \textit{i.e.} that has a larger overlap with the LUMO and higher unoccupied states.
For instance: H$_\text{2}$O, H$_\text{2}$O$_\text{2}$, HF, CO$_\text{2}$, NH$_\text{3}$, have quasi-particle $G_0W_0^{\rm{tc-tc}}$ shifts more positive than the $G_0W_0$ values by more than 600 meV as reported in \autoref{Fig:varQP}.
\begin{table*}
\caption{HOMO quasi-particle energies on Hartree-Fock reference for all levels of theory considered.
Finite box corrections are included for all calculations as specified in \autoref{sec:Comp}.
All values in eV.
}
\label{tab:GWGamma}
\begin{tabularx}{\textwidth} {@{}l *8{>{\centering\arraybackslash}X}@{}}
\hline \hline
molecule & \multicolumn{4}{c}{finite basis set} & \multicolumn{3}{c}{extrapolated results} \\
& HF & $G_0W_0$ & $G_0W_0^{\rm{tc-tc}}$ & $G_0W_0\Gamma$ & $G_0W_0$ & $G_0W_0^{\rm{tc}}$ & $G_0W_0\Gamma$ \\
\hline
H$_{\text{2}}$ & -16.17 & -16.34 & -16.00 & -16.25 & -16.72 & -16.40 & -16.52 \\
Li$_{\text{2}}$ & -4.90 & -5.26 & -5.00 & -5.16 & -5.29 & -5.05 & --- \\
N$_{\text{2}}$ & -17.16 & -16.12 & -15.45 & -16.06 & -16.56 & -15.92 & -16.39 \\
P$_{\text{2}}$ & -10.61 & -11.10 & -10.93 & -10.92 & -11.31 & --- & --- \\
Cl$_{\text{2}}$ & -12.06 & -11.65 & -11.25 & -11.52 & -12.08 & -11.71 & -11.80 \\
HF & -17.63 & -15.83 & -14.98 & -15.72 & -16.29 & -15.43 & -16.18 \\
LiH & -8.14 & -8.12 & -7.39 & -7.87 & -8.26 & -7.50 & -7.94 \\
NaCl & -9.57 & -9.14 & -8.65 & -9.06 & -9.51 & -9.06 & -9.32 \\
ClF & -13.50 & -13.21 & -12.78 & -13.14 & -13.49 & -13.06 & -13.33 \\
CH$_{\text{4}}$ & -14.80 & -14.65 & -14.24 & -14.37 & -14.95 & -14.55 & -14.57 \\
C$_{\text{2}}$H$_{\text{4}}$ & -10.23 & -10.69 & -10.48 & -10.49 & -10.91 & -10.71 & -10.66 \\
C$_{\text{2}}$H$_{\text{2}}$ & -11.09 & -11.50 & -11.25 & -11.26 & -11.73 & -11.49 & -11.43 \\
CH$_{\text{3}}$Cl & -11.82 & -11.60 & -11.18 & -11.37 & -11.90 & -11.50 & -11.56 \\
CH$_{\text{3}}$OH & -12.27 & -11.30 & -10.68 & -11.06 & -11.71 & -11.10 & -11.39 \\
CH$_{\text{3}}$SH & -9.68 & -9.72 & -9.37 & -9.62 & -9.93 & -9.60 & -9.76 \\
SiH$_{\text{4}}$ & -13.17 & -13.12 & -12.70 & -12.94 & -13.40 & -12.98 & -12.88 \\
CO & -15.15 & -14.77 & -14.28 & -14.71 & -15.03 & -14.55 & -14.89 \\
CO$_{\text{2}}$ & -14.77 & -13.94 & -13.33 & -13.83 & -14.35 & -13.75 & -14.16 \\
SiO & -12.65 & -11.72 & -10.99 & -11.63 & -12.05 & -11.30 & -11.78 \\
SO$_{\text{2}}$ & -13.56 & -12.83 & -12.31 & -12.64 & -13.20 & --- & --- \\
CS & -12.78 & -12.37 & -11.88 & -12.29 & -12.63 & -12.14 & -12.47 \\
H$_{\text{2}}$O & -13.85 & -12.61 & -11.87 & -12.55 & -13.10 & -12.30 & -12.92 \\
H$_{\text{2}}$O$_{\text{2}}$ & -13.16 & -11.63 & -10.81 & -11.39 & -12.12 & -11.31 & -11.81 \\
H$_{\text{2}}$S & -10.45 & -10.51 & -10.18 & -10.39 & -10.79 & -10.47 & -10.60 \\
HClO & -12.14 & -11.63 & -11.12 & -11.41 & -11.96 & -11.48 & -11.66 \\
HCN & -13.31 & -13.65 & -13.34 & -13.40 & -13.92 & -13.64 & -13.61 \\
NH$_{\text{3}}$ & -11.70 & -11.10 & -10.51 & -11.04 & -11.45 & -10.86 & -11.28 \\
N$_{\text{2}}$H$_{\text{4}}$ & -11.16 & -10.47 & -9.90 & -10.28 & -10.86 & -10.31 & -10.58 \\
PH$_{\text{3}}$ & -10.45 & -10.64 & -10.37 & -10.47 & -10.89 & -10.62 & -10.66 \\
MAE \textit{vs} CCSD(T) & & & & & 0.46 $\pm$ 0.16& 0.15$\pm$ 0.15 & 0.24$\pm$ 0.19 \\
MSD \textit{vs} CCSD(T) & & & & & -0.46 & 0.06 & -0.24 \\
\hline \hline
\end{tabularx}
\end{table*}
On the other hand, in the $G_0W_0\Gamma$ approximation (shown in \autoref{Fig:SE_PH3_CO} by dashed-dotted lines) there is another effect at play: the inclusion of exchange diagrams in the self-energy redistributes the spectral weight of its resonances across the frequency range.
This effect, for the vast majority of systems considered here (the only exceptions being the unsaturated hydrocarbons C$_{\text{2}}$H$_{\text{4}}$ and C$_{\text{2}}$H$_{\text{2}}$), counteracts the upshift of $G_0W_0^{\rm{tc-tc}}$ in comparison with $G_0W_0$ quasi-particle energies.
In particular the effect is more pronounced for the set of molecules that showed larger quasi-particle shifts from $G_0W_0$ to $G_0W_0^{\rm{tc-tc}}$ (see \autoref{Fig:varQP}).
This comparison can be systematically carried out as shown in \autoref{Fig:varQP} for the whole range of molecules considered: the impact of vertex corrections in the self-energy, \textit{i.e.} in going from the $G_0W_0^{\rm{tc-tc}}$ to the $G_0W_0\Gamma$ approximation, is system dependent.
In particular, systems with a more ionic character (for instance fluorides and chlorides) or with unsaturated chemical bonds (\textit{e.g.} N$_{\text{2}}$, CO, CO$_{\text{2}}$) will have the electron density rather localised around the more electronegative element or in the bonding region respectively.
As electrons become more localised correlations among them become pivotal and necessitate the inclusion of the vertex at the self-energy level for an accurate description.
Comparison with previous literature is not straightforward.
Our observations are in contrast with the well-established behaviour of the local vertex in the $GW \Lambda_{\rm{LDA}}$ approximation, which results in a rigid shift of the quasi-particle energies \cite{Morris2007,Hung2016,Delsole1994}, regardless of the system's details.
$GW\Gamma$ calculations have also been reported by Shirley and Martin \cite{Shirley1993} for neutral atoms and ions.
Specifically, it has been reported that in ionized atoms the importance of exchange effects in the polarisability exceeds that of correlation effects in the self-energy, whereas for (neutral) atoms correlation effects in the self-energy are more relevant. With the very diverse systems considered here, we can not substantiate this trend.
\begin{figure}
\centering
\includegraphics [width=14.0cm] {varQPshift.png}
\caption{Variation of the quasi-particle shift for the $G_0W_0^{\rm{tc-tc}}$ (red squares), $G_0W_0\Gamma$ (black diamonds). The reference quasi-particle shifts are taken to be the $G_0W_0$ values.
Numerical values (in eV) are provided in \autoref{tab:GWGamma}}
\label{Fig:varQP}
\end{figure}
Now we evaluate the basis set corrections for the various levels of theory.
We notice these values are consistenly smaller for $G_0W_0\Gamma$ than for the $G_0W_0^{\rm{tc-tc}}$ level of theory (see the finite basis quasi-particle levels and the extrapolated counterparts in \autoref{tab:GWGamma}).
This observation can be rationalised with a similar diagrammatic analysis as carried out for the paramagnetic electron gas \cite{Maggio2016}.
The argument can be concisely restated as follows: in lowest order there is a cancellation between the bubble and the exchange diagrams in the self-energy (see top and bottom row in \autoref{Fig:SE_2ndO}) because of the sign rule for fermionic loops that assigns a different sign to the two classes of diagrams.
This cancellation is only partial in second order because there are twice more ways to assign a spin variable to each propagator line in the bubble diagram than there are for the exchange diagram (this is a consequence of the spin conservation at each interaction line).
From these considerations the linear extrapolation of the quasi-particle energy will have a smaller slope by a factor $\frac{1}{2}$ if the exchange diagrams in the self-energy are included.
The inclusion of higher order diagrams clearly partially invalidates the previous arguments, since already in third order there are exchange type diagrams that feature the same number of fermionic loops as the bubble diagram (see \autoref{Fig:SE_3rdO} bottom right).
In general terms, we can however still expect a faster convergence over the number of orbitals for the $G_0W_0\Gamma$ level of theory in comparison with the $G_0W_0^{\rm{tc-tc}}$ case.
This turns out to be the case for the vast majority of the systems considered here.
The faster convergence for the $G_0W_0\Gamma$ results can be seen in \autoref{Fig:QP_extrap} for CO and PH$_\textrm{3}$.
\begin{figure}
\centering
\includegraphics [width=7.5cm] {fit.png}
\caption{Quasi-particle HOMO energies for CO (red) and $\textrm{PH}_\textrm{3}$ (blue) as a function of the basis set size.
The symbols correspond to the quasi-particle energies obtained with the default energy cutoff $E_c$ increased by a factor $n=$1.0, 1.3, 1.5 and, for PH$_\textrm{3}$, 2.0.
The energy zero has been set to the $G_0W_0$ HOMO energy (with $n=$1.0) for all levels of theory considered: $G_0W_0$ (circles), $G_0W_0^{\rm{tc-tc}}$ (squares) and $G_0W_0\Gamma$ (diamonds).
Linear extrapolations are shown as a guide to the eye.}
\label{Fig:QP_extrap}
\end{figure}
Finally, we can compare the resulting basis set extrapolated quasi-particle energies with the CCSD(T) results provided in \cite{Bruneval2013}.
We reiterate that this comparison is somewhat problematic, since basis set incompleteness errors were not included in the coupled cluster results, and they are estimated to be over the order of 100 to 150~meV, as already elaborated above.
Since CCSD(T) calculations also include exchange diagrams (as all interactions are antisymmetrised) we expect CCSD(T) to have a similar convergence rate as $G_0W_0\Gamma$ and not as $G_0W_0$.
Neglecting these corrections, it is clear that the standard $G_0W_0$@HF yields much too large IP's.
The $G_0W_0^{\rm{tc-tc}}$ approximation significantly improves on these results, with only few outliers with an error exceeding 300~meV.
These outliers are two polar molecules, with much too small IP, namely HF and LiH, as well as CS, CO and N$_\text{2}$ where the IP is still way too large.
However, the $G_0W_0^{\rm{tc-tc}}$ approximation is not a systematic approximation, since it includes vertex correction in the polarizability only but neglects them in the self-energy. If we keep in mind,
that basis set converged CCSD(T) calculations will result in more negative HOMO energies (about $-100$~meV to $-150$~meV), this could possibly
improve agreement between $G_0W_0\Gamma$ and CCSD(T) and worsen the $G_0W_0^{\rm{tc-tc}}$ results.
The effect of the vertex in the self-energy is to systematically bring the extrapolated IP again above the CCSD(T) reference values \autoref{Fig:QPvsCC}.
Generally, the $G_0W_0\Gamma$ IP's lie between the $G_0W_0$ and $G_0W_0^{\rm{tc-tc}}$ values, which is in agreement with the general assumptions in literature.
Exceptions to this trend are the unsaturated hydrocarbons, SiH$_\text{4}$ and HCN, whose IP's, accidentally, reproduce the CCSD(T) values quite well.
It is interesting to comment on the biggest outliers in \autoref{Fig:QPvsCC}. These are N$_\text{2}$, CO and CS.
A previous study \cite{Koval2014} has established the strong dependence on the used orbitals for carbon monoxide and the nitrogen dimer.
Indeed, if full self-consistency is achieved (with the update of the initial Hartree-Fock orbitals), the resulting IP's are reduced by an excess of 0.8 eV.
Unfortunately, similar studies have not been carried out for CS, however we expect a similar behaviour for this system, given the similarities in electronic structures.
Therefore, for these molecules we expect that an update of the orbitals, or at least of the orbital's energies, more than the inclusion of the vertex will be relevant for an estimate of their IP's.
The mean absolute deviation from the coupled cluster values for the remaining molecules in the set is 0.21 eV for $G_0W_0\Gamma$ and 0.15 eV for $G_0W_0^{\rm{tc-tc}}$, scoring a marked improvement in comparison with the $G_0W_0$ deviation (0.46 eV).
We believe that inclusion of basis set corrections in the CCSD(T) results would improve the agreement between $G_0W_0\Gamma$ and CCSD(T) even further, since this would increase the CCSD(T) IP values by the already quoted 100-150 meV moving them closer to the $G_0W_0\Gamma$.
Obviously this correction would worsen the agreement between CCSD(T) and $G_0W_0^{\rm{tc-tc}}$, which is to be expected, since $G_0W_0^{\rm{tc-tc}}$ is not a well balanced diagrammatic theory.
In Ref. \cite{Gallandi2016} range-separated density functionals have been used as a starting point of the $G_0W_0$ calculation, and these scored a similar improvement in quasi-particle energies as the one reported here by applying the vertex.
In some sense, this starting point is constructed to reproduce the system's IP, whereas the method we employ is completely \textit{ab initio} and, as such, does not require any previous information on the system's electronic structure.
Another well established class of \textit{ab initio} approaches is the algebraic diagrammatic construction (ADC). It offers a general framework to generate the self-energy diagrams at any order in perturbation theory.
The two-particle one-hole Tamm-Dancoff approximation (2ph-TDA) is the simplest ADC method and it includes bubble and particle-hole ladder diagrams in its resummation.
It is quite striking that, albeit this approximation is diagrammatically similar to the $GW\Gamma$ method, its performance for small molecules is unsatisfactory with a mean absolute error (\textit{vs} CCSD(T)) that exceeds 0.6 eV \cite{Corzo2015}.
The inclusion of all self-energy diagrams up to third order, as it is performed in the ADC(3) approximation, largely improves the IP's values.
It is important to point out that in the 2ph-TDA approach the 2p-2h virtual excitations are not included, whereas these effects are implicitly accounted for by the coupling amplitudes in the full ADC(3) \cite{VonNiessen1984}.
These excitations are responsible for introducing correlation effects in the HF ground state \cite{Ring1980} (which is obviously uncorrelated) and are explicitly built in the $GW$ and $GW\Gamma$ approximation thanks to the inclusion of the so-called resonant-antiresonant coupling.
The comparison between our results and those obtained with the electron propagator methods cited above allows us to rationalise the poor performance of the 2ph-TDA.
Specifically, the coupling between resonant and antiresonant virtual transistions is a key feature of $GW$ (and $GW\Gamma$) that is implicitly accounted for in ADC(3).
This coupling seems to be responsible for a dramatic improvement over the quasi-particle energies obtained at the 2ph-TDA level.
On the other hand, the inclusion of other higher order particle-particle ladders in the ADC(3) method appear to have an overall modest impact (with the exceptions discussed above that are poorly described by the HF reference), since its performance on small molecules is comparable to that of the $GW\Gamma$ method.
The comparison with experimental data (\autoref{Fig:QPvsCC}, bottom panel) proceeds along similar lines as above, with the mean absolute error increasing slightly owing to the impact of adiabatic effects on the measurements for a number of systems considered (these are reported in \autoref{tab:G0W0} as non-italicised entries).
\begin{figure}
\centering
\includegraphics [width=14.0cm] {IP.jpg}
\caption{Deviation from the experimental ionisation potential (top panel) and from the CCSD(T) estimate \cite{Bruneval2013} (bottom panel) for the first ionisation potential ($-\epsilon_{\rm{HOMO}}^{\rm{QP}}$) computed at the $G_0W_0$ (blue circles), $G_0W_0^{\rm{tc-tc}}$ (red squares) and $G_0W_0\Gamma$ (black diamonds) levels of theory. All computed values include basis set extrapolation corrections. Horizontal lines are centered at $\pm$0.3 eV.}
\label{Fig:QPvsCC}
\end{figure}
\section{Conclusions}
The main topic of this work is the inclusion of exchange diagrams in both the polarization propagator as well as the self-energy using Hedin's equations.
Commonly this is referred to as the $GW\Gamma$ approximation.
In Hedin's equations the interaction kernel is strictly given by the functional derivative of the self-energy employed in the preceding step with respect to the Green's function.
In the present work, we adopt the Hartree-Fock starting point.
Then from the derivative of the Hartree-term $V_\mathrm{H}$ with respect to the Green's function, one obtains the standard RPA bubble diagrams, and from the derivative of the exchange potential $\Sigma_\mathrm{x}=V_\mathrm{x}$ one obtains the ladder diagrams.
Hence, applying the common $GW$ approximation--- that includes bubbles only but neglects ladder diagrams ---is from the outset a flawed prescription for the Hartree-Fock starting point.
If it were successful, this could only be related to some sort of error cancellation.
An important aspect of the present work is whether such a cancellation exists.
To finally and concisely answer this question, we have performed quasi-particle calculations on top of the Hartree-Fock approximation for a set of 29 small molecules.
As already alluded to above, we have included the ladder diagrams related to the derivative of the exchange potential, in both the polarization propagator and the self-energy.
The applied equations have been derived from Hedin's equations and should be fairly easy to implement in any time-dependent Hartree-Fock, Casida or BSE code.
We have also argued that, although the final equations (\ref{eq:SigmaR}) essentially only rephrase the equation of motion for the Green's function, they have not yet been used in quantum chemistry or computational solid state physics.
This is somewhat astounding, since the implementation is, in principle, straightforward.
All one needs to know are the two-particle eigenstates of the TD-HF equation.
Now let us turn to our results.
The first, already often published, observation is that the $G_0W_0$ approximation is not satisfactory when performed on top of HF.
It systematically overestimates the IP's. This should come as no surprise.
In the $G_0W_0$ approximation, a non-physical polarizability is used, that is calculated without particle-hole ladder diagrams related to the variation of the exchange.
Inclusion of the ladder diagrams in the polarization propagator (which is equivalent to using the polarizability from TD-HF) largely rectifies this issue and massively improves agreement with the CCSD(T) reference values.
The mean absolute error decreases from about 400~meV in $G_0W_0$ to 150~meV in this $G_0W_0^{\mathrm{tc-tc}}$ approximation.
Including the exchange diagrams in the polarizability only is, however, not an entirely well balanced approximation instead, one should include the exchange diagrams also in the self-energy ($G_0W_0\Gamma$ approximation).
The $G_0W_0\Gamma$ values are always in between the standard $G_0W_0$ and the $G_0W_0^{\mathrm{tc-tc}}$ results, as discussed in more detail in the next paragraph.
With a mean absolute deviation of 210~meV compared to the CCSD(T) reference results, $G_0W_0\Gamma$ results are on average slightly worse than the $G_0W_0^{\mathrm{tc-tc}}$ (MAE 150~meV).
This is certainly somewhat unsatisfactory and requires further studies and, in particular, better reference values.
With an error of roughly 100~meV for cc-pVQZ basis sets, CCSD(T) values without basis set extrapolation are not sufficiently accurate to allow for an unambigious comparison.
In passing, we also emphasize a side result of our study, namely that $G_0W_0$ converges differently with respect to basis set size than $G_0W_0\Gamma$.
This is a result of the inclusion of the exchange diagrams in the self-energy as shown in Fig. \ref{Fig:QP_extrap}.
This also means that comparisons at finite basis sets (without basis set corrections) need to be done with caution, since one could observe fortuitous agreement at small basis sets that might not hold up at improved basis sets.
A key result of this study is that changes from $G_0W_0^{\mathrm{tc-tc}}$ to $G_0W_0\Gamma$ are very system-dependent: in some cases, $G_0W_0\Gamma$ recovers the $G_0W_0$ values, in other cases the final values are closer to $G_0W_0^{\mathrm{tc-tc}}$.
As an example, one can consider the case of ammonia and phosphine: they have a comparable geometry and valence electron configuration, with the obvious difference that phosphorus is one period below nitrogen in the periodic table.
In one case (PH$_{\text{3}}$) the introduction of the vertex in the self-energy has a marginal impact on the IP, whereas for NH$_{\text{3}}$ the higher electronegativity of the central atom, as discussed in the previous section, leads to increased electron correlations and to a more pronounced difference between $G_0W_0^{\mathrm{tc-tc}}$ and $G_0W_0\Gamma$.
Albeit these observations are in line with "chemical intuition", they are difficult to generalise, let alone if quantitative estimates are required.
We conclude that there is simply no shortcut to $GW\Gamma$ or general error cancellation; one needs to include the vertex consistently.
All these results are in agreement with previous work on solids, where on average the vertex in the self-energy also compensated the effect of the vertex in the polarization propagator, and as for molecules, the results varied significantly between different solids and even different orbitals \cite{Gruneis2014a}.
Finally, let us comment on the future perspectives: to us it is clear that the effect of self-consistency on the Green's function needs to be studied.
Unfortunately, this is not a simple matter.
If one updates the self-energy and the Green's function, additional terms come into play in the interaction kernel $I$, and ultimately one is then forced to make compromises between computational efficiency and accuracy \cite{Kutepov2016}.
We plan to report a combination of self-consistency and vertex corrections in our forthcoming study.
In summary, we have shown that $G_0W_0\Gamma$@HF improves upon $G_0W_0$@HF.
This implies that the vertex in Hedin's equations is important and can not be neglected.
Given the simplicity of the present approach, we are confident that it will draw interest in the computational quantum chemistry community, as it constitutes an important fundamental step towards accurate quasi-particle calculations.
\section*{Acknowledgements}
Funding by the Austrian Science Fund (FWF) within the Spezialforschungs-
bereich F41 (SFB ViCoM) is gratefully acknowledged.
Computations were predominantly performed on the Vienna Scientific Cluster VSC3.
| de615de54c185cd917cb57e15c11d1fc1780fb46 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Experimental Setup}
\subsection{Arcade Learning Environment (ALE)}
\label{sec:setup-ale-appendix}
For ALE \citep{ale}, the network architecture as described in~\cite{dqn} is used. This consists of $3$ convolutional layers ($32$ filters of size $8\times8$ and stride $4$, $64$ filters of size $4\times4$ and stride $2$, $64$ filters of size $3\times3$ and stride $1$) followed by $1$ hidden layer with $512$ units followed by a linear output layer with one unit for each action. ReLUs are used in each layer, while layer normalization~\citep{layernorm} is used in the fully connected part of the network.
For parameter space noise, we also include a second head after the convolutional stack of layers. This head determines a policy network with the same architecture as the $Q$-value network, except for a softmax output layer. The target networks are updated every $10\,\mathrm{K}$~timesteps. The $Q$-value network is trained using the Adam optimizer \citep{adam} with a learning rate of $10^{-4}$ and a batch size of $32$.
The replay buffer can hold $1\,\mathrm{M}$~state transitions.
For the $\epsilon$-greedy baseline, we linearly anneal $\epsilon$ from $1$ to $0.1$ over the first $1\,\mathrm{M}$~timesteps.
For parameter space noise, we adaptively scale the noise to have a similar effect in action space (see \autoref{sec:adapt-dqn} for details), effectively ensuring that the maximum KL divergence between perturbed and non-perturbed $\pi$ is softly enforced.
The policy is perturbed at the beginning of each episode and the standard deviation is adapted as described in \autoref{sec:adapt} every $50$ timesteps.
Notice that we only perturb the policy head after the convolutional part of the network (i.e. the fully connected part, which is also why we only include layer normalization in this part of the network).
To avoid getting stuck (which can potentially happen for a perturbed policy), we also use $\epsilon$-greedy action selection with $\epsilon=0.01$.
In all cases, we perform $50\,\mathrm{K}$~random actions to collect initial data for the replay buffer before training starts.
We set $\gamma=0.99$, clip rewards to be in $[-1, 1]$, and clip gradients for the output layer of $Q$ to be within $[-1, 1]$.
For observations, each frame is down-sampled to $84\times84$ pixels, after which it is converted to grayscale.
The actual observation to the network consists of a concatenation of $4$ subsequent frames.
Additionally, we use up to $30$ noop actions at the beginning of the episode.
This setup is identical to what is described by~\cite{dqn}.
\subsection{Continuous Control}
\label{sec:setup-continuous-appendix}
For DDPG, we use a similar network architecture as described by~\cite{ddpg}:
both the actor and critic use $2$ hidden layers with $64$ ReLU units each.
For the critic, actions are not included until the second hidden layer.
Layer normalization~\citep{layernorm} is applied to all layers.
The target networks are soft-updated with $\tau=0.001$.
The critic is trained with a learning rate of $10^{-3}$ while the actor uses a learning rate of $10^{-4}$.
Both actor and critic are updated using the Adam optimizer~\citep{adam} with batch sizes of $128$.
The critic is regularized using an $L2$ penalty with $10^{-2}$.
The replay buffer holds $100\,\mathrm{K}$ state transitions and $\gamma=0.99$ is used.
Each observation dimension is normalized by an online estimate of the mean and variance.
For parameter space noise with DDPG, we adaptively scale the noise to be comparable to the respective action space noise (see \autoref{sec:adapt-ddpg}).
For dense environments, we use action space noise with $\sigma=0.2$ (and a comparable adaptive noise scale).
Sparse environments use an action space noise with $\sigma=0.6$ (and a comparable adaptive noise scale).
TRPO uses a step size of $\delta_\text{KL}=0.01$, a policy network of $2$ hidden layers with $32$ $\tanh$ units for the nonlocomotion tasks, and $2$ hidden layers of $64$ $\tanh$ units for the locomotion tasks. The Hessian calculation is subsampled with a factor of $0.1$, $\gamma=0.99$, and the batch size per epoch is set to $5\,\mathrm{K}$~timesteps. The baseline is a learned linear transformation of the observations.
The following environments from OpenAI Gym\footnote{\url{https://github.com/openai/gym}}~\citep{gym} are used:
\begin{itemize}
\item \emph{HalfCheetah} (${\S \subset \mathbb{R}^{17}}$, ${\mathcal{A} \subset \mathbb{R}^{6}}$),
\item \emph{Hopper} (${\S \subset \mathbb{R}^{11}}$, ${\mathcal{A} \subset \mathbb{R}^{3}}$),
\item \emph{InvertedDoublePendulum} (${\S \subset \mathbb{R}^{11}}$, ${\mathcal{A} \subset \mathbb{R}}$),
\item \emph{InvertedPendulum} (${\S \subset \mathbb{R}^{4}}$, ${\mathcal{A} \subset \mathbb{R}}$),
\item \emph{Reacher} (${\S \subset \mathbb{R}^{11}}$, ${\mathcal{A} \subset \mathbb{R}^{2}}$),
\item \emph{Swimmer} (${\S \subset \mathbb{R}^{8}}$, ${\mathcal{A} \subset \mathbb{R}^{2}}$), and
\item \emph{Walker2D} (${\S \subset \mathbb{R}^{17}}$, ${\mathcal{A} \subset \mathbb{R}^{6}}$).
\end{itemize}
For the sparse tasks, we use the following environments from rllab\footnote{\url{https://github.com/openai/rllab}} \citep{duan2016benchmarking}, modified as described by \cite{vime}:
\begin{itemize}
\item \emph{SparseCartpoleSwingup} (${\S \subset \mathbb{R}^{4}}$, ${\mathcal{A} \subset \mathbb{R}}$), which only yields a reward if the paddle is raised above a given threshold,
\item \emph{SparseHalfCheetah} (${\S \subset \mathbb{R}^{17}}$, ${\mathcal{A} \subset \mathbb{R}^{6}}$), which only yields a reward if the agent crosses a distance threshold,
\item \emph{SparseMountainCar} (${\S \subset \mathbb{R}^{2}}$, ${\mathcal{A} \subset \mathbb{R}}$), which only yields a reward if the agent drives up the hill,
\item \emph{SparseDoublePendulum} (${\S \subset \mathbb{R}^{6}}$, ${\mathcal{A} \subset \mathbb{R}}$), which only yields a reward if the agent reaches the upright position, and
\item \emph{SwimmerGather} (${\S \subset \mathbb{R}^{33}}$, ${\mathcal{A} \subset \mathbb{R}^{2}}$), which yields a positive or negative reward upon reaching targets.
\end{itemize}
\subsection{Chain Environment}
\label{sec:setup-chain-appendix}
We follow the state encoding proposed by~\cite{bootstrappeddqn} and use $\phi(s_t) = (\mathbbm{1}\{x \leq s_t\})$ as the observation, where $\mathbbm{1}$ denotes the indicator function.
DQN is used with a very simple network to approximate the $Q$-value function that consists of $2$ hidden layers with $16$ ReLU units.
Layer normalization~\citep{layernorm} is used for all hidden layers before applying the nonlinearity.
Each agent is then trained for up to $2\,\mathrm{K}$ episodes.
The chain length $N$ is varied and for each $N$ three different seeds are trained and evaluated.
After each episode, the performance of the current policy is evaluated by sampling a trajectory with noise disabled (in the case of bootstrapped DQN, majority voting over all heads is performed).
The problem is considered solved if one hundred subsequent trajectories achieve the optimal episode return.
\autoref{fig:chain-env} depicts the environment.
\begin{figure}[h]
\centering
\begin{tikzpicture}[->,>=stealth,auto,node distance=2.2cm,
thick,node/.style={circle,draw}, minimum width=1.2cm,
scale=0.8, every node/.style={scale=0.8}]
\node[node, fill=red!40, label={[red!40]left:$r=0.001$}] (s1) {\large$s_1$};
\node[node, right of=s1] (s2) {\large$s_2$};
\node[node, draw=none, right of=s2] (si) {\large$\ldots$};
\node[node, right of=si] (sN1) {\large$s_{N-1}$};
\node[node, fill=red, text=white, right of=sN1, label={[red]right:$r=1$}] (sN) {\large$s_N$};
\path[->] (s1) edge [loop above] ();
\path[->] (s1) edge [bend left] (s2);
\path[->] (s2) edge [bend left] (si);
\path[->] (si) edge [bend left] (sN1);
\path[->] (sN1) edge [bend left] (sN);
\path[->] (sN) edge [loop above] ();
\path[->] (s2) edge [bend left] (s1);
\path[->] (si) edge [bend left] (s2);
\path[->] (sN1) edge [bend left] (si);
\path[->] (sN) edge [bend left] (sN1);
\end{tikzpicture}
\caption{Simple and scalable environment to test for exploratory behavior~\citep{bootstrappeddqn}.}
\label{fig:chain-env}
\end{figure}
We compare adaptive parameter space noise DQN, bootstrapped DQN~\citep{bootstrappeddqn} (with $K=20$~heads and Bernoulli masking with $p=0.5$), and $\epsilon$-greedy DQN (with $\epsilon$ linearly annealed from $1.0$ to $0.1$ over the first one hundred episodes).
For adaptive parameter space noise, we only use a single head and perturb $Q$ directly, which works well in this setting.
Parameter space noise is adaptively scaled so that $\delta \approx 0.05$.
In all cases, $\gamma=0.999$, the replay buffer holds $100\,\mathrm{K}$ state transitions, learning starts after $5$ initial episodes, the target network is updated every $100$ timesteps, and the network is trained using the Adam optimizer~\citep{adam} with a learning rate of $10^{-3}$ and a batch size of $32$.
\FloatBarrier
\section{Parameter Space Noise for On-policy Methods}
\label{sec:on-policy-appendix}
Policy gradient methods optimize
$\mathbb{E}_{\tau \sim (\pi, p)} [ R(\tau) ]$. Given a stochastic policy $\pi_\theta(a|s)$ with $\theta \sim \mathcal{N}(\phi, \Sigma)$, the expected return can be expanded using likelihood ratios and the reparametrization trick~\citep{kingma2013auto} as
\begin{align}
\nabla_{\phi,\Sigma} \mathbb{E}_\tau [ R(\tau) ]
& = \nabla_{\phi,\Sigma} \mathbb{E}_{\theta \sim \mathcal{N}(\phi,\Sigma)} \left[ \sum_\tau p(\tau|\theta) R(\tau) \right] \\
& = \mathbb{E}_{\epsilon \sim \mathcal{N}(0,I)} \nabla_{\phi,\Sigma} \left[ \sum_\tau p(\tau|\phi + \epsilon \Sigma^{\frac{1}{2}}) R(\tau) \right] \\
& = \mathbb{E}_{\epsilon \sim \mathcal{N}(0,I),\tau} \left[ \sum_{t=0}^{T-1} \nabla_{\phi,\Sigma} \log \pi(a_t|s_t;\phi+\epsilon \Sigma^{\frac{1}{2}}) R_t(\tau) \right] \\
& \approx \frac{1}{N} \sum_{\epsilon^i,\tau^i} \left[ \sum_{t=0}^{T-1} \nabla_{\phi,\Sigma} \log \pi(a_t|s_t;\phi+\epsilon^i \Sigma^{\frac{1}{2}}) R_t(\tau^i)\right]
\end{align}
for $N$ samples $\epsilon^i \sim \mathcal{N}(0,I)$ and $\tau^i \sim (\pi_{\phi + \epsilon^i \Sigma^{\frac{1}{2}}}, p)$, with $R_t(\tau^i) = \sum_{t'=t}^{T} \gamma^{t'-t} r_{t'}^i$. This also allows us to subtract a variance-reducing baseline $b_t^i$, leading to
\begin{equation}
\nabla_{\phi,\Sigma} \mathbb{E}_\tau [ R(\tau) ] \approx \frac{1}{N} \sum_{\epsilon^i,\tau^i} \left[ \sum_{t=0}^{T-1} \nabla_{\phi,\Sigma} \log \pi(a_t|s_t;\phi+\epsilon^i \Sigma^{\frac{1}{2}}) (R_t(\tau^i) - b_t^i) \right].
\end{equation}
In our case, we set $\Sigma := \sigma^2 I$ and use our proposed adaption method to re-scale as appropriate.
\FloatBarrier
\section{Adaptive Scaling}
\label{sec:adapt}
Parameter space noise requires us to pick a suitable scale $\sigma$.
This can be problematic since the scale will highly depend on the specific network architecture, and is likely to vary over time as parameters become more sensitive as learning progresses.
Additionally, while it is easy to intuitively grasp the scale of action space noise, it is far harder to understand the scale in parameter space.
We propose a simple solution that resolves all aforementioned limitations in an easy and straightforward way.
This is achieved by adapting the scale of the parameter space noise over time, thus using a time-varying scale $\sigma_k$.
Furthermore, $\sigma_k$ is related to the action space variance that it induces, and updated accordingly.
Concretely, we use the following simple heuristic to update $\sigma_k$ every $K$ timesteps:
\begin{equation}
\sigma_{k+1} =
\begin{cases}
\alpha \sigma_k, & \text{if } d(\pi, \widetilde{\pi}) < \delta \\
\frac{1}{\alpha} \sigma_k, & \text{otherwise},
\end{cases}
\end{equation}
where $d(\cdot, \cdot)$ denotes some distance between the non-perturbed and perturbed policy (thus measuring in action space), $\alpha \in \mathbb{R}_{>0}$ is used to rescale $\sigma_k$, and $\delta \in \mathbb{R}_{>0}$ denotes some threshold value.
This idea is based on the Levenberg-Marquardt heuristic~\citep{ranganathan2004levenberg}.
The concrete distance measure and appropriate choice of $\delta$ depends on the policy representation.
In the following sections, we outline our choice of $d(\cdot, \cdot)$ for methods that do (DDPG and TRPO) and do not (DQN) use behavioral policies.
In our experiments, we always use $\alpha = 1.01$.
\subsection{A Distance Measure for DQN}
\label{sec:adapt-dqn}
For DQN, the policy is defined implicitly by the $Q$-value function.
Unfortunately, this means that a naΓ―ve distance measure between $Q$ and $\widetilde{Q}$ has pitfalls.
For example, assume that the perturbed policy has only changed the bias of the final layer, thus adding a constant value to each action's $Q$-value.
In this case, a naΓ―ve distance measure like the norm $\lVert{Q - \widetilde{Q}}\rVert_2$ would be nonzero, although the policies $\pi$ and $\widetilde{\pi}$ (implied by $Q$ and $\widetilde{Q}$, respectively) are exactly equal.
This equally applies to the case where DQN as two heads, one for $Q$ and one for $\pi$.
We therefore use a probabilistic formulation\footnote{It is important to note that we use this probabilistic formulation only for the sake of defining a well-behaved distance measure. The actual policy used for rollouts is still deterministic.} for both the non-perturbed and perturbed policies: ${\pi, \widetilde{\pi}: \S \times \mathcal{A} \mapsto [0, 1]}$ by applying the softmax function over predicted $Q$ values: ${\pi(s) = \exp{Q_i(s)} / {\sum_i{\exp{Q_i(s)}}}}$, where $Q_i(\cdot)$ denotes the $Q$-value of the $i$-th action.
$\widetilde{\pi}$ is defined analogously but uses the perturbed $
\widetilde{Q}$ instead (or the perturbed head for $\pi$).
Using this probabilistic formulation of the policies, we can now measure the distance in action space:
\begin{equation}
d(\pi, \widetilde{\pi}) = D_{\text{KL}}\infdivx{\pi}{\widetilde{\pi}},
\end{equation}
where $D_{\text{KL}}\infdivx{\cdot}{\cdot}$ denotes the Kullback-Leibler (KL) divergence.
This formulation effectively normalizes the $Q$-values and therefore does not suffer from the problem previously outlined.
We can further relate this distance measure to $\epsilon$-greedy action space noise, which allows us to fairly compare the two approaches and also avoids the need to pick an additional hyperparameter $\delta$.
More concretely, the KL divergence between a greedy policy $\pi(s, a) = 1$ for $a = \text{argmax}_{a'} Q(s, a')$ and $\pi(s, a) = 0$ otherwise and an $\epsilon$-greedy policy $\widehat{\pi}(s, a) = 1 - \epsilon + \frac{\epsilon}{|\mathcal{A}|}$ for $a = \text{argmax}_{a'} Q(s, a')$ and $\widehat{\pi}(s, a) = \frac{\epsilon}{|\mathcal{A}|}$ otherwise is ${D_{\text{KL}}\infdivx{\pi}{\widehat{\pi}} = -\log{(1 - \epsilon + \frac{\epsilon}{|\mathcal{A}|})}}$, where $|\mathcal{A}|$ denotes the number of actions (this follows immediately from the definition of the KL divergence for discrete probability distributions).
We can use this distance measure to relate action space noise and parameter space noise to have similar distances, by adaptively scaling $\sigma$ so that it matches the KL divergence between greedy and $\epsilon$-greedy policy, thus setting $\delta := -\log{(1 - \epsilon + \frac{\epsilon}{|\mathcal{A}|})}$.
\subsection{A Distance Measure for DDPG}
\label{sec:adapt-ddpg}
For DDPG, we relate noise induced by parameter space perturbations to noise induced by additive Gaussian noise.
To do so, we use the following distance measure between the non-perturbed and perturbed policy:
\begin{equation}
d(\pi, \widetilde{\pi}) = \sqrt{\frac{1}{N} \sum_{i=1}^{N}{\mathbb{E}_{s}\left[\left(\pi(s)_i - \widetilde{\pi}(s)_i\right)^2 \right]}},
\end{equation}
where $\mathbb{E}_s[\cdot]$ is estimated from a batch of states from the replay buffer and $N$ denotes the dimension of the action space (i.e. $\mathcal{A} \subset \mathbb{R}^N$).
It is easy to show that $d(\pi, \pi + \mathcal{N}(0, \sigma^2 I)) = \sigma$.
Setting $\delta:=\sigma$ as the adaptive parameter space threshold thus results in effective action space noise that has the same standard deviation as regular Gaussian action space noise.
\subsection{A Distance Measure for TRPO}
\label{sec:adapt-trpo}
In order to scale the noise for TRPO, we adapt the sampled noise vectors $\epsilon \sigma$ by computing a natural step $H^{-1} \epsilon \sigma$. We essentially compute a trust region around the noise direction to ensure that the perturbed policy $\widetilde{\pi}$ remains sufficiently close to the non-perturbed version via
\begin{eqnarray*}
E_{s\sim \rho_{\widetilde{\theta}}}[D_{\text{KL}}(\pi_{\widetilde{\theta}}(\cdot |s) \| \pi_{\theta}(\cdot|s))] \leq \delta_{\text{KL}}.
\end{eqnarray*}
Concretely, this is computed through the conjugate gradient algorithm, combined with a line search along the noise direction to ensure constraint conformation, as described in Appendix C of \cite{trpo}.
\FloatBarrier
\section{Additional Results on ALE}
\label{sec:results-ale-appendix}
\autoref{fig:atari-full} provide the learning curves for all 21 Atari games.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{figures/atari_full}
\caption{Median DQN returns for all ALE environment plotted over training steps.}
\label{fig:atari-full}
\end{figure}
\autoref{table:full-comparison} compares the final performance of ES after $1\,000\,\mathrm{M}$ frames to the final performance of DQN with $\epsilon$-greedy exploration and parameter space noise exploration after $40\,\mathrm{M}$ frames.
In all cases, the performance is estimated by running $10$ episodes with exploration disabled.
We use the numbers reported by~\cite{es} for ES and report the median return across three seeds for DQN.
\begin{table}[h]
\caption{Performance comparison between Evolution Strategies (ES) as reported by \cite{es}, DQN with $\epsilon$-greedy, and DQN with parameter space noise (this paper). ES was trained on $1\,000\,\mathrm{M}$, while DQN was trained on only $40\,\mathrm{M}$ frames.}
\label{table:full-comparison}
\centering
\begin{tabular}{lrrr}
\toprule
Game & ES & DQN w/ $\epsilon$-greedy & DQN w/ param noise \\
\midrule
Alien & 994.0 & 1535.0 & \textbf{2070.0} \\
Amidar & 112.0 & 281.0 & \textbf{403.5} \\
BankHeist & 225.0 & 510.0 & \textbf{805.0} \\
BeamRider & 744.0 & \textbf{8184.0} & 7884.0 \\
Breakout & 9.5 & \textbf{406.0} & 390.5 \\
Enduro & 95.0 & 1094 & \textbf{1672.5} \\
Freeway & 31.0 & \textbf{32.0} & 31.5 \\
Frostbite & 370.0 & 250.0 & \textbf{1310.0} \\
Gravitar & \textbf{805.0} & 300.0 & 250.0 \\
MontezumaRevenge & \textbf{0.0} & \textbf{0.0} & \textbf{0.0} \\
Pitfall & \textbf{0.0} & -73.0 & -100.0 \\
Pong & \textbf{21.0} & \textbf{21.0} & 20.0 \\
PrivateEye & 100.0 & \textbf{133.0} & 100.0 \\
Qbert & 147.5 & \textbf{7625.0} & 7525.0 \\
Seaquest & 1390.0 & 8335.0 & \textbf{8920.0} \\
Solaris & \textbf{2090.0} & 720.0 & 400.0 \\
SpaceInvaders & 678.5 & 1000.0 & \textbf{1205.0} \\
Tutankham & 130.3 & 109.5 & \textbf{181.0} \\
Venture & \textbf{760.0} & 0 & 0 \\
WizardOfWor & \textbf{3480.0} & 2350.0 & 1850.0 \\
Zaxxon & 6380.0 & \textbf{8100.0} & 8050.0 \\
\bottomrule
\end{tabular}
\end{table}
\FloatBarrier
\section{Additional Results on Continuous Control with Shaped Rewards}
\label{sec:results-continuous-appendix}
For completeness, we provide the plots for all evaluated environments with dense rewards.
The results are depicted in \autoref{fig:dense-full}.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{figures/dense_full}
\caption{Median DDPG returns for all evaluated environments with dense rewards plotted over epochs.}
\label{fig:dense-full}
\end{figure}
The results for \emph{InvertedPendulum} and \emph{InvertedDoublePendulum} are very noisy due to the fact that a small change in policy can easily degrade performance significantly, and thus hard to read.
Interestingly, adaptive parameter space noise achieves the most stable performance on \emph{InvertedDoublePendulum}.
Overall, performance is comparable to other exploration approaches.
Again, no noise in either the action nor the parameter space achieves comparable results, indicating that these environments combined with DDPG are not well-suited to test for exploration.
\FloatBarrier
\section{Additional Results on Continuous Control with Sparse Rewards}
\label{sec:results-sparse-appendix}
The performance of TRPO with noise scaled according to the parameter curvature, as defined in Section~\ref{sec:adapt-trpo} is shown in \autoref{fig:continuous-sparse-trpo-adapt}.
The TRPO baseline uses only action noise by using a policy network that outputs the mean of a Gaussian distribution, while the variance is learned. These results show that adding parameter space noise aids in either learning much more consistently on these challenging sparse environments.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{figures/trpo_curvature}
\caption{Median TRPO returns with three different environments with sparse rewards plotted over epochs.}
\label{fig:continuous-sparse-trpo-adapt}
\end{figure}
\subsection*{References}}
\DeclarePairedDelimiterX{\infdivx}[2]{(}{)}{%
#1\;\delimsize\|\;#2%
}
\newcommand{D_{\text{KL}}\infdivx}{D_{\text{KL}}\infdivx}
\usepackage{bm}
\usepackage{xcolor}
\renewcommand{\vec}[1]{\boldsymbol{#1}}
\newcommand{\norm}[1]{\left\lVert#1\right\rVert}
\renewcommand{\subsectionautorefname}{Section}
\renewcommand{\subsubsectionautorefname}{Section}
\renewcommand{\S}{\mathcal{S}}
\renewcommand{\O}{\mathcal{O}}
\newcommand{\mathcal{G}}{\mathcal{G}}
\newcommand{\mathcal{A}}{\mathcal{A}}
\newcommand{\mathcal{N}}{\mathcal{N}}
\renewcommand{\L}{\mathcal{L}}
\newcommand{\mathbb{R}}{\mathbb{R}}
\newcommand{\mathbb{E}}{\mathbb{E}}
\newcommand{\mathbbm{1}}{\mathbbm{1}}
\newcommand{\change}[1]{#1}
\title{Parameter Space Noise for Exploration}
\author{
Matthias Plappert$^{\dagger \ddagger}$, Rein Houthooft$^{\dagger}$, Prafulla Dhariwal$^{\dagger}$, Szymon Sidor$^{\dagger}$,\\
\textbf{Richard Y.\ Chen$^{\dagger}$, Xi Chen$^{\dagger\dagger}$, Tamim Asfour$^{\ddagger}$, Pieter Abbeel$^{\dagger\dagger}$, and Marcin Andrychowicz$^{\dagger}$}\\
$^{\dagger}$ OpenAI\\
$^{\ddagger}$ Karlsruhe Institute of Technology (KIT)\\
$^{\dagger\dagger}$ University of California, Berkeley \\
Correspondence to \texttt{[email protected]} \\
}
\begin{document}
\maketitle
\begin{abstract}
Deep reinforcement learning (RL) methods generally engage in exploratory behavior through noise injection in the action space. An alternative is to add noise directly to the agent's parameters, which can lead to more consistent exploration and a richer set of behaviors. Methods such as evolutionary strategies use parameter perturbations, but discard all temporal structure in the process and require significantly more samples. Combining parameter noise with traditional RL methods allows to combine the best of both worlds. We demonstrate that both off- and on-policy methods benefit from this approach through experimental comparison of DQN, DDPG, and TRPO on high-dimensional discrete action environments as well as continuous control tasks.
\end{abstract}
\input{sections/introduction}
\input{sections/background}
\input{sections/param_noise}
\input{sections/exp}
\input{sections/related_work}
\input{sections/conclusion}
\begingroup
\small
\bibliographystyle{iclr2018_conference}
\small
\section{Parameter Space Noise for Exploration}
This work considers policies that are realized as parameterized functions, which we denote as $\pi_{\theta}$, with $\theta$ being the parameter vector.
We represent policies as neural
networks but our technique can be applied to arbitrary parametric models.
To achieve structured exploration, we sample from a set of policies by applying additive Gaussian noise to the parameter vector of the current policy: $ \widetilde{\theta} = \theta + \mathcal{N}(0,\sigma^2 I)$.
Importantly, the perturbed policy is sampled at the beginning of each episode and kept fixed for the entire rollout.
For convenience and readability, we denote this perturbed policy as $\widetilde{\pi} := \pi_{\widetilde{\theta}}$ and analogously define $\pi := \pi_\theta$.
\paragraph{State-dependent exploration}
As pointed out by \cite{DBLP:conf/pkdd/RuckstiessFS08}, there is a crucial difference between action space noise and parameter space noise. Consider the continuous action space case. When using Gaussian action noise, actions are sampled according to some stochastic policy, generating ${a_t = \pi(s_t) + \mathcal{N}(0, \sigma^2 I)}$. Therefore, even for a \emph{fixed} state $s$, we will almost certainly obtain a different action whenever that state is sampled again in the rollout, since action space noise is completely \emph{independent} of the current state $s_t$ (notice that this is equally true for correlated action space noise). In contrast, if the parameters of the policy are perturbed at the beginning of each episode, we get $a_t = \widetilde{\pi}(s_t)$. In this case, the same action will be taken every time the same state $s_t$ is sampled in the rollout. This ensures consistency in actions, and directly introduces a dependence between the state and the exploratory action taken.
\paragraph{Perturbing deep neural networks}
It is not immediately obvious that deep neural networks, with potentially millions of parameters and complicated nonlinear interactions, can be perturbed in meaningful ways by applying spherical Gaussian noise.
However, as recently shown by~\cite{es}, a simple reparameterization of the network achieves exactly this.
More concretely, we use layer normalization~\citep{layernorm} between perturbed layers.\footnote{This is in contrast to~\cite{es}, who use virtual batch normalization, which we found to perform less consistently}
Due to this normalizing across activations within a layer, the same perturbation scale can be used across all layers, even though different layers may exhibit different sensitivities to noise.
\paragraph{Adaptive noise scaling}
Parameter space noise requires us to pick a suitable scale $\sigma$.
This can be problematic since the scale will strongly
depend on the specific network architecture, and is likely to vary over time as parameters become more sensitive to noise as learning progresses.
Additionally, while it is easy to intuitively grasp the scale of action space noise, it is far harder to understand the scale in parameter space.
We propose a simple solution that resolves all aforementioned limitations in an easy and straightforward way.
This is achieved by adapting the scale of the parameter space noise over time and relating it to the variance in action space that it induces.
More concretely, we can define a distance measure between perturbed and non-perturbed policy in action space and adaptively increase or decrease the parameter space noise depending on whether it is below or above a certain threshold:
\begin{equation}
\sigma_{k+1} =
\begin{cases}
\alpha \sigma_k & \text{if $d(\pi, \widetilde{\pi}) \leq \delta$,} \\
\frac{1}{\alpha} \sigma_k & \text{otherwise,}
\end{cases}
\end{equation}
where $\alpha \in \mathbb{R}_{>0}$ is a scaling factor and $\delta \in \mathbb{R}_{>0}$ a threshold value.
The concrete realization of $d(\cdot, \cdot)$ depends on the algorithm at hand and we describe appropriate distance measures for DQN, DDPG, and TRPO in \autoref{sec:adapt}.
\paragraph{Parameter space noise for off-policy methods}
In the off-policy case, parameter space noise can be applied straightforwardly since, by definition, data that was collected off-policy can be used. More concretely, we only perturb the policy for exploration and train the non-perturbed network on this data by replaying it.
\paragraph{Parameter space noise for on-policy methods}
Parameter noise can be incorporated in an on-policy setting, using an adapted policy gradient, as set forth by \cite{ruckstiess2008state}. Policy gradient methods optimize
$\mathbb{E}_{\tau \sim (\pi, p)} [ R(\tau) ]$.
Given a stochastic policy $\pi_\theta(a|s)$ with $\theta \sim \mathcal{N}(\phi, \Sigma)$, the expected return can be expanded using likelihood ratios and the re-parametrization trick~\citep{kingma2013auto} as
\begin{align}
\nabla_{\phi,\Sigma} \mathbb{E}_\tau [ R(\tau) ]
\approx \frac{1}{N} \sum_{\epsilon^i,\tau^i} \left[ \sum_{t=0}^{T-1} \nabla_{\phi,\Sigma} \log \pi(a_t|s_t;\phi+\epsilon^i \Sigma^{\frac{1}{2}}) R_t(\tau^i) \right]
\end{align}
for $N$ samples $\epsilon^i \sim \mathcal{N}(0,I)$ and $\tau^i \sim (\pi_{\phi + \epsilon^i \Sigma^{\frac{1}{2}}}, p)$ (see \autoref{sec:on-policy-appendix} for a full derivation). Rather than updating $\Sigma$ according to the previously derived policy gradient, we fix its value to $\sigma^2 I$ and scale it adaptively as described in \autoref{sec:adapt}.
\section{Background}
We consider the standard RL framework consisting of an agent interacting with an environment. To simplify the exposition we assume that the environment is fully observable.
An environment is modeled as a Markov decision process (MDP) and is defined by a set of states $\S$, a set of actions $\mathcal{A}$, a distribution over initial states $p(s_0)$, a reward function $r : \S \times \mathcal{A} \mapsto \mathbb{R}$, transition probabilities $p(s_{t+1}|s_t,a_t)$, a time horizon $T$, and a discount factor $\gamma \in [0,1)$.
We denote by $\pi_\theta$ a policy parametrized by $\theta$, which can be either deterministic, $\pi: \S \mapsto \mathcal{A}$, or stochastic, \change{$\pi: \S \mapsto \mathcal{P}(\mathcal{A})$}.
The agent's goal is to maximize the expected discounted return ${\eta(\pi_\theta)} = \mathbb{E}_\tau[\sum_{t=0}^T \gamma^t r(s_t, a_t)]$, where $\tau = (s_0, a_0, \ldots, s_T)$ denotes a trajectory with ${s_0 \sim p(s_0)}$, ${a_t \sim \pi_\theta(a_t | s_t)}$, and ${s_{t+1} \sim p(s_{t+1} | s_t, a_t)}$. Experimental evaluation is based on the undiscounted return $\mathbb{E}_\tau[\sum_{t=0}^T r(s_t, a_t)]$.\footnote{If $t=T$, we write $r(s_T, a_T)$ to denote the terminal reward, even though it has has no dependence on $a_T$, to simplify notation.}
\subsection{Off-policy Methods}
Off-policy RL methods allow learning based on data captured by arbitrary policies. This paper considers two popular off-policy algorithms, namely Deep Q-Networks (DQN, \cite{dqn}) and Deep Deterministic Policy Gradients (DDPG, \cite{ddpg}).
\paragraph{Deep Q-Networks (DQN)}
DQN uses a deep neural network as a function approximator to estimate the optimal $Q$-value function, which conforms to the Bellman optimality equation: $${Q(s_t, a_t) = r(s_t, a_t) + \gamma \max_{a' \in \mathcal{A}} Q(s_{t+1}, a')}.$$
The policy is implicitly defined by $Q$ as $\pi(s_t) = \text{argmax}_{a' \in \mathcal{A}}Q(s_t, a')$.
Typically, a stochastic $\epsilon$-greedy or Boltzmann policy~\citep{sutton1998introduction} is derived from the $Q$-value function to encourage exploration, which relies on sampling noise in the action space.
The $Q$-network predicts a $Q$-value for each action and is updated using off-policy data from a replay buffer.
\paragraph{Deep Deterministic Policy Gradients (DDPG)} DDPG is an actor-critic algorithm, applicable to continuous action spaces.
Similar to DQN, the critic estimates the $Q$-value function using off-policy data and the recursive Bellman equation:
$${Q(s_t, a_t) = r(s_t, a_t) + \gamma Q\left(s_{t+1}, \pi_\theta(s_{t+1})\right)},$$
where $\pi_\theta$ is the actor or policy.
The actor is trained to maximize the critic's estimated $Q$-values by back-propagating through both networks.
For exploration, DDPG uses a stochastic policy of the form $\widehat{\pi_\theta}(s_t) = \pi_\theta(s_t) + w$, where $w$ is either $w \sim \mathcal{N}(0, \sigma^2 I)$ (uncorrelated) or $w \sim \text{OU}(0, \sigma^2)$ (correlated).\footnote{$\text{OU}(\cdot, \cdot)$ denotes the Ornstein-Uhlenbeck process~\citep{uhlenbeck1930theory}.}
Again, exploration is realized through action space noise.
\subsection{On-policy Methods}
In contrast to off-policy algorithms, on-policy methods require updating function approximators according to the currently followed policy. In particular, we will consider Trust Region Policy Optimization (TRPO, \cite{schulman2015trust}), an extension of traditional policy gradient methods \citep{williams1992simple} using the natural gradient direction \citep{DBLP:journals/ijon/PetersS08,kakade2001natural}.
\paragraph{Trust Region Policy Optimization (TRPO)}
TRPO improves upon REINFORCE~\citep{williams1992simple} by computing an ascent direction that ensures a small change in the policy distribution. More specifically, TRPO solves the following constrained optimization problem:
\begin{eqnarray*}
&\textrm{maximize}_{\theta}& E_{s \sim \rho_{\theta'}, a \sim \pi_{\theta'}} \left[ \frac{\pi_\theta(a|s)}{\pi_\theta'(a|s)} A(s, a)\right]\cr
&\text{s.t.}& E_{s\sim \rho_{\theta'}}[D_{\text{KL}}(\pi_{\theta'}(\cdot |s) \| \pi_{\theta}(\cdot|s))] \leq \delta_{\text{KL}}
\end{eqnarray*}
where $\rho_\theta = \rho_{\pi_\theta}$ is the discounted state-visitation frequencies induced by $\pi_\theta$,
$A(s, a)$ denotes the advantage function estimated by the empirical return minus the baseline, and $\delta_{\text{KL}}$ is a step size parameter which controls how much the policy is allowed to change per iteration.
\section{Conclusion}
\change{In this work, we propose parameter space noise as a conceptually simple yet effective replacement for traditional action space noise like $\epsilon$-greedy and additive Gaussian noise.
This work shows that parameter perturbations can successfully be combined with contemporary on- and off-policy deep RL algorithms such as DQN, DDPG, and TRPO and often results in improved performance compared to action noise.
Experimental results further demonstrate that using parameter noise allows solving environments with very sparse rewards, in which action noise is unlikely to succeed.
Our results indicate that parameter space noise is a viable and interesting alternative to action space noise, which is still the \emph{de facto} standard in most reinforcement learning applications.
}
\section{Related Work}
\label{sec:related_work}
The problem of exploration in reinforcement has been studied extensively.
A range of algorithms~\citep{e3, rmax, ucrl} have been proposed that guarantee near-optimal solutions after a number of steps that are polynomial in the number of states, number of actions, and the horizon time.
However, in many real-world reinforcements learning problems both the state and action space are continuous and high dimensional so that, even with discretization, these algorithms become impractical. In the context of deep reinforcement learning, a large variety of techniques have been proposed to improve exploration
\citep{DBLP:journals/corr/StadieLA15, vime, tang2016exploration, bootstrappeddqn, countbasedexploration, intrisincmotivation, randomizedvaluefuncs}.
However, all are non-trivial to implement and are often computational expensive.
The idea of perturbing the parameters of a policy has been proposed by~\cite{DBLP:conf/pkdd/RuckstiessFS08} for policy gradient methods.
The authors show that this form of perturbation generally outperforms random exploration and evaluate their exploration strategy with the REINFORCE~\citep{DBLP:journals/ml/Williams92} and Natural Actor-Critic~\citep{DBLP:journals/ijon/PetersS08} algorithms.
However, their policies are relatively low-dimensional compared to modern deep architectures, they use environments with low-dimensional state spaces, and their contribution is strictly limited to the policy gradient case.
In contrast, our method is applied and evaluated for both on and off-policy setting, we use high-dimensional policies, and environments with large state spaces.
Our work is also closely related to evolution strategies (ES, \cite{eigen1973ingo, schwefel1977numerische}), and especially neural evolution strategies (NES, \cite{DBLP:conf/icml/YiWSS09, sun09, DBLP:conf/ppsn/GlasmachersSS10, DBLP:conf/gecco/GlasmachersSSWS10, DBLP:conf/gecco/SchaulGS11, DBLP:journals/jmlr/WierstraSGSPS14}).
In the context of policy optimization, our work is closely related to \cite{DBLP:conf/nips/KoberP08} and \cite{DBLP:journals/nn/SehnkeORGPS10}.
More recently, \cite{es} showed that ES can work for high-dimensional environments like Atari and OpenAI Gym continuous control problems.
However, ES generally disregards any temporal structure that may be present in trajectories and typically suffers from sample inefficiency.
Bootstrapped DQN~\citep{bootstrappeddqn} has been proposed to aid with more directed and consistent exploration by using a network with multiple heads, where one specific head is selected at the beginning of each episode.
In contrast, our approach perturbs the parameters of the network directly, thus achieving similar yet simpler (and as shown in \autoref{sec:param-space-exploration}, sometimes superior) exploration behavior.
Concurrently to our work, \cite{fortunato2017noisy} have proposed a similar approach that utilizes parameter perturbations for more efficient exploration.
\section{Introduction}
\label{sec:introduction}
Exploration remains a key challenge in contemporary deep reinforcement learning (RL). Its main purpose is to ensure that the agent's behavior does not converge prematurely to a local optimum. Enabling efficient and effective exploration is, however, not trivial since it is not directed by the reward function of the underlying Markov decision process (MDP). Although a plethora of methods have been proposed to tackle this challenge in high-dimensional and/or continuous-action MDPs, they often rely on complex additional structures such as counting tables \citep{tang2016exploration}, density modeling of the state space \citep{countbasedexploration}, learned dynamics models \citep{vime, achiam2017surprise, DBLP:journals/corr/StadieLA15}, or self-supervised curiosity~\citep{pathakICMl17curiosity}.
An orthogonal way of increasing the exploratory nature of these algorithms is through the addition of temporally-correlated noise, for example as done in bootstrapped DQN \citep{bootstrappeddqn}.
Along the same lines, it was shown that the addition of parameter noise leads to better exploration by obtaining a policy that exhibits a larger variety of behaviors \citep{sun09,es}. We discuss these related approaches in greater detail in Section~\ref{sec:related_work}. Their main limitation, however, is that they are either only proposed and evaluated for the on-policy setting with relatively small and shallow function approximators~\citep{DBLP:conf/pkdd/RuckstiessFS08} or disregard all temporal structure and gradient information~\citep{es, DBLP:conf/nips/KoberP08, DBLP:journals/nn/SehnkeORGPS10}.
This paper investigates how parameter space noise can be effectively combined with off-the-shelf deep RL algorithms such as DQN~\citep{dqn}, DDPG~\citep{ddpg}, and TRPO~\citep{trpo} to improve their exploratory behavior. Experiments show that this form of exploration is applicable to both high-dimensional discrete environments and continuous control tasks, using on- and off-policy methods. Our results indicate that parameter noise outperforms traditional action space noise-based baselines, especially in tasks where the reward signal is extremely sparse.
\section{Experiments}
This section answers the following questions:
\begin{enumerate}
\item[(i)] Do existing state-of-the-art RL algorithms benefit from incorporating parameter space noise?
\item[(ii)] Does parameter space noise aid in exploring sparse reward environments more effectively?
\item[(iii)] How does parameter space noise exploration compare against evolution strategies \change{for deep policies \citep{es}} with respect to sample efficiency?
\end{enumerate}
\change{Reference implementations of DQN and DDPG with adaptive parameter space noise are available online.\footnote{\url{https://github.com/openai/baselines}}}
\subsection{Comparing Parameter Space Noise to Action Space Noise}
\label{sec:param-space-alternative}
The added value of parameter space noise over action space noise is measured on both high-dimensional discrete-action environments and continuous control tasks. For the discrete environments, comparisons are made using DQN, while DDPG and TRPO are used on the continuous control tasks.
\paragraph{Discrete-action environments}
For discrete-action environments, we use the Arcade Learning Environment (ALE, \cite{ale}) benchmark along with a standard DQN implementation.
We compare a baseline DQN agent with $\epsilon$-greedy action noise against a version of DQN with parameter noise.
We linearly anneal $\epsilon$ from $1.0$ to $0.1$ over the first $1$ million timesteps.
For parameter noise, we adapt the scale using a simple heuristic that increases the scale if the KL divergence between perturbed and non-perturbed policy is less than the KL divergence between greedy and $\epsilon$-greedy policy and decreases it otherwise (see \autoref{sec:adapt-dqn} for details).
By using this approach, we achieve a fair comparison between action space noise and parameter space noise since the magnitude of the noise is similar and also avoid the introduction of an additional hyperparameter.
For parameter perturbation, we found it useful to reparametrize the network in terms of an explicit policy that represents the greedy policy $\pi$ implied by the $Q$-values, rather than perturbing the $Q$-function directly. To represent the policy $\pi(a|s)$, we add a single fully connected layer after the convolutional part of the network, followed by a softmax output layer. Thus, $\pi$ predicts a discrete probability distribution over actions, given a state.
We find that perturbing $\pi$ instead of $Q$ results in more meaningful changes since we now define an explicit behavioral policy.
In this setting, the $Q$-network is trained according to standard DQN practices.
The policy $\pi$ is trained by maximizing the probability
of outputting the greedy action accordingly to the current $Q$-network.
Essentially, the policy is trained to exhibit the same behavior as running greedy DQN.
To rule out this double-headed version of DQN alone exhibits significantly different behavior, we always compare our parameter space noise approach against two baselines, regular DQN and two-headed DQN, both with $\epsilon$-greedy exploration.
We furthermore randomly sample actions for the first $50$ thousand timesteps in all cases to fill the replay buffer before starting training.
Moreover, we found that parameter space noise performs better if it is combined with a bit of action space noise (we use a $\epsilon$-greedy behavioral policy with $\epsilon=0.01$ for the parameter space noise experiments).
Full experimental details are described in~\autoref{sec:setup-ale-appendix}.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{figures/atari}
\caption{Median DQN returns for several ALE environment plotted over training steps.}
\label{fig:atari}
\end{figure}
We chose 21 games of varying complexity, according to the taxonomy presented by \citep{bellemare2016unifying}. The learning curves are shown in \autoref{fig:atari} for a selection of games (see \autoref{sec:results-ale-appendix} for full results). Each agent is trained for $40\,\mathrm{M}$~frames. The overall performance is estimated by running each configuration with three different random seeds, and we plot the median return (line) as well as the interquartile range (shaded area).
Note that performance is evaluated on the exploratory policy since we are interested in its behavior especially.
Overall, our results show that parameter space noise often outperforms action space noise, especially on games that require consistency (e.g. Enduro, Freeway) and performs comparably on the remaining ones.
Additionally, learning progress usually starts much sooner when using parameter space noise.
Finally, we also compare against a double-headed version of DQN with $\epsilon$-greedy exploration to ensure that this change in architecture is not responsible for improved exploration, which our results confirm.
Full results are available in \autoref{sec:results-ale-appendix}.
\change{That being said, parameter space noise is unable to sufficiently explore in extremely challenging games like Montezuma's Revenge.
More sophisticated exploration methods like \cite{bellemare2016unifying} are likely necessary to successfully learn these games.
However, such methods often rely on some form of ``inner'' exploration method, which is usually traditional action space noise.
It would be interesting to evaluate the effect of parameter space noise when combined with exploration methods.
}
On a final note, proposed improvements to DQN like double DQN~\citep{hasselt2010double}, prioritized experience replay~\citep{schaul2015prioritized}, and dueling networks~\citep{wang2015dueling} are orthogonal to our improvements and would therefore likely improve results further.
We leave the experimental validation of this theory to future work.
\paragraph{Continuous control environments}
We now compare parameter noise with action noise on the continuous control environments implemented in OpenAI Gym~\citep{gym}.
We use DDPG~\citep{ddpg} as the RL algorithm for all environments with similar hyperparameters as outlined in the original paper except for
the fact that
layer normalization~\citep{layernorm} is applied after each layer before the nonlinearity, which we found to be useful in either case and especially important for parameter space noise.
We compare the performance of the following configurations:
\begin{enumerate*}[label=(\alph*)]
\item no noise at all,
\item uncorrelated additive Gaussian action space noise ($\sigma=0.2$),
\item correlated additive Gaussian action space noise (OrnsteinβUhlenbeck process~\citep{uhlenbeck1930theory} with $\sigma=0.2$), and
\item adaptive parameter space noise.
\end{enumerate*}
In the case of parameter space noise, we adapt the scale so that the resulting change in action space is comparable to our baselines with uncorrelated Gaussian action space noise (see \autoref{sec:adapt-ddpg} for full details).
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{figures/dense}
\caption{Median DDPG returns for continuous control environments plotted over epochs.}
\label{fig:continuous-dense}
\end{figure}
We evaluate the performance on several continuous control tasks.
\autoref{fig:continuous-dense} depicts the results for three exemplary environments.
Each agent is trained for $1\,\mathrm{M}$~timesteps, where $1$ epoch consists of $10$~thousand timesteps.
In order to make results comparable between configurations, we evaluate the performance of the agent every $10$~thousand steps by using no noise for $20$~episodes.
On \emph{HalfCheetah}, parameter space noise achieves significantly higher returns than all other configurations.
We find that, in this environment, all other exploration schemes quickly converge to a local optimum (in which the agent learns to flip on its back and then ``wiggles'' its way forward).
Parameter space noise behaves similarly initially but still explores other options and quickly learns to break out of this sub-optimal behavior.
Also notice that parameter space noise vastly outperforms correlated action space noise on this environment, clearly indicating that there is a significant difference between the two.
On the remaining two environments, parameter space noise performs on par with other exploration strategies.
Notice, however, that even if no noise is present, DDPG is capable of learning good policies.
We find that this is representative for the remaining environments (see \autoref{sec:results-continuous-appendix} for full results), which indicates that these environments do not require a lot of exploration to begin with due to their well-shaped reward function.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{figures/trpo_dense}
\caption{Median TRPO returns for continuous control environments plotted over epochs.}
\label{fig:trpo-dense}
\end{figure}
The results for TRPO are depicted in~\autoref{fig:trpo-dense}. Interestingly, in the \emph{Walker2D} environment, we see that adding parameter noise decreases the performance variance between seeds. This indicates that parameter noise aids in escaping local optima.
\subsection{Does Parameter Space Noise Explore Efficiently?}
\label{sec:param-space-exploration}
The environments in the previous section required relatively little exploration.
In this section, we evaluate whether parameter noise enables existing RL algorithms to learn on environments with very sparse rewards, where uncorrelated action noise generally fails \citep{bootstrappeddqn, achiam2017surprise}.
\paragraph{A scalable toy example}
We first evaluate parameter noise on a well-known toy problem, following the setup described by~\cite{bootstrappeddqn} as closely as possible.
The environment consists of a chain of $N$ states and the agent always starts in state $s_2$, from where it can either move left or right.
In state $s_1$, the agent receives a small reward of $r = 0.001$ and a larger reward $r = 1$ in state $s_N$.
Obviously, it is much easier to discover the small reward in $s_1$ than the large reward in $s_N$, with increasing difficulty as $N$ grows. The environment is described in greater detail in \autoref{sec:setup-chain-appendix}.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{figures/chain_plot}
\caption{Median number of episodes before considered solved for DQN with different exploration strategies. Green indicates that the problem was solved whereas blue indicates that no solution was found within $2\,\mathrm{K}$~episodes. Note that less number of episodes before solved is better.}
\label{fig:chain-plot}
\end{figure}
We compare adaptive parameter space noise DQN, bootstrapped DQN, and $\epsilon$-greedy DQN.
The chain length $N$ is varied and for each $N$ three different seeds are trained and evaluated.
After each episode, we evaluate the performance of the current policy by performing a rollout with all noise disabled (in the case of bootstrapped DQN, we perform majority voting over all heads).
The problem is considered solved if one hundred subsequent rollouts achieve the optimal return. We plot the median number of episodes before the problem is considered solved (we abort if the problem is still unsolved after $2$~thousand episodes).
Full experimental details are available in \autoref{sec:setup-chain-appendix}.
\autoref{fig:chain-plot} shows that parameter space noise clearly outperforms action space noise (which completely fails for moderately large $N$) and even outperforms the more computational expensive bootstrapped DQN.
\change{
However, it is important to note that this environment is extremely simple in the sense that the optimal strategy is to always go right.
In a case where the agent needs to select a different optimal action depending on the current state, parameter space noise would likely work less well since weight randomization of the policy is less likely to yield this behavior.
Our results thus only highlight the difference in exploration behavior compared to action space noise in this specific case.
In the general case, parameter space noise does not guarantee optimal exploration.
}
\paragraph{Continuous control with sparse rewards}
We now make the continuous control environments more challenging for exploration.
Instead of providing a reward at every timestep, we use environments that only yield a non-zero reward after significant progress towards a goal.
More concretely, we consider the following environments from rllab\footnote{\url{https://github.com/openai/rllab}}~\citep{duan2016benchmarking}, modified according to \cite{vime}:
\begin{enumerate*}[label=(\alph*)]
\item \emph{SparseCartpoleSwingup}, which only yields a reward if the paddle is raised above a given threshold,
\item \emph{SparseDoublePendulum}, which only yields a reward if the agent reaches the upright position, and
\item \emph{SparseHalfCheetah}, which only yields a reward if the agent crosses a target distance,
\item \emph{SparseMountainCar}, which only yields a reward if the agent drives up the hill,
\item \emph{SwimmerGather}, yields a positive or negative reward upon reaching targets.
\end{enumerate*}
For all tasks, we use a time horizon of $T=500$ steps before resetting.
\begin{figure}[h]
\centering
\includegraphics[width=\textwidth]{figures/sparse_06}
\caption{Median DDPG returns for environments with sparse rewards plotted over epochs.}
\label{fig:continuous-sparse}
\end{figure}
We consider both DDPG and TRPO to solve these environments (the exact experimental setup is described in \autoref{sec:setup-continuous-appendix}).
\autoref{fig:continuous-sparse} shows the performance of DDPG, while the results for TRPO have been moved to \autoref{sec:results-sparse-appendix}. The overall performance is estimated by running each configuration with five different random seeds, after which we plot the median return (line) as well as the interquartile range (shaded area).
For DDPG, \emph{SparseDoublePendulum} seems to be easy to solve in general, with even no noise finding a successful policy relatively quickly.
The results for \emph{SparseCartpoleSwingup} and \emph{SparseMountainCar} are more interesting: Here, only parameter space noise is capable of learning successful policies since all other forms of noise, including correlated action space noise, never find states with non-zero rewards.
For \emph{SparseHalfCheetah}, DDPG at least finds the non-zero reward but never learns a successful policy from that signal.
On the challenging \emph{SwimmerGather} task, all configurations of DDPG fail.
Our results clearly show that parameter space noise can be used to improve the exploration behavior of these off-the-shelf algorithms.
\change{
However, it is important to note that improvements in exploration are not guaranteed for the general case.
It is therefore necessary to evaluate the potential benefit of parameter space noise on a case-by-case basis.
}
\subsection{Is RL with Parameter Space Noise more Sample-efficient than ES?}
Evolution strategies (ES) are closely related to our approach since both explore by introducing noise in the parameter space, which can lead to improved exploration behavior~\citep{es}.\footnote{\change{To clarify, when we refer to ES in this context, we refer to the recent work by \cite{es}, which demonstrates that deep policy networks that learn from pixels can be trained using ES. We understand that there is a vast body of other work in this field (compare \autoref{sec:related_work}).}}
However, ES disregards temporal information and uses black-box optimization to train the neural network.
By combining parameter space noise with traditional RL algorithms, we can include temporal information as well rely on gradients computed by back-propagation for optimization while still benefiting from improved exploratory behavior.
We now compare ES and traditional RL with parameter space noise directly.
We compare performance on the $21$ ALE games that were used in \autoref{sec:param-space-alternative}.
The performance is estimated by running $10$ episodes for each seed using the final policy with exploration disabled and computing the median returns.
For ES, we use the results obtained by~\cite{es}, which were obtained after training on $1\,000\,\mathrm{M}$ frames.
For DQN, we use the same parameter space noise for exploration that was previously described and train on $40\,\mathrm{M}$ frames.
Even though DQN with parameter space noise has been exposed to $25$ times less data, it outperforms ES on $15$ out of $21$ Atari games (full results are available in \autoref{sec:results-ale-appendix}).
Combined with the previously described results, this demonstrates that parameter space noise combines the desirable exploration properties of ES with the sample efficiency of traditional RL.
\subsection*{References}}
\DeclarePairedDelimiterX{\infdivx}[2]{(}{)}{%
#1\;\delimsize\|\;#2%
}
\newcommand{D_{\text{KL}}\infdivx}{D_{\text{KL}}\infdivx}
\usepackage{bm}
\usepackage{xcolor}
\newcommand{\todo}[1]{\textcolor{red}{[TODO: #1]}}
\renewcommand{\vec}[1]{\boldsymbol{#1}}
\newcommand{\norm}[1]{\left\lVert#1\right\rVert}
\renewcommand{\S}{\mathcal{S}}
\renewcommand{\O}{\mathcal{O}}
\newcommand{\mathcal{G}}{\mathcal{G}}
\newcommand{\mathcal{A}}{\mathcal{A}}
\newcommand{\mathcal{N}}{\mathcal{N}}
\renewcommand{\L}{\mathcal{L}}
\newcommand{\mathbb{R}}{\mathbb{R}}
\newcommand{\mathbb{E}}{\mathbb{E}}
\newcommand{\mathbbm{1}}{\mathbbm{1}}
\title{Parameter Space Noise for Exploration: Supplementary Material}
\author{
Matthias Plappert\thanks{Work done while interning at OpenAI.}\phantom{\footnotesize 1}\textsuperscript{,}\thanks{Correspondence to \texttt{[email protected]}.} \\
Karlsruhe Institute of Technology \\
\And
Rein Houthooft \\
OpenAI \\
\And
Prafulla Dhariwal \\
OpenAI \\
\And
Marcin Andrychowicz \\
OpenAI \\
\And
Pieter Abbeel \\
UC Berkeley, OpenAI \\
}
\begin{document}
\maketitle
| 8403ecfa404bcf68d577203c7a014d089a4dd402 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The theory of Hochschild homology for algebras has a topological analogue, called topological Hochschild homology (THH). Topological Hochschild homology for ring spectra is defined by changing the ground ring in the ordinary Hochschild complex for rings from the integers to the sphere spectrum. For coalgebras, there is a theory dual to Hochschild homology called coHochschild homology. Variations of coHochschild homology for classical coalgebras, or corings, appear in \cite[Section 30]{brzeznski-wisbauer}, and \cite{doi}, for instance, and the coHochschild complex for differential graded coalgebras appears in \cite{hps}. In recent work \cite{hs.coTHH}, Hess and Shipley define a topological version of coHochschild homology (coTHH), which is dual to topological Hochschild homology.
In this paper we develop computational tools for $\mathrm{coTHH}$.
We begin by giving a very general definition for $\mathrm{coTHH}$ in a model category endowed with a symmetric monoidal structure in terms of the homotopy limit of a cosimplicial object. This level of generality makes concrete calculations difficult; nonetheless, we give more accessible descriptions of $\mathrm{coTHH}$ in an assortment of algebraic contexts.
One of the starting points in understanding Hochschild homology is the Hoch\-schild\---Kostant--Rosenberg theorem, which, in its most basic form, identifies the Hochschild homology of free commutative algebras. See \cite{hochschild-kostant-rosenberg} for the classical result and \cite{mccarthy-minasian} for an analogue in the category of spectra. The set up for coalgebras is not as straightforward as in the algebra case and we return to the general situation to analyze the ingredients necessary for defining an appropriate notion of cofree coalgebras. In Theorem \ref{thrm:coTHHcofreeinsymmonoidalmodelcat} we conclude that the Hochschild--Kostant--Rosenberg theorem for cofree coalgebras in an arbitrary model category boils down to an analysis of the interplay between the notion of ``cofree'' and homotopy limits.
We then prove our first computational result, a Hochschild--Kostant--Rosenberg theorem for cofree differential graded coalgebras.
A similar result has been obtained by Farinati and Solotar in \cite{farinati-solotar-ext} and \cite{farinati-solotar} in the ungraded setting.
\begin{thrm} \label{thm:HKR} Let $X$ be a nonnegatively graded cochain complex over a field $k$. Let $S^c(X)$ be the cofree coaugmented cocommutative coassociative conilpotent coalgebra over $k$ cogenerated by $X$.
Then there is a quasi-isomorphism of differential graded $k$-modules
\[
\mathrm{coTHH}(S^c(X)) \simeq \Omega^{S^c(X)\vert k}.\]
where the right hand side is an explicit differential graded $k$-module defined in Section \ref{sect:categoricalhochschildhomology}.
\end{thrm}
The precise definition of $\Omega^{S^c(X)\vert k}$ depends on the characteristic of $k$. Let $U(S^c(X))$ denote the underlying differential graded $k$-modules of the cofree coalgebra $S^c(X)$. If $\mathrm{char}(k) \neq 2$, we can compute $\mathrm{coTHH}(S^c(X))$ as $U(S^c(X)) \otimes U(S^c(\Sigma^{-1} X))$, while in characteristic $2$, we have $\mathrm{coTHH}(S^c(X)) \cong U(S^c(X)) \otimes \Lambda (\Sigma^{-1}X)$, where $\Lambda$ denotes exterior powers.
Definitions of all of the terms here appear in Sections \ref{sect:definitions} and \ref{sect:categoricalhochschildhomology} and this theorem is proved as Theorem \ref{thrm:HKRcofreedg}.
We then develop calculational tools to study $\mathrm{coTHH}(C)$ for a coalgebra spectrum $C$ in a symmetric monoidal model category of spectra; see Section~\ref{sect:thespectralsequence}. Recall that for topological Hochschild homology, the B\"okstedt spectral sequence is an essential computational tool. For a field $k$ and a ring spectrum $R$, the B\"okstedt spectral sequence is of the form
\[
E^2_{*,*} = \mathrm{HH}_*(H_*(R; k)) \Rightarrow H_*(\mathrm{THH}(R); k).
\]
This spectral sequence arises from the skeletal filtration of the simplicial spectrum $\mathrm{THH}(R)_{\bullet}$. Analogously, for a coalgebra spectrum $C$, we consider the Bousfield--Kan spectral sequence arising from the cosimplicial spectrum $\mathrm{coTHH}^\bullet(C)$. We call this the coB\"okstedt spectral sequence. We identify the $E_2$-term in this spectral sequence and see that, as in the $\mathrm{THH}$ case, the $E_2$-term is given by a classical algebraic invariant:
\begin{thm}\label{thm:coBoekstedt} Let $k$ be a field. Let $C$ be a coalgebra spectrum that is cofibrant as an underlying spectrum. The Bousfield--Kan spectral sequence for $\mathrm{coTHH}(C)$ gives a {\em coB\"okstedt spectral sequence} with $E_2$-page
\[ E_2^{s, t}=\mathrm{coHH}_{s,t}^{k}(H_*(C;k)),\]
that abuts to
\[H_{t-s}(\mathrm{coTHH}(C);k).\]
\end{thm}
As one would expect with a Bousfield--Kan spectral sequence, the coB\"okstedt spectral sequence does not always converge. However, we identify conditions under which it converges {completely}. In particular, if the coalgebra spectrum $C$ is a suspension spectrum of a simply connected space $X$, denoted $\Sigma_+^{\infty}X$, the coB\"okstedt spectral sequence converges {completely} to $H_*(\mathrm{coTHH}(\Sigma_+^{\infty}X); k)$ {if a Mittag-Leffler condition is satisfied.}
Further, we prove that for $C$ connected and cocommutative this is a spectral sequence of coalgebras. Using this additional algebraic structure we prove several computational results, including the following. The divided power coalgebra $\Gamma_k[-]$ and the exterior coalgebra $\Lambda_k(-)$ are introduced in detail in Section~\ref{sect:computations}. Note that if $X$ is a graded $k$-vector space concentrated in even degrees and with basis $ ( x_i )_{i\in \mathbb{N}}$, there are identifications
$S^c(X) \cong \Gamma_k[x_1,x_2,\dots]$
and
$\Omega^{S^c(X) \vert k} \cong U(\Gamma_k[x_1,x_2,\dots]) \otimes U(\Lambda_k[z_1,z_2,\dots])$. We improve Theorem \ref{thm:HKR} in Proposition \ref{cohhcomp} by showing that under the above conditions on $X$ there is indeed an isomorphism of coalgebras
$$\mathrm{coHH}(S^c(X)) \cong \Gamma_k[x_1,x_2,\dots]) \otimes \Lambda_k[z_1,z_2,\dots].$$ Hence Theorem \ref{thm:coBoekstedt} and the coalgebra structure on the spectral sequence yield the following result.
\begin{thrm}\label{comp1} Let $C$ be a cocommutative coassociative coalgebra spectrum that is cofibrant as an underlying spectrum, and whose homology coalgebra is
\[H_*(C;k) = \Gamma_{k}[x_1,x_2,\dots],\] where the $x_i$ are cogenerators in nonnegative even degrees,
and there are only finitely many cogenerators in each degree. Then the coB\"okstedt spectral sequence for $C$ collapses at $E_2$, and
\[ E_2\cong E_{\infty} \cong \Gamma_{k}[x_1, x_2, \dots] \otimes \Lambda_{k}(z_1, z_2, \dots) ,\]
with $x_i$ in degree $(0, \deg(x_i))$ and $z_i$ in degree $(1, \deg(x_i))$.
\end{thrm}
This theorem should thus be thought of as a computational analogue of the Hochschild--Kostant--Rosenberg Theorem at the level of homology.
These computational results apply to determine the homology of $\mathrm{coTHH}$ of many suspension spectra, which have a coalgebra structure induced by the diagonal map on spaces. For simply connected $X$, topological coHochschild homology of $\Sigma_+^\infty X$ coincides with topological Hochschild homology of $\Sigma^{\infty}_+\Omega X$, and thus is closely related to algebraic $K$-theory, free loop spaces, and string topology. Our results give a new way of approaching these invariants, as we briefly recall in Section \ref{sect:computations}.
The paper is organized as follows. In Section \ref{sect:definitions}, we define $\mathrm{coTHH}$ for a coalgebra in any model category endowed with a symmetric monoidal structure as the homotopy limit of the cosimplicial object given by the cyclic cobar construction $\mathrm{coTHH}^{\bullet}$. We then give conditions for when this homotopy limit can be calculated efficiently and explain a variety of examples.
In Section \ref{sect:categoricalhochschildhomology}, we develop cofree coalgebra functors. These are not simply dual to free algebra functors, and we are explicit about the conditions on a symmetric monoidal category that are necessary to make these functors well behaved. In this context an identification similar to the Hochschild--Kostant--Rosenberg Theorem holds on the cosimplicial level. We also prove a dg-version of the Hochschild--Kostant--Rosenberg theorem.
In Section 4 we define and study the coB\"okstedt spectral sequence for a coalgebra spectrum. In particular, we analyze the convergence of the spectral sequence, and prove that it is a spectral sequence of coalgebras.
In Section 5 we use the coB\"okstedt spectral sequence to make computations of the homology of $\mathrm{coTHH}(C)$ for certain coalgebra spectra $C$. We exploit the coalgebra structure in the coB\"okstedt spectral sequence to prove that the spectral sequence collapses at $E_2$ in particular situations and give several specific examples.
\subsection{Acknowledgements} The authors express their gratitude to the organizers of the Women in Topology II Workshop and the Banff International Research Station, where much of the research presented in this article was carried out. We also thank Kathryn Hess, Ben Antieau, and Paul Goerss for several helpful conversations. The second author was supported by the National Science Foundation CAREER Grant, DMS-1149408. The third author was supported by the DNRF Niels Bohr Professorship of Lars Hesselholt. The fourth author was supported by the National Science Foundation, DMS-140648. The participation of the first author in the workshop was partially supported by the AWM's ADVANCE grant.
\section{coHochschild homology: Definitions and examples}\label{sect:definitions}
Let $(\mathcal{D}, \otimes, 1)$ be a symmetric monoidal category. We will suppress the associativity and unitality isomorphisms that are part of this structure in our notation. The coHochschild homology of a coalgebra in $\mathcal{D}$ is defined as the homotopy limit of a cosimplicial object in $\mathcal{D}$. In this section, we define coHochschild homology and then discuss a variety of algebraic examples.
\begin{defn}
A {\em coalgebra} $(C, \triangle)$ in $\mathcal{D}$ consists of an object $C$ in $\mathcal{D}$ together with a morphism
$\triangle \colon C \rightarrow C \otimes C,$
called comultiplication, which is coassociative, \emph{i.e.\,\,} satisfies
\[(C \otimes \triangle) \triangle = (\triangle \otimes C)\triangle.\]
Here we use $C$ to denote the identity morphism on the object $C\in\mathcal{D}$.
A coalgebra is called \emph{counital} if it admits a {\em counit} morphism $\epsilon \colon C \rightarrow 1$ such that
\[(\epsilon \otimes C) \triangle = C = (C \otimes \epsilon) \triangle.\]
Finally, a counital coalgebra $C$ is called {\em coaugmented} if there furthermore is a coaugmentation morphism $\eta \colon 1 \rightarrow C$
satisfying the identities
\[ \triangle \eta = \eta \otimes \eta \quad \text{ and } \quad \epsilon \eta = 1.\]
We denote the category of coaugmented coalgebras by $\mathrm{Coalg}_{\D}$.
The coalgebra $C$ is called {\em cocommutative} if
\[\tau_{C,C} \triangle = \triangle,\]
with $\tau_{C,D} \colon C \otimes D \rightarrow D \otimes C$ the symmetry isomorphism. The category of cocommutative coaugmented coalgebras in $\mathcal{D}$ is denoted by $\mathrm{comCoalg}_{\D}$.
\end{defn}
We often do not distinguish between a coalgebra and its underlying object.
Given a counital coalgebra, we can form the associated cosimplicial coHochschild object:
\begin{defn}\label{defn:coTHH}
Let $(C, \triangle, \epsilon)$ be a counital coalgebra in $\mathcal{D}$.
We define a cosimplicial object $\mathrm{coTHH}^{\bullet}(C)$ in $\mathcal{D}$ by setting
\[
\mathrm{coTHH}^{n}(C) = C^{\otimes n+1}
\]
with cofaces
\[\delta_i \colon C^{\otimes n+1} \rightarrow C^{\otimes n+2}, \quad \delta_i=\begin{cases}
C^{\otimes i} \otimes \triangle \otimes C^{n-i}, & 0\leq i \leq n,\\
{\tau_{C, C^{\otimes n+1}}}(\triangle \otimes C^{\otimes n}), & i=n+1,
\end{cases}\]
and codegeneracies
\[\sigma_i \colon C^{\otimes n+2} \rightarrow C^{\otimes n+1}, \quad
\sigma_i = C^{\otimes i+1} \otimes \epsilon \otimes C^{\otimes n-i} \quad \text{for } 0 \leq i \leq n.\]
\end{defn}
Note that if $C$ is cocommutative, $\triangle$ is a coalgebra morphism and hence $\mathrm{coTHH}^{\bullet}(C)$ is a cosimplicial counital coalgebra in $\mathcal{D}$.
Recall that if $\mathcal{D}$ is a model category, by \cite[16.7.11]{hirschhorn} the category $\mathcal{D}^{\Delta}$ of cosimplicial objects in $\mathcal{D}$ is a Reedy framed diagram category, since $\Delta$ is a Reedy category. In particular, simplicial frames exist in $\mathcal{D}$. Hence we can form the homotopy limit $\mathrm{holim}_{\Delta} X^{\bullet}$ for $X \in \mathcal{D}^{\Delta}$. More precisely, we use the model given by Hirschhorn \cite[Chapters 16 and 19]{hirschhorn} and in particular apply \cite[19.8.7]{hirschhorn}, so that $\mathrm{holim}_\Delta(X_\bullet)$ can be computed via totalization as
\[\mathrm{holim}_{\Delta} X_{\bullet} = \Tot Y_{\bullet}\]
where $Y_\bullet$ is a Reedy fibrant replacement of $X_\bullet$. Note that in order to apply \cite[19.8.7]{hirschhorn}, it suffices to have a Reedy framed diagram category structure.
\begin{defn}\label{defn:coTHHholim}
Let $\mathcal{D}$ be a model category and a symmetric monoidal category. Let $(C, \triangle, 1)$ be a counital coalgebra in $\mathcal{D}$. Then the \emph{coHochschild homology} of $C$ is defined by
\[\mathrm{coTHH}(C) = \mathrm{holim}_{\Delta} \mathrm{coTHH}^{\bullet}(C).\]
\end{defn}
We chose to denote coHochschild homology in a category $\mathcal{D}$ as above by $\mathrm{coTHH}$ instead of $\mathrm{coHH}$ to emphasize that the construction relies on the existence of a model structure on $\mathcal{D}$. However, in several of the examples that follow, the category $\mathcal{D}$ does not come from topological spaces or spectra.
For example we will see in \ref{ex:cochBounded} and \ref{ex:ch} that $\mathrm{coTHH}$ coincides with the original definition of coHochschild homology for $k$-modules and more generally cochain coalgebras (see \cite{doi}, \cite{hps}). We will use the standard notation $\mathrm{coHH}(C)$ for the coHochschild homology of a coalgebra $C$ in the category of graded $k$-modules in later sections.
\begin{rem}\label{rem:fibrancyassumptionsforholim}
Let $\mathcal{D}$ be a symmetric monoidal model category. Let $f \colon C \rightarrow D$ be a morphism of counital coalgebras and assume that $f$ is a weak equivalence and $C$ and $D$ are cofibrant in $\mathcal{D}$. Then $f$ induces a degreewise weak equivalence of cosimplicial objects $\mathrm{coTHH}^{\bullet}(C) \rightarrow \mathrm{coTHH}^{\bullet}(D)$. Hence the corresponding Reedy fibrant replacements are also weakly equivalent and $f$ induces a weak equivalence $\mathrm{coTHH}(C) \rightarrow \mathrm{coTHH}(D)$ according to \cite[19.4.2(2)]{hirschhorn}.
\end{rem}
A central question raised by the definition of coHochschild homology is that of understanding the homotopy limit over $\Delta$. In the rest of this section, we discuss conditions under which one can efficiently compute it, including an assortment of algebraic examples which we discuss further in Section \ref{sect:categoricalhochschildhomology}.
For a cosimpicial object $X^{\bullet}$ in $\mathcal{D}$ let $M_n(X^{\bullet})$ be its $n$th matching object given by
\[
M_n(X^{\bullet}) = \lim_{\substack{\alpha \in \Delta([n],[a]), \\\alpha \text{ surjective}, \\a \neq n}} X^a.
\]
The matching object $M_n(X^{\bullet})$ is traditionally also denoted by $M^{n+1}(X^{\bullet})$, but we stick to the conventions in \cite[15.2]{hirschhorn}.
If $\mathcal{D}$ is a model category, $X^{\bullet}$ is called \emph{Reedy fibrant} if for every $n$, the morphism
\[\sigma \colon X^n \rightarrow M_n(X^{\bullet})\]
is a fibration, where $\sigma$ is induced by the collection of maps $\alpha_* \colon X^n \rightarrow X^a$ for $\alpha\in \Delta([n],[a])$.
If $\mathcal{D}$ is a simplicial model category and $X^{\bullet}$ is Reedy fibrant, $\mathrm{holim}_{\Delta} X^{\bullet}$ is weakly equivalent to the totalization of $X^{\bullet}$ defined via the simplicial structure; see \cite[Theorem 18.7.4]{hirschhorn}. In this case the totalization is given by the equalizer
\[\Tot (X^{\bullet}) = \text{eq}\Bigg(
\prod_{n \geq 0} (X^n)^{\Delta^n} \rightrightarrows \prod_{\alpha \in \Delta([a],[b])} (X^b)^{\Delta^a}
\Bigg),\]
where $D^K$ denotes the cotensor of an object $D$ of $\mathcal{D}$ and a simplicial set $K$ and where the two maps are induced by the cosimplicial structures of $X^{\bullet}$ and $\Delta^{\bullet}$.
If the category $\mathcal{D}$ is not simplicial, one has to work with simplicial frames as in \cite[Chapter 19]{hirschhorn}.
In the dual case, if $A$ is a monoid in a symmetric monoidal model category, one considers the cyclic bar construction associated to $A$, \emph{i.e.\,\,} the simplicial object whose definition is entirely dual to the definition of $\mathrm{coTHH}^{\bullet}(C)$. This simplicial object is Reedy cofibrant under mild conditions; see for example \cite[Example 23.8]{shulman}. But the arguments proving Reedy cofibrancy of the simplicial Hochschild object do not dualize to $\mathrm{coTHH}^{\bullet}(C)$: It is neither reasonable to expect the monoidal product to commute with pullbacks, nor to expect that pullbacks and the monoidal product satisfy an analogue of the pushout-product axiom.
Nonetheless, in certain cases $\mathrm{coTHH}^{\bullet}(C)$ can be seen to be Reedy fibrant. In particular Reedy fibrancy is simple when the symmetric monoidal structure is given by the cartesian product.
\begin{lemma} \label{lem:ReedyFibTimes}
Let $\mathcal{D}$ be any category and let $C$ be a coalgebra in $(\mathcal{D}, \times, *)$. Then
\[\sigma \colon \mathrm{coTHH}^n(C) \rightarrow M_n(\mathrm{coTHH}^{\bullet}(C))\]
is an isomorphism for $n\geq 2$. Hence if $\mathcal{D}$ is a model category, then $\mathrm{coTHH}^\bullet(C)$ is Reedy fibrant if $C$ itself is fibrant.
\end{lemma}
\begin{proof}
For surjective $\alpha \in \Delta([n],[n-1])$, let
\[p_{\alpha} \colon M_n(\mathrm{coTHH}^{\bullet}(C)) \rightarrow \mathrm{coTHH}^{n-1}(C)=C^{\times n}\]
be the canonical projection to the copy of $\mathrm{coTHH}^{n-1}(C)$ corresponding to $\alpha$. Then an inverse to $\sigma$ is given by
\begin{multline*}
M_n(\mathrm{coTHH}^{\bullet}(C)) \xto{(p_{\alpha})_{\alpha}} (C^{\times n})^{\times n} \xto{((C^{\times n})^{\times n},C^{\times n} \times (*)^{\times n-1})} (C^{\times n})^{n+1}\\
\xto{(C\times *^{n-1}) \times \dotsb \times (*^{n-1} \times C) \times (*^{n-1} \times C)} C^{\times n+1}.
\end{multline*}
Intuitively, this map sends $(c^1,\dotsc,c^n) \in M_n(\mathrm{coTHH}^{\bullet}(C))$ with $c^i=(c^i_1,\dotsc,c^i_n) \in C^{\times n}$ to
$(c^1_1,\dotsc, c^n_n, c^1_n).$
The Reedy fibration conditions for $n = 1, 0$ follow from the fibrancy of $C$.
\end{proof}
In several algebraic examples, degreewise surjections are fibrations and the following lemma will imply Reedy fibrancy.
\begin{lemma} \label{lem:ReedyFibSurj}
Let $C$ be a counital coalgebra in the category of graded $k$-modules for a commutative ring $k$. If $C$ is coaugmented, or if there is a $k$-module map $\eta\colon k\to C$ such that $\epsilon\eta=C$, then the matching maps
\[ \sigma\colon \mathrm{coTHH}^n(C) \to M_n(\mathrm{coTHH}^\bullet(C)) \]
are surjective.
\end{lemma}
In the rather technical proof we construct an explicit preimage for any element in $M_n(\mathrm{coTHH}^\bullet(C))$. We defer the proof to Appendix \ref{appendix:yuckyproof}. Note that if $k$ is a field, then by choosing a basis for $C$ over $k$ compatibly with the counit, we can always construct the necessary map $\eta$. Thus $\mathrm{coTHH}$ of any counital coalgebra over a field is Reedy fibrant.
\begin{comment}
\begin{proof} From Hirschhorn \cite[15.2.6]{hirschhorn} we know that the matching space $M_n(X^\bullet)$ for a cosimplicial graded $k$-module $X^{\bullet}$ is given by
\[M_n(X^{\bullet}) = \lbrace (x_0,\dotsc, x_{n-1}) \in (X^{n-1})^{\times n} \mid \sigma_i(x_j) = \sigma_{j-1} (x_i) \text{ for $0 \leq i<j \leq n-1$}\rbrace.\]
This space $M_n(X^{\bullet})$ can be obtained by iterated pullbacks:
For $k\leq n-1$, define $M_n^{k}(X^{\bullet})$ by
\[M_n^k(X^{\bullet}) = \lbrace (x_0,\dotsc,x_{k}) \in (X^{n-1})^{\times k+1} \mid \sigma_i(x_j) = \sigma_{j-1} (x_i) \text{ for $0 \leq i<j\leq k$}\rbrace\]
and let $\sigma^{(k)}$ be the map $(\sigma_0,\dotsc,\sigma_k) \colon X^n \rightarrow M_n^k(X^{\bullet})$ . Then $M_n^{n-1}(X^{\bullet}) = M_n(X^{\bullet})$ and $\sigma^{(n-1)}=\sigma$.
We show by induction that for $X^{\bullet} = \mathrm{coTHH}^{\bullet}(C)$ the maps $\sigma^{(k)}$ are all surjective. The map $\sigma^{(0)}=\sigma_0 \colon C\otimes C \to C$ is clearly surjective.
Assume that $\sigma^{(k-1)}$ is surjective, and consider $(x_0,\dotsc,x_k) \in M_n^k(X^{\bullet})$. Since $(x_0,\dotsc,x_{k-1}) \in M_n^{k-1}(X^{\bullet})$, we can choose $y \in C^{\otimes n}$ such that
\[\sigma^{(k-1)}(y)=(x_0,\dotsc,x_{k-1}).\]
We know that $\sigma_i(x_k) = \sigma_{k-1}(x_i) = \sigma_i \sigma_k(y)$ for $i<k$.
Choose a basis $(c_l)_{l\in L}$ of $C$, such that there is exactly one $l_0 \in L$ with $\epsilon(c_{l_0})=1$, and such that $\epsilon(c_l)=0$ for all $l \neq l_0$.
Write $x_k$ and $y$ in terms of tensors of the basis elements as
\[x_k= \sum_{j_0,\dotsc,j_{n-1} \in L} \lambda_{(j_0,\dotsc, j_{n-1})} c_{j_0} \otimes \dotsb \otimes c_{j_{n-1}}\]
and
\[y= \sum_{j_0,\dotsc,j_{n} \in L} \mu_{j_0,\dotsc,j_{n}} c_{j_0} \otimes \dotsb \otimes c_{j_{n}}\]
for some coefficients $\lambda_{j_0,\dotsc,j_{n-1}}$ and $\mu_{j_0,\dotsc,j_{n}}$ in $k$.
If $j_1,\dotsc,j_k\neq l_0$, then $\sigma_i(c_{j_0}\otimes\dotsb \otimes c_{j_{n}})=0$ for $0\leq i\leq k-1$. Hence the coefficient $\mu_{j_0,\dotsc,j_{n}}$ does not contribute to $\sigma^{(k-1)}(y)$ and so we can assume that $\mu_{j_0,\dotsc,j_{n}}=0$ if $j_1,\dotsc,j_{k} \neq l_0$.
Define $z \in C^{\otimes n+1}$ by
\[z= y +\sum_{j_1,\dotsc,j_k \neq l_0} \sum_{j_0,j_{k+1},\dotsc,j_{n-1}} \lambda_{j_0,\dotsc,j_{n-1}} c_{j_0} \otimes \dotsb \otimes c_{j_k} \otimes c_{l_0} \otimes c_{j_{k+1}} \otimes \dotsb \otimes c_{j_{n-1}}.\]
For $i<k$ and any simple tensor of basis elements $c_{j_0}\otimes\dotsb\otimes c_{j_{n}}$ where $j_1,\dotsc, j_k\neq l_0$, $\sigma_i(c_{j_0}\otimes\dotsb\otimes c_{j_{n}})=0$. Hence for $i<k$,
\[\sigma_i(z) = \sigma_i(y) = x_i.\]
We must show that $\sigma_k(z)=x_k$.
Fix an $n$-tuple $(j_0,\dotsc,j_{n-1}) \in L^n$. The terms of $z$ that contribute to the coefficient of $c_{j_0}\otimes \dotsb \otimes c_{j_{n-1}}$ in $\sigma_k(z)$ are those of the form
\[c_{j_0}\otimes \dotsb \otimes c_{j_k}\otimes c_{l_0}\otimes c_{j_{k+1}}\otimes \dotsb \otimes c_{j_{n-1}}.\]
If $j_i \neq l_0$ for all $1\leq i\leq k$, then for any $l\in L$, the coefficient $\mu_{j_0,\dotsc,j_k,l,j_{k+1},\dotsc,j_{n-1}}$ is zero, and thus the coefficient of $c_{j_0}\otimes \dotsb \otimes c_{j_{n-1}}$ in $\sigma_k(z)$ is $\lambda_{j_0,\dotsc,j_{n-1}}$ as required.
Suppose $j_s=l_0$ for some $1\leq s\leq k$. The coefficient of $c_{j_0}\otimes \dotsb \otimes c_{j_{n-1}}$ in $\sigma_k(z)$ is then the coefficient of $c_{j_0}\otimes\dotsb\otimes c_{j_{s-1}}\otimes c_{j_{s+1}}\otimes \dotsb \otimes c_{j_{n-1}}$ in $\sigma_{s-1}\sigma_k(z)$. Similarly, the coefficient of $c_{j_0}\otimes \dotsb \otimes c_{j_{n-1}}$ in $x_k$ is the coefficient of $c_{j_0}\otimes \dotsb\otimes c_{j_{s-1}}\otimes c_{j_{s+1}} \otimes \dotsb\otimes c_{j_{n-1}}$ in $\sigma_{s-1}(x_k)$. Since $(x_0,\dotsc, x_k)\in M_n^k(X^\bullet)$,
\[ \sigma_{s-1}(x_k)=\sigma_{k-1}(x_{s-1})\]
and by construction, $x_{s-1}=\sigma_{s-1}(z).$ Thus $\sigma_{s-1}(x_k)=\sigma_{k-1}\sigma_{s-1}(z)$ and so by the cosimplicial identities,
\[\sigma_{s-1}(x_k)=\sigma_{s-1}\sigma_{k}(z).\]
Thus the coefficients of $c_{j_1}\otimes\dotsb\otimes c_{j_n}$ in $x_k$ and $\sigma_k(z)$ are equal.
\end{proof}
\end{comment}
We now collect a couple of examples of what $\mathrm{coTHH}$ is in various categories.
\begin{ex}\label{ex:coskmod}
Let $\mathcal{D}=(\mathrm{ckmod}, \otimes, k)$ be the symmetric monoidal category of cosimplicial $k$-modules for a field $k$. There is a simplicial model structure on $\mathcal{D}$; see \cite[II.5]{fresse} for an exposition. The simplicial cotensor of a cosimplicial $k$-module $M^{\bullet}$ and a simplicial set $K_{\bullet}$ is given by
\[(M^K)^n= \mathrm{Set}(K_n,M^n)\]
with $\alpha \in \Delta([n],[m])$ mapping $f\colon K_n \rightarrow M^n$ to
\[K_m \xto{\,\alpha\,} K_n \xto{\,f\,} M_n \xto{\,\alpha_*\,} M_m,\]
see \cite[II.2.8.4]{goerss-jardine}.
As in the simplicial case (see \cite[III.2.11]{goerss-jardine}), one can check that the fibrations are precisely the degreewise surjective maps.
Hence by Lemma \ref{lem:ReedyFibSurj} $\mathrm{coTHH}(C)$ can be computed as the totalization, that is, as the diagonal of the bicosimplicial $k$-module $\mathrm{coTHH}^{\bullet}(C)$.
Even though $\mathcal{D}$ is not a monoidal model category, the tensor product of weak equivalences is again a weak equivalence, and hence a weak equivalence of counital coalgebras induces a weak equivalence on $\mathrm{coTHH}$.
\end{ex}
Recall that the dual Dold--Kan correspondence is an equivalence of categories
\[
N^* \colon \mathrm{ckmod} \leftrightarrows\mathrm{dgmod}^{\geq0}(k) \ \colon\! \Gamma^{\bullet}
\]
between cosimplicial $k$-modules and nonnegatively graded cochain complexes over $k$. Given $M^{\bullet} \in \mathrm{ckmod}$, the normalized cochain complex $N^*(M^{\bullet})$ is defined by
\[N^a(M^{\bullet})=\bigcap_{i=0}^{a-1} \ker (\sigma_i \colon M^a \rightarrow M^{a-1})\]
for $a>0$ and $N^0(M^{\bullet}) = M^0$, with differential $d \colon N^a(M^{\bullet}) \rightarrow N^{a+1}(M^{\bullet})$ given by
\[\sum_{i=0}^{a+1} (-1)^i \delta_i.\]
\begin{ex}\label{ex:cochBounded}
Let $k$ be a field. Let $\mathcal{D}$ be the symmetric monoidal category $\mathcal{D}=(\mathrm{dgmod}^{\geq0}(k), \otimes, k)$ of nonnegatively graded cochain complexes over $k$. The dual Dold--Kan correspondence
yields that $\mathcal{D}$ is a simplicial model category, with weak equivalences the quasi-isomorphisms and with fibrations the degreewise surjections.
Hence by Lemma \ref{lem:ReedyFibSurj} we have that $\mathrm{coTHH}^{\bullet}(C)$ is Reedy fibrant for any counital coalgebra $C$ in $\mathcal{D}$.
The totalization of a cosimplicial cochain complex $M^{\bullet}$ is given by
\[\Tot(M^{\bullet}) \cong N^*(\Tot(\Gamma(M^{\bullet}))) = N^*(\mathrm{diag}(\Gamma(M^{\bullet}))).\]
Recall that given a double cochain complex $B_{*,*}$ with differentials $d^h \colon B_{*,*} \rightarrow B_{*+1,*}$ and $d^v\colon B_{*,*} \rightarrow B_{*,*+1}$, we can form the total cochain complex $\Tot_{\mathrm{dgmod}}(B)$ given by
\[\Tot_{\mathrm{dgmod}}(B)_n = \prod_{p+q =n} B_{p,q}.\]
The projection to the component $B_{s, n+1-s}$ of the differential applied to $(b_{p} \in B_{p,n-p})_p$ is given by
$d^h(b_{s-1,n-s+1}) +(-1)^pd^v(b_{s,n-s}).$ A similar construction can be carried out for double chain complexes.
As in the bisimplicial case, we can turn a bicosimpicial $k$-module $X^{\bullet,\bullet}$ into a double complex by applying the normalized cochain functor to both cosimplicial directions. The homotopy groups of $\mathrm{diag} X^{\bullet,\bullet}$ are the homology of the associated total complex, hence
\[\mathrm{coTHH}(C) = \Tot_{\mathrm{dgmod}} N^*(\mathrm{coTHH}^{\bullet}(C))\]
for any counital coalgebra in $\mathcal{D}$. Hence our notion of coHochschild homology coincides with the classical one as in \cite{doi} and \cite{hps}. Again weak equivalences of counital coalgebras induce weak equivalences on $\mathrm{coTHH}$.
\end{ex}
\begin{ex}\label{ex:simpset}
The category $\mathcal{D} = (\mathrm{sSet}, \times, *)$ is a simplicial symmetric monoidal model category with its classical model structure. Every object is cofibrant, and
by Lemma \ref{lem:ReedyFibTimes}, $\mathrm{coTHH}(C)$ can be computed as the totalization of the cosimplicial simplicial set $\mathrm{coTHH}^{\bullet}(C)$ when $C$ is fibrant.
\end{ex}
\begin{ex}\label{ex:skmod}
Let $\mathcal{D} = (\mathrm{skmod}, \otimes, k)$ be the category of simplicial $k$-modules over a field $k$. Every object is cofibrant in $\mathcal{D}$.
Let $C$ be a counital coalgebra. Since every morphism that is degreewise surjective is a fibration (see \cite[III.2.10]{goerss-jardine}), Lemma \ref{lem:ReedyFibSurj} yields that $\mathrm{coTHH}^{\bullet}(C)$ is Reedy fibrant, and $\mathrm{coTHH}(C)$ is given by the usual totalization of cosimplicial simplicial $k$-modules.
\end{ex}
\begin{ex} \label{ex:chBounded}
The category $\mathcal{D}=(\mathrm{dgmod}_{\geq0}(k), \otimes, k)$ of nonnegatively graded chain complexes over a ring $k$ with the projective model structure is both a monoidal model category as well as a simplicial model category via the Dold--Kan equivalence $(N, \Gamma)$ with $(\mathrm{skmod}, \otimes, k)$.
If we work over a field, every object is cofibrant, and for any counital coalgebra $C$ in $\mathcal{D}$ Lemma \ref{lem:ReedyFibSurj} yields that $\mathrm{coTHH}^{\bullet}(C)$ is Reedy fibrant.
However, a word of warning is in order. The definition of coHochschild homology given in Definition \ref{defn:coTHHholim} does not coincide with the usual notion of coHochschild homology of a differential graded coalgebra as for example found in \cite{doi}: Usually, coHochschild homology of a coalgebra $C$ in $k$-modules or differential graded $k$-modules is given by applying the normalized cochain complex functor $N^*$ to $\mathrm{coTHH}^{\bullet}(C)$ and forming the total complex $\Tot_{\mathrm{dgmod}}(N^*(\mathrm{coTHH}^{\bullet}(C)))$ of the double chain complex resulting from interpreting $N^a(\mathrm{coTHH}^{\bullet}(C))$ as living in chain complex degree $-a$.
But for any cosimplicial nonnegatively graded chain complex $M^{\bullet}$,
\[\Tot(M^{\bullet}) \cong N_*(\Tot(\Gamma(M^{\bullet}))),\]
where the second $\Tot$ is the totalization of the simplicial cosimplicial $k$-module $\Gamma(M^{\bullet})$.
As explained in \cite[III.1.1.13]{fresse}, we have an identity
\[N_*(\Tot(\Gamma(M^{\bullet}))) = \tau_*(\Tot_{\mathrm{dgmod}}(N^*(M^{\bullet}))),\]
where $\tau_*(C)$ of an unbounded chain complex $C$ is given by
\[\tau_*(C)_n = \begin{cases}
C_n, & n>0,\\
\ker(C_0 \rightarrow C_{-1}), & n=0,\\
0, n<0.
\end{cases}\]
Hence if $C$ is concentrated in degree zero, so is $\mathrm{coTHH}(C)$, in contrast to the usual definition as for example found in \cite{doi} and \cite{hps}.
Nonetheless, if we assume that $C_0 =0$, the above identification of the totalization yields that the two notions of coHochschild homology coincide.
\end{ex}
\begin{ex}\label{ex:ch}
The category $\mathcal{D}=(\mathrm{dgmod}, \otimes, k)$ of unbounded chain complexes over a commutative ring $k$ is a symmetric monoidal model category with weak equivalences the quasi-isomorphisms and fibrations the degreewise surjective morphisms; see \cite[4.2.13]{hovey}.
Let $C$ be a cofibrant chain complex. A simplicial frame of $C$ is given by the simplicial chain complex $\mathrm{fr}(C)$ which is given in simplicial degree $m$ by the internal hom object $\underline{\mathrm{dgmod}}(N_*(\Delta^m), C)$ (see proof of \cite[5.6.10]{hovey}).
Again, $\mathrm{coTHH}^{\bullet}(C)$ is Reedy fibrant. This yields that the homotopy limit $\mathrm{holim}_{\Delta} \mathrm{coTHH}^{\bullet}(C)$ coincides
with the total complex of the double complex $N^*(\mathrm{coTHH}^{\bullet}(C))$. Again this coincides with the definitions of coHochschild homology in \cite{doi} and \cite{hps}.
\end{ex}
\begin{ex}
Both the category of symmetric spectra as well as the category of $\mathbb{S}$-modules are simplicial symmetric monoidal model categories. Sections \ref{sect:thespectralsequence} and \ref{sect:computations} discuss $\mathrm{coTHH}$ for coalgebra spectra.
\end{ex}
\section{coHochschild homology of cofree coalgebras in symmetric monoidal categories}\label{sect:categoricalhochschildhomology}
We next turn to calculations of coHochschild homology of cofree coalgebras in symmetric monoidal categories. The first step is to specify what we mean by a ``cofree'' coalgebra in our symmetric monoidal category $\mathcal{D}$. Intuitively, such a coalgebra should be given by a right adjoint to the forgetful functor from the category of coalgebras in $\mathcal{D}$ to $\mathcal{D}$ itself. The constructions are the usual ones and can be found for example in \cite[II.3.7]{markl-shnider-stasheff} in the algebraic case, but we specify conditions under which we can guarantee the existence of this adjoint. For an operadic background to these conditions we refer the reader to \cite{ching}.
Let $\mathcal{D}$ be complete and cocomplete and admit a zero object $0$. Given a coalgebra $(C, \triangle)$ in a symmetric monoidal category $\mathcal{D}$ and $n\geq 0$, we denote by $\triangle^n \colon C \rightarrow C^{\otimes n+1}$ the iterated comultiplication defined inductively by
\[\triangle^{0} = C, \qquad \triangle^{n+1} = (\triangle \otimes C^{\otimes n}) \triangle^n.\]
Note that if $(C, \triangle, \epsilon, \eta)$ is a coaugmented coalgebra, the cokernel of the coaugmentation, $\mathrm{coker}\: \eta$, is a coalgebra.
\begin{defn}
We call a (non-coaugmented) coalgebra $C$ {\em conilpotent} if the morphism
\[ (\triangle^n)_{n\geq 0} \colon C \rightarrow \prod_{n\geq {0}} C^{\otimes n+1},\]
that is, the morphism given by $\triangle^n$ after projection to $C^{\otimes n+1}$, factors through
$\bigoplus_{n \geq {0}} C^{\otimes n+1}$
via the canonical morphism
\[\bigoplus_{n \geq {0}} C^{\otimes n+1} \rightarrow \prod_{n \geq {0}} C^{\otimes n+1}.\]
We say that a coaugmented coalgebra $C$ is conilpotent if $\mathrm{coker}\: \eta$ is a conilpotent coalgebra.
We denote the full subcategory of $\mathrm{Coalg}_{\D}$ consisting of coaugmented conilpotent coalgebras by $\mathrm{Coalg}_{\D}^{\mathrm{conil}}$. Similarly, the category of cocommutative conilpotent coaugmented coalgebras is denoted by $\mathrm{comCoalg}_{\D}^{\mathrm{conil}}$.
\end{defn}
\begin{defn}\label{defn:cofreefriendly}
The symmetric monoidal category $(\mathcal{D}, \otimes, 1)$ is called \emph{cofree-friendly}
if it is complete, cocomplete, admits a zero object, and the following additional conditions hold:
\begin{enumerate}
\item For all objects $D$ in $\mathcal{D}$, the functor $D \otimes -$ preserves colimits.
\item Finite sums and finite products are naturally isomorphic, that is for every finite set $J$ and all objects $D_j$, $j \in J$, the morphism
\[\bigoplus_{j \in J} D_j \rightarrow \prod_{j \in J} D_j\]
induced by the identity morphism on each $D_j$ is an isomorphism.
\end{enumerate}
\end{defn}
\begin{prop}
If $\mathcal{D}$ is cofree-friendly, the functor
\[coker \colon \mathrm{Coalg}_{\D}^{\mathrm{conil}} \rightarrow \mathcal{D}, \quad (C,\triangle, \epsilon, \eta) \mapsto \mathrm{coker}\: \eta\]
admits a right adjoint $T^c$, which we call the \emph{cofree coaugmented conilpotent coalgebra functor.}
The underlying object of $T^c (X)$ is given by $\bigoplus_{n \geq 0} X^{\otimes n}$. The comultiplication $\triangle$ is defined via deconcatenation, \emph{i.e.\,\,} by
\addtolength{\jot}{1ex}
\begin{multline*}
\bigoplus_{n\geq 0} X^{\otimes n}\xto{\bigoplus_{n\geq 0} (X^{\otimes n})_{a+b=n}}
\bigoplus_{n \geq 0} \prod_{\substack{ a,b\geq 0 \\ a+b=n}} X^{\otimes n} \xto{\cong}
\bigoplus_{n \geq 0} \bigoplus\limits_{\substack{ a,b\geq 0 \\ a+b=n}} X^{\otimes n} \\
\cong \bigoplus\limits_{a,b\geq 0} X^{\otimes a} \otimes X^{\otimes b}
\xto{\cong}
\bigoplus\limits_{a\geq 0} X^{\otimes a} \otimes \bigoplus\limits_{b\geq 0} X^{\otimes b},
\addtolength{\jot}{-1ex}
\end{multline*}
where the isomorphism on the first line is provided by condition (2) of Definition \ref{defn:cofreefriendly}.
The counit morphism $\pi \colon T^c(X) \rightarrow X$ is the identity on $X$ and the zero morphism on the other summands.
\end{prop}
\begin{ex}\label{ex:kmodcofreefriendly}
The categories of chain or cochain complexes discussed in Examples \ref{ex:cochBounded}, \ref{ex:chBounded} and \ref{ex:ch} are cofree-friendly, and the cofree coalgebra $T^c(X)$ generated by an object in any of these categories is the usual tensor coalgebra.
Similarly, the categories of simplicial or cosimplicial $k$-modules discussed in Examples \ref{ex:coskmod} and \ref{ex:skmod} are cofree-friendly, and the cofree coalgebra generated by $X$ is obtained by applying the tensor coalgebra functor in each simplicial or cosimplicial degree.
\end{ex}
\begin{ex}\label{ex:notcofreefriendlythings}
The categories of simplicial sets, of pointed simplicial sets and of spectra are not cofree-friendly, since finite coproducts and finite products are not isomorphic in any of these categories.
A more abstract way of thinking about this is given by observing that in these categories coaugmented coalgebras can not be described as coalgebras over a cooperad in the usual way: While the category of symmetric sequences in these categories is monoidal with respect to the plethysm that gives rise to the notion of operads, the plethysm governing cooperad structures does not define a monoidal product. This is discussed in detail by Ching \cite{ching}; see in particular Remark 2.10, Remark 2.20 and Remark 2.21.
\end{ex}
Let $G$ be a group with identity element $e$. Recall that a $G$-action on an object $X$ in $\mathcal{D}$ consists of morphisms
\[\phi_g \colon X \rightarrow X, \qquad g \in G,\]
such that
$\phi_g \phi_h = \phi_{gh}$ and $\phi_e = X$.
The {\em fixed points} $X^G$ of this actions are given by the equalizer
\[X^G = \mathrm{eq}(X \xrightrightarrows[(\phi_g)_{g \in G}]{(X)_{g\in G}} \prod_{g\in G} X).\]
\begin{defn}\label{defn:permutationfriendly}
Recall that for any object $X$ of $\mathcal{D}$ the symmetric group $\Sigma_n$ acts on $X^{\otimes n}$ by permuting the factors. We call $\mathcal{D}$ {\em permutation-friendly}
if the morphism
\[ (X^{\otimes a})^{\Sigma_a} \otimes (X^{\otimes b})^{\Sigma_b} \rightarrow (X^{\otimes a+b})^{\Sigma_a \times \Sigma_b},\]
induced by the morphisms $ (X^{\otimes a})^{\Sigma_a} \rightarrow X^{\otimes a} $ and $ (X^{\otimes b})^{\Sigma_b} \rightarrow X^{\otimes b}$,
is an isomorphism for all $X$ in $\mathcal{D}$ and all $a,b\geq 0$.
\end{defn}
\begin{prop}
If $\mathcal{D}$ is cofree-friendly and permutation-friendly, the forgetful functor
\[\mathrm{coker}\:\!\!\colon \mathrm{comCoalg}_{\D}^{\mathrm{conil}} \rightarrow \mathcal{D}, \quad (C,\triangle,\epsilon,\eta) \mapsto \mathrm{coker}\: \eta\]
admits a right adjoint $S^c$, which is called the \emph{cofree cocommutative conilpotent coaugmented coalgebra functor}.
For a given object $X$ of $\mathcal{D}$, define $S^c (X)=\bigoplus_{n \geq 0} (X^{\otimes n})^{\Sigma_n}$. The comultiplication $\triangle$ on $S^c(X)$ is defined by
\addtolength{\jot}{1ex}
\begin{equation*}
\begin{split}
\bigoplus_{n\geq 0} (X^{\otimes n})^{\Sigma_n}
&\xto{\bigoplus_{n\geq 0} ((X^{\otimes n})^{\Sigma_n})_{a+b=n}}
\bigoplus\limits_{n \geq 0} \prod\limits_{\substack{ a,b \geq 0, \\a+b=n}} (X^{\otimes n})^{\Sigma_n}
\xto{\cong}
\bigoplus_{n \geq 0} \bigoplus_{\substack{ a,b \geq 0, \\a+b=n}} (X^{\otimes n})^{\Sigma_n}\\
&\cong \bigoplus\limits_{a,b\geq 0} (X^{\otimes a} \otimes X^{\otimes b})^{\Sigma_{a+b}}
\xto{\bigoplus\limits_{a,b\geq 0}\mathrm{res}_{a,b}}
\bigoplus\limits_{a,b\geq 0} (X^{\otimes a} \otimes X^{\otimes b})^{\Sigma_a \times \Sigma_b}\\
&\xto{\cong}
\bigoplus\limits_{a,b\geq 0} (X^{\otimes a})^{\Sigma_a} \otimes (X^{\otimes b})^{\Sigma_b}
\xto{\cong}
\bigoplus\limits_{a\geq 0} (X^{\otimes a})^{\Sigma_a} \otimes \bigoplus\limits_{b\geq 0} (X^{\otimes b})^{\Sigma_b}.
\end{split}
\addtolength{\jot}{-1ex}
\end{equation*}
The isomorphism on the first line again exists because of condition (2) of Definition \ref{defn:cofreefriendly} and the isomorphism at the beginning of the last line is derived from Definition \ref{defn:permutationfriendly}. The map $\mathrm{res}_{a,b}$ is the map
\[\mathrm{res}_{a,b} \colon (X^{\otimes a} \otimes X^{\otimes b})^{\Sigma_{a+b}} \rightarrow (X^{\otimes a} \otimes X^{\otimes b})^{\Sigma_a \times \Sigma_b}\]
induced by the inclusion $\Sigma_a \times \Sigma_b \rightarrow \Sigma_{a+b}$.
The counit morphism $\pi\colonS^c(X) \rightarrow X$ is given by the identity on $X$ and by the zero morphism on the other summands.
\end{prop}
\begin{ex}\label{ex:permutfriendlythings}
If $k$ is a field,
all the categories discussed in Example~\ref{ex:kmodcofreefriendly} are permutation friendly in addition to being cofree friendly. This means that the categories of nonnegatively graded (co)chain complexes and unbounded chain complexes of $k$-modules and the categories of simplicial and cosimplicial $k$-modules all admit well-defined cofree cocommutative coalgebra functors.
The permutation friendliness of each of these categories can be deduced from the permutation friendliness of the category of graded $k$-modules. A proof for graded $k$-modules concentrated in even degrees or for $k$ a field of characteristic $2$ can be found in Bourbaki \cite[p. IV.49]{bourbaki}. If the characteristic of $k$ is different from $2$ and $M$ is concentrated in odd degrees, $(M^{\otimes n})^{\Sigma_n}$ and $(M^{\otimes a} \otimes M^{\otimes b})^{\Sigma_a \times \Sigma_b}$ can be described explicitly in terms of a basis of $M$.
For an arbitrary graded module $M$, the claim follows from writing $M$ as the direct sum of its even and its odd degree part.
More concretely, if $M$ admits a countable basis $m_i$, $i \geq 1,$ and the $m_i$ are even degree elements, or if the characteristic of $k$ is $2$, we identify $S^c(M)$ with the Hopf algebra $\Gamma_k[m_1, m_2,\dotsc]$. As an algebra, $\Gamma_k[m_1, m_2,\dotsc]$ is the divided power algebra generated by $M$. The isomorphism to $S^c(M)$ is given by identifying the element $\gamma_{j_1}(m_{i_1})\dotsm\gamma_{j_n}( m_{i_n}) \in \Gamma_k[m_1, m_2,\dotsc]$ with the element $\sum_{\sigma \in \mathrm{sh}(j_1,\dotsc,j_n)} \sigma.(m_{i_1}^{\otimes j_1} \otimes \dotsb \otimes m_{i_n}^{j_n}) \in S^c(M),$
where $\sigma$ ranges over the set of $(j_1,\dotsc,j_n)$-shuffles.
The coproduct is given on multiplicative generators by
\[\triangle(\gamma_j(m_i)) = \sum_{a+b=j} \gamma_a(m_i) \otimes \gamma_b(m_i).\]
If $M$ is concentrated in odd degrees and the characteristic of $k$ is different from $2$, we can identify $S^c(M)$ with the Hopf algebra $\Lambda_k(m_1, m_2,\dotsc).$ This is the exterior algebra generated by $M$, and we identify $m_{i_1}\dotsm m_{i_n}$ with $\sum_{\sigma \in \Sigma_n} \mathrm{sgn}(\sigma) \sigma.(m_{1} \otimes \dotsb \otimes m_{n})$. The coproduct is given by
\[\triangle(m_i) = 1 \otimes m_i + m_i \otimes 1.\]
If $k$ is not a field, the cofree cocommutative conilpotent coalgebra cogenerated by a $k$-module $M$ still exists: It is the largest cocommutative subcoalgebra of $T^c(M)$.
\end{ex}
Now that we have established what we mean by cofree coalgebras in $\mathcal{D}$, we turn to an analysis of
coHochschild homology for these coalgebras. These computations should be thought of as a simple case of a dual Hochschild--Kostant--Rosenberg theorem for coalgebras.
Let $(C,\triangle_C, \epsilon_C, \eta_C )$ and $(D,\triangle_D, \epsilon_D, \eta_D)$ be cocommutative coaugmented coalgebras. Recall that their product in $\mathrm{comCoalg}_{\D}$ exists and is given by $C \otimes D$ with comultiplication $ (C\otimes {\tau_{C,D}} \otimes D) (\triangle_C \otimes \triangle_D)$, counit $\epsilon_C \otimes \epsilon_D$ and coaugmentation $\eta_C \otimes \eta_D.$ Here $\tau_{C,D} \colon C \otimes D \rightarrow D \otimes C$ is the symmetry isomorphism switching $C$ and $D$.
The projections $C\otimes D \rightarrow C$ and $C\otimes D \rightarrow D$ are given by $C \otimes \epsilon_D$ and $\epsilon_D \otimes C$.
Since $S^c$ is a right adjoint, we have the following property:
\begin{lemma} \label{lem:symcofreeAndProducts}
If $\mathcal{D}$ is cofree-friendly, permutation-friendly and cocomplete, there is a natural isomorphism
\[S^c(X \times Y) \to S^c(X) \otimes S^c(Y)\]
in $\mathrm{comCoalg}_{\D}$.
It is given by the morphisms
$S^c(X \times Y) \rightarrow S^c(X)$ and $S^c(X \times Y) \rightarrow S^c(Y)$ which are induced by the projections $X \times Y \rightarrow X$ and $X \times Y \rightarrow Y$.
\end{lemma}
Every category $\mathcal{C}$ that admits finite products gives rise to the symmetric monoidal category $(\mathcal{C}, \times, *)$, where $*$ is the terminal object. Every object $X$ in $\mathcal{C}$ is then a counital cocommutative
coalgebra with respect to this monoidal structure: the comultiplication $X \rightarrow X \times X$ is the diagonal, the counit $X \rightarrow *$ is the map to the terminal object. We indicate the monoidal structure we use to form $\mathrm{coTHH}$ with a subscript, so that if $C$ is a coalgebra with respect to $\otimes$, we write $\mathrm{coTHH}_{\otimes}^{\bullet}(C)$.
\begin{prop}\label{prop:abstractHKR}
Let $\mathcal{D}$ be cofree-friendly and permutation-friendly.
Then for any object $X$ in $\mathcal{D}$
\[\mathrm{coTHH}_{\otimes}^{\bullet}(S^c(X)) \cong S^c(\mathrm{coTHH}_{\times}^{\bullet}(X)).\]
\end{prop}
\begin{proof}
Applying Lemma \ref{lem:symcofreeAndProducts} yields an isomorphism in each cosimplicial degree $n$:
\[\mathrm{coTHH}^n_{\otimes}(S^c(X)) =S^c(X)^{\otimes n+1} \cong S^c(X^{\times n+1}) = S^c(\mathrm{coTHH}_{\times}^n(X)).\]
To check that the cosimplicial structures agree, it suffices to show that the comultiplications and counits on both sides agree. The diagonal $X \rightarrow X \times X$ induces a morphism of coalgebras
\[S^c(X) \rightarrow S^c(X \times X).\]
Since $S^c(X)$ is cocommutative, the comultiplication
\[\triangle_{S^c(X)} \colon S^c(X) \rightarrow S^c(X) \otimes S^c(X) \cong S^c(X \times X)\]
is a morphism of coalgebras as well. These maps agree after projection to $X \times X$.
A similar argument yields that $X \rightarrow 0$ induces the counit of $S^c(X)$.
\end{proof}
The following theorem is a consequence of this proposition. This is the best result we obtain about the coHochschild homology of cofree coalgebras without making further assumptions on our symmetric monoidal category $\mathcal{D}$.
\begin{thrm}\label{thrm:coTHHcofreeinsymmonoidalmodelcat}
If $\mathcal{D}$ is a cofree- and permutation-friendly model category, Proposition \ref{prop:abstractHKR} implies that
\[\mathrm{coTHH}_{\otimes}(S^c(X)) \sim \mathrm{holim}_{\Delta}(S^c(\mathrm{coTHH}^{\bullet}_{\times}(X))).\]
\end{thrm}
If $\mathcal{D}$ is a simplicial model category, there is an easy description of $\Tot(\mathrm{coTHH}^{\bullet}_{\times}(X))$ as a free loop object.
\begin{prop}[see \emph{e.g.\,\,} \cite{loday-loops}]\label{prop:coTHHcotensorS1simpmodelcat}%
Let $\mathcal{D}$ be a simplicial model category and $X$ a fibrant object of $\mathcal{D}$.
Then
\[\mathrm{holim}_{\Delta}(\mathrm{coTHH}^{\bullet}_{\times}(X)) \sim X^{\mathbb{S}^1}.\]
\end{prop}
{This result identifies $\mathrm{coTHH}_\times(X)$ as a sort of ``free loop space'' on $X$. This is precisely the case in the category of spaces, as in Example \ref{ex:freeloopspace}.}
\begin{ex}
For the category of cosimplicial $k$-modules we can choose the diagonal as a model for the homotopy limit over $\Delta$. Since the cofree coalgebra cogenerated by a cosimplicial $k$-module is given by applying $S^c$ degreewise, we actually obtain an isomorphism
of coalgebras
\[\mathrm{coTHH}_{\otimes}(S^c(X))\cong S^c(X^{\mathbb{S}^1}).\]
\end{ex}
In the case where $S^c$ is a right Quillen functor, we obtain the following corollary to Theorem \ref{thrm:coTHHcofreeinsymmonoidalmodelcat} and Proposition \ref{prop:coTHHcotensorS1simpmodelcat}.
\begin{cor}
Let $\mathcal{D}$ be a simplicial and a monoidal model category. Let $X$ in $\mathcal{D}$ be fibrant. If there is a model structure on $\mathrm{comCoalg}_{\D}$ such that $S^c$ is a right Quillen functor, then
\[\mathrm{coTHH}_{\otimes}(S^c(X))\cong \mathbb{R}S^c( X^{\mathbb{S}^1}).\]
\end{cor}
In general, however, $S^c$ is not a right Quillen functor. Up to a shift in degrees which is dual to the degree shift in the free commutative algebra functor, the most prominent counterexample is the category of chain complexes over a ring $k$ of positive characteristic: The chain complex $D^2$ consists of one copy of $k$ in degrees $1$ and $2$, with differential the identity. This complex is fibrant and acyclic, but $S^c(D^2)$ is not acyclic. This same counterexample shows $S^c$ is not right Quillen either for unbounded chain complexes or nonnegatively graded cochain complexes.
Nevertheless, a Hochschild--Kostant--Rosenberg type theorem for cofree coalgebras still holds for coHochschild homology in the category of nonnegatively graded cochain complexes $\mathrm{dgmod}^{\geq 0}$ over a field.
We denote the underlying differential graded $k$-module of $S^c(X)$ by $U(S^c(X))$. The desuspension $\Sigma^{-1}X$ of a cochain complex $X$ is the cochain complex with $(\Sigma^{-1}X)^n = X^{n-1}$.
Define the graded $k$-module $\Omega^{S^c(X)\vert k}$ by
\[\Omega^{S^c(X)\vert k} = \begin{cases}
U(S^c(X)) \otimes U(S^c(\Sigma^{-1}X)), & \text{if } \mathrm{char}(k) \neq 2,\\
U(S^c(X)) \otimes \Lambda(\Sigma^{-1}X),& \text{if } \mathrm{char}(k)=2,
\end{cases}
\]
with $\Lambda(M)$ denoting the exterior powers of the cochain complex $M$. We will compare $\Omega^{S^c(X) \vert k}$ with the notion of K\"ahler codifferentials as defined in \cite{farinati-solotar-ext} in Remark \ref{rem:KaehlerFS}.
\begin{thrm}\label{thrm:HKRcofreedg}
Let $k$ be a field. Then for $X$ in $\mathrm{dgmod}^{\geq 0}$ there is a quasi-isomorphism
\[\mathrm{coTHH}_\otimes(S^c(X))\to \Omega^{S^c(X)\vert k}.\]
\end{thrm}
Note that this is an identification of differential graded $k$-modules, not of coalgebras. We determine the corresponding coalgebra structure on $\mathrm{coTHH}_{\otimes}(S^c(X))$ in certain cases in Proposition \ref{cohhcomp}.
This result corresponds to the result of the Hochschild--Kostant--Rosenberg theorem applied to a free symmetric algebra $S(X)$ generated by a chain complex $X$. See, for example, \cite[Theorems 3.2.2 and 5.4.6]{loday}. For a discussion of a coalgebra analogue of K\"ahler differentials we refer the reader to \cite{farinati-solotar-ext}.
The proof is dual to the proof of the corresponding results for Hochschild homology. We follow the line of proof given by Loday \cite[Theorem 3.2.2]{loday}. We begin by proving a couple of lemmas.
First we identify $\mathrm{coTHH}_{\times}(X)$:
\begin{lemma}
For a cochain complex $X$ over a field $k$,
\[\Tot_{\mathrm{dgmod}} (N^* (\mathrm{coTHH}^{\bullet}_{\times}(X) )) \cong X \times \Sigma^{-1}X.\]
\end{lemma}
\begin{proof}
This follows easily from the fact that
$$N^0(\mathrm{coTHH}^{\bullet}_{\times}(X)) = X, \qquad N^1(\mathrm{coTHH}^{\bullet}_{\times}(X)) =0 \times X$$
and $N^a(\mathrm{coTHH}^{\bullet}_{\times}(X)) =0$ for $a\geq 2$.
\end{proof}
We next prove the special case of Theorem \ref{thrm:HKRcofreedg} where the module of cogenerators is one dimensional.
\begin{lemma}\label{lem:HKROneGenerator}
Let $k$ be a field. Let $X$ in $\mathrm{dgmod}^{\geq 0}$ be concentrated in a single nonnegative degree and be one dimensional in this degree. Then there is a quasi-isomorphism
\[\mathrm{coTHH}_{\otimes}(S^c(X)) \rightarrow \Omega^{S^c(X)\vert k}.\]
\end{lemma}
\begin{proof}
First let the characteristic of $k$ be $2$ or $X$ be concentrated in even degree, so that $S^c(X) \cong \Gamma_k[x]$ for a generator $x$. Up to the internal degree induced by $X$, we can use the results of Doi \cite[3.1]{doi} to compute $\mathrm{coTHH}_{\otimes}(S^c(X))$ using a $S^c(X) \otimes S^c(X)$-cofree resolution of $S^c(X)$. Such a resolution is given by
$$\xymatrix{S^c(X) \ar[r]^-{\triangle} & S^c(X) \otimes S^c(X) \ar[r]^-{f} & S^c(X) \otimes X \otimes S^c(X) \ar[r] & 0,}$$
with $f(x^i \otimes x^j) = x^{i-1} \otimes x \otimes x^j - x^i \otimes x \otimes x^{j-1}$, where $x^{-1}=0$. Hence the homology of $\mathrm{coTHH}_{\otimes}(S^c(X))$ is concentrated in cosimplicial degrees $0$ and $1$. An explicit calculation gives that all elements in $N^0(\mathrm{coTHH}_{\otimes}^{\bullet}(S^c(X))) \cong S^c(X)$ are cycles, while generating cycles in $N^1(\mathrm{coTHH}_{\otimes}^{\bullet}(S^c(X))) $
are given by $\sum_{i=1}^{n} i \cdot x^{n-i} \otimes x^i$ for $n\geq 1$. We identify $x^n$ in cosimplicial degree zero with $x^n \otimes 1 \in U(S^c(X)) \otimes \Lambda(\Sigma^{-1}X)$ and $\sum_{i=1}^{n} i \cdot x^{n-i} \otimes x^i$ with $x^{n-1}\otimes \sigma^{-1}x \in U(S^c(X)) \otimes \Lambda(\Sigma^{-1}X).$
Now assume that $X$ is concentrated in odd degree and that the characteristic of $k$ is not $2$, so that $S^c(X) = \Lambda_k(x)$.
Hence a typical element in $\mathrm{coTHH}^a(S^c(X))$ is of the form
$y=x^{t_0} \otimes x^{t_1} \otimes \dotsb \otimes x^{t_a}$ with $t_i \in \lbrace 0,1 \rbrace$. Now $y\in N^a(\mathrm{coTHH}^{\bullet}(S^c(X)))$ if and only if $t_1=\dotsb=t_a = 1$, and all differentials in $N^*(\mathrm{coTHH}^{\bullet}(S^c(X)))$ are trivial. Identifying $1 \otimes x^{\otimes a}$ with $1 \otimes (\sigma^{-1} x)^a \in U(S^c(X)) \otimes U(S^c(\Sigma^{-1}X))$ and $x \otimes x^{\otimes a}$ with $x \otimes (\sigma^{-1}x)^{\otimes a} \in U(S^c(X)) \otimes U(S^c(\Sigma^{-1}X))$ yields the result.
\end{proof}
The general case follows from the interplay of products and the cofree functor and the following result.
\begin{lemma}[Cf. {\cite[p. 719]{michaelis}}]\label{lem:infDimCoalg}
Let $k$ be a field and let $X$ be a graded $k$-module. Then
\[S^c(X) \cong \operatorname*{colim}_{V\subset X \text{\,fin\,dim}} S^c(V),\]
where the canonical projection $\pi_X \colon S^c(X)\cong\operatorname*{colim}_{V\subset X \text{fin\,dim}} S^c(V) \rightarrow X$ is given by the colimit of the maps $S^c(V)\xto{\pi_V} V\to X$.
\end{lemma}
\begin{proof}[Proof of Theorem~\ref{thrm:HKRcofreedg}]
Recall from Example \ref{ex:ch} that $\mathrm{coTHH}(C)$ of a counital coalgebra in $\mathrm{dgmod}^{\geq 0}$ can be computed as
\[\mathrm{coTHH}_{\otimes}(C) = \Tot_{\mathrm{dgmod}} (N^* \mathrm{coTHH}^{\bullet}_{\otimes}(C)).\]
Assume first that $X$ has trivial differential. If $X$ is finite dimensional,
\[\mathrm{coTHH}^{\bullet}_{\otimes}(S^c(X)) \cong \mathrm{coTHH}^{\bullet}_{\otimes}(S^c(X_1)) \otimes ... \otimes \mathrm{coTHH}^{\bullet}_{\otimes}(S^c(X_n))\]
for one dimensional $k$-vector spaces $X_i$. Hence Lemma \ref{lem:HKROneGenerator} and the dual Eilenberg--Zilber map yield the desired quasi-isomorphism. Lemma \ref{lem:infDimCoalg} proves Theorem \ref{thrm:HKRcofreedg} for infinite dimensional cochain complexes $X$ with trivial differential.
If $X$ is any nonnegatively graded cochain complex, applying the result for graded $k$-modules which we just proved
shows that there is a morphism of double complexes
\[N^*(\mathrm{coTHH}^{\bullet}_{\otimes}(S^c(X))) \rightarrow \Omega^{S^c(X)\vert k}\]
which is a quasi-isomorphism on each row, that is, if we fix the degree induced by the grading on $X$. Hence this induces a quasi-isomorphism on total complexes.
\end{proof}
\begin{rem} \label{rem:KaehlerFS}
Farinati and Solotar \cite[Section 3]{farinati-solotar-ext} define a symmetric $C$-bicomodule $\Omega^1_C$ and a coderivation $d \colon \Omega^1_C \rightarrow C$ for any (ungraded) coaugmented coalgebra $C$ such that $(\Omega^1_C, d)$ satisfies the following universal property: Every coderivation $f\colon M \rightarrow C$ from a symmetric $C$-bicomodule $M$ to $C$ factors as
$$f= d \circ \tilde f$$
for a unique $C$-bicomodule morphism $\tilde f$. Farinati and Solotar also give a construction of $(\Omega^1_C, d)$ dual to the construction of the module of K\"ahler differentials associated to a commutative algebra.
To compare this with our definition of $\Omega^{S^c(X)\vert k}$, note that for $X$ concentrated in degree zero
$$\Omega^1_{S^c(X)} \cong S^c(X) \otimes X$$
as a $S^c(X)$-bicomodule. Since $S^c(X)$ is cofree, coderivations into $S^c(X)$ correspond to maps into $X$, and the coderivation $d \colon S^c(X) \otimes X \rightarrow S^c(X)$ is induced by the map $\epsilon \otimes \mathrm{id} \colon S^c(X) \otimes X \rightarrow X$.
If the characteristic of $k$ is different from $2$, this yields that $\Omega^{S^c(X) \vert k}$ coincides with the exterior coalgebra $\Lambda_{S^c(X)}(\Omega^1_{S^c(X)})$ on $\Omega^1_{S^c(X)}$ defined in \cite[Section 6]{farinati-solotar-ext}.
\end{rem}
\begin{rem}
Analogous to Proposition \cite[5.4.6]{loday}, the proof of Theorem \ref{thrm:HKRcofreedg} shows that the Theorem actually holds for any differential graded cocommutative couaugmented coalgebra $C$ whenever the underlying graded cocommutative coaugmented coalgebra is cofree.
\end{rem}
\begin{rem}
A similar result holds for unbounded chain complexes if we define $\Tot_{\mathrm{dgmod}}$ and hence $\mathrm{coTHH}$ as a direct sum instead of a product in each degree. For nonnegatively graded cochain complexes both definitions of $\Tot_{\mathrm{dgmod}}$ agree. However, if we use the homotopically correct definition of $\mathrm{coTHH}$ for unbounded chain complexes via the product total complex, Lemma \ref{lem:HKROneGenerator} doesn't hold for vector spaces $X$ that are concentrated in degree $-1$.
\end{rem}
\section{A coB\"okstedt Spectral Sequence}\label{sect:thespectralsequence}
While in algebraic cases we are able to understand certain good examples of $\mathrm{coTHH}$ by an analysis of the definition, in topological examples we require further tools. In this section, we construct a spectral sequence, which we call the {\em coB\"okstedt spectral sequence.} {Recall that for a field $k$ and a ring spectrum $R$, the skeletal filtration on the simplicial spectrum $\mathrm{THH}(R)_{\bullet}$ yields a spectral sequence
\[
E^2_{*,*} = \mathrm{HH}_*(H_*(R;k)) \Rightarrow H_*(\mathrm{THH}(R);k),
\] called the B\"okstedt spectral sequence \cite{Bo2}. Analogously, for a coalgebra spectrum $C$, we consider the Bousfield--Kan spectral sequence arising from the cosimplicial spectrum $\mathrm{coTHH}^{\bullet}(C)$. We show in Theorem \ref{sscoalgebra} that this is a spectral sequence of coalgebras. Since our spectral sequence is an instance of the Bousfield--Kan spectral sequence, its convergence is not immediate, but in the case where $C$ is a suspension spectrum $\Sigma_{+}^{\infty} X$ with $X$ a simply connected space, {we provide conditions in Corollary~\ref{cor.conv.2} below under which this spectral sequence converges to }$H_{*}(\mathrm{coTHH}(\Sigma_{+}^{\infty} X);k).$
Let $(\mathrm{Spec}, \wedge , \mathbb{S})$ denote a symmetric monoidal category of spectra, such as those given by \cite{ekmm}, \cite{hss}, or \cite{mmss}. The notation $\wedge$ means ``smash product over $\mathbb{S}$.'' Let $C$ be a coalgebra in this category with comultiplication $\triangle\colon C \to C\wedge C$ and counit $\varepsilon\colon C \to \mathbb{S}$. Assume that $C$ is cofibrant as a spectrum so that $\mathrm{coTHH}(C)$ has the correct homotopy type, as in Remark \ref{rem:fibrancyassumptionsforholim}. Note that the spectral homology of $C$ with coefficients in a field $k$ is a graded $k$-coalgebra with structure maps
\begin{align*}
\triangle\colon H_*(C; k) &\xrightarrow{H_*(\triangle)} H_*(C \wedge C; k)\cong H_*(C; k)\otimes_{k} H_*(C; k), \\
\varepsilon\colon H_*(C; k) &\xrightarrow{H_*(\varepsilon)} k .
\end{align*}
\begin{thm}\label{thm:SpecSeq} Let $k$ be a field. Let $C$ be a coalgebra spectrum that is cofibrant as a spectrum. The Bousfield--Kan spectral sequence for the cosimplicial spectrum $\mathrm{coTHH}^\bullet(C)$ gives a \emph{coB\"okstedt spectral sequence} for calculating $H_{t-s}(\mathrm{coTHH}(C);k)$ with $E_2$-page
\[ E_2^{s, t}=\mathrm{coHH}^{k}_{s,t}(H_*(C;k)).\]
given by the classical coHochschild homology of $H_*(C;k)$ as a graded $k$-module.
\end{thm}
\begin{proof}
The spectral sequence arises as the Bousfield--Kan spectral sequence of a cosimplicial spectrum. We briefly recall the general construction. Let $X^\bullet$ be a Reedy fibrant cosimplicial spectrum and recall that
\[\Tot (X^{\bullet}) = \text{eq}\Big(
\prod_{n \geq 0} (X^n)^{\Delta^n} \rightrightarrows \prod_{\alpha \in \Delta([a],[b])} (X^b)^{\Delta^a}
\Big),\]
where $\Delta$ is the cosimplicial space whose $m$-th level is the standard $m$-simplex $\Delta^m$. Let $\mathrm{sk}_n \Delta \subset \Delta$ denote the cosimplicial subspace whose $m$-th level is $ sk_n \Delta^m$, the $n$-skeleton of the $m$-simplex and set
\[\Tot_n (X^{\bullet}) := \text{eq}\Big(
\prod_{m \geq 0} (X^m)^{\mathrm{sk}_n \Delta^m} \rightrightarrows \prod_{\alpha \in \Delta([a],[b])} (X^b)^{\mathrm{sk}_n \Delta^a}
\Big).\]
The inclusions $sk_n \Delta \hookrightarrow sk_{n+1} \Delta$ induce the maps in a tower of fibrations
\begin{align*}
\dotsb \to \Tot_n(X^\bullet) \xrightarrow{p_n} \Tot_{n-1}(X^\bullet) \xrightarrow{} \dotsb \to \text{Tot}_0(X^\bullet) \cong X^0.
\end{align*}
Let $F_n \xrightarrow{i_n} \text{Tot}_n(X^\bullet) \xrightarrow{p_n} \text{Tot}_{n-1}(X^\bullet)$ denote the inclusion of the fiber and consider the associated exact couple:
\[\begin{tikzpicture}
\node(a){$\pi_*(\Tot_*(X^\bullet) ) $};
\node(b)[right of=a, node distance = 3cm]{$\pi_*(\Tot_*(X^\bullet))$};
\node(c)[below of= a,node distance = 1.5cm, right of=a, node distance = 1.5cm]{$\pi_*(F_*).$};
\draw[->](a) to node [above]{$p_*$} (b);
\draw[->](b) to node [below right]{$\partial$} (c);
\draw[->](c) to node [below left]{$i_*$} (a);
\end{tikzpicture}\]
This exact couple gives rise to a cohomologically graded spectral sequence $\{E_r, d_r\}$ with $E_1^{s,t}=\pi_{t-s}(F_{s})$ and differentials $d_r\colon E_r^{s,t} \to E_r^{s+r,t+r-1}$. It is a half plane spectral sequence with entering differentials.
The fiber $F_{s}$ can be identified with $\Omega^s (N^s X^\bullet)$, where
\[N^sX^\bullet=\mathrm{eq}\Big( X^s\xrightrightarrows[*]{(\sigma^0,\dots, \sigma^{s-1})}\prod_{i=0}^{s-1} X^{s-1}\Big)\]
We have isomorphisms
\[E_1^{s,t} =\pi_{t-s}(\Omega^s(N^s X^\bullet)) \cong \pi_{t}(N^s X^\bullet) \cong N^s \pi_t(X^\bullet)\]
and under these isomorphisms the differential $d_1\colon N^s \pi_t(X^\bullet) \to N^{s+1} \pi_t(X^\bullet)$ is identified with $\sum (-1)^i \pi_t(\delta^i)$. Since the cohomology of the normalized complex agrees with the cohomology of the cosimplicial object, we conclude that the $E_2$ term is
\[E_2^{s,t}\cong H^s(\pi_t(X^\bullet), \textstyle\sum (-1)^i \pi_t(\delta^i)).\]
Let $C$ be a coalgebra spectrum and consider the spectral sequence arising from the Reedy fibrant replacement $R(\mathrm{coTHH}^{\bullet}(C)\wedge H k)$. We have isomorphisms
\[
\pi_* R(\mathrm{coTHH}^{n}(C)\wedge H k) \cong \pi_*(\mathrm{coTHH}^{n}(C) \wedge H k) \cong H_*(C;k)^{\otimes_{k}(n+1)}
\]
The map $\pi_*(\delta^i)$ corresponds to the $i$'th coHochschild differential under this identification, therefore
\[E_2^{s,t}= \mathrm{coHH}^{k}_{s,t}(H_*(C;k)). \qedhere\]
\end{proof}
From~\cite[IX 5.7]{BK}, we have the following statement about the convergence of the coB\"okstedt spectral sequence. {See also~\cite[VI.2]{goerss-jardine} for a discussion of complete convergence.} In~\cite[IX 5]{BK}, the authors restrict to $i \geq 1$ so that they are working with groups. Since we have abelian groups for all $i$, that restriction is not necessary here.
\begin{prop}\label{prop:bkconvergence}
If for each $s$ there is an $r$ such that $E_r^{s, s+i} = E_{\infty}^{s, s+i}$, then the coB\"okstedt spectral sequence for $\mathrm{coTHH}(C)$ converges completely to \[\pi_* \Tot R(\mathrm{coTHH}^\bullet(C) \wedge Hk). \]
\end{prop}
There is a natural map
\[P\colon \Tot(R\mathrm{coTHH}^\bullet(C)) \wedge Hk \to \Tot R(\mathrm{coTHH}^\bullet(C) \wedge Hk)\]
arising from the natural construction of a map of the form $\mathrm{Hom}(X,Y)\wedge Z \to \mathrm{Hom}(X,Y\wedge Z)$.
Since $\pi_*(\Tot(R\mathrm{coTHH}^\bullet(C)) \wedge Hk) \cong H_{*}(\mathrm{coTHH}(C);k),$ we have the following.
\begin{cor}\label{cor.conv.1}
If the conditions on $E_r^{s, s+i}$ in Proposition \ref{prop:bkconvergence} hold and the map $P$ described above induces an isomorphism in homotopy, then the coB\"okstedt spectral sequence for $\mathrm{coTHH}(C)$ converges completely to $H_{*}(\mathrm{coTHH}(C);k)$.
\end{cor}
\begin{ex}\label{ex:freeloopspace}
Let $X$ be a simply connected space, let $d\colon X \to X \times X$ denote the diagonal map and let $c\colon X \to *$ be the map collapsing $X$ to a point. We form a cosimplicial space $X^{\bullet}$ with
$X^n = X^{\times (n+1)}$ and with cofaces
\[
\delta_i \colon X^{\times (n+1)} \rightarrow X^{\times (n+2)}, \quad \delta_i=\begin{cases}
X^{\times i} \times d \times X^{\times (n-i)}, & 0\leq i \leq n,\\
{\tau_{X, X^{\times (n+1)}}}(d \times X^{\times n}), & i=n+1,
\end{cases}\]
and codegeneracies
\[
\sigma_i \colon X^{\times (n+2)} \rightarrow X^{\times (n+1)}, \quad
\sigma_i = X^{\times (i+1)} \times c \times X^{\times (n-i)} \quad \text{for } 0 \leq i \leq n.
\]
The cosimplicial space $X^\bullet$ is the cosimplicial space $\mathrm{coTHH}^\bullet(X)$ of Definition \ref{defn:coTHH} for the coalgebra $(X,d,c)$ in the category of spaces; see Example \ref{ex:simpset}. As noted earlier in Proposition \ref{prop:coTHHcotensorS1simpmodelcat}, the cosimplicial space $X^\bullet$ totalizes to the free loop space $\mathcal{L}X$; see Example 4.2 of \cite{bousfield}.
We can consider the spectrum $\Sigma_+^{\infty} X$ as a coalgebra with comultiplication arising from the diagonal map $d\colon X \to X \times X$ and counit arising from $c\colon X \to *$. We have an identification of cosimplical spectra
\[
\mathrm{coTHH}^{\bullet}(\Sigma_+^{\infty} X)=\Sigma^{\infty}_+( X^{\bullet}).
\]
The topological coHochschild homology is the totalization of a Reedy fibrant replacement of the above cosimplicial spectrum. We have a natural map
\[
\Sigma^{\infty}_+ \Tot(X^{\bullet}) \to \Tot(\Sigma^{\infty}_+ X^{\bullet}),
\]
as Malkiewich describes in {Section 2 of \cite{malkiewich}}, which after composing with the map to the totalization of a Reedy fibrant replacement becomes a stable equivalence {for simply connected $X$; see Proposition 2.22} of \cite{malkiewich}. Hence we have a stable equivalence
\[
\Sigma^\infty_+ \mathcal{L}X \xrightarrow{\cong} \mathrm{coTHH}(\Sigma_+^{\infty} X).
\]
\end{ex}
\begin{cor}\label{cor.conv.2}
{Let $X$ be a simply connected space. If for each $s$ there is an $r$ such that $E_r^{s, s+i} = E_{\infty}^{s, s+i}$, the coB\"okstedt spectral sequence arising from the coalgebra $\Sigma_+^{\infty} X$ converges completely to
\[H_{*}(\mathrm{coTHH}(\Sigma_+^{\infty} X);k) \cong H_*(\mathcal{L}X; k).\]
}
\end{cor}
\begin{proof}
Since $X$ is simply connected, by \cite[2.22]{malkiewich}
$H_{*}(\mathrm{coTHH}(\Sigma_+^{\infty} X);k) \cong H_*(\mathcal{L}X; k).$ As in Corollary~\ref{cor.conv.1}, we need to show that there is a weak equivalence
\begin{equation}\label{TotCompare} \Tot (R\mathrm{coTHH}^\bullet(\Sigma_+^{\infty} X)) \wedge Hk \simeq \text{Tot}R(\mathrm{coTHH}^\bullet(\Sigma_+^{\infty} X) \wedge Hk).\end{equation}
First we consider the $n$th stage approximations:
\[\Tot_n (R\mathrm{coTHH}^\bullet(\Sigma_+^{\infty} X)) \wedge Hk \simeq \text{Tot}_n R(\mathrm{coTHH}^\bullet(\Sigma_+^{\infty} X) \wedge Hk).\]
The $n$th stages are weakly equivalent because the derived $\Tot_n$ functor is a finite homotopy limit functor. In spectra, this agrees with a finite homotopy colimit functor and hence the derived $\Tot_n$ commutes with smashing with $Hk$. Since these weak equivalences are compatible for all $n$, it follows that the inverse limits over $n$ of both towers are weakly equivalent. On the right hand side, this inverse limit agrees with the right hand side of Equation \ref{TotCompare}. For the left hand sides to agree, we need to commute the inverse limit with $- \wedge Hk$. This is possible here because of the properties of this specific tower.
In~\cite[2.22]{malkiewich}, and in more detail in an {earlier version~\cite[4.6]{malkiewichv1}}, Malkiewich shows that the connectivity of the fibers of the tower $ \{ \Tot_n R(\mathrm{coTHH}^\bullet(\Sigma_+^{\infty} X)) \}$ tend to infinity, and hence there is a pro-weak equivalence between this tower and the constant tower given by $\{\Tot R(\mathrm{coTHH}^\bullet(\Sigma_+^{\infty} X))\}$. That is, applying homotopy, $\pi_k$, produces a pro-isomorphism of groups for each $k$. By~\cite[8.5]{bousfield}, it follows that there is a pro-isomorphism for each $l$ between the constant tower $\{H_l\Tot(R\mathrm{coTHH}^\bullet(\Sigma_+^{\infty} X))\}$ and the tower $ \{ H_l\Tot_n R(\mathrm{coTHH}^\bullet(\Sigma_+^{\infty} X)) \}$. It follows that there is a pro-weak equivalence
\[ \{\Tot(R\mathrm{coTHH}^\bullet(\Sigma_+^{\infty} X)) \wedge Hk \} \xrightarrow{\simeq} \{\Tot_n(R\mathrm{coTHH}^\bullet(\Sigma_+^{\infty} X)) \wedge Hk \}.\]
By~\cite[III.3.1]{BK}, it follows that the inverse limits are weakly equivalent. That is, there is a weak equivalence
\[\Tot R(\mathrm{coTHH}^\bullet(\Sigma_+^{\infty} X)) \wedge Hk \xrightarrow{\simeq} \lim_n(\Tot_n R(\mathrm{coTHH}^\bullet(\Sigma_+^{\infty} X)) \wedge Hk).\]
This shows that $- \wedge Hk$ commutes with inverse limits here as desired.
\end{proof}
Our next result is to prove that the coB\"okstedt spectral sequence of Theorem \ref{thm:SpecSeq} is a spectral sequence of coalgebras. We exploit this additional structure in Section \ref{sect:computations} to make calculations of topological coHochschild homology in nice cases.
We first describe a functor $\mathrm{Hom}(-,C)\colon \mathrm{Set}^\mathrm{op}\to \mathrm{comCoalg}$, where $C$ is a cocommutative coalgebra. In particular this will give us a cosimplicial spectrum $\textup{Hom}(S^1_{\bullet},C)$ that agrees with the cosimplicial spectrum $\mathrm{coTHH}^{\bullet}(C)$ defined in Definition \ref{defn:coTHH}.
Let $X$ be a set and $C$ be a cocommutative coalgebra spectrum. Let $\mathrm{Hom}(X,C)$ be the indexed smash product
\[ \bigwedge_{x\in X} C.\]
If $f\colon X\to Y$ is a map of sets, we define a map of spectra $\mathrm{Hom}(Y,C)\to \mathrm{Hom}(X,C)$ as the product over $Y$ of the component maps
\[ C\to \bigwedge_{x\in f^{-1}(y)} C\]
given by iterated comultiplication over $f^{-1}(y)$. Here by convention the smash product indexed on the empty set is $\mathbb{S}$, and the map to $C\to \mathbb{S}$ is the counit map. Since $C$ is cocommutative, this map doesn't depend on a choosen ordering for applying the comultiplications and for each $X$, the comultiplication on $C$ extends to define a cocommutative comultiplication on $\mathrm{Hom}(X,C)$. Hence $\mathrm{Hom}(-,C)$ is a functor $\mathrm{Set}^\mathrm{op}\to \mathrm{comCoalg}$. If $X_*$ is a simplicial set, the composite
\[ (\Delta^\mathrm{op})^\mathrm{op} \xto{X_*^\mathrm{op}}\mathrm{Set}^\mathrm{op} \xto{\mathrm{Hom}(-,C)} \mathrm{comCoalg}\]
thus defines a cosimplicial cocommutative coalgebra spectrum.
To obtain the cosimplicial coalgebra yielding $\mathrm{coTHH}(C)$, we take $X_*$ to be the simplicial circle $\Delta^1_{\bullet}/\partial\Delta^1_{\bullet}$. Here $\Delta^1_{\bullet}$ is the simplicial set with $\Delta^1_n=\{x_0,\dotsc,x_{n+1}\}$ where $x_t$ is the function $[n]\to [1]$ so that the preimage of $1$ has order $t$. The face and degeneracy maps are given by:
\[ \delta_i(x_t)= \begin{cases} x_t & t\leq i\\ x_{t-1}& t>i \end{cases}\]
\[ \sigma_i(x_t)=\begin{cases}x_t & t\leq i\\ x_{t+1} & t>i\end{cases}\]
The $n$-simplices of $S^1_*$ are obtained by identifying $x_0$ and $x_{n+1}$ so that $S^1_n=\{x_0,\dotsc, x_n\}$. Comparing the coface and codegeneracy maps in $\mathrm{Hom}(S^1_{\bullet}, C)$ and $\mathrm{coTHH}^\bullet(C)$ shows that the two cosimplicial spectra are the same.
\begin{thm}\label{sscoalgebra}
Let $C$ be a connected cocommutative coalgebra that is cofibrant as a spectrum. Then the Bousfield--Kan spectral sequence described in Theorem \ref{thm:SpecSeq} is a spectral sequence of $k$-coalgebras. In particular, for each $r>1$ there is a coproduct
\[
\psi\colon E_r^{**} \rightarrow E_r^{**} \otimes_{k} E_r^{**},
\]
and the differentials $d_r$ respect the coproduct.
\begin{proof}
Below we will construct the coproduct using natural maps of spectral sequences
\[
E_r^{**} \xto{\ \nabla\ } D_r^{**} \xleftarrow{\ AW\ } E_r^{**} \otimes_{k} E_r^{**}
\]
where the map $AW$ is an isomorphism for $r\geq 2$.
Recall that the cosimplicial spectrum $\mathrm{coTHH}^\bullet(C)$ can be identified with the cosimplicial spectrum Hom($S^1_{\bullet}, C)$. Let $D_r^{**}$ denote the Bousfield--Kan spectral sequence for the cosimplicial spectrum $\mathrm{Hom}((S^1\amalg S^1)_{\bullet}, C)\wedge Hk$. The codiagonal map $\nabla\colon S^1\amalg S^1\to S^1$ is a simplicial map and gives a map $\mathrm{Hom}(S^1_{\bullet},C)\to \mathrm{Hom}((S^1\amalg S^1)_{\bullet},C).$ The map $\nabla\colon E_r^{**} \rightarrow D_r^{**}$ is induced by this codiagonal map. Let $X^{\bullet}$ denote the cosimplicial $Hk$-coalgebra spectrum $\mathrm{coTHH}(C)^{\bullet} \wedge Hk$. Note that the cosimplicial spectrum Hom($(S^1\amalg S^1)_{\bullet}, C) \wedge Hk$ is the cosimplicial spectrum $[n] \to X^n \wedge_{Hk} X^n$. This is the diagonal cosimplicial spectrum associated to a bicosimplicial spectrum, $(p, q) \to X^p \wedge_{Hk} X^q$. We will denote this diagonal cosimplicial spectrum by $\mathrm{diag}(X^{\bullet} \wedge_{Hk} X^{\bullet}$). The spectral sequence $D_r^{* *}$ is given by the Tot tower for the Reedy fibrant replacement $R (\mathrm{diag}(X^{\bullet} \wedge_{Hk} X^{\bullet}$)).
Note that our cosimplicial spectra are in fact cosimplicial $Hk$-modules. We can apply the standard equivalence between $Hk$-modules and chain complexes of $k$-modules \cite{shipley-hz}. By hypothesis, our $Hk$-modules are connective and therefore can be replaced by non-negatively graded chain complexes of $k$-modules. By the Dold--Kan correspondence, non-negatively graded chain complexes of $k$-modules are equivalent to simplicial $k$-modules. This also holds on the level of diagram categories and therefore our cosimplicial $Hk$-module spectra can be identified as cosimplicial simplicial $k$-modules. This string of equivalences is weakly monoidal in the sense of \cite{schwede-shipley-equiv}.
The Bousfield--Kan results in the simplicial setting \cite{ BKpairings, BKquadrant} thus apply, giving us a map of spectral sequences
\[
E_r^{**} \otimes_k E_r^{**} \to D_r^{**}.
\]
By Bousfield and Kan, on the $r=1$ page, this map is given by the Alexander--Whitney map:
\[
E_1^{**} \otimes_{k} E_1^{**} = N^*( \pi_*X^{\bullet}) \otimes_{k} N^*(\pi_*X^{\bullet}) \rightarrow N^*( \pi_*X^{\bullet} \otimes_{k} \pi_*X^{\bullet}) = D_1^{**}.
\]
By the cosimplicial analog of the Eilenberg--Zilber theorem (see, for example, \cite{fresse}) this induces an isomorphism on cohomology
\[
\xymatrix{
H^*(E_1^{*,*}) \otimes_{k} H^*(E_1^{*,*}) \cong H^*(E_1^{*,*}\otimes_{k} E_1^{*,*}) \ar[r]^-{\cong}& H^*(D_1^{*,*}) \cong D_{2}^{**} ,
}
\]
where the first isomorphism is given by the K\"unneth theorem.
This composite is the map $
E_{2}^{*,*} \otimes_{k} E_{2}^{*,*}\xto{\cong} D_{2}^{**} .
$
By induction and repeated application of the K\"unneth theorem, we get equivalences
$ E_{r}^{*,*} \otimes_{k} E_{r}^{*,*} \xto{\cong} D_{r}^{**}$
for all $r>1$. The composite of $\nabla$ with the inverse of this isomorphism gives us the desired coproduct.
\end{proof}
\end{thm}
\section{Computational results}\label{sect:computations}
Now that we have developed the coB\"okstedt spectral sequence and its coalgebra structure, we use these structures to make computations of the homology of $\mathrm{coTHH}(C)$ for certain coalgebra spectra $C$. We also prove that in certain cases this spectral sequence collapses.
We first make some elementary computations of coHochschild homology, which we will later use as input to the coHH to coTHH spectral sequence described in Theorem \ref{thm:SpecSeq}. The coalgebras we consider are the underlying coalgebras of Hopf algebras, specifically, of exterior Hopf algebras, polynomial Hopf algebras and divided power Hopf algebras. As mentioned in Example \ref{ex:permutfriendlythings}, the exterior Hopf algebra and the divided power Hopf algebras give the free cocommutative coalgebras on graded $k$-modules concentrated in odd degree or even degree, respectively. Recall from Example \ref{ex:permutfriendlythings} that $\Lambda_{k}(y_1, y_2, \dotsc)$ denotes the exterior Hopf algebra on generators $y_i$ in odd degrees and that $\Gamma_{k}[x_1, x_2, \dotsc]$ denotes the divided power Hopf algebra on generators $x_i$ in even degrees. Let $k[w_1,w_2,\dotsc]$ denote the polynomial Hopf algebra on generators $w_i$ in even degree. The coproduct is given by $\triangle(w_i^j)=\sum_{k} \binom{j}{k}\,w_i^k\otimes w_i^{j-k}.$
In the following computations, we assume that all the Hopf algebras in question only have finitely many generators in any given degree. We work over a field $k$.
The following computation of coHochschild homology is the main result we use as input for our spectral sequence.
\begin{prop}\label{cohhcomp}
Let the cogenerators $x_i$ be of even nonnegative
degree and let the cogenerators $y_i$ be of odd nonnegative degree. We evaluate the coHochschild homology as a coalgebra for the following cases of free cocommutative coalgebras:
\[
\mathrm{coHH}_*(\Gamma_{k}[x_1, x_2, \dotsc]) = \Gamma_{k}[x_1, x_2, \dotsc] \otimes \Lambda_{k}(z_1, z_2, \dotsc),
\]
where $\deg(z_i) = (1, \deg(x_i))$.
\[
\mathrm{coHH}_*(\Lambda_{k}(y_1, y_2, \dotsc)) = \Lambda_{k}(y_1, y_2, \dotsc) \otimes k[w_1, w_2, \dotsc]
\]
where $\deg(w_i) = (1, \deg(y_i))$.
\end{prop}
Before proving the proposition, we recall some basic homological coalgebra. Consider any coalgebra $C$ over a field $k$. Let $M$ be a right $C$-comodule and $N$ be a left $C$-comodule with corresponding maps $\rho_M\colon M \rightarrow M \otimes C$ and $\rho_N\colon N \rightarrow C \otimes N$. Recall that the cotensor product of $M$ and $N$ over $C$, $M \Box_C N$, is defined as the kernel of the map
\[
\rho_M \otimes 1 - 1 \otimes \rho_N\colon M \otimes N \to M \otimes C \otimes N. \]
The Cotor functors are the derived functors of the cotensor product. In other words, $\mathrm{Cotor}_C^n(M, N) = H^n(M \Box_C \bf{X})$ where $\bf{X}$ is an injective resolution of $N$ as a left $C$-comodule.
Let $C^e = C \otimes C^{op}$. Let $N$ be a $(C, C)$-bicomodule. Then $N$ can be regarded as a right $C^e$-comodule. As shown in Doi \cite{doi}, $\mathrm{coHH}_n(N, C) = \mathrm{Cotor}^n_{C^e}(N, C)$.
\begin{lemma}
For $C = \Gamma_{k}[x_1, x_2, \ldots]$ or $C= \Lambda_{k}(y_1, y_2, \ldots)$,
\[\mathrm{coHH}(C) = C \otimes \mathrm{Cotor}_C(k,k).\]
\end{lemma}
\begin{proof}
As above, $\mathrm{coHH}_n(C) = \mathrm{Cotor}^n_{C^e}(C, C)$. Observe that $C$ is a Hopf algebra. Let $S$ denote the antipode in $C$. As in Doi \cite[Section 3.3]{doi}, we consider the coalgebra map
\[
\nabla\colon C^e \rightarrow C
\]
given by $\nabla(c \otimes d^{op}) = cS(d)$. Given a $(C, C)$-bicomodule $M$, let $M_{\nabla}$ denote $M$ viewed as a right $C$-comodule via the map $\nabla$.
Let $k \rightarrow Y^0 \rightarrow Y^1 \rightarrow \cdots$ be an injective resolution of $k$ as a left $C$-comodule. When $C$ is a Hopf algebra, $(C^e)_{\nabla}$ is free as a right $C$-comodule, and $(C^e)_{\nabla}\Box_C Y^n$ is injective as a left $C^e$-comodule \cite{doi}. The sequence
\[
(C^e)_{\nabla}\Box_Ck \rightarrow (C^e)_{\nabla}\Box_C Y^0 \rightarrow(C^e)_{\nabla}\Box_C Y^1 \rightarrow \cdots
\]
is therefore an injective resolution of $(C^e)_{\nabla}\Box_Ck \cong C$ as a left $C^e$-comodule. Cotensoring over $C^e$ with $C$ we have:
\[
C \Box_{C^e}((C^e)_{\nabla}\Box_C Y^0) \rightarrow C \Box_{C^e}((C^e)_{\nabla}\Box_C Y^1) \rightarrow \cdots
\]
which gives:
\[
C_{\nabla}\Box_C Y^0 \rightarrow C_{\nabla}\Box_C Y^1 \rightarrow \cdots
\]
Therefore $\mathrm{coHH}_n(C) \cong \mathrm{Cotor}^n_{C^e}(C, C) \cong \mathrm{Cotor}^n_C(C_{\nabla}, k).$ In our cases $C$ is cocommutative, so $C_{\nabla}$ has the trivial $C$-comodule structure. Therefore $\mathrm{Cotor}^n_C(C_{\nabla},k) \cong C \otimes \mathrm{Cotor}^n_C(k, k).$
\end{proof}
We now want to compute $\mathrm{Cotor}_C(k, k)$ for the coalgebras $C$ that we are interested in.
\begin{prop}
For $C = \Gamma_{k}[x_1, x_2, \ldots]$, $\mathrm{Cotor}_C(k,k) =\Lambda_{k}(z_1, z_2, \ldots)$ where $\deg(z_i) = (1, \deg(x_i))$. For $C = \Lambda_{k}(y_1, y_2, \ldots)$, $\mathrm{Cotor}_C(k,k) = k[w_1, w_2, \ldots] $ where $\deg(w_i) = (1, \deg(y_i))$.
\end{prop}
\begin{proof}
Note that in both cases $C$ is a cocommutative coalgebra that in each fixed degree is a free and finitely generated $k$-module. Let $A$ denote the Hopf algebra dual of $C$. As in \cite{neisendorfer} we conclude that $\mathrm{Cotor}_C(k, k)$ is a Hopf algebra, which is dual to the Hopf algebra $\mathrm{Tor}^A(k,k)$.
Recall that the Hopf algebra dual of $\Gamma_{k}[x_1, x_2, \dotsc]$ is $k[x_1, x_2, \dotsc]$ and the Hopf algebra $\Lambda_{k}(y_1, y_2, \dotsc)$ is self dual. We recall the classical results:
\begin{align*}
\mathrm{Tor}^{k[x_1, x_2, \ldots]}(k,k) &\cong \Lambda_{k}(z_1, z_2, \ldots),\\
\mathrm{Tor}^{\Lambda_{k}(y_1, y_2, \ldots)}(k,k) &\cong \Gamma_{k}[w_1, w_2, \ldots]
\end{align*}
where the degree of $z_i$ is $(1, \deg(x_i))$ and the degree of $w_i$ is $(1, \deg(y_i))$. The result follows.
\end{proof}
We have now proven Proposition \ref{cohhcomp}. These results provide the input for the coB\"okstedt spectral sequence. They are the starting point for the following theorem, which says that this spectral sequence collapses at $E_2$ in the case where the homology of the input is cofree on cogenerators of even nonnegative degree.
\begin{thrm}\label{comp1} Let $C$ be a cocommutative coassociative coalgebra spectrum that is cofibrant as a spectrum, and whose homology coalgebra is
$$H_*(C;k) = \Gamma_{k}[x_1,x_2,\dots],$$
where the $x_i$ are cogenerators in nonnegative even degrees and there are only finitely many cogenerators in each degree. Then the coB\"okstedt spectral sequence for $C$ collapses at $E_2$, and
\[ E_2\cong E_{\infty} \cong \Gamma_{k}[x_1, x_2, \dots] \otimes \Lambda_{k}(z_1, z_2, \dots) ,\]
with $x_i$ in degree $(0, \deg(x_i))$ and $z_i$ in degree $(1, \deg(x_i))$.
\end{thrm}
\begin{comment}
\begin{proof}
From Theorem \ref{sscoalgebra}, the coB\"okstedt spectral sequence is a spectral sequence of graded coalgebras, so the differential on the $r$th page $d_r\colon E_r^{s,t} \to E_r^{s+r,t+r-1}$ satisfies the graded co-Leibniz rule
\[\triangle \circ d_r =(d_r \otimes \operatorname{id} + \operatorname{id} \otimes d_r)\circ \triangle.\]
We have adopted the Koszul sign convention in the above equation, \emph{i.e.\,\,} if $f, g\colon C\to D$ are graded maps and $x,y$ are homogeneous elements then
\[(f\otimes g)(x\otimes y)=(-1)^{\deg(g) \cdot \deg(x)}f(x) \otimes g(y).\]
By Proposition \ref{cohhcomp} we have an identification of the $E_2$-page:
\[
E_2^{s,t} = \mathrm{coHH}^{\mathbb{F}_p}_{s,t}(\Gamma_{\mathbb{F}_p}[x_1, x_2, \dots]) = \Gamma_{\mathbb{F}_p}[x_1, x_2, \dots] \otimes \Lambda_{\mathbb{F}_p}(z_1, z_2, \dots),
\]
with $x_i$ in degree $(0, \deg(x_i))$ and $z_i$ in degree $(1, \deg(x_i))$.
Recall that an element $x$ in a co-augmented coalgebra is called primitive if $\triangle(x)=x\otimes 1 + 1 \otimes x$. The primitive elements in $\Gamma_{\mathbb{F}_p}[x_1,x_2,\dots]$ are exactly the elements in the submodule $\mathbb{F}_p x_1 \oplus \mathbb{F}_p x_2 \oplus\cdots$,
and the primitive elements in $\Lambda_{\mathbb{F}_p}(z_1,z_2,\cdots)$ are exactly the elements in the submodule $\mathbb{F}_p z_1 \oplus \mathbb{F}_p z_2 \oplus\cdots$.
Moreover, if $A$ and $B$ are co-augmented $k$-coalgebras, then the primitive elements in $(A\otimes_k B, (\operatorname{id} \otimes \tau \otimes \operatorname{id}) \circ(\triangle_A \otimes \triangle_B))$ are the elements of the form $a\otimes 1$ or $1\otimes b$, where $a$ is a primitive element in $A$ or $b$ is a primitive element in $B$. Hence there are no primitive elements in $E^{s,t}_2$ when $s\geq 2$.
Fix $r \geq 2$. We show that $d_r$ is zero because the co-Leibniz rule forces the image of the lowest degree non-vanishing differential to be primitive. First we see that
\begin{align*}
\triangle \circ d_r (1) =(d_r \otimes 1 + 1 \otimes d_r)(1\otimes 1) =d_r(1) \otimes 1 +1 \otimes d_r(1).
\end{align*}
Thus $d_r(1)$ is a primitive element in $E_2^{r,r-1}$ and therefore $d_r(1)=0$. Next assume that all differentials originating from $E_r^{i,j}$ with $i+j<t+s$ vanish. Let $y\in E^{s,t}$. Note that
\[\triangle(y)=y\otimes 1 + 1\otimes y + \sum y_{i1} \otimes y_{i2},\]
where $y_{i1}, y_{i2}\in E^{i,j}$ with $i+j<t+s$. It follows from the inductive hypothesis that
\begin{align*}
\triangle \circ d_r (y)&=(d_r \otimes 1 + 1 \otimes d_r)\circ \triangle(y) =d_r(y) \otimes 1 + 1 \otimes d_r(y).
\end{align*}
Hence $d_r(y)$ is a primitive element in $E^{s+r,t+ r-1}_2$. Since there are no primitive elements in $E_2^{s,t}$ for $s\geq 2$, $d_r=0$ and the spectral sequence collapses at $E_2$.
\end{proof}
\end{comment}
\begin{proof}
By Proposition \ref{cohhcomp},
\[\mathrm{coHH}_*(\Gamma_{k}[x_1, x_2, \dots] ) = \Gamma_{k}[x_1, x_2, \dots] \otimes \Lambda_{k}(z_1, z_2, \dots).\]
A conilpotent coalgebra $C$ in graded $k$-modules is called \emph{cogenerated} by $Y$ if there is a surjection $\pi \colon C \rightarrow Y$ such that
\[ C \xto{(\triangle^{n})_{n}} T^c(C) \xto{T^c(\pi)} T^c(Y)\]
is injective. Every cofree cocommutative coalgebra $S^c(X)$ is cogenerated by $X$ under the canonical projection $\pi\colon S^c(X) \to X$, since the composite
\[S^c(X) \xto{(\Delta^{n})_{n}} T^c(S^c(X)) \xto{T^c(\pi)} T^c(X)\]
is the inclusion of $S^c(X)$ into $T^c(X)$.
Let $p$ be the characteristic of $k$. If $p\neq 2$, $\mathrm{coHH}_*(\Gamma_{k}[x_1, x_2, \dots] )$ is the cofree cocommutative conilpotent coalgebra cogenerated by $\lbrace x_1, x_2, \dots, z_1,z_2, \dots \rbrace$ with respect to the total grading. If $p=2$, it is easy to verify that $\Gamma_{k}[x_1, x_2, \dots] \otimes \Lambda_{k}(z_1, z_2, \dots)$ is cogenerated by $\lbrace x_1, x_2, \dots, z_1,z_2, \dots \rbrace$ as well.
If $C$ is cogenerated by $\pi\colon C \rightarrow Y$, any coderivation $d \colon C \rightarrow C$ is completely determined by $\pi d$: Since
\[\triangle^{n} d = \sum_{i=1}^{n+1}C^{\otimes i-1} \otimes d \otimes C^{\otimes n+1-i}\circ\triangle^{n}\]
and hence
\[ T^c(\pi) (\triangle^{n})_n d = (\sum_{i=1}^{n+1}\pi^{\otimes i-1} \otimes \pi d \otimes \pi^{\otimes n+1-i}\triangle^{n})_n,\]
the injectivity of $T^c(\pi) (\triangle^{n})_n$ yields that $\pi d$ determines $d$.
In particular, if $\pi d =0$, this implies $d=0$.
We now return to the coB\"okstedt spectral sequence. Assume that for some $r \geq 2$ we already know that $d^2,\dotsc, d^{r-1}$ vanish and thus the pages $E_2 \cong \dotsb \cong E_{r-1}$ are as claimed. The differential $d^r$ has bidegree $(r, r-1)$. Since the cogenerators are of bidegrees $(0,t)$ and $(1,t)$ and $E^{s,t}_{r-1} =0$ if $s<0$, the cogenerators cannot be in the image of $d^r$ and we find that $\pi d^r =0$ if $ r\geq 2$.
\end{proof}
Recall that for a space $X$ the suspension spectrum $\Sigma^{\infty} X$ is an $\mathbb{S}$-coalgebra. This provides a wealth of examples for which we can use the computational tools of Theorem \ref{comp1}. Before stating particular results, we discuss the motivation for studying topological coHochschild homology of suspension spectra. Recall the following theorem from Malkiewich \cite{malkiewich} and Hess--Shipley \cite{hs.coTHH}:
\begin{thm}
For $X$ a simply connected space,
\[
\mathrm{THH}(\Sigma^{\infty}_+(\Omega X)) \simeq \mathrm{coTHH}(\Sigma^{\infty}_+X)
\]
\end{thm}
Consequently, for a simply connected space $X$, to understand $\mathrm{THH}$ of $\Sigma^{\infty}_+(\Omega X)$, one can instead compute the $\mathrm{coTHH}$ of the suspension spectrum of $X$.
In some cases this topological coHochschild homology computation is more accessible. In particular, to compute the homology $H_*(\mathrm{THH}(\Sigma^{\infty}_+(\Omega X)); k)$ using the B\"okstedt spectral sequence, the necessary input is the Hochschild homology of the algebra $H_*( \Sigma^{\infty}_+(\Omega X); k)$. To compute $H_*(\mathrm{coTHH}(\Sigma^{\infty}_+X); k)$ the spectral sequence input is the coHochschild homology of the coalgebra $H_*(\Sigma^{\infty}_+X; k)$. In some cases this homology coalgebra is easier to work with than the homology algebra of the loop space, making the topological coHochschild homology calculation a more accessible path to studying the topological Hochschild homology of based loop spaces. For instance, the Lie group calculations in Example \ref{Lie} are more approachable via the coTHH spectral sequence. Topological Hochschild homology of based loop spaces is of interest due to connections to algebraic $K$-theory, free loop spaces, and string topology, which we will now recall.
Recall B\"okstedt and Waldhausen showed that $\mathrm{THH}(\Sigma^{\infty}_+(\Omega X)) \simeq \Sigma^{\infty}_+\mathcal{L} X$, where $\mathcal{L} X$ is the free loop space, $\mathcal{L} X = \mathrm{Map}(S^1, X)$. The homology of a free loop space, $H_*(\mathcal{L} X)$, is the main object of study in the field of string topology \cite{ChasSullivan, CJY}. Since
\[\mathrm{coTHH}(\Sigma^{\infty}_+X)\simeq \Sigma^{\infty}_+\mathcal{L} X\]
for simply connected $X$, topological coHochschild homology gives a new way of approaching this homology.
Topological coHochschild homology of suspension spectra also has connections to algebraic $K$-theory. Waldhausen's $A$-theory of spaces $A(X)$ is equivalent to $K(\Sigma^{\infty}_+\Omega X)$. Recall that there is a trace map from algebraic $K$-theory to topological Hochschild homology
\[
A(X) \simeq K(\Sigma^{\infty}_+\Omega X) \rightarrow \mathrm{THH}(\Sigma^{\infty}_+\Omega X).
\]
When $X$ is simply connected this gives us a trace map to topological coHochschild homology as well:
\[
A(X) \simeq K(\Sigma^{\infty}_+\Omega X) \rightarrow \mathrm{THH}(\Sigma^{\infty}_+\Omega X) \simeq \mathrm{coTHH}(\Sigma^{\infty}_+X).
\]
In the remainder of this section we will make some explicit computations of the homology of coTHH of suspension spectra.
\begin{ex} Let $X$ be a product of $n$ copies of $\mathbb{C}P^{\infty}$. As a coalgebra
\[H_*(\Sigma^{\infty}_+X; k) \cong \Gamma_{k}(x_1, x_2, \ldots x_n)\]
where $\deg(x_i) = 2$. By Theorem \ref{comp1}, there is an isomorphism of graded $k$-modules
\[
H_*(\mathrm{coTHH}(\Sigma^{\infty}_+X); k) \cong \Gamma_{k}(x_1, x_2, \ldots x_n) \otimes \Lambda_{k}(\hat{x}_1, \hat{x}_2, \ldots \hat{x}_n).
\]
\end{ex}
Note that any space having cohomology with polynomial generators in even degrees will give us a suspension spectrum whose homology satisfies the conditions of Theorem \ref{comp1}. Hence this theorem allows us to compute the topological coHochschild homology of various suspension spectra, as in the example below.
\begin{ex}\label{Lie}
Theorem \ref{comp1} directly yields the following computations of topological coHochschild homology. The isomorphisms are of graded $k$-modules.
\begin{align*}
H_*(\mathrm{coTHH}(\Sigma^{\infty}_+BU(n)); k) &\cong \Gamma_{k}(y_1, \ldots y_n) \otimes \Lambda_{k}(\hat{y}_1, \ldots \hat{y}_n) \\
&\qquad\qquad\abs{y_i} = 2i,\ \abs{\hat{y}_i} = 2i-1 \\[2ex]
H_*(\mathrm{coTHH}(\Sigma^{\infty}_+BSU(n)); k)& \cong \Gamma_{k}(y_2, \ldots y_n) \otimes \Lambda_{k}(\hat{y}_2, \ldots \hat{y}_n) \\
& \qquad\qquad \abs{y_i} = 2i,\ \abs{\hat{y}_i} = 2i-1 \\[2ex]
H_*(\mathrm{coTHH}(\Sigma^{\infty}_+BSp(n)); k) &\cong \Gamma_{k}(z_1, \ldots z_n) \otimes \Lambda_{k}(\hat{z}_1, \ldots \hat{z}_n) \\
& \qquad\qquad\abs{z_i} = 4i,\ \abs{\hat{z}_i} = 4i-1
\end{align*}
With some restrictions on the prime $p$, we also immediately get the mod $p$ homology of the topological coHochschild homology of the suspension spectra of the classifying spaces $BSO(2k), BG_2, BF_4,$ $BE_6, BE_7,$ and $BE_8$. Note that these computations also yield the homology of the free loop spaces $H_*(\mathcal{L}BG; \mathbb{F}_p)$ and the homology $H_*(\mathrm{THH}(\Sigma^{\infty}_+G); \mathbb{F}_p)$, for $G = U(n), SU(n), Sp(n), SO(2k),$ $G_2, F_4. E_6, E_7,$ and $E_8$. The topological coHochschild homology calculations also follow directly for products of copies of any of these classifying spaces $BG$.
\end{ex}
| f1465f6ef473290c576897e44e5f54e46b85ae8e | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{One-way definability in the sweeping case}%
\label{sec:characterization-sweeping}
\reviewOne[inline]{From Part 5 onwards, a major point is that an output you can bound is an output you can guess, and a periodic output is an output you don't have to guess, as you can produce it "out of order". On your Figures (and any you might care to add during potential revisions), having a specific visual representation (if possible coherent throughout the paper) to differentiate path sections with empty, bounded, or periodic output will make your point way more intelligible.
}%
\felix[inline]{so this is just about uniformizing the figure for sweeping and for two way ?}%
In the previous section we have shown the implication \PR1 $\Rightarrow$ \PR2 for a functional sweeping
transducer $\cT$. Here we close the cycle by proving the implications \PR2 $\Rightarrow$ \PR3 and \PR3 $\Rightarrow$ \PR1.
In particular, we show how to derive the existence of successful runs admitting a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition
and construct a one-way transducer $\cT'$ that simulates $\cT$ on those runs.
This will basically prove Theorem~\ref{thm:main2} in the sweeping case.
\medskip
\subsection*{Run decomposition.}
We begin by giving the definition of $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition for a run $\rho$ of $\cT$.
Intuitively, a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition of $\rho$ identifies factors of $\rho$ that can be easily simulated
in a one-way manner. The definition below describes precisely the shape of these factors.
First we need to recall the notion of almost periodicity:
a word $w$ is \emph{almost periodic with bound $p$} if $w=w_0~w_1~w_2$
for some words $w_0,w_2$ of length at most $p$ and some word $w_1$
of period at most $p$.
We need to work with subsequences of the run $\rho$ that are induced by
particular sets of locations, not necessarily consecutive.
Recall that $\rho[\ell,\ell']$ denotes the factor of $\rho$
delimited by two locations $\ell\mathrel{\unlhd}\ell'$. Similarly, given any
set $Z$ of locations, we denote by $\rho|Z$ the subsequence of $\rho$
induced by $Z$. Note that, even though $\rho|Z$ might not be a valid
run
\label{rhoZ}
of the transducer, we can still refer to the number of transitions in it
and to the size of the produced output $\out{\rho|Z}$.
Formally, a transition in $\rho|Z$ is a transition from some $\ell$ to $\ell'$,
where both $\ell,\ell'$ belong to $Z$. The output $\out{\rho|Z}$ is the
concatenation of the outputs of the transitions of $\rho| Z$
(according to the order given by $\rho$).
\reviewOne[inline]{Fig. 5 is moderately informative... Figure 17 and 18 are a more inspiring illustration of the intuition behind diagonal and blocks. Their "restricted" counterpart could be simpler to represent.}%
\gabriele[inline]{done}%
\begin{defi}\label{def:factors-sweeping}
Consider a factor $\rho[\ell,\ell ']$ of a run $\rho$ of $\cT$,
where $\ell=(x,y)$, $\ell'=(x',y')$ are two locations with $x \le x'$.
We call $\rho[\ell,\ell']$
\begin{itemize}
\medskip
\item \parbox[t]{\dimexpr\textwidth-\leftmargin}
\vspace{-2.75mm}
\begin{wrapfigure}{r}{7.5cm}
\end{wrapfigure}
%
a \emph{floor} if $y=y'$ is even, namely, if $\rho[\ell,\ell']$
lies entirely on the same level and is rightward oriented;
}\noparbreak
\bigskip
\bigskip
\item \parbox[t]{\dimexpr\textwidth-\leftmargin}
\vspace{-2.75mm}
\begin{wrapfigure}{r}{7.5cm}
\vspace{-22mm}
\input{diagonal-sweeping}
\vspace{-4mm}
\end{wrapfigure}
%
a \emph{$\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-diagonal} if
there is a sequence
$\ell = \ell_0 \mathrel{\unlhd} \ell_1 \mathrel{\unlhd} \dots \mathrel{\unlhd} \ell_{2n+1} = \ell'$,
where each $\rho[\ell_{2i+1},\ell_{2i+2}]$ is a floor,
each $\rho[\ell_{2i}, \ell_{2i+1}]$ produces an output
of length at most $2{\boldsymbol{H}}} %h_{\mathsf{max}}}\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
and the position of each $\ell_{2i}$
is to the left of the position of $\ell_{2i+1}$;
}
\bigskip
\smallskip
\item \parbox[t]{\dimexpr\textwidth-\leftmargin}
\vspace{-2.75mm}
\begin{wrapfigure}{r}{7.5cm}
\vspace{-10mm}
\input{block-sweeping}
\vspace{-5mm}
\end{wrapfigure}
%
a \emph{$\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-block} if
the output produced by $\rho[\ell,\ell']$ is almost periodic with bound $2\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
and the output produced by the subsequence $\rho|Z$,
where $Z=[\ell,\ell'] ~\setminus~ \big([x,x']\times[y,y']\big)$,
has length at most $2{\boldsymbol{H}}} %h_{\mathsf{max}}}\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
}
\vspace{8mm}
\end{itemize}
\end{defi}
Before continuing we give some intuition on the above definitions.
The simplest concept is that of floor, which is a rightward oriented factor of a run.
Diagonals are sequences of consecutive floors interleaved by factors that
produce small (bounded) outputs. We see an example of a diagonal
in Figure \ref{fig:diagonal-sweeping}, where we marked the important
locations and highlighted with thick arrows the two floors that form
the diagonal. The factors of the diagonal that are not floors are
represented instead by dotted arrows.
The third notion is that of a block.
An important constraint in the definition of a block
is that the produce output must be almost periodic, with small enough bound.
In Figure \ref{fig:block-sweeping}, the periodic output is represented by the
thick arrows, either solid or dotted, that go from location $\ell$ to
location $\ell'$.
In addition, the block must satisfy a constraint on the length of
the output produced by the subsequence $\rho|Z$, where
$Z=[\ell,\ell'] ~\setminus~ \big([x,x']\times[y,y']\big)$.
The latter set $Z$ consists of location that are either to the left
of the position of $\ell$ or to the right of the position of $\ell'$.
For example, in Figure \ref{fig:block-sweeping} the set $Z$ coincides
with the area outside the hatched rectangle. Accordingly, the portion
of the subsequence $\rho|Z$ is represented by the dotted bold arrows.
Diagonals and blocks are used as key objects to derive a notion of
decomposition for a run of a sweeping transducer. We formalize this
notion below.
\begin{defi}\label{def:decomposition-sweeping}
A \emph{$\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition} of a run $\rho$ of $\cT$ is a factorization
$\prod_i\,\rho[\ell_i,\ell_{i+1}]$ of $\rho$ into $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-diagonals and $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-blocks.
\end{defi}
\reviewOne[inline]{Fig.6 does not look like a fully factorized run: what happens for example, between the end of $D_1$ and $l_1$? Can a concrete example be devised? This works rather well in Part 8.}%
\gabriele{Improved the figure and the explanation}%
Figure~\ref{fig:decomposition-sweeping} gives an example of such a
decomposition.
\anca{dans Fig 7, les aretes de $B_2$ vont dans la
mauvaise direction}%
\gabriele{I have corrected this. Now the figure is a bit more regular (perhaps too much),
but at least conveys the correct information!}%
Each factor is either a diagonal $D_i$ or a block $B_i$.
Inside each factor we highlight by thick arrows the portions
of the run that can be simulated by a one-way transducer,
either because they are produced from left to right (thus forming diagonals)
or because they are periodic (thus forming blocks).
We also recall from Definition \ref{def:factors-sweeping}
that most of the output is produced inside the
hatched rectangles, since the output produced by a diagonal
or a block outside the corresponding blue or red hatched rectangle
has length at most $2{\hmax^2}} %{{h^2_{\mathsf{max}}}\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Finally, we observe that the locations delimiting the factors of the
decomposition are arranged following the natural order of positions and levels.
All together, this means that the output produced by a run that enjoys
a decomposition can be simulated in a one-way manner.
\input{decomposition-sweeping}
\medskip
\reviewOne[inline]{Part 5: In "From periodicity of inversions to existence of decompositions", the general proof structure in general and transition between lemmas in particular could be made more visually and conceptually explicit, as the global idea of the proof is difficult to grasp at first. A mini "road map" as made in Part 3 for the whole of Part 4 and 5 might be another way to help the reader through these rather intricate proofs.}%
\subsection*{From periodicity of inversions to existence of decompositions.}
Now that we set up the definition of $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition, we turn
towards proving the implication \PR2 $\Rightarrow$ \PR3 of Theorem \ref{thm:main2}.
In fact, we will prove a slightly stronger result than \PR2 $\Rightarrow$ \PR3,
which is stated further below.
Formally, when we say that a run \emph{$\rho$ satisfies \PR2} we mean
that for every inversion $(L_1,\ell_1,L_2,\ell_2)$ of $\rho$, the word
$\out{\tr{\ell_1}} ~ \out{\rho[\ell_1,\ell_2]} ~ \out{\tr{\ell_2}}$
has period $\gcd(|\out{\tr{\ell_1}}|, |\out{\tr{\ell_2}}|) \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
We aim at proving that every run that satisfies \PR2 enjoys a decomposition,
independently of whether other runs do or do not satisfy \PR2:
\begin{prop}\label{prop:decomposition-sweeping}
If $\rho$ is a run of $\cT$ that satisfies \PR2, then $\rho$ admits a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition.
\end{prop}
\felix{a few lines for the roadmap asked by the reviewer, but I don't think they're useful at all}%
\gabriele{I think it is good}%
Let us fix a run $\rho$ of $\cT$ and assume that it satisfies \PR2.
To show that $\rho$ admits a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition, we will identify
the blocks of the decomposition as equivalence classes of a suitable
relation based on inversions (cf.~Definition \ref{def:crossrel}).
Then, we will use combinatorial arguments (notably,
Lemmas \ref{lem:output-minimal-sweeping} and \ref{lem:overlapping})
to prove that the constructed blocks satisfy the desired properties.
Finally, we will show how the resulting equivalence classes form all
the necessary blocks of the decomposition, in the sense that the factors
in between those classes are diagonals.
We begin by introducing the equivalence relation by means of which we can
then identify the blocks of a decomposition of $\rho$.
\begin{defi}\label{def:crossrel}
Let $\ensuremath{\mathrel{\text{\sf S}}}$ be the relation that pairs every two locations $\ell,\ell'$ of $\rho$
whenever there is an inversion $(L_1,\ell_1,L_2,\ell_2)$ of $\rho$ such that
$\ell_1 \mathrel{\unlhd} \ell,\ell' \mathrel{\unlhd} \ell_2$, namely, whenever $\ell$ and $\ell'$
occur within the same inversion.
Let $\simeq$ be the reflexive and transitive closure of $\ensuremath{\mathrel{\text{\sf S}}}$.
\end{defi}
It is easy to see that every equivalence class of $\simeq$ is a convex
subset with respect to the run order on locations of $\rho$.
Moreover, every \emph{non-singleton} equivalence class of $\simeq$ is a
union of a series of inversions that are two-by-two overlapping.
One can refer to Figure~\ref{fig:overlapping}
for an intuitive account of what we mean by two-by-two overlapping: the thick
arrows represent factors of the run that lie entire inside an $\simeq$-equivalence class,
each inversion is identified by a pair of consecutive anchor points with the same
color. According to the run order, between every pair of anchor points with the
same color, there is at least one anchor point of different color: this means that
the inversions corresponding to the two colors are overlapping.
\reviewOne[inline]{Fig.7 might benefit from some $L_1$, $L_2$ appearing, to illustrate the point made in Claim 5.6
\felix[inline]{I don't agree}%
\olivier[inline]{I also disagree. Argued in review1-answer.txt.}%
}%
Formally, we say that an inversion $(L_1,\ell_1,L_2,\ell_2)$ \emph{covers}
a location $\ell$ when $\ell_1 \mathrel{\unlhd} \ell \mathrel{\unlhd} \ell_2$. We say that
two inversions $(L_1,\ell_1,L_2,\ell_2)$ and $(L_3,\ell_3,L_4,\ell_4)$
are \emph{overlapping} if $(L_1,\ell_1,L_2,\ell_2)$ covers $\ell_3$
and $(L_3,\ell_3,L_4,\ell_4)$ covers $\ell_2$ (or the other way around).
\input{overlapping}
The next lemma uses the fact that $\rho$ satisfies \PR2 to deduce that
the output produced inside every $\simeq$-equivalence class has
period at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$. Note that the proof below does not exploit
the fact that the transducer is sweeping.
\begin{lem}\label{lem:overlapping}
If $\rho$ satisfies \PR2 and $\ell\mathrel{\unlhd}\ell'$ are two locations of $\rho$
such that $\ell \simeq \ell'$, then the output $\out{\rho[\ell,\ell']}$
produced between these locations has period at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
\end{lem}
\begin{proof}
The claim for $\ell=\ell'$ holds trivially, so we assume that $\ell\mathrel{\lhd}\ell'$.
Since the $\simeq$-equivalence class that contains $\ell,\ell'$ is non-singleton,
we know that there is a series of inversions
\[
(L_0,\ell_0,L_1,\ell_1) \quad
(L_2,\ell_2,L_3,\ell_3) \quad
\dots\dots \quad
(L_{2k},\ell_{2k},L_{2k+1},\ell_{2k+1})
\]
that are two-by-two overlapping and such that
$\ell_0\mathrel{\unlhd}\ell\mathrel{\lhd}\ell'\mathrel{\unlhd}\ell_{2k+1}$.
Without loss of generality, we can assume that every
inversion $(L_{2i},\ell_{2i},L_{2i+1},\ell_{2i+1})$ is \emph{maximal}
in the following sense: there is no other inversion
$(\tilde L,\tilde\ell,\tilde L',\tilde\ell') \neq (L_{2i},\ell_{2i},L_{2i+1},\ell_{2i+1})$
such that $\tilde\ell \mathrel{\unlhd} \ell_{2i} \mathrel{\unlhd} \ell_{2i+1} \mathrel{\unlhd} \tilde\ell'$.
For the sake of brevity, let $v_i = \out{\tr{\ell_i}}$ and $p_i = | v_i |$.
Since $\rho$ satisfies \PR2 (recall Proposition~\ref{prop:periodicity-sweeping}), we know that, for all $i=0,\dots,n$, the word
\[
v_{2i} ~ \out{\rho[\ell_{2i},\ell_{2i+1}]} ~ v_{2i+1}
\]
has period that divides both $p_{2i}$ and $p_{2i+1}$ and is at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
In order to show that the period of $\out{\rho[\ell,\ell']}$ is also bounded
by $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$, it suffices to prove the following claim by induction on $i$:
\reviewOne[inline]{Claim 5.6, end of page 18: making explicit what words you wish to use Fine-Wilf on might make for a more readable proof.
\olivier[inline]{answered in review1-answer.txt}%
}%
\begin{clm}
For all $i=0,\dots,k$, the word
$\outb{\rho[\ell_0,\ell_{2i+1}]} \: v_{2i+1}$
has period at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ that divides $p_{2i+1}$.
\end{clm}
The base case $i=0$ follows immediately from our hypothesis,
since $(L_0,\ell_0,L_1,\ell_1)$ is an inversion.
For the inductive step, we assume that the claim holds for $i<k$,
and we prove it for $i+1$. First of all, we factorize our word as follows:
\[
\outb{\rho[\ell_0,\ell_{2i+3}]} ~ v_{2i+3}
~=~
\rightward{ \underbracket[0.5pt]{ \phantom{
\outb{\rho[\ell_0,\ell_{2i+2}]} ~
\outb{\rho[\ell_{2i+2},\ell_{2i+1}]} ~ } }%
_{\text{period } p_{2i+1}} }
\outb{\rho[\ell_0,\ell_{2i+2}]}
\overbracket[0.5pt]{ ~ \outb{\rho[\ell_{2i+2},\ell_{2i+1}]} ~
\outb{\rho[\ell_{2i+1},\ell_{2i+3}]} ~
v_{2i+3} }%
^{\text{periods } p_{2i+2} \text{ and } p_{2i+3}}
.
\]
By the inductive hypothesis, the output produced between $\ell_0$
and $\ell_{2i+1}$, extended to the right with $v_{2i+1}$,
has period that divides $p_{2i+1}$.
Moreover, because $\rho$ satisfies \PR2 and
$(L_{2i+2},\ell_{2i+2},L_{2i+3},\ell_{2i+3})$ is an inversion,
the output produced between the locations
$\ell_{2i+2}$ and $\ell_{2i+3}$,
extended to the left with $v_{2i+2}$ and to the
right with $v_{2i+3}$, has period that divides both $p_{2i+2}$ and $p_{2i+3}$.
Note that this is not yet sufficient for applying Fine-Wilf's theorem,
since the common factor $\outb{\rho[\ell_{2i+2},\ell_{2i+1}]}$
might be too short (possibly just equal to $v_{2i+2}$).
The key argument here is to prove that the interval $[\ell_{2i+2},\ell_{2i+1}]$
is covered by an inversion
which is different from those that we considered above, namely,
i.e.~$(L_{2i+2},\ell_{2i+2},L_{2i+1},\ell_{2i+1})$. For example,
$[\ell_2,\ell_1]$ in
Figure~\ref{fig:overlapping} is covered by the inversion
$(L_2,\ell_2,L_1,\ell_1)$.
For this, we have to prove that the anchor points $\ell_{2i+2}$ and $\ell_{2i+1}$
are correctly ordered w.r.t.~$\mathrel{\unlhd}$ and the ordering of positions
(recall Definition~\ref{def:inversion-sweeping}).
First, we observe that $\ell_{2i+2} \mathrel{\unlhd} \ell_{2i+1}$, since
$(L_{2i},\ell_{2i},L_{2i+1},\ell_{2i+1})$ and $(L_{2i+2},\ell_{2i+2},L_{2i+3},\ell_{2i+3})$
are overlapping inversions.
Next, we prove that the position of $\ell_{2i+1}$ is strictly to the left of
the position of $\ell_{2i+2}$. By way of contradiction, suppose that this is
not the case, namely, $\ell_{2i+1}=(x_{2i+1},y_{2i+1})$, $\ell_{2i+2}=(x_{2i+2},y_{2i+2})$,
and $x_{2i+1} \ge x_{2i+2}$. Because $(L_{2i},\ell_{2i},L_{2i+1},\ell_{2i+1})$ and
$(L_{2i+2},\ell_{2i+2},L_{2i+3},\ell_{2i+3})$ are inversions,
we know that $\ell_{2i+3}$ is strictly to the left of $\ell_{2i+2}$
and $\ell_{2i+1}$ is strictly to the left of $\ell_{2i}$.
This implies that $\ell_{2i+3}$ is strictly to the left of $\ell_{2i}$,
and hence $(L_{2i},\ell_{2i},L_{2i+3},\ell_{2i+3})$ is also an inversion.
Moreover, recall that $\ell_{2i} \mathrel{\unlhd} \ell_{2i+2} \mathrel{\unlhd} \ell_{2i+1} \mathrel{\unlhd} \ell_{2i+3}$.
This contradicts the maximality of $(L_{2i},\ell_{2i},L_{2i+1},\ell_{2i+1})$,
which we assumed at the beginning of the proof.
Therefore, we must conclude that $\ell_{2i+1}$ is strictly to the left of $\ell_{2i+2}$.
Now that we know that $\ell_{2i+2} \mathrel{\unlhd} \ell_{2i+1}$ and that $\ell_{2i+1}$ is to the left of $\ell_{2i+2}$,
we derive the existence of the inversion $(L_{2i+2},\ell_{2i+2},L_{2i+1},\ell_{2i+1})$.
Again, because $\rho$ satisfies \PR2, we know that the word
$v_{2i+2} ~ \out{\rho[\ell_{2i+2},\ell_{2i+1}]} ~ v_{2i+1}$ has period at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$
that divides $p_{2i+2}$ and $p_{2i+1}$.
Summing up, we have:
\begin{enumerate}
\item $w_1 ~=~ \outb{\rho[\ell_0,\ell_{2i+1}]} ~ v_{2i+1}$ has period $p_{2i+1}$,
\item $w_2 ~=~ v_{2i+2} ~ \outb{\rho[\ell_{2i+2},\ell_{2i+1}]} ~ v_{2i+1}$
has period $p = \gcd(p_{2i+2},p_{2i+1})$,
\item $w_3 ~=~ v_{2i+2} ~ \outb{\rho[\ell_{2i+2},\ell_{2i+3}]} ~ v_{2i+3}$
has period $p' = \gcd(p_{2i+2},p_{2i+3})$.
\end{enumerate}
We are now ready to exploit our stronger variant of Fine-Wilf's theorem,
that is, Theorem~\ref{thm:fine-wilf}.
We begin with (1) and (2) above.
Let $w = \outb{\rho[\ell_{2i+2},\ell_{2i+1}]}~ v_{2i+1}$ be the common suffix of
$w_1$ and $w_2$. First note that since $p$ divides $p_{2i+2}$, the
word $w$ is also a prefix of $w_2$, thus we can write
$w_2=w\,w'_2$. Second, note that the length of $w$ is at least
$|v_{2i+1}| = p_{2i+1} = p_{2i+1} + p - \gcd(p_{2i+1},p)$. We can
apply now Theorem~\ref{thm:fine-wilf} to $w_1=w'_1\,w$ and
$w_2=w\,w'_2$ and obtain:
\begin{enumerate}
\setcounter{enumi}{3}
\item $w_4 ~=~ w'_1 \: w \: w'_2
~=~ \outb{\rho[\ell_0,\ell_{2i+2}]} ~ v_{2i+2} ~ \outb{\rho[\ell_{2i+2},\ell_{2i+1}]} ~ v_{2i+1}$
has period $p$.
\end{enumerate}
We apply next Theorem~\ref{thm:fine-wilf} to (2) and (3), namely, to
the words $w_2$ and $w_3$ with $v_{2i+2}$ as common factor. It is not
difficult to check that $|v_{2i+2}|=p_{2i+2} \ge p+p'-p''$ with
$p''=\gcd(p,p')$, using the definitions of $p$ and $p'$: we can
write $p_{2i+2}=p''rq=p''r'q'$ with $p=p''r$ and $p'=p''r'$. It suffices to check
that $p''rq+p''r'q' \ge 2(p''r + p''r'-p'')$, hence that $rq+r'q' \ge 2r+2r'-2$. This
is clear if $\min(q,q')>1$. Otherwise the inequality $p_{2i+2} \ge
p+p'-p''$ follows easily because $p=p''$ or $p'=p''$ holds. Hence we obtain
that $w_2$ and $w_3$ have both period $p''$.
Applying once more Theorem~\ref{thm:fine-wilf} to $w_3$ and $w_4$ with
$v_{2i+2}$ as common factor, yields period $p''$ for the word
\[
w_5 ~=~ \outb{\rho[\ell_0,\ell_{2i+2}]} ~ v_{2i+2} ~
\outb{\rho[\ell_{2i+2},\ell_{2i+3}]} ~ v_{2i+3}
\]
Finally, the periodicity is not affected when we remove
factors of length multiple than the period. In particular,
by removing the factor $v_{2i+2}$ from $w_5$,
we obtain the desired word
$\outb{\rho[\ell_0,\ell_{2i+3}]} ~ v_{2i+3}$, whose period
$p''$ divides $p_{2i+3}$. This proves the claim for the inductive step,
and completes the proof of the proposition.
\end{proof}
\medskip
The $\simeq$-classes considered so far cannot be directly used to
define the blocks in the desired decomposition of $\rho$, since the $x$-coordinates of their
endpoints might not be in the appropriate order. The next definition
takes care of this, by enlarging the $\simeq$-classes according to
$x$-coordinates of the anchor points in the equivalence class.
\bigskip\noindent
\begin{minipage}[l]{\textwidth-5.8cm}
\begin{defi}\label{def:bounding-box-sweeping}
Consider a non-singleton $\simeq$-equivalence class $K=[\ell,\ell']$.
Let $\an{K}$
be the restriction of $K$ to the anchor points occurring in some inversion,
and $X_{\an{K}} = \{x \::\: \exists y\: (x,y)\in \an{K}\}$
be the projection of $\an{K}$ on positions.
We define $\block{K}=[\tilde\ell,\tilde\ell']$, where
\begin{itemize}
\item $\tilde\ell$ is the latest location $(\tilde x,\tilde y) \mathrel{\unlhd} \ell$
such that $\tilde x = \min(X_{\an{K}})$,
\item $\tilde\ell'$ is the earliest location $(\tilde x,\tilde y) \mathrel{\unrhd} \ell'$
such that $\tilde x = \max(X_{\an{K}})$
\end{itemize}
(note that the locations $\tilde\ell,\tilde\ell'$ exist since $\ell,\ell'$
are anchor points in some inversion).
\end{defi}
\end{minipage}
\begin{minipage}[r]{5.7cm}
\input{block-construction-sweeping}
\end{minipage}
\smallskip
\begin{lem}\label{lem:bounding-box-sweeping}
If $K$ is a non-singleton $\simeq$-equivalence class,
then $\rho|\block{K}$ is a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-block.
\end{lem}
\begin{proof}
Consider a non-singleton $\simeq$-class $K=[\ell,\ell']$ and let $\an{K}$, $X_{\an{K}}$, and
$\block{K}=[\tilde\ell,\tilde\ell']$
be as in Definition \ref{def:bounding-box-sweeping}.
The reader can refer to Figure \ref{fig:block-construction-sweeping}
to quickly recall the notation.
We need to verify that $\rho[\tilde\ell,\tilde\ell']$ is a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-block
(cf.~Definition \ref{def:factors-sweeping}), namely, that:
\begin{itemize}
\item $\tilde\ell=(\tilde x,\tilde y)$, $\tilde\ell'=(\tilde x',\tilde y')$,
with $\tilde x \le \tilde x'$,
\item the output produced by $\rho[\tilde\ell,\tilde\ell']$
is almost periodic with bound $2\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
\item the output produced by the subsequence $\rho|Z$,
where $Z=[\tilde\ell,\tilde\ell'] ~\setminus~
\big([\tilde x,\tilde x']\times[\tilde y,\tilde y']\big)$,
has length at most $2{\boldsymbol{H}}} %h_{\mathsf{max}}}\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
\end{itemize}
The first condition $\tilde x \le \tilde x'$ follows immediately from
the definition of $\tilde x$ and $\tilde x'$ as $\min(X_{\an{K}})$
and $\max(X_{\an{K}})$, respectively.
Next, we prove that the output produced by the factor
$\rho[\tilde\ell,\tilde\ell']$ is almost periodic with bound $2\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
By Definition \ref{def:bounding-box-sweeping}, we have
$\tilde\ell \mathrel{\unlhd} \ell \mathrel{\lhd} \ell' \mathrel{\unlhd} \tilde\ell'$,
and by Lemma \ref{lem:overlapping} we know that $\out{\rho[\ell,\ell']}$
is periodic with period at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ ($\le 2\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$). So it suffices to show that
the lengths of the words $\out{\rho[\tilde\ell,\ell]}$ and
$\out{\rho[\ell',\tilde\ell']}$ are at most $2\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
We shall focus on the former word, as the arguments for
the latter are similar.
First, we note that the factor $\rho[\tilde\ell,\ell]$
lies entirely to the right of position $\tilde x$, and
in particular, it starts at an even level $\tilde y$. This follows
from the definition of $\tilde\ell$, and whether $\ell$ itself is at
an odd/even level. In particular, the location $\ell$ is either at the same level as $\tilde\ell$,
or just one level above.
Now, suppose, by way of contradiction, that $|\out{\rho[\tilde\ell,\ell]}| > 2\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
We head towards a contradiction by finding a location $\ell'' \mathrel{\lhd} \ell$
that is $\simeq$-equivalent to the first location $\ell$ of the $\simeq$-equivalence
class $K$.
Since the location $\ell$ is either at the same level as $\tilde\ell$,
or just above it,
the factor $\rho[\tilde\ell,\ell]$ is of the form $\alpha\,\beta$,
where $\alpha$ is rightward factor lying on the same level as $\tilde\ell$
and $\beta$ is either empty or a leftward factor on the next level.
Moreover, since $|\out{\rho[\tilde\ell,\ell]}| > 2\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$, we know that
either $|\out{\alpha}| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ or $|\out{\beta}| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Thus, Lemma \ref{lem:output-minimal-sweeping}
says that one of the two factors $\alpha,\beta$ is not output-minimal.
In particular, there is a loop $L_1$, strictly to the right of $\tilde x$,
that intercepts a subfactor $\gamma$ of $\rho[\tilde\ell,\ell]$,
with $\out{\gamma}$ non-empty and output-minimal.
Let $\ell''$ be the first location of the factor $\gamma$.
Clearly, $\ell''$ is an anchor point of $L$ and $\out{\tr{\ell''}}\neq\emptystr$.
Further recall that $\tilde x=\min(X_{\an{K}})$ is the leftmost position of
locations in the class $K=[\ell,\ell']$ that are also anchor points of inversions.
In particular, there is a loop $L_2$ with some anchor point
$\ell''_2=(\tilde x,y''_2)\in \an{K}$, and such that $\tr{\ell''}$ is
non-empty and output-minimal.
Since $\ell'' \mathrel{\lhd} \ell \mathrel{\unlhd} \ell''_2$
and the position of $\ell''$ is to the right of the position of $\ell''_2$,
we know that $(L_1,\ell'',L_2,\ell''_2)$ is also an inversion,
and hence $\ell'' \simeq \ell''_2 \simeq \ell$.
But since $\ell'' \mathrel{\lhd} \ell$, we get a contradiction with the
assumption that $\ell$ is the first location of a $\simeq$-class.
In this way we have shown that $|\out{\rho[\ell_1,\ell]}| \le 2\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
It remains to show that the output produced
by the subsequence $\rho|Z$, where
$Z=[\tilde\ell,\tilde\ell'] ~\setminus~ \big([\tilde x,\tilde x']\times[\tilde y,\tilde y']\big)$,
has length at most $2{\boldsymbol{H}}} %h_{\mathsf{max}}}\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
For this it suffices to prove that
$|\out{\alpha}| \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ for every factor $\alpha$ of
$\rho[\tilde\ell, \tilde\ell']$ that lies
at a single level and either to the left of $\tilde x$ or to the right of $\tilde x'$.
By symmetry, we consider only one of the two types of factors.
Suppose, by way of contradiction, that there is a factor $\alpha$
at level $y''$, to the left of $\tilde x$,
and such that $|\out{\alpha}| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
By Lemma \ref{lem:output-minimal-sweeping} we know that
$\alpha$ is not output-minimal, so there is some loop
$L_2$ strictly to the left of $\tilde x$ that intercepts an
output-minimal subfactor
$\beta$ of $\alpha$ with non-empty output.
Let $\ell''$ be the first location of $\beta$. We know that
$\tilde\ell \mathrel{\lhd} \ell'' \mathrel{\unlhd} \tilde\ell'$. Since the level
$\tilde y$ is even, this means that the level of $\ell''$ is strictly
greater than $\tilde y$. Since we also know that $\ell$
is an anchor point of some inversion, we can take a suitable loop $L_1$
with anchor point $\ell$ and obtain that $(L_1,\ell,L_2,\ell'')$ is an
inversion, so $\ell'' \simeq \ell$. But
this contradicts the fact that $\tilde x$ is the leftmost position of
$\an{K}$.
We thus conclude that $|\out{\alpha}| \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$, and this
completes the proof that $\rho|\block{K}$ is a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-block.
\begin{comment}
We now prove that the output produce by the subsequence $\rho\mid Z$, where
$Z=[\ell_1,\ell_2]\setminus[x_1,x_2]\times[y_1,y_2]$, has length at most
$2 (y_2 - y_1)\cdot \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax} \leq {\boldsymbol{H}}} %h_{\mathsf{max}}} \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax} $.
Suppose, by way of contradiction, that this is not the case, and
that $\big|\out{\rho\mid Z}\big| > 2(y_2 - y_1) \cdot \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Since the set $Z$ spans across at most $y_2 - y_1$ levels and does not cover
the locations inside the ``bounding box'' $[x_1,x_2]\times[y_1,y_2]$,
we could find a factor $\alpha$ that produces an output longer than $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
that lies on a single level $y'$, and either to the left of $x_1$ or to the
right of $x_2$.
Suppose that this output is on the left (the right case is symmetric), and that $\alpha$ is a factor
on some level $y'\in\{y_1+1,\ldots,y_2\}$ that is intercepted by the interval
$I=[1,x_1]$ and that produces an output of length strictly greater than $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
By the contrapositive of Lemma \ref{lemma:simple-loop}, the factor $\alpha$
is captured by some loop $L'=[x'_1,x'_2]$ that is {\sl strongly contained} in
$I=[1,x_1]$, namely, such that $1\le x'_1<x'_2<x_1$, and such that the output of $L'$ at level $y'$ is non-empty.
Let $\ell'=(x',y')$ be the anchor point of $L'$, where $x'=x'_1$ or $x'=x'_2$
depending on whether $y'$ is odd or even. Recall that, by the definition of $[K]$,
there is also a location $\ell=(x,y)$ that strictly precedes $\ell'$ along the
run and that belongs to $H$. This implies that $\ell$ is an, anchor point of some loop $L$.
Moreover, recall that $x'_1\le x'\le x'_2 < x_1$ and $x_1\le x\le x_2$, namely,
$\ell'$ is {\sl strictly to the left} of $\ell$ according to the ordering of
the $x$-coordinates. This shows that $(L,\ell,L',\ell')$ is also an inversion, and
hence $\ell'$ also belongs to $H$.
However, this contradicts the definition of $x_1$ as the minimum of the
$x$-coordinates of the locations in $H$.
A similar contradiction can be obtained in the case where the factor that produces an
output of length strictly greater than $ \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ lies at the right of $x_2$.
By the above arguments, we know that $\big|\out{\rho\mid Z}| \le 2 (y_2 - y_1)\cdot \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax} \leq {\boldsymbol{H}}} %h_{\mathsf{max}}} \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
where $Z=[\ell_1,\ell_2]\setminus[x_1,x_2]\times[y_1,y_2]$, namely, that the second
condition of the definition of block is satisfied.
It remains to verify the first condition, that the output produced by
$\rho[\ell_1,\ell_2]$ is almost periodic with bound $2 \cdot \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Let $\ell'_1$ and $\ell'_2$ be the first and the last locations of the
$\simeq$-equivalence class $K$ (note that $\ell_1\le \ell'_1\le \ell'_2\le\ell_2$).
Recall that Lemma \ref{lem:overlapping} already shows that the output produced between
$\ell'_1$ and $\ell'_2$ is periodic with period at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Thus, we just need to show that the prefix $\out{\rho[\ell_1,\ell'_1]}$ and
the suffix $\out{\rho[\ell'_2,\ell_2]}$ have length at most $2 \cdot \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Suppose that the length of $\out{\rho[\ell_1,\ell'_1]}$ exceeds $ 2 \cdot \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
By the definition of $[K]$, the two locations $\ell_1$ and $\ell'_1$ would be either
on the same level, i.e.~$y_1$, or on adjacent levels, i.e.~$y_1$ and $y_1+1$.
In the following, we show that none of these cases can happen, thus reaching
a contradiction from the assumption $\big|\out{\rho[\ell_1,\ell'_1]}\big| > 2 \cdot \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
If $\ell_1$ were on the same level as $\ell'_1$, then clearly the factor
$\out{\rho[\ell_1,\ell'_1]}$ would lie on a single level and to the right
of $x_1$, and would produce an output longer than $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Then, by using the contrapositive of Lemma \ref{lemma:simple-loop}, we could
find a loop $L'=[x'_1,x'_2]$ strongly contained in the interval $I=[x_1,n]$,
where $n$ is the rightmost position of the input, that captures the factor
$\out{\rho[\ell_1,\ell'_1]}$ with non-empty output.
We recall that $\ell'_1$ is the first location of the $\simeq$-equivalence
class $K$, and hence there is an inversion $(L'_1,\ell'_1,L'',\ell'')$, for some
location $\ell''$ that follows $\ell'_1$ along the run. We then define
$\ell=(x'_1,y_1)$ and we observe that this location is an anchor point
of $L'$ and it strictly precedes $\ell''$.
We thus get that $(L',\ell,L'',\ell'')$ is also an inversion. This, however,
is a contradiction because the inversion $(\ell,\ell'')$ intersects
the $\simeq$-equivalence class $K$, without being contained in it.
Let us now consider the second case, where $\ell_1$ is on level $y_1$ and
$\ell'_1$ is on level $y_1+1$.
Since $\rho[\ell_1,\ell'_1]$ spans across two levels and produces an
output longer than $ 2 \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$, there is a factor
$\alpha$ of $\out{\rho[\ell_1,\ell'_1]}$ that lies entirely on a single
level -- either $y_1$ or $y_1+1$ -- and to the right of $x_1$, and
produces an output longer than $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Then, by reasoning as in the previous case, we can get a contradiction
by finding an inversion $(L,\ell,L'',\ell'')$ that intersects $K$ without
being contained in it.
We have just shown that $\big|\out{\rho[\ell_1,\ell'_1]}\big| \le 2 \cdot \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
By using symmetric arguments, we can also prove that
$\big|\out{\rho[\ell'_2,\ell_2]}\big| \le 2 \cdot \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Finally, we recall that, by Lemma \ref{lem:overlapping}, the period of
$\out{\rho[\ell'_1,\ell'_2]}$ is smaller than $ \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
All together, this shows that the output produced between the
locations $\ell_1$ and $\ell_2$ is almost periodic with bound
$2 \cdot \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ and hence $[K]=(\ell_1,\ell_2)$ is a block.
\end{comment}
\end{proof}
The next lemma shows that blocks do not overlap along the input axis:
\begin{lem}\label{lem:consecutive-blocks-sweeping}
Suppose that $K_1$ and $K_2$ are two different non-singleton $\simeq$-classes
such that $\ell \mathrel{\lhd} \ell'$ for all $\ell \in K_1$ and $\ell' \in K_2$.
Let $\block{K_1}=[\ell_1,\ell_2]$ and $\block{K_2}=[\ell_3,\ell_4]$,
with $\ell_2=(x_2,y_2)$ and $\ell_3=(x_3,y_3)$.
Then $x_2 < x_3$.
\end{lem}
\begin{proof}
Suppose by contradiction that $K_1$ and $K_2$ are as in the
statement, but $x_2 \ge x_3$.
By Definition \ref{def:bounding-box-sweeping},
$x_2=\max(X_{\an{K_1}})$ and $x_3=\min(X_{\an{K_2}})$.
This implies the existence of some inversions
$(L,\ell,L',\ell')$ and $(L'',\ell'',L''',\ell''')$
such that $\ell=(x_2,y)$ and $\ell'''=(x_3,y''')$.
Moreover, since $\ell \mathrel{\unlhd} \ell'''$ and $x_2 \ge x_3$,
we know that $(L,\ell,L''',\ell''')$ is also an inversion,
thus implying that $K_1=K_2$.
\end{proof}
For the sake of brevity, we call \emph{$\simeq$-block} any
factor of the form $\rho|\block{K}$ that is obtained by applying
Definition~\ref{def:bounding-box-sweeping} to a non-singleton $\simeq$-class $K$.
The results obtained so far imply that every location covered by an
inversion is also covered by an $\simeq$-block (Lemma \ref{lem:bounding-box-sweeping}),
and that the order of occurrence of $\simeq$-blocks is the same as the order of positions
(Lemma \ref{lem:consecutive-blocks-sweeping}).
So the $\simeq$-blocks can be used as factors for the $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition
of $\rho$ we are looking for. Below, we show that the remaining
factors of $\rho$, which do not overlap the $\simeq$-blocks, are $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-diagonals.
This will complete the construction of a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition of $\rho$.
Formally, we say that a factor $\rho[\ell,\ell']$
\emph{overlaps} another factor $\rho[\ell'',\ell''']$ if
$[\ell,\ell'] \:\cap\: [\ell'',\ell'''] \neq \emptyset$,
$\ell' \neq \ell''$, and $\ell \neq \ell'''$.
\begin{lem}\label{lem:diagonal-sweeping}
Let $\rho[\ell,\ell']$ be a
factor of $\rho$,
with $\ell=(x,y)$, $\ell'=(x',y')$, and $x\le x'$,
that does not overlap any $\simeq$-block.
Then $\rho[\ell,\ell']$ is a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-diagonal.
\end{lem}
\begin{proof}
Consider a factor $\rho[\ell,\ell']$, with $\ell=(x,y)$, $\ell'=(x',y')$,
and $x\le x'$, that does not overlap any $\simeq$-block.
We will focus on locations $\ell''$ with $\ell \mathrel{\unlhd} \ell''
\mathrel{\unlhd} \ell'$
that are anchor points of some loop with $\out{\tr{\ell''}}\neq\emptystr$.
We denote by $A$ the set of all such locations.
First, we show that the locations in $A$ are monotonic
\reviewOne[inline]{Monotonic$\to$Monotonous?
\olivier[inline]{we keep monotonic}%
}%
w.r.t.~the position order. Formally,
we prove that for all $\ell_1,\ell_2\in A$, if $\ell_1=(x_1,y_1) \mathrel{\unlhd} \ell_2=(x_2,y_2)$, then
$x_1 \le x_2$. Suppose that this were not the case, namely, that $A$ contained two anchor points
$\ell_1=(x_1,y_1)$ and $\ell_2=(x_2,y_2)$ with $\ell_1 \mathrel{\lhd} \ell_2$ and $x_1 > x_2$.
Let $L_1,L_2$ be the loops of $\ell_1,\ell_2$, respectively, and recall that
$\out{\tr{\ell_1}},\out{\tr{\ell_2}}\neq\emptystr$. This means that $(L_1,\ell_1,L_2,\ell_2)$
is an inversion, and hence $\ell_1 \simeq \ell_2$. But this contradicts the hypothesis that
$\rho[\ell,\ell']$ does not overlap any $\simeq$-block.
Next, we identify the floors of our diagonal. Let $y_0,y_1,\dots,y_{n-1}$ be all the {\sl even}
levels that have locations in $A$. For each $i=0,\dots,n-1$, let $\ell_{2i+1}$ (resp.~$\ell_{2i+2}$)
be the first (resp.~last) anchor point of $A$ at level $y_i$. Further let $\ell_0=\ell$ and $\ell_{2n+1}=\ell'$.
Clearly, each factor $\rho[\ell_{2i+1},\ell_{2i+2}]$ is a floor. Moreover, thanks to the previous
arguments, each location $\ell_{2i}$ is to the left of the location $\ell_{2i+1}$.
It remains to prove that each factor $\rho[\ell_{2i},\ell_{2i+1}]$ produces an
output of length at most $2{\boldsymbol{H}}} %h_{\mathsf{max}}}\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$. By construction, $A$ contains no anchor
point at an even level and strictly between $\ell_{2i}$ and $\ell_{2i+1}$.
By Lemma \ref{lem:output-minimal-sweeping} this means that the outputs produced
by subfactors of $\rho[\ell_{2i},\ell_{2i+1}]$ that lie entirely at an {\sl even}
level have length at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Let us now consider the subfactors $\alpha$ of $\rho[\ell_{2i},\ell_{2i+1}]$
that lie entirely at an {\sl odd} level, and let us prove that they produce outputs
of length at most $2\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$. Suppose that this is not the case, namely, that
$|\out{\alpha}| > 2\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$. In this case we show that an inversion
would exist at this level. Formally, we can find two locations $\ell'' \mathrel{\lhd} \ell'''$
in $\alpha$ such that the prefix of $\alpha$ that ends at location $\ell''$ and the suffix
of $\alpha$ that starts at location $\ell'''$ produce outputs of
length greater than $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
By Lemma \ref{lem:output-minimal-sweeping},
those two factors would not be output-minimal,
and hence $\alpha$ would contain disjoint loops $L_1,L_2$ with anchor points $\ell''_1,\ell''_2$
forming an inversion $(L_1,\ell''_1,L_2,\ell''_2)$. But this would imply that $\ell''_1,\ell''_2$
belong to the same non-singleton $\simeq$-equivalence class, which contradicts the hypothesis
that $\rho[\ell,\ell']$ does not overlap any $\simeq$-block.
We must conclude that the subfactors of $\rho[\ell_{2i},\ell_{2i+1}]$
produce outputs of length at most $2\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Overall, this shows that the output produced by each factor $\rho[\ell_{2i},\ell_{2i+1}]$
has length at most $2{\boldsymbol{H}}} %h_{\mathsf{max}}}\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
\end{proof}
\begin{comment}
LICS proof for the 2-way case
\begin{proof}
Suppose by contradiction that there is some $x\in[x_1,x_2]$
such that, for all locations $\ell=(x,y)$ between $\ell_1$ and $\ell_2$,
one of the following conditions holds:
\begin{enumerate}
\item $|\out{\rho\mid Z_\ell^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr135}{\fixed@sra}}\mspace{-2mu}}}| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
where $Z_\ell^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr135}{\fixed@sra}}\mspace{-2mu}} = [\ell,\ell_2] \:\cap\: \big([0,x]\times\bbN\big)$,
\item $|\out{\rho\mid Z_\ell^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr315}{\fixed@sra}}\mspace{-2mu}}}| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
where $Z_\ell^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr315}{\fixed@sra}}\mspace{-2mu}} = [\ell_1,\ell] \:\cap\: \big([x,\omega]\times\bbN\big)$.
\end{enumerate}
We claim first that for each condition above
there is some level $y$ at which it holds. Observe that for the
highest location $\ell$ of the run at position $x$, the set
$Z_\ell^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr135}{\fixed@sra}}\mspace{-2mu}}$ is empty, since the outgoing transition at $\ell$
is rightward. So condition 1 is trivially violated at $\ell$ as above, hence
condition 2 holds by the initial assumption. Symmetrically, condition
1 holds at the lowest location of the run at position $x$.
Let us now compare, for each condition, the levels where it holds.
Clearly, the lower the level of the location $\ell$,
the easier it is to satisfy condition 1, and symmetrically for condition 2.
So, let $\ell=(x,y)$ (resp.~$\ell'=(x,y')$) be the highest (resp.~lowest)
location at position $x$ that satisfies condition 1 (resp.~condition 2).
We claim that $y \ge y'$.
For this, we first observe that $y \ge y'-1$, since otherwise there would
exist a location $\ell=(x,y'')$, with $y < y'' < y'$, violating both conditions 1 and 2.
Moreover, $y$ must be odd, otherwise the transition departing from $\ell=(x,y)$
would be rightward oriented and the location $\ell''=(x,y+1)$ would still
satisfy condition 1, contradicting the fact that $\ell=(x,y)$ was chosen to
be the highest location.
For similar reasons, $y'$ must also be odd, otherwise there would be
a location $\ell''=(x,y'-1)$ that precedes $\ell'$ and satisfies condition 2.
But since $y \ge y'-1$ and both $y$ and $y'$ are odd, we need to have $y\ge y'$.
From the previous arguments we know that in fact $\ell=(x,y)$ satisfies both conditions
1 and 2. We can thus apply Theorem \ref{thm:simon2} to the sets
$Z_\ell^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr315}{\fixed@sra}}\mspace{-2mu}}$ and $Z_\ell^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr135}{\fixed@sra}}\mspace{-2mu}}$, deriving the existence of
two idempotent loops $L_1,L_2$ and two components $C_1,C_2$ of $L_1,L_2$,
respectively, such that
\begin{itemize}
\item $\max(L_2) < x < \min(L_1)$,
\item $\ell_1 \mathrel{\lhd} \an{C_1} \mathrel{\lhd} \ell \mathrel{\lhd} \an{C_2} \mathrel{\lhd} \ell_2$,
\item $\out{\tr{C_1}},\out{\tr{C_2}}\neq\emptystr$.
\end{itemize}
In particular, since $\an{C_1}$ is to the right of $\an{C_2}$ w.r.t.~the order
of positions, we know that $(L_1,C_1,L_2,C_2)$ is an inversion, and hence
$\an{C_1} \simeq \an{C_2}$. But this contradicts the assumption that
$\rho[\ell_1,\ell_2]$ does not overlap with any $\simeq$-block.
\end{proof}
\end{comment}
\begin{comment}
We consider the locations
that are not strictly covered by $\simeq$-blocks, formally, the
set $B = \{ \ell \:\mid\: \nexists\:(\ell_1,\ell_2)\text{ $\simeq$-block s.t.}~\ell_1<\ell<\ell_2\}$.
We equip $B$ with the natural ordering of locations induced by $\rho$.
We now consider some maximal convex subset $C$ of $B$.
Note that the left/right endpoint of $C$ coincides with the first/last
location of the run $\rho$ or with the right/left endpoint of some $\simeq$-block.
Below, we show how to decompose the subrun $\rho\mid C$ into a series of diagonals
and blocks. After this, we will be able to get a full decomposition of $\rho$ by
interleaving the $\simeq$- blocks that we defined above with the diagonals
and the blocks that decompose each subrun $\rho\mid C$.
Let $D_C$ be the set of locations $\ell=(x,y)$ of $C$ such that
there is some loop $L=[x,x']$, with $x'\ge x$, whose trace at level $y$
lies entirely inside $C$ and
produces non-empty output.
We remark that the set $D_C$ may be non-empty. To see this, one can
imagine the existence of two consecutive $\simeq$- blocks (e.g.~$R_1$
and $R_2$ in Figure~\ref{fig:decomposition}) and a loop between them
that produces non-empty output (e.g.~the factor $F_2$). In a more
general scenario, one can find several loops between two consecutive
$\simeq$- blocks that span across different levels.
We can observe however that all the locations in $D_C$ are on even levels.
Indeed, if this were not the case for some $\ell=(x,y) \in D_C$,
then we could select a minimal loop $L=[x_1,x_2]$ such that $x_1\le x\le x_2$
and $\out{\tr{\ell}} \neq \varepsilon$. Since $y$ is odd,
$\ell_1=(x_2,y)$ and $ \ell_2=(x_1,y)$ are anchor points of $L$, and hence $(\ell_1,\ell_2)$ is an inversion.
Since $\ell_1\le\ell\le\ell_2$ and all inversions are covered by
$\simeq$-blocks, there is a $\simeq$-block $(\ell'_1,\ell'_2)$ such
that $\ell'_1\le\ell\le\ell'_2$.
However, as $\ell'_1$ and $\ell'_2$ are at even levels,
$\ell$ must be different from these two locations, and
this would contradict the definition of $D_C$.
Using similar arguments, one can also show that the locations in $D_C$ are
arranged along a ``rising diagonal'', from lower left to upper right.
The above properties suggest that the locations in $D_C$ identify some
diagonals and blocks that form a decomposition of $\rho\mid C$.
The following lemma shows that this is indeed the case, namely, that
any two consecutive locations in $D_C$ form a diagonal or a block.
\begin{lem}\label{lemma:ladder}
Let $\ell_1=(x_1,y_1)$ and $\ell_2=(x_2,y_2)$ be two consecutive locations of $D_C$.
Then, $x_1\le x_2$ and $y_1\le y_2$ and the pair $(\ell_1,\ell_2)$ is a diagonal
or a block, depending on whether $y_1=y_2$ or $y_1<y_2$.
\end{lem}
\begin{proof}
Recall that all the locations in $D_C$ are on even levels, so the lemma holds in particular for $\ell_1$ and $\ell_2$
If $y_1=y_2$, then $(\ell_1,\ell_2)$ is clearly a floor.
So let us assume that $y_1 \neq y_1$.
The fact that $x_1\le x_2$ and $y_1 < y_2$ holds follows from the arguments
given in part of the body that just preceded the lemma.
The first three conditions of the definition of blockare also satisfied because
$\ell_1 $ and $\ell_2 $ are consecutive locations in $D_C$.
Below, we verify that the output produced by the factor $\rho[\ell_1,\ell_2]$
has length at most $(y_2 - y_1 +1)\cdot \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Note that this will prove both the fourth and the fifth conditions:
for $Z=[\ell_1,\ell_2] \:\setminus\: [x_1,x_2]\times[y_1,y_2]$,
the subsequence $\rho\mid Z$ produces an output that is clearly
shorter than that of $\rho[\ell_1,\ell_2]$, and in addition we have
$(y_2 - y_1 + 1) < 2(y_2 - y_1)$.
Suppose, by way of contradiction, that $\out{\rho[\ell_1,\ell_2]} > (y_2 - y_1 +1)\cdot \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Since between $\ell_1$ and $\ell_2$ there are $y_2-y_1+1$ levels,
there is a factor $\alpha$ of $\rho[\ell_1,\ell_2]$ that produces an output longer than
$\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ and that lies on a single level $y$, for some $y_1 \leq y \leq y_2$.
The contrapositive of Lemma~\ref{lemma:simple-loop} implies that
the factor $\alpha$ is captured by some loop $L'=[x'_1,x'_2]$ such that
$x'_1 > x_1$ if $y=y_1$, and $x'_2 < y_2$ if $y=y_2$.
In particular, the location $\ell'=(x'_1,y)$ is an entry point of $L$ and it
belongs to $D_C$. We have just shown that there is $\ell'\in D_C$ such that
$\ell_1<\ell'<\ell_2$. However, this contradicts the hypothesis that $\ell_1$
and $\ell_2$ were consecutive locations in $D_C$.
We must conclude that the output produced by the infix $\rho[\ell_1,\ell_2]$
has length at most $(y_2- y_1 +1) \cdot \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$, and hence $(\ell_1,\ell_2)$ is a block.
\end{proof}
\end{comment}
\smallskip
We have just shown how to construct a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition of the run $\rho$
that satisfies \PR2. This proves Proposition \ref{prop:decomposition-sweeping},
as well as the implication \PR2 $\Rightarrow $ \PR3 of Theorem \ref{thm:main2}.
\medskip
\subsection*{From existence of decompositions to an equivalent one-way transducer.}
We now focus on the last implication \PR3 $\Rightarrow $ \PR1 of Theorem \ref{thm:main2}.
More precisely, we show how to construct a one-way transducer $\cT'$ that simulates
the outputs produced by the successful runs of $\cT$ that admit $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decompositions.
In particular, $\cT'$ turns out to be equivalent to $\cT$ when $\cT$ is one-way definable.
Here we will only give a proof sketch of this construction (as there is no real difference between the sweeping and two-way cases) assuming that $\cT$ is a sweeping
transducer; a fully detailed construction of $\cT'$ from an arbitrary two-way transducer $\cT$
will be given in Section~\ref{sec:characterization-twoway} (Proposition \ref{prop:construction-twoway}),
together with a procedure for deciding one-way definability of $\cT$ (Proposition \ref{prop:complexity}).
\reviewOne[inline]{The structure of Prop 8.6 is all but indistinguishable visually/ This is made all the more aggravating by the fact that there is no corresponding proof in the sweeping case (merely the intuition sketched on Page 23.
\felix[inline]{added a sentence}
\olivier[inline]{answered in review1-answer.txt}%
}%
\begin{prop}\label{prop:construction-sweeping}
Given a functional sweeping transducer $\cT$
a one-way transducer $\cT'$ can be constructed in
$2\exptime$ such that the following hold:
\begin{enumerate}
\item $\cT' \subseteq \cT$,
\item $\dom(\cT')$ contains all words that induce successful runs of $\cT$
admitting $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-de\-com\-po\-si\-tions.
\end{enumerate}
In particular, $\cT'$ is equivalent to $\cT$ iff $\cT$ is one-way definable.
\end{prop}
\begin{proof}[Proof sketch.]
Given an input word $u$, the one-way transducer $\cT'$ needs to guess a successful run $\rho$ of $\cT$ on $u$
that admits a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition. This can be done by guessing the crossing sequences of $\rho$ at each
position,
together with a
sequence of locations $\ell_i$ that identify the factors of a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition of $\rho$.
To check the correctness of the decomposition, $\cT'$ also needs to guess a bounded amount of information
(words of bounded length) to reconstruct the outputs produced by the $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-diagonals
and the $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-blocks. For example, while scanning a factor of the input underlying a diagonal, $\cT'$
can easily reproduce the outputs of the floors and the guessed outputs of factors between them.
In a similar way, while scanning a factor of the input underlying a block, $\cT'$ can simulate
the almost periodic output by guessing its repeating pattern and the bounded prefix and suffix
of it, and by emitting the correct amount of letters, as it is done in Example \ref{ex:running}.
In particular, one can verify that the capacity of $\cT'$ is linear in ${\boldsymbol{H}}} %h_{\mathsf{max}}}\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Moreover, because the guessed objects are of size linear in ${\boldsymbol{H}}} %h_{\mathsf{max}}}\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ and ${\boldsymbol{H}}} %h_{\mathsf{max}}}\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ is a
simple exponential in the size of $\cT$, the size of the one-way transducer $\cT'$ has
doubly exponential size in that of $\cT$.
\end{proof}
\section{The characterization in the two-way case}\label{sec:characterization-twoway}
In this section we generalize the characterization of one-way
definability of sweeping transducers to the general two-way case. As
usual, we fix through the rest of the section a successful run $\rho$
of $\cT$ on some input word $u$.
\medskip
\subsection*{From periodicity of inversions to existence of decompositions.}
We continue by proving the second implication \PR2 $\Rightarrow$ \PR3 of
Theorem \ref{thm:main2} in the two-way case. This requires showing the
existence of a suitable decomposition of a run $\rho$ that
\emph{satisfies} property \PR2. Recall that \PR2 says that for every inversion
$(L_1,\ell_1,L_2,\ell_2)$, the period of the word
$\out{\tr{\ell_1}} ~ \out{\rho[\ell_1,\ell_2]} ~ \out{\tr{\ell_2}}$
divides $\gcd(|\out{\tr{\ell_1}}|, |\out{\tr{\ell_2}}|) \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
The definitions underlying the decomposition of $\rho$ are similar
to those given in the sweeping case:
\begin{defi}\label{def:factors-twoway}
Let $\rho[\ell,\ell']$ be a factor of a run $\rho$ of $\cT$,
where $\ell=(x,y)$, $\ell'=(x',y')$, and $x\le x'$.
We call $\rho[\ell,\ell']$
\begin{itemize}
\medskip
\item \parbox[t]{\dimexpr\textwidth-\leftmargin}{
\vspace{-2.75mm}
\begin{wrapfigure}{r}{8cm}
\vspace{-6mm}
\input{diagonal-twoway}
\vspace{-5mm}
\end{wrapfigure}
%
a \emph{$\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-diagonal}
if for all $z\in[x,x']$, there is a location $\ell_z$ at position $z$
such that $\ell \mathrel{\unlhd} \ell_z \mathrel{\unlhd} \ell'$ and the words
$\out{\rho|Z_{\ell_z}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr135}{\fixed@sra}}\mspace{-2mu}}}$ and
$\out{\rho|Z_{\ell_z}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr315}{\fixed@sra}}\mspace{-2mu}}}$
have length at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
where $Z_{\ell_z}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr135}{\fixed@sra}}\mspace{-2mu}} = [\ell_z,\ell'] \:\cap\: \big([0,z]\times\bbN\big)$
and $Z_{\ell_z}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr315}{\fixed@sra}}\mspace{-2mu}} = [\ell,\ell_z] \:\cap\: \big([z,\omega]\times\bbN\big)$;
}
\bigskip
\medskip
\item \parbox[t]{\dimexpr\textwidth-\leftmargin}{%
\vspace{-2.75mm}
\begin{wrapfigure}{r}{8cm}
\vspace{-6mm}
\input{block-twoway}
\vspace{-6mm}
\end{wrapfigure}
%
a \emph{$\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-block} if the word
$\out{\rho[\ell,\ell']}$ is almost periodic with bound $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
and
$\out{\rho|Z^{\shortleftarrow}}$ and
$\out{\rho|Z^{\shortrightarrow}}$ have length at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
where $Z^{\shortleftarrow} = [\ell,\ell'] \:\cap\: \big([0,x]\times \bbN\big)$
and $Z^{\shortrightarrow} = [\ell,\ell'] \:\cap\: \big([x',\omega]\times \bbN\big)$.
}
\vspace{8mm}
\end{itemize}
\end{defi}
The definition of $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition is copied verbatim from the sweeping case,
but uses the new notions of $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-diagonal and $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-block:
\begin{defi}\label{def:decomposition-twoway}
A \emph{$\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition} of a run $\rho$ of $\cT$ is a factorization
$\prod_i\,\rho[\ell_i,\ell_{i+1}]$ of $\rho$ into $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-diagonals and $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-blocks.
\end{defi}
\noindent
To provide further intuition,
we consider the transduction of Example~\ref{ex:running}
and the two-way transducer $\cT$ that implements it in the most natural way.
Figure~\ref{fig:decomposition-twoway} shows an example of a run of $\cT$ on
an input of the form $u_1 \:\#\: u_2 \:\#\: u_3 \:\#\: u_4$, where
$u_2,\, u_4 \in (abc)^*$, $u_1,\,u_3\nin (abc)^*$, and $u_3$ has even length.
The factors of the run that produce long outputs are highlighted
by the bold arrows. The first and third factors of the decomposition,
i.e.~$\rho[\ell_1,\ell_2]$ and $\rho[\ell_3,\ell_4]$, are diagonals
(represented by the blue hatched areas); the second and fourth factors
$\rho[\ell_2,\ell_3]$ and $\rho[\ell_4,\ell_5]$ are blocks
(represented by the red hatched areas).
To identify the blocks of a possible decomposition of $\rho$,
we reuse the equivalence relation $\simeq$ introduced in Definition \ref{def:crossrel}.
Recall that this is the reflexive and transitive closure of the relation $\ensuremath{\mathrel{\text{\sf S}}}$
that groups any two locations $\ell,\ell'$ that occur between $\ell_1,\ell_2$, for some
inversion $(L_1,\ell_1,L_2,\ell_2)$.
The proof that the output produced inside each $\simeq$-equivalence class is periodic,
with period at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ (Lemma \ref{lem:overlapping}) carries over in the
two-way case without modifications.
Similarly, every $\simeq$-equivalence class can be extended to the left and to the
right by using Definition \ref{def:bounding-box-sweeping}, which
we report here verbatim for the sake of readability, together with an exemplifying figure.
\input{decomposition-twoway}
\bigskip\noindent
\begin{minipage}[l]{\textwidth-5.8cm}
\begin{defi}\label{def:bounding-box-twoway}
Consider a non-singleton $\simeq$-equivalence class $K=[\ell,\ell']$.
Let $\an{K}$
be the restriction of $K$ to the anchor points occurring in some inversion,
and $X_{\an{K}} = \{x \::\: \exists y\: (x,y)\in \an{K}\}$
be the projection of $\an{K}$ on positions.
We define $\block{K}=[\tilde\ell,\tilde\ell']$, where
\begin{itemize}
\item $\tilde\ell$ is the latest location $(\tilde x,\tilde y) \mathrel{\unlhd} \ell$
such that $\tilde x = \min(X_{\an{K}})$,
\item $\tilde\ell'$ is the earliest location $(\tilde x,\tilde y) \mathrel{\unrhd} \ell'$
such that $\tilde x = \max(X_{\an{K}})$
\end{itemize}
(note that the locations $\tilde\ell,\tilde\ell'$ exist since $\ell,\ell'$
are anchor points in some inversion).
\end{defi}
\end{minipage}
\begin{minipage}[r]{5.7cm}
\vspace{-2mm}
\input{block-construction-twoway}
\end{minipage}
\medskip
As usual, we call \emph{$\simeq$-block} any factor of $\rho$ of the form $\rho|\block{K}$
that is obtained by applying the above definition to a non-singleton $\simeq$-class $K$.
Lemma \ref{lem:bounding-box-sweeping}, which shows that $\simeq$-blocks can indeed
be used as $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-blocks in a decomposition of $\rho$, generalizes easily to the
two-way case:
\reviewOne[inline]{Most proofs are quite notation-heavy. A small figure to represent the positions and factors involved is a good thing to have. For example, Lemma 5.8 has Fig.5, but the corresponding (and slightly more involved) Lemme 8.4 hasn't. This seems like a worthwile investment for any long proof, or any time a huge number of notations are involved.
}%
\gabriele[inline]{I think the reviewer is referring to the lack of a figure for the previous definition,
exactly like for Definition 5.7. I put it.}%
\reviewTwo[inline]{
-p.37 l.21-22 the $\ell$ should be replaced with $\tilde{\ell}$
\olivier[inline]{I didn't find where}%
\felix[inline]{second and third line of the list of properties we will prove. I agree.}%
}%
\begin{lem}\label{lem:bounding-box-twoway}
If $K$ is a non-singleton $\simeq$-equivalence class,
then $\rho|\block{K}$ is a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-block.
\end{lem}
\begin{proof}
The proof is similar to that of Lemma \ref{lem:bounding-box-sweeping}. The main
difference is that here we will bound the lengths of some outputs using
a Ramsey-type argument (Theorem \ref{thm:simon2}), instead of output-minimality
of factors (Lemma \ref{lem:output-minimal-sweeping}). To follow the various constructions
and arguments the reader can refer to Figure \ref{fig:block-construction-twoway}.
Let $K=[\ell,\ell']$, $\an{K}$, $X_{\an{K}}$, and $\block{K}=[\tilde\ell,\tilde\ell']$
be as in Definition \ref{def:bounding-box-twoway}, where
$\tilde\ell=(\tilde x,\tilde y)$, $\tilde\ell'=(\tilde x',\tilde y')$,
$\tilde x=\min(X_{\an{K}})$, and $\tilde x'=\max(X_{\an{K}})$.
We need to verify that $\rho|\block{K}$ is a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-block, namely, that:
\begin{itemize}
\item $\tilde x \le \tilde x'$,
\item $\out{\rho[\tilde \ell,\tilde \ell']}$ is almost periodic with bound $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
\item $\out{\rho|Z^{\shortleftarrow}}$ and $\out{\rho|Z^{\shortrightarrow}}$ have length at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
where $Z^{\shortleftarrow} = [\tilde \ell,\tilde \ell'] \:\cap\: \big([0,x]\times \bbN\big)$
and $Z^{\shortrightarrow} = [\tilde \ell,\tilde \ell'] \:\cap\: \big([x',\omega]\times \bbN\big)$.
\end{itemize}
The first condition $\tilde x \le \tilde x'$ follows immediately from
$\tilde x=\min(X_{\an{K}})$ and $\tilde x'=\max(X_{\an{K}})$.
Next, we prove that the output produced by the factor
$\rho[\tilde\ell,\tilde\ell']$ is almost periodic with bound $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
By Definition \ref{def:bounding-box-twoway}, we have
$\tilde\ell \mathrel{\unlhd} \ell \mathrel{\lhd} \ell' \mathrel{\unlhd} \tilde\ell'$,
and by Lemma \ref{lem:overlapping}
we know that $\out{\rho[\ell,\ell']}$ is periodic with
period at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$. So it suffices to bound the length
of the words $\out{\rho[\tilde\ell,\ell]}$ and $\out{\rho[\ell',\tilde\ell']}$.
We shall focus on the former word, as the arguments for the latter
are similar.
First, we show that the factor $\rho[\tilde\ell,\ell]$
lies entirely to the right of position $\tilde x$
(in particular, it starts at an even level $\tilde y$).
Indeed, if this were not the case, there would exist another location
$\ell''=(\tilde x,\tilde y + 1)$, on the same position as $\tilde\ell$,
but at a higher level, such that $\tilde\ell \mathrel{\lhd} \ell'' \mathrel{\unlhd} \ell$.
But this would contradict Definition \ref{def:bounding-box-twoway}
($\tilde\ell$ is the \emph{latest} location $(x,y) \mathrel{\unlhd} \ell$
such that $x = \min(X_{\an{K}})$).
Suppose now that the length of $|\out{\rho[\tilde\ell,\ell]}| >
\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$. We head towards a contradiction by finding a location
$\ell'' \mathrel{\lhd} \ell$ that is $\simeq$-equivalent to the first
location $\ell$ of the $\simeq$-equivalence class $K$. Since the
factor $\rho[\tilde\ell,\ell]$ lies entirely to the right of position
$\tilde x$, it is intercepted by the interval $I=[\tilde x,\omega]$.
So $|\out{\rho[\tilde\ell,\ell]}| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ is equivalent to saying
$|\out{\rho|Z}| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$, where $Z = [\tilde\ell,\ell] \:\cap\:
\big([\tilde x,\omega]\times\bbN\big)$. Then, Theorem
\ref{thm:simon2} implies the existence of an idempotent loop $L$ and
an anchor point $\ell''$ of $L$ such that
\begin{itemize}
\item $\min(L) > \tilde x$,
\item $\tilde\ell \mathrel{\lhd} \ell'' \mathrel{\lhd} \ell$,
\item $\out{\tr{\ell''}}\neq\emptystr$.
\end{itemize}
Further recall that $\tilde x=\min(X_{\an{K}})$ is the leftmost position of
locations in the class $K=[\ell,\ell']$ that are also anchor points of inversions.
In particular, there is an inversion $(L_1,\ell''_1,L_2,\ell''_2)$, with
$\ell''_2=(\tilde x,y''_2) \in K$.
Since $\ell'' \mathrel{\lhd} \ell \mathrel{\unlhd} \ell''_2$
and the position of $\ell''$ is to the right of the position of $\ell''_2$,
we know that $(L,\ell'',L_2,\ell''_2)$ is also an inversion,
and hence $\ell'' \simeq \ell''_2 \simeq \ell$.
But since $\ell'' \neq \ell$, we get a contradiction with the
assumption that $\ell$ is the first location of the $\simeq$-class $K$.
In this way we have shown that $|\out{\rho[\tilde\ell,\ell]}| \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
It remains to bound the lengths of the outputs produced
by the subsequences $\rho|Z^{\shortleftarrow}$ and $\rho|Z^{\shortrightarrow}$,
where $Z^{\shortleftarrow}=[\tilde\ell,\tilde\ell'] \:\cap\: \big([0,\tilde x]\times\bbN\big)$
and $Z^{\shortrightarrow}=[\tilde\ell,\tilde\ell'] \:\cap\: \big([\tilde x',\omega]\times\bbN\big)$.
As usual, we consider only one of the two symmetric cases.
Suppose, by way of contradiction, that $|\out{\rho|Z^{\shortleftarrow}}| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
By Theorem \ref{thm:simon2}, there exist an idempotent loop $L$
and an anchor point $\ell''$ of $L$ such that
\begin{itemize}
\item $\max(L) < \tilde x$,
\item $\tilde\ell \mathrel{\lhd} \ell'' \mathrel{\lhd} \tilde\ell'$,
\item $\out{\tr{\ell''}}\neq\emptystr$.
\end{itemize}
By following the same line of reasoning as before, we recall that
$\ell$ is the first location of the non-singleton class $K$.
From this we derive the existence an inversion $(L_1,\ell''_1,L_2,\ell''_2)$
where $\ell''_1 = \ell$.
We claim that $\ell \mathrel{\unlhd} \ell''$.
Indeed, if this were not the case, then, because $\ell''$ is strictly to the
left of $\tilde x$ and $\ell$ is to the right of $\tilde x$, there would exist
a location $\ell'''$ between $\ell''$ and $\ell$ that lies at position $\tilde x$.
But $\tilde\ell \mathrel{\lhd} \ell'' \mathrel{\unlhd} \ell''' \mathrel{\unlhd} \ell$ would
contradict the fact that $\tilde\ell$ is the {\sl latest} location before
$\ell$ that lies at the position $\tilde x$.
Now that we know that $\ell \mathrel{\unlhd} \ell''$ and that $\ell''$ is to the left of $\tilde x$,
we observe that $(L_1,\ell''_1,L,\ell'')$ is also an inversion, and hence $\ell''\in \an{K}$.
Since $\ell''$ is strictly to the left of $\tilde x$,
we get a contradiction with the definition of $\tilde x$ as leftmost
position of the locations in $\an{K}$.
So we conclude that $|\out{\rho|Z^{\shortleftarrow}}| \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
\end{proof}
The proof of Lemma \ref{lem:consecutive-blocks-sweeping}, which
shows that $\simeq$-blocks do not overlap along the input axis,
carries over in the two-way case, again without modifications.
Finally, we generalize Lemma \ref{lem:diagonal-sweeping} to the new
definition of diagonal, which completes the construction of a
$\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition for the run $\rho$:
\begin{lem}\label{lem:diagonal-twoway}
Let $\rho[\ell,\ell']$ be a
factor of $\rho$,
with $\ell=(x,y)$, $\ell'=(x',y')$, and $x\le x'$,
that does not overlap any $\simeq$-block.
Then $\rho[\ell,\ell']$ is a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-diagonal.
\end{lem}
\begin{proof}
Suppose by way of contradiction that there is some $z \in [x,x']$
such that, for all locations $\ell''$ at position $z$ and between $\ell$ and $\ell'$,
one of the two conditions holds:
\begin{enumerate}
\item $|\out{\rho|Z_{\ell''}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr135}{\fixed@sra}}\mspace{-2mu}}}| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
where $Z_{\ell''}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr135}{\fixed@sra}}\mspace{-2mu}} = [\ell'',\ell'] \:\cap\: \big([0,z]\times\bbN\big)$,
\item $|\out{\rho|Z_{\ell''}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr315}{\fixed@sra}}\mspace{-2mu}}}| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
where $Z_{\ell''}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr315}{\fixed@sra}}\mspace{-2mu}} = [\ell,\ell''] \:\cap\: \big([z,\omega]\times\bbN\big)$.
\end{enumerate}
First, we claim that \emph{each} of the two conditions above are satisfied at
some locations $\ell''\in [\ell,\ell']$ at position $z$.
Consider the highest even level $y''$
such that $\ell''=(z,y'') \in[\ell,\ell']$
(use Figure \ref{fig:diagonal-twoway} as a reference).
Since $z\le x'$, the outgoing transition at $\ell''$ is rightward oriented,
and the set $Z_{\ell''}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr135}{\fixed@sra}}\mspace{-2mu}}$ is empty. This means that
condition (1) is trivially violated at $\ell''$,
and hence condition (2) holds at $\ell''$ by the initial assumption.
Symmetrically, condition (1) holds at the location $\ell''=(z,y'')$,
where $y''$ is the lowest even level with $\ell'' \in[\ell,\ell']$
Let us now compare the levels where the above conditions hold.
Clearly, the lower the level of location $\ell''$,
the easier it is to satisfy condition (1), and symmetrically for condition (2).
So, let $\ell^+=(z,y^+)$ (resp.~$\ell^-=(z,y^-)$) be the highest (resp.~lowest)
location in $[\ell,\ell']$ at position $z$ that satisfies
condition (1) (resp.~condition (2)).
We claim that $y^+ \ge y^-$.
For this, we first observe that $y^+ \ge y^- - 1$, since otherwise there
would exist a location $\ell''=(z,y'')$, with $y^+ < y'' < y^-$, that
violates both conditions (1) and (2).
Moreover, $y^+$ must be odd, otherwise the transition departing from
$\ell^+ = (z,y^+)$ would be rightward oriented and the location $\ell'' = (z,y^+ + 1)$
would still satisfy condition (1), contradicting the definition of highest location $\ell^+$.
For similar reasons, $y^-$ must also be odd, otherwise there would be a location
$\ell'' = (z,y^- - 1)$ below $\ell^-$ that satisfies condition (2).
But since $y^+ \ge y^- - 1$ and both $y^+$ and $y^-$ are odd,
we need to have $y^+ \ge y^-$.
In fact, from the previous arguments we know that the location $\ell''=(z,y^+)$
(or equally the location $(x,y^-)$)
satisfies {\sl both} conditions (1) and (2). We can thus apply Theorem \ref{thm:simon2} to the
sets $Z_{\ell''}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr315}{\fixed@sra}}\mspace{-2mu}}$ and $Z_{\ell''}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr135}{\fixed@sra}}\mspace{-2mu}}$, deriving the existence of
two idempotent loops $L_1,L_2$ and two anchor points $\ell_1,\ell_2$ of $L_1,L_2$,
respectively, such that
\begin{itemize}
\item $\max(L_2) < z < \min(L_1)$,
\item $\ell \mathrel{\lhd} \ell_1 \mathrel{\lhd} \ell'' \mathrel{\lhd} \ell_2 \mathrel{\lhd} \ell'$,
\item $\out{\tr{\ell_1}},\out{\tr{\ell_2}}\neq\emptystr$.
\end{itemize}
In particular, since $\ell_1$ is to the right of $\ell_2$ w.r.t.~the order
of positions, we know that $(L_1,\ell_1,L_2,\ell_2)$ is an inversion, and
hence $\ell_1 \simeq \ell_2$. But this contradicts the assumption that
$\rho[\ell,\ell']$ does not overlap with any $\simeq$-block.
\end{proof}
\medskip
\reviewOne[inline]{Part 5, "From existence of decompositions to an equivalent one-way transducer": so far the paper did a great job hinting at the proof that remains. Here an intuition is given for the construction later detailed in 8.6, but no mention of an intuition for Proposition 9.2: if it is at all possible to do so, this might help the paper's overall flow.
\olivier[inline]{answered in review1-answers.txt}
}%
\subsection*{From existence of decompositions to an equivalent one-way transducer.}
It remains to prove the last implication \PR3 $\Rightarrow$ \PR1 of Theorem~\ref{thm:main2},
which amounts to construct a one-way transducer $\cT'$ equivalent to $\cT$.
Hereafter, we denote by $D$ the language of words $u\in\dom(\cT)$ such that
{\sl all} successful runs of $\cT$ on $u$ admit a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition.
So far, we know that if $\cT$ is one-way definable (\PR1),
then $D=\dom(\cT)$ (\PR3).
As a matter of fact, this reduces the one-way definability problem
for $\cT$ to the containment problem $\dom(\cT) \subseteq D$.
\label{testing-containment}
We will see later (in Section~\ref{sec:complexity})
how the latter problem can be decided
in double exponential space
by further reducing it to checking the emptiness of the
intersection of the languages $\dom(\cT)$ and $D^\complement$,
where $D^\complement$ is the complement of $D$.
Below, we show how to construct a one-way transducer
$\cT'$ of triple exponential size such that
$\cT' \subseteq \cT$ and
$\dom(\cT')$ is the set of all input words that have
{\sl some} successful run admitting a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition
(hence $\dom(\cT')\supseteq D$).
In particular, we will have that
\[
\cT|_D \:\subseteq\: \cT' \:\subseteq\: \cT.
\]
Note that this will prove \PR3 to \PR1, as well as the second
item of Theorem~\ref{thm:main}, since $D=\dom(\cT)$
if and only if $\cT$ is one-way definable.
A sketch of the proof of this construction when
$\cT$ is a sweeping transducer was given at the
end of Section \ref{sec:characterization-sweeping}.
\begin{prop}\label{prop:construction-twoway}
Given a functional two-way transducer $\cT$,
a one-way transducer $\cT'$
satisfying
\[\cT' \subseteq \cT \quad \text{ and } \quad \dom(\cT') \supseteq D\] can be constructed in $3\exptime$.
Moreover, if $\cT$ is sweeping, then $\cT'$
can be constructed in $2\exptime$.
\end{prop}
\begin{proof}
Given an input word $u$, the transducer $\cT'$ will guess (and check)
a successful run $\rho$ of $\cT$ on $u$, together with a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition
$\prod_i \rho[\ell_i,\ell_{i+1}]$.
The latter decomposition will be used by $\cT'$ to simulate the output of
$\rho$ in left-to-right manner, thus proving that $\cT' \subseteq \cT$.
Moreover, $u\in D$ implies the existence of a successful run that can be
decomposed, thus proving that $\dom(\cT') \supseteq D$.
We now provide the details of the construction of $\cT'$.
Guessing the run $\rho$ is standard (see, for instance, \cite{she59,HU79}):
it amounts to guess the crossing sequences $\rho|x$ for
each position $x$ of the input. Recall that this is a bounded
amount of information for each position $x$, since the run is normalized.
As concerns the decomposition of $\rho$, it can be encoded
by the endpoints $\ell_i$ of its factors, that is, by annotating
the position of each $\ell_i$ as the level of $\ell_i$.
In a similar way $\cT'$ guesses the information of whether
each factor $\rho[\ell_i,\ell_{i+1}]$ is a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-diagonal or a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-block.
Thanks to the definition of decomposition
(see Definition~\ref{def:decomposition-twoway} and Figure \ref{fig:decomposition-twoway}),
every two distinct factors span across non-overlapping intervals of positions.
This means that each position $x$ is covered by exactly one factor of
the decomposition. We call this factor the \emph{active factor at position $x$}.
The mode of computation of the transducer will depend on
the type of active factor: if the active factor is a diagonal
(resp.~a block), then we say that $\cT'$ is in \emph{diagonal mode}
(resp.~\emph{block mode}).
Below we describe the behaviour for these two modes of computation.
\smallskip
\par\noindent\emph{Diagonal mode.}~
We recall the key condition satisfied by the diagonal
$\rho[\ell,\ell']$ that is active at position $x$
(cf.~Definition~\ref{def:factors-twoway} and Figure~\ref{fig:diagonal-twoway}):
there is a location $\ell_x=(x,y_x)$ between $\ell$ and $\ell'$ such that the words
$\out{\rho|Z_{\ell_x}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr135}{\fixed@sra}}\mspace{-2mu}}}$ and $\out{\rho|Z_{\ell_x}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr315}{\fixed@sra}}\mspace{-2mu}}}$
have length at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$, where
$Z_{\ell_x}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr135}{\fixed@sra}}\mspace{-2mu}} = [\ell_x,\ell'] \:\cap\: \big([0,x]\times\bbN\big)$
and $Z_{\ell_x}^{\mspace{-2mu}\text{\rotatebox[origin=c]{\numexpr315}{\fixed@sra}}\mspace{-2mu}} = [\ell,\ell_x] \:\cap\: \big([x,\omega]\times\bbN\big)$.
Besides the run $\rho$ and the decomposition, the transducer $\cT'$ will
also guess the locations $\ell_x=(x,y_x)$, that is, will annotate each $x$
with the corresponding $y_x$.
Without loss of generality, we can assume that the function that
associates each position $x$ with the guessed location $\ell_x=(x,y_x)$
is monotone, namely, $x\le x'$ implies $\ell_x\mathrel{\unlhd}\ell_{x'}$.
While the transducer $\cT'$ is in diagonal mode, the goal is to preserve
the following invariant:
\begin{quote}
\em
After reaching a position $x$ covered by the active diagonal,
$\cT'$ must have produced the output of $\rho$ up to location $\ell_x$.
\end{quote}
\noindent
To preserve the above invariant when moving from $x$ to the next
position $x+1$, the transducer should output the word
$\out{\rho[\ell_x,\ell_{x+1}]}$. This word consists of
the following parts:
\begin{enumerate}
\item The words produced by the single transitions of $\rho[\ell_x,\ell_{x+1}]$
with endpoints in $\{x,x+1\}\times\bbN$.
Note that there are at most ${\boldsymbol{H}}} %h_{\mathsf{max}}}$ such words,
each of them has length at most ${\boldsymbol{C}}} %c_{\mathsf{max}}}$, and they can all be determined
using the crossing sequences at $x$ and $x+1$ and the information
about the levels of $\ell_x$ and $\ell_{x+1}$.
We can thus assume that this information is readily available
to the transducer.
\item The words produced by the factors of $\rho[\ell_x,\ell_{x+1}]$
that are intercepted by the interval $[0,x]$.
Thanks to the definition of diagonal, we know that
the total length of these words is at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
These words cannot be determined from the information
on $\rho|x$, $\rho|x+1$, $\ell_x$, and $\ell_{x+1}$
alone, so they need to be constructed while scanning the input.
For this, some additional information needs to be stored.
More precisely, at each position $x$ of the input,
the transducer stores all the outputs produced by the factors of
$\rho$ that are intercepted by $[0,x]$ and that occur {\sl after}
a location of the form $\ell_{x'}$, for any $x'\ge x$ that is
covered by a diagonal.
This clearly includes the previous words when $x'=x$, but also
other words that might be used later for processing other diagonals.
Moreover, by exploiting the properties of diagonals,
one can prove that those words have length at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
so they can be stored with triply exponentially many states.
Using classical techniques, the stored information
can be maintained while scanning the input $u$ using the
guessed crossing sequences of $\rho$.
\item The words produced by the factors of $\rho[\ell_x,\ell_{x+1}]$
that are intercepted by the interval $[x+1,\omega]$.
These words must be guessed, since they depend on a portion
of the input that has not been processed yet.
Accordingly, the guesses need to be stored into memory,
in such a way that they can be checked later. For this, the transducer
stores, for each position $x$, the guessed words that correspond
to the outputs produced by the factors of $\rho$ intercepted by
$[x,\omega]$ and occurring {\sl before} a location of the form
$\ell_{x'}$, for any $x'\le x$ that is covered by a diagonal.
\end{enumerate}
\smallskip
\par\noindent\emph{Block mode.}~
Suppose that the active factor $\rho[\ell,\ell']$ is a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-block.
Let $I=[x,x']$ be the set of positions covered by this factor.
Moreover, for each position $z\in I$, let
$Z^{\shortleftarrow}_z = [\ell,\ell'] \:\cap\: \big([0,z]\times \bbN\big)$
and $Z^{\shortrightarrow}_z = [\ell,\ell'] \:\cap\: \big([z,\omega]\times \bbN\big)$.
We recall the key property of a block
(cf.~Definition~\ref{def:factors-twoway} and Figure~\ref{fig:block-twoway}):
the word $\out{\rho[\ell,\ell']}$ is almost periodic with bound $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
and the words $\out{\rho|Z^{\shortleftarrow}_x}$ and $\out{\rho|Z^{\shortrightarrow}_{x'}}$
have length at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
For the sake of brevity, suppose that $\out{\rho[\ell,\ell']} = w_1\,w_2\,w_3$,
where $w_2$ is periodic with period $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ and $w_1,w_3$
\reviewOne[inline]{$w_1,w_3$ I assume.
\olivier[inline]{fixed (was $w_1,w_2$)}%
}%
have length at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Similarly, let $w_0 = \out{\rho|Z^{\shortleftarrow}_x}$ and $w_4 = \out{\rho|Z^{\shortrightarrow}_{x'}}$.
The invariant preserved by $\cT'$ in block mode is the following:
\begin{quote}
\em
After reaching a position $z$ covered by the active block $\rho[\ell,\ell']$,
$\cT'$ must have produced the output of the prefix of $\rho$
up to location $\ell$, followed by a prefix of $\out{\rho[\ell,\ell']} = w_1\,w_2\,w_3$
of the same length as $\out{\rho|Z^{\shortleftarrow}_z}$.
\end{quote}
\noindent
The initialization of the invariant is done when reaching the left
endpoint $x$. At this moment, it suffices that $\cT'$ outputs
a prefix of $w_1\,w_2\,w_3$ of the same length as
$w_0 = \out{\rho|Z^{\shortleftarrow}_x}$, thus bounded by $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Symmetrically, when reaching the right endpoint $x'$,
$\cT'$ will have produced almost the entire word
$\out{\rho[\ell,\ell']} \, w_1 \, w_2 \, w_3$,
but without the suffix $w_4 = \out{\rho|Z^{\shortrightarrow}_{x'}}$
of length at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Thus, before moving to the next factor of the decomposition, the transducer will
produce the remaining suffix, so as to complete the output
of $\rho$ up to location $\ell_{i_x+1}$.
It remains to describe how the above invariant can be maintained
when moving from a position $z$ to the next position $z+1$ inside $I=[x,x']$.
For this, it is convenient to succinctly represent the word $w_2$
by its repeating pattern, say $v$, of length at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
To determine the symbols that have to be output at each step,
the transducer will maintain a pointer on either $w_1\,v$ or $w_3$.
The pointer is increased in a deterministic way, and precisely
by the amount $|\out{\rho|Z^{\shortleftarrow}_{z+1}}| - |\out{\rho|Z^{\shortleftarrow}_z}|$.
The only exception is when the pointer lies in $w_1\,v$, but its
increase would go over $w_1\,v$: in this case the transducer has
the choice to either bring the pointer back to the beginning of $v$
(representing a periodic output inside $w_2$), or move it to $w_3$.
Of course, this is a non-deterministic choice, but it can be
validated when reaching the right endpoint of $I$.
Concerning the number of symbols that need to be emitted at each
step, this can be determined from the crossing sequences at
$z$ and $z+1$, and from the knowledge of the lowest and highest
levels of locations that are at position $z$ and between
$\ell$ and $\ell'$. We denote the latter levels by
$y^-_z$ and $y^+_z$, respectively.
Overall, this shows how to maintain the invariant of the block mode,
assuming that the levels $y^-_z,y^+_z$ are known, as well as
the words $w_0,w_1,v,w_3,w_4$ of bounded length.
Like the mapping $z \mapsto \ell_z=(z,y_z)$ used in diagonal mode,
the mapping $z \mapsto (y^-_z,y^+_z)$ can be guessed and checked
using the crossing sequences.
Similarly, the words $w_1,v,w_3$ can be guessed just before
entering the active block, and can be checked along the process.
As concerns the words $w_0,w_4$, these can be guessed and checked
in a way similar to the words that we used in diagonal mode.
More precisely, for each position $z$ of the input, the
transducer stores the following additional information:
\begin{enumerate}
\item the outputs produced by the factors of $\rho$ that are
intercepted by $[0,z]$ and that occur after the beginning
$\ell''$ of some block, with $\ell''=(x'',y'')$ and $x''\ge z$;
\item the outputs produced by the factors of $\rho$ that are
intercepted by $[z,\omega]$ and that occur before the ending
$\ell'''$ of a block, where $\ell'''=(x''',y''')$
and $x'''\le z$.
\end{enumerate}
By the definition of blocks, the above words have length
at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ and can be maintained while processing the input
and the crossing sequences.
Finally, we observe that the words, together with the information
given by the lowest and highest levels $y^-_z,y^+_z$, for both $z=x$ and
$z=x'$, are sufficient for determining the content of $w_0$ and $w_4$.
\smallskip
We have just shown how to construct a one-way transducer $\cT' \subseteq \cT$
such that $\dom(\cT') \supseteq D$.
From the above construction it is easy to see that the number of states
and transitions of $\cT'$, as well as the number of letters emitted by
each transition, are at most exponential in $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$. Since $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ is
doubly exponential in the size of $\cT$, this shows that $\cT'$ can
be constructed from $\cT$ in $3\exptime$.
Note that the triple exponential
complexity comes from the lengths of the words that need to be guessed
and stored in the control states, and these lengths are bounded by $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
However, if $\cT$ is a sweeping transducer, then, according to the results
proved in Section \ref{sec:characterization-sweeping}, the bound $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$
is simply exponential. In particular, in the sweeping case
we can construct the one-way transducer $\cT'$ in $2\exptime$.
\end{proof}
\medskip
\subsection*{Generality of the construction.}
We conclude the section with a discussion on the properties of the one-way
transducer $\cT'$ constructed from $\cT$. Roughly speaking, we would like
to show that, even when $\cT$ is not one-way definable, $\cT'$ is somehow
the {\sl best one-way under-approximation of $\cT$}.
However, strictly speaking, the latter terminology is meaningless:
if $\cT'$ is a one-way transducer strictly contained in $\cT$, then
one can always find a better one-way transducer $\cT''$ that satisfies
$\cT' \subsetneq \cT'' \subsetneq \cT$, for instance by extending $\cT'$
with a single input-output pair. Below, we formalize in an appropriate
way the notion of ``best one-way under-approximation''.
We are interested in comparing the domains of transducers, but only up to
a certain amount. In particular, we are interested in languages that are
preserved under pumping loops of runs of $\cT$. Formally, given a language
$L$, we say that $L$ is \emph{$\cT$-pumpable} if $L \subseteq \dom(\cT)$ and
for all words $u\in L$, all successful runs $\rho$ of $\cT$ on $u$, all
loops $L$ of $\rho$, and all positive numbers $n$, the word $\ensuremath{\mathsf{pump}}_L^n(u)$
also belongs to $L$.
Clearly, the domain $\dom(\cT)$ of a transducer $\cT$ is a regular $\cT$-pumpable language.
Another noticeable example of $\cT$-pumpable regular language is the domain
of the one-way transducer $\cT'$, as defined in Proposition \ref{prop:construction-twoway}.
Indeed, $\dom(\cT')$ consists of words $u\in\dom(\cT)$ that induce
successful runs with $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decompositions, and the property of
having a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition is preserved under pumping.
The following result shows that $\cT'$ is the best under-approximation
of $\cT$ within the class of one-way transducers with $\cT$-pumpable domains:
\begin{cor}\label{cor:best-underapproximation}
Given a functional two-way transducer $\cT$, one can construct a one-way transducer $\cT'$ such that
\begin{itemize}
\item $\cT' \subseteq \cT$ and $\dom(\cT')$ is $\cT$-pumpable,
\item for all one-way transducers $\cT''$, if $\cT'' \subseteq \cT$ and $\dom(\cT'')$ is $\cT$-pumpable,
then $\cT'' \subseteq \cT'$.
\end{itemize}
\end{cor}
\begin{proof}
The transducer $\cT'$ is precisely the one defined in Proposition \ref{prop:construction-twoway}.
As already explained, its domain $\dom(\cT')$ is a $\cT$-pumpable language. In particular, $\cT'$
satisfies the conditions in the first item.
For the conditions in the second item, consider a one-way transducer $\cT'' \subseteq \cT$
with a $\cT$-pumpable domain $L=\dom(\cT'')$. Let $\tilde\cT$ be the transducer obtained from
$\cT$ by restricting its domain to $L$. Clearly, $\tilde\cT$ is one-way definable, and one
could apply Proposition \ref{prop:periodicity-twoway} to $\tilde\cT$, using $\cT''$ as a
witness of one-way definability. In particular, when it comes to comparing the outputs of the
pumped runs of $\tilde\cT$ and $\cT''$, one could exploit the fact that the domain $L$ of $\cT''$,
and hence the domain of $\tilde\cT$ as well, is $\cT$-pumpable. This permits to derive
periodicities of inversions with the same bound $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ as before, but only restricted
to the successful runs of $\cT$ on the input words that belong to $L$.
As a consequence, one can define $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decompositions of successful runs of $\cT$
on words in $L$, thus showing that $L \subseteq \dom(\cT')$. This proves that $\cT'' \subseteq \cT'$.
\end{proof}
\section{Basic combinatorics for sweeping transducers}\label{sec:combinatorics-sweeping}
We fix for the rest of the section a functional \emph{sweeping transducer} $\cT$,
an input word $u$, and a (normalized) successful run $\rho$ of $\cT$ on $u$.
\medskip
\subsection*{Pumping loops.}
Loops turn out to be a basic concept for characterizing one-way definability.
Formally, a \emph{loop} of $\rho$ is an interval $L=[x_1,x_2]$ such that $\rho|x_1=\rho|x_2$,
namely, with the same crossing sequences at the extremities.
The run $\rho$ can be pumped at any loop $L=[x_1,x_2]$, and this gives rise
to new runs with iterated factors. Below we study precisely the shape of
these pumped runs.
\begin{defi}[anchor point, trace]
Given a loop $L$ and a location $\ell$ of $\rho$, we say that $\ell$ is an
\emph{anchor point in $L$} if $\ell$ is the first location of some factor
of $\rho$ that is intercepted by $L$;
this factor is then denoted%
\footnote{This is a slight abuse of notation, since the factor $\tr{\ell}$
is not determined by $\ell$ alone, but requires also the knowledge of the loop $L$,
which is usually clear from the context.}
as $\tr{\ell}$ and called the \emph{trace of $\ell$}.
\end{defi}
Observe that a loop can have at most ${\boldsymbol{H}}} %h_{\mathsf{max}}} = 2|Q|-1$ anchor points, since
we consider only normalized runs.
Given a loop $L$ of $\rho$ and a number $n\in\bbN$, we can replicate $n$ times
the factor $u[x_1,x_2]$ of the input, obtaining a new input of the form
\begin{equation}\label{eq:pumped-word}
\ensuremath{\mathsf{pump}}_L^{n+1}(u) ~=~ u[1,x_1]\cdot \big(u[x_1+1,x_2]\big)^{n+1} \cdot u[x_2+1,|u|].
\end{equation}
\reviewTwo[inline]{The notation $u[x_1,x_2]$ is not defined.\olivier[inline]{fixed}}
Similarly, we can replicate $n$ times the intercepted factors $\tr{\ell}$ of $\rho$,
for all anchor points $\ell$ of $L$. In this way we obtain a successful run on $\ensuremath{\mathsf{pump}}_L^{n+1}(u)$
that is of the form
\begin{equation}\label{eq:pumped-run}
\ensuremath{\mathsf{pump}}_L^{n+1}(\rho) ~=~ \rho_0 ~ \tr{\ell_1}^n ~ \rho_1 ~ \dots ~ \rho_{k-1} ~ \tr{\ell_k}^n ~ \rho_k
\end{equation}
where $\ell_1\mathrel{\unlhd}\dots\mathrel{\unlhd}\ell_k$ are all the anchor points in $L$
(listed according to the run order $\mathrel{\unlhd}$), $\rho_0$ is the prefix of $\rho$
ending at $\ell_1$,
$\rho_k$ is the suffix of $\rho$
starting at $\ell_k$, and for all $i=1,\dots,k-1$,
$\rho_i$ is the factor of $\rho$ between
$\ell_i$ and $\ell_{i+1}$.
Note that $\ensuremath{\mathsf{pump}}_L^1(\rho)$ coincides with the original run $\rho$. As a matter of fact,
one could define in a similar way the run $\ensuremath{\mathsf{pump}}_L^0(\rho)$ obtained from removing the loop
$L$ from $\rho$. However, we do not need this, and we will always parametrize the operation
$\ensuremath{\mathsf{pump}}_L$ by a positive number $n+1$.
An example of a pumped run $\ensuremath{\mathsf{pump}}_{L_1}^3(\rho)$ is given in Figure \ref{fig:pumping-sweeping},
together with the indication of the anchor points $\ell_i$ and the intercepted factors $\tr{\ell_i}$.
\input{pumping-sweeping}
\medskip
\subsection*{Output minimality.}
We are interested into factors of the run $\rho$ that lie on a single level
and that contribute to the final output, but in a minimal way, in the sense
that is formalized by the following definition:
\begin{defi}\label{def:output-minimal-sweeping}
Consider a factor $\a=\rho[\ell,\ell']$ of $\rho$.
We say that $\a$ is \emph{output-minimal} if
$\ell=(x,y)$ and $\ell'=(x',y)$, and all loops $L \subsetneq
[x,x']$
produce empty output at level $y$.
\end{defi}
From now on, we set the constant $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax} = {\boldsymbol{C}}} %c_{\mathsf{max}}} |Q|^{\boldsymbol{H}}} %h_{\mathsf{max}}} +1$,
where ${\boldsymbol{C}}} %c_{\mathsf{max}}}$ is the capacity of the transducer, that is, the
maximal length of an output produced on a single transition (recall
that $|Q|^{\boldsymbol{H}}} %h_{\mathsf{max}}}$ is the maximal number of crossing sequences).
As shown below, $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ bounds the length of
the output produced by an output-minimal factor:
\begin{lem}\label{lem:output-minimal-sweeping}
For all output-minimal factors $\alpha$,
$|\out{\alpha}| \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
\end{lem}
\begin{proof}
Suppose by contradiction that $|\out{\alpha}| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$, with
$\a=\rho[\ell,\ell']$, $\ell=(x,y)$ and $\ell=(x',y)$.
Let $X$ be the set of all positions $x''$, with $\min(x,x') < x'' < \max(x,x')$,
that are sources of transitions of $\alpha$ that produce non-empty output.
Clearly, the total number of letters produced by the transitions that depart
from locations in $X\times\{y\}$ is strictly larger than $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}-1$.
Moreover, since each transition emits at most ${\boldsymbol{C}}} %c_{\mathsf{max}}}$ symbols, we have $|X| > \frac{\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}-1}{{\boldsymbol{C}}} %c_{\mathsf{max}}}} = |Q|^{\boldsymbol{H}}} %h_{\mathsf{max}}}$.
Now, recall that crossing sequences are sequences of states of length at most ${\boldsymbol{H}}} %h_{\mathsf{max}}}$.
Since $|X|$ is larger than the number of crossing sequences, $X$ contains two positions
$x_1<x_2$ such that $\rho|x_1=\rho|x_2$. In particular, $L=[x_1,x_2]$
is a loop strictly between $x,x'$
with non-empty output on level $y$.
This shows that $\rho[\ell,\ell']$ is not output-minimal.
\end{proof}
\medskip
\subsection*{Inversions and periodicity.}
Next, we define the crucial notion of inversion. Intuitively, an inversion
in a run identifies a part of the run that is potentially difficult to
simulate in a one-way manner because the order of generating the
output is reversed w.r.t.~the input. Inversions arise naturally in transducers
that reverse arbitrarily long portions of the input, as well as in transducers
that produce copies of arbitrarily long portions of the input.
\label{page-def-inversion}
\begin{defi}\label{def:inversion-sweeping}
An \emph{inversion} of the run $\rho$ is a tuple $(L_1,\ell_1,L_2,\ell_2)$ such that
\begin{enumerate}
\item $L_1,L_2$ are loops of $\r$,
\item $\ell_1=(x_1,y_1)$ and $\ell_2=(x_2,y_2)$
are anchor points of $L_1$ and $L_2$, respectively,
\item $\ell_1 \mathrel{\lhd} \ell_2$ and $x_1 > x_2$
\par\noindent
(namely, $\ell_2$ follows $\ell_1$ in the run,
but the position of $\ell_2$ precedes the position of $\ell_1$),
\item for both $i=1$ and $i=2$, $\out{\tr{\ell_i}}\neq\emptystr$ and $\tr{\ell_i}$ is output-minimal.
\end{enumerate}
\end{defi}
\noindent
The left hand-side of Figure~\ref{fig:inversion-sweeping} gives an example of an inversion,
assuming that the outputs $v_1=\tr{\ell_1}$ and $v_2=\tr{\ell_2}$ are non-empty
and the intercepted factors are output-minimal.
\input{inversion-sweeping}
The rest of the section is devoted to prove the implication \PR1 $\Rightarrow$ \PR2
of Theorem \ref{thm:main2}.
We recall that a word $w=a_1 \cdots a_n$ has \emph{period} $p$ if for every $1\le i\le |w|-p$,
we have $a_i = a_{i+p}$. For example, the word $abc \, abc \, ab$ has period $3$.
We remark that, thanks to Lemma \ref{lem:output-minimal-sweeping},
for every inversion $(L_1,\ell_1,L_2,\ell_2)$, the outputs
$\out{\tr{\ell_1}}$ and $\out{\tr{\ell_2}}$ have length at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
By pairing this with the assumption that the transducer $\cT$ is one-way definable,
and by using some classical word combinatorics, we show that the output
produced between the anchor points of every inversion has period that divides
the lengths of $\out{\tr{\ell_1}}$ and $\out{\tr{\ell_2}}$. In particular, this
period is at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
The proposition below shows a slightly stronger periodicity property,
which refers to the output produced between the anchor points $\ell_1,\ell_2$
of an inversion, but extended on both sides with the words $\out{\tr{\ell_1}}$ and $\out{\tr{\ell_2}}$.
We will exploit this stronger periodicity property later, when dealing with
overlapping portions of the run delimited by different inversions
(cf.~Lemma \ref{lem:overlapping}).
\begin{prop}\label{prop:periodicity-sweeping}
If $\cT$ is one-way definable, then the following property \PR2 holds:
\begin{quote}
For all inversions
$(L_1,\ell_1,L_2,\ell_2)$ of $\rho$, the period $p$ of the word
\[
\out{\tr{\ell_1}} ~ \out{\rho[\ell_1,\ell_2]} ~ \out{\tr{\ell_2}}
\]
divides both $|\out{\tr{\ell_1}}|$ and
$|\out{\tr{\ell_2}}|$. Moreover, $p \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
\end{quote}
\end{prop}
The above proposition
thus formalizes the implication \PR1 $\Rightarrow$ \PR2 of
Theorem \ref{thm:main2}.
Its proof relies on a few combinatorial results.
The first one is Fine and Wilf's theorem~\cite{Lothaire97}.
In short, this theorem says that, whenever two periodic
words $w_1,w_2$ share a sufficiently long factor, then they
have the same period.
Here we use a slightly stronger variant of Fine and Wilf's theorem,
which additionally shows how to align a common factor of the two words $w_1,w_2$
so as to form a third word containing a prefix of $w_1$ and a suffix of $w_2$.
This variant of Fine-Wilf's theorem will be particularly useful in the proof of Lemma~\ref{lem:overlapping}, while for all other applications the classical
statement suffices.
\olivier{This last sentence is outdated, right?}%
\gabriele{This refers to the variant of Fine-Wilf, which still used in the referenced lemma.
But I rephrased so as to make it clear. Let me know if it makes more sense now.}%
\begin{thm}[Fine-Wilf's theorem]\label{thm:fine-wilf}
If $w_1 = w'_1\,w\:w''_1$ has period $p_1$,
$w_2 = w'_2\,w\,w''_2$ has period $p_2$, and
the common factor $w$ has length at least $p_1+p_2-\gcd(p_1,p_2)$,
then $w_1$, $w_2$, and $w_3 = w'_1\,w\,w''_2$ have period $\gcd(p_1,p_2)$.
\end{thm}
The second combinatorial result required in our proof concerns periods of words
with iterated factors, like those that arise from considering outputs of pumped runs,
and it is formalized precisely by the lemma below.
To improve readability, we often highlight the
important iterations of factors inside a word.
\begin{lem}\label{lem:periods}
Assume that $v_0 \: \pmb{v_1^n} \: v_2 \: \cdots \: v_{k-1} \: \pmb{v_k^n} \: v_{k+1}$
has period $p$ for some $n >p$.
Then $v_0 \: \pmb{{v_1}^{n_1}} \: v_2 \: \cdots \: v_{k-1} \: \pmb{{v_k}^{n_k}} \: v_{k+1}$
has period $p$ for all $n_1,\ldots,n_k \in \Nat$.
\end{lem}
\begin{proof}
Assume that
$w=v_0 \: \pmb{v_1^n} \: v_2 \: \cdots \: v_{k-1} \: \pmb{v_k^n} \: v_{k+1}$
has period $p$, and that $n >p$.
Consider an arbitrary factor $v_i^p$ of $w$. Since $v_i^p$ has periods $p$ and $|v_i|$,
it has also period $r=\gcd(p,|v_i|)$.
By Fine-Wilf (Theorem \ref{thm:fine-wilf}),
we know that $w$ has period $r$ as well.
Moreover, since the length of $v_i$ is multiple of $r$,
changing the number of repetitions of $v_i$ inside $w$ does
not affect the period $r$ of $w$.
Since $v_i$ was chosen arbitrarily, this means that, for all $n_1,\dots,n_k\in\bbN$,
$v_0 \: \pmb{{v_1}^{n_1}} \: v_2 \: \cdots \: v_{k-1} \: \pmb{{v_k}^{n_k}} \: v_{k+1}$
has period $r$, and hence period $p$ as well.
\end{proof}
Recall that our goal is to show that the output produced amid every
inversion has period bounded by $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
The general idea is to pump the loops of the inversion and compare the outputs
of the two-way transducer $\cT$ with those of an equivalent one-way
transducer $\cT'$.
The comparison leads to an equation between words with
iterated factors, where the iterations are parametrized by two unknowns
$n_1,n_2$ that occur in opposite order in the left, respectively right
hand-side of the equation.
Our third and last combinatorial result considers a word equation of
this precise form, and derives from it a periodicity property.
\reviewOne[inline]{In Corollary 4.8, the definition of $v^{(n_1,n_2)}$ is hard to process, especially when introduced in the middle of a proof. The notion is used later in the part: it warrants a proper definition and/or clear example either in the preliminaries or at the beginning of Part 4.
\felix[inline]{I don't understand this remark}
\olivier[inline]{answered in review1-answers.txt: the best place is here.}
}
\gabriele[inline]{This reviewer commend has led to a number of corrections, and the removal of Saarela tool}
For the sake of brevity, we use the notation $v^{(n_1,n_2)}$
to represent words with factors iterated $n_1$ or $n_2$ times,
namely, words of the form
$v_0 \: v_1^{n_{i_1}} \: v_2 \: \cdots \: v_{k-1} \: v_k^{n_{i_k}} \: v_{k+1}$,
where the $v_0,v_1,v_2,\dots,v_{k-1},v_k,v_{k+1}$ are fixed words (possibly empty)
and each index among $i_1,\dots,i_k$ is either $1$ or $2$.
\begin{lem}\label{lem:oneway-vs-twoway}
Consider a word equation of the form
\[
v_0^{(n_1,n_2)} \: \pmb{v_1^{n_1}} \: v_2^{(n_1,n_2)} \: \pmb{v_3^{n_2}} \: v_4^{(n_1,n_2)}
~=~
w_0 \: \pmb{w_1^{n_2}} \: w_2 \: \pmb{w_3^{n_1}} \: w_4
\]
where $n_1,n_2$ are the unknowns and $v_1,v_3$ are non-empty words.
If the above equation holds for all $n_1,n_2\in\bbN$,
then
\[
\pmb{v_1} ~ \pmb{v_1^{n_1}} ~ v_2^{(n_1,n_2)} ~ \pmb{v_3^{n_2}} ~ \pmb{v_3}
\]
has period $\gcd(|v_1|,|v_3|)$ for all $n_1,n_2\in\bbN$.
\end{lem}
\begin{proof}
The idea of the proof is to let the parameters $n_1,n_2$ of the equation
grow independently, and apply Fine and Wilf's theorem (Theorem \ref{thm:fine-wilf})
a certain number of times to establish periodicities in overlapping factors of the
considered words.
We begin by fixing $n_1$ large enough so that the factor
$\pmb{v_1^{n_1}}$ of the left hand-side of the equation becomes
longer than $|w_0|+|w_1|$ (this is possible because $v_1$ is non-empty).
Now, if we let $n_2$ grow arbitrarily large, we see that the length of
the periodic word $\pmb{w_1^{n_2}}$ is almost equal to the length of
the left hand-side term
$v_0^{(n_1,n_2)} ~ \pmb{v_1^{n_1}} ~ v_2^{(n_1,n_2)} ~ \pmb{v_3^{n_2}} ~ v_4^{(n_1,n_2)}$:
indeed, the difference in length is given by the constant
$|w_0| + |w_2| + n_1\cdot |w_3| + |w_4|$.
In particular, this implies that $\pmb{w_1^{n_2}}$
covers arbitrarily long prefixes of
$\pmb{v_1} ~ v_2^{(n_1,n_2)} ~ \pmb{v_3^{n_2+1}}$,
which in its turn contains long repetitions of the word $v_3$.
Hence, by Theorem \ref{thm:fine-wilf},
the word
$\pmb{v_1} ~ v_2^{(n_1,n_2)} ~ \pmb{v_3^{n_2+1}}$
has period $|v_3|$.
We remark that the periodicity shown so far holds for a large enough
$n_1$ and for all but finitely many $n_2$, where the threshold for
$n_2$ depends on $n_1$: once $n_1$ is fixed, $n_2$ needs to be larger
than $f(n_1)$, for a suitable function $f$.
In fact, by using
\gabriele{Saarela has been replaced by the new lemma here}
Lemma \ref{lem:periods},
with $n_1$ fixed and $n=n_2$ large enough,
we deduce that the periodicity holds for large enough $n_1\in\bbN$ and
for all $n_2\in\bbN$.
We could also apply a symmetric reasoning: we choose
$n_2$ large enough and let $n_1$ grow arbitrarily large. Doing so, we
prove that for a large enough $n_2$ and for all but finitely many $n_1$,
the word $\pmb{v_1^{n_1+1}} ~ v_2^{(n_1,n_2)} ~ \pmb{v_3}$ is periodic
with period $|v_1|$. As before, with the help of
\gabriele{same here for Saarela...}
Lemma \ref{lem:periods},
this can be strengthened to hold for large enough $n_2\in\bbN$ and for all $n_1\in\bbN$.
Putting together the results proven so far, we get that for all but finitely many $n_1,n_2$,
\[
\rightward{ \underbracket[0.5pt]{ \phantom{ \pmb{v_1^{n_1}} \cdot \pmb{v_1} ~\cdot~
v_2^{(n_1,n_2)} ~\cdot~ \pmb{v_3} } }%
_{\text{period } |v_1|} }
\pmb{v_1^{n_1}} \cdot
\overbracket[0.5pt]{ \pmb{v_1} ~\cdot~ v_2^{(n_1,n_2)} ~\cdot~
\pmb{v_3} \cdot \pmb{v_3^{n_2}} }%
^{\text{period } |v_3|}
.
\]
Finally, we observe that the prefix
$\pmb{v_1^{n_1+1}}\cdot v_2^{(n_1,n_2)}\cdot \pmb{v_3}$
and the suffix
$\pmb{v_1}\cdot v_2^{(n_1,n_2)}\cdot\pmb{v_3^{n_2+1}}$
share a common factor of length at least $|v_1|+|v_3|$.
By Theorem~\ref{thm:fine-wilf},
we derive that $\pmb{v_1^{n_1+1}}\cdot v_2^{(n_1,n_2)}\cdot
\pmb{v_3^{n_2+1}}$ has period $\gcd(|v_1|,|v_3|)$ for all but finitely
many $n_1,n_2$.
Finally, by using again
\gabriele{...and here}
Lemma \ref{lem:periods},
we conclude that the periodicity holds for all $n_1,n_2\in\bbN$.
\end{proof}
\medskip
We are now ready to prove the implication \PR1 $\Rightarrow$ \PR2:
\begin{proof}[Proof of Proposition~\ref{prop:periodicity-sweeping}]
Let $\cT'$ be a one-way transducer equivalent to $\cT$, and consider
an inversion $(L_1,\ell_1,L_2,\ell_2)$ of the successful run $\rho$ of $\cT$
on input $u$.
The reader may refer to Figure \ref{fig:inversion-sweeping}
to get basic intuition about the proof technique.
For simplicity, we assume that the loops $L_1$ and $L_2$ are disjoint,
as shown in the figure. If this were not the case, we would have
at least $\max(L_1) > \min(L_2)$, since the anchor point $\ell_1$
is strictly to the right of the anchor point $\ell_2$.
We could then consider the pumped run $\ensuremath{\mathsf{pump}}_{L_1}^k(\rho)$ for a
large enough $k>1$ in such a way that the rightmost copy of $L_1$
turns out to be disjoint from and strictly to the right of $L_2$.
We could thus reason as we do below, by replacing everywhere
(except in the final part of the proof, cf.~{\em Transferring periodicity to the original run})
the run $\rho$ with the pumped run $\ensuremath{\mathsf{pump}}_{L_1}^k(\rho)$, and the formal
parameter $m_1$ with $m_1+k$.
\smallskip\noindent
{\em Inducing loops in $T'$.}~
We begin by pumping the run $\rho$ and the underlying input $u$,
on the loops $L_1$ and $L_2$, in order to induce new loops $L'_1$
and $L'_2$ that are also loops in a successful run of $\cT'$.
Assuming that $L_1$ is strictly to the right of $L_2$, we define for all numbers $m_1,m_2\in\bbN$:
\[
\begin{array}{rcl}
u^{(m_1,m_2)} &=& \ensuremath{\mathsf{pump}}_{L_1}^{m_1+1}(\ensuremath{\mathsf{pump}}_{L_2}^{m_2+1}(u)) \\[1ex]
\rho^{(m_1,m_2)} &=& \ensuremath{\mathsf{pump}}_{L_1}^{m_1+1}(\ensuremath{\mathsf{pump}}_{L_2}^{m_2+1}(\rho)).
\end{array}
\]
In the pumped run $\rho^{(m_1,m_2)}$, we identify the positions that mark the
endpoints of the occurrences of $L_1,L_2$. More precisely, if $L_1=[x_1,x_2]$
and $L_2=[x_3,x_4]$, with $x_1>x_4$, then the sets of these positions are
\[
\begin{array}{rcl}
X_2^{(m_1,m_2)} &=& \big\{ x_3 + i(x_4-x_3) ~:~ 0\le i\le m_2+1 \big\} \\[1ex]
X_1^{(m_1,m_2)} &=& \big\{ x_1 + j(x_2-x_1) + m_2(x_4-x_3) ~:~ 0\le j\le m_1+1 \big\}.
\end{array}
\]
\smallskip\noindent
{\em Periodicity of outputs of pumped runs.}~
We use now the fact that $\cT'$ is a one-way transducer equivalent to $\cT$.
\reviewTwo[inline]{you fix $\lambda^{m_1,m_2}$ to be a particular run of Tβ on the input $u^{m_1,m_2}$, and then, on line 10, you state that for some particular $m_1$ and $m_2$, this run can be obtained by pumping $\lambda^{k_0,k_0}$. This does not seem true in general: since $T$ is not deterministic, we might just have chosen a different run.}%
\gabriele[inline]{Corrected}%
We first recall (see, for instance, \cite{eilenberg1974automata,berstel2013transductions})
that every functional one-way transducer can be made unambiguous, namely,
can be transformed into an equivalent one-way transducer that admits at
most one successful run on each input.
This means that, without loss of generality, we can assume that $\cT'$ too
is unambiguous, and hence it admits exactly one successful run, say
$\lambda^{(m_1,m_2)}$, on each input $u^{(m_1,m_2)}$.
Since $\cT'$ has finitely many states, we can find, for a large enough number $k_0$
two positions $x'_1<x'_2$, both in $X_1^{(k_0,k_0)}$, such that $L'_1=[x'_1,x'_2]$
is a loop of $\lambda^{(k_0,k_0)}$. Similarly, we can find two positions
$x'_3<x'_4$, both in $X_2^{(k_0,k_0)}$, such that $L'_2=[x'_3,x'_4]$ is a
loop of $\lambda^{(k_0,k_0)}$.
By construction $L'_1$ (resp.~$L'_2$) consists of $k_1\leq k_0$
(resp.~$k_2\leq k_0$) copies of $L_1$ (resp.~$L_2$), and
hence $L'_1,L'_2$ are also loops of $\rho^{(k_0,k_0)}$.
In particular, this implies that for all $n_1,n_2\in\bbN$:
\[
\begin{array}{rcl}
\ensuremath{\mathsf{pump}}_{L'_1}^{n_1+1}(\ensuremath{\mathsf{pump}}_{L'_2}^{n_2+1}(u^{(k_0,k_0)}))
&=& u^{(f(n_1),g(n_2))} \\[1ex]
\ensuremath{\mathsf{pump}}_{L'_1}^{n_1+1}(\ensuremath{\mathsf{pump}}_{L'_2}^{n_2+1}(\rho^{(k_0,k_0)}))
&=& \rho^{(f(n_1),g(n_2))} \\[1ex]
\ensuremath{\mathsf{pump}}_{L'_1}^{n_1+1}(\ensuremath{\mathsf{pump}}_{L'_2}^{n_2+1}(\lambda^{(k_0,k_0)}))
&=& \lambda^{(f(n_1),g(n_2))}.
\end{array}
\]
where $f(n_1)=k_1 n_1+k_0$ and $g(n_2)=k_2 n_2+k_0$.
Now recall that $\rho^{(f(n_1),g(n_2))}$ and $\lambda^{(f(n_1),g(n_2))}$ are
runs of $\cT$ and $\cT'$ on the same word $u^{(f(n_1),g(n_2))}$, and they
produce the same output. Let us denote this output by $w^{(f(n_1),g(n_2))}$.
Below, we show two possible factorizations of $w^{(f(n_1),g(n_2))}$
based on the shapes of the pumped runs $\lambda^{(f(n_1),g(n_2))}$
and $\rho^{(f(n_1),g(n_2))}$.
For the first factorization, we recall that $L'_2$ precedes $L'_1$,
according to the ordering of positions, and that the run
$\lambda^{(f(n_1),g(n_2))}$ is one-way (in particular loops have only one anchor point and one trace). We thus obtain:
\reviewTwo[inline]{eq. 4.3 The exponents should be $n_2 + 1$ and $n_1 + 1$. (same for eq. 4.4)}
\felix[inline]{Yes but not sure it makes sense to keep the +1 everywhere then}
\gabriele[inline]{We agreed to change the two itemized lists below and replace `right border'
by `left border' everywhere. I have implemented this...}
\begin{equation}\label{eq:one-way}
w^{(f(n_1),g(n_2))} ~=~ w_0 ~ \pmb{w_1^{n_2}} ~ w_2 ~ \pmb{w_3^{n_1}} ~ w_4
\end{equation}
where
\begin{itemize}
\item $w_0$ is the output produced by the prefix of $\lambda^{(k_0,k_0)}$
ending at the only anchor point of $L'_2$,
\item $\pmb{w_1}$ is the trace of $L'_2$,
\item $w_2$ is the output produced by the factor of $\lambda^{(k_0,k_0)}$
between the anchor points of $L'_2$ and $L'_1$,
\item $\pmb{w_3}$ is the trace of $L'_1$,
\item $w_4$ is the output produced by the suffix of
$\lambda^{(k_0,k_0)}$ starting at the anchor point of $L'_1$. Hence, $w_0 ~ w_2 ~ w_4$ is the output of $\lambda^{(k_0,k_0)}$.
\end{itemize}
\felix{added the last sentence to avoid the $+1$ or not confusion. In addition to the anchor point/trace usage that we talked about}
For the second factorization, we consider $L'_1$ and $L'_2$ as loops of $\rho^{(k_0,k_0)}$.
We recall that $\ell_1,\ell_2$ are anchor points of the loops $L_1,L_2$ of $\rho$, and that
there are corresponding copies of these anchor points in the pumped
run $\rho^{(f(n_1),g(n_2))}$.
We define $\ell'_1$ (resp.~$\ell'_2$) to be the first (resp.~last)
location in $\rho^{(f(n_1),g(n_2))}$
that corresponds to $\ell_1$ (resp.~$\ell_2$) and that is an anchor point of a copy of $L'_1$ (resp.~$L'_2$).
For example, if $\ell_1=(x_1,y_1)$, with $y_1$ even, then $\ell'_1=\big(x_1+f(n_2)(x_4-x_3),y_1\big)$.
Thanks to Equation \ref{eq:pumped-run} we know that the output produced by
$\rho^{(f(n_1),g(n_2))}$ is of the form
\begin{equation}\label{eq:two-way}
w^{(f(n_1),g(n_2))} ~=~
v_0^{(n_1,n_2)} ~ \pmb{v_1^{n_1}} ~ v_2^{(n_1,n_2)} ~ \pmb{v_3^{n_2}} ~ v_4^{(n_1,n_2)}
\end{equation}
where
\begin{itemize}
\item $\pmb{v_1}=\out{\tr{\ell'_1}}$, where $\ell'_1$ is seen as an anchor point in a copy of $L'_1$,
\item $\pmb{v_3}=\out{\tr{\ell'_2}}$, where $\ell'_2$ is seen as an anchor point in a copy of $L'_2$
\par\noindent
(note that the words $v_1,v_3$ depend on $k_0$, but not on $n_1,n_2$),
\item $v_0^{(n_1,n_2)}$ is the output produced by the prefix of $\rho^{(f(n_1),g(n_2))}$
that ends at $\ell_1'$
(this word may depend on the parameters $n_1,n_2$ since the loops
$L'_1,L'_2$ may be traversed several times before reaching the first occurrence of $\tr{\ell'_1}$),
\item $v_2^{(n_1,n_2)}$ is the output produced by the factor of $\rho^{(f(n_1),g(n_2))}$
that starts at $\ell'_1$
and ends at $\ell'_2$,
\item $v_4^{(n_1,n_2)}$ is the output produced by the suffix of $\rho^{(f(n_1),g(n_2))}$
that starts at $\ell'_2$.
\end{itemize}
Putting together Equations~(\ref{eq:one-way}) and (\ref{eq:two-way}), we get
\begin{equation}\label{eq:one-way-vs-two-way}
v_0^{(n_1,n_2)} ~ \pmb{v_1^{n_1}} ~ v_2^{(n_1,n_2)} ~ \pmb{v_3^{n_2}} ~ v_4^{(n_1,n_2)}
~~=~~
w_0 ~ \pmb{w_1^{n_2}} ~ w_2 ~ \pmb{w_3^{n_1}} ~ w_4 .
\end{equation}
Recall that the definition of inversions (Definition
\ref{def:inversion-sweeping}) states
that the words $v_1,v_3$ are non-empty.
This allows us to apply Lemma \ref{lem:oneway-vs-twoway}, which shows that the word
$\pmb{v_1} ~ \pmb{v_1^{n_1}} ~ v_2^{(n_1,n_2)} ~ \pmb{v_3^{n_2}} ~ \pmb{v_3}$
has period $p=\gcd(|v_1|,|v_3|)$, for all $n_1,n_2\in\bbN$.
Note that the latter period $p$ still depends on $\cT'$, since
the words $v_1,v_3$ were obtained from loops
$L'_1,L'_2$ on the run $\lambda^{(k_0,k_0)}$ of $\cT'$.
However, because each loop $L'_i$ consists of $k_i$ copies of the original loop $L_i$,
we also know that $v_1=(\out{\tr{\ell_1}})^{k_1}$ and $v_3=(\out{\tr{\ell_2}})^{k_2}$.
By Theorem \ref{thm:fine-wilf}, this implies that for all $n_1,n_2\in\bbN$, the word
\[
\big(\out{\tr{\ell_1}}\big) ~
\big(\out{\tr{\ell_1}}\big)^{k_1 n_1} ~
v_2^{(n_1,n_2)} ~
\big(\out{\tr{\ell_2}}\big)^{k_2 n_2} ~
\big(\out{\tr{\ell_2}}\big)
\]
has a period that divides $|\out{\tr{\ell_1}}|$ and $|\out{\tr{\ell_2}}|$.
\smallskip\noindent
{\em Transferring periodicity to the original run.}~
The last part of the proof amounts at showing a similar periodicity property
for the output produced by the original run $\rho$.
By construction, the iterated factors inside $v_2^{(n_1,n_2)}$ in the previous
word are all of the form $v^{k_1 n_1 + k_0}$ or $v^{k_2 n_2 + k_0}$,
for some words $v$. By taking out the constant factors $v^{k_0}$ from the
latter repetitions, we can write $v_2^{(n_1,n_2)}$ as a word with iterated
factors of the form $v^{k_1 n_1}$ or $v^{k_2 n_2}$, namely, as ${v'_2}^{(k_1 n_1, k_2 n_2)}$.
So the word
\[
\big(\out{\tr{\ell_1}}\big) ~
\big(\out{\tr{\ell_1}}\big)^{k'_1} ~
{v'_2}^{(k'_1, k'_2)} ~
\big(\out{\tr{\ell_2}}\big)^{k'_2} ~
\big(\out{\tr{\ell_2}}\big)
\]
is periodic, with period that divides $|\out{\tr{\ell_1}}|$ and $|\out{\tr{\ell_2}}|$,
for all $k'_1 \in \{k_1 n \::\: n\in\bbN\}$ and all $k'_2 \in \{k_2 n \::\: n\in\bbN\}$.
We now apply
\gabriele{Also here Saarela has been replaced by the new lemma}
Lemma \ref{lem:periods},
once with $n=k'_1$ and once with $n=k'_2$,
to conclude that the latter periodicity property holds also for $k'_1=1$ and $k'_2=1$.
This shows that the word
\[
\out{\tr{\ell_1}} ~ \out{\rho[\ell_1,\ell_2]} ~ \out{\tr{\ell_2}}
\]
is periodic, with period that divides $|\out{\tr{\ell_1}}|$ and $|\out{\tr{\ell_2}}|$.
\end{proof}
\section{Combinatorics in the two-way case}\label{sec:combinatorics-twoway}
In this section we develop the main combinatorial techniques required
in the general case.
In particular, we will show how to derive the existence of idempotent
loops with bounded outputs using Ramsey-based arguments, and we will
use this to derive periodicity properties for the outputs produced
between inversions.
As usual, $\rho$ is a fixed successful run of $\cT$ on some input word $u$.
\medskip
\subsection*{Ramsey-type arguments.}
We start with a technique used for bounding the lengths of the outputs
of certain factors, or subsequences of a two-way run. This technique
is a Ramsey-type argument, more precisely it relies on
Simon's ``factorization forest'' theorem~\cite{factorization_forests,factorization_forests_for_words_paper},
which is recalled below. The classical version of Ramsey theorem would
yield a similar result, but without the tight bounds that we get here.
Let $X$ be a set of positions of $\rho$.
A \emph{factorization forest} for $X$ is an unranked tree, where the nodes are
intervals $I$ with endpoints in $X$ and labeled with the corresponding effect $E_I$,
the ancestor relation is given by the containment order on intervals, the leaves are the
minimal intervals $[x_1,x_2]$, with $x_2$ successor of $x_1$ in $X$, and for every
internal node $I$ with children $J_1,\dots,J_k$, we have:
\begin{itemize}
\item $I=J_1\cup\dots\cup J_k$,
\item $E_I = E_{J_1}\odot\dots\odot E_{J_k}$,
\item if $k>2$, then $E_I = E_{J_1} = \dots = E_{J_k}$
is an idempotent of the semigroup $(\cE,\odot)$.
\end{itemize}
Recall that in a normalized run there are at most $|Q|^{{\boldsymbol{H}}} %h_{\mathsf{max}}}}$ distinct
crossing sequences. Moreover, a flow contains at most ${\boldsymbol{H}}} %h_{\mathsf{max}}}$ edges,
and each edge has one of the 4 possible types ${\ensuremath{\mathsf{LL}}},{\ensuremath{\mathsf{LR}}},{\ensuremath{\mathsf{RL}}},\RR$, so they are at most $4^{{\boldsymbol{H}}} %h_{\mathsf{max}}}}$ different flows.
Hence, the effect semigroup $(\cE,\odot)$ has size at most
${\boldsymbol{E}}} %e_{\mathsf{max}}}=4^{{\boldsymbol{H}}} %h_{\mathsf{max}}}}\cdot|Q|^{2{\boldsymbol{H}}} %h_{\mathsf{max}}}}\cdot|Q|^{2{\boldsymbol{H}}} %h_{\mathsf{max}}}} =(2|Q|)^{2{\boldsymbol{H}}} %h_{\mathsf{max}}}}$.
\reviewTwo[inline]{the way you bound the number of flows could be explained more precisely.
Here you just state that there is at most H edges, and that each edge has one of the possible 4 types.
Then, you bound the number of flows by $4^{H}$, which seems to correspond to any choice of types for the H edges.
However, as you state, there is at most H edges, but there could be less.}%
\felix{done}%
Further recall that ${\boldsymbol{C}}} %c_{\mathsf{max}}}$ is the maximum number of letters output by a
single transition of $\cT$.
Like we did in the sweeping case, we define the constant
${\boldsymbol{B = {\boldsymbol{C}}} %c_{\mathsf{max}}} \cdot {\boldsymbol{H}}} %h_{\mathsf{max}}} \cdot (2^{3{\boldsymbol{E}}} %e_{\mathsf{max}}}}+4) + 4{\boldsymbol{C}}} %c_{\mathsf{max}}}}}$
that will be used to bound the lengths of some outputs of $\cT$.
Note that now $\boldsymbol{B}$ is doubly exponential with respect
to the size of $\cT$, due to the size of the effect semigroup.
\felix{added this explanation, it could be misleading otherwise}%
\begin{thm}[Factorization forest theorem \cite{factorization_forests_for_words_paper,factorization_forests}]%
\label{th:simon}
For every set $X$ of positions of $\rho$, there is a factorization forest for $X$
of height at most $3{\boldsymbol{E}}} %e_{\mathsf{max}}}$.
\end{thm}
\begin{wrapfigure}{r}{5cm}
\vspace{-4mm}
\input{ramsey}
\end{wrapfigure}
The above theorem can be used to show that if $\rho$
produces an output longer than $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$, then it contains an idempotent
loop and a trace with non-empty output. Below, we present a result
in the same spirit, but refined in a way that it can be used to
find anchor points inside specific intervals.
To formally state the result, we consider subsequences
of $\r$ induced by sets of locations that are not necessarily
contiguous.
Recall the notation $\rho|Z$ introduced on page~\pageref{rhoZ}:
$\rho|Z$ is the subsequence of $\rho$ induced by the location set $Z$.
For example, Figure \ref{fig:ramsey} depicts a set $Z=[\ell_1,\ell_2]\cap (I\times\bbN)$
by a hatched area, together with the induced subrun $\rho|Z$, represented by
thick arrows.
\smallskip
\begin{thm}\label{thm:simon2}
Let $I=[x_1,x_2]$ be an interval of positions, $K=[\ell_1,\ell_2]$
an interval of locations, and $Z = K \:\cap\: (I\times\bbN)$.
If $|\out{\rho|Z}| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
then there exists an idempotent loop $L$ and an anchor point $\ell$ of $L$ such that
\begin{enumerate}
\item $x_1 < \min(L) < \max(L) < x_2$ (in particular, $L\subsetneq I$),
\item $\ell_1 \mathrel{\lhd} \ell \mathrel{\lhd} \ell_2$ (in particular, $\ell \in K$),
\item $\out{\tr{\ell}} \neq \emptystr$.
\end{enumerate}
\end{thm}
\begin{proof}
Let $I$, $K$, $Z$ be as in the statement,
and suppose that $\big|\out{\rho| Z}\big| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
We define
$Z' = Z ~\setminus~ (\{\ell_1,\ell_2\} \cup \{x_1,x_2\}\times\bbN)$
and we observe that there are at most $2{\boldsymbol{H}}} %h_{\mathsf{max}}} +2$ locations
that are missing from $Z'$. This means that $\rho| Z'$ contains
all but $4{\boldsymbol{H}}} %h_{\mathsf{max}}}+4$
transitions of $\rho|Z$,
and because each transition outputs at most ${\boldsymbol{C}}} %c_{\mathsf{max}}}$ letters, we have
$|\out{\rho| Z'}| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax} - 4{\boldsymbol{C}}} %c_{\mathsf{max}}}\cdot{\boldsymbol{H}}} %h_{\mathsf{max}}} -4{\boldsymbol{C}}} %c_{\mathsf{max}}} = {\boldsymbol{C}}} %c_{\mathsf{max}}}\cdot{\boldsymbol{H}}} %h_{\mathsf{max}}}\cdot 2^{3{\boldsymbol{E}}} %e_{\mathsf{max}}}}$.
\input{ramsey-proof}
For every level $y$, let $X_y$ be the set of positions $x$ such that
$(x,y)$ is the source location of some transition of $\rho|Z'$
that produces non-empty output.
For example, if we refer to Figure~\ref{fig:ramsey-proof},
the vertical dashed lines represent the positions
of $X_y$ for a particular level $y$; accordingly, the circles
in the figure represent the locations of the form $(x,y)$, for
all $x\in X_y$.
Since each transition outputs at most ${\boldsymbol{C}}} %c_{\mathsf{max}}}$ letters,
we have $\sum_y |X_y| > {\boldsymbol{H}}} %h_{\mathsf{max}}}\cdot 2^{3{\boldsymbol{E}}} %e_{\mathsf{max}}}}$.
Moreover, since there are at most ${\boldsymbol{H}}} %h_{\mathsf{max}}}$ levels,
there is a level $y$ (which we fix hereafter) such that $|X_y| > 2^{3{\boldsymbol{E}}} %e_{\mathsf{max}}}}$.
We now prove the following:
\begin{clm}
There are two consecutive loops $L_1=[x,x']$ and $L_2=[x',x'']$
with endpoints $x,x',x''\in X_y$ and such that $E_{L_1}=E_{L_2}=E_{L_1\cup L_2}$.
\end{clm}
\begin{proof}
By Theorem \ref{th:simon},
there is a factorization forest for $X_y$ of height at most $3{\boldsymbol{E}}} %e_{\mathsf{max}}}$.
Since $\rho$ is a valid run, the dummy element $\bot$ of the effect
semigroup does not appear in this factorization forest.
Moreover, since $|X_y| > 2^{3{\boldsymbol{E}}} %e_{\mathsf{max}}}}$, we know that the factorization
forest contains an internal node $L'=[x'_1,x'_{k+1}]$ with $k > 2$ children,
say $L_1=[x'_1,x'_2], \dots, L_k=[x'_k,x'_{k+1}]$.
By definition of factorization forest, the effects
$E_{L'}$, $E_{L_1}$, \dots, $E_{L_k}$ are all equal and idempotent.
In particular, the effect $E_{L'}=E_{L_1}=\dots=E_{L_k}$
is a triple of the form $(F_{L'},c_1,c_2)$,
where $c_i=\rho|x_i$ is the crossing sequence at $x'_i$.
Finally, since $E_{L'}$ is idempotent, we have that $c_1 = c_2$ and
this is equal to the crossing sequences of $\rho$ at the positions
$x'_1,\dots,x'_{k+1}$. This shows that $L_1,L_2$ are idempotent loops.
\end{proof}
Turning back to the proof of the theorem, we know from the above claim that
there are two consecutive idempotent loops $L_1=[x,x']$ and $L_2=[x',x'']$ with the
same effect and with endpoints $x,x',x''\in X_y \subseteq I \:\setminus\: \{x_1,x_2\}$
(see again Figure~\ref{fig:ramsey-proof}).
Let $\tilde\ell_1=(x,y)$ and $\tilde\ell_2=(x'',y)$, and observe that
$\tilde\ell_1,\tilde\ell_2\in Z'$. In particular, $\tilde\ell_1$ and $\tilde\ell_2$
are strictly between $\ell_1$ and $\ell_2$.
Suppose by symmetry that $\tilde\ell_1 \mathrel{\unlhd} \tilde\ell_2$.
Further let $C$ be the component of $L_1\cup L_2$ (or, equally, of $L_1$ or $L_2$)
that contains the node $y$.
Below, we focus on the factors of $\rho[\tilde\ell_1,\tilde\ell_2]$
that are intercepted by $L_1\cup L_2$: these are represented in
Figure~\ref{fig:ramsey-proof} by the thick arrows.
By Lemma~\ref{lem:component2} all these factors correspond to edges
of the same component $C$, namely, they are $(L_1 \cup L_2,C)$-factors.
Let us fix an arbitrary factor $\alpha$ of $\rho[\tilde\ell_1,\tilde\ell_2]$
that is intercepted by $L_1\cup L_2$, and assume that $\alpha=\beta_1 \cdots \beta_k$,
where $\beta_1,\dots,\beta_k$ are the factors intercepted by either $L_1$ or $L_2$.
\begin{clm}
If $\beta,\beta'$ are two factors intercepted by $L_1=[x,x']$ and $L_2=[x',x'']$,
with $E_{L_1}=E_{L_2}=E_{L_1\cup L_2}$, and $\beta,\beta'$ are adjacent in the run
$\rho$ (namely, they share an endpoint at position $x'$), then $\beta,\beta'$
correspond to edges in the same component of $L_1$ (or, equally, $L_2$).
\end{clm}
\begin{proof}
Let $C$ be the component of $L_1$ and $y_1 \mapsto y_2$ the edge of $C$
that corresponds to the factor $\beta$ intercepted by $L_1$.
Similarly, let $C'$ be the component of $L_2$ and $y_3 \mapsto y_4$ the edge of $C'$
that corresponds to the factor $\beta'$ intercepted by $L_2$.
Since $\beta$ and $\beta'$ share an endpoint at position $x'$,
we know that $y_2=y_3$. This shows that $C \cap C' \neq \emptyset$,
and hence $C=C'$.
\end{proof}
The above claim shows that any two adjacent factors $\beta_i,\beta_{i+1}$
correspond to edges in the same component of $L_1$ and $L_2$, respectively.
Thus, by transitivity, all factors $\beta_1,\dots,\beta_k$ correspond to
edges in the same component, say $C'$.
We claim that $C'=C$. Indeed, if $\beta_1$ is intercepted by $L_1$,
then $C'=C$ because $\alpha$ and $\beta_1$ start from the same location
and hence they correspond to edges of the flow that depart from the
same node. The other case is where $\beta_1$ is intercepted by $L_2$,
for which a symmetric argument can be applied.
So far we have shown that every factor of $\rho[\tilde\ell_1,\tilde\ell_2]$ intercepted
by $L_1\cup L_2$ can be factorized into some $(L_1,C)$-factors and some $(L_2,C)$-factors.
We conclude the proof with the following observations:
\begin{itemize}
\item By construction, both loops $L_1,L_2$ are contained in the interval of positions $I=[x_1,x_2]$,
and have endpoints different from $x_1,x_2$.
\item Both anchor points of $C$ inside $L_1,L_2$ belong to the interval of locations
$K\:\setminus\:\{\ell_1,\ell_2\}$.
This holds because
$\rho[\tilde\ell_1,\tilde\ell_2]$ contains a factor $\alpha$ that is intercepted
by $L_1\cup L_2$ and spans across all the positions from $x$ to $x''$, namely,
an ${\ensuremath{\mathsf{LR}}}$-factor.
This factor starts at the anchor point of $C$ inside $L_1$
and visits the anchor point of $C$ inside $L_2$.
Moreover, by construction, $\alpha$ is also a factor of the subsequence $\rho|Z'$.
This shows that the anchor points of $C$ inside $L_1$ and $L_2$ belong to $Z'$,
and in particular to $K\:\setminus\:\{\ell_1,\ell_2\}$.
\item The first factor of $\rho[\tilde\ell_1,\tilde\ell_2]$ that is intercepted by
$L_1\cup L_2$ starts at $\tilde\ell_1=(x,y)$, which by construction is the
source location of some transition producing non-empty output.
By the previous arguments, this factor is a concatenation of $(L_1,C)$-factors
and $(L_2,C)$-factors. This implies that the trace of the anchor point
of $C$ inside $L_1$, or the trace of $C$ inside $L_2$ produces non-empty output.
\qedhere
\end{itemize}
\end{proof}
\medskip
\subsection*{Inversions and periodicity.}
The first important notion that is used to characterize one-way definability
is that of inversion. It turns out that
the
definition of inversion in the sweeping case (see
page~\pageref{page-def-inversion}) can be reused almost verbatim
in the two-way setting. The only difference is that here we require the loops
to be idempotent and we do not enforce output-minimality (we will discuss this
latter choice further below, with a formal definition of output-minimality at hand).
\begin{defi}\label{def:inversion-twoway}
An \emph{inversion} of the run $\rho$ is a tuple $(L_1,\ell_1,L_2,\ell_2)$ such that
\begin{enumerate}
\item $L_1,L_2$ are idempotent loops,
\item $\ell_1=(x_1,y_1)$ and $\ell_2=(x_2,y_2)$
are anchor points inside $L_1$ and $L_2$, respectively,
\item $\ell_1 \mathrel{\lhd} \ell_2$ and $x_1 > x_2$,
\item for both $i=1$ and $i=2$, $\out{\tr{\ell_i}}\neq\emptystr$.
\end{enumerate}
\end{defi}
\input{inversion-twoway}
\noindent
Figure \ref{fig:inversion-twoway} gives an example of an inversion involving
the idempotent loop $L_1$ with anchor point $\ell_1$, and the idempotent
loop $L_2$ with anchor point $\ell_2$. The intercepted factors that form
the corresponding traces are represented by thick arrows; the one highlighted in red
are those that produce non-empty output.
The implication \PR1 $\Rightarrow$ \PR2 of Theorem \ref{thm:main2} in the two-way
case is formalized below exactly as in Proposition~\ref{prop:periodicity-sweeping},
and the proof is very similar to the sweeping case.
More precisely, it can be checked that the proof of the first claim in
Proposition~\ref{prop:periodicity-sweeping} was shown independently of the sweeping assumption
--- one just needs to replace the use of Equation \ref{eq:pumped-run}
with Proposition \ref{prop:pumping-twoway}.
The sweeping assumption was used only for deriving the notion of \emph{output-minimal}
factor, which was crucial to conclude that the period $p$ is bounded by the specific
constant $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
In this respect, the proof of Proposition~\ref{prop:periodicity-twoway}
requires a different argument for showing that $p \le\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$:
\begin{prop}\label{prop:periodicity-twoway}
If $\cT$ is one-way definable, then the following property \PR2 holds:
\begin{quote}
For all inversions $(L_1,\ell_1,L_2,\ell_2)$ of $\rho$,
the period $p$ of the word
\[
\out{\tr{\ell_1}} ~ \out{\rho[\ell_1,\ell_2]} ~ \out{\tr{\ell_2}}
\]
divides both $|\out{\tr{\ell_1}}|$ and $|\out{\tr{\ell_2}}|$.
Moreover, $p \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
\end{quote}
\end{prop}
We only need to show here that $p \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$. Recall that in the
sweeping case we relied on the assumption that the factors
$\tr{\ell_1}$ and $\tr{\ell_2}$ of an inversion are output-minimal,
and on Lemma~\ref{lem:output-minimal-sweeping}. In the general case
we need to replace output-minimality by the following notion:
\begin{defi}\label{def:output-minimal-twoway}
Consider pairs $(L,C)$ consisting of an idempotent loop $L$
and a component $C$ of $L$.
\begin{enumerate}
\item On such pairs, define the relation
$\sqsubset$ by $(L',C') \sqsubset (L,C)$ if
$L'\subsetneq L$ and at least one $(L',C')$-factor
is contained in some $(L,C)$-factor.
\item A pair $(L,C)$ is \emph{output-minimal} if
$(L',C') \sqsubset (L,C)$ implies $\out{\tr{\an{C'}}}=\emptystr$.
\end{enumerate}
\end{defi}
\noindent
Note that the relation $\sqsubset$ is not a partial order in general
(it is however antisymmetric).
Moreover, it is easy to see that the notion of output-minimal
pair $(L,C)$ generalizes that of output-minimal factor introduced
in the sweeping case: indeed, if $\ell$ is the anchor point of a
loop $L$ of a sweeping transducer and $\tr{\ell}$ satisfies
Definition \ref{def:output-minimal-sweeping}, then the pair
$(L,C)$ is output-minimal, where $C$ is the unique component
whose edge corresponds to $\tr{\ell}$.
The following lemma bounds the length of the output trace
$\out{\tr{\an{C}}}$ for an output-minimal pair $(L,C)$:
\begin{lem}\label{lem:output-minimal-twoway}
For every output-minimal pair $(L,C)$, $|\out{\tr{\an{C}}}| \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
\end{lem}
\begin{proof}
Consider a pair $(L,C)$ consisting of an idempotent loop $L=[x_1,x_2]$ and a component $C$ of $L$.
Suppose by contradiction that $|\out{\tr{\an{C}}}|>\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
We will show that $(L,C)$ is not output-minimal.
Recall that $\tr{\an{C}}$ is a concatenation of $(L,C)$-factors, say,
$\tr{\an{C}}=\beta_1\cdots\beta_k$. Let $\ell_1$ (resp.~$\ell_2$) be the
first (resp.~last) location that is visited by these factors. Further let
$K = [\ell_1,\ell_2]$ and $Z = K \:\cap\: (L\times\bbN)$.
By construction, the subsequence $\rho|Z$ can be seen as a concatenation
of the factors $\beta_1,\dots,\beta_k$, possibly in a different order than
that of $\tr{\an{C}}$. This implies that $|\out{\rho|Z}| > \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
By Theorem \ref{thm:simon2}, we know that there exist an idempotent
loop $L'\subsetneq L$ and a component $C'$ of $L'$ such that
$\an{C'} \in K$ and $\out{\tr{\an{C'}}}\neq\emptystr$.
Note that the $(L',C')$-factor that starts at the anchor point
$\an{C'}$ (an ${\ensuremath{\mathsf{LR}}}$- or ${\ensuremath{\mathsf{RL}}}$-factor) is entirely contained
in some $(L,C)$-factor.
This implies that $(L',C') \sqsubset (L,C)$, and thus $(L,C)$ is not output-minimal.
\end{proof}
We remark that the above lemma cannot be used directly to bound the period
of the output produced amid an inversion. The reason is that we cannot
restrict ourselves to inversions $(L_1,\ell_1,L_2,\ell_2)$ that induce
output-minimal pairs $(L_i,C_i)$ for $i=1,2$, where $C_i$ is the unique
component of the anchor point $\ell_i$.
An example is given in Figure~\ref{fig:inversion-twoway},
assuming that the factors depicted in red are the only ones that
produce non-empty output, and the lengths of these outputs exceed $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
On the one hand $(L_1,\ell_1,L_2,\ell_2)$ is an inversion, but
$(L_1,C_1)$ is not output-minimal.
On the other hand, it is possible that $\rho$ contains no other
inversion than $(L_1,\ell_1,L_2,\ell_2)$: any loop strictly contained
in the red factor in $L_1$ will have the anchor point \emph{after}
$\ell_2$.
We are now ready to show the second claim of
Proposition~\ref{prop:periodicity-twoway}.
\input{non-output-minimal-inversion-full}
\begin{proof}[Proof of Proposition~\ref{prop:periodicity-twoway}]
The proof of the second claim requires a refinement of the arguments that involve
pumping the run $\rho$ simultaneously on three different loops. As usual, we
assume that the loops $L_1,L_2$ of the inversion are disjoint (otherwise,
we preliminarily pump one of the two loops a few times).
Recall that the word
\[
\out{\tr{\ell_1}} ~ \out{\rho[\ell_1,\ell_2]} ~ \out{\tr{\ell_2}}
\]
has period $p = \gcd\big( |\out{\tr{\ell_1}}|, |\out{\tr{\ell_2}}|
\big)$, but that we cannot bound $p$ by assuming that $(L_1,\ell_1,L_2,\ell_2)$
is output-minimal.
However, in the pumped run $\rho^{(2,1)}$ we do find inversions with
output-minimal pairs.
For example, as depicted in the right part of Figure~\ref{fig:non-output-minimal-inversion-full},
we can consider the left and right copy of $L_1$ in $\rho^{(2,1)}$,
denoted by $\mathpalette{\overarrow@\shortleftarrowfill@} L_1$ and $\mathpalette{\overarrow@\shortrightarrowfill@} L_1$, respectively.
Accordingly, we denote by $\mathpalette{\overarrow@\shortleftarrowfill@}\ell_1$ and $\mathpalette{\overarrow@\shortrightarrowfill@}\ell_1$ the left and
right copy of $\ell_1$ in $\rho^{(2,1)}$.
Now, let $(L_0,C_0)$ be any {\sl output-minimal} pair such that $L_0$
is an idempotent loop, $\out{\tr{\an{C_0}}}\neq\emptystr$, and either
$(L_0,C_0)=(\mathpalette{\overarrow@\shortleftarrowfill@} L_1,C_1)$ or $(L_0,C_0) \sqsubset (\mathpalette{\overarrow@\shortleftarrowfill@} L_1,C_1)$.
Such a loop $L_0$ is represented in
Figure~\ref{fig:non-output-minimal-inversion-full} by
the red vertical stripe. Further let $\ell_0=\an{C_0}$.
We claim that either $(L_0,\ell_0,L_2,\ell_2)$ or $(\mathpalette{\overarrow@\shortrightarrowfill@} L_1,\mathpalette{\overarrow@\shortrightarrowfill@}\ell_1,L_0,\ell_0)$
is an inversion of the run $\rho^{(2,1)}$, depending on whether $\ell_0$ occurs
before or after $\ell_2$.
First, note that all the loops $L_0$, $L_2$, $\mathpalette{\overarrow@\shortrightarrowfill@} L_1$ are idempotent
and non-overlapping;
more precisely, we have
$\max(L_2) \le \min(L_0)$ and $\max(L_0) \le \min(\mathpalette{\overarrow@\shortrightarrowfill@} L_1)$.
Moreover, the outputs of the traces $\tr{\ell_0}$, $\tr{\mathpalette{\overarrow@\shortrightarrowfill@}\ell_1}$,
and $\tr{\ell_2}$ are all non-empty.
So it remains to distinguish two cases based on the ordering of the anchor
points $\ell_0$, $\mathpalette{\overarrow@\shortrightarrowfill@}\ell_1$, $\ell_2$.
If $\ell_0 \mathrel{\lhd} \ell_2$, then $(L_0,\ell_0,L_2,\ell_2)$ is an inversion.
Otherwise, because $(\mathpalette{\overarrow@\shortrightarrowfill@} L_1,\mathpalette{\overarrow@\shortrightarrowfill@}\ell_1,L_2,\ell_2)$ is an inversion, we know that
$\mathpalette{\overarrow@\shortrightarrowfill@}\ell_1 \mathrel{\lhd} \ell_2 \mathrel{\unlhd} \ell_0$, and hence $(\mathpalette{\overarrow@\shortrightarrowfill@} L_1,\mathpalette{\overarrow@\shortrightarrowfill@}\ell_1,L_0,C_0)$
is an inversion.
Now, we know that
$\rho^{(2,1)}$ contains the inversion $(\mathpalette{\overarrow@\shortrightarrowfill@} L_1,\mathpalette{\overarrow@\shortrightarrowfill@} \ell_1,L_2,\ell_2)$, but also
an inversion involving the output-minimal pair $(L_0,C_0)$, with $L_0$ strictly
between $\mathpalette{\overarrow@\shortrightarrowfill@} L_1$ and $L_2$.
For all $m_0,m_1,m_2$, we define $\rho^{(m_0,m_1,m_2)}$ as the run obtained from
$\rho^{(2,1)}$ by pumping $m_0,m_1,m_2$ times the loops $L_0,\mathpalette{\overarrow@\shortrightarrowfill@} L_1,L_2$, respectively.
By reasoning as we did in the proof of Proposition~\ref{prop:periodicity-sweeping}
(cf.~{\em Periodicity of outputs of pumped runs}), one can show that there are
arbitrarily large output factors of $\rho^{(m_0,m_1,m_2)}$ that are produced
within the inversion on $\ell_0$ (i.e.~either $(L_0,\ell_0,L_2,\ell_2)$ or $(\mathpalette{\overarrow@\shortrightarrowfill@} L_1,\mathpalette{\overarrow@\shortrightarrowfill@}\ell_1,L_0,\ell_0)$)
and that are periodic with period $p'$ that divides $|\out{\tr{\ell_0}}|$.
In particular, by Lemma~\ref{lem:output-minimal-twoway}, we know that
$p' \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
Moreover, large portions of these factors are also produced within the
inversion $(\mathpalette{\overarrow@\shortrightarrowfill@} L_1,\mathpalette{\overarrow@\shortrightarrowfill@}\ell_1,L_2,\ell_2)$, and hence
by Theorem~\ref{thm:fine-wilf} they have period $\gcd(p,p')$.
To conclude the proof we need to transfer the periodicity property
from the pumped runs $\rho^{(m_0,m_1,m_2)}$ to the original run $\rho$.
This is done exactly like in Proposition~\ref{prop:periodicity-sweeping}
by relying on
\gabriele{Usual replacement of Saarela...}
Lemma \ref{lem:periods}:
we observe that the periodicity
property holds for large enough parameters $m_0,m_1,m_2$, hence for
all values of the parameters, and in particular
for $m_0 = m_1 = m_2 = 1$. This shows that the word
\[
\out{\tr{\ell_1}} ~ \out{\rho[\ell_1,\ell_2]} ~ \out{\tr{\ell_2}}
\]
has period $\gcd(p,p') \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
\end{proof}
So far we have shown that the output produced amid every inversion of a
run of a one-way definable two-way transducer is periodic, with period
bounded by $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ and dividing the lengths of the trace outputs of
the inversion. This basically proves the implication \PR1 $\Rightarrow$ \PR2
of Theorem \ref{thm:main2}.
In the next section we will follow a line of arguments similar to that of
Section \ref{sec:characterization-sweeping} to prove the remaining
implications \PR2 $\Rightarrow$ \PR3 $\Rightarrow$ \PR1.
\section{Complexity of the one-way definability problem}\label{sec:complexity}
In this section we analyze the complexity of the problem of deciding whether
a transducer $\cT$ is one-way definable. We begin with the case of a functional
two-way transducer. In this case, thanks to the results presented in
Section~\ref{sec:characterization-twoway} page \pageref{testing-containment},
we know that $\cT$ is one-way
definable if and only if $\dom(\cT) \subseteq D$, where $D$ is the language of words
$u\in\dom(\cT)$ such that all successful runs of $\cT$ on $u$ admit a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition.
In particular, the one-way definability problem reduces to an emptiness problem
for the intersection of two languages:
\[
\cT \text{ one-way definable}
\qquad\text{if and only if}\qquad
\dom(\cT) \cap D^\complement = \emptyset.
\]
The following lemma exploits the characterization of Theorem~\ref{thm:main2}
to show that the language $D^\complement$ can be recognized by a non-deterministic
finite automaton $\cA$ of triply exponential size w.r.t.~$\cT$. In fact,
this lemma shows that the automaton recognizing $D^\complement$ can be constructed
using doubly exponential {\sl workspace}. As before, we gain an exponent when
restricting to sweeping transducers.
\reviewOne[inline]{Lemma 9.1: D is a poor choice of notation for the automaton given that there already is a language D. Might A (same font as current D) be an adequate choice?
\gabriele[inline]{Done. BTW. if anyone has changed the macros \cA, \cT, etc. mind that
now they are rendered as simple capital letters instead as with mathcal.
I like this, and it is more uniform than using caligraphic A for automata
and normal T for transducers... But we should check we are consistent!}%
}%
\begin{lem}\label{lem:D-complement}
Given a functional two-way transducer $\cT$, an NFA $\cA$ recognizing
$D^\complement$ can be constructed in $2\expspace$.
Moreover, when $\cT$ is sweeping, the
NFA $\cA$ can be constructed in $\expspace$.
\end{lem}
\begin{proof}
Consider an input word $u$. By Theorem~\ref{thm:main2} we know that
$u\in D^\complement$ iff there exist a successful run $\rho$ of $\cT$
on $u$ and an inversion $\cI=(L_1,\ell_1,L_2,\ell_2)$ of $\rho$ such that
no positive number $p \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ is a period of the word
\[
w_{\rho,\cI} ~=~
\outb{\tr{\ell_1}} ~ \outb{\rho[\ell_1,\ell_2]} ~ \outb{\tr{\ell_2}}.
\]
The latter condition on $w_{\rho,\cI}$ can be rephrased as follows:
there is a function $f:\{1,\dots,\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}\} \rightarrow \{1,\dots,|w_{\rho,\cI}|\}$
such that $w_{\rho,\cI}\big(f(p)\big) \neq w_{\rho,\cI}\big(f(p)+p\big)$
for all positive numbers $p\le\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
In particular, each of the images of the latter function $f$, that is,
$f(1),\dots,f(\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax})$, can be encoded by a suitable marking of the
crossing sequences of $\rho$. This shows that the run $\rho$, the
inversion $\cI$, and the function $f$ described above can all be
guessed within space $\cO(\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax})$: $\r$ is guessed on-the-fly, the
inversion is guessed by marking the anchor points, and for $f$ we only
store two symbols and a counter $\le\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$, for each $1 \le i \le \boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$.
That is, any state of $\cA$ requires doubly
exponential space, resp.~simply exponential space, depending on whether $\cT$ is arbitrary
two-way or sweeping.
\end{proof}
As a consequence of the previous lemma, the emptiness problem for the language
$\dom(\cT) \cap D^\complement$, and thus the one-way definability problem for $\cT$,
can be decided in $2\expspace$ or $\expspace$, depending on whether $\cT$ is
two-way or sweeping:
\begin{prop}\label{prop:complexity}
The problem of deciding whether a functional two-way transducer
$\cT$ is one-way definable is in $2\expspace$. When $\cT$ is
sweeping, the problem is in $\expspace$.
\end{prop}
\reviewOne[inline]{The transition from 9.2 to 9.3 is a tad abrupt. The kind of argument you make before Corollary 8.7 to explain that this makes your result more robust or tight than I expected in first approach could be useful here.
\gabriele[inline]{Done, see below}%
}%
\medskip
The last result of the section shows that functional two-way transducers
are close to be the largest class for which a characterization
of one-way definability is feasible: as soon as we consider
arbitrary transducers (including non-functional ones),
the problem becomes undecidable.
\begin{prop}\label{prop:undecidability}
The one-way definability problem for \emph{non-functional}
sweeping transducers is undecidable.
\end{prop}
\begin{proof}
The proof uses some ideas and variants of constructions provided in \cite{Ibarra78},
concerning the proof of undecidability of the equivalence problem for one-way
non-functional transducers.
We show a reduction from the Post Correspondence Problem (PCP).
A \emph{PCP instance} is described by two finite alphabets $\Sigma$ and $\Delta$
and two morphisms $f,g:\Sigma^*\then\Delta^*$. A \emph{solution} of such an instance
is any non-empty word $w\in\Sigma^+$ such that $f(w)=g(w)$. We recall that the problem
of testing whether a PCP instance has a solution is undecidable.
Below, we fix a tuple $\tau=(\Sigma,\Delta,f,g)$ describing a PCP instance and we
show how to reduce the problem of testing the {\sl non-existence of solutions} of
$\tau$ to the problem of deciding {\sl one-way definability} of a relation computed
by a sweeping transducer.
Roughly, the idea is to construct a relation $B_\tau$ between words over a suitable
alphabet $\Gamma$ that encodes all the {\sl non-solutions} to the PCP instance
$\tau$ (this is simpler than encoding solutions because the presence of errors
can be easily checked). The goal is to have a relation $B_\tau$ that
(i) can be computed by a sweeping transducer and (ii) coincides with
a trivial one-way definable relation when $\tau$ has no solution.
We begin by describing the encodings for the solutions of the PCP instance.
We assume that the two alphabets of the PCP instance, $\Sigma$ and $\Delta$,
are disjoint and we use a fresh symbol $\#\nin \Sigma\cup\Delta$.
We define the new alphabet $\Gamma = \Sigma\cup\Delta\cup\{\#\}$ that will
serve both as input alphabet and as output alphabet for the transduction.
We call \emph{encoding} any pair of words over $\Gamma$ of the form
$(w\cdot u,w\cdot v)$, where $w\in\Sigma^+$, $u\in\Delta^*$, and $v\in\{\#\}^*$.
We will write the encodings as vectors to improve readability, e.g., as
\[
\lbinom{w\cdot u}{w\cdot v} \ .
\]
We denote by $E_\tau$ the set of all encodings and we observe that $E_\tau$
is computable by a one-way transducer (note that this transducer needs
$\varepsilon$-transitions).
We then restrict our attention to the pairs in $E_\tau$ that are encodings
of valid solutions of the PCP instance.
Formally, we call \emph{good encodings} the pairs in $E_\tau$ of the form
\[
\lbinom{w\cdot u}{w\cdot\#^{|u|}}
\qquad\qquad\text{where } u = f(w) = g(w) \ .
\]
All the other pairs in $E_\tau$ are called \emph{bad encodings}.
Of course, the relation that contains the good encodings is not computable
by a transducer. On the other hand, we can show that the complement
of this relation w.r.t.~$E_\tau$ is computable by a sweeping transducer.
Let $B_\tau$ be the set of all bad encodings.
Consider $(w\cdot u,w\cdot \#^m)\in E_\tau$, with $w\in\Sigma^+$,
$u\in\Delta^*$, and $m\in\bbN$, and we observe that this pair belongs to
$B_\tau$ if and only if one of the following conditions is satisfied:
\begin{enumerate}
\item $m<|u|$,
\label{enc1}
\item $m>|u|$,
\label{enc2}
\item $u\neq f(w)$,
\label{enc3}
\item $u\neq g(w)$.
\label{enc4}
\end{enumerate}
We explain how to construct a sweeping transducer $\cS_\tau$ that computes $B_\tau$.
Essentially, $\cS_\tau$ guesses which of the above conditions holds and processes
the input accordingly. More precisely, if $\cS_\tau$ guesses that the first condition
holds, then it performs a single left-to-right pass, first copying the prefix $w$
to the output and then producing a block of occurrences of the symbol $\#$ that is
shorter than the suffix $u$. This task can be easily performed while reading
$u$: it suffices to emit at most one occurrence of $\#$ for each position in $u$,
and at the same time guarantee that, for at least one such position, no occurrence
of $\#$ is emitted. The second condition can be dealt with by a similar strategy:
first copy the prefix $w$, then output a block of $\#$ that is longer than
the suffix $u$. To deal with the third condition, the transducer $\cS_\tau$
has to perform two left-to-right passes, interleaved by a backward pass that
brings the head back to the initial position.
During the first left-to-right pass, $\cS_\tau$ copies the prefix $w$ to the output.
During the second left-to-right pass, it reads again the prefix $w$, but this time
he guesses a factorization of it of the form $w_1\:a\:w_2$.
On reading $w_1$, $\cS_\tau$ will output $\#^{|f(w_1)|}$.
After reading $w_1$, $\cS_\tau$ will store the symbol $a$ and move to the position
where the suffix $u$ begins. From there, it will guess a factorization of $u$
of the form $u_1\:u_2$, check that $u_2$ does not begin with $f(a)$, and
emit one occurrence of $\#$ for each position in $u_2$.
The number of occurrences of $\#$ produced in the output is thus
$m=|f(w_1)| + |u_2|$, and the fact that $u_2$ does not begin with $f(a)$
ensures that the factorizations of $w$ and $u$ do not match, i.e.
\[ m\neq|f(w)| \]
Note that the described behaviour does not immediately guarantee that $u\neq f(w)$.
Indeed, it may still happen that $u=f(w)$, but as a consequence $m\neq |u|$.
This case is already covered by the first and second condition,
so the computation is still correct in the sense that it produces only
bad encodings.
On the other hand, if $m$ happens to be the same as $|u|$,
then $|u| = m \neq |f(w)|$ and thus $u\neq f(w)$.
A similar behaviour can be used to deal with the fourth condition.
\smallskip
We have just shown that there is a sweeping non-functional transducer $\cS_\tau$
that computes the relation $B_\tau$ containing all the bad encodings.
Note that, if the PCP instance $\tau$ admits no solution, then all encodings
are bad, i.e., $B_\tau=E_\tau$, and hence $B_\tau$ is one-way definable.
It remains to show that when $\tau$ has a solution, $B_\tau$ is not one-way
definable. Suppose that $\tau$ has solution $w\in\Sigma^+$ and let
$\big(w\cdot u,\:w\cdot \#^{|u|}\big)$ be the corresponding good encoding,
where $u=f(w)=g(w)$.
Note that every exact repetition of $w$ is also a solution, and hence
the pairs $\big(w^n\cdot u^n,\:w^n\cdot \#^{n\cdot|u|}\big)$ are also
good encodings, for all $n\ge 1$.
Suppose, by way of contradiction, that there is a one-way transducer $\cT$
that computes the relation $B_\tau$.
For every $n,m\in\bbN$, we define the encoding
\[
\alpha_{n,m} ~=~
\lbinom{w^n\cdot u^m}{w^n\cdot \#^{m\cdot|u|}}
\]
and we observe that $\alpha_{n,m} \in B_\tau$ if and only if $n\neq m$
(recall that $w\neq\emptystr$ is the solution of the PCP instance $\tau$ and $u=f(w)=g(w)$).
Below, we consider bad encodings like the above ones,
where the parameter $n$ is supposed to be large enough.
Formally, we define the set $I$ of all pairs of indices $(n,m)\in\bbN^2$
such that (i) $n\neq m$ (this guarantees that $\alpha_{n,m}\in B_\tau$)
and (ii) $n$ is larger than the number $|Q|$ of states of $\cT$.
We consider some pair $(n,m)\in I$ and we choose a successful
run $\rho_{n,m}$ of $\cT$ that witnesses the membership of
$\alpha_{n,m}$ in $B_\tau$, namely, that reads the input
$w^n\cdot u^m$ and produces the output $w^n\cdot \#^{m\cdot|u|}$.
We can split the run $\rho_{n,m}$ into a prefix $\mathpalette{\overarrow@\shortleftarrowfill@}\rho_{n,m}$
and a suffix $\mathpalette{\overarrow@\shortrightarrowfill@}\rho_{n,m}$ in such a way that $\mathpalette{\overarrow@\shortleftarrowfill@}\rho_{n,m}$
consumes the prefix $w^n$ and $\mathpalette{\overarrow@\shortrightarrowfill@}\rho_{n,m}$ consumes the remaining
suffix $u^m$.
Since $n$ is larger than the number of state of $\cT$, we can find a
factor $\hat\rho_{n,m}$ of $\mathpalette{\overarrow@\shortleftarrowfill@}\rho_{n,m}$ that starts and ends with
the same state and consumes a non-empty exact repetition of $w$,
say $w^{n_1}$, for some $1\le n_1\le |Q|$.
We claim that the output produced by the factor $\hat\rho_{n,m}$
must coincide with the consumed part $w^{n_1}$ of the input.
Indeed, if this were not the case, then deleting the factor $\hat\rho_{n,m}$
from $\rho_{n,m}$ would result in a new successful run that reads
$w^{n-n_1}\cdot u^m$ and produces $w^{n-n_2}\cdot \#^{m\cdot|u|}$
as output, for some $n_2\neq n_1$. This however would contradict
the fact that, by definition of encoding, the possible outputs produced
by $\cT$ on input $w^{n-n_1}\cdot u^m$ must agree on the prefix $w^{n-n_1}$.
We also remark that, even if we do not annotate this explicitly, the number
$n_1$ depends on the choice of the pair $(n,m)\in I$. This number, however,
range over the fixed finite set $J = \big[1,|Q|\big]$.
We can now pump the factor $\hat\rho_{n,m}$ of the run $\rho_{n,m}$ any
arbitrary number of times. In this way, we obtain new successful runs of
$\cT$ that consume inputs of the form
$w^{n+k\cdot n_1}\cdot u^m$ and produce outputs of the form
$w^{n+k\cdot n_1}\cdot \#^m$, for all $k\in\bbN$.
In particular, we know that $B_\tau$ contains all
pairs of the form $\alpha_{n+k\cdot n_1,m}$.
Summing up, we can claim the following:
\begin{clm}
There is a function $h:I\then J$ such that, for all pairs $(n,m)\in I$,
\[
\big\{ (n+k\cdot h(n,m),m) ~\big|~ k\in\bbN \big\} \:\subseteq\: I \ .
\]
\end{clm}
\noindent
We can now head towards a contradiction. Let $\tilde n$
be the maximum common multiple of the numbers $h(n,m)$,
for all $(n,m)\in I$. Let $m=n+\tilde n$ and observe that
$n\neq m$, whence $(n,m)\in I$. Since $\tilde n$ is a multiple
of $h(n,m)$, we derive from the above claim that the pair
$(n+\tilde n,m) = (m,m)$ also belongs to $I$.
However, this contradicts the definition of $I$, since we
observed earlier that $\alpha_{n,m}$ is a bad encoding if
and only if $n\neq m$.
We conclude that $B_\tau$ is not one-way definable
when $\tau$ has a solution.
\end{proof}
\section{Conclusions}\label{sec:conclusions}
It was shown in \cite{fgrs13} that it is decidable whether
a given two-way transducer can be implemented by some one-way
transducer. However, the provided algorithm has non-elementary complexity.
The main contribution of our paper is a new algorithm that solves
the above question with elementary complexity, precisely in $2\expspace$.
The algorithm is based on a characterization of those transductions,
given as two-way transducers, that can be realized by one-way
transducers. The flavor of our characterization is
different from that of \cite{fgrs13}. The approach of
\cite{fgrs13} is based on a variant of Rabin and Scott's construction
\cite{RS59} of one-way automata, and on local modifications of
the two-way run. Our approach instead relies on the global notion
of \emph{inversion} and more involved combinatorial arguments.
The size of the one-way transducer that we obtain is triply exponential
in the general case, and doubly exponential in the sweeping case, and
the latter is optimal.
\felix{add last sentence. Maybe we can do more on future work ? The characterization of sweeping in term of rational relations maybe ?}%
\gabriele{I did not like the sentence in the end, so I moved it here and rephrased (and it was not future work!)}%
The approach described here was adapted to characterize functional two-way transducers
that are equivalent to some sweeping transducer with either known or
unknown number of passes (see~\cite{bgmp16}, \cite{Bas17} for details).
\reviewOne[inline]{The conclusion gives no clue towards future works, either current, considered or at all possible.
\olivier[inline]{added the following paragraph}%
}%
Our procedure considers non-deterministic transducers,
both for the initial two-way transducer, and
for the equivalent one-way transducer, if it exists.
\gabriele{Improved and mentioned that determinism does not bring much to our problem}%
Deterministic two-way transducers are as expressive as
non-deterministic functional ones. This means that
starting from deterministic two-way transducers would address
the same problem in terms of transduction classes, but could
in principle yield algorithms with better complexity.
\gabriele{Added this tiny remark}%
We also recall that Proposition \ref{prop:lower-bound}
gives a tight lower bound for the size of a one-way transducer
equivalent to a given deterministic sweeping transducer. This
latter result basically shows that there is no advantage in
considering one-way definability for deterministic variants
of sweeping transducers, at least with respect to the size
of the possible equivalent one-way transducers.
\gabriele{slightly rephrased}%
A variant of the one-way definability problem asks whether a given
two-way transducer is equivalent to some \emph{deterministic} one-way transducer.
A decision procedure for this latter problem is obtained by combining
our characterization with the classical algorithm that determines whether a one-way
transducer can be determinized \cite{cho77,bealcarton02,weberklemm95}.
In terms of complexity, it is conceivable that a better algorithm
may exist for deciding definability by deterministic one-way transducers,
since in this case one can rely on structural properties
that characterize
deterministic transducers. \anca{I deleted the last phrase (``on the
contrary, our approach etc'') because at this point, the reader
should have remembered that we do only guessing.}
\section{Introduction}\label{sec:introduction}
Since the early times of computer science, transducers have been
identified as a fundamental computational model. Numerous
fields of computer science are ultimately concerned with
transformations, ranging from databases to image
processing, and an important challenge is to perform transformations with
low costs, whenever possible.
The most basic form of transformations is obtained using devices that
process an input with finite memory
and produce outputs during the processing. Such
devices are called finite-state transducers. Word-to-word finite-state
transducers were considered in very early work in formal language
theory~\cite{sch61,ahu69,eilenberg1974automata}, and it was soon clear that they are much more
challenging than finite-state word acceptors (classical finite-state
automata). One essential difference between transducers and automata
over words is that the capability to process the
input in both directions strictly increases the expressive power in the case
of transducers, whereas this does not for
automata~\cite{RS59,she59}. In other words, two-way word transducers
are strictly more expressive than one-way word transducers.
We consider in this paper functional transducers, that compute
functions from words to words. Two-way word transducers capture very
nicely the notion of regularity in this setting. Regular word
functions, in other words, functions computed by functional two-way transducers%
\footnote{We know from \cite{EH98,de2013uniformisation} that deterministic
and non-deterministic functional two-way transducers have the
same expressive power, though the non-deterministic variants are
usually more succinct.},
inherit many of the characterizations and algorithmic properties of the robust
class of regular languages. Engelfriet and Hoogeboom~\cite{EH98}
showed that monadic second-order definable graph transductions,
restricted to words, are equivalent to two-way transducers --- this
justifies the
notation ``regular'' word functions, in the spirit of classical
results in automata theory and logic by B\"uchi, Elgot, Rabin and
others. Recently, Alur and Cern\'{y}~\cite{AlurC10} proposed an enhanced
version of one-way transducers called streaming transducers, and
showed that they are equivalent to the two previous models. A streaming
transducer processes the input word from left to right, and stores
(partial) output words in some given write-only registers, that are
updated through concatenation and constant updates.
\reviewOne[inline]{Few reasons are given for a specific focus towards two-way transducers as opposed to equivalent models.
\olivier[inline]{I answered to the editor that the main reason is that we are more familiar with it, so it seems natural. But we shouldn't mention it here.}
}%
\anca[inline]{I agree not to say it here, but I would not say this
is the real reason - and I changed the answer. The real reason (for
me) is that pumping SST is much less intuitive.}%
Two-way transducers raise challenging questions about resource
requirements. One crucial resource is the number of times the
transducer needs to re-process the input word. In particular, the case
where the input can be processed in a single pass, from left to right,
is very attractive as it corresponds to the setting of \emph{streaming},
where the (possibly very large) inputs do not need to be stored in order
to be processed. The~\emph{one-way definability} of a functional
two-way transducer, that is, the question
whether the transducer is equivalent to some one-way transducer, was
considered quite recently: \cite{fgrs13} shows that one-way
definability of string transducers is a decidable property. However, the decision procedure
of~\cite{fgrs13} has non-elementary complexity, which raises the
question about the intrinsic complexity of this problem.
In this paper we provide an algorithm of elementary complexity solving
the one-way definability problem.
Our decision algorithm has single or doubly exponential space complexity,
depending on whether the input transducer is sweeping (namely, it performs reversals
only at the extremities of the input word) or genuinely two-way.
We also describe an algorithm that constructs an equivalent one-way
transducer, whenever it exists, in doubly or triply exponential time,
again depending on whether the input transducer is sweeping or
two-way. For the
construction of an equivalent one-way transducer we obtain a doubly
exponential lower bound, which is tight for sweeping transducers. Note that for the decision problem,
the best lower bound known is only polynomial space~\cite{fgrs13}.
\reviewOne[inline]{The subclass of sweeping transducers is central to
the paper: not only is the algorithm of better complexity in this
particular case, it is also used as an easy case to prepare for the
general proof. However, beyond this property, it is unclear whether
this class has other interesting properties, or has garnered
interest in the part. This information should, ideally, be made
available in the introduction. \olivier[inline]{added a paragraph
below.}%
}%
\reviewOne[inline]{Were deterministic two-way word transducers
considered at all (Introduction)? If not, do you plan to consider
them (Conclusion)? \olivier[inline]{added the footnote in intro,
and a paragraph in conclusion}%
\felix[inline]{I think the question is "do you have a better
complexity if the given transducer is deterministic?" like
usually. We maybe also need to add something like : One can notice
that if we want to decide definability by \emph{deterministic}
one-way transducers, an efficient procedure is known since
Choffrut78 \olivier[inline]{you can complete the conclusion if you
want. To me it is sufficient.}%
}%
}%
\olivier{added this paragraph}%
\gabriele{Changed a bit (without the reviewer comment the paragraph was a bit isolated)}%
\anca{changed a bit more}%
Our initial interest in sweeping transducers was the fact that they
provide a simpler setting for characterizing one-way
definability. Later it turned out that sweeping transducers enjoy interesting and
useful connections with streaming transducers: they have the same
expressiveness as streaming transducers where concatenation of
registers is disallowed. The connection goes even further, since the number of
sweeps corresponds exactly to the number of
registers~\cite{bgmp16}. The results of this paper were refined
in~\cite{bgmp16}, and used to determine the minimal number of
registers required by functional streaming transducers without
register concatenation.
\subsection*{Related work.}
Besides the papers mentioned above, there are
several recent results around the expressivity and the resources of
two-way transducers, or
equivalently, streaming transducers.
First-order definable transductions were shown to be equivalent to
transductions defined by aperiodic
streaming transducers~\cite{FiliotKT14} and to aperiodic two-way
transducers~\cite{CartonDartois15}. An effective characterization of
aperiodicity for one-way transducers was obtained in~\cite{FGL16}.
Register minimization for right-appending deterministic streaming
transducers
was shown to be
decidable in~\cite{DRT16}. An algebraic characterization of (not necessarily
functional) two-way transducers over unary alphabets was
provided in~\cite{CG14mfcs}. It was shown that in this case
sweeping transducers have the same expressivity.
The expressivity of non-deterministic input-unary
or output-unary two-way
transducers was investigated in~\cite{Gui15}.
In \cite{Smith14} a pumping lemma for two-way transducers is proposed,
and used to investigate properties of the output languages of two-way
transducers. In this paper we also rely on pumping arguments over runs
of two-way transducers, but we require loops of a particular form,
that allows to identify periodicities in the output.
\olivier{TODO ref Ismael to be improved before resubmission}%
The present paper unifies results on one-way definability obtained
in~\cite{bgmp15} and~\cite{bgmp17}.
Compared to the conference versions, some combinatorial proofs have
\gabriele{I have removed the reference to Saarela}%
been simplified,
and the complexity of the procedure presented in~\cite{bgmp15} has been
improved by one exponential.
\anca{please check. I deleted the sentence on the
under-approximation, because I don't find it understandable at this point.}%
\olivier{added this paragraph}%
\subsection*{Overview.}
Section~\ref{sec:preliminaries} introduces basic
notations for two-way and sweeping transducers, and Section~\ref{sec:overview}
states the main result and provides a roadmap of the proofs. For
better readability our paper
is divided in two parts: in the first part we consider the easier case of
sweeping transducers, and in the second part the general case. The two
proofs are similar, but the general case is more involved since we
need to deal with special loops (see
Section~\ref{sec:loops-twoway}). Since the high-level ideas of the
proofs are the same and the sweeping case illustrates them in a
simpler way, the proof in that setting is a preparation for the
general case. Both proofs have
the same structure: first we introduce some combinatorial arguments
(see Section~\ref{sec:combinatorics-sweeping} for the sweeping case
and Section~\ref{sec:combinatorics-twoway} for the general case), then we provide the
characterization of one-way definability (see
Section~\ref{sec:characterization-sweeping} for the sweeping case and
Section~\ref{sec:characterization-twoway} for the general
case). Finally, Section~\ref{sec:complexity} establishes the complexity
of our algorithms.
\section{The structure of two-way loops}\label{sec:loops-twoway}
While loop pumping in a sweeping transducer is rather simple,
we need a much better understanding when it comes to pump loops of
unrestricted two-way transducers.
This section is precisely devoted to untangling the structure of two-way loops.
We will focus on specific loops, called idempotent,
that generate repetitions with a ``nice shape'', very similar
to loops of sweeping transducers.
We fix throughout this section a functional two-way transducer $\cT$,
an input word $u$, and a (normalized) successful run $\rho$ of $\cT$ on $u$.
As usual, ${\boldsymbol{H}}} %h_{\mathsf{max}}}=2|Q|- 1$ is the maximal length of a crossing sequence
of $\rho$, and ${\boldsymbol{C}}} %c_{\mathsf{max}}}$ is the maximal number of letters output by a single
transition.
\reviewTwo[inline]{
I have some questions concerning the definition of the flows and effects.
Initially you fix a transducer T, a word u and a run $\rho$ of T on u, and you define the flow corresponding to an interval I of $\rho$. On line 10, you say that βwe consider effects that are not necessarily associated with intervals of specific runsβ. Then you define the set of all flows, and all effects.
If I understand correctly, a flow has to correspond to an interval I, whereas an effect is a triple containing a flow (corresponding to an interval I) and two crossing sequences (that might not correspond to I).
My first question is: wouldnβt it be more natural to have either
1. also flows that do not correspond to an interval,
2. or only effects that correspond to intervals?
I realise that in case 1, defining the flows in a way for Lemma 6.5 to still hold seems complicated, but I was wondering if you had considered it.
Second, I was wondering if it wouldnβt be easier to fix the set of nodes of the flows to $\{0,\dots , 2|Q| - 2 \}$.
This would allow a smoother definition of $G \cdot G'$ (l.25): as of now, the set of nodes of $G \cdot G'$ is not stated explicitly. If I am not mistaken, it is the set of nodes that are part of G and admit an outgoing edge, or are part of $G'$ and admit an incoming edge (which is not that complicated, but wouldn't be needed at all if we always have all the nodes).
\olivier[inline]{addressed in review2-answer.txt}%
}%
\medskip
\subsection*{Flows and effects.}
We start by analyzing the shape of factors of $\rho$ intercepted
by an interval $I=[x_1,x_2]$.
We identify four types of factors $\alpha$ intercepted by $I$
depending on the first location $(x,y)$ and the last location
$(x',y')$:
\begin{itemize}
\item $\alpha$ is an ${\ensuremath{\mathsf{LL}}}$-factor if $x=x'=x_1$,
\item $\alpha$ is an $\RR$-factor if $x=x'=x_2$,
\item $\alpha$ is an ${\ensuremath{\mathsf{LR}}}$-factor if $x=x_1$ and $x'=x_2$,
\item $\alpha$ is an ${\ensuremath{\mathsf{RL}}}$-factor if $x=x_2 $ and $x'=x_1$.
\end{itemize}
In Figure \ref{fig:intercepted-factors} we see that $\a$ is an
${\ensuremath{\mathsf{LL}}}$-factor, $\b,\delta$ are ${\ensuremath{\mathsf{LR}}}$-factors, $\z$ is an $\RR$-factor,
and $\g$ is an ${\ensuremath{\mathsf{RL}}}$-factor.
\begin{defi}\label{def:flow}
Let $I = [x_1,x_2]$ be an interval of $\rho$ and $h_i$ the length of
the crossing sequence $\rho|x_i$, for both $i=1$ and $i=2$.
The \emph{flow} $F_I$ of $I$ is the directed graph with set of nodes
$\set{0,\dots,\max(h_1,h_2)-1}$ and set of edges consisting of all
$(y,y')$ such that there is a factor of $\rho$ intercepted by $I$
that starts at location $(x_i,y)$ and ends at location $(x_j,y')$,
for $i,j\in\{1,2\}$.
The \emph{effect} $E_I$ of $I$ is the triple $(F_I,c_1,c_2)$,
where $c_i=\r|x_i$ is the crossing sequence at $x_i$.
\end{defi}
\noindent
For example, the interval $I$ of Figure \ref{fig:intercepted-factors}
has the flow graph $0\mapsto 1\mapsto 3\mapsto 4\mapsto 2\mapsto 0$.
It is easy to see that every node of a flow $F_I$ has at most one
incoming and at most one outgoing edge. More precisely, if $y<h_1$ is
even, then it has one outgoing edge (corresponding to an ${\ensuremath{\mathsf{LR}}}$- or ${\ensuremath{\mathsf{LL}}}$-factor
intercepted by $I$), and if it is odd it has one incoming edge (corresponding
to an ${\ensuremath{\mathsf{RL}}}$- or ${\ensuremath{\mathsf{LL}}}$-factor intercepted by $I$). Similarly, if $y<h_2$ is
even, then it has one incoming edge (corresponding to an ${\ensuremath{\mathsf{LR}}}$- or $\RR$-factor),
and if it is odd it has one outgoing edge (corresponding to an ${\ensuremath{\mathsf{RL}}}$- or
$\RR$-factor).
In the following we consider effects that are not necessarily
associated with intervals of specific runs. The definition of such effects
should be clear: they are triples consisting of a graph (called flow)
and two crossing sequences of lengths $h_1,h_2 \le {\boldsymbol{H}}} %h_{\mathsf{max}}}$, with sets of
nodes of the form $\{0,\ldots,\max(h_1,h_2)-1\}$,
that satisfy the in/out-degree properties stated above.
It is convenient to distinguish the edges in a flow based on the
parity of the source and target nodes. Formally, we partition any
flow $F$ into the following subgraphs:
\begin{itemize}
\item $F_{\ensuremath{\mathsf{LR}}}$ consists of all edges of $F$ between pairs of even nodes,
\item $F_{\ensuremath{\mathsf{RL}}}$ consists of all edges of $F$ between pairs of odd nodes,
\item $F_{\ensuremath{\mathsf{LL}}}$ consists of all edges of $F$ from an even node to an odd node,
\item $F_\RR$ consists of all edges of $F$ from an odd node to an even node.
\end{itemize}
We denote by $\cF$ (resp.~$\cE$) the set of all flows (resp.~effects)
augmented with a dummy element $\bot$. We equip both sets $\cF$ and $\cE$ with
a semigroup structure, where the corresponding products $\circ$ and $\odot$ are
defined below (similar definitions appear in \cite{Birget1990}).
Later we will use the semigroup structure to identify the \emph{idempotent loops},
that play a crucial role in our characterization of one-way definability.
\begin{defi}\label{def:product}
For two graphs $G,G'$, we denote by $G\cdot G'$ the graph with edges of
the form $(y,y'')$ such that $(y,y')$ is an edge of $G$ and $(y',y'')$ is an
edge of $G'$, for some node $y'$ that belongs
to both $G$ and $G'$.
Similarly, we denote by $G^*$ the graph with edges $(y,y')$
such that there exists a (possibly empty) path in $G$ from $y$ to $y'$.
The product of two flows $F,F'$ is the unique flow $F\circ F'$ (if it exists) such that:
\begin{itemize}
\item $(F\circ F')_{\ensuremath{\mathsf{LR}}} = F_{\ensuremath{\mathsf{LR}}} \cdot (F'_{\ensuremath{\mathsf{LL}}} \cdot F_\RR)^* \cdot F'_{\ensuremath{\mathsf{LR}}}$,
\item $(F\circ F')_{\ensuremath{\mathsf{RL}}} = F'_{\ensuremath{\mathsf{RL}}} \cdot (F_\RR \cdot F'_{\ensuremath{\mathsf{LL}}})^* \cdot F_{\ensuremath{\mathsf{RL}}}$,
\item $(F\circ F')_{\ensuremath{\mathsf{LL}}} = F_{\ensuremath{\mathsf{LL}}} ~\cup~ F_{\ensuremath{\mathsf{LR}}} \cdot (F'_{\ensuremath{\mathsf{LL}}} \cdot F_\RR)^* \cdot F'_{\ensuremath{\mathsf{LL}}} \cdot F_{\ensuremath{\mathsf{RL}}}$,
\item $(F\circ F')_\RR = F'_\RR ~\cup~ F'_{\ensuremath{\mathsf{RL}}} \cdot (F_\RR \cdot F'_{\ensuremath{\mathsf{LL}}})^* \cdot F_\RR \cdot F'_{\ensuremath{\mathsf{LR}}}$.
\end{itemize}
If no flow $F\circ F'$ exists with the above properties,
then we let $F\circ F'=\bot$.
The product of two effects $E=(F,c_1,c_2)$ and $E'=(F',c'_1,c'_2)$ is either the effect
$E\odot E' = (F\circ F',c_1,c'_2)$ or the dummy element $\bot$, depending on whether
$F\circ F'\neq \bot$ and $c_2=c'_1$.
\end{defi}
\noindent
For example, let $F$ be the flow of interval $I$
in Figure \ref{fig:intercepted-factors}. Then
$(F \circ F)_{\ensuremath{\mathsf{LL}}}=\set{0\mapsto 1, 2\mapsto 3}$,
$(F \circ F)_\RR=\set{1\mapsto 2, 3\mapsto 4}$, and
$(F \circ F)_{\ensuremath{\mathsf{LR}}}=\set{4\mapsto 0}$
--- one can quickly verify this with the
help of Figure \ref{fig:pumping-twoway}.
It is also easy to see that $(\cF,\circ)$ and $(\cE,\odot)$ are finite semigroups, and that
for every run $\r$ and every pair of consecutive intervals $I=[x_1,x_2]$ and $J=[x_2,x_3]$ of $\r$,
$F_{I\cup J} = F_I \circ F_J$ and $E_{I\cup J} = E_I \odot E_J$.
In particular, the function $E$ that associates each interval $I$ of $\rho$ with the
corresponding effect $E_I$ can be seen as a semigroup homomorphism.
\medskip
\subsection*{Loops and components.}
Recall that a loop is an interval $L=[x_1,x_2]$
with the same crossing sequences at $x_1$ and $x_2$.
We will follow techniques similar to those presented in Section \ref{sec:combinatorics-sweeping}
to show that the outputs generated in non left-to-right manner are essentially periodic.
However, differently from the sweeping case, we will consider only special types
of loops:
\begin{defi}\label{def:idempotent}
A loop $L$ is \emph{idempotent} if $E_L = E_L \odot E_L$ and $E_L\neq\bot$.
\end{defi}
\noindent
For example, the interval $I$ of Figure \ref{fig:intercepted-factors} is a loop,
if one assumes that the crossing sequences at the borders of $I$ are the same.
By comparing with Figure \ref{fig:pumping-twoway}, it is easy to see that $I$
is not idempotent. On the other hand, the loop consisting of 2 copies of $I$
is idempotent.
\input{pumping-twoway}
As usual, given a loop $L=[x_1,x_2]$ and a number $n\in\bbN$, we can
introduce $n$ new copies of $L$ and connect the intercepted factors
in the obvious way.
This results in a new run $\ensuremath{\mathsf{pump}}_L^{n+1}(\rho)$ on the word $\ensuremath{\mathsf{pump}}_L^{n+1}(u)$.
Figure \ref{fig:pumping-twoway} shows how to do this for $n=1$ and $n=2$.
Below, we analyze in detail the shape of the pumped run $\ensuremath{\mathsf{pump}}_L^{n+1}(\rho)$
(and the produced output as well) when $L$ is an {\sl idempotent} loop.
We will focus on idempotent loops because pumping non-idempotent loops
may induce permutations of factors that are difficult to handle.
For example, if we consider again the non-idempotent loop $I$ to the
left of Figure \ref{fig:pumping-twoway}, the factor of the run between
$\beta$ and $\gamma$ (to the right of $I$, highlighted in red) precedes
the factor between $\gamma$ and $\delta$ (to the left of $I$, again in red),
but this ordering is reversed when a new copy of $I$ is added.
When pumping a loop $L$, subsets of factors intercepted by $L$ are glued
together to form factors intercepted by the replication of $L$.
The notion of component introduced below identifies
groups of factors that are glued together.
\begin{defi}\label{def:component}
A \emph{component} of a loop $L$ is any strongly
connected component of its flow $F_L$
(note that this is also a cycle, since
every node in it has in/out-degree $1$).
Given a component $C$, we denote by
$\min(C)$ (resp.~$\max(C)$) the minimum (resp.~maximum)
node in $C$.
We say that $C$ is \emph{left-to-right} (resp.~\emph{right-to-left})
if $\min(C)$ is even (resp., odd).
An \emph{$(L,C)$-factor} is a factor of the run that is
intercepted by $L$ and that corresponds to an edge of $C$.
\end{defi}
\noindent
We will usually list the $(L,C)$-factors based on their order of occurrence in the run.
For example, the loop $I$ of Figure \ref{fig:pumping-twoway} contains
a single component $C=0\mapsto 1\mapsto 3\mapsto 4\mapsto 2\mapsto 0$
which is left-to-right.
Another example is given in Figure \ref{fig:many-components}, where the
loop $L$ has three components $C_1,C_2,C_3$ (colored in blue, red,
and green, respectively):
$\a_1,\a_2,\a_3$ are the $(L,C_1)$-factors, $\b_1,\b_2,\b_3$ are
the $(L,C_2)$-factors, and $\g_1$ is the unique $(L,C_3)$-factor.
\input{many-components}
\medskip
Below, we show that the levels of each component of a loop (not necessarily idempotent)
form an interval.
\begin{lem}\label{lem:component}
Let $C$ be a component of a loop $L=[x_1,x_2]$.
The nodes of $C$ are precisely the levels in the interval $[\min(C),\max(C)]$.
Moreover, if $C$ is left-to-right (resp.~right-to-left), then $\max(C)$
is the smallest level $\ge \min(C)$
such that between $(x_1,\min(C))$ and $(x_2,\max(C))$ (resp.~$(x_2,\min(C))$ and $(x_1,\max(C))$)
there are equally many ${\ensuremath{\mathsf{LL}}}$-factors and $\RR$-factors intercepted by $L$.
\end{lem}
\begin{proof}[Proof idea]
The proof of this lemma is rather technical and deferred to
Appendix~\ref{app:proof-component}, since the lemma is not at the core of the proof of the main
result. Let us first note that with the definition of $\max(C)$ stated
in the lemma it is rather easy to see that the interval
$[\min(C),\max(C)]$ is a union of cycles (i.e., components). This
can be shown by arguing that every node in $[\min(C),\max(C)]$ has
in-degree and out-degree one. What is much less obvious is that
$[\min(C),\max(C)]$ is connected, thus consists of a single cycle.
\input{edges}
The crux is thus to show that the nodes visited by every cycle of the flow
(or, equally,
every component) form an interval. For this, we use an induction based
on portions of the cycle, namely, on {\sl paths} of the flow.
The difficulty underlying the formalization of the inductive invariant comes from the
fact that, differently from cycles, paths of a flow may visit sets of levels that do
not form intervals.
An example is given in Figure \ref{fig:edges}, which represents some edges of a flow
forming a path from $\mathpalette{\overarrow@\shortrightarrowfill@} y_i$ to $\mathpalette{\overarrow@\shortrightarrowfill@} y_i +1$ and covering a non-convex set of nodes:
note that there could be a large gap between the nodes $\mathpalette{\overarrow@\shortrightarrowfill@} y_i$ and $\mathpalette{\overarrow@\shortleftarrowfill@} y_i$ due
to the unbalanced numbers of ${\ensuremath{\mathsf{LL}}}$-factors and $\RR$-factors below $\mathpalette{\overarrow@\shortleftarrowfill@} y_i$.
Essentially, the first part of the proof of the lemma amounts at identifying the sources
$\mathpalette{\overarrow@\shortrightarrowfill@} y_i$ (resp.~$\mathpalette{\overarrow@\shortleftarrowfill@} y_i$) of the ${\ensuremath{\mathsf{LR}}}$-factors (resp.~${\ensuremath{\mathsf{RL}}}$-factors), and at showing
that the latter factors are precisely of the form $\mathpalette{\overarrow@\shortrightarrowfill@} y_i \rightarrow \mathpalette{\overarrow@\shortleftarrowfill@} y_{i-1}+1$
(resp.~$\mathpalette{\overarrow@\shortleftarrowfill@} y_i \rightarrow \mathpalette{\overarrow@\shortrightarrowfill@} y_i +1$).
Once these nodes are identified, we show by induction on $i$ that every two
consecutive nodes $\mathpalette{\overarrow@\shortrightarrowfill@} y_i$ and $\mathpalette{\overarrow@\shortrightarrowfill@} y_i +1$ must be connected by a path whose
intermediate nodes form the interval $[\mathpalette{\overarrow@\shortleftarrowfill@} y_{i-1}+1,\mathpalette{\overarrow@\shortleftarrowfill@} y_i]$.
Finally, we argue that every cycle (or component) $C$ visits all and only
the nodes in the interval $[\min(C),\max(C)]$.
\end{proof}
The next lemma describes the precise shape and order of the intercepted factors
when the loop $L$ is idempotent.
\begin{lem}\label{lem:component2}
If $C$ is a left-to-right (resp.~right-to-left) component
of an {\sl idempotent} loop $L$, then the $(L,C)$-factors are in the following order:
$k$ ${\ensuremath{\mathsf{LL}}}$-factors (resp.~$\RR$-factors), followed by one ${\ensuremath{\mathsf{LR}}}$-factor (resp.~${\ensuremath{\mathsf{RL}}}$-factor),
followed by $k$ $\RR$-factors (resp.~${\ensuremath{\mathsf{LL}}}$-factors), for some $k \ge 0$.
\end{lem}
\begin{proof}
Suppose that $C$ is a left-to-right component of $L$.
We show by way of contradiction that $C$ has only one ${\ensuremath{\mathsf{LR}}}$-factor
and no ${\ensuremath{\mathsf{RL}}}$-factor. By Lemma~\ref{lem:component} this will yield
the claimed shape. Figure~\ref{fig:notidem} can be used as a reference
example for the arguments that follow.
We begin by listing the $(L,C)$-factors.
As usual, we order them based on their occurrences in the run $\rho$.
Let $\gamma$ be the first $(L,C)$-factor that is not an ${\ensuremath{\mathsf{LL}}}$-factor,
and let $\beta_1,\dots,\beta_k$ be the $(L,C)$-factors that precede $\gamma$
(these are all ${\ensuremath{\mathsf{LL}}}$-factors).
Because $\gamma$ starts at an even level, it must be an ${\ensuremath{\mathsf{LR}}}$-factor.
Suppose that there is another $(L,C)$-factor, say $\zeta$, that comes
after $\gamma$ and it is neither an $\RR$-factor nor an ${\ensuremath{\mathsf{LL}}}$-factor.
Because $\zeta$ starts at an odd level, it must be an ${\ensuremath{\mathsf{RL}}}$-factor.
Further let $\delta_1,\dots,\delta_{k'}$ be the intercepted $\RR$-factors
that occur between $\gamma$ and $\zeta$.
We claim that $k'<k$, namely, that the number of $\RR$-factors between
$\gamma$ and $\zeta$ is strictly less than the number of ${\ensuremath{\mathsf{LL}}}$-factors
before $\gamma$. Indeed, if this were not the case, then, by Lemma
\ref{lem:component}, the level where $\zeta$ starts would not belong to
the component $C$.
Now, consider the pumped run $\rho'=\ensuremath{\mathsf{pump}}_L^2(\rho)$, obtained by adding a new
copy of $L$. Let $L'$ be the loop of $\rho'$ obtained from the union
of $L$ and its copy. Since $L$ is idempotent, the components of $L$ are isomorphic
to the components of $L'$. In particular, we can denote by $C'$ the component
of $L'$ that is isomorphic to $C$.
Let us consider the $(L',C')$-factors of $\rho'$. The first $k$ factors
are isomorphic to the $k$ ${\ensuremath{\mathsf{LL}}}$-factors $\beta_1,\ldots,\beta_k$ from $\rho$.
However, the $(k+1)$-th element has a different shape: it is isomorphic to
$\gamma~\beta_1~\delta_1~\beta_2~\cdots~\delta_{k'}~\beta_{k'+1}~\zeta$,
and in particular it is an ${\ensuremath{\mathsf{LL}}}$-factor.
This implies that the $(k+1)$-th edge of $C'$ is of the form $(y,y+1)$,
while the $(k+1)$-th edge of $C$ is of the form $(y,y-2k)$.
This contradiction comes from having assumed the existence of
the ${\ensuremath{\mathsf{RL}}}$-factor $\zeta$, and is illustrated in Figure~\ref{fig:notidem}.
\end{proof}
\input{notidem}
\begin{rem}
Note that every loop in the sweeping case is
idempotent. Moreover, the $(L,C)$-factors are precisely the factors
intercepted by the loop $L$.
\end{rem}
\medskip
\subsection*{Pumping idempotent loops.}
To describe in a formal way the run obtained by pumping an idempotent loop,
we need to generalize the notion of anchor point in the two-way case
(the reader may compare this with the analogous definitions in
Section~\ref{sec:combinatorics-sweeping} for the sweeping case).
Intuitively, the anchor point of a component $C$ of an idempotent loop $L$
is the source location of the unique ${\ensuremath{\mathsf{LR}}}$- or ${\ensuremath{\mathsf{RL}}}$-factor intercepted by
$L$ that corresponds to an edge of $C$ (recall Lemma~\ref{lem:component2}):
\begin{defi}\label{def:anchor}
Let $C$ be a component of an idempotent loop $L = [x_1,x_2]$.
The \emph{anchor point} of $C$ inside $L$, denoted%
\footnote{In denoting the anchor point --- and similarly the trace --- of a component $C$
inside a loop $L$, we omit the annotation specifying $L$, since this is
often understood from the context.}
$\an{C}$, is either the location $(x_1,\max(C))$ or the location
$(x_2,\max(C))$, depending on whether $C$ is left-to-right or right-to-left.
\end{defi}
\noindent
We will usually depict anchor points by black circles (like, for instance, in Figure \ref{fig:many-components}).
It is also convenient to redefine the notation $\tr{\ell}$ for representing
an appropriate sequence of transitions associated with each anchor point
$\ell$ of an idempotent loop:
\begin{defi}\label{def:trace}
Let $C$ be a component of some idempotent loop $L$, let $\ell=\an{C}$
be the anchor point of $C$ inside $L$, and let
$i_0 \mapsto i_1 \mapsto i_2 \mapsto \dots \mapsto i_k \mapsto i_{k+1}$
be a cycle of $C$, where $i_0=i_{k+1}=\max(C)$.
For every $j=0,\dots,k$, further let $\beta_j$ be the factor intercepted
by $L$ that corresponds to the edge $i_j \mapsto i_{j+1}$ of $C$.
The \emph{trace} of $\ell$ inside $L$ is the run $\tr{\ell} = \beta_0 ~ \beta_1 ~ \cdots ~ \beta_k$.
\end{defi}
\noindent
Note that $\tr{\ell}$ is not necessarily a factor of the original run
$\rho$. However, $\tr{\ell}$ is indeed a run, since $L$ is a loop and
the factors $\b_i$ are concatenated according to the flow. As we will see below,
$\tr{\ell}$ will appear as (iterated) factor of the pumped version of $\rho$,
where the loop $L$ is iterated.
As an example, by referring again to the components $C_1,C_2,C_3$ of
Figure~\ref{fig:many-components}, we have the following traces:
$\tr{\an{C_1}}=\alpha_2\:\alpha_1\:\alpha_3$,
$\tr{\an{C_2}}=\beta_2\:\beta_1\:\beta_3$, and
$\tr{\an{C_3}}=\gamma_1$.
The next proposition shows the effect of pumping idempotent loops. The
reader can note the similarity with the sweeping case.
\begin{prop}\label{prop:pumping-twoway}
Let $L$ be an idempotent loop of $\rho$ with components $C_1,\dots,C_k$,
listed according to the order of their anchor points:
$\ell_1=\an{C_1} \mathrel{\lhd} \cdots \mathrel{\lhd} \ell_k=\an{C_k}$.
For all $n\in\bbN$, we have
\[
\ensuremath{\mathsf{pump}}_L^{n+1}(\rho) ~=~
\rho_0 ~ \tr{\ell_1}^n ~ \rho_1 ~ \cdots ~ \rho_{k-1} ~ \tr{\ell_k}^n ~ \rho_k
\]
where
\begin{itemize}
\item $\rho_0$ is the prefix of $\rho$ that ends at the first anchor point $\ell_1$,
\item $\rho_k$ is the suffix of $\rho$ that starts at the last anchor point $\ell_k$,
\item $\rho_i$ is the factor $\rho[\ell_i,\ell_{i+1}]$, for all $1\le i<k$.
\end{itemize}
\end{prop}
\begin{proof}
Along the proof we sometimes refer to Figure \ref{fig:many-components} to
ease the intuition of some definitions and arguments.
For example, in the left hand-side of Figure \ref{fig:many-components},
the run $\rho_0$ goes until the first location marked by a black circle;
the runs $\rho_1$ and $\rho_2$, resp., are between the first and the
second black dot, and the second and third black dot; finally, $\rho_3$
is the suffix starting at the last black dot. The pumped run
$\ensuremath{\mathsf{pump}}_L^{n+1}(\rho)$ for $n=2$ is depicted to the right of the figure.
Let $L=[x_1,x_2]$ be an idempotent loop and, for all $i=0,\dots,n$, let
$L'_i=[x'_i,x'_{i+1}]$ be the $i$-th copy of the loop $L$ in the pumped run
$\rho'=\ensuremath{\mathsf{pump}}_L^{n+1}(\rho)$, where $x'_i = x_1 + i\cdot (x_2-x_1)$
(the ``$0$-th copy of $L$'' is the loop $L$ itself).
Further let $L'=L'_0\cup\dots\cup L'_n = [x'_0,x'_{n+1}]$, that is,
$L'$ is the loop of $\rho'$ that spans across the $n+1$ occurrences of $L$.
As $L$ is idempotent, the loops $L'_0,\dots,L'_n$ and $L'$ have all the
same effect as $L$.
In particular, the components of $L'_0,\dots,L'_n$, and $L'$ are isomorphic
to and in same order as those of $L$.
We denote these components by $C_1,\dots,C_k$.
We let $\ell_j=\an{C_j}$ be the anchor point of each component $C_j$ inside
the loop $L$ of $\rho$
(these locations are marked by black dots in the left hand-side
of Figure \ref{fig:many-components}).
Similarly, we let $\ell'_{i,j}$ (resp.~$\ell'_j$)
be the anchor point of $C_j$ inside the loop $L'_i$ (resp.~$L'$).
From Definition \ref{def:anchor}, we have that either $\ell'_j=\ell'_{1,j}$ or $\ell'_j=\ell'_{n,j}$,
depending on whether $C_j$ is left-to-right or right-to-left (or, equally, on whether $j$ is odd or even).
Now, let us consider the factorization of the pumped run $\rho'$
induced by the locations $\ell'_{i,j}$, for all $i=0,\dots,n$ and for $j=1,\dots,k$
(these locations are marked by black dots in the right hand-side of the figure).
By construction, the prefix of $\rho'$ that ends at location $\ell'_{0,1}$
coincides with the prefix of $\rho$ that ends at $\ell_1$,
i.e.~$\rho_0$ in the statement of the proposition.
Similarly, the suffix of $\rho'$ that starts at location $\ell'_{n,k}$ is isomorphic
to the suffix of $\rho$ that starts at $\ell_k$, i.e. $\rho_k$ in the statement.
Moreover, for all odd (resp.~even) indices $j$, the factor
$\rho'[\ell'_{n,j},\ell'_{n,j+1}]$ (resp.~$\rho'[\ell_{0,j},\ell_{0,j+1}]$) is isomorphic
to $\rho[\ell_j,\ell_{j+1}]$, i.e.~the $\rho_j$ of the statement.
The remaining factors of $\rho'$ are those delimited by the pairs of locations
$\ell'_{i,j}$ and $\ell'_{i+1,j}$, for all $i=0,\dots,n-1$ and all $j=1,\dots,k$.
Consider one such factor $\rho'[\ell'_{i,j},\ell'_{i+1,j}]$,
and assume that the index $j$ is odd (the case of an even $j$ is similar).
This factor can be seen as a concatenation of factors intercepted by $L$
that correspond to edges of $C_j$ inside $L'_i$.
More precisely, $\rho'[\ell'_{i,j},\ell'_{i+1,j}]$ is obtained by concatenating
the unique ${\ensuremath{\mathsf{LR}}}$-factor of $C_j$ --- recall that by Lemma \ref{lem:component2}
there is exactly one such factor --- with an interleaving of the ${\ensuremath{\mathsf{LL}}}$-factors
and the $\RR$-factors of $C_j$.
As the components are the same for all $L'_i$'s and for $L$, this corresponds
precisely to the trace $\tr{\ell_j}$ (cf.~Definition \ref{def:trace}).
Now that we know that $\rho'[\ell'_{i,j},\ell'_{i+1,j}]$ is isomorphic to $\tr{\ell_j}$,
we can conclude that
$\rho'[\ell'_{0,j},\ell'_{n,j}] \:=\:
\rho'[\ell'_{0,j},\ell'_{1,j}] ~ \dots ~ \rho'[\ell'_{n-1,j},\ell'_{n,j}]$
is isomorphic to $\tr{\ell_j}^n$.
\end{proof}
\section*{Acknowledgment}
\noindent The authors wish to acknowledge fruitful discussions with
Emmanuel Filiot, Isma\"el Jecker and Sylvain Salvati. We also thank
the referees for their very careful reading and the suggestions for improvement.
\bibliographystyle{abbrv}
\section{One-way definability: overview}\label{sec:overview}
In this section we state our main result, which is the existence of an
elementary algorithm for checking whether a two-way transducer is
equivalent to some one-way transducer. We call such transducers
\emph{one-way definable}. Before stating our result, we
start with a few examples illustrating the reasons that may prevent a
transducer to be one-way definable.
\begin{exa}\label{ex:one-way-definability}
We consider two-way transducers that accept any input $u$
from a given regular language $R$ and produce as output the word $u\,u$.
We will argue how, depending on $R$, these transducers may or may not be one-way definable.
\begin{enumerate}
\item If $R=(a+b)^*$, then there is no equivalent one-way transducer,
as the output language is not regular.
If $R$ is finite, however, then the transduction mapping $u\in R$ to $u\,u$
can be implemented by a one-way transducer that stores the input $u$
(this requires at least as many states as the cardinality of $R$),
and outputs $u\,u$ at the end of the computation.
\item A special case of transduction with finite domain is obtained from the language
$R_n = \{ a_0 \, w_0 \, \cdots a_{2^n-1} \, w_{2^n-1} \::\: a_i\in\{a,b\} \}$,
where $n$ is a fixed natural number, the input alphabet is $\{a,b,0,1\}$,
and each $w_i$ is the binary encoding of the index $i=0,\dots,2^n-1$
(hence $w_i\in\{0,1\}^n$).
According to Proposition~\ref{prop:lower-bound} below, the transduction
mapping $u\in R_n$ to $u\,u$ can be implemented by a two-way transducer
of size $\cO(n^2)$, but every equivalent one-way transducer
has size (at least) doubly exponential in $n$.
\item Consider now the periodic language $R=(abc)^*$.
The function that maps $u\in R$ to $u\,u$ can be easily implemented by a
one-way transducer: it suffices to output alternatively $ab$, $ca$, $bc$
for each input letter, while checking that the input is in $R$.
\end{enumerate}
\end{exa}
\begin{exa}\label{ex:running}
We consider now a slightly more complicated transduction
that is defined on input words of the form $u_1 \:\#\: \cdots \:\#\: u_n$,
where each factor $u_i$ is over the alphabet $\Sigma=\{a,b,c\}$.
The associated output has the form
$w_1 \:\#\: \cdots \:\#\: w_n$, where each $w_i$
is either $u_i \: u_i$ or just $u_i$, depending on whether or not
$u_i\in (abc)^*$ and $u_{i+1}$ has even length, with $u_{n+1}=\emptystr$
by convention.
\noindent
The natural way to implement this transduction is by means
of a two-way transducer that performs multiple passes
on the factors of the input:
a first left-to-right pass is performed on
$u_i \,\#\, u_{i+1}$ to produce the first copy of $u_i$
and to check whether $u_i\in (abc)^*$ and $|u_{i+1}|$ is even; if so,
a second pass on $u_i$ is performed to produce
another copy of $u_i$.
\noindent
Observe however that the above transduction can also be implemented by a one-way
transducer, using non-determinism: when entering a factor $u_i$, the transducer
guesses whether or not $u_i\in (abc)^*$ and $|u_{i+1}|$ is even;
depending on this it outputs either $(abc\,abc)^{\frac{|u_i|}{3}}$
or $u_i$, and checks that the guess is correct while proceeding to
read the input.
\end{exa}
\noindent
\medskip
The main result of our paper is an elementary algorithm that decides
whether a functional transducer is one-way definable:
\begin{thm}\label{thm:main}
There is an algorithm that takes as input a functional two-way
transducer $\cT$ and outputs in $3\exptime$ a \emph{one-way} transducer
$\cT'$ satisfying the following properties:
\begin{enumerate}
\item $\cT'\subseteq\cT$,
\item $\dom(\cT')=\dom(\cT)$ if and only if $\cT$ is one-way definable.
\item $\dom(\cT')=\dom(\cT)$ can be checked in $2\expspace$.
\end{enumerate}
Moreover, if $\cT$ is a sweeping transducer, then $\cT'$ can be
constructed in $2\exptime$
and $\dom(\cT')=\dom(\cT)$ is decidable in $\expspace$.
\end{thm}
\begin{rem}
The transducer $\cT'$ constructed in the above theorem is
in a certain sense maximal: for every
$v \in \dom(\cT) ~\setminus~ \dom(\cT')$
and every one-way transducer $\cT''$ with
$\dom(\cT') \subseteq \dom(\cT'') \subseteq \dom(\cT)$ there exists
some witness input $v'$ obtained from $v$ such that
$v' \in \dom(\cT) ~\setminus~ \dom(\cT'')$. We will make this more
precise at the end of Section~\ref{sec:characterization-twoway}.
\end{rem}
We also provide a two-exponential lower bound for the size of the equivalent transducer.
As the lower bound is achieved by a sweeping transduction (even a deterministic one),
this gives a tight lower bound on the size of any one-way transducer equivalent to
some sweeping transducer.
\begin{prop}\label{prop:lower-bound}
\gabriele{I have added the fact that the sweeping transducers are even deterministic}%
There is a family $(f_n)_{n\in\bbN}$ of transductions such that
\begin{enumerate}
\item $f_n$ can be implemented by a deterministic sweeping transducer of size $\cO(n^2)$,
\item $f_n$ can be implemented by a one-way transducer,
\item every one-way transducer that implements $f_n$
has size $\Omega(2^{2^n})$.
\end{enumerate}
\end{prop}
\begin{proof}
The family of transformations is precisely the one described in
Example~\ref{ex:one-way-definability}~(2), where $f_n$ maps inputs of the form
$u = a_0 \, w_0 \, \cdots \, a_{2^n-1} ~ w_{2^n-1}$ to outputs of the form $u\,u$,
where $a_i\in\{a,b\}$ and $w_i\in\{0,1\}^n$ is the binary encoding of $i$.
A deterministic sweeping transducer implementing $f_n$ first checks that the
binary encodings $w_i$, for $i=0,\dots,2^n-1$, are correct.
This can be done with $n$ passes:
the $j$-th pass uses $\cO(n)$ states to check the correctness of
the $j$-th bits of the binary encodings.
Then, the sweeping transducer performs two additional passes to
copy the input twice. Overall, the sweeping transducer has size $\cO(n^2)$.
As already mentioned, every one-way transducer that implements $f_n$ needs
to remember input words $u$ of exponential length in order to output $u\,u$, which
roughly requires doubly exponentially many states.
A more formal argument providing a lower bound for the size of a one-way
transducer implementing $f_n$ goes as follows.
First of all, one observes that given a one-way transducer $\cT$,
the language of its outputs,
i.e., $L^{\text{out}}_\cT = \{ w \::\: (u,w)\in\sL(\cT) \text{ for
some } u\}$
is regular. More precisely, if $\cT$ has size $N$, then the language
$L^{\text{out}}_\cT$ is recognized by an automaton of size linear in $N$.
Indeed, while parsing $w$, the automaton can guess an input word $u$
and a run on $u$,
together with a factorization of $w$ in which the $i$-th
factor corresponds to the output of the transition
on the $i$-th letter of $u$. Basically, this requires
storing as control states the transition rules of $\cT$ and the
suffixes of outputs.
Now, suppose that the function $f_n$ is implemented by a one-way
transducer $\cT$ of size $N$. The language
$L^{\text{out}}_\cT = \{ u\,u \::\: u\in\dom(f_n) \}$ is then
recognized by an automaton of size $\cO(N)$.
Finally, we recall a result from \cite{GlaisterShallit96}, which shows that,
given a sequence of pairs of words $(u_i,v_i)$, for $i=1,\dots,M$,
every non-deterministic automaton that separates the language
$\{u_i\,v_i \::\: 1\le i\le M\}$ from the language
$\{u_i\,u_j \::\: 1\le i\neq j\le M\}$ must have at least $M$ states.
By applying this result to our language $L^{\text{out}}_\cT$, where
$u_i=v_i$ for all $i=1,\dots,M=2^{2^n}$, we get that $N$ must be at
least linear in $M$, and hence $N \in \Omega(2^{2^n})$.
\end{proof}
\bigskip
\gabriele{improved a bit, but we may do better}%
The proof of Theorem~\ref{thm:main} will be developed in the next sections.
The main idea is to decompose a run of the two-way transducer $\cT$
into factors that can be easily simulated in a one-way manner. We defer the
formal definition of such a decomposition to Section \ref{sec:characterization-sweeping},
while here we refer to it simply as a ``$\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition'', where
$\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ is a suitable number computed from $\cT$.
The reader can refer to Figure~\ref{fig:decomposition-sweeping} on page
\pageref{fig:decomposition-sweeping}, which provides some intuitive
account of a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition for a sweeping run.
Roughly speaking,
each factor
of a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition either already looks like a run of a
one-way transducer (e.g.~the factors $D_1$ and $D_2$ of Figure~\ref{fig:decomposition-sweeping}),
or it produces a periodic output, where the period is bounded by $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$
(e.g.~the factor between $\ell_1$ and $\ell_2$).
Identifying factors that look like runs of one-way transducers is rather easy.
On the other hand, to identify factors with periodic outputs we rely on
a notion of ``inversion'' of a run. Again, we defer the formal definition
and the important combinatorial properties of inversions
to Section \ref{sec:combinatorics-sweeping}.
The reader can refer to Figure \ref{fig:inversion-sweeping}
on page \pageref{fig:inversion-sweeping}
for an example of an inversion of a run of a sweeping transducer.
Intuitively, this is a portion of run that is potentially difficult to simulate
in a one-way manner, due to existence of long factors of the output that are
generated following the opposite order of the input.
Finally, the complexity of the decision procedure in Theorem~\ref{thm:main}
is analyzed in Section~\ref{sec:complexity}.
\reviewOne[inline]{As neither inversions nor blocks are defined by Page 8, the presence of Theorem 3.6 is not as enlightening as one might hope. If the notion of inversion and their relevance is made quite clear at a glance of Page 11, the definition of run decomposition is not as intuitive. Hence, Theorem 3.6 cannot be appreciated before Page 17. It might be a good idea to give an intuition about inversion and diagonal/blocks before 3.6 (this solution being preferable, but made challenging by how involved the notion of decomposition seems to be in a first read).
\felix[inline]{added the references to the corresponding figures}%
\olivier[inline]{and Gabriele added some intuitions on decompositions}%
}%
\subsection*{Roadmap.}
In order to provide a roadmap of our proofs, we state below the equivalence
between the key properties related to one-way definability, inversions of runs,
and existence of decompositions:
\begin{thm}\label{thm:main2}
Given a functional two-way transducer $\cT$,
an integer $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$ can be computed such that the following are equivalent:
\begin{itemize}
\item[\PR1)] $\cT$ is one-way definable,
\item[\PR2)] for every successful run of $\cT$ and every inversion in it,
the output produced amid
the inversion has period at most $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$,
\item[\PR3)] every input has a successful run of $\cT$ that admits a $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition.
\end{itemize}
\end{thm}
As the notions of inversion and $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decomposition are simpler to formalize
for sweeping transducers, we will first prove the theorem assuming that $T$ is
a sweeping transducer; we will focus later on unrestricted two-way transducers.
Specifically, in Section~\ref{sec:combinatorics-sweeping} we introduce the basic combinatorics
on words and the key notion of inversion for a run of a sweeping transducer, and we prove the
implication \PR1 $\Rightarrow$ \PR2.
In Section~\ref{sec:characterization-sweeping} we define $\boldsymbol{B}} %{\cmax \cdot \hmax \cdot 2^{3\emax}$-decompositions of runs of
sweeping transducers, prove the implication \PR2 $\Rightarrow$ \PR3, and sketch a proof of
\PR3 $\Rightarrow$ \PR1 (as a matter of fact, this latter implication can be proved in a way
that is independent of whether $\cT$ is sweeping or not, which explains why we
only sketch the proof in the sweeping case).
Section~\ref{sec:loops-twoway} lays down the appropriate definitions
concerning loops of two-way transducers, and analyzes in detail the effect of
pumping special idempotent loops.
In Section~\ref{sec:combinatorics-twoway} we further develop the combinatorial arguments
that are used to prove the implication \PR1 $\Rightarrow$ \PR2 in the general case.
Finally, in Section~\ref{sec:characterization-twoway} we prove the implications
\PR2 $\Rightarrow$ \PR3 $\Rightarrow$ \PR1 in the general setting,
and show how to decide the condition $\dom(\cT')=\dom(\cT)$ of Theorem \ref{thm:main}.
\section{Preliminaries}\label{sec:preliminaries}
We start with some basic notations and definitions for two-way
automata and transducers. We assume that every input word $w=a_1\cdots a_n$
has two special delimiting symbols $a_1 = \vdash$
and $a_n = \dashv$ that do not occur elsewhere: $a_i \notin \{\vdash,\dashv\}$
for all $i=2,\dots,n-1$.
A \emph{two-way automaton} is a tuple
$\cA=(Q,\Sigma,\vdash,\dashv,\Delta,I,F)$, where
\begin{itemize}
\item $Q$ is a finite set of states,
\item $\Sigma$ is a finite alphabet (including $\vdash, \dashv$),
\item $\Delta \subseteq Q \times \Sigma \times Q \times \set{\ensuremath{\mathsf{left}},\ensuremath{\mathsf{right}}}$
is a transition relation,
\item $I,F\subseteq Q$ are sets of initial and final states, respectively.
\end{itemize}
By convention, left transitions on $\vdash$ are not
allowed; on the other hand, right transitions on $\dashv$ are allowed, but, as we will see,
they will necessarily appear as last transitions of successful runs.
A \emph{configuration} of $\cA$ has the form $u\,q\,v$,
with $uv \in \vdash\, \S^* \, \dashv$
and $q \in Q$. A configuration $u\,q\,v$ represents the
situation where the current state of $\cA$ is $q$ and its head reads the first
symbol of $v$ (on input $uv$). If $(q,a,q',\ensuremath{\mathsf{right}}) \in \Delta$,
then there is a transition from any configuration of the form
$u\,q\,av$ to the configuration $ua\,q'\,v$; we denote such a transition
by $u\,q\,av \trans{a,\ensuremath{\mathsf{right}}} ua\,q'\,v$.
Similarly, if $(q,a,q',\ensuremath{\mathsf{left}}) \in \Delta$,
then there is a transition from any configuration of the form
$ub\,q\,av$ to the configuration $u\,q'\,bav$,
denoted as $ub\,q\,av \trans{a,\ensuremath{\mathsf{left}}} u\,q'\,bav$.
A \emph{run} on $w$
is a sequence of transitions.
It is \emph{successful} if it starts in an initial configuration
$q\, w$, with $q\in I$, and ends in a final configuration $w\,q'$,
with $q' \in F$ --- note that this latter configuration does not allow
additional transitions. The \emph{language} of $\cA$ is the set of words
that admit a successful run of $\cA$.
The definition of \emph{two-way transducers} is similar to that of two-way automata,
with the only difference that now there is an additional output alphabet $\Gamma$ and the transition
relation is a finite subset of $Q \times \Sigma \times \Gamma^* \times Q \times \{\ensuremath{\mathsf{left}},\ensuremath{\mathsf{right}}\}$,
which associates an output over $\Gamma$ with each transition of the underlying two-way automaton.
For a two-way transducer $\cT=(Q,\Sigma,\vdash,\dashv,\Gamma,\Delta,I,F)$,
we have a transition of the form $ub\,q\,av \trans{a,d \, \mid w} u'\,q'\,v'$, outputting $w$,
whenever $(q,a,w,q',d)\in\Delta$ and either $u'=uba ~\wedge~ v'=v$ or
$u'=u ~\wedge~ v'=bav$, depending on whether $d=\ensuremath{\mathsf{right}}$ or $d=\ensuremath{\mathsf{left}}$.
The \emph{output} associated with a run
$\rho = u_1\,q_1\,v_1 \trans{a_1,d_1 \mid w_1}<\qquad> \dots
\trans{a_n,d_n \mid w_n}<\qquad> u_{n+1}\,q_{n+1}\,v_{n+1}$
of $\cT$ is the word $\out{\rho} = w_1\cdots w_n$. A transducer $\cT$
defines a relation $\sL(\cT)$
\anca{I was wondering if we ever use this notation, but
I found it on page 10... Maybe it is the only place...
consisting of all pairs $(u,w)$ such that
$w=\out{\rho}$, for some successful run $\rho$ on $u$.
The \emph{domain} of $\cT$, denoted $\dom(\cT)$,
is the set of input words that have a successful run.
For transducers $\cT,\cT'$, we write $\cT' \subseteq \cT$
to mean that $\dom(\cT') \subseteq \dom(\cT)$ and the transductions
computed by $\cT,\cT'$ coincide on $\dom(\cT)$.
A transducer is called \emph{one-way} if it does not contain transition
rules of the form $(q,a,w,q',\ensuremath{\mathsf{left}})$. It is called \emph{sweeping} if
it can perform reversals only at the borders of the input word.
A transducer that is equivalent to some one-way
(resp.~sweeping) transducer is called \emph{one-way definable}
(resp.~\emph{sweeping definable}).
The \emph{size} of a transducer takes into
account both the state space and the transition relation, and thus
includes the length of the output of each transition.
\medskip
\subsection*{Crossing sequences.}
The first notion that we use throughout the paper is that of crossing sequence.
We follow the convenient presentation from \cite{HU79}, which appeals to a
graphical representation of runs of a two-way transducer, where each configuration
is seen as a point (location) in a two-dimensional space.
Let $u=a_1\cdots a_n$ be an input word (recall that $a_1=\vdash$ and $a_n=\dashv$)
and let $\rho$ be a run of a two-way automaton (or transducer) on $u$.
The \emph{positions} of $\rho$ are the numbers from $0$ to $n$, corresponding
to ``cuts'' between two consecutive letters of the input. For example,
position $0$ is just before the first letter $a_1$,
position $n$ is just after the last letter $a_n$,
and any other position $x$, with $1\le x<n$, is between
the letters $a_x$ and $a_{x+1}$.
We will denote by $u[x_1,x_2]$ the factor of $u$ between
the positions $x_1$ and $x_2$ (both included).
\olivier{added}%
\gabriele{Please check this reformulation in terms of configurations}%
Each configuration $u\,q\,v$ of a two-way run $\r$
has a specific position associated with it. For technical reasons
we need to distinguish leftward and rightward transitions. If the configuration $u\,q\,v$
is the target of a rightward transition
then the position associated with $u\,q\,v$ is $x=|u|$.
The same definition also applies when $u\,q\,v$ is the initial configuration,
for which we have $u=\emptystr$ and $x=|u|=0$.
Otherwise, if $u\,q\,v$ is the target of a leftward transition
then the position associated with $u\,q\,v$ is $x=|u|+1$. Note that in
both cases, the letter read by the transition leading to $u\,q\,v$ is
$a_x$.
\anca{rewritten a bit}%
A \emph{location} of $\rho$ is any pair $(x,y)$,
where $x$ is the position of some configuration of $\rho$ and
$y$ is any non-negative integer for which there are at least
$y+1$ configurations in $\rho$ with the same position $x$.
The second component $y$ of a location is called \emph{level}.
For example, in Figure \ref{fig:run} we represent a possible
run of a two-way automaton together with its locations $(0,0)$,
$(1,0)$, $(2,0)$, $(2,1)$, etc.
Each location is naturally associated with a configuration, and thus a state.
Formally, we say that $q$ is the \emph{state at location $\ell=(x,y)$} in $\rho$,
and we denote this by writing $\rho(\ell)=q$,
if the $(y+1)$-th configuration of $\rho$ with position $x$ has state $q$.
Finally, we define the \emph{crossing sequence} of $\rho$ at position $x$
as the tuple $\rho|x=(q_0,\dots,q_h)$, where the $q_y$'s are all the states
at locations of the form $(x,y)$, for $y=0,\dots,h$.
\input{run}
As shown in Figure~\ref{fig:run}, a two-way run can be represented
as an path between locations annotated by the associated states.
We observe in particular that if a location $(x,y)$ is the target
of a rightward transition, then this transition has read the symbol $a_x$;
similarly, if $(x,y)$ is the target of a leftward transition, then
the transition has read the symbol $a_{x+1}$.
We also observe that, in any successful run $\rho$, every crossing
sequence has odd length and every rightward (resp.~leftward) transition
reaches a location with even (resp.~odd) level.
In particular, we can identify four types of transitions between locations,
depending on the parities of the levels (the reader may refer again to
Figure~\ref{fig:run}):
\begin{center}
\begin{tikzpicture}[baseline=0,scale=0.9]
\draw (-1,1.75) node (node1) {$\phantom{\!+\!1}~(x,2y)$};
\draw (3,1.75) node (node2) {$(x\!+\!1,2y')~\phantom{\!+\!1}$};
\draw (-1,0) node (node3) {$(x,2y\!+\!1)$};
\draw (3,0) node (node4) {$(x\!+\!1,2y'\!+\!1)$};
\draw (9,1.75) node (node5) {$\phantom{\!+\!1}~(x,2y)$};
\draw (9,2.5) node (node6) {$(x,2y\!+\!1)$};
\draw (9,0) node (node7) {$(x,2y\!+\!1)$};
\draw (9,0.75) node (node8) {$(x,2y\!+\!2)$};
\draw (node1) edge [->] node [above=-0.05, scale=0.9] {\small $a_{x+1},\ensuremath{\mathsf{right}}$} (node2);
\draw (node4) edge [->] node [above=-0.05, scale=0.9] {\small $a_{x+1},\ensuremath{\mathsf{left}}$} (node3);
\draw (node5.east) edge [->, out=0, in=0, looseness=2]
node [right=0, scale=0.9] {\small $a_{x+1},\ensuremath{\mathsf{left}}$} (node6.east);
\draw (node7.west) edge [->, out=180, in=180, looseness=2]
node [left=0, scale=0.9] {\small $a_x,\ensuremath{\mathsf{right}}$} (node8.west);
\end{tikzpicture}
\vspace{1mm}
\end{center}
Hereafter, we will identify runs with the corresponding annotated paths between locations.
It is also convenient to define a total order $\mathrel{\unlhd}$ on the locations of a run $\rho$
by letting $\ell_1 \mathrel{\unlhd} \ell_2$ if $\ell_2$ is reachable from $\ell_1$ by following the
path described by $\rho$ --- the order $\mathrel{\unlhd}$ on locations is called
\emph{run order}.
Given two locations $\ell_1 \mathrel{\unlhd} \ell_2$ of a run $\rho$, we write $\rho[\ell_1,\ell_2]$
for the factor of the run that starts in $\ell_1$ and ends in $\ell_2$. Note that the latter
is also a run and hence the notation $\outb{\rho[\ell_1,\ell_2]}$ is permitted.
\gabriele{Parity of levels needs to be preserved}%
We will often reason with factors of runs up to isomorphism, that is, modulo
shifting the coordinates of their locations while preserving the parity of the levels.
Of course, when the last location of (a factor of) a run $\rho_1$
coincides with the first location of (another factor of) a run $\rho_2$,
then $\rho_1$ and $\rho_2$ can be concatenated to form a longer run,
denoted by $\rho_1 \rho_2$.
\olivier{added, see review 2 below}%
\gabriele{rephrased a bit, and I think corrected a problem with negative numbers.
But do we really need to say this?}%
This operation can be performed even if the two locations,
say $(x_1,y_1)$ and $(x_2,y_2)$, are different, provided that
$y_1=y_2\bmod 2$: in this case it suffices to shift the positions
(resp.~levels) of the locations of the first run $\rho_1$ by $x_2$ (resp.~$y_2$)
and, similarly, the positions (resp.~levels) of the locations of the second
run $\rho_2$ by $x_1$ (resp.~$y_1$).
\felix{it's not really enough, with our current definition the first location of $\rho_2$ is $(x,0)$ is $x$ is the position of the cut. So we cannot even concatenate $\rho_1$ and $\rho_2$. I have an alternate paragraph where the locations of the subruns use the levels of the original run.}%
\gabriele{I don't understand your comment Felix. Does it still apply?}%
\reviewTwo[inline]{
There seems to be some problems in you definition of crossing sequences.
First, according to your definition, the state at location $(x,y)$ is the state reached by the $(y+1)$-th transition that crosses x. In order for this to be consistent with Figure 1, you should swap u and u' in your definition of a position crossed by a transition.
\olivier[inline]{Correct, there was a mistake there! Check my fix above.}%
Second, since, following your definition, the state at a given location always is the target of a transition, the first state of the run has no location (it should be added explicitly).
\olivier[inline]{Correct, also fixed}%
Third, not allowing more freedom in the second components of the locations seems to cause some problems while considering subruns.
Let me explain what I mean with an example.
Let us suppose that we split the run $\rho$ of your Figure 1 into two runs $\rho_1$ and $\rho_2$, where $\rho_2$ is composed of the last three transitions.
Then the locations of the states of the run $\rho_1$ (considered as a whole run) are equal to the locations of the same states considered as part of $\rho$.
However, the locations of the states of the run $\rho_2$ (considered as a whole run) are (1,0), (2,0), (3,0), (4,0), whereas in $\rho$, the locations of the corresponding states are (1,2), (2,2), (3,0), (4,0).
As a consequence, $\rho_1$ and $\rho_2$ can not be concatenated to reform $\rho$, since to do so we need that βthe last location of [\dots] $\rho_1$ coincides with the first location of [\dots] $\rho_2$β (p.5 l.15).
\olivier[inline]{Added a sentence to explicit}%
Finally (and this one is not a real problem), you introduce precise notions, but sometimes do not use them in the following parts of the paper: for example, p.4, βnever uses a left transition from position xβ corresponds to βno transition crosses position xβ. Why not use the latter ?
\olivier[inline]{``no left transition on $\vdash$'' comes before the definition of ``crosses'' and is more explicit, so I would not change.}%
}%
\anca{there is no ``crosses'' anymore, right?}%
\begin{wrapfigure}{r}{5.1cm}
\input{intercepted-factors}
\end{wrapfigure}
\medskip
\subsection*{Intercepted factors.}
For simplicity, we will denote by $\omega$
the maximal position of the input word.
We will consider \emph{intervals of positions} of
the form $I=[x_1,x_2]$, with $0 \le x_1<x_2 \le \omega$.
The \emph{containment} relation $\subseteq$ on intervals is
defined expected, as $[x_3,x_4] \subseteq [x_1,x_2]$ if $x_1\le x_3 < x_4\le x_2$.
A \emph{factor} of a run $\rho$ is a contiguous subsequence of $\rho$.
A factor of $\rho$ \emph{intercepted} by an interval
$I=[x_1,x_2]$ is a maximal factor that visits only
positions $x\in I$, and never uses a left transition from
position $x_1$ or a right transition from position $x_2$.
Figure~\ref{fig:intercepted-factors} on the right shows
the factors $\alpha,\beta,\gamma,\delta,\zeta$ intercepted
by an interval $I$.
The numbers that annotate the endpoints of the factors
represent their levels.
\medskip
\subsection*{Functionality.}
We say that a transducer is \emph{functional} (equivalently, one-valued, or single-valued)
if for each input $u$, at most one output $w$ can be
produced by any possible successful run on $u$.
Of course, every deterministic transducer is functional, while the opposite
implication fails in general.
To the best of our knowledge determining the precise complexity of the determinization
of a two-way transducer (whenever an equivalent deterministic one-way transducer exists)
is still open. From classical bounds on determinization of finite automata, we only know
that the size of a determinized transducer may be exponential in the worst-case.
One solution to this question, which is probably
not the most efficient one, is to
check one-way definability: if an equivalent one-way transducer is
constructed, one can check in $\ptime$ if it can be
determinized~\cite{cho77,weberklemm95,bealcarton02}.
The following result, proven in Section \ref{sec:complexity},
is the reason to consider only functional transducers:
{
\renewcommand{\thethm}{\ref{prop:undecidability}}
\begin{prop}[
The one-way definability problem for \emph{non-functional}
sweeping transducers is undecidable.
\end{prop}
}
Unless otherwise stated, hereafter we tacitly assume that all transducers
are functional. Note that functionality is a decidable property, as shown below.
The proof of this result is similar to the decidability proof for the equivalence problem
of deterministic two-way transducers~\cite{Gurari80}, as it reduces the functionality
problem to the reachability problem of a $1$-counter automaton of exponential size.
A matching $\pspace$ lower bound follows by a reduction of the emptiness problem
for the intersection of finite-state automata \cite{Kozen77}.
\begin{prop}\label{prop:functionality-pspace}
Functionality of two-way transducers can be decided in polynomial
space. This problem is $\pspace$-hard already for sweeping transducers.
\end{prop}
A (successful) run of a two-way transducer is called \emph{normalized}
if it never visits two locations with the same position, the same
state, and both either at even or at odd level. It is easy to see that
if a successful run $\rho$ of a \emph{functional} transducer visits two locations
$\ell_1=(x,y)$ and $\ell_2=(x,y')$ with the same state
$\rho(\ell_1)=\rho(\ell_2)$ and with $y=y' \mod 2$, then the output
produced by $\rho$ between $\ell_1$ and $\ell_2$ is empty:
otherwise, by repeating the non-empty factor $\rho[\ell_1,\ell_2]$, we would
contradict functionality. So, by deleting the factor
$\rho[\ell_1,\ell_2]$ we obtain a successful run that produces the
same output. Iterating this operation leads to an equivalent,
normalized run.
Normalized runs are interesting because their crossing sequences have
bounded length (at most ${\boldsymbol{H}}} %h_{\mathsf{max}}} = 2|Q|-1$). Throughout the paper we
will implicitly assume that successful runs are normalized. The
latter property can be easily checked on crossing sequences.
| f323656664f8d3dcfb3212ab38bc5518b342a7bd | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Introduction}
This article follows up on the ideas in \cite{I.CH} and \cite{R.P} on monotonicity eigenvalues which are relative
to continuous domains. The purpose of this work is to explore in the case of weighted directed graphs, some familiar
facts of monotonicity proved on domains of $\mathbb{R}^n$ and on compact Riemannian manifolds. Specially, our main question is:
"can one study the behavior of the eigenvalues of a special operator under perturbations on finite graphs?"
First, we consider a finite, directed and connected graph $G$ with non symmetric edge weights. Then, we introduce
the associated non symmetric Laplacian $\Delta_G$. We introduce the self-adjoint operator $S_G=\Delta_G+\Delta^*_G$, such that it
is easier to examine its spectrum thanks to selfadjointness.
We give some spectral properties of $\Delta_G$ and we show that the real part of its eigenvalues can coincide with the eigenvalues of $\dfrac{1}{2}S_G$. Secondly, we study the monotonicity of eigenvalues relative to vertices or edges of $G$. We prove that the $k^{th}$ eigenvalue $\lambda_k$ of $S_G$ is decreasing only in a class of graphs called here \textit{flower-like-graphs}, and it is monotone
increasing in the set of edges. These results are inspired by the classical results going back from M. Fiedler
\cite{M.F} and P. Kurasov, G. Malenova, S. Naboko \cite{Kuras} on the first nonzero eigenvalue of a simple graph.
We extend these results for the higher eigenvalues $\lambda_k$ of our special operator. In the second part
of our work, we try to establish and improve the Proposition 2.1 of \cite{B.B} for a Riemannian manifold $M$ which
gives upper bounds on the higher eigenvalues $\lambda_k(M)$ in terms of Dirichlet eigenvalues on components of a
partition of $M$.
Let us briefly outline the contents of this article. We shall start with a short section of preliminaries consisting
on some basic properties of the non symmetric Laplacian $\Delta_G$ on $G$ and the associated Green formula.
In section \ref{monot}, we establish a generalization of some monotonicity eigenvalue results. Furthermore, we study
the example of flower-like graphs and insist on the interest of the subgraph concept. Hence, we can remark that
these considerations of graph help to give an upper bound of the eigenvalues of a simple tree. This section includes
also similar Weyl and Cauchy theorems for the matrices \cite{H.J}.
In section \ref{comp}, we are interested on the study of the eigenvalues of the Dirichlet Laplacian.
We involve a comparison on the eigenvalues to use the decomposition of $G$ into two components $A$ and $B$ and
we give an upper bound on the eigenvalues of $G$ in terms of Dirichlet eigenvalues of $A$ and $B$.
\section{Preliminaries}
We will review in this section some basic definitions and introduce the notation used in the article.
They are introduced in \cite{bal} for the infinite graph.
\subsection{Notion of Graphs}
We call oriented or directed graph, the couple $G=(V,\vec{E})$, where $V$ is a set of vertices,
and $\vec{E}\subset V\times V$ is a set of directed edges. For two vertices $x,y$ of $V$,
we denote by $(x,y)$ the edge that connects $x$ to $y$, we also say that $x$ and $y$ are \textit{ neighbors}.
For all $x\in V$, we set:
\begin{itemize}
\item $E=\left\{\{x,y\},~(x,y)\in\vec{E}\text{ or } (y,x)\in\vec{E}\right\}$
\item $V_x^+=\left\{y\in V,~~(x,y)\in\vec{E}\right\}$
\item $V_x^-=\left\{y\in V,~~(y,x)\in\vec{E}\right\}$
\item $V_x=V_x^+\cup V_x^-$.
\end{itemize}
The valency on $G$ is given by:
$$v(x)=\#V_x~~\text{ for all } x\in V.$$
We introduce some definitions given in \cite{CN}, \cite{bal}, \cite{R.D}, \cite{N.T} for the case of symmetric graphs.
\begin{defini}
\begin{itemize}
\item A path between two vertices $x$ and $y$ in $V$ is a finite set of directed edges
$(x_1,y_1);~(x_2,y_2);..;(x_n,y_n),~n\geq 1$ such that
$$x_1=x,~y_n=y \text{ and } x_i=y_{i-1}~~\forall~2\leq i\leq n.$$
\item $G=(V,\vec {E})$ is called connected if two vertices are always related by a path.
\item $G=(V,\vec {E})$ is called strongly connected if there is for all vertices $x,y$ a path from $x$ to $y$
and one from $y$ to $x$.
\begin{exem}
The cycle graph $C_n=\{0,1,..,n-1\}$, with
$$(0,1)\in\vec{E},~(1,2)\in\vec{E},..,(n-2,n-1)\in\vec{E},~(n-1,0)\in\vec{E}$$
is strongly connected.
\end{exem}
\item Define for a finite subset $\Omega$ of ~$V$, the interior, the vertex boundary and the edge boundary
of $\Omega$ respectively by:
$$\overset{\circ}{\Omega}=\big\{y\in \Omega,~ V_y \subset\Omega\big\}$$
$$\partial\Omega=\big\{y\in\Omega^c,~y\in V_x~\text{ for some } x\in \Omega\big\}$$
$$\partial_E \Omega=\big\{(x,y)\in \vec{E}:~ (x\in \Omega,~y\in \Omega^{c} )~~or ~~(x\in \Omega^{c},~y\in \Omega)\big\}.$$
\end{itemize}
\end{defini}
We remark that the strong connectedness of $G$ assures that:
\begin{equation}\label{cnx}
\forall ~x\in V, ~~\#V_x^+\neq 0~and~\#V_x^-\neq 0.
\end{equation}
In this work we suppose that $G$ is \textbf{finite}, \textbf{connected} and satisfies the \textbf{Hypothesis} (\textbf{\ref{cnx}}). In this work, we take the following definition.
\begin{defi}\label{an}
\textbf{Directed weighted Graph}: A weighted graph $(G,b)$ is the data of a graph $(V,\vec{E})$ and a weight
$b:V\times V\to \mathbb{R}_+$ satisfying the following conditions:
\begin{itemize}
\item $b(x,x)=0$ for all $x\in V$, ~~(no loops in $\vec{E}$)
\item $b(x,y)>0$ iff $(x,y)\in \vec{E}$
\item Assumption $(\beta)$.
\end{itemize}
\end{defi}
\textbf{Assumption $(\beta)$}: for all $x\in V$,~ $\beta^+(x)=\beta^-(x)$\\
where $$\beta^+(x)=\sum_{y\in V_x^+ }b(x,y)\text{ and }\beta^-(x)=\sum_{y\in V_x^-}b(y,x).$$
The weight $\beta_G$ on a vertex $x\in V$ is given by:
$$\beta_G(x)=\beta^+(x)+\beta^-(x)=2\beta^+(x).$$
\begin{rem}
The Assumption $(\beta)$ is natural, it looks like the Kirchhoff's law in the electrical networks.
\end{rem}
The weighted graph is \textit{symmetric} if for all $x,y\in V$, $b(x,y)=b(y,x)$,\\
as a consequence $(x,y)\in \vec{E}\Rightarrow (y,x)\in\vec{E}$ (the graph is symmetric).\\
In addition, we consider a weight $m$ on $V$:
$$m:V\to \mathbf{R}^*_+.$$
\subsection{Functional spaces}
Let us introduce the following function spaces associated to the graph $G$:
$$\mathcal{C}_m(V)=\{f:V\to \mathbb{C} \}$$
endowed with the following inner product:
$$(f,g)_m=\sum_{x\in V}m(x)f(x)\overline{g(x)}.$$
We define its associated norm by:
$$\|f\|_m=\sqrt{(f,f)_m}.$$
A particular case called normalized is for $m = \beta^+$.\\
For a subset $U$ of $V$, Let
$$\mathcal{C}_m(U)=\{f\in \mathcal{C}_m(V),~ f \text{ with support in U}\}.$$
The weights $m$ and $b$ are called simple if they are constant equal to $1$ on $V$ and $E$ respectively.
We denote by $G^{s}$ the simple graph (with simple weights).
\section{Laplacian on directed graphs}
For a weighted connected directed graph $(G,b)$, we introduce the combinatorial Laplacians:
\begin{defini}
\begin{itemize}
\item We define the Laplacian $\Delta_G$ on $\mathcal{C}_m(V)$ by:
$$\Delta_G f(x)=\frac{1}{m(x)}\sum_{y\in V_x^+}b(x,y)\big(f(x)-f(y)\big).$$
\item In particular, if for all $x\in V$, $\beta^{+}(x)=m(x)$, the Laplacian is said to be the
normalized Laplacian and is defined on $\mathcal{C}_{\beta^+}(V)$ by:
$$\tilde{\Delta}_Gf(x)=\frac{1}{\beta^+(x)}\sum_{y\in V^+_x}b(x,y)\big(f(x)-f(y)\big).$$
\item For any operator $A$ on $\mathcal{C}_m(V)$, the Dirichlet operator $A^{D}_U$,
where $U $ is a subset of $ V$, is defined by:
$$f \text{ is with support in }U,~~ A^{D}_U(f)=A(f)\vert_ U.$$
\end{itemize}
\end{defini}
Thanks to Hypothesis $(\beta)$ the adjoint of $\Delta$ has a simple expression.
\begin{defi}
\textbf{Adjoint of an operator}: The adjoint operator $\Delta^*$ of $\Delta$ is defined by:
$$\forall \phi,\psi \in \mathcal{C}_m(V), ~( \Delta \psi, \phi) = (\psi, \Delta^*\phi).$$
\end{defi}
\begin{prop} Let $f$ be a function of $\mathcal{C}_m(V)$, we have
$$\Delta^*_G f(x)=\frac{1}{m(x)}\sum_{y\in V_x^-}b(y,x)\big(f(x)-f(y)\big).$$
\end{prop}
\begin{demo}
The following calculation for all $f,g \in \mathcal{C}_m(V)$ gives:
\begin{align*}
(\Delta_G f,g)_m=&\sum_{(x,y)\in \vec{E}}b(x,y)\big( f(x)-f(y)\big) \overline{g(x)}\\
=&\sum_{x\in V}\overline{g(x)}f(x)\sum_{y\in V^{+}_x}b(x,y)-\sum_{(y,x)\in \vec{E}}b(y,x) \overline{g(y)}f(x)\\
=&\sum_{x\in V}\overline{g(x)}f(x)\sum_{y\in V^{-}_x}b(y,x)-\sum_{(y,x)\in \vec{E}}b(y,x) \overline{g(y)}f(x)\\
=&\sum_{x\in V} f(x)\sum_{y\in V_x^-}b(y,x)\big( \overline{g(x)-g(y)\big)}\\
=&(f,\Delta^*_Gg)_m.
\end{align*}
\end{demo}
The Green's formula is one of the main tools when we are working with the symmetric Laplace operator.
In the following we establish it for the non symmetric Laplacian.
\begin{lemm} \textbf{Green's Formula}.
Let $f$ and $g$ be two functions of $\mathcal{C}_m(V)$. Then
$$(\Delta_G f,g)_m+(\Delta^*_Gf,g)_m=\sum_{(x,y)\in \vec{E}}b(x,y)\big( f(x)-f(y)\big) \big( \overline{g(x)-g(y)}\big).$$
\end{lemm}
\begin{demo}
The proof is a simple calculation:
\begin{align*}
(\Delta_G f,g)_m+(\Delta^*_Gf,g)_m=&\sum_{(x,y)\in \vec{E}} b(x,y)\big(f(x)-f(y)\big)\overline{g(x)}\\
+&\sum_{(y,x)\in \vec{E}} b(y,x)\big(f(x)-f(y)\big)\overline{g(x)}\\
=&\sum_{(x,y)\in \vec{E}} b(x,y)\Big(f(x)\overline{g(x)}+f(x)\overline{g(x)}-f(y)\overline{g(x)} -f(x)\overline{g(y)}\Big)\\
=&\sum_{(x,y)\in \vec{E}} b(x,y)\big( f(x)-f(y)\big) \big(\overline{g(x)-g(y)}\big).
\end{align*}
\end{demo}
\begin{defi}{\textbf{Special Laplacian.}}
We define a special Laplacian $S_G$ as the sum of the two non self-adjoint Laplacians $\Delta_G$ and $\Delta^{*}_G$, given by:
\begin{align*}
S_Gf(x)&=(\Delta_G+\Delta^{*}_G)f(x)\\
&=\frac{1}{m(x)}\sum_{y\in V_x^{+}\cup V_x^{-}}\big(b(x,y)+b(y,x)\big)\big(f(x)-f(y)\big)\\
&=\frac{1}{m(x)}\sum_{y\in V_x}a(x,y)\big(f(x)-f(y)\big)
\end{align*}
where $a(x,y)=b(x,y)+b(y,x)$ for any $x,y\in V$.
\end{defi}
\begin{rem}
\begin{enumerate
\item $S_G$ is a symmetric operator on $\mathcal{C}_m(V)$, because $(\Delta^{*}_G)^*=\Delta_G$.
\item $S_G$ is a positive operator: for all $f\in \mathcal{C}_m(V)$, \begin{align*}
(S_Gf,f)= &(\Delta_Gf,f)+(\Delta_G^*f,f) \\
= & \sum_{(x,y)\in \vec{E}} b(x,y)\big| f(x)-f(y)\big|^2 \\
\geq & 0.
\end{align*}
\end{enumerate}
\end{rem}
In our discussion on the study of eigenvalues of a self-adjoint operator, it is natural to introduce
the different characterization by variational principles \cite{Has}.
\subsection{Variational principles and Properties}
Let $A$ be a bounded from below self-adjoint operator. The eigenvalues of $A$ can be characterized by three fundamental variational principles:
the Rayleigh's principle, the Poincar\'e-Ritz
max-min principle and the Courant-Fischer-Weyl principle applied to the Rayleigh
quotients $\mathcal{R}(f)=\dfrac{(Af,f)}{(f,f)}$, $f\neq 0$.
\\
Let us arrange the eigenvalues of $A$ as
$$\lambda_1\leq \lambda_2\leq...\leq\lambda_n$$
counted according to their multiplicities. \\
In this case we have :
\begin{enumerate
\item The \textit{Rayleigh's principle} states:
\begin{equation}
\lambda_k=\min_{f\neq0,~(f,f_i)=0 \atop i=1,...,k-1}\mathcal{R}(f)
\label{rai}
\end{equation}
where $f_i$ are eigenvectors corresponding to the eigenvalues $\lambda_i$ and the minimum is reached
at the eigenvector $f_k$.
\item The \textit{Poincar\'e-Ritz principle} establishes:
\begin{equation}
\lambda_k=\min_{dim \Omega=k}~\max_{f\in \Omega,~f\neq 0}\mathcal{R}(f).
\label{poin}
\end{equation}
\item The \textit{Courant-Fischer-Weyl principle} is given in the form:
\begin{equation}
\lambda_k=\max_{dim \Omega=k-1}~\min_{f\perp \Omega\atop~f\neq 0}\mathcal{R}(f).
\label{fish}
\end{equation}
\end{enumerate}
The result below establishes a link between the eigenvalues of $S_G$ and $\Delta_G$.
We assume that the eigenvalues of $\Delta_G$ are ordered as follows respectively:
$$\mathcal{R}e(\lambda_1(\Delta_G))\leq\mathcal{R}e(\lambda_2(\Delta_G))..\leq\mathcal{R}e(\lambda_n(\Delta_G)).$$
\begin{lemm}\label{delta}
$$2\mathcal{R}e(\lambda_n(\Delta_G))\leq\lambda_n(S_G)$$
\end{lemm}
\begin{demo}
Let $f$ be an eigenfunction associated to $\lambda_n(\Delta_G)$, we have
\begin{align*}
\lambda_n(S_G)\geq&\dfrac{(S_G f,f)_m}{(f,f)_m}\\
\geq & \dfrac{(\Delta_G f,f)_m+\overline{(\Delta_G f,f)}_m}{(f,f)_m} \\
\geq & 2\mathcal{R}e(\lambda_n(\Delta_G)).
\end{align*}
\end{demo}
\begin{rem}
In a particular case, the previous inequalities are strict. Let us consider the following example, where $m\equiv 1$,
\begin{figure}[ht]
\begin{center}
\includegraphics*[height=2.5cm,width=3cm]{nsymN.pdf}
\end{center}
\caption{Graph with real spectrum}
\end{figure}
we have $\sigma(S_G)=\{0,3,7\}$ and $\sigma(\Delta_G)=\{0,2,3\}$.
\end{rem}
We introduce in the following a particular case of graphs whose $\sigma(S_G)=2\mathcal{R}e\big(\sigma(\Delta_G)\big)$.
\begin{exem}
Let us consider the simple cycle graph $C_3$, see the Figure \ref{cyc}, we have ,
$$\sigma(\Delta_{C_3})=\left\{0,~\frac{3}{2}+i\frac{\sqrt{3}}{2},~\frac{3}{2}-i\frac{\sqrt{3}}{2}\right\}\text{ and } \sigma(S_{C_3})=\left\{0,3,3\right\}.$$
\begin{figure}[ht]
\begin{center}
\includegraphics*[height=2cm,width=3cm]{cycle.pdf}
\end{center}
\caption{Cycle graph}
\label{cyc}
\end{figure}
\end{exem}
In the following proposition we determine the spectrum of the non symmetric Laplacian $\Delta_{C_n}$. We follow the same approach as Grigoryan for the symmetric Laplacian \cite{A.G} page 49.
\begin{prop}
The eigenvalues of $\Delta_{C_n}=I-P$ are as follows:
\begin{enumerate
\item If $n$ is odd then the eigenvalues are $\lambda=0$ (simple) and $\lambda=1-e^{\pm i\frac{2l\pi }{n}}$ for all $l=1,..,\frac{n-1}{2}$ (simple).
\item If $n$ is even then the eigenvalues are $\lambda=0,~2$ (simple) and $\lambda=1-e^{\pm i\frac{2l\pi}{n}}$ for all $l=1,..,\frac{n}{2}-1$ (simple).
\end{enumerate}
\end{prop}
\begin{demo}
To compute the eigenvalues of $\Delta_{C_n}=I-P$, it is sufficient to determine the spectrum of $P$. Let $\alpha$ be an eigenvalue of the operator $Pf(k+1)= f(k)$, for $k=0,..,n-1$, which leads to $f(k)=\alpha^k f(0)$ but $f(n)=f(0)$ thus $\alpha^n=1$. As $f$ is an eigenfunction $(f(0)\neq 0)$, then $\alpha=e^{\pm i\theta}$. As $f$ is n-periodic provided $n\theta$
is a multiple of $2\pi$, hence, $$\theta=\frac{2l\pi}{n},$$
where $l$ is an integer of $(0,\frac{n}{2})$.
\end{demo}
Observe that an interesting corollary concerning the spectra of $S_{C_n}$ and $\Delta_{C_n}$.
\begin{coro}
$$\sigma(S_{C_n})=2\mathcal{R}e\big(
\sigma(\Delta_{C_n})\big).$$
\end{coro}
\begin{demo}
We refer the Lemma 2.7 of \cite{A.G}, we remark that the eigenvalues of the operator $\dfrac{1}{2}(P+P^*)f(k)=\dfrac{1}{2}\big(f(k+1)+f(k-1)\big)$ coincide with the real part of the eigenvalues of $P$.
\end{demo}
Using the Green's formula, we establish some properties of the spectrum on any graph $G$ .
\begin{prop}
\begin{enumerate
\item $0$ is a simple eigenvalue of $\tilde{S}_G$ and $S_G$.
\item All the eigenvalues of $\tilde{S}_G$ are contained in $[0,4]$.
\item The real part of the eigenvalues of $\tilde{\Delta}_G$ are also contained in $[0,2]$.
\end{enumerate}
\end{prop}
\begin{demo}
\begin{enumerate
\item As in the case of undirected graph \cite{A.G}, we have for all $f\in\mathcal{C}_{\beta^+}(V)$,
$$(\tilde{S}_Gf,f)_{\beta^+}=\sum_{(x,y)\in \vec{E}} b(x,y)|f(x)-f(y)|^2.$$
Clearly, the constant function is an eigenfunction of $0$. Assume now that f is an eigenfunction
of the eigenvalue $0$. By the connectedness of $G$, $f$ is constant, which will imply that $0$ is a simple eigenvalue.
It is similar for $\tilde{S}_G$.
\item It is sufficient to prove that $\tilde{S}_G$ is bounded by $4$ because $\tilde{S}_G$ is non negative by the Green's formula. In fact, for all $f\in \mathcal{C}_{\beta^+}(V)$
and thanks to Assumption $(\beta)$, we obtain
\begin{align*}
(\tilde{S}_Gf,f)_{\beta^+}=&\sum_{(x,y)\in \vec{E}} b(x,y)|f(x)-f(y)|^2\\
\leq &2\sum_{(x,y)\in \vec{E}} b(x,y)\big||f(x)|^2+|f(y)|^2\big|\\
\leq &2\sum_{(x,y)\in \vec{E}} b(x,y)|f(x)|^2+2\sum_{(x,y)\in \vec{E}} b(x,y)|f(y)|^2\\
\leq &2\sum_{x\in V}\sum_{y\in V_x^+} b(x,y)|f(x)|^2+2\sum_{y\in V}\sum_{x\in V_x^-} b(x,y)|f(y)|^2\\
\leq &2\sum_{x\in V}|f(x)|^2\beta^+(x)+2\sum_{y\in V}|f(y)|^2\beta^-(y)\\
\leq & 4(f,f)_{\beta^+}.
\end{align*}
\item We deduce directly our inclusion thanks to the Lemma \ref{delta}.
\end{enumerate}
\end{demo}
\section{Domain monotonicity of eigenvalues}\label{monot}
The purpose of this part is to give an overview of some results concerning the monotonicity with regard to
the domain, of eigenvalues of $S_G$, the special self-adjoint Laplacian associated to directed graphs
with non symmetric edge weights. We could be concerned with the related question:\\
Does a given eigenvalue increases or decreases under a given perturbation of $G$?
\subsection{Definitions on $G$}
Before discussing the study of variation of eigenvalues, let us recall some basic definitions:
let $G=(V,\vec{E})$ be a graph,
\begin{itemize}
\item The graph $G_1=(V, \vec{E}_1)$ is called a partial graph of $G$, if $\vec{E}_1$ is included in $\vec{E}$.
\item A graph $H=(V_H,\vec{E}_H)$ is called a subgraph of $G=(V_G,\vec{E}_G)$ if $V_H\subset V_G$
and $\vec{E}_H=\big\{(x,y) ;~x,y \in V_H~~\big\}\cap \vec{E}_G $.
\item A graph $(V_U,\vec{E}_U)$ is called a part of a graph $G=(V_G,\vec{E}_G)$ if $V_U\subset V_G$ and
$\vec{E}_U=\big\{(x,y),~~x,y\in V_U\big\}\subset \vec{E}_G.$
\end{itemize}
\begin{rem}
A subgraph is a part of $G$ but the converse is not true, for example let us give the following undirected graphs,
see the Figure \ref{exp}.
\begin{figure}[ht]
\begin{center}
\includegraphics*[height=2cm,width=9cm]{exp}
\end{center}
\caption{$H$ part of $G$}
\label{exp}
\end{figure}
\\
$H$ is a part of $G$, but not a subgraph.
\end{rem}
\begin{rem}
A confusion between a subgraph and a part of a graph can create a false interpretation on the
monotonicity of eigenvalue.
\end{rem}
\subsection{Monotonicity relative to vertices } We study the monotonicity of eigenvalues under the variation of the set of vertices.\\
By the Courant-Fischer-Weyl principle we establish the following statement.
\begin{theo}\label{thm1}
Let $H=(V_H,\vec{E}_H)$ be a connected subgraph of a graph $G=(V_G,\vec{E}_G)$, $\sharp V_G=n$, then for any
$1 \leq k\leq \sharp V_H=r$ :
$$\lambda_k(S_H) \leq \lambda_{n-r+k}(S_G).$$
\end{theo}
\begin{demo}
Let $f_1,~f_2,...,f_{k}$ be $k$ eigenfunctions associated to $\lambda_1(S_H),~\lambda_2(S_H),...,\lambda_{k}(S_H)$;
and $F=\{f_1,~f_2,...,f_{k},\delta_1,..,\delta_{n-r}\}$, $\delta_1,..,\delta_{n-r}$ are the Dirac measures on $G$ relative to the vertices in $V_G\setminus V_H$.
It is clear that $\dim F=k+n-r$.
Then using (\ref{fish}), we obtain ;
$$ \lambda_{n-r+k+1}(S_G) \geq \min_{\varphi\in F^\perp\setminus\{0\}}\dfrac{(S_G \varphi,\varphi)_m
}{(\varphi,\varphi)_m }$$
hence $\exists\varphi_k\in F^\perp$ with support in $H$ such that
\begin{align*}
\lambda_{n-r+k+1}(S_G)( \varphi_k,\varphi_k)_m &\geq(S_G \varphi_k,\varphi_k)_m\\
&=(S_H \varphi_k,\varphi_k)_m\\
&\geq \lambda_{k+1}(S_H)(\varphi_k,\varphi_k)_m.
\end{align*}
\end{demo}
For studying the behavior of eigenvalues relative to perturbations, we propose a special construction of graphs.
\begin{defi}
Let $G=(V_G,\vec{E}_G)$ be a weighted graph. $G$ is called $H$-flower-like with respect to the
subgraph $H=(V_H,\vec{E}_H)$ of $G$ if there exists $(H_i)_{i\in I}$ a family
of subgraphs of $G$ such that:
\begin{enumerate}
\item $V_G=V_{H}\cup (\displaystyle{\uplus_i V_{\overset{\circ}{{H_i}}}})$
\item $\forall ~i,j\in I,~~~x\in V_{\overset{\circ}{{H_i}}},~~y\in V_{\overset{\circ}{{H_j}}}, ~i\neq j~\Rightarrow~x\notin V_y$.
\item $\forall~i\in I,~\exists ~x_i\in V_G,~V_H\cap V_{H_i}=\{x_i\}$.
\end{enumerate}
\end{defi}
\begin{figure}[ht]
\begin{center}
\includegraphics*[height=6cm,width=7cm]{fig}
\end{center}
\caption{H-flower-like graph}
\end{figure}
These special graphs are used to create a rule of monotonicity of $\lambda_k$ under a given graph perturbation.
\begin{theo} \label{AA}
Let $G=(V_G,\vec{E}_G)$ be a $H$-flower-like graph, we have then for any $1\leq k\leq \sharp V_H=r$,
\begin{equation}\label{Hfl}
\lambda_k(S_H) \geq \lambda_k(S_G).
\end{equation}
\end{theo}
\begin{demo}
We use the variational principle (\ref{rai}). Let be $f_1,~f_2,...,f_{k-1}$ eigenfunctions of $G$ in $\mathcal{C}_m(V)$
associated to $\lambda_1(S_G),~\lambda_2(S_G),...,\lambda_{k-1}(S_G)$,
and $g_j$ be the eigenfunction associated to $\lambda_j(S_H)$, for $j=1,..,k$. We define a function $\phi_j$ on $V_G$ by:
\begin{equation*}
\phi_j (x)=
\left\{
\begin{aligned}
& g_j (x)~~~~~~~~~~~~~~\text{ if } x\in V_H\\
& g_j (x_1)~~~~~~~~~~~~\text{ if } x\in V_{H_1}\setminus\{x_1\}\\
& . \\
& . \\
& g_j (x_r)~~~~~~~~~~~~\text{ if } x\in V_{H_r}\setminus \{x_r\}
\end{aligned}
\right.
\end{equation*}
where $\{x_s\}=V_H\cap V_{H_s}, ~\forall s\in I=\{1,..,r\}$\\
Let $F=Span\{\phi_j\}_{1\leq j \leq k}$, $\dim F=k$ so, there exist $k$ reals $\alpha_1,\alpha_2,...,\alpha_k$
not all equal to zero, satisfying:
$$\Big(\sum_{j=1}^{k}\alpha_j \phi_j~,~f_l\Big)_m=0~~~~~\forall~~ 1 \leq l \leq k-1.$$
Therefore the function $\psi(x)=\displaystyle{\sum_{j=1}^{k}}\alpha_j \phi_j(x)\neq 0 $ is orthogonal
to $f_l$ for all $1 \leq l \leq k-1$, we define for $x\in H$, $\psi'(x)=\displaystyle{\sum_{j=1}^{k}}\alpha_j g_j(x)$.
Then we get :
\begin{align*}
\lambda_k(S_G)(\psi,\psi)_m & \leq (S_G \psi,\psi)_m \\
&=(S_H \psi',\psi')_m ~~\text{ because $\psi$ is constant on each $H_i$ }\\
&\leq\lambda_k(S_H)(\psi',\psi')_m
\end{align*}
as
\begin{align*}
(\psi,\psi)_m~=~ & \sum_{x\in V_G}m(x)|\psi(x)|^{2}\\
\geq~& \sum_{x\in V_H}m(x) |\psi'(x)|^{2}\\
=~&(\psi',\psi')_m
\end{align*}
then
$$\lambda_k(S_G)(\psi',\psi')_m ~~\leq ~~\lambda_k(S_H)(\psi',\psi')_m.$$
\end{demo}
The above results have several important consequences, for instance on a tree seen as a flower-like graph.
\begin{coro}
Let $G^{s}$ be a simple symmetric tree and $2q=\displaystyle{\max_{x\in V}v(x)}$ then the eigenvalues
of $G^{s}$ satisfy for all $k=1,..,q$:
$$\lambda_k(S_{G^{s}})\leq 2$$
and
$$\lambda_{q+1}(S_{G^{s}})\leq 2(q+1).$$
\end{coro}
\begin{demo}
Clearly there exists a symmetric star graph $S_q$ with $q+1$ vertices seen as a subgraph of $G^{s}$. It can be
considered as a $S_q$-flower-like graph. Therefore $G^{s}$ satisfies the assumptions of the Proposition \ref{AA}.
Hence we get the result because the spectrum of $S_q$ is : $0,2,...,2,2(q+1)$, see \cite{T.M}.
\end{demo}
In the following, we will show more general results : instead of adding only one vertex and one edge,
we would also add a graph
\begin{coro}
Let $G$ be a graph with $n$ vertices, let $G_1$ connected to G by a single edge
Then for $k=1,..,n$;
$$\lambda_k(S_G) \geq \lambda_k(S_{G_1}).$$
\end{coro}
\begin{rem}
The previous corollary is an immediate consequence of the Proposition \ref{AA}. In addition this is an
interesting generalization of Proposition 2 in \cite{Kuras} which shows that $\lambda_2(S_G) \geq \lambda_2(S_{G_1})$
where $G_1$ is obtained from $G$ by adding one edge between one vertex of $G$ and one new vertex.
\end{rem}
\subsection{Monotonicity relative to edges}
We apply the Weyl and interlacing Theorems for matrices to study how the spectrum of the special
operator $S_G$ of a directed graph changes under adding an edge or a set of edges.\\
Now we have several opportunities to refer to the following basic observation about subspace intersections
(see \cite{H.J} page 235).
\begin{lemm}\label{lemm}
Let $W$ be a finite dimensional vector space and let $S_1$, $S_2$,..,$S_k$ be subspaces of $W$, if
$$\delta=\dim S_1+...+\dim S_k -(k-1) \dim W \geq 1 $$
then~~$\dim (S_1\cap ..\cap S_k)\geq \delta \text{ and hence } S_1\cap ..\cap S_k \text{ contains } \delta$
linearly independent vectors, in particular , it contains a nonzero vector.
\end{lemm}
\begin{obs}
Let $S_G$ be the special self-adjoint operator with eigenvalues $\lambda_1(S_G) \leq \lambda_2(S_G)\leq...
\leq\lambda_n(S_G)$. Then the ordered eigenvalues of $-S$ are
$\lambda_1(-S_G) \leq \lambda_2(-S_G)\leq...\leq\lambda_n(-S_G)$, that is, $\lambda_k(-S_G)=-\lambda_{n-k+1}(S_G),~k = 1, . . . , n.$
\end{obs}
We show that the $k^{th}$ eigenvalue $\lambda_k(S_G)$ is monotonously increasing functions of the set of edges.
The following results are the generalizations of the Proposition 1 and the Proposition 2 in \cite{Kuras}.
\begin{prop} \label{Aaa}
Let $G=(V,\vec{E})$ be a connected finite weighted graph $(\sharp V=n)$, consider the partial
graphs $G_1=(V,\vec{E}_1)$ and $G_2=(V,\vec{E}_2)$ where $E=E_1\sqcup E_2$ (disjoint union).
Then for all $k=1,...,n$ and $r,s=1,2$, $r\neq s$:
\begin{equation}\label{equ}
\lambda_k(S_G) \leq \lambda_{k+j}(S_{G_s})+ \lambda_{n-j}(S_{G_r})~~~~j=0,...,n-k
\end{equation}
and
$$\lambda_k(S_G) \geq \lambda_{k-j+1}(S_{G_s})+ \lambda_{j}(S_{G_r})~~~~j=1,...,k.$$
\end{prop}
\begin{demo}
Let $g_k$,~ $h_k$, and $f_k$ be the eigenfunctions associated to $\lambda_k(S_{G_1})$, $\lambda_k(S_{G_2})$
and $\lambda_k(S_G)$ respectively for $k=1,..,n$. \\
For $j=0,...,n-k$, we define $S_1=Span\{g_1,...,g_{k+j}\}$, $S_2=Span\{h_1,...,h_{n-j}\}$ and $S_3=Span\{f_k,...,f_n\}$
by the Lemma (\ref{lemm}) there exists a non zero function $\psi$ in $S_1\cap S_2 \cap S_3$
and so we will have
\begin{align}\label{pos}
\lambda_k(S_G)(\psi,\psi)_m &\leq (S_{G} \psi,\psi)_m\\
&=\sum_{(x,y)\in \vec{E}}b(x,y)\big|\psi(x)-\psi(y)\big|^{2}\notag\\
&= (S_{G_s} \psi,\psi)_{m}+(S_{G_r} \psi,\psi)_{m}\notag\\
&\leq \lambda_{k+j}(S_{G_s})(\psi,\psi)_{m}+ \lambda_{n-j}(S_{G_r})(\psi,\psi)_{m}\notag\\
&\leq\big( \lambda_{k+j}(S_{G_s})+ \lambda_{n-j}(S_{G_r})\big)(\psi,\psi)_m\notag.
\end{align}
In the following we apply the equality (\ref{equ}) to the operator $-S_{G} $ because the inequalities
(\ref{pos}) does not depend on positivity of $b$, and by:
$$\lambda_{k}(-S_G)=-\lambda_{n-k+1}(S_G).$$
We obtain by re-indexing:
$$\lambda_k(S_G) \geq \lambda_{k-j+1}(S_{G_s})+ \lambda_{j}(S_{G_r})~~~~j=1,...,k.$$
\end{demo}
We can easily deduce,
\begin{coro} \label{Aa}
Let $G=(V,\vec{E})$ be a connected finite graph, $G_1=(V,\vec{E}_1)$ and $G_2=(V,\vec{E}_2)$ two partial
graphs of $G$ where $\sharp V=n \text{ and }E=E_1\sqcup E_2$, then for all $k=1,...,n$ and
$r,s=1,2$, $r\neq s$:
$$\lambda_{k}(S_{G_s}) \leq \lambda_k(S_G) \leq \lambda_{k}(S_{G_s})+ \lambda_{n}(S_{G_r}).$$
\end{coro}
\begin{demo}
By applying the proposition \ref{Aaa} to $j=0$ and $j=1$ respectively, we obtain the result because
$\lambda_{1}(S_{G_r})=0.$
\end{demo}
In other words, adding a subset of edges to $\vec{E}$ while keeping the same set of vertices always induces
an increasing of the $k^{th}$ eigenvalue or keeps it unchanged.
\begin{coro}\label{cor}
Let $G=(V,\vec{E})$ be a connected finite graph with $n$ vertices, and $G_1$ a graph obtained by adding a set
of edges to $G$ then for all $k=1,...,n$:
$$\lambda_{k}(S_G) \leq \lambda_k(S_{G_1}).$$
\end{coro}
\section{Comparison eigenvalues of Dirichlet Laplacian on graphs}\label{comp}
In this section, we present some results about the spectrum comparison between the Laplacian and the Dirichlet Laplacian.
The purpose of this part is to find the relation between the usual vertex weight on a subgraph $H$ of $G$ and its
boundary weight to compare eigenvalues.\\
This is done by establishing a clear and explicit link between the eigenvalues and the Dirichlet eigenvalues on $G$.
In the following proposition, we treat the Dirichlet Laplacian case :
$$\forall x\in U,~~S^D_U f(x)=\frac{1}{m(x)}\sum_{y\in V_x\atop y\in V_G}a(x,y)\big(f(x)-f(y)\big).$$
By the same techniques used in the Lemma \ref{delta}, we can show the following Lemma.
\begin{lemm}
$$\lambda_1(S^D_U)\leq 2\mathcal{R}e(\lambda_1(\Delta^D_U))$$
and
$$\lambda_n(S^D_U)\geq 2\mathcal{R}e(\lambda_n(\Delta^D_U)).$$
\end{lemm}
\begin{demo}
Let $f$ and $g$ be eigenfunctions associated to $\lambda_1(\Delta^D_U)$ and $\lambda_n(\Delta^D_U)$ respectively.
By the variational principle of $\lambda_1(S^D_U)$ and $\lambda_n(S^D_U)$, we have
$$\lambda_1(S^D_U)\leq\dfrac{(S^D_U f,f)_m}{(f,f)_m}=\dfrac{(\Delta^D_U f,f)_m}{(f,f)_m}+
\dfrac{\overline{(\Delta^D_U f,f)}_m}{(f,f)_m}=\lambda_1(\Delta^D_U)+\overline{\lambda_1(\Delta^D_U)}$$
and
$$\lambda_n(S^D_U)\geq\dfrac{(S^D_U g,g)_m}{(g,g)_m}=\dfrac{(\Delta^D_U g,g)_m}{(g,g)_m}
+\dfrac{\overline{(\Delta^D_U g,g)}_m}{(g,g)_m}=\lambda_n(\Delta^D_U)+\overline{\lambda_n(\Delta^D_U)}$$
\end{demo}
In the same spirit as the Cauchy interlacing theorem concerning hermitian bordered matrices (see [9]
theorem 4.3.28 for a generalized statement) one can prove the following.
\begin{prop}\label{thm2}
Consider a connected subgraph $H=(V_H,\vec{E}_H)$ of $G=(V_G,\vec{E}_G)$, $(\#V_G=n, \#V_H=r)$,
then the eigenvalues on $G$ of the normalized Laplacian $S_G$ satisfies:
$$\lambda_k(S^D_H) \leq \lambda_{k+n-r}(S_G).$$
\end{prop}
\begin{demo}
Let $h_1,..,h_r$ and $f_1,..,f_n$ be the eigenfunctions associated to $S^{D}_H$ and $S_G$ respectively,
define the function $g_k$ for $k=1,..,r$ by:
\begin{equation*}
g_k=~\left\{
\begin{aligned}
&h_k~~~~on~~V_H \\
&0~~~~otherwise
\end{aligned}
\right.
\end{equation*}
Let $1\leq k\leq r$ and fix $S_1=Span\{g_k,...,g_{r}\}$ and $S_2=Span\{f_1,...,f_{k+n-r}\}$
by the Lemma \ref{lemm} there exists a function $\psi$ in $S_1\cap S_2$. Since $\psi \in S_1$, it has the form
\begin{equation*}
\psi=~\left\{
\begin{aligned}
&g~~~~on~~V_H \\
&0~~~~otherwise
\end{aligned}
\right.
\end{equation*}
for some $g\in Span\{g_k,...,g_{r}\}$. Observe that:
\begin{align*}
\lambda_k(S^D_H)( g,g)_{m} &\leq(S^{D}_{H} g,g)_{m}\\
&\leq (S_G \psi,\psi)_{m}\notag\\
\end{align*}
since $( g,g)_{m}=(\psi,\psi)_{m}$, we get
$$\lambda_k(S^D_H)\leq\dfrac{(S_G\psi,\psi)_{m}}{(\psi,\psi)_{m}}$$
then,
$$\lambda_k(S^D_H) \leq \lambda_{k+n-r}(S_G).$$
\end{demo}
We deduce easily from the Proposition \ref{thm2} an estimation of the eigenvalues of $S_G$ and $\Delta_G$,
thanks to Theorem \ref{thm1}, as follows:
\begin{coro}
Consider a connected subgraph $H=(V_H,\vec{E}_H)$ of $G=(V_G,\vec{E}_G)$, $(\#V_G=n, \#V_H=r)$,
then the eigenvalues on $G$ satisfies:
$$ \lambda_{k+n-r}(S_G)\geq \max\big(\lambda_k(S_H),\lambda_k(S^D_H)\big).$$
\end{coro}
\begin{coro}
Consider a connected subgraph $H=(V_H,\vec{E}_H)$ of $G=(V_G,\vec{E}_G)$, $(\#V_G=n, \#V_H=r)$,
then:
$$ \lambda_{n}(S_G)\geq \max\Big(2\mathcal{R}e\big((\lambda_r(\Delta_H)\big),2\mathcal{R}e\big(\lambda_r(\Delta^D_H)\big)\Big).$$
\end{coro}
\begin{coro}
Consider a connected subgraph $H=(V_H,\vec{E}_H)$ of the cycle graph $C_n$, $(\#V_H=r)$,
then:
$$\mathcal{R}e\big((\lambda_n(\Delta_{C_n})\geq \max\Big(\mathcal{R}e\big(
(\lambda_r(\Delta_{H})\big),\mathcal{R}e\big(\lambda_r(\Delta^D_{H})\big)\Big).$$
\end{coro}
In the following Proposition we prove how to give an upper bound for $\lambda_{k+l}(S_G)$ in
terms of a Rayleigh quotient. We give a discrete version of the Proposition 2.1 \cite{B.B} by applying
the Poincar\'e min-max principle. The methods we use follow closely the arguments given in B. Benson \cite{B.B}
in the case of the Laplacian of Riemannian manifolds. Next we provide an upper bound of the eigenvalues of $G$
according to Dirichlet eigenvalues on such repartition of $G$ as in the Figure \ref{partition}.
\begin{prop}
Let $G=(V,\vec{E})$ a finite connected graph, $U=(V_U,\vec{E}_U)$ a part of $G$ and $A=(V_A,\vec{E}_A)$,
$B=(V_B,\vec{E}_B)$ two subgraphs satisfying the following conditions:
\begin{enumerate
\item $V=V_A\sqcup V_B=V_{\overset{\circ}{A}}\sqcup V_{\overset{\circ}{B}}\sqcup V_U$ ~~~(\text{ disjoint union})
\item $\vec{E}=\vec{E}_A\sqcup\vec{E}_B\sqcup \vec{E}_U$
\item $\partial_{\vec{E}} A=\partial_{\vec{E}} B=\vec{E}_U$.
\end{enumerate}
\begin{figure}[ht]
\begin{center}
\includegraphics*[height=5cm,width=7cm]{repart}
\end{center}
\caption{Repartition of $G$}
\label{partition}
\end{figure}
Then we get for $1 \leq k,l \leq \min(\sharp V_{\overset{\circ}{A}},\sharp V_{\overset{\circ}{B}})$:
$$\lambda_{k+l}(S_G)\leq \max \big(\lambda_{k}(S^D_{\overset{\circ}{A}}),\lambda_l(S^D_{\overset{\circ}{B}})\big).$$
\end{prop}
\begin{demo}
Let $h_{k}$ and $g_l$ be the eigenvectors associated to $\lambda_{k}(S^D_{\overset{\circ}{A}})$ and
$\lambda_l(S^D_{\overset{\circ}{B}})$ respectively, for $1 \leq k \leq \sharp V_{\overset{\circ}{A}}$ and $1\leq l\leq\sharp V_{\overset{\circ}{B}}$.\\
We define $f$ on $\mathcal{C}_{m}(\overset{\circ}{A}\oplus\overset{\circ}{B})$ by:
\begin{equation*}
f(x)=~\left\{
\begin{aligned}
&h_{k}(x)~~~~~if~x\in V_{\overset{\circ}{A}} \\
&g_l(x)~~~~~~~~if~x\in V_{\overset{\circ}{B}} \\
&0~~~~~~~~~~~~~if~x\in V_U.
\end{aligned}
\right.
\end{equation*}
We have
$$(S f,f)_{m}=
(S^D_{\overset{\circ}{A}} h_{k},h_{k})_{m}+(S^D_{\overset{\circ}{B}} g_l,g_l)_{m}.$$
Using the Poincar\'e min-max principle (\ref{poin}), we obtain:
\begin{align*}
\lambda_{k+l}(S_G)\leq\max_{f\in F_{\overset{\circ}{A}}+F_{\overset{\circ}{B}}}\dfrac{(S_G f,f)_{m}}{( f,f)_{m}}& =\max_{h\in F_{\overset{\circ}{A}},g\in F_{\overset{\circ}{B}}}\dfrac{ (S^D_{\overset{\circ}{A}} h,h)_{m}+(S^D_{\overset{\circ}{B}} g,g)_{m}}{( h,h)_{m}+(g,g)_m}\\
&\leq \dfrac{ \lambda_{k}(S^D_{\overset{\circ}{A}})(h,h)_{m}+\lambda_l(S^D_{\overset{\circ}{B}})( g,g)_{m}}{( h,h)_{m}+(g,g)_m}\\
&\leq \max\big(\lambda_{k}(S^D_{\overset{\circ}{A}}),\lambda _l(S^D_{\overset{\circ}{B}})\big).
\end{align*}
Hence
$$\lambda_{k+l}(S_G)\leq \max\big(\lambda_{k}(S^D_{\overset{\circ}{A}}),\lambda _l(S^D_{\overset{\circ}{B}})\big).$$
\end{demo}
\begin{rem}
The previous Proposition remains true in the particular case of the Laplacian $\tilde{S}_G$.
\end{rem}
An estimate of $\lambda_{2}(S_G)$ can also be obtained with respect to $\lambda_{1}(\Delta^D_{\overset{\circ}{B}})$
and $\lambda_{1}(\Delta^D_{\overset{\circ}{B}})$.
\begin{coro}
Under the same hypothesis of the previous Proposition we have
$$\lambda_{2}(S_G)\leq \max \Big(2\mathcal{R}e\big(\lambda_{1}(\Delta^D_{\overset{\circ}{A}})\big),2\mathcal{R}e\big(\lambda_1(\Delta^D_{\overset{\circ}{B}})\big)\Big).$$
\end{coro}
\begin{coro}
Under the same hypothesis of the previous Proposition, from the cycle graph $C_n$, we have
$$\mathcal{R}e\big(\lambda_{2}(\Delta_{C_n})\leq \max \Big(\mathcal{R}e\big(\lambda_{1}(\Delta^D_{\overset{\circ}{A}})\big),\mathcal{R}e\big(\lambda_1(\Delta^D_{\overset{\circ}{B}})\big)\Big).$$
\end{coro}
\textbf{Acknowledgement}: I take this opportunity to express my gratitude to
my PhD advisors Colette Ann\'e and Nabila Torki-Hamza for all the fruitful discussions,
helpful suggestions and their guidance during this work. This work was financially supported by
the "PHC Utique" program of the French Ministry of Foreign Affairs and Ministry of higher education
and research and the Tunisian Ministry of higher education and scientific research in the CMCU project
number 13G1501 " Graphes, G\'eom\'etrie et th\'eorie Spectrale". Also I would like to thank the Laboratory
of Mathematics Jean Leray of Nantes (LMJL) and the research unity (UR/13ES47) of Faculty of Sciences of
Bizerte (University of Carthage) for its financial and its continuous support.
| 855965b2e171207d3ad8a9d64d58e88989ef6517 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{Sec:Into}
Consider the Brownian excursion standardized to have length 1 and conditioned to be positive.
This is also the Brownian bridge conditioned to stay positive, or the 3-dimensional Bessel process conditioned to hit zero at
$t=1$, and its %
finite-dimensional distributions are given by formula \eqref{eq:RevuzYor} below.
We let $\mathbb B^{ex} =(\mathbb B^{ex}_t)_{t\in[0,1]}$ denote this process throughout.
Brownian excursion has been extensively investigated in the literature. See for example \citet{bertoin94path} and
\citet{revuz99continuous}. It also appears in asymptotic analysis
of various combinatorial problems, see for example
\citet{pitman06combinatorial} and \citet{janson2007brownian}.
The purpose of this note is to introduce a ``dual representation" that ties the
Laplace transforms of finite-dimensional distributions of
Brownian excursion and another Markov process, denoted by $(X_t)_{t\geq 0}$ throughout,
with
state space $[0,\infty)$ and transition probabilities
\begin{equation*}
\mathds{P}(X_t \in dy\mid X_s =x)=p_{t-s}(x,y) dy ,\; 0\leq s<t, x\geq 0
\end{equation*}
with
\begin{equation}
\label{X-transitions}
p_t(x,y) =\frac{2 t \sqrt{y}}{\pi
\bb{ (y-x)^2+2 (x+y)t^2+t^4
}},\;t>0, y\geq 0.
\end{equation}
This is a positive self-similar Markov process
that arises as the tangent process at the boundary of support of
so-called $q$-Brownian motions and $q$-Ornstein--Uhlenbeck processes; see \citet[Proposition 2.2]{Bryc-Wang-2015} and \citet[Theorem 3.1]{wang2016extremes}.
It can also be obtained from the construction in \citet{biane98processes} applied to the
1/2-stable free L\'evy process
by including appropriate drift. The
derivation of the transition probability density function, following Biane's approach,
can be found in \citet[Section 3]{Bryc-Wang-2015}.
Our main result is the following
identity which was needed in \cite{Bryc-Wang-2017ASEP},
where we investigated, by essentially computing the Laplace transforms, the fluctuations of asymmetric simple exclusion processes with open boundaries
in the steady state. These processes are representative non-equilibrium models
\cite{derrida2007non,derrida2004asymmetric}
that have attracted much
attentions recently in probability and mathematical physics.
Let $\EEx{x}{\cdot}$ denote the expectation with respect to the law of $(X_t)_{t\ge0}$ starting at
$X_0 =
x>0$.
\begin{theorem}
\label{T1.5}
For $d\in\mathbb N$, let $s_0=0<s_1<s_2<\dots<s_d$ and
$t_0=0\le t_1<\dots<t_d\le 1=t_{d+1}$. Then,
\begin{multline}
\label{Excursion-dual}
\EE{\exp\pp{-\summ k1d (s_k-s_{k-1})\mathbb B^{ex}_{t_k}}} \\=
\frac1{\sqrt{2\pi}} \int_0^\infty \EEx{x}{\exp \pp{-\frac12\sum_{k=0}^{d}
(t_{k+1}-t_{k})X_{s_k}}}\sqrt{x}\,dx.
\end{multline}
\end{theorem}
The left-hand side of~\eqref{Excursion-dual} is the Laplace transform of the joint distribution of Brownian excursion, while
the right-hand side is the Laplace transform of the
process $(X_t)_{t\geq 0}$
with the arguments and time indices
interchanged. %
On the right-hand side, the initial distribution of
$(X_t)_{t\geq 0}$
is the stationary
$\sigma$-finite measure $(x/(2\pi))^{1/2}\indd{x>0}dx$.
We are aware of only a couple of results that connect Laplace transforms of stochastic processes by interchanging the argument and time parameters. One such result is
the formula for the joint generating function of the finite Asymmetric Simple Exclusion Process in \cite[Theorem 1]{Bryc-Wesolowski-2015-asep}.
Another result of this type is the formula \citet[Eq.~(2)]{Bertoin-Yor2001subordinators} for the univariate Mellin transform of a positive self-similar Markov process,
see also \citet{hirsch2012temporally}.
As an immediate consequence of Theorem~\ref{T1.5} however, a family of such dualities can be derived easily for {\em
generalized Brownian meanders} defined in \eqref{eq:Gen-meander} below, including some of the generalized Bessel meanders considered
in \citep[Section 3.7]{Mansuy-Yor-2008}.
The dualities, see Theorem~\ref{T-gen-meander} and Corollary \ref{coro:3}, have
similar forms as in \eqref{Excursion-dual}, but they differ in the choice of the $\sigma$-finite measure for the initial law on
the right-hand side of~\eqref{Excursion-dual}.
Due to its importance,
here we state the duality formula for Brownian meander $(\mathbb B^{me}_t)_{0\leq t\leq 1}$, which is a special case $\delta = 1$ of %
Corollary \ref{coro:3} below. Recall that Brownian meander can be defined as the
Brownian motion conditioned to stay positive over the time interval $[0,1]$. See for example
\citet{bertoin94path,Pitman-1999,pitman06combinatorial}.
We also need this formula for our analysis in \cite{Bryc-Wang-2017ASEP} for asymmetric simple exclusion processes.
\begin{theorem}
\label{coro:excursion}
For
$d\in\mathbb N$, let $s_0=0<s_1<s_2<\dots<s_d$ and $t_0=0\leq t_1<t_2<\dots
<t_d\le 1=t_{d+1}$. Then
\begin{multline}
\label{Meander-dual}
\EE{\exp\left(-\sum_{k=1}^d (s_{k}-s_{k-1})\mathbb B^{me}_{1-t_k}\right)}\\=
\frac1{\sqrt{2\pi}} \int_0^\infty \EEx{x}{\exp \pp{-\frac12\sum_{k=0}^{d}
(t_{k+1}-t_{k})X_{s_k}}}\frac{1}{\sqrt{x}}\,dx .
\end{multline}
\end{theorem}
The paper is organized as follows. In Section~\ref{Sec:Exc} we recall some known facts about Brownian excursion and prove
Theorem~\ref{T1.5}. In Section~\ref{Sec:randomized} we prove the dual representations for
generalized Brownian meanders, with
Brownian meander as a special case.
\section{Proof of dual representation of Brownian excursion}
\label{Sec:Exc}
We first recall some facts on Brownian excursion $\mathbb B^{ex}$.
For our purposes it is convenient to define it as a
Markov process that starts at $\mathbb B^{ex}_0=0$, ends at $\mathbb B^{ex}_1=0$,
and has transition probabilities
\begin{equation*}\label{Transitions-Excursion}
P(\mathbb B^{ex}_t\in dy\mid \mathbb B^{ex}_s=x)=\begin{cases}
\sqrt{8\pi} \ell_t(y)\ell_{1-t}(y) &\mbox{ if $s=0<t<1, x=0, y> 0$}\\
\\
\displaystyle g_{t-s}(x,y)\frac{\ell_{1-t}(y)}{\ell_{1-s}(x)}
&\mbox{ if $0<s<t<1, x>0,y> 0$},\\
\end{cases}
\end{equation*}
with
\begin{equation*}
\label{RY-ell}
\ell_t(y) = \frac1{\sqrt{2\pi t^3}}\cdot y\exp\pp{-\frac{y^2}{2t}}\indd {y>0}
\end{equation*}
and
\begin{equation*}
\label{RY-q}
g_t(y_1,y_2) = \frac1{\sqrt{2\pi t}}\bb{\exp\pp{-\frac{(y_1-y_2)^2}{2t}} - \exp\pp{-\frac{(y_1+y_2)^2}{2t}}}\indd{y_1,y_2>0}.
\end{equation*}
Equivalently, the joint probability density function at time points $0<t_1<\cdots<t_d<1$ is
\begin{equation}
\label{eq:RevuzYor}
f_{t_1,\dots,t_d}(y_1,\dots,y_d) = \sqrt{8\pi}\ell_{t_1}(y_1) \ell_{1-t_d}(y_d)\prodd k1{d-1}g_{t_{k+1}-t_k}(y_k,y_{k+1}).
\end{equation}
See %
\citet[page 76]{ito125jr}, \citet{durrett1977weak}, %
or
\citet[page 464]{revuz99continuous}.
\begin{proof}[Proof of Theorem~\ref{T1.5}]
The proof consists of rewriting the right-hand side of \eqref{Excursion-dual}.
We first assume $t_1>0$ and $t_d<1$.
With $t_0=0$, we have, using transition density functions of the process $X$ and letting $d\vec x$ denote $dx_0\cdots dx_d$,
\begin{align}
\int_0^\infty &\EEx{x_0}{\exp \pp{-\frac12\sum_{k=0}^d
(t_{k+1}-t_{k})X_{s_k}}}\sqrt{x_0}\,dx_0 \nonumber\\
&= \int_{{\mathbb R}_+^{d+1}}d\vec x \sqrt{x_0} \exp\pp{-\frac12\summ k0d (t_{k+1}-t_{k})x_k}\prodd k1d p_{s_k-s_{k-1}}(x_{k-1},x_{k})\nonumber
\\
&= 2^{d+1}\int_{{\mathbb R}_+^{d+1}}d\vec x \exp\pp{-\frac12\summ k0d (t_{k+1}-t_{k})x_k^2} \times x_0^2 \prod_{k=1}^d x_kp_{s_k-s_{k-1}}(x_{k-1}^2,x_{k}^2),\label{iiint}
\end{align}
where we apply in the last equality changes of variables from $x_k$ to $x^2_k$, $k=0,\dots,d$.
We derive a different expression of the products of transition density functions, thanks to the factorization. Namely, recalling~\eqref{X-transitions}, we write
\begin{align*}
p_s(x^2,y^2) &=\frac{2s y}{\pi (s^2+(y-x)^2)(s^2+(y+x)^2)}\\
&=\frac{1}{2\pi x}\left(\frac{s}{s^2+(y-x)^2}-\frac{s}{s^2+(y+x)^2} \right).\end{align*}
We now use the elementary Laplace transform
$$
\frac{s}{s^2+a^2}=\int_0^\infty e^{-s y} \cos (a y) dy$$
to get
\begin{align*}
y p_s(x^2,y^2) &=\frac{y}{2\pi x}\int_0^\infty e^{-s z} \bb{\cos ((y-x) z)-\cos ((y+x) z)}dz\\
&= \frac{1}{\pi} \frac{y}{ x}\int_0^\infty e^{-s z} \sin (xz)\sin(yz)dz.
\end{align*}
Therefore, writing $d\vec y = dy_1\cdots dy_d$,
\begin{multline*}
x_0^2 \prod_{k=1}^d x_kp_{s_k-s_{k-1}}(x_{k-1}^2,x_{k}^2) = \frac1{\pi^d}\int_{{\mathbb R}_+^d}d\vec y \, \exp\pp{-\summ k1d(s_k-s_{k-1})y_k} \\
\times x_0\sin(x_0y_1) \times x_d \sin(x_dy_d) \times \prodd k1{d-1}\sin(x_ky_{k+1})\sin(x_ky_k).
\end{multline*}
Then, noticing that the integrand is non-negative and interchanging the order of integrations, we rewrite the right-hand side
of \eqref{iiint} as %
\begin{align}
\frac{2^{d+1}}{\pi^d} & \int_{{\mathbb R}_+^{d}}d\vec y \,\exp\pp{-\summ k1d(s_k-s_{k-1})y_k}\label{eq:interchanging}\\
& \times \int_0^\infty x_0e^{-t_1x_0^2/2}\sin(x_0y_1)dx_0
\times \int_0^\infty x_de^{-(1-t_d)x_d^2/2}\sin(x_dy_d)dx_d\nonumber\\
&\times \prodd k1{d-1}\int_0^\infty e^{-(t_{k+1}-t_k)x_k^2/2}\sin(x_ky_{k+1})\sin(x_ky_k)dx_k.\nonumber
\end{align}
By straightforward calculation we have
\begin{equation*}\label{eq:sin}
\int_0^\infty x e^{-t x^2/2}\sin (xy) dx = \frac yt\int_0^\infty e^{-tx^2/2}\cos(xy)dx = \pi\ell_t(y).
\end{equation*}
In the last step above we used the %
formula for the cosine transform
\begin{equation*}\label{eq:cos}
\int_0^\infty e^{-tx^2/2}\cos(ax)dx
= \sqrt{\frac\pi2}\frac1{t^{1/2}}e^{-a^2/(2t)},
\end{equation*}
which follows from the characteristic function of a Gaussian distribution with mean zero and variance $1/t$. We also have
\begin{align*}
\int_0^\infty & e^{-t x^2/2}\sin(xy_1)\sin(xy_2)dx \\
\ & = \frac12\int_0^\infty e^{-tx^2/2} \bb{\cos(x(y_1-y_2))-\cos(x(y_1+y_2))}dx\\
&= \sqrt{\frac\pi8}\frac1{t^{1/2}}\pp{e^{-(y_1-y_2)^2/(2t)} - e^{-(y_1+y_2)^2/(2t)}} = \frac\pi2 g_t(y_1,y_2).
\end{align*}
Thus, recalling~\eqref{eq:RevuzYor}, expression~\eqref{eq:interchanging} now becomes,
\begin{multline*}
4\pi\int_{{\mathbb R}^d}d\vec y\, \exp\pp{-\summ k1d(s_k-s_{k-1})y_k}
\ell_{t_1}(y_1)\ell_{1-t_d}(y_d)\prodd k1{d-1}g_{t_{k+1}-t_k}(y_k,y_{k+1}) \\
= \sqrt{2\pi}\EE{\exp\pp{-\summ k1d (s_k-s_{k-1}) \mathbb B^{ex}_{t_k}}}.
\end{multline*}
We have proved the theorem with $0<t_1,t_d<1$.
Now assume $t_d = 1$ and $t_1>0$. %
Recall that $\mathbb B^{ex}_1 = 0$. If $d=1$, then the left-hand side of~\eqref{Excursion-dual} becomes 1, and the right-hand side becomes
$(\sqrt{2\pi})^{-1}\int_0^\infty e^{-x/2}\sqrt xdx = 1$.
If $d>1$, %
we see that the desired identity is reduced to the same type of identity with
fewer arguments, namely with $0<s_1<\cdots<s_{d-1}$ and $0<t_1<\cdots<t_{d-1}<1 = t_d$. Such an identity has been proved in the
first part of the proof.
It remains to prove
the theorem
for $t_1 = 0$. If $d=1$, then the left-hand side
of~\eqref{Excursion-dual} is 1 and the right-hand side equals
\[
\frac1{\sqrt{2\pi}}\int_0^\infty \EEx{x}{e^{-X_{s_1}/2}}\sqrt xdx
=
\frac1{\sqrt{2\pi}}\int_0^\infty e^{-x/2}\sqrt xdx = 1,
\]
where we used the fact that $X$ is a Markov process with stationary distribution $\sqrt x \indd{x\ge 0}dx$ (as $\sqrt xp_s(x,y) = \sqrt y p_s(y,x)$ for $x,y\ge 0$, $s>0$).
For $d\ge 2$, consider
$\wt s_k := s_{k+1}-s_1$, $\wt t_k := t_{k+1}, k=0,\dots,d-1$ and $\wt t_d := 1$. In this way, the left-hand side of~\eqref{Excursion-dual} becomes
$\mathds E\sbb{\exp\spp{-\summ k1{d-1} (\wt s_k-\wt s_{k-1})\mathbb B^{ex}_{\wt t_k}}}$, and
the right-hand side becomes
\begin{multline*}
\frac1{\sqrt{2\pi}} \int_0^\infty \EEx{x}{\exp \pp{-\frac12\sum_{k=0}^{d-1}
(\wt t_{k+1}-\wt t_{k})X_{\wt s_k+s_1}}}\sqrt{x}\,dx\\
= \frac1{\sqrt{2\pi}} \int_0^\infty \EEx{x}{\exp \pp{-\frac12\sum_{k=0}^{d-1}
(\wt t_{k+1}-\wt t_{k})X_{\wt s_k}}}\sqrt{x}\,dx.
\end{multline*}
Above in the last step we used the stationarity of $X$ again.
Since necessarily $\wt t_1
=t_2
>0$, when $d\ge 2$ the desired identity %
\eqref{Excursion-dual}
becomes an identity for Laplace transform of ($d-1$)-dimensional distribution of Brownian excursion at time points $0<\wt t_1<\cdots<\wt t_{d-1}\le 1$ that we have
already
proved before.
This completes the proof.
\end{proof}
\section{Application to generalized Brownian meanders}
\label{Sec:randomized} In this section, we let $\nu(dv)$ be a probability measure on $[0,1]$ such that $\nu(\{0\})=0$, and
consider the
generalized Brownian meander $(\mathbb{B}\topp\nu_t)_{0\leq t\leq 1}$ defined by
\begin{equation}
\label{eq:Gen-meander}
\mathbb{B}\topp\nu_t =\frac1{\sqrt V}\mathbb B^{ex}_{Vt}, \; t\in[0,1],
\end{equation}
where $V$ is a random variable with law $\nu$ and independent from $\mathbb B^{ex}$.
Explicit examples,
including Brownian meander,
will be discussed
later in this section.
Noting that the function $v\mapsto e^{-x/v} v^{-3/2}$ is bounded on $(0,1]$ for $x>0$, we define
\begin{equation}
\label{eq:phi_nu}
\varphi_\nu(x)=\frac{\sqrt{x}e^{x/2}}{\sqrt{2\pi}}\int_{(0,1]} e^{-\tfrac{x}{2v}}v^{-3/2} \nu(dv).
\end{equation}
We have the following dual representation of $\mathbb{B}\topp\nu$. %
\begin{theorem}
\label{T-gen-meander}
For %
$d\in\mathbb N$, let $s_0=0<s_1<s_2<\dots<s_d$ and $t_0=0\leq t_1<t_2<\dots<t_d\le 1=t_{d+1}$.
Then
\begin{multline*}
\EE{\exp\left(-\sum_{k=1}^d (s_{k}-s_{k-1})\mathbb{B}\topp\nu_{1-t_k}\right)}\\=
\int_0^\infty \EEx{x}{\exp \pp{-\frac12\sum_{k=0}^{d}
(t_{k+1}-t_{k})X_{s_k}}}\varphi_\nu(x)dx . %
\end{multline*}
\end{theorem}
\begin{proof}
The key step is the following identity
\begin{multline}
\label{eq:pre-V}
\EE{ \exp \pp{-\summ k1d \frac{s_k-s_{k-1}}{\sqrt v}\mathbb B^{ex}_{v(1-t_k)}} }
\\=
\frac{1}{\sqrt{2\pi}}\int_0^\infty %
\sqrt xe^{x/2}\frac{e^{-\tfrac x{2v}}}{v^{3/2} }
\EEx{x}{\exp \pp{-\frac12\summ k0d (t_{k+1}-t_k)X_{ s_k} }} dx
\end{multline}
for $0<v\leq 1$.
Once we establish~\eqref{eq:pre-V}, integrating with respect to $\nu$ on both sides yields the desired result, by Fubini's theorem.
Now we prove~\eqref{eq:pre-V}.
We first
rewrite
\eqref{Excursion-dual}, with $t_k$ replaced by $1-t_k$, as %
processes $(\mathbb B^{ex}_t)_{t\in[0,1]}$ and $(\mathbb B^{ex}_{1-t})_{t\in[0,1]}$ have the same law. So~\eqref{Excursion-dual} becomes
\begin{multline}\label{aadvark}
\EE{\exp \pp{-\summ k1d (s_k-s_{k-1})\mathbb B^{ex}_{1-t_k}}}\\
= \frac1{\sqrt{2\pi}} \int_0^\infty \EEx{x}{\exp \pp{-\frac12\summ k1d (t_{k+1}-t_k)X_{s_k} - \frac12(1-(1-t_1))X_0}}\sqrt x\, dx.
\end{multline}
This identity remains valid, if we replace the increasing sequence $(s_k)_{k=0,\dots,d}$ with %
$(s_k/\sqrt{v})_{k=0,\dots,d}$ and replace the increasing sequence $(t_k)_{k=1,\dots,d+1}$ with %
$(\wt t_k)_{k=1,\dots,d+1}$ with $\wt t_k := 1-v(1-t_k)$ (formally we just replace $s_k$ by $s_k/\sqrt v$ and $1-t_k$ by $v(1-t_k)$, but we also need to verify the same monotonicity of the replaced sequences, and that $\wt t_1 \ge 0, \wt t_d\le 1$). In this way we get
\begin{multline*}
\EE{\exp \pp{-\summ k1d \frac{s_k-s_{k-1}}{\sqrt v}\mathbb B^{ex}_{v(1-t_k)}}}\\
= \frac1{\sqrt{2\pi}} \int_0^\infty \EEx{x}{\exp \pp{-\frac12\summ k1d v(t_{k+1}-t_k)X_{\frac{s_k}{\sqrt v}} - \frac12(1-v(1-t_1))X_0}}\sqrt x \, dx.
\end{multline*}
Thanks to the self-similarity of process $X$,
\[
((X_{\lambda s})_{s\ge 0}, \mathds{P}_x) \stackrel{d}= \pp{\lambda^2(X_s)_{s\ge 0},
\mathds{P}_{x/\lambda^2}}, x>0, \lambda>0,
\]
where $\mathds{P}_x(\cdot) = \mathds{P}(\cdot\mid X_0 = x)$ (see \citet{Bryc-Wang-2015}), %
%
we have
\begin{align*}
&\int_0^\infty \EEx{x}{\exp \pp{-\frac12\summ k1d v(t_{k+1}-t_k)X_{\frac{s_k}{\sqrt v}} - \frac12(1-v(1-t_1))X_0}}\sqrt xdx\\
& = \int_0^\infty \EEx{vx}{\exp\pp{-\frac12\summ k1d(t_{k+1}-t_k)X_{s_k}-\frac12\pp{{1-v(1-t_1)}}\frac{X_0}v}}\sqrt xdx\\
& = \int_0^\infty \EEx{vx}{\exp\pp{-\frac12\summ k0d(t_{k+1}-t_k)X_{s_k}-\frac12\pp{\frac1v -1}X_0}}\sqrt xdx\\
& = \int_0^\infty \EEx{x}{\exp\pp{-\frac12\summ k0d(t_{k+1}-t_k)X_{s_k}}}e^{-x\pp{1/v -1}/2}\frac{\sqrt x}{v^{3/2}}dx.
\end{align*}
This yields~\eqref{eq:pre-V}, and completes the proof.
\end{proof}
Next, we specialize Theorem~\ref{T-gen-meander} to the case of Beta distribution.
The formula shall involve the confluent hypergeometric function
\[
\psi(\alpha,\beta,x) = \frac1{\Gamma(\alpha)}\int_0^\infty e^{-xu}u^{\alpha-1}(1+u)^{\beta-\alpha-1}du, \alpha>0,
\beta\in\mathds{R},
x>0.
\]
See for example~\citep[page 268]{lebedev65special}.
\begin{corollary}
\label{coro:gen-meander} For %
$d\in\mathbb N$, let $s_0=0<s_1<s_2<\dots<s_d$ and $t_0=0\le t_1<\dots<t_d\le 1=t_{d+1}$. If
$\nu(dv)=v^{\alpha-1}(1-v)^{\beta-1}/B(\alpha,\beta)$ with $\alpha,\beta>0$, then
\begin{multline}
\label{Meander-alpha-dual}
\EE{\exp\left(-\sum_{k=1}^d (s_{k}-s_{k-1})\mathbb B\topp{\nu}_{1-t_k}\right)}
\\
= \frac{\Gamma(\alpha+\beta)}{\sqrt {2\pi}\Gamma(\alpha)} \int_0^\infty \EEx{x}{\exp \pp{-\frac12\sum_{k=0}^{d}(t_{k+1}-t_{k})X_{s_k}}} \psi\pp{\beta,\frac{5}2-\alpha,\frac x2}\sqrt x \,dx.
\end{multline}
\end{corollary}
\begin{proof}
We apply Theorem~\ref{T-gen-meander} with
\begin{align}
\varphi_\nu(x) & =
\frac1{\sqrt{2\pi}}\frac{\sqrt x}{B(\alpha,\beta)}\int_0^1e^{-x(1/v-1)/2}v^{\alpha-5/2}(1-v)^{\beta-1}dv\nonumber\\
\label{eq:d+d'=3}
& = \frac1{\sqrt{2\pi}}\frac{\sqrt{x}}{B(\alpha,\beta)}\int_0^\infty e^{-ux/2}(1+u)^{3/2-(\alpha+\beta)}u^{\beta-1}du\\
& = \frac{\sqrt x\Gamma(\beta)}{\sqrt{2\pi} B(\alpha,\beta)}\psi\pp{\beta,\frac{5}2-\alpha,\frac x2},\nonumber
\end{align}
where in the second step we applied a change of variable $u = 1/v-1$.
\end{proof}
Our results are also related to
generalized Bessel meanders that were introduced in \citet[Sections 3.6 and 3.7]{Mansuy-Yor-2008}.
There are three equivalent ways to define a generalized Bessel meander %
$\mathbb B\topp{\delta,\delta'}$ for $\delta,\delta'>0$.
(This is the process under the law $M^{\delta,\delta'}$ in \citet{Mansuy-Yor-2008}, and strictly speaking the authors did not give it
a name there but only mentioned `generalized meanders' in the section titles.)
Firstly, it can be defined as a
randomized Bessel %
bridge
of
dimension $\delta+\delta'$,
using Beta$(\delta/2,\delta'/2)$ distribution.
For this approach, see \citet[Theorem 3.12]{Mansuy-Yor-2008}. Secondly, it can be defined as $((R_t^2+
(R_t')^2)^{1/2})_{t\in[0,1]}$, where $R$ is a $\delta$-dimensional Bessel bridge (pinned down at $R_0 = 0$ and $R_1 = 0$),
$R'$ is a $\delta'$-dimensional Bessel process starting from 0, and $R$ and $R'$ are independent.
Thirdly, the law $M^{\delta,\delta'}$ of $\mathbb B\topp{\delta,\delta'}$, viewed as a probability measure on $C([0,1])$, is absolutely continuous with respect to the %
probability measure
$P_0^{{\rm BES}(\delta+\delta')}$ on $C([0,1])$ induced by a $(\delta+\delta')$-dimensional Bessel process starting from 0; more precisely
\begin{equation}\label{eq:Mdd}
M^{\delta,\delta'} = \frac{c_{\delta,\delta'}}{X_1^\delta}P_0^{{\rm BES}(\delta+\delta')},\quad \delta,\delta'>0,
\end{equation}
with %
\begin{equation}\label{eq:cdd}
c_{\delta,\delta'} = M^{\delta,\delta'}(X_1^\delta) = \frac{2^{\delta /2}\Gamma((\delta+\delta')/2)}{\Gamma(\delta'/2)}.
\end{equation}
Here, for $\omega$ from the canonical space $C([0,1])$, $X_1(\omega) = \omega_1$.
For the second and third characterizations, see \citet[Theorem 3.9]{Mansuy-Yor-2008};
the corresponding formula \eqref{eq:cdd} in Theorem 3.9 therein has a typo and is corrected here.
The generalized Bessel meanders $\mathbb B^{\delta,\delta'}$ and the generalized Brownian meanders $\mathbb B\topp\nu$ with Beta distribution $\nu$ in Corollary \ref{coro:gen-meander} are in general different processes. However,
since a Bessel bridge of dimension 3 is a Brownian excursion, generalized Bessel meanders %
\eqref{eq:Mdd} with \begin{equation}\label{eq:dd}
\delta\in(0,3) \mbox{ and } \delta' = 3-\delta,
\end{equation}
become a special
case of the generalized Brownian meanders introduced in \eqref{eq:Gen-meander}. This case covers a couple of examples
investigated in the literature.
In particular,
it is known that $\mathbb B\topp {1,2}$ is the Brownian meander $\mathbb B^{me}$ and in this case the relation~\eqref{eq:Mdd}
is
originally due to \citet{imhof84density}. The process $\mathbb B\topp {2,1}$, known as {\em Brownian co-meander}, has also been investigated before,
and the relation~\eqref{eq:Mdd}
is due to \citet{biane87processus}. See also~\citet[Theorem 7.4.1]{Yen-Yor-2013}.
Our Corollary \ref{coro:gen-meander}
simplifies
and takes the following form.
\begin{corollary}\label{coro:3}
In the notations of %
Corollary \ref{coro:gen-meander}, we have for all $\delta\in(0,3)$,
\begin{multline}
\label{Meander-alpha-dual1}
\EE{\exp\left(-\sum_{k=1}^d (s_{k}-s_{k-1})\mathbb B\topp{\delta,3-\delta}_{1-t_k}\right)}
\\= \frac1{2^{\delta/2}\Gamma(\delta/2)} \int_0^\infty \EEx{x}{\exp \pp{-\frac12\sum_{k=0}^{d}(t_{k+1}-t_{k})X_{s_k}}}x^{\delta/2-1} dx.
\end{multline}
\end{corollary}
\begin{proof}
From \eqref{eq:d+d'=3}
with $\alpha=\delta/2$, $\beta=3/2-\alpha$ it follows that
\begin{align*}
\varphi_\nu(x) & = \frac{\sqrt x}{\sqrt{2\pi}}\frac{\Gamma(\alpha+\beta)}{\Gamma(\alpha)\Gamma(\beta)}\int_0^\infty e^{-ux/2}u^{\beta-1}
du
= \frac{2^{\beta}\Gamma(\alpha+\beta)}{\sqrt{2\pi}\Gamma(\alpha)}x^{1/2-\beta}\\& = \frac{x^{\delta/2-1}}{2^{\delta/2}\Gamma(\delta/2)}.
\end{align*}
\end{proof}
\begin{remark}
When $d=1$,
expression
\eqref{Meander-alpha-dual1} becomes
\[
\EE{ e^{-s
\mathbb B\topp{\delta,3-\delta}_{t}}}=%
\frac1{2^{\delta/2}\Gamma(\delta/2)} \int_0^\infty\EEx{x}{e^{-tX_s/2}} e^{-(1-t)x/2}x^{\delta/2-1}\,dx,
\]
This differs from \citet[Eq.~(2)]{Bertoin-Yor2001subordinators}, who developed a dual representation of Laplace transforms
of univariate distributions in the form of univariate integrals
\[
\EE{ e^{-s U_t}}=\EE{ e^{-t R_s}}, s,t>0.
\]
But, clearly, there is some similarity in how the roles of arguments and times are interchanged.
\end{remark}
\begin{remark}\label{Joseph}
It is natural to interpret %
Beta$(0,3/2)$
and Beta$(3/2,0)$ as degenerate laws.
The limiting case of %
Beta$(\delta/2,3/2-\delta/2)$ as %
$\delta\uparrow3$ is $V=1$,
so with
$\mathbb B\topp{3,0}=\mathbb B^{ex}$
Corollary \ref{coro:3} can be viewed as an extension
of Theorem~\ref{T1.5}.
On the other hand, as $\delta\downarrow 0$,
one can check that
$B\topp{ \delta,3-\delta}$
converges to the 3-dimensional Bessel process.
In this limit, the right-hand side of \eqref{Meander-alpha-dual} becomes an undefined expression. %
The 3-dimensional Bessel process is also considered in \citet[Theorem 6.1]{hirsch2012temporally}.
Laplace transforms of squared Bessel processes have been investigated in \citet{pitman82decomposition}. However, we do not see immediate connection between our
identities
to the results there.
\end{remark}
\subsection*{Acknowledgement}
The authors thank Jim Pitman for insightful comments on %
several early versions of the paper, and a few key references including in particular \citet{Mansuy-Yor-2008},
which helped us improve significantly the paper. The authors also thank Joseph Najnudel for several inspiring discussions.
WB's research was supported in part by the Charles Phelps Taft Research Center at the University of Cincinnati. YW's research was partially supported by NSA grant H98230-16-1-0322 and Army Research Laboratory grant W911NF-17-1-0006.
\bibliographystyle{apalike}
| 7a6e4cd3db82644710b6b4877ff5e26f95c55838 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Introduction}
The Brain Computer Interface (BCI) community today's priority is to assure the system robustness and its usability. It is quite a difficult task, considering the abundant inter and intra-subject variability. The major obstacle lies in the large spectrum of sources of variability during BCI usage, ranging from (i) imperfect recording conditions e.g. environmental noise, humidity, static electricity etc. \cite{Maby2016} to (ii) the fluctuations in the userβs psycho-physiological states, due to e.g., fatigue, motivation or attention \cite{JeunetPredictingPatterns}. There are yet more improvements to be done for a system ready to be easily used in real life conditions \cite{Wolpaw2002BraincomputerControl}.
BCI systems showed quite an improvement with adaptive methods, i.e. adapting the machine to the changeable brain signals of the user during a BCI task. Currently, adaptation is mainly done by using different signal processing techniques without including human factors \cite{Makeig2012}. However, if the users do not understand how to manipulate a BCI system, or are not motivated to make necessary effort for such manipulation, then they are not able to produce stable and distinct EEG patterns. In that case, no signal processing algorithm would be able to decode such signals \cite{Lotte2013}. Thus, for designing a BCI, ignoring certain information about the users, e.g. their skills, cognitive abilities and motivations, may represent one of the major drawbacks for the advancement of BCIs.
A potential improvement in BCI is to acknowledge how difficult it can be to learn to produce mental commands (a very atypical skill) without a proper feedback about the progress one has made.
In every discipline, a certain feedback on ones performance is necessary to enable learning, as shown in the earliest work about Operant Conditioning and Reinforcement Learning \cite{Skinner1938}. Notably, this question was studied by behaviorists for decades on animals, using rewards e.g. food, as extrinsic motivation to promote desired behavior. As humans have more complex cognitive functions, a more effective way to promote learning is in a social context, with a tutor who would prepare and adapt a task according to the student's competences. The tutor's feedback and well organized tasks would lead the disciple to gradually build up knowledge and skills, to feel confident and to be intrinsically motivated, or to be in the Zone of Proximal Development (ZPD)\cite{Vygotsky1978}. Derived from cognitive developmental theories \cite{Vygotsky1978} and refined through instructional design theories \cite{Keller1987,Malone1987}, intrinsic motivation is to be a substantial element for learning. Thus, it is important to carefully design the feedback if we want to encourage learning and optimal performance.
Unfortunately, for long this was not the case in BCI community, as BCI systems were improved mostly with novel machine learning techniques \cite{Makeig2012}. The result of neglecting the feedback design led to often monotonous and repetitive content, further discouraging the user, and leading to reduced skill and impaired performance \cite{Cho2004a,Kleih2010}, thus highly affecting the system's accuracy. Potentially, instructional design theories could add a missing piece for designing optimal BCI feedback \cite{Lotte2013}.
There have been extensive literature describing higher BCI user performance and experience using game-like feedback \cite{Ron-angevin2009BraincomputerTechniques,Scherer2015}.
Immersive and game-like environments attract users' attention, induce intrinsic motivation, thus promote learning and performance with less effort and frustration -- for a review see \cite{Lumsden2016}.
Even using extrinsic motivation such as monetary reward can encourage users to perform better \cite{Kleih2010}. Some studies showed that user's belief on their performance with biased feedback induced motivation and thus higher performance \cite{Barbero2010BiasedInterfaces.}. Hence, sometimes it is worth to trade the system's accuracy to the perceived, subjective user's feeling of control.
Keeping that into account, a way to promote efficiency and motivation while respecting the principles of instructional design leads us to the Theory of Flow introduced by Csikszentmihalyi in \cite{Csikszentmihalyi1975}. He was fascinated by the capacity of artists to be in a state of enjoyment while effortlessly focused on a task so immersive that one looses the perception of time, of self and of basic human needs (hunger, sleep etc.). When in the flow state, people are absorbed in an activity, their focused awareness is narrowed, they lose self-consciousness, and they feel in control of their environment. As a consequence, they often perform to the best of their capacity with a sense of automaticity and a high level of confidence. Studies report flow experience in numerous activities including rock climbing, dancing, chess, reading, etc. \cite{Csikszentmihalyi1975,Csikszentmihalyi1989}.
Another pertinent element which encourages intrinsic motivation and is showed to be in relation with flow, is music \cite{Croom2015}. Recent studies showed that music has an ergogenic effect on humans, i.e. physical enhancement while performing a physical activity \cite{Anshel1978}. In \cite{Karageorghis2010ErgogenicExercise} was reported that Haile Gebrselassie, an athlete who broke 10 000m world record in 1998, paced his running on music he was listening to, i.e. synchronous music. There is evidence that synchronous music, as a strong motivational effect, directly enhances physical performance \cite{Simpson2006} while asynchronous (background) music induces flow when accomplishing a task \cite{Pates2003EffectsPlayers,Pain2011Pre-CompetitionSoccer}. Most of all, background music with medium tempo (speed) has showed highest impact on flow \cite{Karageorghis2008PsychologicalExercise}.
To be in the state of flow, a task needs to have the following requirements:
\begin{itemize}
\item \textbf{To be immersive}, with attractive visual/audio stimuli to maintain the user's attention. The principle of preserving flow with aesthetically pleasing and ergonomic content have been researched largely in the context of human computer interaction \cite{Webster1993} and Internet navigation, e.g. e-learning \cite{Esteban-millat2014ModellingEnvironment};
\item \textbf{To adapt} the task difficulty with the userβs skills, i.e. an easy task might be boring as a difficult one might be frustrating, hence finding the golden middle is the way of feeling in control and keeping the motivation. Such difficulty adaptations were found in games, to keep the gamer in flow \cite{Bulitko2012FlowModel}, or during teaching activities \cite{Clement2015Multi-ArmedJEDM} to improve learning and keep the student in the ZPD \cite{Vygotsky1978}.
\item \textbf{To have clear goals and immediate feedback / rewards}; aspired for educational purposes \cite{Heutte2016}, so that learning becomes an enjoying and autotelic (self-rewarding) process \cite{Ninaus2015GameTask}.
\end{itemize}
Therefore, in order to improve BCI users' performance, learning, and experience, it seems promising to try to guide them towards the state of flow. This is what we start to explore in this paper.
In particular, our research question is: \textit{Does flow improve BCI user performance?} We chose to manipulate flow in a ludic BCI environment with 2 factors: 1) Feedback adaptation, i.e. perceived difficulty adaptation to the user skills, and 2) Asynchronous music to encourage the user. Thus, our following hypotheses are:
\textbf{H1.} Adapting the feedback improves flow, thus improves performance.
\textbf{H2.} Asynchronous music improves flow, thus improves performance.
In consonance with the Flow theory, we presented a motor imagery (MI) BCI task in an open-source 3D video game (TuxRacer \footnote{https://extremetuxracer.sourceforge.io/}).
We investigated the effects of these two flow factors on user's flow state as well as on user performance, i.e. classification accuracy.
\section*{Materials and Methods}
\emph{Manipulating Flow:} In order to fulfill Flow theory requirements, we considered the following:
\begin{itemize}
\item \textbf{An immersive and ludic environment}, here the \textbf{TuxRacer} video game was adjusted for a 2-class Motor Imagery (MI) BCI. The game depicts a ski course, in which a virtual penguin, Tux -- controlled by the player -- slides through various slopes and has to catch as much fish as possible. With the BCI adjustments, Tux was maneuvered with kinesthetic imagination of either left or right hand, see Figure \ref{fig:tux_score}.
\item \textbf{The adaptation} of the feedback bias, i.e. users were made to believe they performed differently from what they really did, in order to be in the flow state. If they had poor performances they were positively biased to a higher degree than if they had fairly good performances. However, when the performances were too good, then the users were slightly negatively biased, so that the task would not seem too easy. This was achieved by adaptively increasing or decreasing the classifier output, i.e. the decoding of MI commands would seem different from what it was in reality.
\item \textbf{Asynchronous music} consisted of 3 songs with medium tempo (120-160 beats per min), played in the background during the BCI task. 15 persons voted on social media for songs which would motivate them while playing TuxRacer. The selected songs are "Epic" by Alexey Anisimov (113s), "Confident \& Successful" by MFYM (168s) and "Acoustic Corporation" by OAP (132s), all available on Jamendo\footnote{https://www.jamendo.com/}.
\item \textbf{Clear goals with immediate audio and visual feedback}, i.e. to collect maximum points by manipulating Tux to move either left or right to catch fish.
The feedback is clear -- once caught, the fish disappears with a brief audio stimuli stressing that the target was reached.
\end{itemize}
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.7\columnwidth]{img/tux_score.jpg}
\caption{\small{Participant using MI commands to play TuxRacer, e.g. imagining right hand movement to catch fish on the right.}~\label{fig:tux_score}}
\end{figure}
\hspace{0.5cm}\emph{Experimental design:} We created a 2 (\textit{adapt} \textit{vs} \textit{no-adapt}) by 2 (\textit{music} \textit{vs} \textit{no-music}) mixed factorial design, i.e. a between-subject adaptation factor, and a within-subject background music factor.
\hspace{0.5cm}\emph{Protocol:} 28 healthy subjects, naive to BCI, participated in the \textasciitilde2 hour-long experiment (5 women, mean age: 25.23 years, SD: 2.98). The first 30 minutes consisted of (i) signing a consent form, (ii) installing of a 32 channels Brain Product LiveAmp EEG, (iii) instructions given to the user and preparation, (iv) \textasciitilde10 minutes system calibration (40 trials of 7s) with the standard 2-class MI BCI (left/right hand) Graz protocol \cite{PfurtschellerMotorCommunication}. In the Graz protocol, the user was presented with arrows indicating the left or right side, to instruct the participant to imagine a left or right hand movement. Afterwards, each participant took part in 2 counterbalanced conditions of \textasciitilde 20mins each with TuxRacer, (a) with and (b) without background music. 3 songs were repeated to accompany the music condition of 6 runs (1 song per 2 runs). Each condition comprised of $6\times 3$min-runs, with 22 trials per run (11 for left and 11 for right hand, in random order), see Figure \ref{fig:protocool}. Each trial consisted in performing left/right hand MI to move Tux in order to catch fish on the left/right of the ski course, respectively. There were 7 closely arranged fish per trial, to be caught within 3 seconds. During 5-second long breaks between trials, the BCI controls were disabled so that Tux would return in a neutral position (center on the ski course) and participants could rest. The study was approved by the Inria ethics committee, COERLE (ComitΓ© opΓ©rationnel d'Γ©valuation des risques lΓ©gaux et Γ©thiques).
\begin{figure}[!htbp]
\centering
\includegraphics[width=1\columnwidth]{img/protocool3.jpg}
\caption{\small{The experiment started with \textasciitilde 10min calibration \cite{PfurtschellerMotorCommunication}, followed by 2 conditions: either with or without music -- 6 runs of 3 minutes per condition. The adapt group received an adapted (biased) feedback, contrary to the no-adapt group. Adaptation is symbolized by the magnet. Both groups were asked to fill EduFlow2 \cite{Heutte2016} questionnaires for the flow state assessment and BMRI \cite{Karageorghis1999} questionnaire for investigating the quality of music.}~\label{fig:protocool}}
\end{figure}
\hspace{0.5cm}\emph{Questionnaires:} Prior to the experiment, a Swedish Flow Proneness Questionnaire (SFPQ) \cite{Ullen2011} was sent to subjects to fill in at home. This 5 points Likert scale questionnaire measures flow proneness -- flow as a person's trait. To estimate to which extent users were in the state of flow, they were asked to fill in the EduFlow2 questionnaire \cite{Heutte2016} after each condition (\textit{music} or \textit{no-music}). The EduFlow2 measures flow state through 4 dimensions: cognitive control, immersion, selflessness and autotelism -- a self rewarding experience. To have a measure of the quality and motivation of the selected music, the participants also filled a dedicated questionnaire, the Brunel Music Rating Inventory (BMRI) \cite{Karageorghis1999}.
\hspace{0.5cm}\emph{Signal processing}: The acquired EEG was band-bass filtered with a Butterworth temporal filter between 8 and 30Hz. We computed the band power using a 1s time window sliding every 1/16th s. We used a set of Common Spatial Patterns (CSP) spatial filters to reduce the 32 original channels down to 6 "virtual" channels that maximize the differences between the two class motor imagery \cite{Ramoser2000}. A probabilistic SVM (Support Vector Machine) with a linear kernel was used to classify the data between left and right classes (regularization parameter C = 1). That way, the output of the SVM between 0 and 1, indicated a class recognized with a certain degree of confidence, e.g. 1 means that the right- hand class was recognized with high confidence.
\hspace{0.5cm}\emph{Performances}: The online performance corresponds to the peak accuracy of the classifier that controlled the video game, i.e. the highest classification accuracy over all trials' time windows. The offline performance was computed afterwards with a 4-folds cross validation, i.e. regarding only the data recorded during the interaction with the video game for training and testing. In other words, data recorded during the Graz protocol was not used to compute offline performances. We used a LDA (Linear Discriminant Analysis) for the offline classification, since it is less computationally demanding. Both for online and offline analyses, one accuracy score was computed over the music / no\_music condition (i.e. 6 runs of 22 trials).
\hspace{0.5cm}\emph{Game controls}: The TuxRacer game was controlled \textit{via} a virtual joystick. When a right hand movement was recognized (SVM output of 1) the virtual joystick was tilted toward the right at its maximum angle, 45 degrees. Inversely, when a left hand movement was recognized (SVM output of 0), the virtual joystick was tilted 45 degrees to the left. Between 0 and 1, the values of the virtual joystick were mapped linearly (from minus 45 degrees to 45 degrees). Thanks to this simple virtual joystick, we did not need to modify the usual input commands to the complex BCI ones in the game. Basically, the virtual joystick can act as or replace the usual computer controls, such as keys on the keyboard. Our freely available source code\footnote{https://github.com/conphyture/LSL2joy} could be used to control any (linux) joystick-based game with a BCI.
\hspace{0.5cm}\emph{Game modifications}: We designed the BCI TuxRacer game so that its timing and structure mirror that of the Graz motor imagery BCI protocol \cite{PfurtschellerMotorCommunication}, but in an immersive and motivating environment. We modified the shape of the terrain, curving it alike a bobsleigh course. Consequently, by the force of gravity, Tux would slide back to the middle of the screen between trials, when the commands were deactivated. Between trials, Tux would still be skiing towards the following trial with constant speed, enabling the users to see the next fish.
We fixed the position of the fish on the ski course edges, so that the targets were equidistant from Tux at the beginning of each trial, i.e. same distance from the center of the ski course. The reason for this is to enable the user to provide equivalent potential effort for both MI classes (left/right hand). By assuring a constant speed for Tux, a race (run) always lasted 3 minutes.
\hspace{0.5cm}\emph{Game adaptation}: The \textit{no\_adapt} group was the first to participate in the experiment. Thanks to the fact that there was a correlation between the user's flow state and the performances in the control group, we empirically calculated a performance level (classifier accuracy) for which users felt most in flow. We used that value as an attractor or a quasi-flow value to lure Tux in. At each instant ($1/16^{th}$sec sliding window) we would retrieve the classifier output and add to it a value which would push Tux a half way towards our attractor. This value was determined intuitively from Flow theory, to keep the difficulty in the "golden middle". Consequently, when user performances were very poor, Tux was boosted to a higher extent towards the attractor, i.e. in this case users were helped (positively biased) more than when their performances were fairly good. However, when the performances were too good, the perceived performances were deteriorated (feedback was negatively biased). The "flow" function would then be: $f(s_i) = s_i + \frac{(a-s_i)}{2} , s_i \in [-1,1] $, where $s_i$ stands for user skill, which is given by the classifier output and scaled to ease the computation (-1 for left, 1 for right), for all instants $i$ within 22 trials, ($i =1,..16Hz\times 66s$). Finally, $a = 0.79$ denotes the attractor for the right class ($a = -0.79$ for left class).
\section*{Results}
The normal distribution of all the data was verified using a Shapiro-Wilk normality test.
\hspace{0.5cm}\emph{Flow-factor's influence on EduFlow2}: We tested the effects of our mixed factorial design on each of the 4 dimensions measured by the EduFlow2 questionnaires using a Markov Chain Monte Carlo (MCMC) method \cite{Hadfield2010b}. The MCMC showed a significant difference between \emph{adapt} and \emph{no-adapt} along the $1^{st}$ dimension (p < 0.01). Participants in the adapt group reported higher cognitive control (mean: 5.38, SD: 0.84) compared to the no-adapt group (mean: 4.49, SD: 0.83), see Figure \ref{fig:eduflow}.
There was no significant difference between groups regarding their flow trait, measured with SFPQ (1-way ANOVA, p = 0.25). ANCOVA tests showed that it was not a confounding factor for neither EduFlow2 nor performances.
There was no difference between groups regarding BMRI (1-way ANOVA, p = 0.53). Mean score: 15.80, SD: 4.14 -- maximum score with the questionnaire we distributed: 25.3. There was no correlation between BMRI scores and flow (p = 0.54) nor with user performance, online (p = 0.78) or offline (p = 0.20).
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.9\columnwidth]{img/eduflow_D1.pdf}
\caption{\small {The EduFlow2 score (7 Likert scale) for the first dimension (cognitive control) depending on the between-subject factor \textit{adapt} and on the within-subject factor \textit{music}. Users were in higher cognitive control in the adapt condition (left).}~\label{fig:eduflow}}
\end{figure}
\hspace{0.5cm}\emph{Flow-factor's influence on performance}: The question whether our conditions could directly improve the online performances was tested with a 2-way ANOVA. There was a significant interaction between music and adaptation (\textit{p<0.05}). Music had a significant effect on the online performance (mean with music: 0.62, SD: 0.09, mean with no\_music: 0.65, SD: 0.11, \textit{p<0.05}) but adaptation had not (\textit{p=0.08}), see Figure \ref{fig:perfs}. A post-hoc Tukey analysis reveals that the one significant interaction occurs in the no\_adapt condition, between music (mean: 0.64, SD: 0.11) and no\_music (mean: 0.68, SD: 0.13) (\textit{p<0.001}).
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.9\columnwidth]{img/perfs.pdf}
\caption{\small {The peak performance during the completion of the video game depending on the between-subject factor \textit{adapt} and on the within-subject factor \textit{music}. In the no\_adapt condition (right), users had better online performances without music.}~\label{fig:perfs}}
\end{figure}
\hspace{0.5cm}\emph{Correlation between EduFlow2 \& performance}:
There was no correlation between flow (mean of all the EduFlow2 dimensions) and online performance (\textit{p=0.12}), however there was a positive correlation between flow and offline performance (Pearson coefficient: 0.35, \textit{p<0.01}), see Figure \ref{fig:correlation}. More precisely, offline performances are significantly correlated with two dimensions of flow: the $2^{nd}$ -- immersion (p<0.01, Pearson coefficient: 0.38) and the $4^{th}$ -- autotelism, (p<0.05, Pearson coefficient: 0.32). We corrected the p-values for multiple comparisons with false discovery rate \cite{Noble2009}.
\begin{figure}[!htbp]
\centering
\includegraphics[width=0.9\columnwidth]{img/correlation.pdf}
\caption{\small {Positive correlation between EduFLow2 scores (mean of 4D) and offline performance (Pearson coef:0.35, \textit{p<0.01}).}~\label{fig:correlation}}
\end{figure}
\section*{Discussion}
\textbf{H1. validated}: Adapting the task difficulty to users skill improved one dimension of flow state, cognitive control. People who faced a challenge better suited to their skill felt more in control. Thus, taking into account user's predispositions could lead to a greater user experience.
\textbf{H2. in contradiction}: Not only the presence of a background music had no effect on flow, but it deteriorated the online performance. Therefore, this result contradicts our second hypothesis.
\hspace{0.5cm}\emph{Music pace mismatch}. As opposed to what we expected, we could not directly improve performance by manipulating the flow factors we chose (adaptation and music). The latter could be explained by the songs we chose, since the motivational qualities of the music (measured with the BMRI questionnaire) were not very high and not correlated to any dimension of flow. Instead of picking those songs from the public domain, users may have been more motivated should they have chosen their own music. The decrease of performance in the music condition might come from the mismatch between the rhythm of the music and the pace of the game, i.e. with the pace of the imagined hands movements. Indeed, some users shared informally that they were imagining playing their musical instrument as MI commands and that the songs further disturbed their pace.
\hspace{0.5cm}\emph{Different training environment.} There was no correlation between flow state (EduFlow2) and online performances. That could be due to the differences between the calibration environment (Graz protocol) and the game, e.g. the first being minimalistic and the latter a 3D video-game. Moreover, as the calibration was done without music, maybe the performances online were better without it because the EEG signals might have changed, therefore the classifier could not recognize them anymore.
\hspace{0.5cm}\emph{Flow increases with performances.} These later assumptions are strengthened by the fact that there was a positive correlation between flow and the offline performances, when only game data was taken into consideration. The state of flow was then positively correlated with users' performance: the feeling of immersion and the autotelic experience (i.e. the completion of the task was self-rewarding) increased with the offline performance. Hence, not only encouraging a state of flow would produce BCIs more pleasing to the users, but it might also benefit the accuracy of the system. We still have to identify the direction of the correlation though: does flow state increases performances or do good performances increase flow state?
Overall, the discrepancy in our results could stress that flow is a complex phenomenon, and however beneficial to obtaining better BCI, the emerging interaction between its components should be more thoroughly investigated.
\section*{Conclusion}
By investigating means to improve BCI user performance and usability through instructional design theories, we came across the Flow Theory. This theory, which describes an optimal user state, showed to improve performances in many fields. We hypothesized that the state of flow could benefit BCIs. In a MI BCI task, we manipulated flow by adapting the perceived difficulty and by adding a background music. We used an immersive environment, a 3D video game, TuxRacer (the modification can be found online \footnote{https://github.com/jelenaLis/tux-modifs)}).
Our main findings show that the adaptation increases one of the dimensions of flow -- cognitive control, and that user's offline performances are positively correlated with flow. In the future we could attempt to better suit the adaptation of the task to the users: it could be biased adaptively over time, across several sessions, following the progress of the user. We could also try to account for the amount of effort that the user puts into the completion of the task in order to better comprehend such complex phenomena. For example, measuring workload could facilitate the assessment of the challenge that users are facing and computationally predict the state of flow\cite{Bulitko2012FlowModel}.
According to the literature, we chose asynchronous music with medium tempo to follow the BCI task. Unexpectedly, the background music impeded the performances of the user. This result stresses the importance of the choice of music to accompany a task. One explanation could lie in the very BCI paradigm we chose. Indeed, a motor imagery task might share similarities with actual physical activity, where it had been shown that \textit{synchronous} music could effectively stimulate the sensory-motor cortex\cite{Hardy2013}. Hence, a future work would consist in synchronizing music to game's cues (e.g. trials sequences) or to user's motor imagery pace. Such music, generated in real time, might enhance the flow state and intrinsic motivation. Concurrently, we should verify if the user is musically educated, as in some cases users imagined playing instruments as MI commands, and because musicians elicit different brain activity in motor areas\cite{Luo2012}.
Flow is not only a promising research direction to improve BCI systems, but it raises a new question: should we put all our efforts in favoring the machine accuracy, or rather the human subjective experience?
\smallbreak
\emph{Acknowledgements}: This work was supported by the French National Research Agency with the REBEL project and grant ANR-15-CE23-0013-01. The authors would like to thank Jean Heute for his help with the EduFlow questionaire.
\setlength{\bibitemsep}{0pt}
\printbibliography
\end{document}
| 10d5d13eb4b539289bb71ad1034e617b64a6ce24 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
It has been found recently that for any 4d ${\cal N}=2$ SCFT, there is a protected sector described by a 2d chiral algebra or more precisely a vertex operator algebra (VOA) \cite{Beem:2013sza}. The VOA corresponding to 4d ${\cal N}=2$ SCFT includes the operators in the Higgs branch chiral ring and more generally the so-called Schur operators \cite{Gadde:2011ik,Gadde:2011uv}. Correlation functions in this sector are meromorphic and do not change under exactly marginal deformations.
The 2d chiral algebra\footnote{We use the vertex operator algebra and chiral algebra interchangeably.} is constructed from the 4d theory as follows. First, pick a two-dimensional slice in the four-dimensional space $\mathbb{R}^2 \subset \mathbb{R}^4$ with complex coordinates $(z, \bar{z})$. Then choose a set of particular operators ${\cal O}(z, \bar{z})$ living on this plane $\mathbb{R}^2$ annihilated by a combination of Poincare and conformal supercharges $\mathfrak{Q} = Q + S$. At the origin these operators are the Schur operators. The operator product expansion (OPE) of these operators turns out to be meromorphic up to the $\mathfrak{Q}$-exact piece. Therefore by passing through the $\mathfrak{Q}$-cohomology, the operators ${\cal O}(z, \bar{z})$ form a meromorphic chiral algebra or VOA.
Under this 4d ${\cal N}=2$ SCFT/VOA correspondence, the 2d Virasoro central charge is given in terms of 4d central charge as
\begin{align}
c_{2d} = -12 c_{4d} \ .
\end{align}
When there is a (non-R) global symmetry $\mathfrak{g}$, the symmetry in 2d is enhanced to affine symmetry $\hat{\mathfrak{g}}_{k_{2d}}$ with the level given by the 4d flavor central charge as
\begin{align}
k_{2d} = - \frac{1}{2} k_{4d} \ .
\end{align}
The (super)-character of the VOA gives the Schur limit of the superconformal index
\begin{align}
{\rm Tr}_{{\cal V}} (-1)^F q^{L_0} = {\cal I}_{\mathrm{Schur}}(q) \ ,
\end{align}
where ${\cal V}$ denotes the VOA or the vacuum module of the corresponding chiral algebra.
We review the definition of the superconformal index and its various limits in appendix \ref{app:n2idx}.
Various aspects of the VOA corresponding to the 4d ${\cal N}=2$ SCFT has been studied. The topological field theory (TQFT) structure of the chiral algebra for the class ${\cal S}$ theory has been investigated in \cite{Beem:2014rza, Lemos:2014lua}. The chiral algebra for the (generalized) Argyres-Douglas theory \cite{Cordova:2015nma, Xie:2016evu, Creutzig:2017qyf} and also for the ${\cal N}=3$ theories \cite{Nishinaka:2016hbw, Lemos:2016xke} has been identified. Bounds on the central charges have been obtained using the chiral algebra structure \cite{Liendo:2015ofa, Lemos:2015orc}. The effect of the defect operators has been studied in \cite{Cordova:2016uwk, Cordova:2017mhb}.
It has also been conjectured that it is possible to reconstruct the Higgs branch \cite{BRVOA} and also Macdonald index from the chiral algebra \cite{Song:2016yfd}. Moreover, the Coulomb branch index is related to the modular transformation of the VOA \cite{Fredrickson:2017yka}.
Our primary focus in this paper is to understand the Schur sector and the Higgs branch of the generalized Argyres-Douglas theory \cite{Argyres:1995jj,Argyres:1995xn} that can be engineered from M5-branes \cite{Gaiotto:2009hg,Gaiotto:2009we, Xie:2012hs, Wang:2015mra}.
They are interesting because the corresponding chiral algebra/VOA for a subset of such AD theories are particularly simple \cite{Xie:2016evu}. In this paper, we focus on the subset of the AD theories, where 1) there is no mass parameters associated to the irregular singularity $J^b[k]$, and 2) there is no exactly marginal deformations. It is widely believed that every exactly marginal operator in $4d$ ${\cal N}=2$ SCFT comes from the gauge coupling. Therefore the existence of a marginal coupling signals that this theory can be decomposed into decoupled SCFTs by setting the gauge coupling to zero. We will focus on the ``non-decomposable" AD theories.\footnote{For example, $D_4^6[3] \equiv (D_4^6[3], \varnothing) $ theory has an exactly marginal operator. This theory can be decomposed into three copies of the $(A_1^2[1], F)$ theory (also called the $H_1 = (A_1, A_3)$ theory in the literatures) coupled via $SU(2)$ gauge group. The associated VOA is conjectured in \cite{Buican:2016arp}.}
The main tools we use are the conjectural vertex operator algebra and the topological field theory structure suggested by
the M5-brane construction of the class ${\cal S}$ theory theories. The Schur index for any class ${\cal S}$ theory can be obtained by using the 4d/2d correspondence \cite{Gadde:2009kb,Gadde:2011ik,Gadde:2011uv,Gaiotto:2012xa} between 4d SCFT and 2d topological quantum field theory (TQFT). This correspondence has also been extended to the case with irregular punctures \cite{Buican:2015ina,Buican:2015tda,Song:2015wta, Buican:2017uka}, which allows us to obtain the index for the AD theory. It has also been found to be consistent with the vacuum character of the chiral algebra for the AD theory. The result also agrees with the IR computation in the Coulomb branch \cite{Cordova:2015nma, Cecotti:2015lab, Cordova:2016uwk, Cordova:2017ohl} done by using the BPS monodromy operator \cite{Cecotti:2010fi,Iqbal:2012xm}. The full superconformal index has been computed using the renormalization group flow from certain ${\cal N}=1$ gauge theory to the ${\cal N}=2$ AD theory \cite{Maruyoshi:2016tqk, Maruyoshi:2016aim, Agarwal:2016pjo}, which gives a consistent result with the chiral algebra.
We consider a class of AD theories labeled as $(J^b[k], Y)$.\footnote{When $b=h^\vee$ and $Y=F$, this also is identical to the $D_p(J)$ theory of \cite{Cecotti:2012jx,Cecotti:2013lda} with some value $p$.} This class of SCFTs is engineered using 6d ${\cal N}=(2,0)$ theory of type $J \in ADE$ on a Riemann sphere with one irregular singularity labeled by $J^b[k]$, and one regular
singularity labeled by $Y$ which also specifies a nilpotent orbit\footnote{In this paper, we use the Higgs branch label for a puncture, i.e. a full regular puncture corresponds to a trivial nilpotent orbit, while null regular puncture corresponds to
the principal nilpotent orbit.} of $J$. When there is no regular puncture (or $Y=\varnothing$), we use the notation $J^b[k]\equiv(J^b[k], \varnothing)$ to denote the corresponding theory.
When the regular puncture is full ($Y=F$) or null ($Y=\varnothing$), the VOAs for the corresponding theories $(J^b[k], F)$ and $J^b[k]$ have been conjectured in \cite{Xie:2016evu}. On the other-hand, in \cite{Beem:2014rza} it was shown that the associated VOA for the theories given by other type of punctures can be obtained via quantum Drinfeld-Sokolov reduction \cite{Drinfeld:1984qv, Bershadsky:1989mf,Bershadsky:1989tc,Feigin:1990pn}. Combining the two, we obtain the following:
\begin{itemize}
\item
The VOA for the $(J^b[k], Y)$ theory is given by the W-algebra ${\cal W}^{k_{2d}}(J, Y)$,
where $k_{2d}=-h+{b\over b+k}$ with $h$ being the dual Coxeter number of $J$. Here ${\cal W}^{k_{2d}}(J, Y)$ is the W-algebra obtained via quantum Drinfeld-Sokolov reduction of the Kac-Moody algebra $\widehat{J}_{k_{2d}}$ using the nilpotent orbit associated to $Y$:
\begin{align}
\widehat{J}_{k_{2d}}= {\cal W}^{k_{2d}}(J, F) \leadsto {\cal W}^{k_{2d}}(J, Y)
\end{align}
Here $F$ denotes the full puncture carrying the flavor symmetry $J$.
\end{itemize}
As was conjectured in \cite{Beem:2014rza}, the VOA also allows us to extract the Higgs branch of the 4d theory. Therefore we have:
\begin{itemize}
\item
The Higgs branch of theory $(J^b[k], Y)$ is identified with the associated variety $X_{\cal V}$ \cite{MR3456698} of ${\cal W}^{k_{2d}}(J, Y)$:\footnote{This relation is not strictly true for a general 4d ${\cal N}=2$ SCFT, but we provide evidence for the examples we consider in this paper. See \cite{Beem:2014rza} for more detail.}
\begin{align}
\boxed{ {\cal M}_{\mathrm{Higgs}} = X_{\cal V} }
\end{align}
The associated variety $X_{\cal V}$ is given in terms of the closure of a nilpotent orbit, which depends on the choice of $k$ and $Y$ as:
\begin{itemize}
\item If $k>0$, the associated variety is
\begin{align}
X_{{\cal V}}= N \cap S_Y \ ,
\end{align}
where $N$ is the principal nilpotent cone, and $S_Y$ is the Slodowy slice defined using the nilpotent orbit corresponding to $Y$.
\item If $k<0$, the associated variety is given as
\begin{equation}
X_{{\cal V}}=X_M \cap S_Y \ ,
\end{equation}
where $X_M = \overline{\mathbb{O}[k]}$ is the closure of a certain nilpotent orbit which \emph{depends on $k$}.
\end{itemize}
\end{itemize}
The goal of the current paper is to verify the conjectures above. We test them by computing the Schur and Hall-Littlewood indices of the $(J^b[k], Y)$ theory and comparing them with the character and the associated variety \cite{BRVOA,MR3456698} of the chiral algebra ${\cal W}^{k_{2d}}(J, Y)$. We also utilize the 3d mirror symmetry whenever it is available.
One of the key result of the current paper is the following:
\begin{itemize}
\item
The Schur index of the theory $(J^b[k], Y)$ is given by the vacuum character of algebra ${\cal W}^{k_{2d}}(J, Y)$. In particular, when $b=h$ and $Y=F$ (the full puncture), the character has a surprisingly simple formula in terms of plethystic exponential,
\begin{align}\label{eq:IdxFull}
\boxed{
{\cal I}_{(J^h[k], F)} (q; \vec{z}) = \mathrm{PE} \left[ \frac{q - q^{h+k}}{(1-q)(1-q^{h+k})}\chi^F_{\textrm{adj}} (\vec{z}) \right]
} \ .
\end{align}
When there is no regular puncture, or equivalently $Y=\varnothing$, the Schur index is given by
\begin{align}
\boxed{
{\cal I}_{J^h [k]}(q) = \frac{1}{\prod_{i=1}^r (q^{d_i}; q)} \mathrm{PE} \left[ - \frac{q^{h+k}}{1-q^{h+k}} \chi_{\textrm{adj}} (q^{\vec\rho}) \right]
} \ .
\end{align}
Here $d_i$ are the degrees of the Casimirs of $J$ and $\vec{\rho}$ is the Weyl vector. The plethystic exponential is defined as
\begin{equation}
\mathrm{PE}\left[x\right] = \exp\left(\sum_{n=1}^{\infty}\frac{1}{n}x^n\right).
\end{equation}
We also give a similar expression for the general puncture of type $Y$ in section \ref{sec:Schur}.
\end{itemize}
It was proven that our formula for the Schur index \eqref{eq:IdxFull} for the $(J_h[k], F)$ theory is indeed identical to the vacuum character of the corresponding VOA \cite{kac2017remark}.
This paper is organized as follows. In section \ref{sec:ADthy}, we review some basic facts about the AD theories we consider. In section \ref{sec:ADVOA}, we describe the corresponding
vertex operator algebras of our theories. In section \ref{sec:Schur}, we describe the TQFT approach to the Schur index and thereby giving a physical derivation of the character formula for the associated chiral algebra. In section \ref{sec:Higgs}, we study the Higgs branch from three perspectives:
the associated variety of the vertex operator algebra, the 3d mirror symmetry, and the TQFT approach. We find these different approaches give consistent results. Finally, we conclude with a remark in section \ref{sec:Conclusion}. In the appendix, we review the definition of the superconformal index and its limits and also list the nilpotent orbits which can appear as the Higgs branch of the AD theories considered in this paper.
\section{Generalized Argyres-Douglas theories from M5-branes} \label{sec:ADthy}
One can engineer four-dimensional $\mathcal{N}=2$ Argyres-Douglas SCFT by putting six dimensional ${\cal N}=(2,0)$ theory of type $J$ on a two-dimensional Riemann surface with the following configurations: a) Sphere with an
irregular singularity; b) Sphere with one irregular singularity and one regular singularity.
The regular singularity is classified in terms of the
nilpotent orbits\footnote{We use the Higgs branch label, i.e. a trivial nilpotent orbit represents a full puncture.} \cite{Gaiotto:2009hg,Chacaltana:2012zy} of the Lie algebra $J$.
The classification of the irregular singularity is related to the classification of positive grading of Lie algebra $J$ \cite{Xie:2012hs, Wang:2015mra, Xie:2017vaf}.
\subsection{Irreducible irregular singularity}
\textbf{Irregular singularity} Let us start with the 6d ${\cal N}=(2,0)$ theory of type $J$ and compactify it on a Riemann surface $\Sigma$,
to get a four dimensional $\mathcal{N}=2$ theory. The Coulomb branch of
the 4d theory is described by a Hitchin system defined on $\Sigma$.
The Hitchin system involves a holomorphic one-form $\Phi$ which is called the Higgs field,
and the irregular singularity is defined by the following singular boundary condition:
\begin{equation}
\Phi={T_k\over z^{2+{k\over b}}}+\ldots\,.
\end{equation}
Here $T_k$ is a regular semi-simple element of $J$. The allowed value of $b$ has been classified in \cite{Wang:2015mra} and summarized in table \ref{table:isolatedsingularitiesALEfib}. We label these irregular singularities as $J^b[k]$. The mass parameters are
identified with the parameters appear as the coefficient of the first order pole. We list the type of irregular singularities that do not admit any mass parameter in table \ref{table:constraint}.
We call them as irreducible irregular singularities.
One can get an AD theory using a single irreducible irregular singularity of the above type, and the central charge for such theory is shown in table \ref{table:noflavor:centralcharges}.
One can also find the Coulomb branch spectrum by studying the Seiberg-Witten geometry.
\begin{table}[h]
\begin{center}
\begin{tabular}{ |c|c|c|c| }
\hline
$J$& Singularity & $b$ &$\mu_0$ \\ \hline
$A_{N-1}$ &$x_1^2+x_2^2+x_3^N+z^k=0$& $N$& $(N-1)(k-1)$\\ \hline
$~$& $x_1^2+x_2^2+x_3^N+x_3 z^k=0$ & $N-1$& $N(k-1)+1$\\ \hline
$D_N$ & $x_1^2+x_2^{N-1}+x_2x_3^2+z^k=0$ & $2N-2$ &$N(k-1)$ \\ \hline
$~$ &$x_1^2+x_2^{N-1}+x_2x_3^2+z^k x_3=0$& $N$& $2k(N-1)-N$ \\ \hline
$E_6$ & $x_1^2+x_2^3+x_3^4+z^k=0$&12 & $6k-6$ \\ \hline
$~$ & $x_1^2+x_2^3+x_3^4+z^k x_3=0$ &9& $8k-6$ \\ \hline
$~$ & $x_1^2+x_2^3+x_3^4+z^k x_2=0$ &8& $9k-6$ \\ \hline
$E_7$ & $x_1^2+x_2^3+x_2x_3^3+z^k=0$& 18 & $7k-7$ \\ \hline
$~$ & $x_1^2+x_2^3+x_2x_3^3+z^kx_3=0$ &14& $9k-7$ \\ \hline
$E_8$ & $x_1^2+x_2^3+x_3^5+z^k=0$&30 & $8k-8$ \\ \hline
$~$ & $x_1^2+x_2^3+x_3^5+z^k x_3=0$ &24& $10k-8$ \\ \hline
$~$ & $x_1^2+x_2^3+x_3^5+z^k x_2=0$ & 20 &$12k-8$ \\ \hline
\end{tabular}
\end{center}
\caption{3-fold singularities corresponding to our irregular punctures \cite{Wang:2015mra}, where $\mu_0$ is the dimension of charge lattice. When $b=h$ with h the dual Coxeter number, such
theories are also called $(J, A_{k-1})$ which were first studied in \cite{Cecotti:2010fi}. }
\label{table:isolatedsingularitiesALEfib}
\end{table}
\begin{table}[h]
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
${\cal T}$ &$~$&${\cal T}$&$~$ \\ \hline
$A_{N-1}^N[k]$ &$(k,N)=1$& $A_{N-1}^{N-1}[k]$ &$\text{No solution}$\\ \hline
$D_{N}^{2N-2}[k]$ &$k\neq 2n$& $D_{N}^{N}[k]$&$N=2^m(2i+1),k\neq 2^m n$\\ \hline
$E_{6}^{12}[k]$ &$k\neq 3n$& $E_6^{9}[k]$ &$k\neq 9n$\\ \hline
$E_{6}^8[k]$ &$\text{No solution}$& $E_{7}^{18}[k]$ &$k\neq 2n$\\ \hline
$E_7^{14}[k]$ &$k\neq 2n,n>1$& $E_{8}^{30}[k]$ &$k\neq 30n$\\ \hline
$E_{8}^{24}[k]$ &$k\neq 24n$& $E_{8}^{20}[k]$ &$k\neq 20 n$ \\ \hline
\end{tabular}
\end{center}
\caption{Constraint on $k$ so that the irregular singularity $J^b[k]$ has no flavor symmetry. Here $n \in \mathbb{Z}$.}
\label{table:constraint}
\end{table}
\begin{table}[h]
\begin{center}
\begin{tabular}{ |c|c|c|c| }
\hline
${\cal T}$ &$c_{4d}$&${\cal T}$&$c_{4d}$ \\ \hline
$A_{N-1}^N[k]$ &${(N-1)(k-1)(N+k+Nk)\over 12(N+k)}$& $A_{N-1}^{N-1}[k]$ &${(Nk-N+1)(N+k+Nk-1)\over 12(N-1+k)}$\\ \hline
$D_{N}^{2N-2}[k]$ &${N(k-1)(-2-k+2N+2kN)\over 12(-2+k+2N)}$& $D_{N}^{N}[k]$ &${((N-1)2k-N)(N+k(2N-1))\over 12(k+N)}$\\ \hline
$E_{6}^{12}[k]$ &${(k-1)(12+13k)\over 2(12+k)}$& $E_6^{9}[k]$ &${(4k-3)(13k+9)\over 6(9+k)}$\\ \hline
$E_{6}^8[k]$ &${(3k-2)(13k+8)\over4(8+k)}$& $E_{7}^{18}[k]$ &${7(k-1)(19k+18)\over 12(18+k)}$\\ \hline
$E_7^{14}[k]$ &${(9k-7)(19k+14)\over 12(14+k)}$& $E_{8}^{30}[k]$ &${2(k-1)(30+31k)\over 3(30+k)}$\\ \hline
$E_{8}^{24}[k]$ &${(5k-4)(24+31k)\over 6(24+k)}$& $E_{8}^{20}[k]$ &${(3k-2)(20+31k)\over3(20+k)}$\\ \hline
\end{tabular}
\end{center}
\caption{Central charges for the AD theories without flavor symmetries. Note that this formula is only true for certain choice of $k$ and $N$ so that the irregular puncture is irreducible.}
\label{table:noflavor:centralcharges}
\end{table}
\newpage
\noindent \textbf{Regular singularity} One can also add a regular singularity on top of the irregular singularity to obtain Argyres-Douglas theories carrying a non-abelian flavor symmetry.
The regular singularity is classified by the embeddings $\Lambda_Y$ of $SU(2)$ into $J$, or equivalently by the nilpotent orbits of the Lie algebra $J$. The flavor symmetry associated to a puncture of type $Y$ is given by the commutant of $\Lambda_Y(\sigma^+)$ in $J$.
We list some special orbits in table \ref{special}.
\begin{table}[h]
\begin{center}
\begin{tabular}{|c|c|c|}
\hline
Nilpotent orbit & Dimension & Flavor symmetry \\ \hline
Maximal (Principal) & $\textrm{dim}J-\textrm{rank}(J)$ & N/A \\ \hline
Sub-regular &$ \textrm{dim}J-\textrm{rank}(J)-2$& ~ \\ \hline
Minimal& $d_1$ & ~ \\ \hline
Trivial & 0 & $J$ \\ \hline
\end{tabular}
\end{center}
\caption{Some special regular singularities.
Here $d_1$ is equal to one plus the number of positive roots not orthogonal to the highest root. }
\label{special}
\end{table}
We label our theory as $(J^b[k], Y)$ Here $Y$ represents the regular puncture (or its corresponding nilpotent orbit) and $J^b[k]$ denotes the irreducible irregular puncture (has no mass parameter). We often denote $F$ as the full (regular) puncture and $\varnothing$ as the null (or the absense of) puncture. We write the central charges\footnote{Here we use the normalized the flavor central charge $k_F$ so that a free hypermultiplet has $k_{SU(2)}=1$.} for such theories in table \ref{table:centralcharge:ADmatter}.
For $J=A_{N-1}$, the regular singularities are classified by the partitions of $N$ or Young Tableaux $Y=[n_1^{h_1},\ldots, n_s^{h_s}]$ with $\sum_i h_i n_i = N$, and the
corresponding flavor symmetry is given as
\begin{equation}
G_Y=\left(\prod_{i=1}^s U(h_i)\right)/ U(1).
\end{equation}
For the classification of the regular punctures and the corresponding flavor symmetries of other Lie algebras, see \cite{Chacaltana:2012zy}.
\begin{table}[!htb]
\begin{center}
\begin{tabular}{ |c|c|c|}
\hline
Theory &$c_{4d}$&$\frac{1}{2} k_F$ \\ \hline
$(A_{N-1}^N[k],F)$ &$\frac{1}{12}{(N+k-1)(N^2-1)}$& ${N(N+k-1)\over N+k}$\\ \hline
$(A_{N-1}^{N-1}[k],F)$&${(N+1)[N^2+N(k-2)+1]\over 12}$&${(N-1)^2+kN\over N+k-1}$\\ \hline
$(D_N^{2N-2}[k],F)$ &${1\over 12}N(2N-1)(2N+k-3)$ & ${(2N-2)(2N+k-3)\over 2N-2+k}$ \\ \hline
$(D_N^{N}[k],F)$ &${(2N-1)[2k(N-1)+N(2N-3)]\over 12}$ & ${2k(N-1)+N(2N-3)\over N+k} $ \\ \hline
$(E_{6}^{12}[k],F)$ &${13(k+11)\over 2}$ &${12(k+11)\over k+12}$ \\ \hline
$(E_{6}^{9}[k],F)$& ${13\over 6}(33+4k)$ &$12-{9\over k+9}$ \\ \hline
$(E_{6}^{8}[k],F)$&${13\over 4}(22+3k)$ &$12-{8\over k+8}$ \\ \hline
$(E_{7}^{18}[k],F)$&${133\over 12}(17+k)$ &${18(k+17)\over k+18}$ \\ \hline
$(E_7^{14}[k],F)$&${19\over 12}(119+9k)$ &$18-{14\over k+14}$ \\ \hline
$(E_{8}^{30}[k],F)$ & ${62\over 3} (29+k)$ &${30(k+29)\over k+30}$ \\ \hline
$(E_{8}^{24}[k],F)$ &${31\over 6} (116+5 k)$ & $30-{24\over k+24}$ \\ \hline
$(E_{8}^{20}[k],F)$&${31\over3} (58+3 k)$ &$30-{20\over k+20}$ \\ \hline
\end{tabular}
\end{center}
\caption{The central charge $c_{4d}$ and flavor central charge $k_F$ for the Argyres-Douglas matter. }
\label{table:centralcharge:ADmatter}
\end{table}
\noindent \textbf{Rank 1 SCFTs} One interesting class of 4d ${\cal N}=2$ SCFTs is the rank 1 SCFTs labeled by $H_0$, $H_1$, $H_2$, $D_4$, $E_6$, $E_7$, $E_8$. They can be realized as the world-volume theory of a single D3-brane probing the 7-brane singularities in F-theory. They have the global symmetry $\varnothing$, $SU(2)$, $SU(3)$, $SO(8)$, $E_6$, $E_7$, $E_8$ respectively. A noticeable feature is that the Higgs branch of each theory is given by the minimal nilpotent orbit of the flavor group, which is the same as the centered 1-instanton moduli space.
These theories can be realized in our setup as $(A_1^2[3])$,$(A_1^2[1], F)$,$(A_2^3[-1], F)$, $(D_4^4[-3], F)$, $(E_6^9[-8], F)$, $(E_7^{14}[-13], F)$, $(E_8^{24}[-23], F)$ respectively.
\subsection{Theory with 3d Lagrangian mirror}
We can reduce 4d $\mathcal{N}=2$ SCFTs to 3d and flow to IR to obtain 3d $\mathcal{N}=4$ SCFTs. For a 3d $\mathcal{N}=4$ SCFT A, it
is often possible to find a mirror SCFT B. The essential feature of 3d mirror symmetry is that the Coulomb branch of theory A is identified with the Higgs branch
of theory B, and vice versa. Since the 4d Higgs branch is the same as the Higgs branch of the 3d SCFT A obtained via dimensional reduction, we can use the Coulomb branch of the 3d mirror B to study the 4d Higgs branch. For example, if B admits a Lagrangian description, we can compute the Higgs branch index of A through computing the Coulomb branch index of B.
\begin{figure}[h!]
\centering
\includegraphics[width=6.0in]{enhanced.eps}
\caption{3d Mirror descriptions for some Argyres-Douglas theories considered in this paper, whose flavor symmetry is $SU$. Here, the number $n$ assigned to the edges mean that there are $n$ copies of the bifundamental hypermultiplet. As usual, the overall $U(1)$ factor has to be modded out.
$A$ is the general structure of a quiver whose flavor symmetry would be a simple $SU$ group. $B-F$ are examples.
We also list the flavor central charge here, and the corresponding theory in terms of the label $(J^b[k],F)$.}
\label{3dmirror}
\end{figure}
For $J=A_{N-1}$, we find the list of AD theories that admit Lagrangian 3d mirrors, as given in figure \ref{3dmirror}.
The idea for finding these theories is as follows: first, we use the method of Gaiotto-Witten \cite{Gaiotto:2008ak} to find the 3d ${\cal N}=4$ quivers whose flavor symmetry on the Coulomb branch is $SU(N)$. Then we use the method of \cite{Xie:2012hs} to find a Hitchin system realization where the Higgs field has integral order poles. From the Hitchin system, we can derive the 4d Coulomb branch spectrum. Finally, we search for a realization that fits into the class of theories we label as $(J^b[k],F)$ with $J=A_{N-1}$.
\subsection{Reducible irregular singularity}
In the last subsection, we considered the theory defined by an irregular singularity which does not admit any mass parameter.
For a general irregular singularity, it was found in \cite{Xie:2017vaf}
that it is useful to represent our theory by an auxiliary punctured sphere.
Let us consider 6d $A_{N-1}$ $(2,0)$ theory on a sphere with the following irregular singularity
\begin{align}
\Phi={T\over z^{2+{p\over q}}}+\ldots \ .
\end{align}
Here $(p,q)$ are co-prime and $T$ is regular semi-simple, and $N=rq$ or $N=rq+1$ with $r$ an integer. One can also add a regular singularity of $A_{N-1}$ type. The range of $(p,q)$ is restricted to
$q>0,~ p \geq -q +1$.
Depending on values of $p, q$, it is shown that one
can have the following representations in terms of auxiliary punctured sphere \cite{Xie:2017vaf}:
\begin{itemize}
\item If $p=0$, we get the usual class ${\cal S}$ theory on a sphere with regular singularity.
\item If $p=1, q=1$, the theory can be represented by a sphere with $r$ marked points of the black type $Y=[1]$.
\item If $p>1, q=1$, one can represent the theory by a sphere with $r$ marked points of the black type $Y=[1]$, and one extra red marked point representing the regular singularity.
\item If $q\neq 1$, one can represent the theory by a sphere with $r$ marked points of the black type, one marked point of the red type representing the regular singularity, and one marked point of the blue type.
\end{itemize}
For each marked point, one associate a Young Tableaux $Y$ with varying size (except for the class ${\cal S}$ case where the Young Tableaux has the fixed size determined by $J$). We refer to \cite{Xie:2017vaf} for more detail.
In the previous sections, we have considered the case of $r=1$ and we will be only considering this case in the rest of the current paper (notice that in general this condition implies that the irregular singularity has neither mass parameter nor exact marginal deformation). Therefore our theory can be represented by the following three punctured sphere: one trivial blue marked point, one red marked point representing regular singularity, and one black marked point of type $Y=[1]$.
We hope that the above representation of more general AD theories and the S-duality proposed in \cite{Xie:2017vaf} can be helpful in studying the indices of those theories.
\section{Vertex operator algebra of the AD theories} \label{sec:ADVOA}
We have introduced a class of AD theories labeled as $(J^b[k], Y)$. Here $J^b[k]$ represents an irregular singularity without any mass parameter, and $Y$ is an arbitrary
regular singularity. We also assume that the theory is `indecomposable,' namely has no exactly marginal operators.
It was shown in \cite{Beem:2013sza} that for any four-dimensional
$\mathcal{N}=2$ SCFT, one can associate a two-dimensional chiral algebra or vertex operator algebra. The basic correspondence is as follows:
\begin{itemize}
\item The 2d Virasoro central charge $c_{2d}$ is given in terms of the conformal anomaly $c_{4d}$ of the 4d theory as
\begin{eqnarray}
c_{2d} = -12 c_{4d} \ .
\end{eqnarray}
\item The global symmetry algebra $\mathfrak{g}$ becomes an affine Kac-Moody algebra $\hat{\mathfrak{g}}_{k_{2d}}$ and the level of affine Kac-Moody algebra $k_{2d}$ is given by the 4d the flavor central charge $k_F$ as
\begin{equation}
k_{2d}=-\frac{1}{2} k_F \ .
\label{eq:centralchargerelation}
\end{equation}
\item The (normalized) vacuum character of the chiral algebra/VOA is identical to the Schur index of the 4d theory:
\begin{align}
\chi_0(q) = {\cal I}_{\textrm{Schur}}(q) \ .
\end{align}
\end{itemize}
The VOAs corresponding to the theories $(J^b[k], F)$ or $J^b[k] \equiv (J^b[k], \varnothing)$ with irreducible irregular puncture $J^b[k]$ were conjectured in \cite{Xie:2016evu}.
Along with the same line as in the case of the usual class ${\cal S}$ theory \cite{Beem:2014rza}, we can further conjecture VOA for the general choice of the puncture $Y$:
\begin{Conjecture}
The VOA corresponding to $(J^b[k], Y)$ is given by the W-algebra ${\cal W}^{k_{2d}}(J, Y)$.
Here $k_{2d}=-h+{b\over b+k}$ and ${\cal W}^{k_{2d}}(J, Y)$ algebra is the quantum Drinfeld-Sokolov reduction \cite{Drinfeld:1984qv, Bershadsky:1989mf,Bershadsky:1989tc,Feigin:1990pn} of the Kac-Moody algebra
$\widehat{J}_{k_{2d}}$ using the nilpotent orbit (or the $SU(2)$ embedding into $J$) associated with $Y$.
\end{Conjecture}
Two extreme situations where $Y$ being $F$(full) or $\varnothing$(null) have been considered in \cite{Xie:2016evu}, and the corresponding VOAs are the affine Kac-Moody algebra $\widehat{J}$ and the standard $W$-algebra associated with $J$.
For a general choice of $Y$, we obtain the VOA for the corresponding theory via quantum Drinfeld-Sokolov (qDS) reduction.
The central charge of the vertex operator algebra ${\cal W}^{k}(J, Y)$ is given as \cite{deBoer:1993iz}
\begin{align}
c_{2d}=\textrm{dim}J_0 -{1\over 2} \textrm{dim} J_{1/2}-{12\over k+h} \big|\rho-(k+h)x \big|^2 \ ,
\label{2dcentral}
\end{align}
where $h$ is the dual Coxeter number, and $\rho$ is the longest root.
Let us explain the notation: for a nilpotent orbit, one has a $su(2)$ triple $(x,e,f)$ so that $[x,e]=e,~[x,f]=-f, [e,f]=x$. Here $x$ is a semi-simple
element of the Lie algebra $J$, and $f$ is a nilpotent element. This can be also obtained by choosing a $su(2)$ embedding $\Lambda_Y$ into $J$ given by the choice of $Y$. It gives $(x, e, f) = (\Lambda_Y(\sigma^3), \Lambda_Y(\sigma^+), \Lambda_Y(\sigma^-))$. Now, $J$ has an eigenspace decomposition under the adjoint action of $x$ as:
\begin{align}
J=\bigoplus_{j\in {1\over 2}\mathbb{Z}} J_j \ ,
\end{align}
Let $J^f$ be the centralizer of $f$ in $J$, which also has a grading by the adjoint action of $x$:
\begin{equation}
J^f= \bigoplus_{j \in \frac{1}{2} \mathbb{Z}} J_{-j}^f
\end{equation}
For us, $f$ is an element of the nilpotent orbit associated to $Y$. The algebra ${\cal W}^{k_F}(J, Y)$ is strongly generated by $\textrm{dim}J^f$ fields $X^a(z)$, where $a$ runs over
a basis of $J^f$. The generator $X^a(z)$ has a conformal weight $1+j$ if $X^a\in J_{-j}^f$.
The level $k_{2d}$ of an affine Kac-Moody algebra $\widehat{J}_{k_{2d}}$ (for $J=ADE$) is called \emph{admissible}
if it can be written in the following way
\begin{equation}
k_{2d} =-h+{p\over q},~~(p,q)=1,~~p\geq h.
\end{equation}
So only for $b=h$, the level of our conjectural VOA is admissible.
The vacuum character of the corresponding affine Kac-Moody algebra with admissible level and the $W$-algebra obtained from qDS are given in \cite{Kac:1988qc} and \cite{Frenkel:1992ju} respectively.
There is one simple consistency check we can do: the 4d central charge can be computed using the method presented in \cite{Xie:2012hs,Xie:2013jc} and can be used to check
with 2d central charge formula (\ref{2dcentral}).
\paragraph{Examples}
Let us study several interesting class of examples.
$(A_1, ADE)$\footnote{Here $(G_1, G_2)$ means that the corresponding BPS quiver is a product of two Dynkin diagrams for $G_1$ and $G_2$.} theories are a class of well studied Argyes-Douglas theories.
The $(A_1, A_{2N})$ theory can be realized by the choice $J=A_1, b=2, k=2N+1$ with $Y=\varnothing$ being trivial. The $(A_1, D_{2N-1})$ theory can be realized from $J=A_1, b=2, k=2N-3$ with $Y=F$ being the full puncture.
The $(A_1, E_6)$ theory can be realized from $J=A_2, b=3, k=4$ with $Y=\varnothing$, and the $(A_1, E_8)$ theory can be realized from $J=A_2, b=3, k=5$ with $Y=\varnothing$. The corresponding VOAs are listed in table \ref{A1ADE}.
In the following, we will find VOAs for other AD theories in this class.
\textbf{Example 1}: Consider the case with $J=A_{N-1},~ b=N,~k=1$, and $Y=[N-1,1]$. This theory is identical to the $(A_1, A_{2N-1})$ theory and has $U(1)$ global symmetry (except for $N=2$, which has enhanced $SU(2)$ symmetry).
The corresponding VOA is ${\cal W}^{-{N^2\over N+1}}(A_{N-1}, Y)$.
\textbf{Example 2}: Consider the theory defined by data $J=A_{N-1}, ~b=N,~k=-1$ and $Y=[N-2,1,1]$. It gives the $(A_1, D_{2N-2})$ theory which has $SU(2)\times U(1)$ flavor symmetry (except for $N=3$, which has enhanced $SU(3)$ symmetry).
The corresponding VOA is ${\cal W}^{{-N^2+2N\over N-1}}(A_{N-1}, Y)$.
\textbf{Example 3}: Choose $J=A_{N-1}, ~b=N-1,~(k,b)=1$ and $Y$ to be trivial.
The same theory can be also obtained from $J=A_{N-1},~b=N,~k=N-1$, and a simple puncture $Y=[N-1,1]$. This also belongs to our list. The $(A_1, E_7)$ theory is identical to $(A_2^2[3], [2, 1])$. Using the other description, we find that the corresponding VOA is ${\cal W}^{-{13\over5}}(A_2, [2, 1])$.
The complete results are summarized in table \ref{A1ADE}. We are able to recover what was found in \cite{Creutzig:2017qyf}.
These theories can also be realized in other ways, for example $(A_1, A_{2N})$ theory can be realized using $J=A_{2N}$. In general, there are many different M5-brane realizations of the identical 4d AD theory. Whenever this happens, we find isomorphisms between W-algebras.
\begin{table}[h]
\begin{center}
\begin{tabular}{|c|c|c|c|}
\hline
$(G, G')$ & $(J^b[k], Y)$ &VOA&$Y$ \\ \hline
$(A_1, A_{2N})$ & $(A_1^2 [2N+1], Y) $ &${\cal W}^{-2+{2\over 2N+3}}(A_1, Y)$& $[2]$ \\ \hline
$(A_1, A_{2N-1})$& $(A_{N-1}^N[1], Y)$ & ${\cal W}^{{-N^2\over N+1}}(A_{N-1}, Y)$&$ [N-1,1]$ \\ \hline
$(A_1, D_{2N-1})$& $(A_1^2 [2N-3], Y)$ &${\cal W}^{-2+{2\over 2N-1}}(A_1, Y)$& $[1,1]$\\ \hline
$(A_1, D_{2N-2})$& $(A_{N-1}^N[-1], Y)$ &${\cal W}^{{-N^2+2N\over N-1}}(A_{N-1}, Y)$& $[N-2,1^2]$ \\ \hline
$(A_1, E_{6})$&$(A_2^3[4], Y)$ & ${\cal W}^{-{18\over 7}}(A_2, Y)$& $[3]$ \\ \hline
$(A_1, E_{7})$&$(A_2^2[3], Y) $ &${\cal W}^{-{13\over5}}(A_2, Y)$& $[2,1]$ \\ \hline
$(A_1, E_{8})$&$(A_2^3[5], Y)$ & ${\cal W}^{-{21\over 8}}(A_2, Y)$& $[3]$ \\ \hline
\end{tabular}
\end{center}
\caption{The VOA of $(A_1, ADE)$ theories. Here $Y=[n_1,n_2,\ldots, n_s]$ means that the nilpotent orbit has the Jordan block with size $n_1,\ldots, n_s$. }
\label{A1ADE}
\end{table}
\paragraph{Schur index and the vacuum character}
The Schur index is given by the vacuum character of the corresponding VOA.
Here, let us consider the case $b=h$ and $(k,h)=1$, then the VOA for the $J^h[k])$ theory is the same as the $W^J[h+k,h]$ minimal model.
The vacuum character takes the following simple form \cite{Kac:1988qc, Frenkel:1992ju, Bouwknegt:1992wg}\footnote{For $J=A_{N-1}$, the character is already considered in \cite{Cordova:2015nma}.}
\begin{equation}\label{eq:chaofminimalmodel}
\chi(q)={1\over \eta(\tau)^r}\sum_{\omega \in \hat{W}} \epsilon(w)q^{{1\over 2 p p'}|p'w(\Lambda^{+}+\rho)-p(\Lambda^{-}+\rho)|^2},
\end{equation}
where $r$ is the rank of the Lie algebra $J$, $\rho$ is the Weyl vector, $\hat{W}$ is the affine Weyl group, and $\epsilon(w)$ is the signature of the affine Weyl group element.
$\Lambda^{+}\in P_{+}^{0}$ and $\Lambda^{-}\in P_{+}^{k}$ are principle admissible weights such that:
\begin{equation}
{1\over 12}r h(h+1)(p'-p)^2=|p'(\Lambda^{+}+\rho)-p(\Lambda^{-}+\rho)|^2.
\end{equation}
We have the solution: $\Lambda^{+}=0$ and $\Lambda^{-}=(k,0,\ldots,0)$. Substituting into the character formula, we get the Schur index of our 4d theory $J^h[k]$.
In the next section, we will show that TQFT structure of AD theory dictates the Schur index of $J^h[k]$ theory to be the simple form
\begin{equation}
{\cal I}_{J^h [k]} (q) = \frac{1}{\prod_{i=1}^r (q^{d_i}; q)} \mathrm{PE} \left[ - \frac{q^{h+k}}{1-q^{h+k}} \chi_{\textrm{adj}} (q^{\vec\rho}) \right] \ ,
\end{equation}
where $d_i$ are the degrees of the Casimirs of $J$.
We have checked that this formula matches with equation \ref{eq:chaofminimalmodel} up to a high power in $q$.
It is viable to compute the vacuum character of ${\widehat{J}}_{-\frac{1}{2} k_F}$ hence the Schur index of $(J^b[k],F)$ by using the Kac-Wakimoto formula. However, we propose a much compact formula for $b=h$ case,
\begin{equation}
{\cal I}_{(J^h[k], F)} = \mathrm{PE} \left[ \frac{q - q^{h+k}}{(1-q)(1-q^{h+k})}\chi^F_{\textrm{adj}} (\vec{z}) \right] \ .
\end{equation}
A derivation of this formula is given in section \ref{sec:Schur}.
\section{Schur index and TQFT} \label{sec:Schur}
In this section, we derive a universal formula for the Schur index of the AD theory using the TQFT structure of the index, which provides a strong check for the identification of the corresponding VOA.
\subsection{Wave function for the irregular puncture}
As an intermediate step, we first write the Schur index of the pure Yang-Mills theory in a TQFT form. Even though the superconformal index is properly defined only for a conformal theory, there is increasing evidence that the Schur index makes sense even for a non-conformal theory. See for example \cite{Cordova:2015nma, Cecotti:2015lab, Cordova:2016uwk}.
The Schur index for the pure YM theory of gauge group $J$ is given by
\begin{eqnarray}
{\cal I}_{\textrm{SYM}}^{J} = (q; q)^r \oint [d\vec{z}] \prod_{\vec{\alpha} \in \Delta_{J}} (q \vec{z}^{\vec{\alpha}}; q)^2 \ ,
\end{eqnarray}
where $r$ is the rank of the gauge group and $\Delta_J$ is the set of all roots. We write the measure factor as
\begin{eqnarray}
[d\vec{z}] = \frac{1}{|W_J|} \prod_{i=1}^r \frac{dz_i}{ 2\pi i z_i} \prod_{\vec{\alpha} \in \Delta_J} (1-\vec{z}^{\vec\alpha}) \ ,
\end{eqnarray}
where $|W_J|$ denotes the order of Weyl group of $J$. Here we used the short-hand notation $\vec{z}^{\vec{\alpha}} \equiv \prod_i z_i^{\alpha_i}$.
The pure YM theory can be obtained from 6d ${\cal N}=(2, 0)$ theory. When $J$ is simply-laced, we pick $J$-type 6d ${\cal N}=(2, 0)$ theory and compactify on a sphere with two identical irregular punctures. The irregular puncture, which we denote as $I_J \equiv {J^{h}[-h+1]}$ (here $h$ is the dual Coxeter number of $J$), realizes a singularity of the form
\begin{eqnarray}
\Phi (z) = \frac{T}{z^{2-\frac{h-1}{h}}} + \cdots = \frac{T}{z^{1+\frac{1}{h}}} + \cdots \ .
\end{eqnarray}
From the TQFT structure of the index \cite{Gadde:2009kb,Gadde:2011ik,Gadde:2011uv,Gaiotto:2012xa}, the Schur index should be given by
\begin{eqnarray} \label{eq:SchurYMTQFT}
{\cal I}_{SYM}^{J} = \sum_{\vec \lambda} \psi_{\vec \lambda}^{I_{J}} \psi_{\vec \lambda}^{I_{J}} \ .
\end{eqnarray}
The wave function for the puncture is given by \cite{Song:2015wta}
\begin{eqnarray} \label{eq:SchurYMWaveFtn}
\psi_{\vec \lambda}^{I_{J}} = \oint [d\vec{z}] \mathrm{PE} \left[ -\frac{q}{1-q} \chi_{\textrm{adj}}(\vec{z}) \right] \chi_{\vec \lambda}(\vec{z}) \ ,
\end{eqnarray}
where $\chi_{\vec \lambda} (\vec{z}) $ is the character (or Schur function) of $J$ for the representation $\vec \lambda$. This is an analog of the Gaiotto state \cite{Gaiotto:2009ma,Keller:2011ek} in the AGT correspondence \cite{Alday:2009aq}.
It is easy to verify that \eqref{eq:SchurYMWaveFtn} reproduces the Schur index for the pure YM. Plugging in the wave function to the RHS of \eqref{eq:SchurYMTQFT}, we get
\begin{align}
\begin{split}
\sum_{\vec \lambda} \psi_{\vec \lambda}^{I_{J}} \psi_{\vec \lambda}^{I_{J}}
&= \oint [d\vec{z}][d\vec{z'}] \mathrm{PE} \left[ - \frac{q}{1-q} \left( \chi_{\textrm{adj}} (\vec{z}) + \chi_{\textrm{adj}} (\vec{z'})\right) \right] \chi_{\vec \lambda} (\vec{z}) \chi_{\vec \lambda} (\vec{z'}) \\
&= \oint [d\vec{z}] \mathrm{PE} \left[ - \frac{2q}{1-q} \chi_{\textrm{adj}}(\vec{z}) \right] = {\cal I}_{SYM}^{J} \ .
\end{split}
\end{align}
Here we used the relation $\sum_{\vec \lambda} \chi_{\vec \lambda}(\vec{z})\chi_{\vec \lambda}(\vec{z'}) = \frac{1}{\Delta_J(z)}\delta (\vec{z}-\vec{z'})$, where $\Delta_J (z)$ is the Haar measure.
The wave functions for the $I_{SU(N)}$ are evaluated when $N=2, 3$ \cite{Song:2015wta}. For $N=2$, we get
\begin{align}
\psi_\lambda^{I_{SU(2)}}(q)
&= \begin{cases} (-1)^{\frac{\lambda}{2}} q^{\frac{1}{2} \frac{\lambda}{2} (\frac{\lambda}{2}+1)} & \mbox{$\lambda$ even}, ~~ \\
0 & \mbox{$\lambda$ odd}.
\end{cases}
\end{align}
When $N=3$, we get
\begin{eqnarray}
\psi^{I_{SU(3)}}_{(\lambda_1, \lambda_2)}(q) =
\begin{cases}
q^{k(k+1)+\ell(\ell+1)+k \ell} & \mbox{if $\lambda_1 = 3k, \lambda_2 = 3\ell$} , \\
-q^{k^2+\ell^2-1+(k-1)(\ell-1)} & \mbox{if $\lambda_1 = 3k-2, \lambda_2 = 3 \ell-2 $} , \\
0 & \mbox{otherwise} ,
\end{cases}
\end{eqnarray}
where $k, \ell \in \mathbb{Z}_{\ge 0}$.
The wave function for the irregular puncture $I_{N, k} \equiv A_{N-1}^N[k]$ is conjectured to be given by \cite{Song:2015wta}
\begin{eqnarray}
\psi_{\vec\lambda}^{I_{N, k}} (q) = \psi_{\vec\lambda}^{I_{SU(N)}}(q^{N+k}) \ .
\end{eqnarray}
Here, we further conjecture that the wave function for the irreducible irregular puncture $J^h[k]$ (with $b=h$) with no flavor symmetry (where the condition is summarized in table \ref{table:constraint}) is given by
\begin{eqnarray} \label{eq:IrrSchur}
\boxed{ \psi_{\vec \lambda}^{J^h[k]}(q) = \psi_{\vec{\lambda}}^{I_J} (q^{h+k}) } \ .
\end{eqnarray}
In the following, we compute the Schur indices for a number of examples assuming the relation \eqref{eq:IrrSchur} and find that it agrees with the vacuum character of the VOA.
\subsection{AD theories of type $J^{h}[k]$}
Let us consider the theory of type $J^{h}[k] \equiv (J^h[k], \varnothing)$, where $h$ is the dual Coxeter number of $J$. As before, we can write the wave function for the irregular puncture realizing the analog of Gaiotto-Whittaker state for the pure Yang-Mills as
\begin{eqnarray}
\psi_{\vec{\lambda}}^{I_J} (q) = \oint [d\vec{z}] \mathrm{PE} \left[ - \frac{q}{1-q} \chi_{\textrm{adj}} (\vec{z}) \right] \chi_{\vec{\lambda}}(\vec{z}) \ ,
\end{eqnarray}
where $\vec{\lambda}$ is a Dynkin label for the representation of $J$. From the TQFT structure, we should have
\begin{eqnarray}
{\cal I}_{J^{h} [k]} (q) = \sum_{\vec \lambda} C_{\vec\lambda}^{-1}(q) \psi_{\vec \lambda}^{J^h[k]}(q)
= \sum_{\vec \lambda} \frac{\chi_{\vec\lambda}(q^{\vec\rho})}{\prod_{i=1}^r (q^{d_i}; q)} \psi_{\vec \lambda}^{J^h[k]}(q) \ ,
\end{eqnarray}
where $r$ is the rank of group $J$, and $d_i$ are the degrees of the Casimirs of $J$. We conjecture that the wave function for the general irregular puncture is given by
\begin{eqnarray}
\psi_{\vec \lambda}^{J^h[k]}(q) = \psi_{\vec{\lambda}}^{I_J} (q^{h+k}) \ .
\end{eqnarray}
If we assume this relation, the above expression for the index can be evaluated easily using the completeness of the characters. We get
\begin{align} \label{eq:SchurAA}
\begin{split}
{\cal I}_{J^h [k]} (q) &= \frac{1}{\prod_{i=1}^r (q^{d_i}; q)} \oint [d\vec{z}] \mathrm{PE} \left[ - \frac{q^{h+k}}{1-q^{h+k}} \chi_{\textrm{adj}} (\vec{z}) \right] \sum_{\vec\lambda} \chi_{\vec\lambda} (q^{\vec\rho}) \chi_{\vec\lambda} (\vec{z}) \\
&= \frac{1}{\prod_{i=1}^r (q^{d_i}; q)} \mathrm{PE} \left[ - \frac{q^{h+k}}{1-q^{h+k}} \chi_{\textrm{adj}} (q^{\vec\rho}) \right] \ .
\end{split}
\end{align}
Here we used the completeness of the character $\sum_{\vec\lambda} \chi_{\vec\lambda} (\vec{z})\chi_{\vec\lambda} (\vec{w}) = \delta(\vec{z}-\vec{w})$.
The character of the adjoint evaluated at $\vec{z}=q^{\vec\rho}$ is given by
\begin{eqnarray}
\chi_{\textrm{adj}} (q^{\vec\rho}) = \sum_{i=1}^{r} \chi_{d_i - 1} (q) = \sum_{i=1}^r [2 d_i -1]_q \ ,
\end{eqnarray}
where $\chi_j$ denotes the character of the spin-$j$ representation of $SU(2)$.
\paragraph{Example: $A_{N-1}^N [k] = (A_{N-1}, A_{k-1})$ theory}
This theory is engineered by putting 6d ${\cal N}=(2, 0)$ theory of type $J=A_{N-1}$ on a sphere with $I_{N, k} \equiv A_{N-1}^N[k]$ puncture.
It was conjectured \cite{Cordova:2015nma,Song:2015wta} that the Schur index of this theory is given by the vacuum character of the $W(k, k+N)$ minimal model. The vacuum character for the $W(k, k+N)$ minimal model is given in \cite{Andrews1999a} as
\begin{align}
\chi_0^{W(k, k+N)}(q) = \left( \frac{(q^{k+N}; q^{k+N})}{(q;q)} \right)^{k-1} \prod_{a=1}^{k-1} (q^{N+a}; q^{k+N})^a (q^a; q^{k+N})^{k-a} \ .
\end{align}
We can rewrite the character using PE as
\begin{align}\label{eq:Wchar}
\begin{split}
\chi_0^{W(k, k+N)}(q) &= \mathrm{PE} \left[ (k-1)\left( \frac{q}{1-q} - \frac{q^{k+N}}{1-q^{k+N}} \right) - \sum_{a=1}^{k-1} \frac{a q^{N+a} + (k-a)q^a }{1-q^{k+N}} \right] \\
&= \mathrm{PE} \left[ \frac{(q-q^k)(q-q^N)}{(1-q)^2 (1-q^{k+N})}\right] \ .
\end{split}
\end{align}
Note that the character is symmetric under the exchange of $k$ and $N$.
Now, let us prove that \eqref{eq:SchurAA} indeed reproduces the corresponding character. We have
\begin{align}
\begin{split}
{\cal I}_{A_{N-1}^N [k]} (q)
&= \frac{1}{{\prod_{i=2}^N (q^i; q)}} \mathrm{PE} \left[ - \frac{q^{N+k}}{1-q^{N+k}} \chi_{\textrm{adj}} (q^{\vec{\rho}}) \right] \ .
\end{split}
\end{align}
Note that $\chi_{\textrm{adj}}^{SU(N)} (q^{\vec\rho}) = ([N]_q)^2 - 1$, where $[N]_q = \frac{q^{N/2}-q^{-N/2}}{q^{1/2}-q^{-1/2}}$. Putting this back to the expression above, we obtain
\begin{align}
\begin{split}
{\cal I}_{A_{N-1}^N [k]} (q) &= \mathrm{PE} \left[ \frac{q^2 + \cdots + q^N}{1-q}- \frac{q^{N+k}}{1-q^{N+k}} \left( \frac{q^{-N} \left(q - q^N\right) \left(1 - q^{N+1}\right)}{(1-q)^2} \right) \right] \\
&= \mathrm{PE} \left[ \frac{(1-q^{N-1})(q^2 - q^{k+1})}{(1-q)^2 (1-q^{k+N})} \right] \ .
\end{split}
\end{align}
This is exactly equal to the expression in \eqref{eq:Wchar}, so the proof is done.
\paragraph{Example: $D_N^{2N-2}[k]$ theory}
When $J = D_N$, we have
\begin{eqnarray}
\chi_{\textrm{adj}} (q^{\vec\rho}) = \frac{q^{1-2N}(1-q^N)(1-q^{2N-1})(q^2+q^N)}{(1-q)(1-q^2)} \ ,
\end{eqnarray}
so that we obtain
\begin{align}
{\cal I}_{D_N^{2N-2} [k]} (q) &= \mathrm{PE} \left[ \frac{q^2 + q^4 + \cdots q^{2N-2} + q^N}{1-q} - \frac{q^{2N-2+k}}{1-q^{2N-2+k}} \frac{q^{1-2N}(1-q^N)(1-q^{2N-1})(q^2+q^N)}{(1-q)(1-q^2)} \right] \nonumber \\
&= \mathrm{PE} \left[ \frac{(1-q^N)(q^2+q^N)(1 - q^{k-1})}{(1-q)(1-q^2)(1-q^{2N-2+k})} \right] \ .
\end{align}
This is expected to be the closed form formula for the character of the corresponding VOA ${\cal W}({D_N})$ with the central charge given by
\begin{align}
c_{2d} = - \frac{r(k-1)(h + k +hk)}{h+k} \ ,
\end{align}
where $r=N$, $h=2N-2$.
\subsection{AD theories of type $(J^{h}[k], Y)$}
\paragraph{$(J^h[k], F)$ theory}
The $(J^h[k], F)$ theory is engineered by adding a full regular puncture in addition to $J^h[k]$ puncture on a sphere. From the TQFT structure, it is straight-forward to write the Schur index once we know the corresponding wave functions for the punctures. The wave function for the irregular puncture $J^h[k]$ is given as in \eqref{eq:IrrSchur}. The wave function for the full regular puncture is given by
\begin{align}
\psi_{\vec\lambda}(\vec{z}) = \mathrm{PE} \left[ \frac{q}{1-q} \chi_{\textrm{adj}}(\vec{z}) \right] \chi_{\vec\lambda}(\vec{z}) \ .
\end{align}
Now, combining the above two, we obtain
\begin{align}
\begin{split}
I_{(J^b[k], F)} &= \sum_{\vec\lambda} \psi_{\vec \lambda}^{J^h[k]}(q) \psi_{\vec \lambda} (\vec{z}) \\
&= \oint [d\vec{x}] \mathrm{PE} \left[ - \frac{q^{h+k}}{1-q^{h+k}} \chi_{\textrm{adj}} (\vec{x}) \right] \mathrm{PE} \left[ \frac{q}{1-q} \chi_{\textrm{adj}} (\vec{z}) \right] \sum_{\vec\lambda} \chi_{\vec\lambda} (\vec{x}) \chi_{\vec\lambda} (\vec{z}) \\
&= \mathrm{PE} \left[ \frac{q - q^{h+k}}{(1-q)(1-q^{h+k})}\chi^F_{\textrm{adj}} (\vec{z}) \right] \ .
\end{split}
\end{align}
This is expected to be the character for the VOA given by the affine Lie algebra $J_{-\frac{1}{2} k_F}$. We have checked this result against a number of examples as we discussed in the previous section.
Our conjecture here has been proven recently in \cite{kac2017remark}.
\paragraph{$(J^h[k], Y)$ theory}
From the Jacobson-Morozov theorem every nilpotent orbit are in one-to-one correspondence with the $SU(2)$ embedding. Therefore, one can label the puncture by a $SU(2)$ embedding $\Lambda$ to $J$. Let us denote $\Lambda_Y$ to be the $SU(2)$ embedding into $J$. When $J=A_{N-1}$, the $SU(2)$ embeddings are labelled by partitions of $N$. One can obtain the theory corresponding to the general puncture by starting with the full regular puncture and performing a partial closure. From the field theory side, this means that we Higgs the $(J^h[k], F)$ theory by giving a nilpotent vev $\Lambda_Y(\sigma^+)$ to the moment map operator for the flavor group $J$. In terms of the VOA, it can be realized as a quantum Drinfeld-Sokolov reduction by the $SU(2)$ embedding $\Lambda_Y$.
The TQFT prescriptions for the general regular punctures are studied in various places \cite{Gadde:2011ik,Gadde:2011uv,Lemos:2012ph,Mekareeya:2012tn,Gadde:2013fma}. Let us decompose the adjoint representation of $J$ in terms of the $SU(2)$ embedding $\Lambda_Y$:
\begin{align} \label{eq:slodowy}
\textrm{adj} \to \bigoplus_j R_j \otimes V_j \ ,
\end{align}
where $R_j$ is a representation of the commutant of $\Lambda_Y$ in $J$, and the $V_j$ is the spin-$j$ irreducible representation of $SU(2)$. The $\oplus_j R_j$ gives the Slodowy slice of $J$ given by $Y$.
Now, for the Schur index, the wave function for the regular puncture of type $Y$ is given by
\begin{align}
\psi^{Y}_{\vec\lambda}(\vec{a}) = \mathrm{PE} \left[\sum_j \frac{q^{j+1}}{1-q} {\rm Tr}_{R_j} (\vec{a}) \right] \chi_{\vec\lambda} ( \vec{a}q^{\Lambda_Y} ) \ ,
\end{align}
where $q^{\Lambda_Y}$ is defined as the image of the map $\Lambda_Y: SU(2) \to J$. For example, when $J=SU(9)$ and $\Lambda_Y$ is given by the partition $[3, 2, 2, 1, 1]$, the commutant is given by $S[U(1)U(2)^2]$ and we set $\vec{a}q^{\Lambda_Y} = (a q^1, a q^0, a q^{-1}, b q^{\frac{1}{2}}, b q^{-\frac{1}{2}}, cq^{\frac{1}{2}}, cq^{-\frac{1}{2}}, d, e )$ with $abcde = 1$.
From this, it is easy to compute the Schur index for the $(J^h[k], Y)$ theory. We obtain
\begin{align}
\label{eq:schur1ir1r}
{\cal I}_{(J^b[k], Y)} &= \sum_{\vec\lambda} \psi_{\vec \lambda}^{J^h[k]}(q) \psi^{Y}_{\vec\lambda}(\vec{a}) \nonumber \\
&= \oint [d\vec{x}] \mathrm{PE} \left[ - \frac{q^{h+k}}{1-q^{h+k}} \chi_{\textrm{adj}} (\vec{x})
+ \sum_j \frac{q^{j+1}}{1-q} {\rm Tr}_{R_j} (\vec{a}) \right] \sum_{\vec\lambda} \chi_{\vec\lambda} (\vec{x}) \chi_{\vec\lambda} ( \vec{a}q^{\Lambda_Y} ) \nonumber \\
&= \mathrm{PE} \left[\sum_j \frac{q^{j+1}}{1-q} {\rm Tr}_{R_j} (\vec{a}) - \frac{q^{h+k}}{1-q^{h+k}} \chi_{\textrm{adj}} (\vec{a}q^{\Lambda_Y}) \right] \ .
\end{align}
When $Y$ is the null puncture $N$ or the $\Lambda_Y$ is given by the principal embedding, the commutant of $\Lambda_Y$ is empty and $\textrm{adj} \to \oplus_{d_i} V_{d_i}$ where $d_i$ are the degrees of Casimirs. In this case, $q^{\Lambda_Y} = q^{\vec{\rho}}$, so that we reproduce the result of \eqref{eq:SchurAA}.
\paragraph{Example: $(A_{N-1}^N[k], [N]) = A_{N-1}^N[k]$ theory} Let us consider the example of $(A_{N-1}^N[k], Y) $ with the puncture $Y$ fully closed to a null puncture. The theory should be the same as the theory on a sphere with only one irregular puncture of type $A_{N-1}^N[k] (=I_{N,k})$.
Following equation \eqref{eq:schur1ir1r}, the Schur index is given by
\begin{equation}
{\cal I}_{(A_{N-1}^N[k],N)}= \mathrm{PE}\left[\sum_{i=2}^{N}\frac{q^i}{1-q}-\frac{q^{N+k}}{1-q^{N+k}}\chi_{\textrm{adj}}^{SU(N)} (q^{\vec{\rho}})\right]={\cal I}_{A_{N-1}^N[k]},
\end{equation}
which is the same as the Schur index of $A_{N-1}^N[k]$ theory.
\paragraph{Example: $(A_{N-1}^N[k], [N-1,1]) $ theory} Let us consider the example of $(A_{N-1}^N[k], S)=(I_{N, k}, S)$ theory.
This theory is engineered by putting 6d ${\cal N}=(2, 0)$ theory of type $J = A_{N-1}$ on a sphere with irregular puncture of type $A_{N-1}^N[k]$ (=$I_{N, k}$) and a simple regular puncture. A simple puncture $S$ has $\Lambda_S$ given by the partition $[N-1,1]$ for $SU(N)$. The commutant is $U(1)$ and $\textrm{adj} \to \oplus_{j=0}^{N-2}V_{j} \oplus ((U(1)_2 \oplus U(1)_{-2}) \otimes V_{N/2 - 1})$. Its Schur index is,
\begin{equation}
\label{eq:simplepunc}
{\cal I}_{(A_{N-1}^N[k],S)}(z) = \mathrm{PE}\left[\sum_{i=1}^{N-1}\frac{q^i}{1-q} +\frac{q^{\frac{N}{2}}}{1-q}(z^2+z^{-2})-\frac{q^{N+k}}{1-q^{N+k}}\chi_{\textrm{adj}}^{SU(N)} (zq^{\frac{N-1}{2}},\cdots,zq^{-\frac{N-1}{2}},z^{-1})\right].
\end{equation}
When $k=1$, $(A_{N-1}^N[1], [N-1,1])$ is equivalent with $(A_1, A_{2N-1})$ theory and the corresponding VOA is ${\cal W}^{{-N^2\over N+1}}(A_{N-1}, [N-1,1])$. For example, the Schur index of $(A_{2}^3[1], [N-1,1])=(A_1,A_5)$ can be worked out based on equation \eqref{eq:simplepunc},
\begin{equation}
\begin{split}
&{\cal I}_{(A_1, A_5)}(z) = 1+q+(z^2+z^{-2})q^{\frac{3}{2}} + 3q^2 + (2z^2+2z^{-2})q^{\frac{5}{2} }+ (z^4+5+z^{-4})q^3+\cdots.\\
\end{split}
\end{equation}
The results match perfectly with the vacuum character of ${\cal W}^{{- \frac{9}{4}}}(A_{2}, [2,1])$ \cite{Cordova:2015nma,Buican:2015ina,Song:2015wta}.
\paragraph{Example: $(A_{N-1}^N[k], F) = (I_{N, k}, F)$ theory}
Let us consider the example of $(A_{N-1}^N[k], F)=(I_{N, k}, F)$ theory.
This theory is engineered by putting 6d ${\cal N}=(2, 0)$ theory of type $J = A_{N-1}$ on a sphere with irregular puncture of type $A_{N-1}^N[k]$ (=$I_{N, k}$) and a full regular puncture \cite{Xie:2012hs,Xie:2013jc}. When $N$ and $k$ are coprime, this theory is conjectured to have the VOA \cite{Beem:2013sza,Beem:2014rza} given by $\widehat{\mathfrak{su}}(N)_{-\frac{N(N-1+k)}{N+k}}$ \cite{Xie:2016evu}. From the TQFT structure, the Schur index is given by
\begin{eqnarray}
{\cal I}_{(A_{N-1}^N [k], F)} (\vec{z}) = \mathrm{PE} \left[ \frac{q - q^{N+k}}{(1-q)(1-q^{N+k})} \chi_{\textrm{adj}}^{SU(N)} (\vec{z}) \right] \ .
\end{eqnarray}
For example, when $N=2n+1$ and $k=-2n+1$, we get
\begin{eqnarray}
{\cal I}_{(A_{2n}^{2n+1}[-2n+1], F)} = \mathrm{PE} \left [ \frac{q}{1-q^2} \chi_{\textrm{adj}}^{SU(2n+1)} (\vec{z}) \right] \ ,
\end{eqnarray}
which proves the conjecture made in \cite{Xie:2016evu}.
When $N=2$, this theory is identical to $(A_1, D_{2+k})$. Especially, when $k=2n-1$, from the TQFT and wave functions, we obtain
\begin{eqnarray}
{\cal I}_{(A_1^2 [2n-1], F)} (z) = \frac{1}{(q z^{\pm 2, 0}; q)} \sum_{m\ge 0} (-1)^m q^{\frac{m(m+1)}{2} (2n+1)} \chi_{2m} (z) \ ,
\end{eqnarray}
which exactly agrees with the vacuum character of $\widehat{\mathfrak{su}}(2)_{-\frac{4n}{2n+1}}$.
For $N=3$, when $(N, k)=1$, the VOA is given by $\widehat{\mathfrak{su}}(3)_{-\frac{3(k+2)}{k+3}}$ for $k>-2$. When $k=-2$, we get $c_{4d}=0$, which means that the theory is trivial. We find the explicit form of the Schur indices to be
\begin{align}
{\cal I}_{(A_2^3 [-2], F) } &= 1 \\
{\cal I}_{(A_2^3 [-1], F) } &= 1+8 q+36 q^2+128 q^3+394 q^4+1088 q^5+2776 q^6+6656 q^7 +\ldots \ , \\
{\cal I}_{(A_2^3 [1], F) } &= 1+8 q+44 q^2+192 q^3+718 q^4+2400 q^5+7352 q^6+20992 q^7 +\ldots \ , \\
{\cal I}_{(A_2^3 [2], F) } &= 1+8 q+44 q^2+192 q^3+726 q^4+2456 q^5+7640 q^6+22176 q^7 +\ldots \ , \\
{\cal I}_{(A_2^3 [4], F)} &= 1+8 q+44 q^2+192 q^3+726 q^4+2464 q^5+7704 q^6+22520 q^7 +\ldots \ ,
\end{align}
where we set $z_i=1$ for simplicity.
\section{Higgs branch and associated variety of VOA} \label{sec:Higgs}
In this section, we study Higgs branches of the generalized Argyres-Douglas theories from the vertex operator algebra (VOA) or chiral algebra perspective. For a given VOA ${\cal V}$, it is possible to construct the associated variety $X_{\cal V}$, which was studied by Arakawa \cite{MR3456698}. It has been conjectured that the associated variety of the VOA agrees with the Higgs branch \cite{BRVOA}. We give some evidence for the case of Argyres-Douglas theories by computing the Hilbert series of the Higgs branch and compare against the one obtained from the TQFT description of the Hall-Littlewood index. We also use the 3d mirror symmetry to check some of the examples.
\subsection{Higgs branch from the associated variety of VOA}
For a VOA, one can associate a Poisson variety \cite{MR3456698}. The construction goes as follows. Recall that a vertex algebra ${\cal V}$ is a vector space
equipped with an element $\textbf{1} \in {\cal V}$ (or $\ket{0} \in {\cal V}$) called the vacuum, $T \in \textrm{End}(V)$ called the translation operator, and a linear map
\begin{equation}
Y(\cdot,z): {\cal V}\rightarrow \textrm{End}({\cal V})[[z,z^{-1}]],~~a \mapsto Y(a, z)=a(z)=\sum_{n\in Z} a_{(n)} z^{-n-1}.
\end{equation}
called the state-operator correspondence, such that
\begin{align}
\begin{split}
& \textbf{1}(z)=id_{\cal V}, \\
& a_{(n)}b=0~\textrm{for}~n \gg 0 \\
& a_{(n)}\textbf{1}=0~\textrm{for}~n\geq 0,~~\textrm{and}~~a_{(-1)}\textbf{1}=a \\
& (Ta)(z)=[T, a(z)]={d\over dz}a(z) \\
& (z-w)^n[a(z), b(w)]=0~\textrm{in}~\textrm{End}({\cal V})~\textrm{for}~n\gg0.
\end{split}
\end{align}
There is a filtration $F^{\bullet} {\cal V}$ on any vertex algebra \cite{Li2005}: Set $F^0{\cal V}={\cal V}$, and
\begin{equation}
F^p {\cal V}=\textrm{Span} \{a_{(-i-1)}b; a\in {\cal V},~b\in F^{p-i}{\cal V},~i\geq 1\} \ .
\end{equation}
Then we have a decreasing filtration
\begin{equation}
{\cal V}=F^0{\cal V}\supset F^1 {\cal V} \supset F^2 {\cal V} \supset \ldots \ .
\end{equation}
The subspace $F^1 {\cal V}$ is the linear span of $a_{(-2)} b$ with $a,b \in {\cal V}$. Now, we define
\begin{equation}
R_{\cal V}={\cal V}/F^1{\cal V}.
\end{equation}
This is called the Zhu's $C_2$-algebra of ${\cal V}$ \cite{zhu1996modular}.
In short, $R_{\cal V}$ is the set of states that are only generated by the $(-1)$-modes of any generators in a VOA.
$R_{\cal V}$ has a Poisson structure and the Poisson bracket is given by
\begin{equation}
\{a, b\}=a_{(0)} b~~\textrm{for}~a,b\in R_{\cal V}
\end{equation}
From now on, we assume $R_{\cal V}$ is finitely generated as a ring, and define the associated variety of ${\cal V}$ as
\begin{equation}
X_{\cal V}=\textrm{Spec} R_{\cal V} \ .
\end{equation}
One can also define associated variety for a (non-vacuum) module of a vertex algebra.
Now, we conjecture that for the Argyres-Douglas theories, the Higgs branch chiral ring is given by $R_{\cal V}$ and the Higgs branch is given by the associated variety $X_{\cal V}$ \cite{BRVOA}:
\begin{Conjecture}
Zhu's algebra $R_{\cal V}$ is identical to the Higgs branch chiral ring of the Argyres-Douglas theories. For this case, the Higgs branch is given by the associated variety $X_{\cal V}$.
\begin{align}
{\cal M}_{\rm{Higgs}} = X_{\cal V} \ , \qquad \mathbb{C}[{\cal M}_{\rm{Higgs}}] = R_{\cal V}.
\end{align}
\end{Conjecture}
For a general 4d ${\cal N}=2$ SCFT, the above relation has to be modified, since $R_{\cal V}$ is in general (conjectured to be) given by the Hall-Littlewood (HL) chiral ring \cite{BRVOA}. The HL chiral ring is generated by the operators contributing to the HL index, which is not in general identical to the Higgs branch operators. But they are indeed identical when the 4d ${\cal N}=2$ SCFT is given by a quiver gauge theory with no loops \cite{Gadde:2011uv}. We conjecture this is also the case for the AD theories.
\paragraph{Examples}
Let us consider some simple examples. The VOA $(A_1, A_{2n})$ theory is given by a simple Virasoro minimal model. For these theories, $L_{-1}\ket{0}$ is a null state, therefore $R_{\cal V}$ is trivial and $X_{\cal V}$ is a point.
The VOA for the $(A_1, D_{2n+1})$ theory, is given by $\widehat{\mathfrak{su}}(2)_{-\frac{4n}{2n+1}}$. In this case, we have 3 generators $J^+, J^0, J^-$. So $R_{\cal V}$ is generated by $J^+_{-1}, J^0_{-1}, J^-_{-1}$. But there is a relation $L_{-2} \ket{0} \sim (J^+_{-1} J^-_{-1} +J^0_{-1}J^0_{-1})\ket{0}$ which is coming from the Sugawara construction of the Virasoro generators from the affine Lie algebra. Since $L_{-2} \ket{0} \in F^1 {\cal V}$, it has to be modded out.\footnote{The vertex operator algebra or conformal vertex algebra contains an element called the conformal vector $\omega = L_{-2} \ket{0}$, which is mapped to the stress tensor: $Y(\omega, z) = L(z)$} This gives us
\begin{align}
R_{\cal V} = \mathbb{C}[x, y, z]/\langle xy+z^2\rangle \ .
\end{align}
We see that the associated variety
\begin{align}
X_{\cal V} = \textrm{Spec}(R_{\cal V}) = \{x, y, z \in \mathbb{C} ~|~ xy+z^2=0 \} = \mathbb{C}^2/\mathbb{Z}_2
\end{align}
is identical to the Higgs branch of the $(A_1, D_{2n+1})$ theory.
\paragraph{Non-maximal punctures}
Now let us consider the VOA $W^{k_{2d}}(J, Y)$ associated to our AD theories $(J^b[k], Y)$, where we partially close the puncture as the one labeled by $Y$. We
also assume that the level $k_F$ is admissible, i.e. $b=h$. The associated variety for this case has been already determined by Arakawa \cite{MR3456698}, and the
answer takes the following form
\begin{equation}
X_{\cal V}=X_M\cap S_Y.
\end{equation}
Here $X_M$ is the associated variety of the Kac-Moody algebra $\widehat{J}_{k_{2d}}$, and $S_Y$ is the Slodowy slice defined using the nilpotent element of $Y$.
Recall that the Slodowy slice defined by $Y$ is given by $S_Y = \oplus_j R_j$ under the decomposition $\mathrm{adj} \to \bigoplus_j R_i \otimes V_j$. The space $X_{\cal V}$ is identical to the Higgs branch obtained by the partial closure of the puncture given by $Y$, which is the commutant of the nilpotent element $\Lambda_Y(\sigma^+)$ inside the Higgs branch $X_{\cal V}$.
The associated variety $X_M$ is given by the closure of a nilpotent orbit $\mathbb{O}[k_{2d}]$, depending only on the denominator of $k_{2d} = -\frac{1}{2} k_F$.
If the level is given in the following form
\begin{equation}
k_{2d} =-h+{p\over q},~~(p,q)>1, p\geq h,
\end{equation}
the associated variety is given by
\begin{align}
X_M = \overline{\mathbb{O}}_q \ .
\end{align}
Here the nilpotent orbits $\mathbb{O}_q$ for each $J$ is given in tables \ref{table:ADorbits}, \ref{table:E6orbits}, \ref{table:E7orbits}, \ref{table:E8orbits}, reproduced from \cite{MR3456698}.
\subsection{Examples} \label{sec:HiggsEx}
In this subsection, we focus on theories with $J=A_{N-1}$, $b=N$. This gives the corresponding chiral algebra to be the Kac-Moody algebra $\widehat{\mathfrak{su}}(N)_{k_{2d}}$ with level taking the following form
\begin{equation}
k_{2d}=-N+{N\over N+k},~~(k,N)=1.
\end{equation}
If $k>0$, then the associated variety is the principal nilpotent cone of the $A_{N-1}$ Lie algebra.
This is identical to the Higgs branch of the (3d ${\cal N}=4$) quiver gauge theory
\begin{align}
(1) - (2) - (3) - \ldots - (N-1) - [N] \ .
\end{align}
Here, each nodes of $(n)$ means $U(n)$ gauge group and $[N]$ refers to $SU(N)$ flavor group.
If $k<0$, then the associated variety is the closure of the following nilpotent orbit labelled by a partition of $N$
\begin{equation}
[\underbrace{q,\ldots, q}_r, s],~~s\leq q-1 \ ,
\end{equation}
where $q=N+k$. The nilpotent orbit can be identified with the Higgs branch of a quiver gauge theory associated with the transpose of the partition above, which is $[\underbrace{r+1, \ldots, r+1}_s,\underbrace{r,\ldots, r}_{q-s}, ]$. The quiver is:
\begin{align}
(r)-(2r)-\ldots-((q-s)r)-((q-s)r+r+1)-\ldots-(N-(r+1))-[N].
\end{align}
\paragraph{$\mathbf{k>0}$}
According to Arakawa's result, the Higgs branch of these theories is identified with nilpotent cone of $A_{N-1}$ Lie algebra.
\paragraph{$\mathbf{k=-1}$}
For this case, $k_{2d} = -N + \frac{N}{N-1}$ so that $q=N-1 < h^\vee = N$. The associated variety for the VOA is given by the nilpotent orbit labelled by the partition $(N-1, 1)$. This gives the following quiver described by the transpose of the partition $(2, 1, 1, \ldots, 1)$:
\begin{eqnarray}
(1) - (2) - (3) - \ldots - (N-2) - [N]
\label{kequalminusone}
\end{eqnarray}
The Higgs branch of the above quiver gauge theory is the closure of sub-regular nilpotent orbit.
\paragraph{$\mathbf{k=-N+2}$}
For this case, $k_{2d} = -N +\frac{N}{2}$. We choose $N=2n+1$ to be an odd number. Then the nilpotent orbit is given by the partition $(2, 2, \ldots, 2, 1)$. The transpose of the partition is given by $(2N+1, N)$, which gives the quiver
\begin{eqnarray}
(N) - [2N+1] \ .
\label{mini}
\end{eqnarray}
The Higgs branch of the above quiver gauge theory is identical to the Higgs branch of the $(A_{N-1}^N[-N+2], F)$ theory.
\paragraph{Minimal nilpotent orbits}
Let us consider some theories whose corresponding VOAs have non-admissible levels.
The Higgs branch of the $(D_4^4[-3], F)$, $(E_6^9[-8],F)$, $(E_7^{14}[-13],F)$, $(E_8^{24}[-23],F)$ theories are given by the minimal nilpotent orbit $D_4, E_6, E_7, E_8$.
These are the same theories as the rank 1 SCFTs with $D_4, E_{6, 7, 8}$ global symmetries.
Affine vertex operator algebras with non-admissible levels have been studied recently \cite{Arakawa:2016ad, Arakawa:2016aa}, and the corresponding associated variety is exactly the minimal nilpotent orbit.
\subsection{Check using 3d mirror}
It is possible to write down the 3d mirror quiver $B$ for the class of theory $(J^b[k], Y)$ with
$J=A_{N-1}, b=N, (k,N)=1$ and $Y=F$ being a full puncture (see section \ref{sec:ADthy}). The Coulomb branch of
this mirror quiver B (in the IR limit) should give the Higgs branch of the original 4d theory $A$. It is often possible to find another 3d quiver $C$ whose Higgs branch would
be the same as the IR Coulomb branch of theory $B$. In turn, the Higgs branch of the quiver $C$ coincides with the Higgs branch of original 4d theory $A$. Therefore, the quiver $C$ can be very useful
to check Arakawa's result. We will discuss several examples here, and leave the full check of Arakawa's result to an interested reader.
Consider a theory given by $\mathbf{J=A_{2N},~b=2N+1,~k=-2N+1}$, and a full regular puncture $Y=F$.
The 3d mirror of this theory is described by the following quiver \cite{Xie:2016evu} (here we ungauge the $U(1)$ gauge group, see figure \ref{3dmirror}):
\begin{align}
\begin{split}
&(1)-(2)-\ldots-(N)-(N)-\ldots-(2)-(1)\\
&~~ ~~~~~~~~~~~~~~~~~~~~~~| ~~~~~~~~ | \\
&~~ ~~~~~~~~~~~~~~~~~~~~~[1] ~~~~~~[1]
\end{split}
\end{align}
The mirror of the above quiver can be found using the Hanany-Witten construction, and it is exactly as the one shown in \eqref{mini}.
The Higgs branch of this quiver (which is identical to the Higgs branch of the original 4d theory) is exactly given by the closure of nilpotent orbit labelled as $[N+1, N]$.
Let us now consider the theory with $\bf{J=A_{N-1},~b=N,~k=1}$ and a full regular puncture ($Y=F$). Its 3d mirror is given as below (see also figure \ref{3dmirror}):
\begin{align}
[N]-(N-1)-\ldots-(2)-(1)
\end{align}
This quiver is self-mirror, and the Coulomb branch is given by the nilpotent cone of the $\mathrm{su}(N+1)$ Lie algebra. This agrees with the associated variety of the corresponding VOA.
Finally, Let's now consider the theory with $\bf{(J=A_{N-1},~b=N,~k=-1)}$ and a full regular puncture, its 3d mirror is (see figure. \ref{3dmirror} and we ungauge the $U(1)$ gauge group.)
\begin{align}
\begin{split}
&[N-2]-(N-2)-(N-2)-(N-3)-\ldots-(2)-(1)\\
&~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~| \\
&~~ ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~[1]
\end{split}
\end{align}
Using the Hanany-Witten construction, one can find its mirror as shown in \eqref{kequalminusone}.
\subsection{Check using TQFT}
The Hall-Littlewood limit of the superconformal index is defined as
\begin{align}
{\cal I}_{HL}(t) = {\rm Tr}_{{\cal H}_{HL}} (-1)^F t^{E-R} \ ,
\end{align}
where the trace is over the states satisfying $E-2R-r=0$ and $j_1=0$. The Hall-Littlewood (HL) index is conjectured to be the same as the Hilbert series on the Higgs branch for a large class of theories.\footnote{This fails for the quiver gauge theory with loops.} We will use the TQFT description to compute the HL indices for the AD theories and verify against the direct computation for the nilpotent orbits for the possible cases. From the TQFT we also derive concise expressions for the Hilbert series of the Higgs branch for certain AD theories.
\subsubsection{Wave function for the irregular puncture}
We follow similar strategy as in the case of Schur index. First, we come up with the wave function for the Gaiotto state (minimally irregular singularity) realizing the pure Yang-Mills (YM) theory. Second, we try to guess the wave functions for other irregular singularities.
The Hall-Littlewood index for the pure YM theory with gauge group $J$ is given by
\begin{eqnarray}
{\cal I}_{HL}^J = \oint [d\vec{z}] I_{\textrm{vec}} (\vec{z}) = \oint [d\vec{z}] \mathrm{PE} \left[ -t \chi_{\textrm{adj}} (\vec{z}) \right] = (1-t)^r \oint [d\vec{z}] \prod_{\vec{\alpha} \in \Delta_J} (1-t \vec{z}^{\vec\alpha}) \ .
\end{eqnarray}
We find a closed form expression for the $G=SU(N)$ case as
\begin{eqnarray}
{\cal I}_{HL}^{SU(N)} = \prod_{n=1}^{N-1} (1 - t^{2n+1}) \ .
\end{eqnarray}
From the TQFT structure of the index, we should be able to obtain the index from the sphere with two irregular punctures of the same type $I_J \equiv I_J[-h+1]$. Therefore, we write the index as
\begin{eqnarray}
{\cal I}_{HL}^{J} = \sum_{\vec\lambda} \psi_{\vec\lambda}^{I_{J}} \psi_{\vec\lambda}^{I_{J}} \ ,
\end{eqnarray}
where the wave function for the $I_{J}$ puncture is given by
\begin{eqnarray}
\psi_{\vec\lambda}^{I_{J}} =\oint [d\vec{z}] \mathrm{PE} \left[ -t \chi_{\textrm{adj}} (\vec{z}) \right] \psi_{\vec\lambda}(\vec{z}) = \oint [d\vec{z}] P^{HL}_{\vec\lambda} (\vec{z}) \ .
\end{eqnarray}
And the wave function for the full regular puncture is given as
\begin{align}
\psi_{\vec\lambda}(\vec{z}) = \mathrm{PE} \left[ t \chi_{\textrm{adj}} (\vec{z}) \right] P^{HL}_{\vec\lambda}(\vec{z}) \ ,
\end{align}
where $P^{HL}_{\vec\lambda}(\vec z)$ is the (normalized) Hall-Littlewood polynomial. We have chosen our normalization of the Hall-Littlewood polynomial so that
\begin{align}
\oint [d\vec{z}] I_{\textrm{vec}}(\vec{z}) \psi_{\vec{\lambda}}(\vec{z}) \psi_{\vec{\mu}}(\vec{z})
= \oint [d\vec{z}] \mathrm{PE} \left[ t \chi_{\textrm{adj}} (\vec{z}) \right] P^{HL}_{\vec\lambda}(\vec{z}) P^{HL}_{\vec\mu}(\vec{z})
= \delta_{\vec\lambda \vec\mu} \ .
\end{align}
Using the orthonormality of the Hall-Littlewood polynomial under the vector multiplet measure, we get
\begin{align}
\sum_{\vec \lambda} \psi_{\vec\lambda}^{I_J} \psi_{\vec\lambda}^{I_J} = \oint [d\vec{z}][d\vec{z'}] P^{HL}_{\vec\lambda}(\vec{z}) P^{HL}_{\vec\lambda}(\vec{z'})
= \oint [d\vec{z}]\mathrm{PE} \left[ -t \chi_{\textrm{adj}} (\vec{z}) \right] = {\cal I}^J_{HL} \ ,
\end{align}
so that our wave function for the irregular puncture $I_J$ is indeed consistent.
Let us write some explicit examples. For $J=A_1$, we get ($I_{N, k} \equiv A_{N-1}^N[k]$)
\begin{align}
\psi_{\lambda}^{I_{2, -1}} =
\begin{cases}
\sqrt{1-t^2} & (\lambda = 0) \ , \\
-t \sqrt{1-t} & (\lambda = 2) \ .
\end{cases}
\end{align}
When $G=A_2$, we obtain
\begin{align}
\begin{split}
\psi_{(\lambda_1, \lambda_2)}^{I_{3, -2}} =
\begin{cases}
\sqrt{(1-t^2)(1-t^3)} & (\lambda_1, \lambda_2) = (0, 0) , \\
-t (1-t^2) & (\lambda_1, \lambda_2) = (1, 1), \\
-t^3 (1-t) & (\lambda_1, \lambda_2) = (2, 2), \\
t^2 \sqrt{(1-t)(1-t^2)} & (\lambda_1, \lambda_2) = (3, 0), (0, 3), \\
0 & \textrm{otherwise} \ .
\end{cases}
\end{split}
\end{align}
We conjecture the wave function for $k > 0$ to be
\begin{align} \label{eq:HLpsi0}
\psi_{\vec\lambda}^{J^b[k]} = \prod_{i=1}^{r} (1-t^{d_i})^{\frac{1}{2}} \delta_{\vec{\lambda}, \vec{0}} \ ,
\end{align}
where $r=\textrm{rank}(J)$ and $d_i$ are the degrees of the Casimirs of $J$.
When $k=-1$, we find the wave function to be
\begin{align} \label{eq:HLpsim1}
\psi_{\vec{\lambda}}^{J^b[k]} =
\begin{cases}
\prod_{i=1}^r (1-t^{d_i})^{\frac{1}{2}} & \vec{\lambda} = \vec{0} \ , \\
-t^{h - 1}(1-t) \prod_{i=1}^{r-2} (1-t^{d_i})^\frac{1}{2} & \vec{\lambda} = \textrm{adjoint} \ .
\end{cases}
\end{align}
We have checked that the above expressions to be consistent with our previous results and known results.
It is not obvious to us how to obtain the wave function (and thereby the index) for the general value of $k$.
\subsubsection{AD theory of type $(J^h[k], Y)$}
From the TQFT structure, the Hall-Littlewood index for the $(J^b [k], F)$ theory can be written as
\begin{eqnarray}
{\cal I}_{(J^b [k], F)} = \sum_{\vec \lambda} \psi_{\vec \lambda}^{J^b [k]} \psi_{\vec \lambda} (\vec{z}) \ .
\end{eqnarray}
Plugging in the expression \eqref{eq:HLpsi0} to above, it is straight-forward to obtain (for $k>0$)
\begin{eqnarray} \label{eq:HLidxA}
{\cal I}_{(J^b [k], F)} = \frac{\prod_{i=1}^r (1-t^{d_i}) }{(1-t)^r \prod_{\vec{\alpha} \in \Delta_{J}} (1 - t \vec{z}^{\vec{\alpha}})} \ ,
\end{eqnarray}
where we used the fact that $P^{HL}_{\vec{\lambda} = \vec{0}} (\vec{z}) = \prod_{i=1}^r (1-t^{d_i})^\frac{1}{2} $.
Here $r = \textrm{rank}(J)$ and $d_i$ are the degrees of the Casimirs of $J$. Note that there is no dependence on $k$, which reflects the fact that the Higgs branch for each $k>0$ is the same. This agrees with the result of \cite{MR3456698} that the associated variety of the VOA for $k>0$ is always given by a principal nilpotent cone. We claim that this expression is the closed-form expression for the Higgs branch (and equivalently the Coulomb branch since it is self-mirror) of the 3d ${\cal N}=4$ theory $T[J]$ of \cite{Gaiotto:2008ak} with $J \in ADE$. This result also agrees with that of \cite{Cremonesi:2014kwa}.
\paragraph{Example: $(A_{N-1}^N[k], F) = (I_{N, k}, F)$ theory}
From the TQFT structure, the Hall-Littlewood index for the $(A_{N-1}^N[k], F) = (I_{N, k}, F)$ theory for $k>0$ can be written as
\begin{eqnarray} \label{eq:HLidxA}
{\cal I}_{(A_{N-1}^N[k], F)} = \mathrm{PE} \left[ t \chi_{\textrm{adj}} (\vec{z}) \right] \prod_{n=2}^N (1-t^n) = \frac{\prod_{n=2}^N (1-t^n)}{(1-t)^{N-1} \prod_{\vec\alpha \in \Delta_{SU(N)}} (1-t \vec{z}^{\vec\alpha})} \ .
\end{eqnarray}
This index is indeed the same as the Hilbert series of the Higgs branch of $T[SU(N)]$ theory. This is also the same as the Hilbert series of the Coulomb branch because the $T[SU(N)]$ theory is self-mirror.
\paragraph{Example: Non-principal orbits}
Let us consider when the associated variety is given in terms of a non-principal orbit. For $k=-1$, one can obtain the Hall-Littlewood index using \eqref{eq:HLpsim1}. We find that when $J$ is of $A$ or $D$ type, the index agrees with the Hilbert series of the sub-regular nilpotent orbit \cite{Hanany:2016gbz}.
For example, consider the $(A_2^3[-1], F) = (A_2, A_2) = (A_1, D_4)$ theory. The corresponding VOA is given by $\widehat{\mathfrak{su}}(3)_{-\frac{3}{2}}$. The wave function for $I_{3, -1}$ is given by
\begin{eqnarray}
\psi_{(\lambda_1, \lambda_2)}^{I_{3, -1}} =
\begin{cases}
\sqrt{(1-t^2)(1-t^3)} & (\lambda_1, \lambda_2) = (0, 0) \ , \\
- t^2 (1-t) & (\lambda_1, \lambda_2) = (1, 1) \ .
\end{cases}
\end{eqnarray}
It reproduces the correct Hall-Littlewood index of the $(A_2, A_2)=(A_1, D_4)$ theory. We know that the Higgs branch of this theory is given by the minimal nilpotent orbit of $SU(3)$, which is identical to the 1-instanton moduli space of $SU(3)$.
\paragraph{Non-maximal punctures}
As in the Schur case, it is straight-forward to obtain the theory for the general regular puncture $Y$ via nilpotent Higgsing. The adjoint representation of $J$ is now decomposed in terms of the commutant of the $SU(2)$ embedding $\Lambda_Y$ as $\textrm{adj} \to \bigoplus_j R_j \otimes V_j$. Now, the wave function for the regular puncture of type $Y$ is given by a concise closed form as
\begin{align}
\psi^Y_{\vec \lambda} (\vec{a}) = \mathrm{PE} \left[ \sum_j t^{j+1} {\rm tr}_{R_j} (\vec{a}) \right] P^{HL}_{\vec\lambda} (\vec{a}t^{\Lambda_Y}) \ ,
\end{align}
where various symbols are the same as the ones defined around \eqref{eq:slodowy}. Therefore, the index for the $(J^b[k], Y)$ theory with $k>0$ is given by
\begin{align}
{\cal I}_{(J^b[k], Y)} = \mathrm{PE} \left[ \sum_j t^{j+1} {\rm tr}_{R_j} (\vec{a}) - \sum_{i=1}^r t^{d_i}\right] = \frac{\prod_{i=1}^r (1-t^{d_i})}{\prod_j \prod_{\vec{w} \in R_j} (1-t^{j+1} \vec{a}^{\vec{w}})} \ .
\end{align}
Here $\vec{w} \in R_j$ denotes the weight vectors of the representation $R_j$. This formula agrees with the one obtained in \cite{Cremonesi:2014kwa}.
\section{Conclusion} \label{sec:Conclusion}
We systematically studied the chiral algebra/vertex operator algebra corresponding to a large class of Argyres-Douglas theories labeled as $(J^b[k], Y)$. Here $J^b[k]$ denotes an irregular singularity carrying no mass parameters, and $Y$ represents a regular singularity. We restricted to the case where there is no exactly marginal deformations. In particular, all the theories of the form $(A_1, \Gamma)$ with $\Gamma \in ADE$ belongs to this class, so that we recover the result of \cite{Cordova:2015nma,Creutzig:2017qyf}.
We found a surprisingly simple formula for the Schur index (which is equal to the vacuum character of the corresponding chiral algebra) when $b=h$ utilizing the TQFT description of the index. The Higgs branch of a theory in this class is identified with the associated variety of the corresponding VOA, and we also found a concise formula for its Hilbert series or the Hall-Littlewood index.
Various nilpotent orbits appear as the Higgs branch of an AD theory. However, not all of them appear. For the $su(N)$ Lie algebra, only the nilpotent orbits of the form $[q,\ldots, q, s]$ appear. It would be interesting to find AD theories whose Higgs branch are given by other nilpotent orbits, or find some obstructions possibly along the line of \cite{Shimizu:2017kzs}.
AD theories in our list can be used to construct various new SCFTs by gauging the flavor symmetry whenever available (they are called the AD matter in \cite{Xie:2017vaf}). The Schur indices of these general SCFTs can be found using the Schur index of the AD matter. These indices can be used to check the S-duality of the Argyres-Douglas theory proposed in \cite{Buican:2014hfa, Xie:2017vaf}. However, not all of the AD theories can be constructed from the gluing of the AD theories considered in this paper, and it is interesting to find the indices of other AD matters used in \cite{Xie:2017vaf}.
We derived many results for the theory $(J^b[k], Y)$ when $b=h$, where the level of the corresponding chiral algebra is admissible. However, our results also suggests
that similar results should also hold for $b<h$ where the level for the corresponding VOA is not admissible. It would be interesting to study them further.
Also, it would be interesting to find more refined partition functions, such as the Macdonald index, Lens space index and the full superconformal index for the AD theories in our class.
\acknowledgments
We thank Chris Beem and Leonardo Rastelli for discussions and correspondence. JS and WY would like to thank the organizers of the workshop ``Exact Operator Algebras in Superconformal Field Theories" and Perimeter Institute for hospitality. JS also thanks Kavli IPMU for hospitality where this paper is finalized.
The work of JS is supported in part by the US Department of Energy under UCSD's contract de-sc0009919 and also by Hwa-Ahm foundation.
The work of DX and WY is supported by Center for Mathematical Sciences and Applications at Harvard University.
| a03a2e669d401508b9f3d444362c07941a0cb6ff | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Density functional theory (DFT) \cite{Hohenberg_1964_864, Parr_1994_} is a powerful computational tool in chemistry and physics. Typically DFT calculations are conducted in the context of the Born-Oppenheimer approximation, where the nuclei are treated as fixed point charges when solving the electronic Schr{\"o}dinger equation. However, the quantum mechanical nature of the nuclei and the non-Born-Oppenheimer effects between electrons and nuclei are important for many challenging physical and chemical problems, such as proton-coupled electron transfer \cite{Hammes-Schiffer_2015_8860}. Multicomponent DFT, which enables the quantum mechanical treatment of more than one type of particle, such as electrons and protons, was developed to tackle these problems \cite{Capitani_1982_568, Shigeta_1998_659, Kreibich_2001_2984, Chakraborty_2008_153001}. Analogs to the Hohenberg-Kohn theorems \cite{Hohenberg_1964_864} have been proven for multicomponent systems \cite{Capitani_1982_568, Kreibich_2001_2984, Kreibich_2008_22501}, and the Kohn-Sham formalism \cite{Kohn_1965_1133}, in which the kinetic energy functional is approximated with electronic and nuclear orbitals, has been extended to multicomponent systems \cite{Shigeta_1998_659, Udagawa_2006_244105, Pak_2007_4522, Kreibich_2001_2984, Kreibich_2008_22501, Chakraborty_2008_153001, Chakraborty_2009_124115}. This paper centers on the nuclear-electronic orbital (NEO) \cite{Webb_2002_4106} implementation of multicomponent DFT, denoted NEO-DFT \cite{Pak_2007_4522, Chakraborty_2008_153001}, applied to systems in which specific hydrogen nuclei are treated quantum mechanically, and at least two other nuclei are treated as classical point charges to avoid difficulties with translations and rotations.
A major challenge in the Kohn-Sham implementation of multicomponent DFT is the design of accurate and practical electron-proton correlation functionals. When electron-proton correlation is neglected, the proton densities are highly over-localized, leading to unphysical results for most properties of interest. Efforts to approximate the electron-proton correlation functional by analyzing the connectivity between the electron-proton pair density and the total wave function or certain orbitals \cite{Kreibich_2001_2984, Chakraborty_2008_153001, Sirjoosingh_2011_2689, Sirjoosingh_2012_174114, Kreibich_2008_22501} have improved proton densities but are computationally expensive or not easily transferable to general systems. Recent attempts \cite{Imamura_2008_735, Udagawa_2014_52519} to develop simple electron-proton correlation functionals related to the Colle-Salvetti formulation \cite{Colle_1975_329}, which forms the basis of the Lee-Yang-Parr (LYP) electron-electron correlation functional \cite{Lee_1988_785}, relied on problematic approximations. For example, Ref. \cite{Imamura_2008_735} has the incorrect sign of electron-proton correlation in the paper and incorrectly assumed a norm condition that leads to a singularity problem. Moreover, these previous attempts only applied those functionals as energetic corrections after the self-consistent-field (SCF) procedure. As a result, these approaches \cite{Imamura_2008_735, Udagawa_2014_52519} do not influence the highly over-localized proton densities and thus are not useful for describing any properties that rely on the proton density. In this paper, we develop an electron-proton correlation functional based on the Colle-Salvetti formulation using well-justified approximations and apply this new functional within the SCF procedure, producing accurate proton densities at a low computational cost.
\section{Theory}
Consider a multicomponent system with $N_e$ electrons and $N_p$ protons, where the subscripts $e$ and $p$ denote electrons and protons for all variables hereafter. Within the framework of multicomponent DFT, the total energy can be expressed as
\begin{equation}
\label{eq:DFTenergy}
\begin{aligned}
E[\rho_e,\rho_p]=&(T_s[\rho_e]+T_s[\rho_p])+(V_e^{ext}[\rho_e]+V_p^{ext}[\rho_p])\\&+(J_{ee}[\rho_e]+J_{pp}[\rho_p]+J_{ep}[\rho_e,\rho_p])\\&+(E_{e}^{xc}[\rho_e]+E_{p}^{xc}[\rho_p])+(E^{epc}[\rho_e,\rho_p]),
\end{aligned}
\end{equation}
where the quantities in each parenthesis represent the non-interacting reference kinetic energy, external potential energy, mean-field Coulomb interactions, same-particle exchange-correlation (xc) energy, and electron-proton correlation (epc) energy, respectively. The electron and proton densities are defined in terms of the electron and proton orbitals, $\phi_{e}^{i}(\mathbf{r}_e)$ and $\phi_{p}^{I}(\mathbf{r}_p)$:
$\rho_e(\mathbf{r}_e)=\sum_i^{N_e}|\phi_e^i(\mathbf{r}_e)|^2$,
$\rho_p(\mathbf{r}_p)=\sum_{I}^{N_p}|\phi_p^{I}(\mathbf{r}_p)|^2$.
The Kohn-Sham equations for electrons and protons obtained by minimizing the total energy are
\begin{subequations}
\begin{align}
\bigg(-\frac{1}{2}\nabla^2_e+v_e(\mathbf{r}_e)\bigg)\phi_e^i(\mathbf{r}_e)=&\varepsilon_e^i\phi_e^i(\mathbf{r}_e),\\
\bigg(-\frac{1}{2m_p}\nabla^2_p+v_p(\mathbf{r}_p)\bigg)\phi_p^{I}(\mathbf{r}_p)=&\varepsilon_p^{I}\phi_p^{I}(\mathbf{r}_p),
\end{align}
\end{subequations}
in atomic units with the effective potentials for the electrons and protons defined as
\begin{subequations}
\begin{align}
v_e(\mathbf{r}_e)=&v_e^{ext}(\mathbf{r}_e)+v_e^{J_{ee}}(\mathbf{r}_e)+v_e^{J_{ep}}(\mathbf{r}_e)\\&\notag+v_e^{xc}(\mathbf{r}_e)+v_e^{epc}(\mathbf{r}_e),\\
v_p(\mathbf{r}_p)=&v_p^{ext}(\mathbf{r}_p)+v_p^{J_{pp}}(\mathbf{r}_p)+v_p^{J_{ep}}(\mathbf{r}_p)\\&\notag+v_p^{xc}(\mathbf{r}_p)+v_p^{epc}(\mathbf{r}_p).
\end{align}
\end{subequations}
Each potential term is obtained by taking the functional derivative of the corresponding energy term in Eq. \eqref{eq:DFTenergy} with respect to either the electron or the proton density. The electron-electron exchange-correlation functional is defined the same as in standard electronic DFT, and previously developed functionals can be used. Because quantum protons are spatially localized in molecular systems, the proton-proton exchange and correlation energies are negligible, and can be approximated with the diagonal Hartree-Fock exchange terms to eliminate self-interaction error. For systems with multiple protons, the proton orbitals can be assumed to be only singly occupied with high spin.
As discussed above, the epc functional $E^{epc}[\rho_e,\rho_p]$ is the most critical term in the multicomponent Kohn-Sham formalism. The approach that completely neglects this term produces highly over-localized proton densities and therefore is not a useful tool for accurately describing the quantum behavior of protons. Herein we develop a multicomponent epc functional, denoted epc17, based on an analog of the Colle-Salvetti formulation but utilizing different approximations from previous attempts \cite{Imamura_2008_735, Udagawa_2014_52519} in an effort to produce accurate proton densities. The epc17 functional is a multicomponent counterpart to the well-known LYP functional \cite{Lee_1988_785} in conventional electronic DFT, excluding the density gradient terms in its current form, and it is implemented within the SCF procedure.
The derivation of our density functional begins with an analog of the Colle-Salvetti ansatz \cite{Colle_1975_329}, which approximates the total wave function for a multicomponent electron-proton system as:
\begin{equation}
\label{eq:CSansatz}
\begin{aligned}
&\Psi(\mathbf{x}_{1,e},\mathbf{x}_{2,e},\dots\mathbf{x}_{N_e,e},\mathbf{r}_{1,p},\mathbf{r}_{2,p},\dots\mathbf{r}_{N_p,p})\\=&\Psi^{FCI}_{e}(\mathbf{x}_{1,e},\mathbf{x}_{2,e},\dots\mathbf{x}_{N_e,e})\Psi^{FCI}_{p}(\mathbf{r}_{1,p},\mathbf{r}_{2,p},\dots\mathbf{r}_{N_p,p})\\
&\cdot\prod_{i\in N_e,\,I\in N_p}[1-\varphi(\mathbf{r}_{i,e},\mathbf{r}_{I,p})],
\end{aligned}
\end{equation}
where $\mathbf{x}_{i,e}$ denotes the combined electronic variable for the spatial coordinate $\mathbf{r}_{i,e}$ and spin $\sigma_{i,e}$, and only the spatial coordinate $\mathbf{r}_{I,p}$ is used for protons as the nuclei are assumed to be high spin. FCI denotes full configuration interaction for a particular type of particle, which takes into account all of the exchange and correlation effects between particles of this type but only includes a mean-field Coloumb potential from the other type of particle. Therefore, the product of the electron and proton FCI wave functions (i.e., Eq. \eqref{eq:CSansatz} without the last correlation factor) is the mean-field FCI electron-proton wave function \cite{Cassam-Chenaie_2017_52}. The electron-proton correlation effects are included in the last Jastrow factor, with the function $\varphi(\mathbf{r}_{i,e},\mathbf{r}_{I,p})$ defined as
\begin{equation}
\label{eq:CSfactor}
\begin{aligned}
\varphi(\mathbf{r}_{i,e},\mathbf{r}_{I,p})=e^{-\beta^{2}(\mathbf{R})r^{2}}[1-\xi(\mathbf{R})(1-r)],
\end{aligned}
\end{equation}
where $r=|\mathbf{r}|=|\mathbf{r}_{i,e}-\mathbf{r}_{I,p}|$, and $\mathbf{R}$ is the center of mass for the two particles, which is approximately the same as $\mathbf{r}_{I,p}$ because of the significantly larger mass of the proton compared to the electron. In this equation, $\beta(\mathbf{R})$ and $\xi(\mathbf{R})$ are two functions that will be determined below. This ansatz is chosen because it fulfills the electron-proton cusp condition as $\mathbf{r}_{i,e}\rightarrow\mathbf{r}_{I,p}$ \cite{Bingel_1967_54} and reduces to the exact mean-field FCI electron-proton wave function when the electrons and protons are far apart.
Analogous to the original Colle-Salvetti formulation for electron-electron correlation, the electron-proton pair density matrix, which was the second-order reduced density matrix in the electronic case, can be approximated by the product of the mean-field pair density matrix and a correlation factor,
\begin{equation}
\label{eq:PairDensity}
\begin{aligned}
P_{2,ep}(\mathbf{x}_{1,e}^{\prime},\mathbf{r}_{1,p}^{\prime};&\mathbf{x}_{1,e},\mathbf{r}_{1,p})\approx P_{2,ep}^{FCI}(\mathbf{x}_{1,e}^{\prime},\mathbf{r}_{1,p}^{\prime};\mathbf{x}_{1,e},\mathbf{r}_{1,p})\\&\cdot[1-\varphi(\mathbf{r}_{1,e}^{\prime},\mathbf{r}_{1,p}^{\prime})][1-\varphi(\mathbf{r}_{1,e},\mathbf{r}_{1,p})],
\end{aligned}
\end{equation}
where the mean-field electron-proton pair density matrix is the product of the associated electron and proton first-order reduced density matrices: $P_{2,ep}^{FCI}(\mathbf{x}_{1,e}^{\prime},\mathbf{r}_{1,p}^{\prime};\mathbf{x}_{1,e},\mathbf{r}_{1,p})=P_{1,e}^{FCI}(\mathbf{x}_{1,e}^{\prime};\mathbf{x}_{1,e})P_{1,p}^{FCI}(\mathbf{r}_{1,p}^{\prime};\mathbf{r}_{1,p})$.
The electron-proton correlation energy is formally defined as the difference between the exact total energy associated with the wave function in Eq. \eqref{eq:CSansatz} and the energies associated with the electron and proton mean-field FCI wave functions after removing the double counting of the mean-field electron-proton Coulomb interaction energy:
\begin{equation}
\label{eq:EcorrelationDef}
\begin{aligned}
E^{epc}=&E-(E^{FCI}_{e}+E^{FCI}_{p}-J_{ep})\\
=&\mathrm{Tr[}(-\frac{1}{2}\nabla_{e}^{2}+v_{e}^{ext})(P_{1,e}-P_{1,e}^{FCI})]+\mathrm{Tr[}V_{ee}(P_{2,e}\\&-P_{2,e}^{FCI})]+\mathrm{Tr[}(-\frac{1}{2m_{p}}\nabla_{p}^{2}+v_{p}^{ext})(P_{1,p}-P_{1,p}^{FCI})]\\&+\mathrm{Tr[}V_{pp}(P_{2,p}-P_{2,p}^{FCI})]-\int\mathrm{d}\mathbf{x}_{1,e}\mathrm{d}\mathbf{r}_{1,p}\frac{1}{r}\\&\cdot[P_{2,ep}(\mathbf{x}_{1,e},\mathbf{r}_{1,p})-P_{2,ep}^{FCI}(\mathbf{x}_{1,e},\mathbf{r}_{1,p})].
\end{aligned}
\end{equation}
Here $V_{ee}$ and $V_{pp}$ are two-particle Coulomb repulsion operators, $P_{1,e}$ and $P_{2,e}$ are the electron first- and second-order reduced density matrices associated with the total wave function in Eq. \eqref{eq:CSansatz}, and $P_{1,p}$ and $P_{2,p}$ are the proton first- and second-order reduced density matrices defined analogously. Moreover, the pair density $P_{2,ep}(\mathbf{x}_{1,e},\mathbf{r}_{1,p})$ denotes the diagonal elements of the electron-proton pair density matrix.
Assuming that the electron and proton first- and second-order reduced density matrices are reasonably well approximated by those from mean-field FCI, the electron-proton correlation energy reduces to the last term in Eq. \eqref{eq:EcorrelationDef}:
\begin{equation}
\label{eq:EcorrelationSimp}
\begin{aligned}
E^{epc}
=&-\int\mathrm{d}\mathbf{x}_{1,e}\mathrm{d}\mathbf{r}_{1,p}\frac{1}{r}P_{2,ep}^{FCI}(\mathbf{x}_{1,e},\mathbf{r}_{1,p})\\&\cdot[\varphi^2(\mathbf{r}_{1,e},\mathbf{r}_{1,p})-2\varphi(\mathbf{r}_{1,e},\mathbf{r}_{1,p})].
\end{aligned}
\end{equation}
Equating the electron and proton densities associated with the full wave function in Eq. \eqref{eq:CSansatz} to those associated with the mean-field FCI electron-proton wave function, which is a direct consequence of the above approximation for the second-order reduced density matrices, leads to an approximate relationship between the undetermined functions $\xi(\mathbf{R})$ and $\beta(\mathbf{R})$:
\begin{widetext}
\begin{equation}
\label{eq:ApproxPhi}
\begin{aligned}
\xi(\mathbf{R}) \approx \frac{-9.2 \beta^2(\mathbf{R})+14.1 \beta(\mathbf{R})\sqrt{1.0 \beta^2(\mathbf{R})-2.0 \beta(\mathbf{R})+1.2} +12.0 \beta(\mathbf{R})}{5.0 \beta^2(\mathbf{R})-8.0 \beta(\mathbf{R})+3.8}.
\end{aligned}
\end{equation}
\end{widetext}
In the original electronic Colle-Salvetti formulation, $\beta(\mathbf{R})$ is assumed to be inversely proportional to the Wigner-Seitz radius $r_{s,e}\propto\rho_e^{-1/3}(\mathbf{R})$ from a uniform electron gas model. In the multicomponent electron-proton case, however, $\beta(\mathbf{R})$ must depend on both the electron and proton densities. Thus, we define a geometric mean Wigner-Seitz radius $r_{s,ep}=(r_{s,e}r_{s,p})^{1/2}$, where $r_{s,p}$ is defined analogously to $r_{s,e}$, and assume $\beta(\mathbf{R})$ to be inversely proportional to the local mean $r_{s,ep}(\mathbf{R})$
\begin{equation}
\label{eq:CorrelationLength}
\begin{aligned}
\beta(\mathbf{R})\propto\frac{1}{r_{s,ep}(\mathbf{R})}\propto\rho_e^{1/6}(\mathbf{R})\rho_p^{1/6}(\mathbf{R}).
\end{aligned}
\end{equation}
Note that Refs. \cite{Imamura_2008_735,Udagawa_2014_52519} assumed that $\beta(\mathbf{R})\propto\rho_e^{1/3}(\mathbf{R})$ in deriving their expressions for electron-proton correlation energy; however, such neglect of the dependence on the proton density does not seem to be physically reasonable.
Although $\xi(\mathbf{R})$ and $\beta(\mathbf{R})$ have been determined, the epc functional in Eq. \eqref{eq:EcorrelationSimp} is still not practical because it involves an expensive six-dimensional integration over both electron and proton coordinates. We continue following the strategy of the Colle-Salvetti formalism and transform the coordinates \{$\mathbf{x}_{1,e},\mathbf{r}_{1,p}$\} to \{$\mathbf{R}$,$\mathbf{r}$\}. After making use of the mathematical relation
\begin{equation}
\label{eq:EqualityRelation}
\begin{aligned}
\int\mathrm{d}\mathbf{r}&\frac{e^{-\beta^2r^2}}{r}F(\mathbf{r})=4\pi F(0)\int re^{-\beta^2r^2}\mathrm{d}r\\&+\frac{2\pi}{3}(\nabla_{r}^{2}F(\mathbf{r}))|_{r=0}\int r^{3}e^{-\beta^2r^2}\mathrm{d}r+O(r^{6}),
\end{aligned}
\end{equation}
and Eq. \eqref{eq:CSfactor}, the electron-proton correlation energy becomes
\begin{widetext}
\begin{equation}
\label{eq:Ecorr}
\begin{aligned}
E^{epc}
\approx& -4\pi\int\mathrm{d}\mathbf{R}P_{2,ep}^{FCI}(\mathbf{R},0)\int\mathrm{d}re^{-2\beta^{2}(\mathbf{R})r^{2}}[1-\xi(\mathbf{R})(1-r)]^{2}r\\
&+8\pi\int\mathrm{d}\mathbf{R}P_{2,ep}^{FCI}(\mathbf{R},0)\int\mathrm{d}re^{-\beta^{2}(\mathbf{R})r^{2}}[1-\xi(\mathbf{R})(1-r)]r\\
&-2\pi/3\int\mathrm{d}\mathbf{R}(\nabla_{r}^{2}P_{2,ep}^{FCI}(\mathbf{R},\mathbf{r}))|_{r=0}\int\mathrm{d}re^{-2\beta^{2}(\mathbf{R})r^{2}}[1-\xi(\mathbf{R})(1-r)]^{2}r^{3}\\
&+4\pi/3\int\mathrm{d}\mathbf{R}(\nabla_{r}^{2}P_{2,ep}^{FCI}(\mathbf{R},\mathbf{r}))|_{r=0}\int\mathrm{d}re^{-\beta^{2}(\mathbf{R})r^{2}}[1-\xi(\mathbf{R})(1-r)]r^{3}.
\end{aligned}
\end{equation}
\end{widetext}
In this paper, we neglect the last two terms involving the second-order derivatives of the pair density and leave the inclusion of these terms for future development. Therefore, we are able to obtain a simple form for the epc functional. Substituting Eq. \eqref{eq:ApproxPhi} into the first two terms of Eq. \eqref{eq:Ecorr} and integrating leads to the final form of our epc17 functional, which we approximate as:
\begin{equation}
\label{eq:FunctionalForm}
\begin{aligned}
E^{epc}=-\int\mathrm{d}\mathbf{R}\frac{\rho_{e}(\mathbf{R})\rho_{p}(\mathbf{R})}{a-b\rho_{e}^{1/2}(\mathbf{R})\rho_{p}^{1/2}(\mathbf{R})+c\rho_{e}(\mathbf{R})\rho_{p}(\mathbf{R})},
\end{aligned}
\end{equation}
with three parameters to be determined below. Although the derivation of the epc17 functional is based on the wave function in Eq. \eqref{eq:CSansatz} and its associated density matrices, these are never explicitly constructed. This simple type of functional has only two fundamental variables, namely electron density $\rho_e$ and proton density $\rho_p$. Therefore, it is a local density approximation (LDA) form for both electrons and protons. Note that $\rho_e$ is the total electron density for both closed-shell and open-shell cases, and $\rho_p$ is the total proton density, where the protons are assumed to be high spin.
\section{results}
We have implemented the epc17 functional in GAMESS \cite{Schmidt_1993_1347} and applied this approach to the molecules \ce{FHF-} and \ce{HCN}, which have been extensively studied for testing purposes \cite{Sirjoosingh_2015_214107, Brorsen_2015_214108, Culpitt_2016_44106}. The heavy nuclei are treated as point charges fixed to the positions obtained by optimizing the geometry at the conventional electronic DFT level, and the hydrogen nucleus, as well as all electrons, are treated quantum mechanically. The def2-QZVP \cite{Weigend_2005_3297} electronic basis set and an even-tempered 8{\it s}8{\it p}8{\it d} proton basis set centered at the optimized hydrogen position with exponents running from $2\sqrt{2}$ to $32$ are used for all calculations. The B3LYP functional \cite{Becke_1988_3098, Lee_1988_785} is used for electron-electron exchange-correlation. Our epc17 functional with parameters $a$, $b$, and $c$ optimized to be 2.35, 2.4, and 3.2, is used for electron-proton correlation. Note that the parameters are fit to accurate proton densities for \ce{FHF-}, analogous to the Colle-Salvetti electron correlation formula, where the parameters were fit to the properties of the He atom. As for most parameterized electronic exchange-correlation functionals, the values of the parameters could significantly influence the final results. This parameter dependence will be investigated more thoroughly in future work.
This NEO-DFT/epc17 method is compared to the NEO-DFT/no-epc method, which includes no electron-proton correlation, as well as a grid-based method \cite{Marston_1989_3571} that is considered to be a reference for these electronically adiabatic systems. In the grid-based method, the total electronic energy is calculated using conventional electronic DFT/B3LYP for the hydrogen nucleus positioned at each grid point on a three-dimensional grid spanning the relevant region for the proton density, and the three-dimensional Schr{\"o}dinger equation is solved numerically for the proton using the Fourier grid Hamiltonian method \cite{Marston_1989_3571}. The proton density is defined to be the square of the proton wave function calculated in this manner. The grid method is numerically exact when the Born-Oppenheimer approximation is valid, as for these model systems.
\begin{figure}
\begin{center}
\includegraphics[scale=0.28]{FHFSumPicNew.eps}
\caption{\label{fig:FHF}On-axis and off-axis proton densities for \ce{FHF-} calculated with the grid-based reference method(solid black line), the NEO-DFT/no-epc method (dashed red line), and the NEO-DFT/epc17 method (dotted blue line). The midpoint of the F---F bond is set as the origin.}
\end{center}
\end{figure}
\begin{figure}
\begin{center}
\includegraphics[scale=0.28]{HCNSumPicNew.eps}
\caption{\label{fig:HCN}On-axis and off-axis proton densities for \ce{HCN} calculated with the grid-based reference method(solid black line), the NEO-DFT/no-epc method (dashed red line), and the NEO-DFT/epc17 method (dotted blue line). The position of the carbon atom is set as the origin. The expectation values for the proton coordinate on the principal axis are 2.001, 2.053, 2.028 a.u., respectively.}
\end{center}
\end{figure}
Figure \ref{fig:FHF} shows the on-axis and off-axis proton densities for \ce{FHF-}, where the on-axis proton density is the slice along the axis connecting the two heavy nuclei, and the off-axis proton density is a slice perpendicular to this axis passing through the midpoint of the F---F bond. The NEO-DFT/no-epc method predicts a highly over-localized proton density, whereas the NEO-DFT/epc17 method is in excellent agreement with the grid-based reference method. Despite visual appearances, all of these proton densities are normalized in three-dimensional space, where multiplication by the volume element more heavily weights the regions further from the origin. In contrast, the results from previous papers were rescaled to be normalized in one dimension for the on-axis density \cite{Sirjoosingh_2015_214107, Culpitt_2016_44106, Brorsen_2015_214108} but contained significant errors because of a poor description of the off-axis proton density. The current epc17 functional corrects this problem and exhibits excellent agreement for both on-axis and off-axis proton densities with proper three-dimensional normalization. To our knowledge, no other electron-proton correlation functional can describe the proton densities with proper normalization in three dimensions even qualitatively accurately.
Without further parameterization, we applied the NEO-DFT/epc17 method to \ce{HCN}, an asymmetric case where the hydrogen atom is located at one end of the molecule. The results for this molecule are shown in Figure \ref{fig:HCN}. The NEO-DFT/epc17 method also provides greatly improved results compared to the NEO-DFT/no-epc method for both on-axis and off-axis proton densities, although minor differences can be observed. Note that altering the $a$, $b$, and $c$ parameters within the epc17 functional did not lead to significant improvement, implying transferability across different molecules for this functional, although additional molecules will be studied for further testing. Inclusion of the last two terms in Eq. \eqref{eq:Ecorr} may further improve this type of asymmetric case.
\begin{figure}
\begin{center}
\includegraphics[scale=0.45]{VepcForProtonNew.eps}
\caption{\label{fig:Vepc}Electron-proton correlation potential in the proton Kohn-Sham equation obtained from the epc17 functional as a function of the electron and proton densities. Because high proton density leads to high potential energy, the proton tends to become delocalized to reduce the potential energy.}
\end{center}
\end{figure}
We investigated the reason for the more delocalized proton densities obtained with the epc17 functional by plotting the epc proton potential $v^{epc}_p$ as a function of the electron and proton densities in Figure \ref{fig:Vepc}. The ranges of the variables $\rho_e$ and $\rho_p$ are chosen to be in the physical regime for hydrogen atoms in a typical molecular environment. This plot shows that when the proton density increases, the epc potential for the proton also increases. Driven by this extra potential, the proton tends to become more delocalized to minimize the area with high proton density and hence decrease the potential energy. This property explains why such a simple LDA functional can effectively delocalize the proton density.
Thus, this LDA type of electron-proton correlation functional should be viewed as a first important step, with the understanding that gradient corrections will be critical for future functional development. Moreover, because the high and low density limits of electron-proton correlation are unknown, the parameterized form developed herein focuses on the physically meaningful regions for molecular systems, and the asymptotic high and low density limits are most likely not correct. Further analysis of these limits and development of functional forms that satisfy such constraints are other directions for future development.
In addition, similar to the LYP functional, the epc17 functional does not explicitly include the kinetic correlation effects. These effects could be included with an adiabatic connection formula, \cite{Langreth_1975_1425, Imamura_2002_6458} as in a previous electron-proton correlation functional from our group. \cite{Sirjoosingh_2011_2689, Sirjoosingh_2012_174114} However, because the epc17 functional is parameterized to fit numerically exact densities or energies, these effects are implicitly included in a semiempirical manner. Another challenge in the development of electronic functionals has been the accurate inclusion of exchange terms for electrons. In contrast to electronic functionals, however, electron-proton functionals do not need to consider exchange between an electron and a proton because they are not identical particles. Given the extensive literature aimed at improving electronic functionals, future efforts will explore alternative models and treatments for electron-proton correlation functionals.
\section{Conclusions}
In this paper, we developed an electron-proton correlation functional analogous to the LYP electron correlation functional based on the Colle-Salvetti formulation. The implementation of this epc17 functional within the SCF procedure for multicomponent DFT produces accurate proton densities that are in good agreement with reference data for two representative molecules. To our knowledge, such accurate proton densities have not been achieved with any previous multicomponent DFT formulations. Moreover, the NEO-DFT/epc17 method is computationally inexpensive with the same formal scaling as conventional electronic DFT and is therefore promising for a wide range of future applications. For example, recent work based on the epc17 functional has been used to calculate proton affinities and to investigate the impact of proton delocalization on optimized geometries. \cite{Brorsen_2017_3488,Footnote} In addition, the NEO-DFT/epc17 method will enable the inclusion of nuclear quantum effects in calculations of p\textit{K}\textsubscript{a}'s, reaction paths, and reaction dynamics, as well as tunneling splittings and vibronic couplings. Therefore, this development lays the foundation for a wide range of potential new directions.
\begin{acknowledgments}
This work was supported by the National Science Foundation Grant No. CHE-1361293.
\end{acknowledgments}
\section*{Supplementary material}
See supplementary material for the complete derivation of the epc17 functional.
\section*{References}
| 5bd78f6ce040027be8af8f2358369593e311e65b | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Clusters of galaxies, the most massive objects in the Universe, are continuously growing, both by the steady accretion of matter from their surrounding environment, and by occasional mergers with smaller sub-clusters. The diffuse intergalactic gas accreted by clusters is rapidly shock heated, giving rise to the hot ($10^7$--$10^8$~K) X-ray emitting intra-cluster medium (ICM) that pervades clusters. The ICM is in approximate virial equilibrium, with the outer boundary of the virialized region - the virial radius - being approximately equal to $1.3r_{200}$, where within $r_{200}$ the mean enclosed mass density of the cluster is 200 times the critical density of the Universe at the cluster redshift \citep{lacey1993}. Galaxy clusters are also unique astrophysical laboratories that allow us to study nucleosynthesis and the chemical enrichment history of the Universe \citep[see][]{werner2008}. The deep gravitational potential wells of galaxy clusters hold all of the metals ever produced by stars in member galaxies, making them archaeological treasure troves to study the integrated history of star formation \citep{deplaa2007,mernier2016}. The dominant fraction of the metals in clusters currently resides within the hot ICM, which constitutes $\gtrsim 70$\% of the baryonic mass content for systems above $1.4\times10^{14}~M_{\odot}$ \citep{giodini2009}. However, when and how these metals were injected into the intergalactic medium is not well understood.
Most of the line emission from metals in the ICM arises from K- and L-shell transitions of highly ionized elements \citep[see][]{bohringer2010}. Because the ICM is in collisional ionization equilibrium and is optically thin, the equivalent widths of the emission lines can be converted directly into elemental abundances. The strongest line emission in the X-ray band is produced by the K-shell transitions of helium-like iron, making it an excellent tracer of chemical enrichment.
It has been known for about 40 years that a significant portion of the hot plasma in the central regions of galaxy clusters (the inner $\sim0.3r_{200}$) has been enriched by iron produced in stars to about one-third to one-half of the Solar value \citep{mushotzky1978,mushotzky1981}. In the central regions of clusters with strongly peaked ICM density distributions the abundance of iron is also peaked \citep[e.g.][]{degrandi2004}, but decreases with radius to about one-third Solar \citep[assuming the Solar abundances of][]{asplund2009} beyond about $0.2r_{200}$ \citep{leccardi2008}. Due to the low X-ray surface brightness in the outskirts of clusters, metal abundance measurements beyond one-half of the virial radius of clusters remain sparse.
The best measurements of the Fe abundance distribution at large radii were performed using the {\emph{Suzaku}}\ Key Project data (1 Ms observation along 8 azimuthal directions) of the Perseus cluster, which provided 78 data points outside of the cluster core ($r>0.25r_{200}$). These data revealed a remarkably uniform iron abundance, as a function of radius and azimuth, that is statistically consistent with a constant value of $Z_{\rm Fe} = 0.314 \pm 0.012$ Solar \citep[using the Solar abundance scale of ][]{asplund2009} out to $r_{200}$ \citep{werner2013}. Subsequent {\emph{Suzaku}}\ observations of the Virgo cluster extended these measurements to elements other than iron indicating an uniform chemical composition throughout the cluster volume \citep{simionescu2015}. The observed homogeneous distribution suggests that most of the metal enrichment of the ICM occurred before the cluster formed and its entropy distribution became stratified, preventing further efficient mixing. A key prediction of this early enrichment scenario is that the ICM in all massive clusters should be uniformly enriched to a similar level \citep{werner2013,fabjan2010,biffi2017}.
In order to test these predictions, we have analysed all archival observations of nearby galaxy clusters observed with {\it Suzaku} for which data extend to $r\sim r_{200}$ and robust measurements based on the Fe-K lines can be performed at $r > 0.25r_{200}$. (Because the Fe-K complex is by far the strongest line complex in the X-ray spectrum, the word metallicity will in this paper refer to and will be used interchangeably with the Fe abundance.) Our sample spans a redshift range $z=0.017$--$0.183$ and a temperature range of about $2.5$--$9$~keV \citep[the corresponding range of virial masses is about 1.4--$14\times10^{14}~M_\odot$;][]{arnaud2005}. The selected temperature range permits metallicity measurement using the Fe-K lines, allowing us to largely avoid multi-temperature biases arising from the measurements of the Fe-L complex \citep{buote2000}.
Sect.~\ref{sect:metObsanal} describes the data analysis, spectral modeling, and the treatment of the X-ray background. In~Sect.~\ref{sect:metResults}, we present the results. Finally, in Sect.~\ref{discussion} and \ref{sect:metConclusions}, we briefly discuss the implications of these results and draw our conclusions.
\begin{table}
\scriptsize
\centering
\caption{Details of \emph{Suzaku} observations used in the analysis. The
columns show, respectively, the target name, the {\emph{Suzaku}}\ observation ID, the date of the
observation and the clean exposure time.}
\label{tab:metData}
\begin{tabular}{lccc}
\hline\hline
\textbf{Name}&\textbf{Obs. ID}&\textbf{Obs. date}&\textbf{Exposure (ks)}\\
\hline
ABELL 262 CENTER &802001010&2007-08-17& 10.9\\
ABELL 262 OFFSET 1$^\dagger$ &802079010&2007-08-06& 11.0\\
ABELL 262 OFFSET 2 &802080010&2007-08-08& 14.6\\
NCG 669$^\dagger$ &804049010&2009-07-05& 50.3\\
A262 NE1 &808108010&2014-02-14& 8.4 \\
A262 NE2$^\dagger$ &808109010&2014-02-15& 1.4 \\
A262 NE3$^\dagger$ &808110010&2014-02-15& 9.8 \\
A262 NE4 &808111010&2014-02-16& 31.0\\
A262 E1 &808112010&2014-02-06& 9.2\\
A262 E2$^\dagger$ &808113010&2014-02-07& 5.4\\
A262 E3 &808114010&2014-02-13& 10.9\\
A262 E4$^\dagger$ &808115010&2014-02-13& 9.8\\
\hline
ABELL 1795$^\dagger$ &800012010&2005-12-10&13.1\\
ABELL 1795 Near North$^\dagger$ &800012020&2005-12-10&24.4\\
ABELL 1795 Far North &800012030&2005-12-11&5.9\\
ABELL 1795 Near South &800012040&2005-12-11&11.2\\
ABELL 1795 Far South &800012050&2005-12-12&12.5\\
A1795\_FAR\_NORTHEAST$^\dagger$ &804082010&2009-06-28&4.7\\
A1795\_FAR\_SOUTHWEST &804083010&2009-06-29&8.3\\
A1795\_FAR\_WEST &804084010&2009-06-26&8.5\\
A1795\_NEAR\_WEST &804085010&2009-06-27&4.1\\
\hline
A1689-OFFSET1 &803024010&2008-07-23&9.1\\
A1689-OFFSET2 &803025010&2008-07-24&7.5 \\
A1689-OFFSET3 &803026010&2008-07-25&8.9\\
A1689-OFFSET4$^\dagger$ &803027010&2008-07-26&6.2 \\
ABELL 1689 (OFFSET)$^\dagger$ &808089010&2013-06-27&17.5\\
ABELL 1689 (OFFSET) &808089020&2013-06-30&6.4\\
ABELL 1689 (OFFSET) &808089030&2013-12-31&53.2\\
ABELL 1689 (OFFSET) &808089040&2014-01-13&7.1\\
\hline
HYDRA A-1$^\dagger$ &805007010&2010-11-08& 6.9 \\
HYDRA A-2 &805008010&2010-11-09& 6.9 \\
HYDRA A SE &807087010&2012-06-07& 5.5 \\
HYDRA A FAR SE &807088010&2012-06-04& 5.6 \\
HYDRA A FAR N &807089010&2012-06-05& 5.6 \\
HYDRA A SW$^\dagger$ &807090010&2012-11-10& 39.8 \\
HYDRA A OUT &807091010&2012-06-07& 0.3 \\
\hline
ABELL 2029$^\dagger$ &802060010&2008-01-08&27.6\\
A2029\_1 &804024010&2010-01-28&3.5\\
A2029\_2 &804024020&2010-01-28&2.6\\
A2029\_3 &804024030&2010-01-28&6.3\\
A2029\_4 &804024040&2010-01-29&1.9\\
A2029\_5$^\dagger$ &804024050&2010-01-30&3.5\\
\hline
A2142 &801055010&2007-01-04&12.0 \\
A2142 OFFSET 1 &802030010&2007-08-04&7.1 \\
A2142 OFFSET 2 &802031010&2007-09-15&57.7 \\
A2142 OFFSET 3 &802032010&2007-08-29&7.0 \\
FILAMENT OF GALAXIES &805029010&2010-07-29&19.9 \\
\hline
ABELL 2204 &801091010&2006-09-17&14.3 \\
A2204\_FIELD\_1$\dagger$ &805056010&2010-09-01&5.9 \\
A2204\_FIELD\_2 &805057010&2010-08-27&6.9 \\
A2204\_FIELD\_3 &805058010&2010-08-28&6.8 \\
\hline
A133\_W &805019010&2010-06-07&50.0 \\
A133\_N &805020010&2010-06-05&50.2 \\
A133\_E$^\dagger$ &805021010&2010-06-09&51.6 \\
A133\_S$^\dagger$ &805022010&2010-06-08&51.1 \\
A133\_FIELD\_1 &808081010&2013-12-19&53.6 \\
A133\_FIELD\_2 &808082010&2013-12-20&50.6 \\
A133\_FIELD\_3 &808083010&2013-12-05&51.9 \\
A133\_FIELD\_4$^\dagger$ &808084010&2013-12-06&52.5 \\
\hline
SWIFT J0250.7+4142 &709006010&2014-08-03&82.2 \\
AWM7$^\dagger$ &801035010&2006-08-07&19.0 \\
AWM7 EAST OFFSET &801036010&2006-08-05&38.5 \\
AWM7 WEST OFFSET$^\dagger$ &801037010&2006-08-06&39.8 \\
AWM7 EAST OFFSET &802044010&2008-01-27&85.6 \\
AWM7 SOUTH OFFSET$^\dagger$ &802045010&2008-01-29&31.3 \\
AWM7 SOUTH OFFSET$^\dagger$ &802045020&2008-02-23&91.2 \\
AWM7 45' EAST &806008010&2011-08-07&36.9 \\
AWM7 27' SOUTH$^\dagger$ &806009010&2012-02-18&35.0 \\
AWM7 45' SOUTH &806010010&2012-02-17&34.4 \\
AWM7 NW1 &808023010&2014-02-17&14.9 \\
AWM7 NW2$^\dagger$ &808024010&2014-02-17&35.3 \\
AWM7 SE1 &808025010&2014-02-18&16.8 \\
AWM7 SE2$^\dagger$ &808026010&2014-02-19&35.3 \\
\hline\hline
\multicolumn{4}{l}{$^\dagger$observations influenced by SWCX}
\end{tabular}
\end{table}
\begin{table*}
\centering
\caption{Central coordinates, redshifts, mean temperatures, and values of $r_{200}$ for the clusters in our sample. The central coordinates are from the MCXC \citep{piffaretti2011}, except for A~1689, which are from the RASS-BSC \citep{voges1999}.}
\label{tab:metClusters}
\begin{tabular}{l|llccc}
\hline\hline
&RA (J2000)& dec (J2000)& $z$&$kT_{\rm ref}^{\dagger}$ (keV)&$r_{200, {\rm ref}}\,({\rm Mpc})$\\
\hline
\textbf{A~262} &01h 52m 46.8s &+36$^{\circ}$ 09' 05'' &0.017&2.3& 1.52 \citep{neill2001}\\
\textbf{A~1795} &13h 48m 53.0s &+26$^{\circ}$ 35' 44'' &0.063&6.2& 1.90 \citep{bautz2009}\\
\textbf{A~1689} &13h 11m 29.5s &$-$01$^{\circ}$ 20' 14'' &0.183&8.6& 2.50 \citep{umetsu2008}\\
\textbf{Hydra~A} &09h 18m 06.5s &$-$12$^{\circ}$ 05' 36'' &0.054&3.8& 1.48 \citep{sato2012}\\
\textbf{A~2029} &15h 10m 55.0s &+05$^{\circ}$ 43' 14'' &0.077&7.9& 1.92 \citep{walker2012b}\\
\textbf{A~2204} &16h 32m 46.5s &+05$^{\circ}$ 34' 14'' &0.152&6.4& 1.84 \citep{reiprich2009}\\
\textbf{A~2142} &15h 58m 20.6s &+27$^{\circ}$ 13' 37'' &0.091&8.5& 2.48 \citep{akamatsu2011}\\
\textbf{A~133} &01h 02m 42.1s &$-$21$^{\circ}$ 52' 25'' &0.057&4.0& 1.60 \citep{morandi2014}\\
\textbf{AWM~7} &02h 54m 29.5s &+41$^{\circ}$ 34' 18'' &0.017&3.7& 1.47 \citep{walker2014}\\
\hline\hline
\multicolumn{4}{l}{$^\dagger$\citet{ikebe2002}}
\end{tabular}
\end{table*}
\section{Observations and Data Analysis}
\label{sect:metObsanal}
The details of the {\emph{Suzaku}}\ observations for each of the 9~clusters
analyzed in this study are shown in~Tab.~\ref{tab:metData}. For each
cluster we analyzed the data from all available X-Ray Imaging Spectrometers
(XIS 0, 1, 2\footnote{XIS2 was lost to a likely micrometeoroid hit on 2006
November~9, and therefore its data are available only for the observations
from before this date.}, 3).
\subsection{Data Reduction}
\label{subs:metDataReduction}
We obtained the initial cleaned event lists using the standard criteria
provided by the XIS~team\footnote{Arida,~M., XIS Data Analysis,\\
http://heasarc.gsfc.nasa.gov/docs/suzaku/analysis/abc/node9.html}. There is
a gradual increase in the number of flickering pixels in the XIS~detectors
with time, which may affect the measurements, if unaccounted for. We used
maps provided by the XIS~team\footnote{The current maps, as well as the recipe for the removal
process, are available at
http://www.astro.isas.ac.jp/suzaku/analysis/xis/nxb\_new/} to remove the flickering pixels from the cluster
observations, as well as from the night Earth observations, which were later used
to create the non-X-ray background (NXB) data products.
We checked for likely solar wind charge-exchange (SWCX) emission
contamination using the WIND Solar Wind Experiment data\footnote{The
data are available at: ftp://space.mit.edu/pub/plasma/wind/kp\_files/},
following the analysis of~\citet{fujimoto2007}. In the case of affected
pointings, marked in~Tab.~\ref{tab:metData}, we only used data above
1.5~keV, where no SWCX emission lines are expected.
We filtered out times of low geomagnetic cut-off rigidity (COR$>$6~GV).
Ray-tracing simulations of spatially uniform extended emission were used to
perform vignetting corrections \citep{ishisaki2007}. For the XIS~1 data
obtained after the reported charge injection level increase on
2011~June~1st, we have excluded two adjacent rows on either side of the
charge-injected rows. (The standard is to exclude one row on either side.)
\subsection{Image analysis}
\label{subs:metImaging}
We extracted images from all XIS~detectors in the $0.7-7.0$~keV energy band, removing $\sim30$~arcsec regions around the edges. We extracted
instrumental background images in the same energy band from
flickering-pixel-subtracted night Earth observations using the tool
\textsc{xisnxbgen}. We subtracted the background images from the cluster
images before applying the vignetting correction. The resulting mosaics for
each cluster are shown
in~Figs.~\ref{fig:a262portrait}--\ref{fig:awm7portrait}.
\subsection{Point Source Detection}
\label{subs:metPsources}
The initial identification of the point sources was carried out using the
\textsc{ciao} tool \textsc{wavdetect}. We used a single wavelet radius of
1~arcmin, which is approximately matched to the half-power radius of the X-ray telescopes on
{\emph{Suzaku}}. For each cluster we created a candidate set of point
sources assuming a source with a radius of 1~arcmin at each of the positions
identified by \textsc{wavdetect}. We then calculated X-ray
surface brightness profiles centered on the coordinates in
Tab.~\ref{tab:metClusters}, excluding the candidate set of point sources,
and fitted an isotropic $\beta$-model to the surface
brightness profile of each cluster:
\begin{equation}
S_X=S_0\left[1+\left(\frac{r}{r_c}\right)^2\right]^{(-3\beta+0.5)}+S_{X, \rm
bkg},
\label{eqn:beta}
\end{equation}
where $r$ is the distance from the cluster center and the free parameters
are the normalization $S_0$, the core radius $r_c$ and $\beta$. $S_{X, \rm bkg}$ is
the surface brightness of the X-ray background, which is assumed to be
constant across the whole area of the cluster. The best-fit parameters for
the individual clusters are shown in~Tab.~\ref{tab:metSx}.
\begin{table}
\centering
\caption{$\beta$ model parameters for each cluster. See Eqn.~\ref{eqn:beta}
for the interpretation of the individual columns. Parameters $S_0$ and $S_{X, \rm bkg}$
are in the units of counts/s/pixel$^2$ and the core radius $r_c$ is in {\emph{Suzaku}}\
pixels, where $1\,\text{arcmin}=7.2\,\text{pixels}$.}
\label{tab:metSx}
\begin{tabular}{l|cccc}
\hline\hline
&$S_0\times10^{5}$&$r_c$&$\beta$&$S_{X, \rm bkg}\times10^7$\\
\hline
\textbf{A~262} &$1.45\pm0.05$&$18.3\pm0.7$&$0.51\pm0.01$&$0.98\pm0.035$\\
\textbf{A~1795} &$7.78\pm0.26$&$14.7\pm0.4$&$0.70\pm0.01$&$0.90\pm0.04$\\
\textbf{A~1689} &$4.40\pm0.43$&$10.6\pm0.8$&$0.78\pm0.02$&$1.18\pm0.03$\\
\textbf{Hydra~A} &$3.88\pm0.23$&$14.1\pm0.8$&$0.67\pm0.02$&$1.03\pm0.06$\\
\textbf{A~2029} &$9.52\pm0.37$&$14.1\pm0.5$&$0.69\pm0.01$&$1.37\pm0.04$\\
\textbf{A~2204} &$4.44\pm0.17$&$16.5\pm0.7$&$0.93\pm0.03$&$2.89\pm0.08$\\
\textbf{A~2142} &$4.58\pm0.11$&$20.3\pm0.6$&$0.70\pm0.01$&$8.15\pm0.05$\\
\textbf{A~133} &$0.96\pm0.07$&$13.3\pm0.9$&$0.73\pm0.02$&$0.71\pm0.04$\\
\textbf{AWM~7} &$2.37\pm0.05$&$29.4\pm0.8$&$0.55\pm0.01$&$0.75\pm0.04$\\
\hline\hline
\end{tabular}
\end{table}
We divided the mosaic images of the individual clusters by the best-fit
surface brightness models and used the resulting residual images to identify
by eye sources with radii larger than 1~arcmin. In these cases, we manually
increased the sizes of the sources in question by the appropriate amount.
The resulting updated sets of point sources (including substructures and artifacts that can appear at chip edges), which were excluded from the
subsequent spectral analysis, are shown with magenta circles in
Figs~\ref{fig:a262portrait} -- \ref{fig:awm7portrait}.
\subsection{Spectral Analysis}
\label{subs:metSpectral}
For each cluster, we extracted spectra from a series of concentric annular
regions centered on the respective cluster's centre (see
Tab.~\ref{tab:metClusters}). The width of the annuli was set to be at least
3~arcmin, with each annulus containing at least 3500~cluster counts, allowing us, in principle, to
measure the Fe~abundance with a relative uncertainty of at most $20\%$.
The resulting annuli are shown in yellow
in~Figs.~\ref{fig:a262portrait}--\ref{fig:awm7portrait}.
Instrumental background spectra were created using Night Earth observations.
We rebinned each spectrum to a minimum of one count per bin,
employing the extended C-statistic \citep{cash1979,arnaud1996} in the
fitting. For each spectrum, we constructed an individual response matrix
with a resolution of 16~eV, spanning the $0.2-9.5$~keV energy band\footnote{These
rebinned matrices require $\sim25$~times less disk space compared to the default
choice of 2~eV resolution and the full $0.2-16.0$~keV {\emph{Suzaku}}\ energy band,
allowing us to model all the spectra for a given cluster simultaneously and
speeding up the analysis, without compromising the accuracy of the results.}.
We used the task \textsc{xissimarfgen} (version 2010-11-05) to create
ancillary response files (assuming uniform emission from a circular region
with the radius of 20~arcmin).
We used \textsc{xspec} \citep[][version 12.9.0]{arnaud1996} to model the
spectra. For each cluster we modeled all spectra simultaneously, using the
$0.7-7.0$~keV band for the front illuminated XIS~0, XIS~2 and
XIS~3 detectors, and the $0.6-7.0$~keV band for the back illuminated XIS~1, except for observations with possible SWCX contamination (see
Tab.~\ref{tab:metData}), where we used the $1.5-7.0$~keV energy band. We
modeled the ICM emission in each annulus as a single temperature plasma in collisional
ionisation equilibrium using the absorbed \textsc{apec (ATOMDB 3.0.3)} model
\citep{smith2001}.
For a given annulus, we used a single temperature and metallicity. Normalizations were allowed to vary among individual observations,
but were tied among the detectors in a single observation; in other
words, all spectra from a given observation were members of a single fitting
group in \textsc{xspec}. We used the abundance table of~\citet{asplund2009}
in the analysis. Galactic absorption was set to the average column
along the line of sight inferred from the Leiden/Argentine/Bonn Survey
\citep{kalberla2005}. The uncertainties in
all derived parameters were determined using Markov Chain Monte Carlo (MCMC)
simulations.
After removing the burn in period and thinning each chain, we used the mean and the standard deviation as the value and the uncertainty of each derived parameter, respectively.
\begin{table*}
\setlength{\extrarowheight}{4pt}
\centering
\caption{The CXFB model parameters for the individual clusters. The four
CXFB model components we used are the power-law component (PL), the Galactic
halo (GH), the hot foreground component (HF) and the Local Hot Bubble (LHB).
Subscript n stands for normalization, kT for temperature and ind for index.
Normalizations are in units of $\int n_en_H\,dV\times\frac{10^{-14}}{4\pi\left[D_A(1+z)\right]^2}\frac1{20^2\pi}\,\text{cm}^{-5}\,\text{arcmin}^{-2}$.}
\label{tab:metCxfb}
\begin{tabular}{l|cccccccc}
\hline\hline
&PL\textsubscript{ind}&PL\textsubscript{n}$\times10^{4}$&GH\textsubscript{kT}&GH\textsubscript{n}$\times10^3$&HF\textsubscript{kT}&HF\textsubscript{n}$\times10^4$&LHB\textsubscript{kT}&LHB\textsubscript{n}$\times10^4$\\
\hline
\textbf{A~262} &$1.43_{-0.05}^{+0.05}$&$9.38_{-0.64}^{+0.67}$&$0.18_{-0.01}^{+0.01}$&$2.82_{-0.19}^{+0.13}$&$0.94_{-0.05}^{+0.04}$&$2.59_{-0.22}^{+0.12}$&$0.103_{-0.002}^{+0.002}$&$9.01_{-0.30}^{+0.18}$\\
\textbf{A~1795} &$1.38_{-0.08}^{+0.05}$&$10.07_{-0.79}^{+0.84}$&$0.22_{-0.01}^{+0.01}$&$0.91_{-0.08}^{+0.18}$&N/A&N/A&$0.10_{-0.01}^{+0.01}$&$49.10_{-0.36}^{+0.26}$\\
\textbf{A~1689} &$1.34_{-0.03}^{+0.04}$&$9.16_{-0.41}^{+0.47}$&$0.18_{-0.02}^{+0.01}$&$2.43_{-0.40}^{+0.50}$&$0.59_{-0.03}^{+0.05}$&$2.11_{-0.32}^{+0.64}$&$0.10_{-0.01}^{+0.01}$&$13.65_{-0.20}^{+0.23}$\\
\textbf{Hydra~A} &$1.39_{-0.09}^{+0.07}$&$8.61_{-0.90}^{+1.15}$&$0.12_{-0.01}^{+0.01}$&$4.45_{-1.30}^{+1.40}$&$0.79_{-0.16}^{+0.11}$&$1.91_{-0.16}^{+0.36}$&$0.10_{-0.01}^{+0.01}$&$9.53_{-1.20}^{+1.20}$\\
\textbf{A~2029} &$1.25_{-0.08}^{+0.07}$&$7.01_{-0.63}^{+0.69}$&$0.18_{-0.02}^{+0.01}$&$6.18_{-1.05}^{+1.15}$&$0.58_{-0.01}^{+0.02}$&$13.95_{-1.30}^{+0.90}$&$0.10_{-0.01}^{+0.01}$&$13.4_{-0.37}^{+0.19}$\\
\textbf{A~2204} &$1.49_{-0.10}^{+0.10}$&$12.52_{-1.47}^{+1.62}$&$0.227_{-0.005}^{+0.008}$&$10.63_{-2.25}^{+0.45}$&$0.61_{-0.02}^{+0.02}$&$10.25_{-0.12}^{+0.22}$&$0.13_{-0.01}^{+0.06}$&$17.50_{-2.00}^{+1.08}$\\
\textbf{A~2142} &$1.37_{-0.05}^{+0.06}$&$9.19_{-0.61}^{+0.64}$&$0.146_{-0.004}^{+0.004}$&$4.88_{-0.55}^{+55}$&$0.64_{0.02}^{+0.06}$&$2.51_{-0.22}^{+0.14}$&$0.09_{-0.01}^{+0.01}$&$7.33_{0.30}^{+0.60}$\\
\textbf{A~133} &$1.54_{-0.07}^{+0.10}$&$10.37_{-0.95}^{+1.27}$&$0.147_{-0.006}^{+0.012}$&$1.93_{-0.20}^{+0.58}$&$0.61_{-0.16}^{+0.06}$&$1.18_{-0.20}^{+0.35}$&$0.08_{-0.01}^{+0.01}$&$7.33_{-0.90}^{+0.65}$\\
\textbf{AWM~7} &$1.52_{-0.05}^{+0.02}$&$11.5_{-0.72}^{+0.30}$&$0.150_{-0.002}^{+0.004}$&$3.60_{-0.23}^{+0.13}$&$0.78_{-0.02}^{+0.02}$&$4.16_{-0.12}^{+0.14}$&$0.09_{-0.01}^{+0.01}$&$9.65_{-0.07}^{+0.10}$\\
\hline\hline
\end{tabular}
\end{table*}
\subsection{Modeling the X-ray Foreground and Background}
\label{subs:metCxfb}
At large clustercentric radii, the cosmic X-ray foreground/background (CXFB)
makes up a dominant fraction of the total X-ray emission, requiring careful
modeling.
Our spectral model for the CXFB included four components -- an
absorbed power law (PL) due to the unresolved point sources
\citep{deluca2004}, an absorbed thermal component modeling the Galactic
halo emission \citep[GH,][]{kuntz2000}, a potential 0.6~keV foreground component
that we will from now on refer to as the hot foreground
\citep[HF,][]{masui2009,yoshino2009}, and an unabsorbed thermal component
modeling the emission from the local hot bubble \citep[LHB,][]{sidher1996}.
In order to better constrain the low-temperature CXFB components, we used
the X-Ray Background
Tool\footnote{http://heasarc.gsfc.nasa.gov/cgi-bin/Tools/xraybg/xraybg.pl},
which calculates the average X-ray background spectra from the ROSAT All-Sky
Survey diffuse background maps. For each cluster we obtained spectra from
six independent circular regions with radii $R=1.3r_{200}$ evenly
surrounding the cluster so that each touches the outer edge of a circle with radius $r=1.3r_{200}$ centered on the cluster core, as well as two neighbouring background regions. The distance from the cluster core ensures that the
spectra are not significantly contaminated by emission from the ICM. This
setup, as opposed to using a single annular region, allowed us to
assign separate absorptions to each of the six regions, which can
potentially significantly influence the modeling at {\it ROSAT}\ energies,
$0.7-2.0$~keV.
\begin{figure*}
\centering
\includegraphics[width=.6\textwidth]{metals_noSmooth.pdf}
\caption{Iron abundance measurements in our cluster sample plotted as a function of radius scaled to $r_{200}$. On average, the iron abundances peak in the cores of the clusters and decrease as a function of radius, flattening at radii $r>0.25r_{200}$. The average metallicity is shown as blue solid line. The dashed line shows the best fit metallicity reported by \citet{werner2013} for the Perseus cluster. }
\label{fig:radial}
\end{figure*}
As stated before, for each cluster we modeled all spectra simultaneously, including the CXFB model. To limit the systematic effects that might
potentially influence the fit, we first separately determined the
PL~parameters, that were kept fixed during the subsequent modeling. To do
this, we used the spectra from the outermost annulus in a given cluster in
the high energy band, $2.0-7.0$~keV, where the PL~component is dominant at
large clustercentric radii. For six of the clusters in our sample,
A~133, A~1689, A~1795, A~2029, A~2142, and A~2204, the outermost
annulus covers only regions outside $r_{200}$. In these cases, we assumed
no significant cluster emission to contribute to the spectrum and therefore
the high energy fit included only the PL~model. For the remaining clusters,
A~262, AWM~7 and Hydra~A, we accounted for potential cluster emission by
including an \textsc{apec} component in the high energy fit, fixing its
temperature to $kT=2$~keV and keeping the
normalizations in the individual observations free. The best-fit
PL~parameters are shown in the first two columns of Tab.~\ref{tab:metCxfb}.
In the subsequent spectral modeling, the remaining CXFB parameters were
kept free and tied among all spectra from a given cluster, with the
exception of the metallicity and the redshift of the three thermal
components, which were fixed at unity and zero, respectively.
For each cluster, we tied
together the ICM~temperatures and Fe~abundances in the neighboring annuli
where required in order to obtain a statistically significant constraint. The final CXFB model parameters for the individual clusters
are listed in Tab.~\ref{tab:metCxfb}.
\begin{figure*}
\centering
\includegraphics[width=.6\textwidth]{metallicities_noIso_noPlCrit.pdf}
\caption{Robust measurements (see text for details) of the iron abundances at $r>0.25r_{200}$ in
the individual clusters. The clusters have been ordered by mass from the
least to the most massive. The blue stripe marks the 68\% confidence interval
around the constant fit to these data, $Z_{\rm Fe}=0.316\pm0.012$~Solar.
The red stripe shows the confidence interval around the best fit iron abundance
reported by \citet{werner2013} for the Perseus cluster, $Z_{\rm Fe}=0.314\pm0.012$~Solar.}
\label{fig:metConservative}
\end{figure*}
\subsection{Criteria For Robust Metallicity Measurements}
\label{subs:metTrimming}
In each cluster we formally obtained profiles of temperature and metallicity
out to the outermost radii reached by the observations. However, due to the low
surface brightness of the ICM, the measurements at large radii may be
significantly influenced by systematic uncertainties, such as potential
variations of the CXFB model throughout the cluster. To address our main
scientific questions, it is therefore crucial to only use the metallicity
measurements which we are confident about. We used the following criteria to
identify these measurements:
\begin{itemize}
\item We only used the measurements at radii $r>0.25r_{200}$ to avoid the
central metallicity peak observed in most cool-core galaxy clusters. We used a
similar radial range in the Perseus cluster
\citep{werner2013}, where metallicity measurements inside $20'$
($r_{200}=82'$) were discarded.
\item We only used annuli with ICM-signal-to-background (ISB) ratios
around the Fe-K lines higher than 10\%. This is defined as the ratio
of the total number of modeled counts received from the ICM to the sum
of modeled counts from the CXFB and the instrumental background, in a 1~keV
wide energy band centered on the appropriately redshifted Fe-K line (rest
energy $E=6.7\,\text{keV}$). The top right panels in
Figs.~\ref{fig:a262portrait}--\ref{fig:awm7portrait} show the profiles of
the ISB~ratios for the individual annuli (in blue), as well as the ratios
for the individual observations (in red). The 10\% threshold is broadly
consistent with the measurements in the outermost regions of the Perseus
cluster \citep{werner2013,urban2014}.
\item Finally, we only used the annuli where the contamination from the neighbouring regions due to the wings of the broad point spread function (PSF) of the telescopes is small. The half-power diameter (HPD) of the X-ray Telescopes (XRT) on board of {\emph{Suzaku}}\ is
$\sim2'$, which causes a fraction of the emission from an object to be
registered elsewhere on the detector. Addressing this issue is especially
important in the cool core clusters at relatively large distances
($z\gtrsim0.1$), since the emission from the metal-rich X-ray surface
brightness peak may bias spectral measurements out to larger radii.
To test for this, we used {\it Chandra}\ surface brightness profiles with high
spatial resolution relative to {\emph{Suzaku}}\ (binned to $\sim4''$), which we
convolved with a simple Gaussian model for the {\emph{Suzaku}}\ point-spread
function with a HPD of 2~arcmin. Using this model, for each of our annuli
we calculated the fraction of emission that we expected to come from the
other annuli, and removed those where it exceeded 10\%. Only the two most distant clusters in our sample, A~2204 and A~1689, were affected. For both systems, we removed the measurements
immediately outside $0.25r_{200}$ from the subsequent analysis.
\end{itemize}
\begin{table}
\begin{center}
\caption{Measurement radii and the best fit metallicities measured in our cluster sample. }
\begin{tabular}{lc|c}
\hline\hline
\textbf{cluster}&\textbf{r/r\textsubscript{200}}&\textbf{Z/Z}$_{\odot}$\\
\hline
\multirow{3}{*}{\textbf{AWM7}}&$0.285\pm0.015$&$0.386\pm0.035$\\
&$0.32\pm0.02$&$0.261\pm0.043$\\
&$0.37\pm0.03$&$0.404\pm0.071$\\
\hline
\multirow{3}{*}{\textbf{Hydra A}}&$0.30\pm0.06$&$0.268\pm0.055$\\
&$0.42\pm0.06$&$0.243\pm0.072$\\
&$0.54\pm0.06$&$0.361\pm0.115$\\
\hline
\textbf{A262}&$0.28\pm0.02$&$0.370\pm0.053$\\
\hline
\multirow{3}{*}{\textbf{A133}}&$0.31\pm0.06$&$0.454\pm0.071$\\
&$0.43\pm0.06$&$0.260\pm0.083$\\
&$0.56\pm0.07$&$0.321\pm0.120$\\
\hline
\multirow{2}{*}{\textbf{A2204}}&$0.66\pm0.13$&$0.413\pm0.076$\\
&$0.93\pm0.13$&$0.268\pm0.147$\\
\hline
\multirow{3}{*}{\textbf{A1795}}&$0.29\pm0.06$&$0.306\pm0.029$\\
&$0.40\pm0.05$&$0.341\pm0.056$\\
&$0.52\pm0.06$&$0.145\pm0.064$\\
\hline
\multirow{2}{*}{\textbf{A2029}}&$0.34\pm0.07$&$0.355\pm0.087$\\
&$0.55\pm0.14$&$0.276\pm0.094$\\
\hline
\multirow{2}{*}{\textbf{A2142}}&$0.31\pm0.06$&$0.358\pm0.042$\\
&$0.43\pm0.06$&$0.354\pm0.064$\\
\hline
\textbf{A1689}&$0.62\pm0.12$&$0.353\pm0.131$\\
\hline
\multirow{3}{*}{\textbf{Coma$^{\dagger}$}}
&$0.25\pm0.02$&$0.285\pm0.035$\\
&$0.29\pm0.02$&$0.281\pm0.059$\\
&$0.34\pm0.03$&$0.240\pm0.072$\\
&$0.40\pm0.03$&$0.327\pm0.068$\\
&$0.50\pm0.08$&$0.260\pm0.056$\\
&$0.71\pm0.13$&$0.41\pm0.18$\\
\hline\hline
\textbf{Average metallicity}&&$0.316\pm0.012$\\
\hline\hline
\end{tabular}\\
\end{center}
$^\dagger$\citet{simionescu2013}
\label{metaltable}
\end{table}
\section{Results}
\label{sect:metResults}
The best fit normalizations, temperatures and metallicities for the individual clusters are shown in the bottom panels of Figures \ref{fig:a262portrait}--\ref{fig:awm7portrait}. Most of the systems in our sample are so-called cooling core clusters with bright, relatively cool, metal-rich cores. To the iron abundance measurements in this work, we also added the iron abundances measured for the non-cool core Coma cluster by \citet{simionescu2013}.
Fig. \ref{fig:radial} shows all metallicity measurements in our cluster sample plotted as a function of radius scaled to $r_{200}$. The average metallicity (shown with the solid blue line) peaks in the central region and decreases as a function of radius, flattening at radii $r>0.25r_{200}$.
We tested our results for biases associated with possible multi-temperature structure by fitting the data both in the full spectral band and above 2 keV. At radii $r>0.25r_{200}$ the two fits provided consistent results, indicating that cooler temperature components do not contribute significantly to the observed emission in the cluster outskirts.
After excluding the measurements at $r<0.25r_{200}$ and using all criteria outlined in the previous section, we are left with
26~individual metallicity measurements from 10~different clusters. These measurements are shown in Table~\ref{metaltable} and Fig. \ref{fig:metConservative} where the clusters have been ordered by mass from the least to the most massive. There is no evidence for any trend in metallicity as a function of cluster mass.
The measurements are consistent with being constant at $Z_{\rm Fe}=0.316\pm0.012$~Solar,
with $\chi^2=28.85$ for 25 degrees of freedom. This best fit value is statistically consistent with $Z_{\rm Fe}=0.314\pm0.012$ Solar reported for the Perseus cluster \citep{werner2013}, shown as a dashed line in Fig. \ref{fig:radial}.
\section{Discussion}
\label{discussion}
We find that across our sample of 10 clusters of galaxies the Fe abundances measured outside the central regions ($r>0.25r_{200}$) are consistent with a constant value, $Z=0.316\pm0.012$ Solar (Fig. \ref{fig:radial}). The metallicity measurements also show no significant trend with temperature (Fig. \ref{fig:metConservative}).
Based on the uniform iron abundance distribution in the Perseus cluster, both as a function of radius and azimuth, statistically consistent with a constant value of $Z_{\rm Fe} = 0.314\pm0.012$ Solar out to $r_{200}$, \citet{werner2013} proposed that most of the metal enrichment of the intergalactic medium occurred before clusters formed, probably more than ten billion years ago ($z>2$), during the period of maximal star formation and black hole activity. A key prediction of the early enrichment scenario is that the ICM in all massive clusters should be uniformly enriched to a similar level. Previous indications for a uniform ICM enrichment include the small cluster to cluster scatter in the Fe abundance observed within $r_{500}$ \citep{matsushita2011,leccardi2008} and the observed pre-enrichment of the ICM between the clusters Abell 399/401 \citep{fujita2007}. Our observation of a constant iron abundance at large radii across a sample of 26 independent measurements for ten massive clusters further confirms this early enrichment scenario. This early enrichment could have been driven by galactic winds \citep{deyoung1978} which would be strongest around the peak of star formation and AGN activity \citep[redshifts $z\sim2-3$][]{madau1996,brandt2005}.
Recent numerical simulations by \citet{fabjan2010} and \citet{biffi2017} indicate that while star-formation and supernova feedback are unable to enrich the intergalactic medium uniformly, simulations which also include feedback from AGN produce remarkably homogeneous metallicity distribution in the ICM out to large radii. They show that the uniform metallicity is the result of a widespread
displacement of metal-rich gas by powerful AGN outbursts that occure {\it before} the epoch of maximal star-formation and AGN activity. \citet{biffi2017} conclude that early AGN feedback acting on high-redshift ($z > 2$) small haloes, with shallow gravitational potential wells, was particularly efficient in spreading and mixing the metals. Given the complexity of the physics of the chemical enrichment processes, these simulation results should probably be considered tentative. However, our measurements provide an important anchor with which the results of these and future simulations can be compared, bringing more understanding into the process of chemical enrichment.
The constant ratios of abundances of several elements observed throughout the Virgo cluster \citep{simionescu2015} as well as in the radial profiles of 44 clusters observed out to intermediate radii with {\it XMM-Newton} \citep{mernier2017} reveal that, during the early period of metal enrichment, the products of core-collapse and type Ia supernovae were well mixed. The estimated ratio between the number of SN Ia and the total number of supernovae enriching the ICM is about 15--20\%, generally consistent with the metal abundance patterns in our own Galaxy and only marginally lower than the SN Ia contribution estimated for the cluster cores \citep{simionescu2015}.
The most direct way to confirm the early enrichment scenario is to measure the core-excluded metallicity of clusters as a function of redshift. Contrary to initial findings \citep{Balestra2007,maughan2008,anderson2009,baldi2012}, under the early enrichment scenario, there should be no substantial redshift evolution in the ICM metallicity outside the central regions of clusters, out to $z \sim 2$. Recent results \citep{andreon2012,ettori2015,mcdonald2016, mantz2017} indicate that most metals in the ICM were already in place at $z = 1$, consistent with the picture of an early enrichment.
At various overdensities, the chemical enrichment might proceed on different time scales or with different initial mass functions, resulting in a trend with cluster mass. Within the mass range probed by our sample (factor of $\sim10$), there is no evidence for dependence of ICM metallicity on total cluster mass. A more thorough analysis of the mass dependence will require reliable measurements of absolute abundances in low mass clusters and groups of galaxies, which are often made difficult for current CCD instruments by multi-temperature structure in the ICM \citep{simionescu2015,simionescu2017}. A lack of trend with cluster mass would either indicate a rate of metal enrichment in the early Universe that is independent of the density
contrast between different regions, or a very high efficiency of mixing on large scales.
If the ICM at large radii is clumpy and multiphase \citep{simionescu2011,urban2014,simionescu2017} then its best fit metallicity, derived using a single temperature model, might be biased \citep{avestruz2014}. The best fit Fe abundance is the most significantly biased at temperatures around 1~keV, where its value is determined based on the Fe-L lines, which are very sensitive to the underlying temperature structure \citep[see][]{buote2000}. The metallicities of the clusters in our sample are determined using the Fe-K lines and depending on the temperature structure could be biased by at most 30 per cent \citep[both toward higher and lower values;][]{rasia2008,simionescu2009b,gastaldello2010}. The fact that the spectral fits in the full band and above 2~keV give consistent results (see Section~\ref{sect:metResults}) indicates that if substantially cooler, denser clumps are present in the ICM, they do not contribute significantly to the observed emission measure and the metal budget.
In the near future, the metallicities of groups and cooler clusters could be further studied with the {\it Astrosat} satellite \citep{singh2014}. The low earth orbit and the small inclination of the orbital plane of {\it Astrosat} provide a low and stable background environment that is required for cluster outskirts studies. The large field of view provides a sufficient grasp, enabling mapping the faint X-ray emission in the outskirts of nearby clusters that span large angular scales in the sky. Such observations will further test the possible mass-dependence of metallicity. Deep observations with {\it XMM-Newton} and {\it Chandra} will allow us to precisely determine the metallicity outside the cores of high redshift clusters, providing further constraints on the redshift evolution of metallicity. Observations with high spectral resolution obtained with the {\it X-ray Astronomy Recovery Mission (XARM)} will allow more accurately measured relative abundance ratios for clusters at low redshifts, testing our models of nucleosynthesis. In the further future, missions like {\it Athena} \citep{nandra2013} or {\it Lynx}\footnote{https://wwwastro.msfc.nasa.gov/lynx/} will allow detailed studies of metal abundances in high redshift clusters, providing comprehensive understanding of the metal cycle in the Universe.
\section{Conclusions}
\label{sect:metConclusions}
Here, we report 26~independent
metallicity measurements in the outskirts ($r>0.25r_{200}$) of ten nearby galaxy clusters.
These measurements are consistent with a constant
value $Z_{\rm Fe}=0.316\pm0.012$~Solar. No significant trend of metallicity versus temperature or mass is observed.
Our results corroborate the conclusions drawn from previous metallicity measurements at large radii in the Perseus cluster
\citep[][]{werner2013}. In particular, they confirm the predictions of an early
enrichment scenario, where the majority of metal enrichment occurs before
the cluster formation, at $z>2$.
\section*{Acknowledgments}
This work was supported in part by NASA grants NNX12AE05G and NNX13AI49G, and by the US Department of Energy under contract number DE-AC02-76SF00515, as well as by the Lend\"ulet LP2016-11 grant awarded by the Hungarian Academy of Sciences. The authors thank the {\it Suzaku} operation team and Guest Observer Facility, supported by JAXA and NASA.
\bibliographystyle{mnras}
| eafea86c0a6476cbfb7e058833e3d149b5d8f678 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Variational, oscillation and jump inequalities have been the subject of many
recent articles in probability, ergodic theory and harmonic
analysis. The first variational inequality was proved by L\'epingle \cite{Lep76} for martingales. Using L\'epingle's result, Bourgain \cite{Bou89} is the first one who to obtain corresponding variational estimates for the Birkhoff ergodic averages and then directly deduce pointwise convergence results without previous knowledge that pointwise convergence holds for a dense subclass of functions, which are not available in some ergodic models. Bourgain's work has initiated a new research direction in ergodic theory and harmonic analysis. In \cite{CJRW2000,CJRW2002,JKRW98,JRW03,JSW08}, Jones and his collaborators systematically studied jump and variational inequalities for ergodic averages and truncated singular integrals. Since then many other publications came to enrich the literature on this subject. Analogs are also true for corresponding maximal operators which were known. A number of phenomena show that variational, jump and oscillation operators seems to play the same role as maximal operators in harmonic analysis.
\par
In this paper we consider variation, oscillation and $\lambda$-jump operators associated
with approximate identities. The variational inequality gives us the characterization of $H^p$ for $p\in(\frac n{n+1},1]$. We obtain estimates for the oscillation and $\lambda$-jump acting on Hardy space. A counterexample show that oscillation and $\lambda$-jump operators are too small to characterize Hardy space. Before we present our main results we recall some definitions and known results.
Let $\mathcal I$ be a subset of $\mathbb R^+$, $\mathfrak{a}=\{a_t: t\in \mathcal I\}$ be a family of complex numbers and $\rho\ge1$. The $\rho$-variation norm of the family $\mathfrak{a}$ is defined by
\begin{equation}\nonumber
\|\mathfrak{a}\|_{v_\rho(\mathcal I)}=\sup\big(\sum_{k\geq1}
|a_{t_k}-a_{t_{k-1}}|^\rho\big)^{\frac{1}{\rho}},
\end{equation}
where the supremum runs over all finite decreasing sequences $\{t_k\}$ in $\mathcal I$. We denote the norm $v_\rho(\mathbb R^+)$ by $v_\rho$ for short. It is trivial that
\begin{equation}\nonumber
\|\mathfrak{a}\|_{L^\infty(\mathcal I)}:=\sup_{t\in \mathcal I}|a_t|
\le |a_{t_0}|+\|\mathfrak{a}\|_{v_\rho(\mathcal I)}
\end{equation}
for any $t_0\in \mathcal I$ and $\rho\ge1$.
\par
Given a family of Lebesgue measurable functions $\mathcal F(x)=\{F_t(x):t\in \mathcal I\}$, the value of the $\rho$-variation function $\mathscr V_q(\mathcal F)$ of the family $\mathcal F$ at $x$ is defined by
\begin{equation}\nonumber
\mathscr V_\rho(\mathcal F)(x)=\|\{F_t(x)\}\|_{v_\rho(\mathcal I)},\quad \rho\ge1.
\end{equation}
Specially, suppose $\mathscr{A}=\{{A}_t\}_{t>0}$ is a family of operators, the $\rho$-variation operator related $\mathscr A$ is simply defined as
$$\mathscr V_\rho(\mathscr Af)(x)=\|\{A_t(f)(x)\}_{t>0}\|_{v_\rho}.$$
It is easy to observe that for any fixed $x\in\mathbb R^n$, if $\mathscr V_\rho(\mathscr Af)(x)<\infty$, then $\lim\limits_{t\rightarrow0^+}A_t(f)(x)$ and $\lim\limits_{t\rightarrow+\infty}A_t(f)(x)$ exist. In particular, if $\mathscr V_\rho(\mathscr Af)$ belongs to some function spaces such as $L^p$ or $L^{p,\infty}$, then the sequence converges almost everywhere without any additional condition. This is why mapping property of $\rho$-variation operator is so interesting in probability, ergodic theory and harmonic analysis. In 1976, L\'{e}pingle \cite{Lep76} showed that the $\rho$-variation operator related to a bounded martingale sequence is a bounded operator on $L^p$ for $1<p<\infty$ and $\rho>2$. These estimates can fail for $\rho\le2$, see \cite{JW04,Q98}. So, we need the oscillation operator to substitute the $2$-variation operator.
\par
For each fixed decreasing sequence $\{t_i\}$ in $\mathbb R^+$, we also define the oscillation operator related to $\mathscr A$
\begin{equation}\nonumber
\mathscr O(\mathscr Af)(x)=\big(\sum_{i=1}^\infty\sup_{t_{i}\le\varepsilon_{i}<\varepsilon_{i+1}\le t_{i+1}}|A_{\varepsilon_{i+1}}f(x)-A_{\varepsilon_{i}}f(x)|^2\big)^{1/2}.
\end{equation}
\par
We also study the $\lambda$-jump operator. For $\lambda>0$, the value of the $\lambda$-jump function for $\mathcal F$ at $x$ is defined by
$$N_\lambda(\mathcal F)(x)=\sup\big\{N\in\Bbb N:\ \exists\ s_1<\varepsilon_1\leq s_2<\varepsilon_2\leq\dotsc\leq s_N<\varepsilon_N\ \text{such that\ } |F_{\varepsilon_k}(x)-F_{s_k}(x)|>\lambda\big\}.$$
Similarly, we define the $\lambda$-jump operator related to $\mathscr A$ as
$N_\lambda(\mathscr Af)(x)=N_\lambda(\{A_tf\}_{t>0})(x)$. Obviously, if $\lim\limits_{t\rightarrow0^+}A_tf(x)$ and $\lim\limits_{t\rightarrow+\infty}A_tf(x)$ exist, then $N_\lambda(\mathscr Af)(x)<\infty$ for any $\lambda>0$. Moreover, for $\lambda>0$ and $\rho\ge1$
\begin{equation}\label{contr ineq}
\lambda[N_\lambda(\mathscr Af)(x)]^{1/\rho}\leq C_\rho\mathscr V_\rho(\mathscr Af)(x).
\end{equation}
\par
Let $\phi\in\mathscr S$ with $\int \phi dx=1$, $\phi_t(x)=\frac1{t^n}\phi(\frac xt)$, denote function family $\{\phi_t\ast f(x)\}_{t>0}$ by $\Phi\star f(x)$. Let $f$ be a tempered distribution, we define maximal function $M_\phi$ by
$$
M_\phi f(x)=\sup_{t>0}|(f\ast\phi_t)(x)|.
$$
\begin{definition}
Let $0<p<\infty$. A distribution $f$ belongs to $H^p$ if the maximal function $M_\phi f$ is in $L^p$.
\end{definition}
\par
The main results of this paper are the following three theorems.
\begin{theorem}\label{SchHp}
For any $\rho>2$, there exists $C_{\rho}>0$ such that
\begin{equation}\label{VISchHp}
\|\mathscr V_\rho(\Phi\star f)\|_{L^p}\le C_{\rho}\|f\|_{H^p},\ \ \frac n{n+1}<p\le1.
\end{equation}
Moreover, for $\frac n{n+1}<p<\infty$, the following conditions are equivalent:
\begin{description}
\item (i) There is a $\phi\in\mathscr S$ with $\int\phi dx\neq0$ so that $M_\phi f\in L^p$.
\item (ii) For any $\rho>2$, there is a $\phi\in\mathscr S$ with $\int\phi dx\neq0$ so that $\phi\ast f$ and $\mathscr V_\rho(\Phi\star f)$ in $L^p$.
\end{description}
\end{theorem}
In the above result, the variation operator is used to characterize $H^p$ spaces,
it is natural to ask if the analogue for the oscillation operator holds.
\begin{theorem}\label{oHp}
For any $\{t_i\}\searrow0$ and $p\in(\frac n{n+1},1]$, there exists a positive constant $C_p$ such that
\begin{equation}\label{OISchHp}
\|\mathscr O(\Phi\star f)\|_{L^p}\le C_p\|f\|_{H^p}.
\end{equation}
Moreover, there exists $\{t_i\}\searrow0$, $\phi\in\mathscr S$ with $\int\phi dx\neq0$ and $f\in \mathscr S$ such that $\mathscr O(\Phi\star f)\in L^p$ and $M_\phi(f)\notin L^p$ for any $p\in(0,1]$.
\end{theorem}
We also apply the above result on variation to provide estimates
for the $\lambda$-jump operator.
\begin{theorem}\label{ljrhp}
If $\rho>2$, then the $\lambda$-jump operator $N_\lambda(\Phi\star f)$ satisfies
\begin{equation}
\|\lambda N_\lambda(\Phi\star f)^{1/\rho}\|_{L^p}\le C_{\rho}\|f\|_{H^p},\ \ \frac n{n+1}<p\le1,
\end{equation}
uniformly in $\lambda>0$. Moreover, there exists $\phi\in\mathscr S$ with $\int\phi dx\neq0$ and $f\in \mathscr S$ such that $\|\lambda N_\lambda(\Phi\star f)^{1/\rho}\|_{L^1}<\infty$ uniformly in $\lambda>0$ and $M_\phi(f)\notin L^1$.
\end{theorem}
The paper is organized as follows. In Section 2, we prove that $\rho$-variation, oscillation and $\lambda$-jump operators related to approximate identities are of strong type $(p,p)$ for $1<p<\infty$ and weak type $(1,1)$. Using the strong $L^p(p>1)$ estimates for the $\rho$-variation operator and the atom decomposition of Hardy space, we show Theorem \ref{SchHp} in Section 3. We consider the estimate for the oscillation operator associated to approximate identities acting on Hardy space and show that the oscillation is not proper to characterize Hardy spaces in Section 4. Finally, we present a jump inequality for the $\lambda$-jump operator on Hardy space and conjecture that its improvement holds, we also illustrate that the $\lambda$-jump operator can not be used to characterize $H^1$.
\section{Operators related to approximate identities on $L^p$}
To study the mapping property of variation, oscillation and $\lambda$-jump operators related to approximate identities on $H^p$, we need strong $L^p$ estimates for $1<p<\infty$. We use the argument in \cite{JSW08} and include all details for completeness although they are trivial. The following lemmas will be used later.
\begin{lemma}(\cite[Lemma 1.3]{JSW08})\label{pdN} Let $\mathscr A=\{A_t\}_{t>0}$ be a family of operators. Then
$$
\lambda\sqrt{N_\lambda(\mathscr A f)}\le C[S_2(\mathscr A f)+\lambda \sqrt{N^d_{\lambda/3}(\mathscr A f)}],
$$
where $N^d_\lambda(\mathscr A f)=N_\lambda(\{A_{2^k}f\})$ and $S_2(\mathscr Af)=(\sum_j\|\mathscr Af\|_{v_2(2^j,2^{j+1}]}^2)^{1/2}$.
\end{lemma}
Let $\sigma$ be a compactly supported finite Borel measure and satisfying
\begin{equation}\label{fts}
|\hat{\sigma}(\xi)|\le C|\xi|^{-b}, \ \ for\ some\ b>0.
\end{equation}
$\sigma_t$ is given by $<\sigma_t,f>=\int f(tx)d\sigma$.
\begin{lemma}(\cite[Theorem 1.1,\ Lemma 6.1]{JSW08})\label{NdS2}
Let $\mathfrak U=\{\mathfrak U_k\}$ where $\mathfrak U_kf=f\ast\sigma_{2^k}$. If $\sigma$ satisfies \eqref{fts}, then
$$
\|\lambda\sqrt{N_\lambda(\mathfrak U f)}\|_{L^p}\le C_p\|f\|_{L^p}
$$
uniformly in $\lambda>0$. Moreover, let $\mathscr A=\{A_t\}_{t>0}$ where $A_tf=f\ast\sigma_t$. If $\sigma$ satisfies \eqref{fts} for some $b>1/2$, then
$$
\|S_2(\mathscr Af)\|_{L^p}\le C_p\|f\|_{L^p}
$$
holds for $\min\{2n/(n+2b-1),(2b+1)/2b\}<p<\max\{2n/(n-2b-1),2b+1\}$.
\end{lemma}
\begin{lemma}(\cite[Theorem 1.1]{JSW08})\label{Ndw11}
If $\phi$ satisfies
\begin{equation}
\int_{\mathbb R^n}|\phi(x+y)-\phi(x)|dx\le C|y|^{-b}
\end{equation}
for some $b>0$, then for any $\alpha>0$ we have
\begin{equation}
|\{x:\lambda\sqrt{N_\lambda^d(\Phi\star f)(x)}>\alpha\}|\le \frac C\alpha\|f\|_{L^1}
\end{equation}
uniformly in $\lambda>0$.
\end{lemma}
We now state a theorem on the $\lambda$-jump operator associated with approximate identities.
\begin{theorem}\label{ljphip}
Let $\phi\in\mathscr S$ with $\int\phi dx\neq0$. For $1<p<\infty$, there exists a positive constant $C_p$ such that
\begin{equation}\label{JISchp}
\|\lambda\sqrt{N_\lambda(\Phi\star f)}\|_{L^p}\le C_p\|f\|_{L^p},
\end{equation}
uniformly in $\lambda>0$. Moreover, for any $\alpha>0$,
\begin{equation}\label{JISch1}
|\{x:\lambda\sqrt{N_\lambda(\Phi\star f)(x)}>\alpha\}|\le \frac C\alpha\|f\|_{L^1},
\end{equation}
uniformly in $\lambda>0$.
\end{theorem}
\begin{proof}
Clearly,
$\hat{\phi}\in\mathscr S$, which means that for any $N\in\mathbb N$ there is a $C_N$ such that
$$
|\hat{\phi}(\xi)|\le C_{N}(1+|\xi|)^{-N}.
$$
Therefore $|\hat{\phi}(\xi)|\le C_{n}|\xi|^{-\frac{n+1}2}$. By Lemma \ref{NdS2}, we have
\begin{equation}
\|\lambda\sqrt{N_\lambda^d(\Phi\star f)}\|_{L^p}\le C_p\|f\|_{L^p}\ \ \text{and}\ \ \|S_2(\Phi\star f)\|_{L^p}\le C_p\|f\|_{L^p},\ \ 1<p<\infty.
\end{equation}
By Lemma \ref{pdN}, we get \eqref{JISchp}. We turn to the proof of \eqref{JISch1}. Lemma \ref{pdN} and Lemma \ref{Ndw11} imply that it suffices to prove
\begin{align}\label{S2w1}
|\{x:S_2(\Phi\star f)(x)>1\}|\le C\|f\|_{L^1}.
\end{align}
We perform the Calder\'{o}n-Zygmund decomposition of $f$ at height $1$ and write $f=g+b$. We just need to establish the following two estimates:
\begin{equation}\label{sg11}
|\{x:S_2(\Phi\star g)(x)>1/2\}|\le C
\|f\|_{L^1},
\end{equation}
and
\begin{equation}\label{sb11}
|\{x:S_2(\Phi\star b)(x)>1/2\}|\le C
\|f\|_{L^1}.
\end{equation}
As usual the known $L^p$ bounds for $S_2$ allows us to obtain (\ref{sg11}):
\begin{align*}
|\{x:S_2(\Phi\star g)(x)>1/2\}|&\le C\int_{\mathbb R^n}|S_2(\Phi\star g)(x)|^2dx\\
&\le C \int_{\mathbb R^n}|g(x)|^2dx\le C\|f\|_{L^1}.
\end{align*}
\par
To show \eqref{sb11}, we write $b=\sum_jb_j$ precisely, where each $b_j$ is supported in a dyadic cube $Q_j$. We denote by $l(Q_j)$ the side length of $Q_j$.
Let $\tilde{Q}_j$ be the cube with sides parallel to the axes having the same center as $Q_j$ and having side length $8l(Q_j)$, write $\tilde{Q}=\bigcup \tilde{Q}_j$. Obviously,
\begin{align*}
|\tilde{Q}|\le\sum_j |\tilde{Q}_j|\le C\sum_j |Q_j|\le C\|f\|_{L^1}.
\end{align*}
We still need to prove that
\begin{align*}\label{wsbc11}
|\{x\in\tilde{Q}^c:S_2(\Phi\star b)(x)>1/2\}|\le C\|f\|_{L^1}.
\end{align*}
Let $x_j$ be the center of $Q_j$. Using the moment condition of $b_j$ and H\"{o}lder's inequality, we get
\begin{align*}
\|\Phi\star b(x)\|_{v_2(2^i,2^{i+1}]}\le\sum_j\int_{Q_j}|b_j(y)|\|\Phi(x-y)-\Phi(x-x_j)\|_{v_2(2^i,2^{i+1}]}dy.
\end{align*}
For $x\in \tilde{Q}^c$ and $x_j,y\in Q_j$, it is clear that $|x-x_j-\theta (y-x_j)|\sim |x-x_j|$ for any $\theta\in [0,1]$. Then,
\begin{align*}
\|\Phi(x-y)-\Phi(x-x_j)\|_{v_2(2^i,2^{i+1}]}
&\le \|\Phi(x-y)-\Phi(x-x_j)\|_{v_1(2^i,2^{i+1}]}\\
&\le \int_{2^i}^{2^{i+1}}|\frac d{dt}[\phi_t(x-y)-\phi_t(x-x_j)]|dt\\
&\le C|y-x_j|\int_{2^i}^{2^{i+1}}\frac{1}{t^{n+2}}\big(1+\frac{|x-x_j|}t)^{-(n+2)}dt,
\end{align*}
where we use the following well-known fact
\begin{equation}\label{vr1i}
\|\mathfrak a\|_{v_\rho}\le \|\mathfrak a\|_{v_1}\le \int_0^\infty|\mathfrak a'(t)|dt,
\end{equation}
see (39) in \cite{JSW08}. Consequently,
\begin{equation}\label{eoV2i}
\|\Phi(x-y)-\Phi(x-x_j)\|_{v_2(2^i,2^{i+1}]}\le Cl(Q_j)\min\{2^{i+1}|x-x_j|^{-n-2}, 2^{-i(n+1)}\}.
\end{equation}
Finally, by \eqref{eoV2i} and H\"{o}lder's inequality,
\begin{align*}
|\{x\in\tilde{Q}^c:S_2(\Phi\star b)(x)>1/2\}|&\le C
\int_{\tilde{Q}^c}(\sum_{i\in\mathbb Z}\|\Phi\star b(x)\|_{v_2(2^i,2^{i+1}]}^2)^{\frac12}dx\\
&\le C\int_{\tilde{Q}^c}\sum_j\int_{Q_j}|b_j(y)|[\sum_{i\in\mathbb Z}\|\Phi(x-y)-\Phi(x-x_j)\|_{v_2(2^i,2^{i+1}]}^2]^{\frac12}dydx\\
&\le C\sum_jl(Q_j)\int_{Q_j}|b_j(y)|dy\int_{\tilde{Q_j}^c}\frac{dx}{|x-x_j|^{n+1}}\\
&\le C\|f\|_{L^1}.
\end{align*}
This completes the proof of Theorem \ref{ljphip}.
\end{proof}
To obtain the variational inequality, we present a lemma reducing variational inequalities to jump inequalities, which is a generalization of Bourgain's argument in \cite{Bou89}.
\begin{lemma}(\cite[Lemma 2.1]{JSW08})\label{jcv}
Suppose that $p_0<q<p_1$ and that for $p_0<p<p_1$ the inequality
$$
\sup_{\lambda>0}\|\lambda[N_\lambda(\mathcal Tf)]^{1/q}\|_{L^p}\le C\|f\|_{L^p}
$$
holds for all $f$ in $L^p$. Then we have for $q<\rho$,
$$
\|\mathscr V_\rho(\mathcal Tf)\|_{L^p}\le C(p,\rho)\|f\|_{L^p}
$$
for $f\in L^p$, $p_0<p<p_1$.
\end{lemma}
As a result of above lemma, the following variational inequality holds:
\begin{theorem}\label{SchLp}
For any $\rho>2$ and $1<p<\infty$, there exists $C_{p,\rho}>0$ such that
\begin{equation}\label{VISchp}
\|\mathscr V_\rho(\Phi\star f)\|_{L^p}\le C_{p,\rho}\|f\|_{L^p}.
\end{equation}
Moreover, for any $\alpha>0$,
\begin{equation}\label{VISch1}
|\{x:\mathscr V_\rho(\Phi\star f)(x)>\alpha\}|\le \frac C{\alpha}\|f\|_{L^1}.
\end{equation}
\end{theorem}
\begin{proof}
Clearly, \eqref{JISchp} and Lemma \ref{jcv} imply \eqref{VISchp}. \eqref{VISch1} can be proved as \eqref{S2w1}.
\end{proof}
\begin{theorem}\label{OSchLp}
For any $\{t_i\}\searrow0$ and $1<p<\infty$, there exists $C_p>0$ such that
\begin{equation}\label{OISchp}
\|\mathscr O(\Phi\star f)\|_{L^p}\le C_{p}\|f\|_{L^p}.
\end{equation}
Moreover, for any $\alpha>0$,
\begin{equation}\label{OISch1}
|\{x:\mathscr O(\Phi\star f)(x)>\alpha\}|\le \frac C{\alpha}\|f\|_{L^1}.
\end{equation}
\end{theorem}
\begin{proof} Let $k_i$ be the smallest integer such that $2^{k_i}$ greater than or equal to $t_i$. The long oscillation operator is given by
\begin{equation}\nonumber
\mathscr O_L(\Phi\star f)(x)=\big(\sum_{i}\sup_{k_{i+1}\le l\le m\le k_i}|\phi_{2^l}\ast f(x)-\phi_{2^m}\ast f(x)|^2\big)^{1/2}.
\end{equation}
The oscillation inequality follows from the pointwise estimate
\begin{align*}
\mathscr O(\Phi\star f)(x)\le C[S_2(\Phi\star f)(x)+\mathscr O_L(\Phi\star f)(x)].
\end{align*}
Strong $L^p$ estimates and weak $(1,1)$ estimate for $S_2(\Phi\star f)$ have been established as above. For the long oscillation $\mathscr O_L(\Phi\star f)$, we borrow some notations and results from \cite[pp.6724]{JSW08}. For $j\in\mathbb Z$ and $\beta=(m_1,\cdots,m_n)\in\mathbb Z^n$, we denote the dyadic cube $\prod_{k=1}^n(m_k2^j,(m_k+1)2^j]$ in $\mathbb R^n$ by $Q_\beta^j$, and the set of all dyadic cubes with side length $2^j$ by $\mathcal D_j$. The conditional expectation of a local integrable $f$ with respect to $\mathcal D_j$ is given by
$$
\mathbb E_jf(x)=\sum_{Q\in \mathcal D_j}\frac1{|Q|}\int_{Q}f(y)dy\cdot\chi_{Q}(x)
$$
for all $j\in\mathbb Z$.
Note that $\mathscr O_L$ satisfies
\begin{equation}\nonumber
\mathscr O_L(\Phi\star f)\le \mathscr O_L(\mathscr D f)+\mathscr O_L(\mathscr Ef),
\end{equation}
where
$$
\mathscr Df=\{\phi_{2^k}\ast f-\mathbb E_kf\}_k \quad \text{and}\quad \mathscr Ef=\{\mathbb E_kf\}_k.
$$
Following inequalities are oscillation inequalities for dyadic martingales (see \cite{JKRW98}),
$$
|\{x:\mathscr O_L(\mathscr{E}f)(x)>\alpha\}|\le \frac C{\alpha}\|f\|_{L^1}\ \text{and}\ \|\mathscr O_L(\mathscr{E}f)\|_{L^p}\leq C_p\|f\|_{L^p},\ 1<p<\infty.
$$
Next, observe that
\begin{equation}\nonumber
\mathscr O_L(\mathscr D f)\le C\big(\sum_{k\in\mathbb Z}|\phi_k\ast f-\mathbb E_kf|^2\big)^{1/2}:=\mathcal Sf.
\end{equation}
Jones {\it et al} \cite{JSW08} have established the following weak-type $(1,1)$ bound and $L^p$ bounds for $\mathcal S$,
\begin{equation}\label{S1p}
|\{x:\mathcal Sf(x)>\alpha\}|\le \frac C{\alpha}\|f\|_{L^1}\ \text{and}\ \|\mathcal Sf\|_{L^p}\le C_p\|f\|_{L^p},\ \ 1<p<\infty,
\end{equation}
see also \cite{DHL16}.
This completes the proof of Theorem \ref{OSchLp}.
\end{proof}
\begin{remark} With $\phi$ a radial function,
Campbell {\it et al} \cite{CJRW2000,CJRW2002} obtained Theorem \ref{SchLp} and Theorem \ref{OSchLp} by rotation method.
\end{remark}
\section{Variational characterization of $H^p$}
In what follows we shall use the well-known atom decomposition of $H^p$. So, we present the definition of $(p,q)$-atom.
\begin{definition}
Let $0<p\le 1\le q\le\infty$, $p\neq q$. A function $a(x)\in L^q$ is called a $(p,q)$-atom with the center at $x_0$, if it satisfies the following conditions:
\begin{description}
\item (i) Supp $a\subset B(x_0,r)$;
\item (ii) $\|a\|_{L^q}\le |B(x_0,r)|^{\frac1q-\frac1p}$;
\item (iii) $\int a(x)dx=0$.
\end{description}
\end{definition}
\begin{lemma}(\cite{L78})
Let $0<p\le 1$. Given a distribution $f\in H^p$, there exists a sequence of $(p,q)$-atoms with $1\le q\le\infty$ and $q\neq p$, $\{a_k\}$, and a sequence of scalars $\{\lambda_k\}$ such that
$$
f=\sum_{k}\lambda_ka_k\ \ \text{in}\ \ H^p.
$$
\end{lemma}
\textit{Proof of Theorem \ref{SchHp}}.\ \ In proving Theorem \ref{SchHp} we consider first the inequality \eqref{SchHp}.
By \cite[Theorem 1.1]{YZ08}, we just need to prove that there exists a positive constant $C$ such that
$\|\mathscr V_\rho(\Phi\star a)\|_{L^p}\le C$ for any $(p,2)$ atom $a$ and $p\in(\frac{n}{n+1},1]$.
\par
Suppose $a$ is supported in a cube $Q$, $x_0$ is the center of $Q$, write $\tilde{Q}=8Q$. By H\"{o}lder's inequality and Theorem \ref{SchLp},
$$
\int_{\tilde{Q}}\mathscr V_\rho(\Phi\star a)^p(x)dx\le |\tilde{Q}|^{1-\frac p2}\|\mathscr V_\rho(\Phi\star a)\|_{L^2}^p\le C|Q|^{1-\frac p2}\|a\|^p_{L^2}\le C.
$$
\par
To deal with $x\in (\tilde{Q})^c$ one uses the cancelation condition of $a$ and Minkowski's inequality,
\begin{align*}
&\mathscr V_\rho(\Phi\star a)(x)\\
&=\sup_{\{\varepsilon_k\}\searrow0}\bigg(\sum_{k}\bigg|\int_{\mathbb R^n}\bigg\{\big[\phi_{\varepsilon_k}(x-y)-\phi_{\varepsilon_{k+1}}(x-y)\big]-\big[\phi_{\varepsilon_k}(x-x_0)-\phi_{\varepsilon_{k+1}}(x-x_0)\big]\bigg\}a(y)dy\bigg|^\rho\bigg)^{\frac1\rho}\\
&\le\int_{Q}|a(y)| \sup_{\{\varepsilon_k\}\searrow0}\bigg(\sum_{k}\bigg|\big[\phi_{\varepsilon_k}(x-y)-\phi_{\varepsilon_k}(x-x_0)\big]-\big[\phi_{\varepsilon_{k+1}}(x-y)-\phi_{\varepsilon_{k+1}}(x-x_0)\big]\bigg|^\rho\bigg)^{\frac1\rho}dy\\
&\le\int_{Q}|a(y)|\|\Phi(x-y)-\Phi(x-x_0)\|_{v_\rho}dy.
\end{align*}
For $x\in (\tilde{Q})^c$, $y\in Q$ and $\theta\in(0,1)$, we have $|x-x_0+\theta (y-x_0)|\sim|x-x_0|$. Note that $\phi(\frac{x-y}t)$ is a smooth function of $t$ on $(0,\infty)$ for any fixed $x,y$. Applying \eqref{vr1i} and the mean value theorem, we estimate
\begin{align*}
\nonumber\|\Phi(x-y)-\Phi(x-x_0)\|_{v_\rho}\le&\|\Phi(x-y)-\Phi(x-x_0)\|_{v_1}\\
\le& C|y-x_0|\int_0^\infty\frac1{t^{n+2}}\big(1+\frac{|x-x_0|}t)^{-(n+2)}dt\\
\nonumber\le& C\frac{|y-x_0|}{|x-x_0|^{n+1}}\int_0^\infty\frac{t^n}{(1+t)^{n+2}}dt\\
\nonumber\le&C\frac{|y-x_0|}{|x-x_0|^{n+1}}.
\end{align*}
Hence, we obtain the desired bound
\begin{align*}
\int_{(\tilde Q)^c}\mathscr V_\rho(\Phi\star a)^p(x)dx\le C\int_{|x-x_0|\ge 8l(Q)}\frac{l^p(Q)}{|x-x_0|^{(n+1)p}}dx\big(\int_Q|a(y)|dy\big)^p\le C,
\end{align*}
finishing the proof of \eqref{SchHp}.
\par
We now turn to the equivalence of two conditions in Theorem \ref{SchHp}. Note that $M_\phi f(x)\le |\phi\ast f(x)|+\mathscr V_\rho(\Phi\star f)(x)$. Thus, $(ii)$ implies $(i)$. Conversely, $M_\phi f\in L^p$ means $f\in H^p$ for $\frac n{n+1}<p\le 1$ and $f\in L^p$ for $1<p<\infty$. Also, the pointwise estimate $\phi\ast f(x)\le M_\phi f(x)$ shows $\phi\ast f\in L^p$ for $\frac n{n+1}<p<\infty$. From \eqref{SchHp} and \eqref{VISchp}, we obtain $\mathscr V_\rho(\Phi\star f)\in L^p$, finishing the proof of Theorem \ref{SchHp}.
\section{Oscillation on $H^p$}
Estimates for the oscillation operator acting on $H^p$ are trivial. It suffices to show that there exists a positive constant $C$ such that
$\|\mathscr O(\Phi\star a)\|_{L^p}\le C$ for any $(p,2)$ atom $a$ and $p\in(\frac{n}{n+1},1]$. By Theorem \ref{OSchLp} and H\"{o}lder's inequality, we have
$$
\int_{\tilde{Q}}\mathscr O(\Phi\star a)^p(x)dx\le C|Q|^{1-\frac p2}\|\mathscr O(\Phi\star a)\|_{L^2}^p\le C|Q|^{1-\frac p2}\|a\|_{L^2}^p\le C.
$$
Next we consider the integral of $\mathscr O(\Phi\star a)^p$ on $(\tilde{Q})^c$. Note that $\mathscr O(\Phi\star a)(x)\le C\mathscr V_2(\Phi\star a)(x)$. Applying the same argument, we obtain
\begin{equation}\nonumber
\int_{(\tilde{Q})^c}\mathscr O(\Phi\star a)^p(x)dx\le C\int_{(\tilde{Q})^c}\mathscr V_2(\Phi\star a)^p(x)dx\le C\int_{|x-x_0|\ge 8l(Q)}\frac{l^p(Q)}{|x-x_0|^{(n+1)p}}\big(\int_Q|a(y)|dy\big)^pdx \le C.
\end{equation}
Hence, we get $\|\mathscr O(\Phi\star f)\|_{L^p}\le C\|f\|_{H^p}$.
\par
We now turn to the negative result in Theorem \ref{oHp}. Let $\phi(x)=f(x)=e^{-x^2}$ for $x\in\mathbb R$, $\phi_t\ast f(x)=\frac1{\sqrt{t^2+1}}e^{-\frac{x^2}{1+t^2}}$ , $\phi\ast f(x)=\frac{\sqrt2}{2}e^{-\frac{x^2}{2}}\in L^p$ for any $p\in(0,1]$. Define $F(s,x)=se^{\frac{x^2}{s^2}}$ for $(s,x)\in(1,+\infty)\times \mathbb R^+$. Clearly, for fixed $x\in[0,\frac{\sqrt{2}}2]$, $F(s,x)$ is increasing respect to $s$; for fixed $x\in(\frac{\sqrt{2}}2,+\infty)$, $F(s,x)$ is decreasing on $(1,\sqrt2x]$ and increasing on $(\sqrt2x,+\infty)$. Note that $M_\phi(f)$ is even. Therefore,
\begin{equation}\nonumber
M_\phi(f)(x)=
\begin{cases}
e^{-x^2},&\mbox{$x\in[-\frac{\sqrt{2}}2,\frac{\sqrt{2}}2]$,}\\
\frac1{\sqrt{2e}}\frac1x, &\mbox{ $x\in(-\infty,-\frac{\sqrt{2}}2)\bigcup(\frac{\sqrt{2}}2,+\infty)$.}
\end{cases}
\end{equation}
Obviously, $M_\phi(f)\notin L^p$ for any $p\in(0,1]$.
\par
For the oscillation of $\Phi\star f$, we take $t_n=\frac1n$ and use the following pointwise estimate
\begin{align*}
\mathscr O(\Phi\star f)(x)&\le \sum_n\sup_{\frac1{n+1}\le\varepsilon_{n+1}<\varepsilon_n\le \frac1n}|\phi_{\varepsilon_{n+1}}\ast f(x)-\phi_{\varepsilon_{n}}\ast f(x)|.
\end{align*}
For fixed $x\in[0,\frac{\sqrt2}2]$, $\frac 1se^{-\frac{x^2}{s^2}}$ is decreasing on $(1,+\infty)$. So,
\begin{align*}
\mathscr O(\Phi\star f)(x)&\le \sum_n|\phi_{\frac1{n+1}}\ast f(x)-\phi_{{\frac1n}}\ast f(x)|\\
&=\phi_0\ast f(x)-\phi_1\ast f(x)=e^{-x^2}-\frac{\sqrt2}2e^{-\frac{x^2}2}.
\end{align*}
For fixed $x\in(\frac{\sqrt2}2,1)$, $\frac1se^{-\frac{x^2}{s^2}}$ is increasing on $(1,\sqrt2x]$ and decreasing on $(\sqrt2x,\sqrt2]$. We estimate
\begin{align*}
\mathscr O(\Phi\star f)(x)&\le \sum_{\frac1{n+1}\le \sqrt{2x^2-1}}\sup_{\frac1{n+1}\le\varepsilon_{n+1}<\varepsilon_n\le \frac1n}|\phi_{\varepsilon_{n+1}}\ast f(x)-\phi_{\varepsilon_{n}}\ast f(x)|\\
&+\sum_{\frac1n> \sqrt{2x^2-1}}\sup_{\frac1{n+1}\le\varepsilon_{n+1}<\varepsilon_n\le \frac1n}|\phi_{\varepsilon_{n+1}}\ast f(x)-\phi_{\varepsilon_{n}}\ast f(x)|\\
&\le \sqrt{\frac2e}\frac1x-e^{-x^2}-\frac{\sqrt2}2e^{-\frac{x^2}2}.
\end{align*}
For fixed $x\in[1,+\infty)$, $\frac1se^{-\frac{x^2}{s^2}}$ is increasing on $(1,\sqrt2]$. Hence
\begin{align*}
\mathscr O(\Phi\star f)(x)&\le \sum_n|\phi_{\frac1{n+1}}\ast f(x)-\phi_{{\frac1n}}\ast f(x)|\\
&=\phi_1\ast f(x)-\phi_0\ast f(x)=\frac{\sqrt2}2e^{-\frac{x^2}2}-e^{-x^2}.
\end{align*}
Obviously,$\mathscr O(\Phi\star f)$ is even and $\mathscr O(\Phi\star f)\in L^p$ for any $p\in(0,1]$, completing the proof of Theorem \ref{oHp}.
\section{$\lambda$-jump on $H^p$}
\textit{Proof of Theorem \ref{ljrhp}}.\ \
It is clear that for $\lambda>0$ and $\rho\ge1$
\begin{equation}\nonumber
\lambda[N_\lambda(\Phi\star f)(x)]^{1/\rho}\leq C_\rho\mathscr V_\rho(\Phi\star f)(x).
\end{equation}
Consequently, we have
\begin{equation}\nonumber
\|\lambda[N_\lambda(\Phi\star f)]^{1/\rho}\|_{L^p}\leq C_\rho\|\mathscr V_\rho(\Phi\star f)\|_{L^p}\le C_\rho\|f\|_{H^p},\ \ p\in(\frac{n}{n+1},1],
\end{equation}
uniformly in $\lambda>0$.
\par
For counterexample, we take $\phi(x)=f(x)=e^{-x^2}$. Obviously, $f\notin H^p$ and $M_\phi(f)\notin L^p$ for any $p\in(0,1]$.
\par
When $\lambda\ge1$, we have $N_\lambda(\Phi\star f)(x)\equiv0$ and $\|\lambda[N_\lambda(\Phi\star f)]^{1/\rho}\|_{L^p}<\infty$ uniformly in $\lambda>0$ for any $p\in(0,1]$.
\par
When
$\frac1{\sqrt e}\le\lambda<1$, we get $N_\lambda(\Phi\star f)(x)\le e^{-x^2}\lambda^{-1}$ for $|x|\le \sqrt{-\ln\lambda}$ and $N_\lambda(\Phi\star f)(x)=0$ for $\sqrt{-\ln\lambda}<x$. Hence, $\|\lambda[N_\lambda(\Phi\star f)]^{1/\rho}\|_{L^p}<\infty$ uniformly in $\lambda>0$ for any $p\in(0,1]$.
\par
When $0<\lambda<\frac1{\sqrt e}$, we obtain
\begin{equation}\nonumber
N_\lambda(\Phi\star f)(x)\le C
\begin{cases}
e^{-x^2}\lambda^{-1},&\mbox{$x\in[-\frac{\sqrt{2}}2,\frac{\sqrt{2}}2]$,}\\
\sqrt{\frac 2e}\frac1{\lambda |x|}- e^{-x^2}\lambda^{-1}, &\mbox{ $x\in(-\frac1{\lambda\sqrt{2e}},-\frac{\sqrt{2}}2)\bigcup(\frac{\sqrt{2}}2,\frac1{\lambda\sqrt{2e}})$,}\\
0,&\mbox{ $x\in(-\infty,-\frac1{\lambda\sqrt{2e}})\bigcup(\frac1{\lambda\sqrt{2e}},+\infty)$.}
\end{cases}
\end{equation}
One can establish the following $L^p$ bounds:
\begin{align*}
\|\lambda[N_\lambda(\Phi\star f)]^{1/\rho}\|_p^p&\le C \lambda^{p(1-1/\rho)}\int_0^{\frac{\sqrt2}2}e^{-px^2/\rho}dx+C\lambda^{p(1-1/\rho)}\int_{\frac{\sqrt2}2}^{\frac1{\lambda\sqrt{2e}}}x^{-\frac p{\rho}}dx\\
&\le C+C\lambda^{p-1}.
\end{align*}
Consequently, $\|\lambda[N_\lambda(\Phi\star f)]^{1/\rho}\|_{L^1}<\infty$ uniformly in $\lambda>0$ for $\rho\in(1,\infty)$.\qed
\par
Theorems 1.4 suggests the following improvement:
\begin{conjecture}
For $\frac n{n+1}<p\le 1$, there exists $C_{p}>0$ such that
\begin{equation}\nonumber
\|\lambda\sqrt{N_\lambda(\Phi\star f)}\|_{L^p}\le C_{p}\|f\|_{H^p},
\end{equation}
uniformly in $\lambda>0$.
\end{conjecture}
In the case of analogous for variation, oscillation and $\lambda$-jump operators, we know the conjecture above is possible. However, our current techniques do not allow us to prove it.
| c64dc9a85a7c5bc9146ccba6d3f2620ca0393336 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{int}
The proper accounting for spectral lines is a keypoint of hot-plasma modelling, for radiative transfer (mostly through Rosseland mean opacities), and more generally for absorption and emission spectroscopy, where they are used as a diagnostic tool to infer the temperature (from ratios of line intensities) and density (from line widths) of plasmas. The applications range from inertial confinement fusion, magnetic confinement fusion (for the radiative losses in the W-coated tiles of divertor) and astrophysics. Heating by fusion reactions deep within stellar cores produces thermal X-ray radiation and models of stellar structure and evolution are very sensitive to radiative transfer and opacity. Matter at extreme densities and temperatures gives us a chance to study important and exotic physical processes: plasmon neutrinos as a test of electro-weak theory at low-energies, turbulent energy transport in high-gravity environments, dark matter in the form of axions \cite{schlattl99} and/or WIMPS, \emph{etc.} \cite{vinyoles15}. Exploration of extreme physics in star interiors is possible thanks to asteroseismology. The Cepheid stars (named after $\delta$ Cephei) are known to be used to estimate the stellar distances. They have been better understood after a revision by a factor two of the opacity for species of atomic number $Z>2$ \cite{simon82} confirmed by the first measurements of absorption coefficients \cite{dasilva92}. The $\beta$ Cephei-type stars (which should not be confused with Cepheid variables), also known as $\beta$ Cepheids and named after $\beta$ Cephei (or Alfirk) are now of primary interest due to the correlation between the frequencies of the excited modes and the mass of these stars for a given age. Such stars are now used to estimate the age of several young open clusters \cite{balona97}. In $\beta$ Cephei-type stars, the iron-group (chromium, iron and nickel) opacity peak (when plotting Rosseland opacity with respect to temperature) excites acoustic modes through the ``kappa-mechanism'' \cite{eddington26a,eddington26b}.
With the recent revision of solar chemical abundances \cite{asplund09}, the standard stellar model fails to reproduce the helioseismic results. Among the possible explanations, an increase of 5 to 20 \% of the opacity in the solar radiative zone would be sufficient to reconcile the modelling and the asteroseismic observations. Several absorption spectroscopy measurements were performed recently on the Z machine at Sandia National Laboratory by Bailey {\emph et al.} \cite{bailey09} at conditions similar to those at the boundary between the convective and radiative zones of the Sun. In that region, iron is responsible for 25 \% of the total opacity. One of those experiments, published in 2015 \cite{bailey15} revealed that, in the spectral range between 7 and 12.7 \AA, the opacity inferred from the measurements was higher than the opacity predicted by all the best codes in the world by 30 to 400 \%.
Highly charged iron produces some of the brightest X-ray emission lines from hot astrophysical objects, including galaxy clusters and stellar coronae, and dominates the emission of the Sun at wavelengths near 15 \AA. The Fe XVII spectrum is, however, poorly fitted by the best astrophysical models. A particular problem has been that the strongest Fe XVII line is weaker than predicted. This has affected the interpretation of observations by the Chandra and XMM-Newton orbiting X-ray missions \cite{paerels03}, fuelling a continuing controversy over whether this discrepancy is caused by an incomplete modelling of the plasma environment in these objects or by shortcomings in the treatment of the underlying atomic physics.
Of course, iron is not the only element we are interested in. To explore the history of star formation in our galaxy, cosmochronology estimates the ages of the coolest white dwarf stars. Measuring relative line shapes of H$_{\beta}$, H$_{\gamma}$, and H$_{\delta}$ enables one to determine white dwarf photospheric and atmospheric conditions \cite{wiese72}.
In Sec. \ref{rad}, the different processes contributing to radiative opacity are introduced, and the detailed opacity code SCO-RCG is presented. Throughout the last years, the code was used to interpret several laser or Z-pinch experiments, as shown in Sec. \ref{exp}. In Sec. \ref{env}, we discuss the role of opacities for the modelling of stellar envelopes, and present the kappa mechanism, responsible for the pulsations of particular stars. In Sec. \ref{bou}, the importance of the opacity at the boundary between convective and radiative zones of the Sun (also named ``tachocline'' \cite{spiegel92}) is outlined, and theoretical efforts dedicated to understanding the experiment performed recently on the Z machine at Sandia National Laboratory are detailed in Sec. \ref{bai}. The enigma of the ratio between two important iron lines, characteristic of the corona of the Sun and of other stars such as $\alpha$ Aurigae (Capella) or $\beta$ Canis Majoris (Procyon), is briefly mentioned in Sec. \ref{3c3}, and the need for Stark-broadening modelling of hydrogen Balmer lines is recalled in Sec. \ref{whi}.
\subsection{Radiative opacity}\label{rad}
The spectral opacity is the photo-absorption cross-section per mass unit, usually expressed in cm$^2$/g. The total frequency-dependent opacity can be calculated as the sum of the contributions of different processes: photo-excitation $\kappa_{\mathrm{bb}}$, photo-ionization $\kappa_{\mathrm{bf}}$, inverse Bremsstrahlung $\kappa_{\mathrm{ff}}$ and photon scattering $\kappa_{\mathrm{scat}}$. It is then given by the following expression \cite{mayer47}:
\begin{equation}
\kappa'(h\nu)=\kappa(h\nu)\left(1-e^{-\left(h\nu/k_BT\right)}\right)+\kappa_{\mathrm{scat}}(h\nu),
\end{equation}
\noindent where $h$ the Planck constant, $k_B$ the Boltzmann constant, $T$ the temperature, $\nu$ the photon frequency and
\begin{equation}
\kappa(h\nu)=\kappa_{\mathrm{bb}}(h\nu)+\kappa_{\mathrm{bf}}(h\nu)+\kappa_{\mathrm{ff}}(h\nu).
\end{equation}
\noindent Photo-excitation and de-excitation can be described as
\begin{equation}
X_i^{q+}+h\nu \leftrightharpoons X_j^{q+},
\end{equation}
\noindent where $X_i^{q+}$ is an ion with charge $q$ in an excitation state $i$. The signature of the emitted or absorbed photon $h\nu$ is a spectral line. The relevant atomic parameter for the direct and inverse processes is the oscillator strength ($f$), and one has \cite{pain17b}
\begin{equation}
\kappa_{\mathrm{bb}}(h\nu)=\frac{1}{4\pi\epsilon_0}\frac{\mathcal{N}_A}{A}\frac{\pi e^2h}{mc}\sum_{i\rightarrow j}\mathcal{P}_if_{i\rightarrow j}\Psi_{i\rightarrow j}(h\nu),
\end{equation}
\noindent where $\mathcal{P}_i$ is the population of initial level $i$, $f_{i\rightarrow j}$ the oscillator strength and $\Psi_{i\rightarrow j}$ the profile of the spectral line corresponding to the transition $i\rightarrow j$, accounting for broadening mechanisms (Doppler, Stark, ...). $\epsilon_0$ is the dielectric constant, $\mathcal{N}_A$ the Avogadro number, $e$ and $m$ represent respectively the electron charge and mass, $c$ the speed of light and $A$ the atomic mass of the considered element. Photo-ionization is a process that occurs when a bound electron $e^-$ is ejected after absorption of a photon
\begin{equation}
X_i^{q+}+h\nu \leftrightharpoons X_j^{(q+1)+}+e^-.
\end{equation}
\noindent The inverse process is radiative recombination (RR), occuring when a free electron recombines with an ion along with emission of a photon. This can occur via an intermediate excited state as
\begin{equation}
e^-+X_i^{q+}\leftrightharpoons \left(X_j^{(q-1)+*}\right)\leftrightharpoons \left\{\begin{array}{ll}
e^-+X_k^{q+} & AI\\
X_m^{(q-1)+}+h\nu & DR
\end{array}\right..
\end{equation}
\noindent A colliding electron excites the target and attaches to form the short-lived autoionizing state (*), which may be doubly excited, but not necessarily: it is only required for it to lie above the first ionization limit, and to be able to decay radiatively to bound states. The intermediate state decays either by autoionization (AI) where the electron becomes free and the target drops to ground state, or by dielectronic recombination (DR) where the electron gets bound by emitting a photon. Photo-ionization resonances can be seen in absorption spectra and DR resonances in emission spectra. The opacity involves also two other processes, inverse Bremsstrahlung or free-free absorption, and photon-electron scattering. Bremsstrahlung refers to the radiation emitted by an electron slowing down in the electromagnetic field of an ion. The inverse process occurs when a free electron and an ion absorb a photon
\begin{equation}
h\nu+\left[X_i^{q+}+e^-(\epsilon)\right]\rightarrow X_j^{q+}+e^-\left(\epsilon'\right),
\end{equation}
\noindent $\epsilon$ and $\epsilon'$ being the energies of the free electron before and after the photo-absorption. Calculations of the free-free cross-section involve quantities related to the elastic-scattering matrix elements for electron-impact excitation of ions. The detailed (fine-structure) opacity code SCO-RCG \cite{porcherot11,pain15a,pain15b} enables one to compute precise opacities for the calculation of accurate Rosseland means (see Sec. \ref{ross}). The (super-)configurations are generated by the SCO code \cite{blenski00} on the basis of a statistical fluctuation theory (see the details below) and a self-consistent computation of atomic structure is performed for all the configurations. In such a way, each configuration has its own set of wavefunctions. The latests are determined in a single-configuration approximation. One peculiarity of the code is that it does not rely on the ``isolated atom'' picture, but on a realistic atom-in-plasma modelling (equation of state). Relativistic effects are taken into account in the Pauli approximation. The DLA (Detailed Line Accounting) part of the spectrum is performed using an adapted version of the RCG routine from Cowan's suite of atomic-structure and spectra codes \cite{cowan81}. The RCG source code was used for decades by spectroscopists, it has many available options and is well documented. In SCO-RCG, criteria are defined to select transition arrays that can be treated line-by-line. The data required for the calculation of the detailed transition arrays (Slater, spin-orbit and dipolar integrals) are obtained from SCO, providing in this way a consistent description of the plasma screening effects on the wavefunctions. Then, the level energies and the lines are calculated by RCG. The computation starts with an average-atom calculation in local thermodynamic equilibrium (LTE), which provides the average populations of the subshells. We then build a list of super-configurations of the kind
\begin{equation}\label{sc}
(1s)^{p_1}(2s)^{p_2}...\left(n_{k-1}\ell_{k-1}\right)^{p_{k-1}}\left(\prod_{i=k}^Nn_i\ell_i\right)^{p_k},
\end{equation}
\noindent each super-configuration containing $k$ supershells and $N$ subshells, where $n_N\ell_N$ is the last highest-energy subshell found by the average-atom calculation at the given density and temperature. A super-configuration is made of supershells (a supershell being a group of subshells) populated by electrons. In Eq. (\ref{sc}), the first $(k-1)$ supershells are ordinary subshells and the last $(N-k+1)$ subshells are gathered in the same supershell. We use the LTE fluctuation theory \cite{perrot00} around the average-atom non-integer populations in order to determine the range of variation of the populations $p_k$, $k$=1, $N$ and therefore the possible list of configurations (if $p_N$ is equal to zero) or super-configurations (if $p_N$ is strictly positive). The super-configurations are then sorted according to their Boltzmann weights, estimated using the average-atom wavefunctions. We only keep the $N$ (super-)configurations having the highests weights. The strength of this approach is that it enables one to take into account many highly excited states and satellite lines. The populations of those states may be small, but their number is so huge that they can play a significant role in the opacity. The orbitals in the Rydberg supershell are chosen in a way so that they weakly interact with inner orbitals. A DLA calculation is performed if possible and necessary for all the transition arrays starting from the configuration; DLA computations are carried out only for pairs of configurations giving rise to less than 800,000 lines. In other cases, transition arrays are represented by Gaussian profiles in the UTA (Unresolved Transition Array \cite{bauche79,bauche82}) or SOSA (Spin-Orbit Split Array \cite{bauche85}) formalisms. If the Rydberg supershell contains at least one electron, then transitions starting from the super-configuration are taken into account by the Super Transition Array (STA) model \cite{bar89}. The amount of detailed calculations performed in SCO-RCG is now largely dominant.
The PRTA (Partially Resolved Transition Array) model was recently implemented. It enables one to replace many statistical transition arrays by small-scale DLA calculations. We have extended this approach to the STA formalism \cite{bar89}, consisting in omitting the Rydberg supershell in the computation, and adding its contribution to the widths of all lines. The contribution of the Rydberg supershell is included as a Gaussian ``dressing function'' \cite{pain15a}. We have also the possibility to replace this dressing function by a coarse grained configurationally resolved profile, following the CRSTA (Configurationally Resolved Transition Array) method \cite{kurzweil16}.
\section{Interpretation of experiments}\label{exp}
In the experiment by Davidson \emph{et al.} \cite{davidson88}, the laser beams were produced by a neodymium-doped glass laser system, HELEN, delivering about 100 J of 1.06 $\mu$m radiation in 300 ps, which is converted to about 60 J of 0.53 $\mu$m light incident on a target, in each of its two beams. One beam is used to directly irradiate a thin layer of gold (500 \AA) producing X-rays that serve as an heating source for an aluminum sample (0.2-0.5 $\mu m$). A layer of plastic is included (0.25 $\mu m$), which is largely transparent to these X rays but remains at a higher density than the critical density associated with the laser radiation, in order to ensure that the sample is heated entirely by X rays and not by laser light. The sample foil itself is clad in plastic (0.3 $\mu$m) to constrain the foil expansion and produce a more uniform density. The measured quantity is the transmisison of the sample, obtained from
\begin{equation}
T(h\nu)=\frac{I(h\nu)}{I_0},
\end{equation}
\noindent where $I(h\nu)$ is the attenuated intensity of the probe (backlighter) radiation and $I_0$ the reference intensity. The transmission can be related to opacity, for an homogeneous plasma in which reabsorption processes can be neglected, by Beer-Lambert-Bouguer's law:
\begin{equation}
T(h\nu)=e^{-\rho L\kappa(h\nu)},
\end{equation}
\noindent where $\rho$ is the density of the material and $L$ its thickness. Figure \ref{Al_Eidmann2+davidson3} (left) shows that the spectrum can be interpreted with a rather high accuracy (except maybe as concerns the main feature around 1530 eV) using SCO-RCG at a single temperature and a single density.
The spectrally-resolved transmission of aluminum was measured in the range of 70 to 280 eV at $\rho$=0.01 g/cm$^3$ and $T$=22 eV \cite{winhart95,winhart96}. For this purpose, the iodine laser ASTERIX IV (200 J energy with pulse duration 0.4 ns at 400 nm) was focused into a spherical gold cavity with a diameter of 3 mm. The generated radiation with a temperature $T_R\approx$ 60 eV heated thin tamped absorber foils, which were probed by the radiation of a backlighter source. In contrast to the X-ray range (around 1 keV) in which many experiments were performed using crystal spectroscopy, the spectral range below 1 keV is much less explored. It requires reliable XUV diagnostics and sufficient suppression of self-emission of the heated sample. On the other hand, it is this spectral range, which determines the Rosseland and Planck mean opacities, \emph{i.e.} the radiative transfer. As can be seen in Fig. \ref{Al_Eidmann2+davidson3} (right), the spectrum can be analyzed taking into account a temperature variation of 6 eV.
\begin{figure}[h]
\vspace{3mm}
\centering
\includegraphics[width=7.2cm]{figure1.jpg}
\hspace{2mm}
\includegraphics[width=7cm]{figure2.jpg}
\caption{Left: aluminum transmission spectrum measured by Davidson \emph{et al.} \cite{davidson88} on HELEN laser facility in the United Kingdom and interpreted by SCO-RCG at $T$=37 eV and $\rho$=0.01 g/cm$^3$. The areal mass is equal to 54 $\mu$g/cm$^{2}$. Right: aluminum spectrum measured by Winhart \emph{et al.} \cite{winhart95,winhart96} on ASTERIX IV laser facility in Germany and interpreted by SCO-RCG at $\rho$=0.01 g/cm$^3$ and averaged over four temperatures ($T$=18, 20, 22 and 24 eV) in order to simulate the gradients. The areal mass is equal to 30 $\mu$g/cm$^{2}$.}
\label{Al_Eidmann2+davidson3}
\end{figure}
X-ray transmission spectrum of copper was measured at ``Laboratoire d'Utilisation des Lasers Intenses'' (LULI) in France at the LULI2000 laser facility with an improved design of indirect heating \cite{dozieres15}. The sample is a thin foil of mid-Z material inserted between two gold cavities heated by two 300 J, frequency-doubled (\emph{i.e.} the wavelength of the laser is $\lambda$=0.526 $\mu$m for a better X-ray conversion efficiency) nanosecond laser beams. A third laser beam irradiates a gold foil to create a spectrally continuous X-ray source (backlight) used to probe the sample. Figure \ref{calculs_tir69} shows an interpretation of the transmission of a multi-layer sample made of different materials: C (70 nm)/Al (38 nm)/Cu (12 nm)/Al (38 nm)/C (70 nm).
\begin{figure}[h]
\vspace{7mm}
\centering
\includegraphics[width=10cm]{figure3.jpg}
\caption{Copper $2p-nd$, $n$=3, 4 ... (around 975 eV, 1075 eV, ...) and aluminum $1s-n'p$, $n'$=2, 3, ... (around 1530 eV, 1675 eV, ...) absorption structures measured by Dozi\`eres \emph{et al.} measured at LULI laser facility \cite{dozieres15}. Comparison between experiment and our SCO-RCG calculation at $T$=27 eV and $\rho$= 0.01 g/cm$^3$. The areal mass of copper is equal to 15 $\mu$g/cm$^{2}$ and that of aluminum to 14 $\mu$g/cm$^{2}$.}
\label{calculs_tir69}
\end{figure}
There are several benefits in resorting to Z-pinch radiation for opacity measurements, including relatively large cm-scale lateral sample sizes and relatively long 3-5 ns experiments durations. Indeed, these characteristics enhance sample uniformity. The spectrally resolved transmission through a CH-tamped NaBr foil was measured by Bailey \emph{et al.} at the Sandia National Laboratories (SNL) and published in 2003 \cite{bailey03}. The Z-pinch produced the X rays for both the heating and backlight sources. There is a good agreement between observed and synthetic (SCO-RCG) spectra in the region were the $n$=2 to 3, 4 transitions in bromine ionized into the M shell are exprected to occur (see Fig. \ref{BaileyT2_new_fr_2bis}). The ratio between the two main structures ($2p_{1/2}\rightarrow 3d_{3/2}$ and $2p_{3/2}\rightarrow 3d_{5/2}$) is highly sensitive to relativistic configuration interaction (\emph{i.e.} Coulomb interaction between relativistic subconfigurations of a non-relativistic configuration).
\begin{figure}[h]
\vspace{7mm}
\centering
\includegraphics[width=8cm]{figure4.jpg}
\caption{Bromine $2p-3d$ absorption structures of an NaBr plasma measured by Bailey \emph{et al.} \cite{bailey02,bailey03}. Comparison between experiment and our SCO-RCG calculations at $n_e$=3$\times$10$^{21}$ cm$^3$, averaged over three temperatures: $T$= 39, 42 and 45 eV. The areal densities obtained from the He$_{\gamma}$ and He$_{\delta}$ Na line fits are respectively 1.0$\times$10$^{17}$ and 1.8$\times$10$^{17}$ atoms/cm$^{2}$.}
\label{BaileyT2_new_fr_2bis}
\end{figure}
\begin{figure}[h]
\vspace{2mm}
\centering
\includegraphics[width=7.2cm]{figure5.jpg}
\caption{The Z machine at SNL. \copyright Sandia National Laboratories.}
\label{Z}
\end{figure}
In 2007, Bailey \emph{et al.} reported on iron transmission measurements at inferred $T$= 156 eV and $n_e$=6.9$\times$10$^{21}$ cm$^{-3}$ over the photon energy range $h\nu\approx$ 800-1800 eV \cite{bailey07}. The samples consisted of an Fe/Mg mixture fully tamped on both sides by a 10 $\mu m$ thick parylene-N (C$_8$H$_8$). The Fe/Mg mixture was fabricated by depositing 10 alternating Mg and Fe layers. The challenges of high-temperature opacity experiments were overcome here using the dynamic hohlraum X-ray source at the SNL Z facility (see Fig. \ref{Z}). The process entails accelerating an annular tungsten Z-pinch plasma radially inward onto a cylindrical low density CH$_2$ foam, launching a radiating shock propagating toward the cylinder axis. Radiation trapped by the tungsten plasma forms a hohlraum and a sample attached on the top diagnostic aperture is heated during $\approx$ 9 ns when the shock is propagating inward and the radiation temperature rises above 200 eV. The radiation at the stagnation is used to probe the sample. The experimental spectrum was well reproduced by many fine-structure opacity codes \cite{bailey07} (see Fig. \ref{fitBaileywithorwithoutSCO} for the comparison with SCO-RCG). The features around 12.4 \AA~ were not reproduced by any of the involved codes.
\begin{figure}[h]
\vspace{2mm}
\centering
\includegraphics[width=12cm]{figure6.jpg}
\caption{Iron transmission spectrum measured by Bailey \emph{et al.} \cite{bailey07} on the SNL Z facility. Comparison between experiment and our SCO-RCG calculation at $T$=150 eV and $\rho$=0.058 g/cm$^3$. The areal mass, measured by Rutherford backscattering, is equal to 54 $\mu$g/cm$^{2}$.}
\label{fitBaileywithorwithoutSCO}
\end{figure}
\vspace{2mm}
Hot, solid density aluminum plasmas were generated using a 4$\times$10$^{18}$ W/cm$^2$, 350 fs laser pulse at ELFIE laser facility in LULI in France \cite{dervieux15}. Ultra-high intensity (UHI) laser pulses are efficient tools to reach states of matter associated with stellar opacities. The mechanisms whereby the laser-accelerated electrons propagate and deposit their energy through the target are different from those in the ns-pulse heating. The non-thermal, high-energy electron density is usually high enough so that, in addition to the direct collisions with the target particles, the dominant heating process is the ohmic dissipation of the inductive return current formed by collisional background electrons. Time-integrated K-shell spectra and time-resolved He$_\beta$ line emission were used to infer the plasma parameters following the laser irradiation. A suite of simulation tools was employed to describe the laser-solid interaction and the subsequent radiative-hydrodynamic processes. The space-time integrated intensity spectrum is well reproduced by SCO-RCG for a mean temperature of 310 eV (see Fig. \ref{fig_uhi}). The intensity is computed here as
\begin{equation}
I(h\nu)=B(h\nu)\times\left[1-e^{-\rho L_1\kappa(h\nu)}\right]\times S_e\times\Delta t/h,
\end{equation}
\noindent where $B(h\nu)$ (erg/cm$^2$/sr) represents Planck's distribution function, $\rho$ (g/cm$^ 3$) is the density of the material, $L_1$ (cm) its reabsorption length, $S_e$ (cm$^3$) the emissive surface, $\Delta t$ (s) the emission duration and $h$ the Planck constant expressed in eV.s. The structure around 1720 eV corresponds to Ly$_{\alpha}$ and the structure around 1850 eV to He$_{\beta}$.
\vspace{6mm}
\begin{figure}[h]
\vspace{8mm}
\centering
\includegraphics[width=8cm]{figure7.jpg}
\caption{Measured emission of aluminum ``buried layers'' heated by an ultra-short laser \cite{dervieux15} (emissive volume: 400 $\mu m^2$ $\times$ 0.5 $\mu m$, duration: 3 ps) compared to our SCO-RCG prediction at $T$=310 eV and solid density ($\rho_0$=2.7 g/cm$^3$).}
\label{fig_uhi}
\end{figure}
\section{Stellar envelopes}\label{env}
\subsection{$\kappa$ mechanism}
The opacity of the ``iron group'' (Cr, Fe, Ni and Cu) for $T\approx$ 2-3$\times$10$^5$ K and $\rho\approx$ 10$^{-7}$-10$^{-6}$ g/cm$^3$ is particularly important for envelopes of $\beta$ Cephei ($\beta$ Canis Majoris) stars, such as $\nu$ Eridani, $\gamma$ Pegasi, $\beta$ Crucis, $\beta$ Centauri, \emph{etc.}, which are hot blue-white stars of spectral class B. Their mass is from 8 to 20 M$_{\odot}$ and their magnitude from +3.16 to +3.27. $\beta$ Cephei are pulsating stars, progenitors of type II supernovae, and their period is about 4.57 hours. They pulsate through the $\kappa$ mechanism, due to M-shell transitions in elements of the iron group, which induce, when plotting Rosseland mean opacity as a function of $\log$(temperature in K), an opacity bump. The shape and position of this bump are very sensitive to the mass, the metallicity (fraction of mass that is not in hydrogen or helium) and the age of the considered star, showing that a precise determination of these three parameters as well as the proper knowledge of the opacities are required to understand the structure and evolution of the star. The first difficulty in interpreting their oscillation spectrum comes from the fact that some modes are observed but not predicted by stellar models.
Opacity drives the pulsations of many variable stars. In cases where the opacity increases with $T$, the atmosphere becomes unstable against pulsations. The kappa mechanism proceeds in several steps:
\vspace{2mm}
1. The inward motion of a layer of the star tends to compress the layer and increase the density $\rho$.
\vspace{2mm}
2. The layer becomes more opaque, the flux from the deeper layers gets stuck in the high opacity ($\kappa$) region.
\vspace{2mm}
3. This heat increase causes a build-up of pressure that pushes the layer back out again.
\vspace{2mm}
4. The layer expands, cools and becomes more transparent to radiation.
\vspace{2mm}
5. Energy and pressure beneath the layer diminish.
\vspace{2mm}
6. The layer falls inward and the cycle repeats.
\vspace{2mm}
There is a renewal in the kappa mechanism studies since the launch of COROT (2006) which is the first satellite dedicated to the development of asteroseismology \cite{michel08} and search for exoplanets. This mechanism is responsible for the pulsational instability of stars between 1.5 and 20 M$_{\odot}$. In the external optically thick layers, the diffusion approximation is justified and the variation of radiation flux $F_R$ is related to luminosity $L_r$ by
\begin{equation}
\delta\mathrm{div}F_R=\frac{1}{4\pi r^2}\frac{d\delta L_r}{L_r},
\end{equation}
\noindent where
\begin{equation}
\frac{\delta L_r}{L_r}=\frac{dr}{d\ln T}-\frac{d\kappa}{\kappa}+A\left(\frac{\delta T}{T}+\frac{\delta r}{r}\right).
\end{equation}
\noindent The first term describes the radiative dissipation and in fact stabilizes the star, the term $\delta T/T$ (sometimes referred to as $\gamma$ mechanism) describes the direct influence of the temperature variation on the luminosity (this effect is usually small), $\kappa$ represents the Rosseland mean opacity, $A$ is a constant, and the last term contributes to the instability because the radiating area is reduced in case of the compression \cite{turck09}. The $\kappa$ mechanism will work if the opacity varies more quickly with radius than the other terms. Two kinds of excitation are possible: the opacity bump due to the ionization of He (He II or He III), this is the so-called Eddington valve \cite{shapley14,eddington18,eddington26a,eddington26b}, at $\log T\approx$ 4.5 ($T$ in K), excites the evolved Cepheids, RR Lyrae (low mass evolved stars) and the $\delta$ Scuti stars (1.5-2.5 M$_{\odot}$ main-sequence stars) or the opacity bump is due to the M shell of iron at $\log T\approx$ 5.2, which excites the $\beta$ Cephei (see Fig. \ref{Alfirk}) stars (7-20 M$_{\odot}$ main-sequence stars), ``Slowly Pulsating B'' stars (SPB, 3-9 M$_{\odot}$ main-sequence stars) and ``B-type subdwarf'' (sdB, helium core, 0.5-1.4 M$_{\odot}$). The latter are representative of a late stage in the evolution of some stars, caused when a red giant star looses its outer layer before the core begins to fuse helium. The sdB stars are expected to become white dwarfs without going through any more giant stages. Some of them pulsate with a period from 900 to 600 seconds (such as V361 Hydrae in the Hydra constellation), and others with a period 45 to 180 minutes. Their mass is from 2 to 6 M$_{\odot}$. SPB stars are subject to gravity modes (``g'' modes), also connected to iron opacity. SPB and $\beta$ Cephei-type stars are interesting pulsators that might put important constraints on rotation and magnetic fields, two key actors for massive stars.
\begin{figure}[h]
\vspace{2mm}
\centering
\includegraphics[width=6.2cm]{figure8.jpg}
\caption{$\beta$ Cephei star. \copyright NASA, Palomar observatory.}
\label{Alfirk}
\end{figure}
\subsection{Comparisons of Rosseland means}\label{ross}
The Rosseland mean opacity is the harmonic mean opacity averaged over the derivative of the Planck function with respect to the temperature:
\begin{equation}
\frac{1}{\kappa_R}=\int_0^{\infty}\frac{W_R(u)}{\kappa'(u)}du,
\end{equation}
\noindent where
\begin{equation}
u=\frac{h\nu}{k_BT}\;\;\;\; \mathrm{and}\;\;\;\; W_R(u)=\frac{15u^4e^{-u}}{4\pi^4\left[1-e^{-u}\right]^2}.
\end{equation}
\noindent The ATOMIC-code \cite{fontes15,colgan16} ``full'' calculations allow single-configuration fine-structure detail to be included in a relatively computationally inexpensive manner. They thus include intermediate coupling within a configuration but not complete interaction among the set of configurations. Such a simplification allows one to include the contributions from many thousands of configurations. A second set of ATOMIC semi-relativistic multi-configuration calculations was also performed using a reduced list of configurations. These calculations were designed to include the transitions in the photon range of current interest ($h\nu$ lower than 250 eV), $\Delta n$=0,1, 2 and $n_{\mathrm{max}}=5$ ($n$ is the principal quantum number), referred to as ATOMIC n5. One observes that ATOMIC ``full'' and ATOMIC n5 agree within 20 \% and that, in all cases, ATOMIC n5 values are smaller than those of ATOMIC ``full''. SCO-RCG, which is not limited to $n\leq$ 5, is generally closer to ATOMIC ``full'' than to ATOMIC n5 (see Fig. \ref{compa_2mg+compa_2mg_1}, left and right, as well as tables \ref{atomic1} and \ref{atomic2}) \cite{turck16}, which is consistent with the fact that configuration interaction plays a less important role than state completeness in the present case. The values of OP (Opacity Project) \cite{op1,op2}, are usually lower. The Opacity Project, an international collaboration of about 25 scientists from 6 countries, started in response to a plea \cite{simon82}. The atomic data are available in the database TOPbase \cite{topbase} and the monochromatic opacities and Rosseland mean opacities are available at the OPServer \cite{opserver} at the Ohio Supercomputer Center.
\begin{figure}[h]
\vspace{5mm}
\centering
\includegraphics[width=7cm]{figure9.jpg}
\hspace{2mm}
\includegraphics[width=7cm]{figure10.jpg}
\caption{Comparison between SCO-RCG and ATOMIC ``full'' computations \cite{fontes15,colgan16,turck16} of iron opacity at $T$=23 eV and $\rho$=2 mg/cm$^3$. The ATOMIC calculations were kindly provided by J. Colgan.}
\label{compa_2mg+compa_2mg_1}
\end{figure}
\begin{table}[h]
\caption{Values of Rosseland mean opacity calculated by OP, ATOMIC and SCO-RCG codes for conditions of stellar envelopes.}
\centering
\begin{tabular}{cccccc}\hline
\textbf{$T$ (K)} & \textbf{$n_e$ (cm$^{-3}$)} & \textbf{$\rho$ (g/cm$^{3}$)} & \textbf{OP (cm$^2$/g)} & \textbf{ATOMIC ``full'' (cm$^2$/g)} & \textbf{SCO-RCG (cm$^2$/g)}\\\hline
125800 & 10$^{17}$ & 1.35$\times$10$^{-6}$ & 25 & 64 & 63\\
177827 & 3.16$\times$10$^{17}$ & 3.44$\times$10$^{-6}$ & 358 & 683 & 674\\
199473 & 10$^{17}$ & 9.52$\times$10$^{-7}$ & 354 & 487 & 500\\
251190 & 10$^{18}$ & 8.85$\times$10$^{-6}$ & 1270 & 1359 & 1313\\
295553 & 3.16$\times$10$^{17}$ & 2.44$\times$10$^{-6}$ & 232 & 232 & 122\\\hline
\end{tabular}
\label{atomic1}
\end{table}
\begin{table}[h]
\caption{Rosseland mean opacities computed with two versions of ATOMIC and SCO-RCG codes at conditions that can be reached experimentally on a laser facility. These conditions are interesting since they provide an average ionization close to the one of real stellar envelopes.}
\centering
\begin{tabular}{ccccc}\hline
\textbf{$T$ (eV)} & \textbf{$\rho$ (g/cm$^{3}$)} & \textbf{ATOMIC full (cm$^2$/g)} & \textbf{ATOMIC n5 (cm$^2$/g)} & \textbf{SCO-RCG (cm$^2$/g)}\\\hline
21 & 2$\times$10$^{-3}$ & 19266 & 14361 & 17853\\
22 & 2$\times$10$^{-3}$ & 19613 & 14910 & 18435\\
23 & 2$\times$10$^{-3}$ & 19508 & 15205 & 18510\\
25 & 2$\times$10$^{-3}$ & 18384 & 15094 & 17550\\\hline
\end{tabular}
\label{atomic2}
\end{table}
\noindent The observed modes do not agree, as concerns their stability, with modes deduced from the OP or OPAL tables \cite{iglesias96}, which suggests that those opacities are underestimated \cite{daszynska10}. A recent study \cite{salmon12} pointed out that as much as 50 \% increase in astrophysical opacities was necessary to solve B star pulsation problems in the Magellanic cloud. Iglesias \cite{iglesias15b} showed that an increase of population I stars (metallicity of 0.02) near the iron-group bump involves an order of magnitude multiplier on the nickel opacity and more than an overall doubling of the iron opacity. Although there are uncertainties in theoretical opacities (spectral-line models, configuration-interaction accounting, mixture models, ...), these errors seem unable to explain such large opacity increases. Furthermore, code comparisons do show differences in spectral details, but also suggest that current advanced implementations of photon-absorption theories are unlikely to produce large overall discrepancies in the iron-group opacities for matter conditions relevant to B stars.
Moravveji \cite{moravveji15} produced new opacity tables with enhanced iron and nickel contributions to the Rosseland mean opacity by 75 \% each, and computed, with the MESA (Modules for Experiments in Stellar Astrophysics) code \cite{paxton15}, three dense grids of evolutionary models for Galactic O- and B-type stars covering from 2.5 to 25 M$_{\odot}$ from zero-age main sequence until effective temperature $T_{\mathrm{eff}}$=10000 K after the core hydrogen exhaustion. The authors carried out non-adiabatic mode stability analysis and found that $\approx$ 75 \% enhancement, only in the iron opacity, is sufficient to consistently reproduce the observed position of late O-type $\beta$ Cepheids and eight hybrid B-type pulsators on the Kiel diagram, \emph{i.e.} $\log T_{\mathrm{eff}}$ vs $\log g$, where $g$ represents the stellar surface gravity \cite{langer14}.
\section{The boundary of the radiative / convective zones of the Sun}\label{bou}
The Sun is our closest star and is thus used as a benchmark to study other stars. Its radius, luminosity and mass are known with a great accuracy. However, some doubts are raised on the relevance of the used microscopic physics. In the past decade, the photospheric abundances of the Sun have been revised several times by many observers \cite{asplund05,asplund09,yang16}. The standard solar models constructed with the new low-$Z$-metal abundances disagree with helioseismic results and detected neutrino fluxes \cite{turck01a,turck01b,antia06,guzik08,serenelli09,turck11}. For instance, a discrepancy between helioseismic observations and predictions by standard solar models appeared in the solar sound speed profile (sound speed versus radius). This discrepancy, of about 20 times the vertical error bar, raises some questions on the solar radiative transfer. In addition, the Sun has an inner radiative heat conduction zone that gives way to a convective zone nearer the surface. The different contributions to the opacity of iron calculated with SCO-RCG at conditions of the boundary of the convective zone (BCZ) are displayed in Fig. \ref{contrib_bcz+sco-rcg_bcz} (left) as well as a comparison between our SCO-RCG and ATOMIC codes (right). The Rosseland means are rather close, although some differences exist around $h\nu$=800 eV (maximum of the derivative of Planck's distribution with respect to temperature).
\begin{figure}[h]
\vspace{4mm}
\centering
\includegraphics[width=7cm]{figure11.jpg}
\hspace{2mm}
\includegraphics[width=7cm]{figure12.jpg}
\caption{Different contributions to the opacity of iron calculated with SCO-RCG at conditions of the boundary of the convective zone (left), {\em \emph{i.e.}} $T$=192.91 eV and $n_e$=10$^{23}$ cm$^{-3}$, and a comparison between the present SCO-RCG and ATOMIC total opacities \cite{fontes15,colgan16} at the same conditions (right). The ATOMIC calculations were kindly provided by J. Colgan. The OP Rosseland mean at these conditions is much lower and equal to 854 cm$^2$/g.}
\label{contrib_bcz+sco-rcg_bcz}
\end{figure}
Solar models typically find a location of the boundary between these zones that differs significantly from the measured one (by more than 13 standard deviations \cite{basu08}). Several hypotheses can explain this difference:
\vspace{2mm}
$\bullet$ macroscopic processes in the radiative zone are not taken into account in the energetic balance of the Sun.
\vspace{2mm}
$\bullet$ the radiative transfer calculations are not accurate, either as concerns the Rosseland mean value that could be underestimated or in the treatment of the radiative acceleration which limits the gravitational settling and could lead to incorrect internal abundances. The radiative acceleration of species $k$ reads
\begin{equation}
g(k)=\frac{F_R}{c}\frac{M}{M_k}\kappa_R\gamma(k),
\end{equation}
\noindent where $F_R$ is the radiation flux, $M$ the total mass of the star, $M_k$ the mass of species $k$ and
\begin{equation}
\gamma(k)=\int\frac{\kappa_{\nu}(k)}{\kappa_{\nu}(\mathrm{total})}f_{\nu}d\nu,
\end{equation}
\noindent where $\kappa_{\nu}(k)$ is the monochromatic opacity of species $k$ and $f_{\nu}$ is a function of the frequency and of the temperature. The unexplained discrepancies could also be due to all these effects simultaneously. The heavy elements significantly contribute to opacity even if they are present only at few percents in mass fraction in the solar mixture, which is mainly constituted of hydrogen and helium. The most important contributors are:
\vspace{2mm}
$\bullet$ iron, which contributes to the total opacity (including hydrogen and helium) at a level of 20 \% in most of the radiative zone because it is always partially ionized;
\vspace{2mm}
$\bullet$ oxygen, which becomes partially ionized at 0.6 R$_{\odot}$ and plays a major role at the basis of the convective zone. The increase of its opacity contribution triggers the convection;
\vspace{2mm}
$\bullet$ silicon, which contributes about 10 \% at temperatures below 10 MK.
\vspace{2mm}
Rotation, enhanced diffusion, convection overshoot and magnetic field contribute to explain the discrepancy. The too low helium subsurface abundance in enhanced diffusion models can be improved by the mixing caused by rotation and magnetic fields \cite{yang16}. The problem of the depth of the convective zone in rotating models can be resolved by accounting for convection overshoot. The latest phenomenon refers to convection carrying matter beyond and unstable region into a stratified, stable region. Overshoot is due to the momentum of the convective material, which carries the material beyond the unstable region. The heat of the Sun's thermonuclear fusion is carried outward by radiation in the deep interior radiation zone and by convective circulation in the outer convection zone. However, cool sinking matter from the surface penetrates farther into the radiative zone than theory would suggest. The Asplund-Grevesse-Sauval rotation model \cite{asplund09} including overshooting reproduces the seismically inferred sound-speed and density profiles and the convection zone depth, but fails to reproduce the surface helium abundance and neutrino fluxes.
More generally, the solar abundance controversy inspires many investigations in the microscopic physics. For instance, Mussack and D\"appen examined the correction to the proton-proton reaction rate due to dynamic screening effects \cite{mussack11}.
Stars generate low-mass weakly interacting particles which are responsible for energy losses, such as the axions, emitted by the Primakoff effect \cite{schlattl99}. Axions are elementary particles postulated in 1977 to solve the strong CP problem in quantum chromodynamics. They are a possible component of cold dark matter. It was reported in 2014 that axions have been detected as a seasonal variation in observed X-ray emission that would be expected from conversion in the Earth's magnetic field of axions streaming from the Sun \cite{fraser14}. In non-hadronic axion models, where axions couple to electrons, the solar axion flux is completely dominated by the ABC reactions (Atomic recombination and deexcitation, Bremsstrahlung and Compton). The ABC flux was computed from available libraries of monochromatic photon radiative opacities by exploiting the relations between axion and photon emission cross sections. These results turn to be $\approx$ 30 \% larger than previous estimates due to atomic recombination (free-bound electron transitions) and deexcitation (bound-bound), which where not taken into account \cite{redondo13}.
The spectra of silicon, in particular with high resolution, are helpful for deducing the abundance of silicon in stars. In the past decades, the K-shell absorption lines of silicon ions in various astrophysical objects have been extensively observed with high-resolution spectrometers of the XMM-Newton, Chandra, and Suzaku space missions. A typical example of a spectrometer is the High-energy Transmission Grating Spectrometer (HETGS) on board the Chandra X-ray Observatory. With the HETGS, silicon absorption lines were observed in the spiral galaxy NGC 3783, in the black hole candidate Cygnus X-1, in the ultracompact X-ray Dipper 4U 1916-05, in the Seyfert 1 galaxy NGC 5548, in the bright Seyfert 1 galaxy MCG 6-30-15 and NGC 3516, in the Seyfert 1 active galactic nucleus NGC4051, \emph{etc.} \cite{xiong16}. Silicates are an important component of cosmic matter \cite{savin12}. They form in the winds of AGB (asymptotic giant branch) stars. The AGB is a region of the Hertzsprung-Russell diagram populated by evolved cool luminous stars. This is a period of stellar evolution that is experienced by all low- to intermediate-mass stars (0.6-10 M$_{\odot}$) late in their lives. An AGB star appears, for a large part of its life, as a bright red giant with a luminosity thousands of times greater than the Sun. Silicates are processed in the diffuse interstellar medium. They are also present in dust of protoplanetary and debris disks where they help regulating thermal exchanges. In circumstellar environments, evidence for crystalline silicates is found both around AGB stars and in disks around Ae/Be stars, T Tauri stars and brown dwarfs \cite{savin12}. They are also found in cometary environments, in asteroid spectra and in interplanetary dust particles. Accurate interpretation of their absorption lines requires detailed theoretical models \cite{wei08}.
\begin{table}[h]
\caption{Temperature and free-electron density at different depths of the Standard Solar Model computed with the MESA code \cite{paxton15} for the Asplund \emph{et al.} composition \cite{asplund09}.}
\centering
\begin{tabular}{ccc}\hline
\textbf{Solar radius fraction ($r/R_{\odot}$)} & \textbf{$T$ (eV)} & \textbf{$n_e$ (cm$^ {-3}$)}\\\hline
0.5 & 340 & 8$\times$10$^{23}$\\
0.6 & 270 & 2.5$\times$10$^{23}$\\
0.7 & 200 & 10$^{23}$\\\hline
\end{tabular}
\label{radius}
\end{table}
Oxygen and magnesium are also important for the modelling of the Sun (their respective opacities are shown in Fig. \ref{oxygen-magnesium}).
\begin{figure}[h]
\vspace{6mm}
\centering
\includegraphics[width=8cm]{figure13.jpg}
\caption{Opacity of oxygen and magnesium computed by SCO-RCG for $T$=174.92 eV and a total density of the mixture of $\rho$=0.16 g/cm$^3$ (the partial densities of oxygen and magnesium are 0.321 and 0.356 g/cm$^3$ respectively), corresponding to solar radius $R$=0.726.}
\label{oxygen-magnesium}
\end{figure}
Reproducing the solar interior is a real challenge because one tries to reproduce the charge state distribution of the different elements together with the free-electron density $n_e$ at the targeted conditions, that means density greater than that of a solid and high temperature (see table \ref{radius}).
\section{Attempts to understand the enigmatic photo-absorption experiment on iron performed by Bailey \emph{et al.} at Sandia National Laboratory}\label{bai}
As mentioned in Section \ref{int}, Bailey \emph{et al.} published in 2007 a measurement, performed on the Z-pinch facility, of the iron transmission at conditions lower than those of the basis of the convective zone ($T$=156 eV and $\rho$=0.058 g/cm$^3$) \cite{bailey07}. This first measurement agrees well with most of the theoretical predictions. Then they have increased the temperature up to 196 eV and reached free-electron densities of several times 10$^{21}$ cm$^{-3}$. However, for this last measurement, an unexplained gap exists between the measurement and the theoretical calculations \cite{bailey15}.
\subsection{Effect of density and temperature}
Density and temperature were inferred by Nagayama \emph{et al.} \cite{nagayama16}, in the experiment described in Ref. \cite{bailey15}, by analysis of Stark profiles of magnesium K-shell lines. As shown in Fig. \ref{fig_Mg_tn}, the line profiles (intensity, asymmetry, wings, ...) are very sensitive to the plasma conditions, and can therefore be used to diagnose the plasma temperature and density. In the SNL experiment, the analysis of the K-shell transmission spectra of Mg leads to $T$=182 eV and $n_e$=10$^{22}$ cm$^{-3}$ (which corresponds roughly to $\rho$=0.17 g/cm$^3$) \cite{nagayama16}.
\begin{figure}[h]
\vspace{6mm}
\centering
\includegraphics[width=8cm]{figure14.jpg}
\caption{Magnesium opacity computed by SCO-RCG at $T$=165 eV, $n_e$=9$\times$10$^{21}$ cm$^{-3}$, $T$=195 eV, $n_e$=4$\times$10$^{22}$ cm$^{-3}$ and $T$=195 eV, $n_e$=6$\times$10$^{22}$ cm$^{-3}$.}
\label{fig_Mg_tn}
\end{figure}
Figure \ref{fig1+2} displays the opacity computed by SCO-RCG in two different conditions: $T$=182 eV, $n_e$=3.1$\times$10$^{22}$ cm$^{-3}$ and $T$=195 eV, $n_e$=4$\times$10$^{22}$ cm$^{-3}$. The amplitude of the structures is better reproduced at $T$=182 eV, $n_e$=3.1$\times$10$^{22}$ cm$^{-3}$, but the predicted opacity level is much lower than the experimental one in both cases. However, the highest opacity values (peaks around 10.7 and 11.5 \AA) at $T$=182 eV, $n_e$=3.1$\times$10$^{22}$ cm$^{-3}$ are consistent with the experimental ones.
\begin{figure}[h]
\vspace{2mm}
\centering
\includegraphics[width=7.5cm]{figure15.jpg}
\hspace{2mm}
\includegraphics[width=7.5cm]{figure16.jpg}
\caption{Opacity computed by SCO-RCG in two different conditions: $T$=182 eV, $n_e$=3.1$\times$10$^{22}$ cm$^{-3}$ and $T$=195 eV, $n_e$=4$\times$10$^{22}$ cm$^{-3}$.}
\label{fig1+2}
\end{figure}
The dominant ions in the spectrum (see Fig. \ref{ions}) at $T$=182 eV and $n_e$=3.1$\times$10$^{22}$ cm$^{-3}$ are Fe XVII (24.9 \%), Fe XVIII (33 \%) and Fe XIX (21.7 \%).
\begin{figure}[h]
\vspace{2mm}
\centering
\includegraphics[width=8cm]{figure17.jpg}
\caption{Opacity of the dominant ions in the spectrum Fe XVII, Fe XVIII and Fe XIX at $T$=182 eV and $n_e$=3.1$\times$10$^{22}$ cm$^{-3}$ ($\approx$ 0.17 g/cm$^3$) computed by SCO-RCG. The opacity of each charge state is not weighted by the ionic fraction.}
\label{ions}
\end{figure}
The contributions to the total opacity (photo-excitation, photo-ionization, inverse Bremsstrahlung and scattering of a photon by a free electron) at $T$=182 eV and $n_e$=3.1$\times$10$^{22}$ cm$^{-3}$ are represented in Fig. \ref{contributions}. The most important processes are photo-excitation and photo-ionization.
\begin{figure}[h]
\vspace{3mm}
\centering
\includegraphics[width=8cm]{figure18.jpg}
\caption{Contributions to the total opacity at $T$=182 eV and $n_e$=3.1$\times$10$^{22}$ cm$^{-3}$ ($\approx$ 0.17 g/cm$^3$) computed by SCO-RCG.}
\label{contributions}
\end{figure}
Figure \ref{Fe_195eV+zoom} shows the effect of density on the computed spectrum at $T$=195 eV. At the highest density ($n_e$=6$\times$10$^{22}$ cm$^{-3}$), the gaps between the structures (for instance between the three F-like 2p-4d features around 11.5 \AA), are less pronounced than at the lower electron density $n_e$=4$\times$10$^{22}$ cm$^{-3}$ and the wings of lines are broader, exhibiting shoulders. However, even at the highest density, the discrepancy between theory and experiment remains prominent.
\vspace{2mm}
\begin{figure}[h]
\centering
\includegraphics[width=11.5cm]{figure19.jpg}
\hspace{10mm}
\includegraphics[width=11.5cm]{figure20.jpg}
\caption{Effect of density on the iron opacity at $T$=195 eV (SCO-RCG computations). The experimental spectrum is also represented \cite{bailey15}.}
\label{Fe_195eV+zoom}
\end{figure}
\subsection{Effect of highly excited states}
In multiply-charged ion plasmas, a significant number of electrons may occupy high-energy orbitals. These ``Rydberg'' electrons, when they act as spectators, are responsible for a number of satellites of X-ray absorption or emission lines, yielding an effective broadening of the resonance lines (red wings). The contribution of such satellite lines may be important, because of the high degeneracy of the relevant excited configurations which give a large Boltzmann weight. However, it is in general difficult to take those configurations into account since they are likely to give rise to a large number of lines. We proposed to model the perturbation induced by the spectators in a detailed way, inspired by the Partially Resolved Transition Array approach recently published by Iglesias \cite{iglesias12a,iglesias12c}, and extended to the STA theory \cite{pain15a,wilson15}. This approach, based on the additivity of the variances from active and spectator electrons in the STA theory, consists in a reduced detailed-line-accounting calculation, omitting these Rydberg electrons, the effect of the latests being included through an additional shift and broadening of the lines, expressed in terms of canonical partition functions, key-ingredients of the STA model. The resulting method can \emph{a priori} be used in any detailed-configuration/line-accounting opacity code.
In the case of an iron plasma at $T$=182 eV and $n_e$=10$^{22}$ cm$^{-3}$ (conditions of the recent experiment by Bailey \emph{et al.} \cite{bailey15}), the statistical modelling of satellites [Be] $2p^5(5s\cdots 8d)^1\rightarrow 2p^44d^1(5s\cdots 8d)^1$ fills significantly the gaps around 11.4 and 11.5 \AA~ (see Fig. \ref{figure_loisel2014_quad}). The total contribution of satellites (blue curve) is very similar to the resonance transition (red curve), except that it is slightly shifted and broadened. These spectator electrons are very weakly bound to the ion, so that they do not interact much with the core electrons and perturb very weakly the transition. In our previous approach, the satellite lines due to the Rydberg electrons were treated in the STA theory, and yielded an overestimation of the opacity in the gap between the three main features of the F-like $2p\rightarrow 4d$ lines.
\clearpage
\vspace{2mm}
\begin{figure}[h]
\centering
\includegraphics[width=7.5cm]{figure21.jpg}
\hspace{2mm}
\includegraphics[width=7.5cm]{figure22.jpg}
\caption{Modelling of satellites [Be] 2p$^5$(5s $\cdots$ 8d)$^1\rightarrow$ 2p$^4$4d$^1$(5s $\cdots$ 8d)$^1$ with SCO-RCG code.}
\label{figure_loisel2014_quad}
\end{figure}
\subsection{Comparison to cold opacity}
Figure \ref{cold_opacity} displays a comparison betweeen the experimental spectrum \cite{bailey15} and two cold opacities published by Henke \emph{et al.} \cite{henke93} and Chantler \cite{chantler95}.
\begin{figure}[h]
\vspace{6mm}
\centering
\includegraphics[width=8cm]{figure23.jpg}
\caption{Comparison between cold opacities published by Henke \emph{et al.} \cite{henke93} and Chantler \cite{chantler95} and the experimental spectrum \cite{bailey15}. The SCO-RCG opacity at $T$=182 eV and $n_e$=3.1$\times$10$^{22}$ cm$^{-3}$ is also represented.}
\label{cold_opacity}
\end{figure}
As was also pointed out by Iglesias \cite{iglesias15a}, recalling that the oscillator strength density is conserved, a smoothed cross-section obtained by the following convolution is constrained to be smaller than the cold opacity:
\begin{equation}
\kappa(\nu)=\int_{-\infty}^{\infty}d\nu'G\left(\nu-\nu'\right)\kappa\left(\nu'\right)\leq\kappa_{\mathrm{cold}},
\end{equation}
\noindent where $G$ represents physical (Doppler, Stark, Van der Waals neutral-atom collisions, autoionization) and instrumental broadenings. The largest theoretical opacity is obtained from ions with a full L shell, which reproduces the cold data at small wavelength. The analysis revealed other puzzling aspects of the data: the integrated absorption suggests a full L shell, inconsistent with the measured spectrum, which displays strong spectral lines from F- and O-like configurations, and the average slope of the absorption disagrees with the one of the cold data $\kappa_{\mathrm{cold}}$. Clearly, the Sandia data exceed $\kappa_{\mathrm{cold}}$ at lower wavelengths.
\subsection{R-matrix photo-ionization}
Recently \cite{nahar16}, Nahar \emph{et al.} reported extensive R-matrix calculations \cite{nahar11} of unprecedented complexity for iron ion Fe XVII, with a wavefunction expansion of 99 Fe XVIII LS core states from $n\leq 4$ complexes (equivalent to 218 fine-structure levels) and found a large enhancement in background photo-ionization cross-sections (up to orders of magnitude) in addition to strongly peaked photo-excitation-of-core resonances (see the black curve of Fig. \ref{comp_1}).
\begin{figure}[h]
\vspace{6mm}
\centering
\includegraphics[width=8cm]{figure24.jpg}
\caption{Comparison between the present SCO-RCG (distorted-wave) in blue and R-matrix \cite{nahar11} photo-ionization cross-sections (black curve) obtained from \cite{topbase}. The experimental spectrum is also represented (red curve). The resonances in the black curve (PEC: photo-excitation of core) are the signature of transitions to a state coupled to the continuum (also known as Beutler-Fano autoionizing resonances \cite{cowan81}).}
\label{comp_1}
\end{figure}
\begin{table}[h]
\caption{$\kappa_R$ calculated with different opacity codes relative to experiment.}
\centering
\begin{tabular}{cc}\hline
\textbf{Source} & \textbf{$\kappa_R$ relative to experiment} \\\hline
OP \cite{op1,op2} & 0.59 \\
R-matrix \cite{nahar11,nahar16} & 0.51 \\
ATOMIC \cite{fontes15,colgan16} & 0.60 \\
OPAS \cite{blancard12,mondet15,lepennec15} & 0.70 \\
SCO-RCG \cite{pain15a} & 0.64 \\
SCRAM \cite{hansen07} & 0.77 \\
TOPAZ \cite{iglesias03,iglesias04} & 0.62 \\
Cold \cite{henke93} & 0.75 \\\hline
\end{tabular}
\label{rmatrix}
\end{table}
Figure \ref{iglesias_wings} displays total and photo-ionization opacities of ground $1s^22s^22p^5$ and excited $1s^22s^22p^53p$ configurations. As pointed out by Iglesias \cite{iglesias16}, it clearly illustrates the fact that the photo-ionization from $3p$ is important and must absolutely be taken into account in the opacity. The point here is that the latest contribution is included in all opacity codes relying on a distorted-wave modelling of photo-ionization, but was not included in the previous R-matrix computations.
\begin{figure}[h]
\vspace{8mm}
\centering
\includegraphics[width=8cm]{figure25.jpg}
\caption{Total and photo-ionization opacities of configurations $1s^22s^22p^5$ and $1s^22s^22p^53p$ (SCO-RCG calculations).}
\label{iglesias_wings}
\end{figure}
It appears \cite{iglesias16}, that major opacity models already include cross-sections that are equivalent to the enhancements reported by the R-matrix method. The fact that the R-matrix calculations neglected important cross-sections (\emph{e.g.} photo-ionization involving $3p$ subshell, as mentioned above) explains why the resultant opacity is lower than other models in the spectral range measured in transmission experiments relevant to the solar interior (see table \ref{rmatrix}). The missing processes help understand the lower Fe XVII R-matrix opacity relative to other models making it discrepant with the SNL measurements \cite{blancard16}. The opacities from all models are larger than R-matrix over the whole energy range of the experiment except for a few narrow features. Moreover, they all reasonably agree, except R-matrix, at higher energies corresponding to the L-shell photo-ionization continuum. The R-matrix calculations consider photo-absorption for 283 LS terms of Fe XVII corresponding to the configurations
\begin{equation}
\begin{array}{l}
2s^22p^5n\ell\;\;\;\;\mathrm{for}\;\;\;\; n=2-10\\
2s^22p^6n\ell\;\;\;\;\mathrm{for}\;\;\;\; n=3,4\;\;\;\;\mathrm{plus}\;\;\;\; n\ell=5s
\end{array}
\end{equation}
\noindent excluding levels with $J>8$. For the target ion, R-matrix calculations includes the Fe XVIII configurations
\begin{equation}
\begin{array}{l}
2s^22p^5\\
2s^22p^4n\ell\;\;\;\;\mathrm{for}\;\;\;\; n=3,4\\
2s2p^6\\
2s2p^5n\ell\;\;\;\;\mathrm{for}\;\;\;\; n=3 \;\;\;\;\mathrm{plus}\;\;\;\; n\ell=4s,4p\\
\end{array}
\end{equation}
\noindent which generate 99 LS-terms. It follows that Nahar and Pradhan omitted the photo-ionizations
\begin{equation}
\begin{array}{ll}
2s^22p^5n\ell+h\nu\rightarrow 2s^22p^5n\ell\epsilon p, & n\ell=4p,4f \;\;\;\;\mathrm{and}\;\;\;\; n\ge 5\\
2s2p^6n\ell+h\nu\rightarrow 2p^6n\ell\epsilon p, & \mathrm{all}\;\;\;\; n\\
2s^22p^5n\ell+h\nu\rightarrow 2s2p^5n\ell\epsilon\ell', & n\ge 5\\
2s2p^6n\ell+h\nu\rightarrow 2s2p^5n\ell\epsilon\ell', & n\ell=4p,4f \;\;\;\;\mathrm{and}\;\;\;\; n\ge 5.
\end{array}
\end{equation}
\noindent with $\ell'=s,d$. Opacity models like ours (relying on distorted-wave approximation) treat the absorption processes described by Nahar and Pradhan as inner-shell bound-bound transitions into discrete autoionizing levels or as bound-free transitions. It is emphasized that all approaches satisfying the $f$-sum rule \cite{thomas25,kuhn25,cowan81} yield the same total absorption strength but only if they use the same set of initial and final configurations. The Thomas-Reiche-Kuhn sum rule reads:
\begin{equation}
\sum_{i,j}f_{ij}+\sum_i\int_{I_i}^ {\infty}\frac{df_{i,\epsilon}}{d\epsilon}d\epsilon=Z-Z^*,
\end{equation}
\noindent where $I_i$ is the ionization potential of subshell $i$ and $Z^*$ the average ionization ($Z-Z^*$ representing therefore the total number of bound electrons). Differences in the opacity can still occur due to several aspects of atomic-structure calculations, such as level populations, and/or details in broadening of spectral lines or photo-ionization resonances. Nahar and Pradhan's R-matrix calculations, however, did not include all target ion states consistent with their initial configurations and thus neglected important cross-sections. The R-matrix calculations also limited the initial configuration list compared with current opacity models. Some of the missing transitions impact the L-shell photo-ionization continuum and help explain the lower opacity from R-matrix relative to other models in the spectral range of the SNL experiments. Effects from other missing atomic data involve numerous weak features in spectral regions of low photo-absorption, such as PEC (photo-excitation-of-core) resonances. Proponents of close-coupling methods argue that such calculations are in principle better than other schemes (e.g.; distorted wave), but that may not be true in the plasma environment. In plasmas, fluctuating electric fields and collisions between ions and electrons may significantly impact, and even destroy, the coherence required for the interference effects included in close-coupling methods. It remains an open question if the improvements in cross-sections reported using R-matrix methods for isolated atoms are really relevant for the modelling of atomic processes in plasmas \cite{iglesias16,nahar16b}.
\subsection{Autoionization}
The process of autoionization, described in Sec. \ref{rad}, is likely to provide an additional broadening to spectral lines. The autoionization rate
\begin{equation}
A_{if}^a=\sum_j\left|\langle\psi_f,j;J_T,M_T\left|\sum_{i<j}\frac{1}{r_{ij}}\right|\psi_i\rangle\right|^2,
\end{equation}
\noindent where $J_T$ and $M_T$ are the total angular momentum and its projection, $j$ the quantum number of the free electron, $\psi_i$ and $\psi_f$ being the wavefunctions of the initial and final configurations respectively. The autoionization rate can be calculated as a configuration average \cite{sampson09,gao10}. We developed such an approach in the Multi-Configuration Dirac-Fock (MCDF) code developed by Bruneau \cite{bruneau83,bruneau84}, and our results are very close to the ones obtained with the Flexible Atomic Code (FAC) \cite{gu08}.
High populations in metastable states can also contribute to a process called ladder ionization, which becomes important when the electron temperature is too small to support significant collisional ionization from the ground state but large enough to support ionization from excited states that lie much closer to the continuum limit. The collisional ionization flux and charge state distribution therefore depend on the degree of state detail. In addition, some mid-shell ions support excitation-ionization processes for states formed by single excitation from inner subshells, such as the $1s^ 22s^22p^ 6n\ell$ states in F-like iron, which are autoionizing for $n\ge 6$ \cite{ralchenko16}. The values of the autoionization rates for Fe XVII $1s^ 22s^12p^6nd$ $\rightarrow$ Fe XVIII $1s^22s^22p^5\epsilon\ell$ and Fe XVIII $1s^22s^12p^5nd$ $\rightarrow$ Fe XIX $1s^22s^22p^4\epsilon\ell$, for $n$=1 to 20, are plotted in figure \ref{auger_hansen}. The rates can reach several 10$^ {13}$ s$^ {-1}$, which is at maximum of the order of one tenth of collisional widths. This is not negligible, but clearly not sufficient to broaden significantly the main absorption features.
\begin{figure}[h]
\vspace{4mm}
\centering
\includegraphics[width=8cm]{figure26.jpg}
\caption{Autoionization rate for Fe XVII $1s^22s^12p^6nd$ $\rightarrow$ Fe XVIII $1s^22s^22p^5\epsilon\ell$ and Fe XVIII $1s^22s^12p^5nd$ $\rightarrow$ Fe XIX $1s^22s^22p^4\epsilon\ell$, $n$=1, 20.}
\label{auger_hansen}
\end{figure}
Figure \ref{figure_1s22s12p66d_bis_2} represents transition $1s^22s^12p^66d-1s^22s^12p^54d6d$ with and without broadening due to autoionization process $1s^22s^12p^66d-1s^22s^22p^5\epsilon\ell$ computed using FAC code \cite{gu08}. The gap is slightly filled, but the initial configuration is a rather highly excited one, and its probability, and therefore its contribution to the spectrum, are not dominant. We checked that the effect of ionization/recombination is much smaller on the resonance transition $1s^22s^22p^6\rightarrow 1s^22s^22p^54d^1$.
\begin{figure}[h]
\vspace{6mm}
\centering
\includegraphics[width=8cm]{figure27.jpg}
\caption{Transition $1s^22s^12p^66d-1s^22s^12p^54d6d$ with and without broadening originating from autoionization process $1s^22s^12p^66d-1s^22s^22p^5\epsilon\ell$.}
\label{figure_1s22s12p66d_bis_2}
\end{figure}
We can see in Fig. \ref{ato} that the dispersion of autoionization rates between relativistic subconfigurations is likely to be rather important (see table \ref{tab_auto}).
\begin{figure}[h]
\vspace{5mm}
\centering
\includegraphics[width=12cm]{figure28.jpg}
\caption{Values of the autoionization rates between relativistic subconfigurations of $2p^33s^13p^2-2p^43s^1\epsilon\ell$. The dashed line represents the value between non-relativistic configurations $2p^33s^13p^2$ and $2p^43s^1\epsilon\ell$. Orbital $n\ell j$ is noted $n\ell_-$ if $j=\ell-1/2$ and $n\ell_+$ if $j=\ell+1/2$.}
\label{ato}
\end{figure}
\begin{table}[h]
\caption{Values of the configuration-averaged autoionization rate (s$^{-1}$) computed with FAC \cite{gu08} and MCDF \cite{bruneau83,bruneau84} codes for different channels.}
\centering
\begin{tabular}{ccc}\hline
\textbf{Autoionization process} & \textbf{MCDF configurations} & \textbf{FAC configurations} \\\hline
$1s^22s^22p^43d^2$ β $1s^22s^22p^5\epsilon\ell$ & 7.146$\times$10$^{13}$ & 7.194$\times$10$^{13}$\\
$1s^22s^22p^43d^14d^1$ β $1s^22s^22p^5\epsilon\ell$ & 1.291$\times$10$^{13}$ & 1.374$\times$10$^{13}$\\
$1s^22s^22p^33d^2$ β $1s^22s^22p^4\epsilon\ell$ & 1.166$\times$10$^{14}$ & 1.221$\times$10$^{14}$\\
$1s^22s^22p^33p^14d^1$ β $1s^22s^22p^4\ell$ & 2.110$\times$10$^{13}$ & 2.214$\times$10$^{13}$\\
$1s^22s^22p^33s^13p^13d^1$ β $1s^22s^22p^43s^1\epsilon\ell$ & 1.894$\times$10$^{13}$ & 2.033$\times$10$^{13}$\\
$1s^22s^22p^43s^13p^14d^1$ β $1s^22s^22p^43p^1\epsilon\ell$ & 7.316$\times$10$^{11}$ & 6.621$\times$10$^{11}$\\
$1s^22s^22p^43p^2$ β $1s^22s^22p^5\epsilon\ell$ & 7.594$\times$10$^{13}$ & 8.156$\times$10$^{13}$\\
$1s^22s^22p^33s^13p^2$ β $1s^22s^22p^43s^1\epsilon\ell$ & 1.106$\times$10$^{14}$ & 1.204$\times$10$^{14}$\\
$1s^22s^22p^33s^13p^2$ β $1s^22s^22p^43p^1\epsilon\ell$ & 6.691$\times$10$^{13}$ & 7.484$\times$10$^{13}$\\\hline
\end{tabular}
\label{tab_auto}
\end{table}
\subsection{Breit interaction and QED corrections}
The Breit operator includes Coulomb repulsion, magnetic interaction and retardation in the electron-electron interaction due to finite value of the speed of light. The Breit Hamiltonian reads
\begin{equation}
h_B=\frac{1}{r_{12}}-\frac{\vec{\alpha}_1.\vec{\alpha}_2}{r_{12}}\cos\left(\omega_{12}r_{12}\right)+\left(\vec{\alpha}.\vec{\nabla}\right)_1\left(\vec{\alpha}.\vec{\nabla}\right)_2\frac{\cos\left(\omega_{12}r_{12}\right)-1}{\omega_{12}^2r_{12}},
\end{equation}
\noindent where $\alpha_i$ are the 4$\times$4 Dirac matrices, $\omega_{12}$ is the frequency of the exchange photon and the electron-electron interaction is expressed in the Coulomb (velocity) gauge \cite{grant74,grant07}.
In the present work, we consider two quantum electrodynamics (QED) corrections, responsible for the Lamb shift: vacuum polarization and self-energy (the corresponding Feynman diagrams are presented in Fig. \ref{fd}). Vacuum polarization is due to creation and annihilation of virtual electron-positron pairs in the field of the nucleus. The first term of order $\alpha(\alpha Z)$ is evaluated as the expectation value of the Uehling's potential \cite{uehling35,frolov12}. The self-energy (SE) represents the interaction of the electron with its own radiation field. In quantum field theory, this interaction corresponds to an electron emitting a virtual photon, which is then reabsorbed by the electron. In the early 1970s, Mohr proposed an atomic self-energy formulation \cite{mohr74} for hydrogenic atoms within the bound-state Furry formalism in a suitable form for numerical evaluation:
\begin{equation}\label{foncf}
E_{n\ell j}^{\mathrm{SE}}\left(\alpha Z\right)=\frac{\left(\alpha Z\right)^4}{\pi n^3\alpha}F_{n\ell j}\left(\alpha Z\right),
\end{equation}
\noindent where $F$ is a slowly varying function of $\alpha Z$. For $s$ and $p$ orbitals, $F$ is evaluated using a development in powers of $(Z\alpha)$ and $\ln(Z\alpha)$ for $Z\leq 10$ \cite{erickson77} and an interpolation in the tabulated values of Mohr \cite{mohr74,mohr75,mohr82} for $Z>10$. For $n$=3 and 4 we take the fit published by Curtis \cite{curtis85} and the results of Le Bigot \emph{et al.} \cite{lebigot01}. Calculation of many-electron radiative corrections is still one of the most difficult problems to deal with for high-precision level prediction. There have been no generalization of the self-energy calculations to arbitrary $N-$electron systems. Without exact solutions, atomic-structure codes use an approximation to the self-energy that consists of evaluating the exact hydrogenic formulas of Mohr and successors for an effective charge $Z_{\mathrm{eff}}$ in order to account for screening and multiple-electron interactions. The screening contribution to the self-energy is defined as
\begin{equation}
E_{n\ell j}^{\mathrm{SE}}\left(\alpha Z_{\mathrm{eff}}\right)-E_{n\ell j}^{\mathrm{SE}}\left(\alpha Z\right)=\frac{\alpha^3}{\pi n^3}\left(Z_{\mathrm{eff}}^4F_{n\ell j}\left(\alpha Z_{\mathrm{eff}}\right)-Z^4F_{n\ell j}\left(\alpha Z\right)\right),
\end{equation}
\noindent where $Z_{\mathrm{eff}}$ takes into account screening by the other electrons. Such an approach is not as accurate as Welton's approach \cite{welton48}, but provides satisfactory results, when the effective charges are deduced from the average radius of the $n\ell j$ orbital \cite{pain17a}. We can see in Fig. \ref{mcdf_Fe_Breit_APS2016} that the shift due to Breit and QED correction is of the order of 2 eV. Because of the width of the experimental structures, it seems difficult to assert that such a shift brings a significant improvement of the computed line energies.
\begin{figure}[h]
\vspace{2mm}
\centering
\includegraphics[width=90pt]{figure29.jpg}
\hspace{1cm}
\includegraphics[width=90pt]{figure30.jpg}
\caption{Feynman diagrams for self-energy (left) and vacuum polarization (right).}
\label{fd}
\end{figure}
\begin{figure}[h]
\vspace{6mm}
\centering
\includegraphics[width=8cm]{figure31.jpg}
\caption{Iron [Be] $2p^5 - 2p^44d$ structures of iron with and without Breit interaction and QED corrections. The calculations are perfomed with our MCDF code \cite{bruneau83,bruneau84}. SCO-RCG is quasi-relativistic and has no account for QED. Breit interaction is accounted for only partially.}
\label{mcdf_Fe_Breit_APS2016}
\end{figure}
\section{Diagnostic utility of the 3C/3D line ratio in Fe XVII: another issue about iron...}\label{3c3}
Fe XVII X-ray emission is present in the coronae of the Sun, Capella ($\alpha$ Aurigae, located at 42.2 light-years from the Earth in the Auriga (or Cocher) constellation, third brightest star in the north hemisphere after Sirius and Arcturus) and Procyon. The latest, also known as $\alpha$ Canis Minoris, is the brightest star in the constellation of Canis Minor. It is a binary star, consisting of a white main-sequence star of spectral type F5 and a white dwarf companion named Procyon B, or NGC 4635. Two of the most distinct lines observed from Fe XVII are the resonance 3C $1s^22s^22p^53d$ $^1P_1$ $\rightarrow$ $1s^22s^22p^6$ $^1S_0$ and intercombination 3D $1s^22s^22p^53d$ $^3D_1$ $\rightarrow$ $1s^22s^22p^6$ $^1S_0$ lines. Values of the relative intensity $r$ of these two lines between 1.6 and 2.8 have been measured in non-flaring active regions of the solar corona. Values in the range 2.6 to 2.8 have been measured from Capella and a value of 1.8 has been reported for Procyon. However, there is a large discrepancy between observations and theory. Table \ref{3c3d} contains the values of the oscillator strengths of 3C and 3D lines calculated with our MCDF code \cite{bruneau83,bruneau84} in the ``transition state'' approximation within two different gauges: Coulomb (velocity) and Babushkin (length) \cite{grant74,grant07}. The self-consistent-field MCDF equations are obtained by requiring that a particular functional is stationary with respect to radial wavefunctions. Such a functional involves elements of the Hamiltonian matrix weighted by coefficients, usually referred to as ``generalized weights''. In the ``Average Level'' (AL) mode, the generalized weights are chosen to be the same for all the Atomic State Functions (ASF). Therefore, resolution of the self-consistent-field equations and diagonalization of the Hamiltonian matrix can be performed separately. However, if the initial state of a transition has a much larger degeneracy than the final state, the wavefunctions will describe preferentially the initial state (and \emph{vice versa}). In that case, orbital relaxation is not taken into account and transition energies may be inaccurate. The AL method can be improved using the Slater transition-state method \cite{slater74,bruneau83}, to equilibrate the weights between the initial and final configurations.
In order to understand the discrepancy between experimental and theoretical values of $r$, a number of effects have been investigated, such as blending with an inner shell satellite line from Fe XVI, cascades from higher levels, non-Maxwellian electron distributions or configuration interaction \cite{fournier05}. The discrepancies between theory and these measurements prompted Bernitt \emph{et al.} \cite{bernitt12} to combine X-ray Free-Electron Laser (XFEL) at the Linac Coherent Light Source (LCLS) facility with an EBIT (Electron Beam Ion Trap). Well-defined X-ray pulses scanned the excitation energies of the associated states while simultaneously measuring the emission spectrum. The aim was to understand the discrepancies with the ratio measured in previous EBIT experiments. The LCLS experiment and the previous EBIT measurements are quite different in nature. The EBIT plasmas were electron-impact dominated, with many levels being simultaneously excited. The LCLS measurements were laser-driven and tuned to excite one transition at a time, producing a two-level system. The fluorescence intensity ratio of 3D and 3C was measured at $r$=2.61 $\pm$ 0.23, averaged over two independent measurement periods, after accounting for the blending of 3D with the Fe XVI line 3C. Loch \emph{et al.} \cite{loch15} performed a non-equilibrium modelling of the XFEL experiment and obtained a reduction of the predicted 3C/3D line ratio, but the discrepancy is still not fully understood.
\begin{table}[h]
\caption{Values of the oscillator strengths of 3C and 3D lines calculated with our MCDF code \cite{bruneau83,bruneau84} in the ``transition state'' approximation: $^1$ Coulomb (velocity) gauge, $^2$ Babushkin (length) gauge \cite{grant74,grant07}. The value of the ratio $r$ is not very sentitive to the choice of the gauge.}
\centering
\begin{tabular}{cccc}\hline
\textbf{Transition} & \textbf{Energy (eV)} & \textbf{Oscillator strength$^1$} & \textbf{Oscillator strength$^2$}\\\hline
3C & 811.86 & 1.820727$\times$10$^{-1}$ & 1.853965$\times$10$^{-1}$\\
3D & 826.55 & 8.297194$\times$10$^{-1}$ & 8.432271$\times$10$^{-1}$\\\hline\hline
Ratio $r$ & & 4.557 & 4.548\\\hline
\end{tabular}
\label{3c3d}
\end{table}
More generally, the emission arising from He-like iron lines and their dielectronic satellites, has been observed at low resolution in a number of clusters of galaxies. More precisely, using Gabriel's notation \cite{gabriel72}, the ratio $G=(x+y+z)/w$ of the intercombination ($x$ and $y$) and forbidden ($z$) to the resonance ($w$) lines arising from the $n$=2 level of the He-like ion, is a sensitive temperature diagnostics \cite{swartz93}.
\section{Stark effect, white dwarfs and Balmer lines}\label{whi}
The diagnostic value of Stark-broadened lines such as H$_{\beta}$ ($n$=2 to $n$=4) has long been investigated by the experimentalists. It is well known that these lines play also a significant role in inferring the plasma conditions at the photospheres of high-gravity astronomical objects \cite{dimitrijevic15}. Particularly for the white dwarfs (WD), those gravities are $\approx$ 10$^4$ times higher than that of the Sun and their correspondingly higher electron densities ($n_e\approx$ 10$^{16}$-10$^{18}$ cm$^{-3}$) lead to significant Stark broadening of the lines. Exploiting this sensitivity, astronomers use the measured widths of these lines to infer the surface gravities of these WD and thereby determine their masses. Figure \ref{AquariiSirius} (left) represents an artist view of the cataclysmic variable binary star AE Aquarii, consisting of an ordinary star in close orbit around a magnetic white dwarf. The white dwarf has 63 \% of the Sun mass, but a radius of only about 1 \% of the Sun. The white dwarf in the AE Aquarii system is the first star of its type known to give off pulsar-like pulsations that are powered by its rotation and particle acceleration. It has the shortest known spin period of any white dwarf, completing a full revolution every 33.08 seconds. Figure \ref{AquariiSirius} (right) represents Sirius A and its white dwarf companion star Sirius B. The distance between Sirius A and Sirius B varies from 8.2 to 31.5 AU (1 AU=1.49597870$\times$10$^{8}$ km). Sirius, also know as the ``Dog star'', reflecting its prominence in its constellation, Canis Major, is the brightest star in the Earth's night sky.
\begin{figure}[h]
\vspace{2mm}
\centering
\includegraphics[width=6.2cm]{figure32.jpg}
\hspace{5mm}
\includegraphics[width=4.2cm]{figure33.jpg}
\caption{Left: artist view of the cataclysmic variable binary star AE Aquarii, which consists of an ordinary star in close orbit around a magnetic white dwarf. \copyright Casey Reed (NASA). Right: Sirius A and its white dwarf companion star Sirius B (the small star on the left). \copyright NASA, ESA, H. Bond (STScI), and M. Barstow (University of Leicester).}
\label{AquariiSirius}
\end{figure}
The spectroscopic method is the most widely-used technique and is responsible for determining parameters for tens of thousands of WD. Obtaining high-accuracy estimates of those parameters is crucial for the use of WD, in order to determine the age of the universe, constrain the mass of supernovae progenitors and probe properties of dark matter axions.
Wiese \emph{et al.} measured the Balmer lines H$_{\alpha}$ , H$_{\beta}$ , H$_{\gamma}$ and H$_{\delta}$ in high-current, wall-stabilized arcs in hydrogen \cite{wiese72}. In 2009, Tremblay and Bergeron \cite{tremblay09} developed a new treatment of line broadening including non ideal effects, using the Hummer-Mihalas equation of state. They found a revision of 500-1000 K in temperature and 5-10 \% mass for the population of WD.
More recently Falcon \emph{et al.} \cite{falcon13,montgomery15} developed a platform on the Z machine at SNL to measure H$_{\beta}$ line shapes in photospheric conditions ($T\approx$ 1 eV, $n_e\approx$ 10$^{17}$ cm$^{-3}$), using the powerful X-ray capability of the machine to drive plasma formation in a gas cell. They measured hydrogen Balmer lines in emission and, for the first time, in absorption.
Up to now, SCO-RCG provided a precise Stark modeling only for hydrogen- and helium-like ions \cite{pain16}. As an example, the hydrogen $H_{\beta}$ profile computed by SCO-RCG at the conditions of Wiese \emph{et al.} is shown in Fig. \ref{white_dwarf+hbeta}, right. For more complex ions, a cruder approach is used \cite{porcherot11}.
\begin{figure}[h]
\vspace{4mm}
\centering
\includegraphics[width=7.5cm]{figure34.jpg}
\hspace{2mm}
\includegraphics[width=7.3cm]{figure35.jpg}
\caption{Left: comparison between the H$_{\beta}$ profile measured by Wiese at $T$=1.15 eV and $n_e$=8.3$\times$10$^{16}$ cm$^{-3}$ and the experimental result obtained at SNL at $n_e$=5.76$\times$10$^{16}$ cm$^{-3}$ and proton density $n_H$=1.55$\times$10$^{17}$ cm$^{-3}$ (Shot z2084). Right: calculation with our line-shape code (the electric field is chosen equal to 2$\times$10$^{-5}$ atomic units and the profile is convolved with a Lorentzian of HWHM=1.5$\times$10$^{-4}$ atomic units. The spectrum of Wiese is also represented.}
\label{white_dwarf+hbeta}
\end{figure}
It is worth mentioning that the existence of white dwarfs with a carbon \cite{dufour07} and oxygen \cite{kepler09} atmospheres was reported. We can expect that several charge states are present (not only hydrogen-like oxygen). Therefore, it is important, in that case, to be able to compute accurate Stark profiles for ions with more than two electrons \cite{gilleron}.
\clearpage
\section{Conclusion}
The detailed opacity code SCO-RCG has a great potential for astrophysical applications. It is also able to provide accurate Rosseland mean opacities and we have shown the interpretation of several laser and Z-pinch spectroscopy experiments. Large changes in stellar opacities may make it necessary to consider new or overlooked physical phenomena absent in present opacity models. This stresses the importance of laboratory experiments to guide and validate theory.
Recent measurements of the iron photo-absorption relevant to the solar interior show significant discrepancies with theoretical results. The data, in the absence of unidentified systematic experimental errors, reveal surprising unobserved photo-absorption phenomena in plasmas. A conjecture to reconcile the discrepancies is that some transitions are missing in the calculations. Another possibility is a redistribution of oscillator strengths not included in the models. A full configuration-interaction treatment, not performed in SCO-RCG calculations, may transfer oscillator strength to higher energies. As mentioned by Iglesias \cite{iglesias15b} the observed enhanced absorption seems incompatible with theory since calculations satisfy the Thomas-Reiche-Kuhn sum rule and preserve oscillator strength density and particle number. In this work, the disagreement between experiment and theory was examined, and although none of our investigations provide an explanation (density effects, highly excited states, line broadening, ...), this enigmatic spectrum remains definitely an exciting mystery and stimulates many new developments in opacity modelling and computation.
\vspace{5mm}
{\bf Acknowledgments}
The authors would like to thank all the authors of Refs. \cite{davidson88,winhart95,winhart96,dozieres15,bailey03,bailey07,dervieux15} for the experimental spectra as well as J. Colgan for providing the ATOMIC spectra.
| 2b571080f159e9f088dcc5e6d16adb6d02e9f6ea | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction} In the last few years, spintronics has emerged as a new field of research, utilizing spin currents, instead of electrons, as one of mediator for transport. It has proved to be quite promising for new types of fast electronic devices over conventional electrical charge-based semiconductor devices with advantages of high storage density, increased data processing speed and low power consumption.\cite{Graf2} To enhance the efficiency of spintronic devices, the carrier for the concerned materials should be fully spin polarized. Half-metallic ferromagnets (HMFs) are great source for highly spin polarized current and thus are ideal materials for spintronic based applications. Half-metallicity is found in magnetic materials in which one of the spin-polarized sub band has a band gap and the other exchange-split sub band has a non zero density of states at the Fermi level(E$_{F}$), thus in these materials the electrical conduction takes place from one spin channel exclusively. The discovery of half-metallicity in half-Heusler alloy NiMnSb,\cite{PhysRevLett.50.2024} by de Groot et al. in 1983 has intensified the interest in Heusler alloys for applications in the field of spintronics. Heusler family of compounds are usually of two types (i) a ternary material with stoichiometry 1:1:1 known as half-Heusler compounds (XYZ) (ii) also a ternary materials but with stoichiometry 2:1:1 known as full-Heusler compounds ($\mathrm{X_2YZ}$). There is another class of Heusler compounds which are discovered recently with stoichiometry 1:1:1:1, and are known as equiatomic quaternary Heusler alloys (XX'YZ), where X, X' and Y are transition metals and Z is a main group element.\cite{:/content/aip/journal/apr2/3/3/10.1063/1.4959093} The half-Heusler alloys crystallize in $C1_b$ structure (space group \#216, $F\bar{4}3m$) with prototype MgAgAs and the full Heusler alloys crystallize in the cubic space group $Fm\bar{3}m$ (\#225), with $\mathrm{Cu_2 MnAl}$ ($L2_1$) as prototype, whereas the βquaternaryβ Heusler alloys crystallize in cubic space group (\#216) with LiMgPdSn as prototype (Y-type).\cite{Graf20111} Among the full Heusler compounds, considerable attention has been paid to the Co based compounds because of their high spin polarization and high Curie temperature,{\cite{Graf2, :/content/aip/journal/jap/116/20/10.1063/1.4902831,PhysRevB.91.104408, :/content/aip/journal/jap/102/3/10.1063/1.2767229,:/content/aip/journal/jap/101/2/10.1063/1.2409775}} which make them more suitable for applications in spintronics. According to Julli\`{e}re model,\cite {JULLIERE1975225} the high spin polarization is very advantageous for getting high tunneling magneto-resistance ratios in magnetic tunnel junctions. The quaternary Heusler alloys CoRhMnZ (Z = Ga, Sn and Sb) have been studied in detail by both theoretical and experimental methods and were found to be HMF by ab-initio calculations.\cite{0953} CoRhMnGe, however, has not been studied from experimental front. Hence keeping in view the increased interest and applications of Co-based Heusler alloys, we present a comprehensive study of electronic, structural, magnetic and transport properties of CoRhMnGe (CRMG) Heusler alloy.
CRMG alloy is found to exist in the ordered cubic Heusler structure (Y-type) with space group $F\bar{4}3m$ (\#216). We have done a detailed local structure analysis using extended X-ray absorption fine structure (EXAFS) spectroscopy. The saturation magnetization value $M_s$ at 5 K is found to be 4.9 $\mu_{B}/f.u.$ which is in close agreement with the value predicted by Slater-Pauling rule (5 $\mu_{B}/f.u.$) for half-metallic materials. The Curie temperature ($\mathrm{T_{C}}$) was found to be $\sim 760$ K which is highest among the reported CoRhMnZ (Z = Ga, Sn, Sb) alloy.\cite {0953} We have also performed a systematic ab-initio calculations to study the electronic structure, magnetism, mechanical and transport properties of the alloy. Ab-initio simulations also predict half-metallic nature for this alloy. Total energy and lattice dynamics calculations suggest that the alloy is chemically and mechanically stable against external pressure. Effect of pressure on magnetism and half metallicity is studied at length.
\section{Experimental Details}
\subsection{Sample Synthesis}
The polycrystalline alloy CRMG was prepared by arc melting the stoichiometric amounts of constituent elements (at least 99.9\% purity) in water cooled copper hearth under high purity argon atmosphere. To further reduce the contamination a Ti ingot was used as an oxygen getter. 2\% extra Mn was taken to compensate the weight loss due to Mn evaporation during melting. The ingot formed was flipped and melted several times for better homogeneity, The final weight loss was less than 1\%.
\subsection{Characterization}
X-ray diffraction (XRD) pattern was taken at room temperature using Xβpert pro diffractometer with $\mathrm{Cu-K\alpha}$ radiation to study the crystal structure of the sample. XRD analysis was done with the help of FullProf suite which uses the least square refinement between the experimental and calculated intensities. Rietveld method is used to optimize the $\chi$-square function given by:
\begin{equation}
\chi^2= w_i\Sigma_i{(y_{iobs}-y_{ical})^2}
\end{equation}
where $w_i$ is the inverse of the variance associated with the $i^{th}$ observation i.e $\sigma^2(y_{iobs})$ and $y_{iobs}$ and $y_{ical}$ are the observed and calculated scattering intensities for a diffraction angle $2\theta_i$.\cite{RR}
It also contains GFourier program, which is used to calculate and visualize the electron density within the unit cell.
EXAFS (Extended X-ray Absorption Fine Structure) measurements on CRMG were done to probe the local structure surrounding the Co, Ge and Mn sites. The X-ray absorption spectroscopy (XAS) measurements have been carried out at the Energy-Scanning EXAFS beam-line (BL-9) in transmission mode at the INDUS-2 Synchrotron Source (2.5 GeV, 200 mA) at Raja Ramanna Centre for Advanced Technology (RRCAT), Indore, India.\cite{2014AIPC.1591..649P,1742-6596-493-1-012032} This beam-line operates in the energy range of 4-25 keV. The beam-line optics consists of a Rh/Pt coated collimating meridional cylindrical mirror and the collimated beam reflected by the mirror is monochromatized by a Si(111) (2d=6.2709 \AA ) based double crystal monochromator (DCM). The second crystal of the DCM is a sagittal cylinder used for horizontal focusing while a Rh/Pt coated bendable post mirror facing down is used for vertical focusing of the beam at the sample position. Three ionization chambers (300 mm length each) were used for data collection in the transmission mode; one for measuring incident flux ($I_0$), one for measuring transmitted flux ($I_T$) and the third for measuring EXAFS spectrum of a reference metal foil for energy calibration. Appropriate gas pressure and gas mixture have been chosen to achieve 10-20\% absorption in first ionization chamber and 70-90\% absorption in second one to improve the signal to noise ratio. Rejection of the higher harmonics content in the X-ray beam was performed by the second mirror. The absorption coefficient, $\mu$ was obtained using the relation:
\begin{equation}
I_T = I_0 e^{-\mu x}
\end{equation}
where, $x$ is the thickness of the absorber. Powder samples of appropriate weight, estimated to obtain a reasonable edge jump, were taken and was mixed thoroughly with cellulose powder to obtain total weight of 100 mg. Subsequently, homogeneous pellets of $15$ mm diameter were prepared using an electrically operated hydraulic press. However, the grain size of the particles was of the order of 50 micron, which resulted in very bad data. The sample was later grounded continuously using a mortar-pestle for 3 hours to reduce the particle size to less than 4 microns. This fine powder was dispersed on the scotch tape and larger particles are brushed out. A reasonable edge jump was obtained by folding the scotch tape.
Magnetization isotherms at 5 K and 300 K were obtained using a vibrating sample magneto meter (VSM) attached to the physical property measurement system (PPMS) (Quantum design) for fields up to 50 kOe. Thermo-magnetic curves in the high temperature region (300 K -1000 K) were taken in VSM with attached high temperature oven under a field of 100 Oe. Electrical resistivity measurements were done using four-probe method in PPMS.
\section{Computational Details}
Ab-initio calculations were performed to study the electronic structure, magnetic and mechanical properties using density functional theory (DFT) implemented within Vienna ab-initio simulation package (VASP)\cite{VASP} with a projected augmented wave (PAW) basis.\cite{PAW} The electronic exchange-correlation potential due to Perdew-Bueke-Ernzerhof (PBE) is used within generalized gradient approximation (GGA) scheme. A $24^3$ {\bf k}-mesh is used to perform the Brillouin zone integration within the tetrahedron method. A plane wave energy cut-off of $269$ eV is used for all the calculations. ALL the structure are fully relaxed, with total energies (forces) converged to values less than 10$^{-6}$ eV ($0.01$ eV/\r{A}).
Transport properties are calculated using Boltzmann transport theory implemented within the BoltzTrap code\cite{Madsen200667} under constant relaxation time approximation of the charge carriers. These calculations are performed using experimental lattice parameter to make a direct comparison with the measured resistivity.
Simulating accurate transport properties usually require a much finer k-mesh as well as plane-wave energy cut-off. As such $46^3$ k-mesh along with a plane wave cut-off of 500 eV are used for transport study. Because, simulated electrical conductivity within BoltzTrap are more reliable at relatively higher temperature ($> 100$ K), we compared our calculated results with those of experiment between the range $100-400$ K.
\begin{figure}[t]
\centering
\includegraphics[scale=0.75]{fig1.pdf}
\caption{Relative energy ($\Delta$E$_{rel.}$) of different non-degenerate atomic configurations with respect to Type3 configuration, calculated at experimental (a$_{elp}$) and relaxed (a$_{rlp}$) lattice parameters. Three distinct configurations, Type1, Type2 and Type3 are shown by the primitive unit cell (a), (b) and (c) respectively and are detailed in the text. }
\label{fig1}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{fig2.pdf}
\caption{Band Structure and the Density of states (DoS) for CRMG at a$_{rlp}$. The system shows a halfmetallic characteristics having a finite DoS in spin up channel at E$_{F}$ and a non-zero indirect band gap (($\Delta$E$_{\text{g}})_{\downarrow} = 0.23$ eV) in the spin down channel at E$_{F}$. Dashed line represents the Fermi level (E$_F$).}
\label{fig3}
\end{figure}
\section{Results and Discussion}
\subsection{Structural stability}
\label{struc-stability}
\par
Quaternary Heusler alloys exhibit LiMgPdSn-type crystal structure (\# $216$) whose primitive cell contains four atoms at the Wyckoff positions, 4a(0, 0, 0), 4b(1/2, 1/2, 1/2), 4c(1/4, 1/4, 1/4) and 4d(3/4, 3/4, 3/4). The preferred occupation depends upon the electronegativities and sizes of individual atoms. Hence, we have calculated the total energies corresponding to different atomic configurations. We have found three non-degenerate atomic configurations, obtained by interchanging constituent atoms at various Wyckoff positions such as Type1, Type2 and Type3 with decreasing stability. The energetics and corresponding configurations are shown in Fig. {\ref{fig1}}. The most stable configuration (Type1) corresponds to the occupation: Co at 4d, Rh at 4c, Mn at 4b and Ge at 4a Wyckoff sites. On the other hand in Type2 configuration, Co sits at 4b, Rh at 4d, Mn at 4c and Ge at 4a sites. The least stable structure, Type 3, is formed when Co occupies at 4d, Rh at 4b, Mn at 4c and Ge at 4a Wyckoff sites respectively.
\subsection{Electronic structure, magnetic and mechanical properties}
\par
In this section, spin polarized electronic structure, magnetic properties and the mechanical stability of CRMG alloy are presented using {\it{ab-initio}} simulations.
Figure \ref{fig3} shows spin polarized band structure and density of states calculated at theoretically relaxed lattice parameter (a$_{rlp} = 5.92 \mathrm{\AA}$). CRMG clearly shows an indirect band gap of $0.23$ eV from $\Gamma \rightarrow X$ in the minority spin channel and finite DoS at E$_{F}$ in the majority spin channel leading to the ideal 100$\%$ spin polarized current in the system.
Experimental lattice parameter (a$_{elp}$) is measured to be 5.89 $\mathrm{\AA}$ which is not very different from that of a$_{rlp}$. Calculated band structure at a$_{elp}$ looks very similar to that of Fig. \ref{fig3} with a band gap of $0.23$ eV.
The calculated total magnetic moment (m$_{tot}$) is $4.94\ \mu_{B}/f.u.$ at a$_{rlp}$, which follows Slater-Pauling rule, {\cite {PhysRevB.83.184428,PhysRevB.66.174429,Slat1,Paul1}} according to which, $m = (N_v - 24) \mu_B$,
where $N_v$ is the number of valence electrons per unit cell. Our experimentally measured total moment is $4.9\ \mu_{B}/f.u.$. We have also calculated the Curie temperature (T$_C$) using the model presented in Refs. \onlinecite{0022-3727-40-6-S01, doi:10.1063/1.3062812}. The calculated T$_C$ comes out to be $915$ K ($917$ K) with a$_{elp}$ (a$_{rlp}$), which compares fairly well with those measured by us ($\sim$ $760$ K).
We have also checked the mechanical stability of the alloy by performing lattice dynamics calculations. The linear response parameter e.g. elastic constants (C$_{ij}$) are nothing but the proportionality factor of stress and strain of the crystal under applied force. Calculated values of $C_{ij}$ for CRMG at equilibrium lattice constant (a$_{rlp}$) are C$_{11} = 259.83$, C$_{12} = 167.41$, C$_{44} = 98.00$ N/m$^2$. The condition for mechanical stability of cubic crystals, the so called Born-Huang criteria,{\cite{BHC}} is defined as,
\[
C_{11} > 0, \; C_{44} > 0, \; C_{11} > C_{12}, \; (C_{11}+2C_{12}) > 0,
\]
These criteria very well hold in the present case and hence CRMG is mechanically stable. The other elastic moduli such as Bulk modulus (B), shear modulus (G), Young's modulus (Y) and the anisotropy factor (A) are also calculated and are given by $198.22$, $72.83$, $194.65$ and $2.12$ respectively.
\begin{table}[t]
\begin{ruledtabular}
\begin{tabular}{|c| c| c| c| c| c|}
a (\AA) & Pressure & n$_{\uparrow}$(E$_{F}$)& n$_{\downarrow}$(E$_{F}$) & ($\Delta$E$_{g}$)$_{\downarrow}$
& Nature \\
\hline
5.00 & 327.08 & 1.40 & 2.10 & no-gap & Metallic\\
\hline
5.10 & 252.96 & 1.13 & 1.92 & no-gap & Metallic\\
\hline
5.20 & 193.39 & 0.92 & 1.86 & no-gap & Metallic\\
\hline
5.30 & 145.24 & 0.90 & 1.86 & no-gap & Metallic\\
\hline
5.40 & 106.54 & 0.87 & 1.48 & no-gap & Metallic\\
\hline
5.50 & 75.32 & 0.87 & 1.04 & no-gap & Metallic\\
\hline
5.60 & 50.30 & 0.89 & 0.49 & no-gap & Metallic\\
\hline
5.70 & 30.27 & 0.94 & 0.00 & 0.32 & HM\\
\hline
5.80 & 14.57 & 1.00 & 0.00 & 0.28 & HM\\
\hline
5.85 & 8.09 & 1.03 & 0.00 & 0.26 & HM\\
\hline
5.89(a$_{elp}$) & 3.49 & 1.05 & 0.00 & 0.25 & HM\\
\hline
5.92(a$_{rlp}$) & 0.00 & 1.07 & 0.00 & 0.23 & HM\\
\hline
6.00 & -6.86 & 1.12 & 0.05 & no-gap & Metallic\\
\end{tabular}
\end{ruledtabular}
\caption{Effect of pressure on the electronic properties of CRMG. Upto 5.70\AA\ (pressure of 30.27 GPa), system retains its halfmetallic characteristics with the variable band-gap (($\Delta$E$_{g}$)$_{\downarrow}$) and DoS (n$_{\uparrow}$(E$_{F}$) and n$_{\downarrow}$(E$_{F}$)). A phase transition from halfmetallic (HM) to metallic occurs below 5.70\AA (i.e. above 30.27 GPa). Band gap, DoS and pressure are measured in eV, states/eV-f.u. and GPa respectively.}
\label{tab1}
\end{table}
\begin{table}[t]
\begin{ruledtabular}
\begin{tabular}{|c| c| c| c| c| c|}
\hline
a (\AA) & m$_{Co}$ & m$_{Rh}$ & m$_{Mn}$
& m$_{Ge}$ & m$_{tot}$\\
\hline
5.00 & 0.69 & 0.38 & 2.44 & 0.02 & 3.52 \\
\hline
5.10 & 0.76 & 0.39 & 2.65 & 0.01 & 3.80 \\
\hline
5.20 & 0.87 & 0.42 & 2.81 & 0.01 & 4.12 \\
\hline
5.30 & 0.97 & 0.45 & 2.95 & 0.01 & 4.38 \\
\hline
5.40 & 1.06 & 0.48 & 3.06 & 0.01 & 4.60 \\
\hline
5.50 & 1.14 & 0.49 & 3.15 & 0.00 & 4.77 \\
\hline
5.60 & 1.20 & 0.50 & 3.21 & -0.01 & 4.89 \\
\hline
5.70 & 1.20 & 0.47 & 3.26 & -0.01 & 4.91 \\
\hline
5.80 & 1.20 & 0.44 & 3.31 & -0.02 & 4.92 \\
\hline
5.85 & 1.20 & 0.42 & 3.33 & -0.03 & 4.93 \\
\hline
5.89(a$_{elp}$) & 1.20 & 0.41 & 3.36 & -0.03 & 4.93 \\
\hline
5.92(a$_{rlp}$) & 1.20 & 0.39 & 3.37 & -0.03 & 4.94 \\
\hline
6.00 & 1.20 & 0.37 & 3.42 & -0.04 & 4.95 \\
\hline
\end{tabular}
\end{ruledtabular}
\caption{Pressure effect on the atom projected and total magnetic moment (in Bohr magneton unit) of CRMG. All the constituent transition metal elements are ferromagnetically aligned. Notably, the system retains its ferromagnetic characteristics through out the pressure range i.e. there is no magnetic phase transition.}
\label{tab2}
\end{table}
\begin{table*}
\begin{ruledtabular}
\begin{tabular}{|c| c| c| c| c| c| c| c| c|}
\hline
a (\AA) & C$_{11}$ & C$_{12}$ & C$_{44}$ & B & G & Y & A & Born-Huang Criteria\\
\hline
5.00 & 939.74 & 1043.46 & 480.81 & 1008.88 & -215.75 & -696.94 & -9.27 & $\text{\sffamily X}$\\
\hline
5.10 & 821.35 & 899.80 & 418.18 & 873.65 & -113.98 & -357.49 & -10.66 & $\text{\sffamily X}$\\
\hline
5.20 & 695.08 & 754.68 & 360.70 & 734.81 & -60.94 & -188.02 & -12.10 & $\text{\sffamily X}$\\
\hline
5.30 & 612.80 & 613.87 & 309.33 & 613.51 & 66.25 & 191.85 & -578.19 & $\text{\sffamily X}$\\
\hline
5.40 & 561.04 & 494.75 & 264.21 & 516.85 & 118.96 & 331.46 & 7.97 & $\checkmark$\\
\hline
5.50 & 513.71 & 399.18 & 225.00 & 437.36 & 131.85 & 359.43 & 3.93 & $\checkmark$\\
\hline
5.60 & 456.75 & 320.94 & 189.77 & 366.21 & 126.70 & 340.79 & 2.79 & $\checkmark$\\
\hline
5.70 & 392.69 & 263.83 & 156.95 & 306.78 & 110.51 & 295.99 & 2.44 & $\checkmark$\\
\hline
5.80 & 327.07 & 215.85 & 128.44 & 252.92 & 92.33 & 246.93 & 2.31 & $\checkmark$\\
\hline
5.85 & 298.53 & 195.20 & 115.58 & 229.64 & 84.12 & 224.91 & 2.24 & $\checkmark$\\
\hline
5.89(a$_{elp}$) & 277.01 & 179.72 & 105.88 & 212.15 & 77.89 & 208.20 & 2.18 & $\checkmark$\\
\hline
5.92(a$_{rlp}$) & 259.83 & 167.41 & 98.00 & 198.22 & 72.83 & 194.65 & 2.12 & $\checkmark$\\
\hline
6.00 & 224.37 & 142.13 & 81.56 & 169.54 & 62.22 & 166.32 & 1.98 & $\checkmark$ \\
\hline
\end{tabular}
\end{ruledtabular}
\caption{Effect of pressure on elastic constants (C$_{ij}$), Bulk modulus (B), Shear modulus (G), Young's modulus (Y) and anisotropy factor (A) for CRMG. The alloy is mechanically stable upto $5.40\ \mathrm{\AA}$, satisfying Born-Huang criteria. Any further decrease in $a$ (increase in pressure) leads instability. The symbol, $\checkmark$($\text{\sffamily X}$) represents whether the Born-Huang criteria is satisfied or not. All the elastic parameters are in GPa unit. }
\label{tab3}
\end{table*}
\subsection{Pressure effect on electronic structure and mechanical stability}
\par
Due to high value of spin polarization, the half-metallic materials are frequently used in spintronic devices in the form of thin films or multilayer. The lattice parameters of pristine bulk materials usually change when the films or multilayer are grown on appropriate substrates. As such, it is important to check the effect of lattice parameter variation (or pressure) on the electronic structure of bulk material which can give some hints for the surface properties.
Effect of hydrostatic pressure on CRMG alloy has been investigated theoretically by reducing the value of lattice constant from its equilibrium value. Pressure effect on electronic structure, magnetism and mechanical properties are summarized in Table~ \ref{tab1}, \ref{tab2} and \ref{tab3} respectively. The alloy retains its halfmetallic characteristics in a limited range of pressure (upto 5.70$\mathrm{\AA}$ which corresponds to 30.27 GPa pressure) with variable band-gap and DoS value in the minority and majority spin channels at E$_{F}$. Below 3.2$\%$ reduction of lattice parameter (i.e. $<$5.70$\mathrm{\AA}$) with respect to a$_{elp}$, DoS in the minority spin channel grows tremendously and resulted in a complete suppression of the band gap and consequently a phase transition from halfmetallic to metallic state (Table I).
There is no magnetic transition seen in the considered range of pressure and the system retains its ferromagnetic characteristics, but with decreasing magnitude in m$_{tot}$ (Table II). Thus, the system is robust against magnetic transition.
Table III summarizes the effect of pressure on the mechanical properties of CRMG alloy. Elastic constant, C$_{ij}$ and various elastic moduli are affected very significantly. As the pressure increases, the elastic moduli B, G, Y and A reduces. Upto $5.40\ \mathrm{\AA}$(i.e. $106.5$ GPa), CRMG satisfies the Born-Huang criteria but below $5.40\ \mathrm{\AA}$, the alloy becomes mechanically unstable.
\begin{figure}[b]
\centering
\includegraphics[width=0.9\linewidth]{fig3.pdf}
\caption{Rietveld refined XRD pattern of CRMG alloy. $\mathrm{Y_{obs}}$ and $\mathrm{Y_{calc}}$ are the observed and calculated scattering intensities. }
\label{XRD}
\end{figure}
\begin{figure}[b]
\centering
\includegraphics[width=0.8\linewidth]{fig4.pdf}
\caption{Measured Crystal Structure of CRMG alloy }
\label{fig5}
\end{figure}
\subsection{X-ray diffraction}
Figure {\ref{XRD}} shows the Rietveld refinement of the room temperature XRD pattern using Fullprof suite. It is clear from the pattern that the alloy CRMG exhibits a cubic structure. The quaternary Heusler alloys exhibit LiMgPdSn-type structure whose primitive cell contains four atoms on the wyckoff positions 4a, 4b, 4c and 4d. Out of various possibilities for the occupancy of four atoms at different sites, only three atomic configurations turn out to be energetically distinct for this structure.\cite{PhysRevB.92.224413,0953} Due to the underlying symmetry of $F\bar{4}3m$ (\#216) space group, interchange of atoms between 4a and 4b or 4c and 4d Wyckoff positions does not change the overall symmetry and hence the total energy of the compound. The presence of super-lattice reflections in the XRD pattern confirms CRMG alloy to crystallize in cubic Heusler structure. The lattice parameter was found to be 5.89\ $\mathrm{\AA}$, which compares fairly well with theory (a$_{rlp}=5.92~ \mathrm{\AA}$). The best fit of the observed intensities was obtained when Co, Mn, Rh and Ge atoms were assigned the Wyckoff positions 4d, 4b, 4c and 4a respectively, which corresponds to a Type1 structure (Fig. \ref{fig1}). The measured crystal structure of CRMG alloy is shown in Fig. \ref{fig5}. The super-lattice reflections are not very intense, which can be due to some disorder in the system. In CRMG, X(Co), Y(Mn) and Z(Ge) atoms are from the same period and their atomic scattering factors are nearly identical. In such cases, it is difficult to find out the extent of disorder using XRD. EXAFS is helpful in determining the short range chemical environment around the atoms. Thus, to further probe the local surroundings of Co, Ge and Mn site, EXAFS measurements have been performed on CRMG alloy, as discussed later.
The electronic densities of the constituent elements at different vertical cut are shown in Fig. \ref{fig:CRMGElectrondensityfinal}(a-d), which is generated from the XRD refinement using GFourier program in Fullprof suite. Figure \ref{fig:CRMGElectrondensityfinal}(c) and Fig. \ref{fig:CRMGElectrondensityfinal}(d) clearly suggest that most of the charge is distributed around Rh site. Co and Mn are surrounded by intermediate charge while Ge has a charge density which is less than that of Rh but greater than that of Co and Mn atomic sites. Thus the position of heavier element in the unit cell is reflected by the most dense electron density contour.
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{fig5.pdf}
\caption{Electronic density of individual atoms in the unit cell of CRMG alloy at (a) z/c = 0, (b) z/c = 0.25, (c) z/c = 0.5 and (d) z/c = 0.75}
\label{fig:CRMGElectrondensityfinal}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=0.9\linewidth]{fig6.pdf}
\caption{Normalized EXAFS spectra of (a) Co, (b) Mn and (c) Ge of CRMG sample. The insets of (a) and (b) show the XANES spectrum of CRMG at Co and Mn K-edge (red line) along with the corresponding metal foil reference (black line).}
\label{fig:EXAFS1}
\end{figure}
\subsection{EXAFS Analysis}
Figure \ref{fig:EXAFS1} shows the normalized EXAFS spectra of (a) Co (b) Mn and (c) Ge for CRMG alloy. The insets of (a) and (b) show the XANES (X-ray absorption near edge structures) spectrum of CRMG at Co and Mn K-edge (red line). We have checked the possibility of oxidation of the samples by comparing the XANES spectra of the samples with that of standard metal foils and commercial powders of respective oxides. The spectra of pure metal foil reference is shown by black line in the insets. We found that the edge positions and the XANES spectra of CRMG sample agree fairly well with those of the pure metal foils and do not resemble that of the oxides, thus ruling out the presence of oxides. In order to take care of the oscillations in the absorption spectra, the absorption function $\chi(E)$ is calculated as follows \cite{Dc}
\begin{equation}
\chi(E) = \frac{\mu(E)-\mu(E_0)}{\Delta{\mu(E_0)}}
\end{equation}
Here $E_0$ is absorption edge energy, $\Delta{\mu(E_0)}$ is the bare atom background and $\mu(E)$ is the step in value at the absorption edge. The $k$-dependent absorption coefficient $\chi(k)$ is then calculated using the relation,
\begin{equation}
k = \sqrt{\frac{2m(E-E_0)}{\hbar^2}}
\end{equation}
where, m
is the electronic mass. $\chi(k)$ is then weighted by $k^3$ to amplify the oscillation at high k. The $k^3*\chi(k)$ spectra for Co, Mn and and Ge k-edge are shown in Fig. {\ref{fig:EXAFS2}}. The $k^3 \chi(k)$ functions are also Fourier transformed in the R-space to generate the
$\chi(R)$ spectra in terms of the real distances from the center of the absorbing atom.
The set of EXAFS data analysis programme available within IFEFFIT software package has been used for final data analysis.\cite{NEWVILLE1995154} This includes background reduction and Fourier transform to derive the $\chi(R)$ versus R
spectra from the absorption spectra (using ATHENA software), generation of the theoretical EXAFS spectra starting from an assumed crystallographic structure and finally fitting of experimental data with the theoretical spectra using ARTEMIS software.
\begin{figure}
\centering
\includegraphics[width=1.0\linewidth]{fig7.pdf}
\caption{$k^3*\chi(K)$ vs. k spectra for Co, Mn and Ge of k- edge of CRMG.}
\label{fig:EXAFS2}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[width=1.0\linewidth]{fig8.pdf}
\caption{Fourier transformed EXAFS spectra $\chi(R)$ of CRMG at (a) Co, (b) Mn and (c) Ge K-edge (scatter points) and theoretical fit (red solid line).}
\label{fig:EXAFS3}
\end{figure}
The $\chi(R)$
versus R spectra generated (Fourier transform range is $k=3.0-10.0$ $\mathrm{\AA^{-1}}$ for the CRMG sample) using the methodology described above are shown in Fig. \ref{fig:EXAFS3}
for CRMG measured at Co, Mn and Ge K-edge. The fitting strategy adopted here was to simultaneously fit multiple data set with multiple edges.\cite{Bainsla2015509,PhysRevB.65.184431} Co, Mn and Ge K edges are fitted simultaneously with common fitting parameters. This fitting reduces the number of independent parameters below the Nyquist criteria and enhances the statistical significance of the fitting model. The goodness of fit has been determined by the value of the $R_{factor}$ defined by:
\small
\begin{equation}
R_{factor}=\sum_{i} {\frac {{[\Im m(\chi_{dat}(r_i)-\chi_{th}(r_i)]^2}+{{[Re(\chi_{dat}(r_i)-\chi_{th}(r_i)}]^2}}{{[\Im m(\chi_{dat}(r_i))]^2}+{[Re(\chi_{dat}(r_i))]^2}}}
\end{equation}
\normalsize
where, $\chi_{dat}$ and $\chi_{th}$ refer to the experimental and theoretical $\chi(r)$ values respectively and $Re$ and $\Im m$ refer to the real and imaginary parts of the respective quantities. In addition to the single scattering paths, multiple scattering paths are also used for the fitting purpose.
The structural parameters (atomic coordination and lattice parameters) of CRMG used for simulation of theoretical EXAFS spectra of the sample have been obtained from XRD results. The best fitted $\chi(R)$ versus $R$
spectra (fitting range R=1.2-5.0 $\mathrm \AA$) of the sample are shown in Fig. \ref{fig:EXAFS3}
along with the experimental data for the measurements carried out at Co, Mn and Ge K-edges. The bond distances, co-ordination numbers (including scattering amplitudes) and disorder (Debye-Waller) factors $(\sigma^2)$, which give the mean square fluctuations in the distances, have been used as fitting parameters. All the three spectra are fitted without site anti-site disorder. Thus, the structural characterization of the CRMG performed using EXAFS analysis reveals a well ordered structure for CRMG (no antisite disorder). All the relevant fitting parameters are shown in Table \ref{tab4}.
\begin{table*}
\centering
\caption{Bond length(R), coordination number(N) and Debye-Waller or disorder factor ($\sigma^2$) obtain by EXAFS fitting for CRMG at Co, Mn and Ge K edge}
\begin{tabular}{|c|c|c|c|c|c|c|c|c|c|c|c|}
\hline
\multicolumn{4}{|c|}{\textbf{Co edge}}&
\multicolumn{4}{c}{\textbf{Mn edge}}&
\multicolumn{4}{|c|}{\textbf{Ge edge}} \\
\hline \rule[-2.5ex]{0pt}{5.5ex} \hspace*{.3cm}{\textbf{Path}} \hspace*{0.17cm}&\hspace*{.17cm}{\textbf{R(\AA)}}\hspace*{0.17cm}&\hspace*{.17cm}{\textbf{N}}\hspace*{0.17cm}&\hspace*{.17cm}{$\boldsymbol\sigma^2$}\hspace*{0.17cm}&\hspace*{.17cm}{\textbf{Path}}\hspace*{0.17cm}&\hspace*{.17cm}{\textbf{R(\AA)}}\hspace*{0.17cm}&\hspace*{.17cm}{\textbf{N}}\hspace*{0.17cm}&\hspace*{.17cm}{$\boldsymbol\sigma^2$}\hspace*{0.17cm}&\hspace*{.17cm}{\textbf{Path}}\hspace*{0.17cm}&\hspace*{.17cm}{\textbf{R(\AA)}}\hspace*{0.17cm}&\hspace*{.17cm}{\textbf{N}}\hspace*{0.17cm}&{$\boldsymbol\sigma^2$} \\
\hline \rule[-2ex]{0pt}{5.5ex} $\boldsymbol{\mathrm{Co-Ge}}$ &$2.40\pm0.01$ &4 &$0.0085\pm0.001$ &$\boldsymbol{\mathrm{Mn-Ge}}$ &$2.40\pm0.01$ &4 &$0.001\pm0.0008$ &$\boldsymbol{\mathrm{Ge-Co}}$ &$2.39\pm0.01$ &4 &$0.02\pm0.002$ \\
\hline \rule[-2ex]{0pt}{5.5ex} $\boldsymbol{\mathrm{Co-Rh}}$ &$2.40\pm0.01$ &4& $0.035\pm0.002$& $\boldsymbol{\mathrm{Mn-Rh}}$& $2.40\pm0.01$& 4& $0.001\pm0.0008$& $\boldsymbol{\mathrm{Ge-Mn}}$& $2.39\pm0.01$& 4& $0.0057\pm0.001$ \\
\hline \rule[-2ex]{0pt}{5.5ex} $\boldsymbol{\mathrm{Co-Mn}}$ &$2.92\pm0.01$ &6 &$0.0142\pm0.001$ &$\boldsymbol{\mathrm{Mn-Co}}$ &$2.89\pm0.02$ &6 &$0.0167\pm0.001$ &$\boldsymbol{\mathrm{Ge-Rh}}$ &$2.92\pm0.01$ &6 &$0.0199\pm0.0008$ \\
\hline \rule[-2ex]{0pt}{5.5ex} $\boldsymbol{\mathrm{Co-Co}}$ &$4.05\pm0.02$ &12 &$0.0255\pm0.0009$ &$\boldsymbol{\mathrm{Mn-Mn}}$ &$4.05\pm0.01$ &12 &$0.0321\pm0.001$ &$\boldsymbol{\mathrm{Ge-Ge}}$ &$4.11\pm0.03$ &12 &$0.0128\pm0.006$ \\
\hline \rule[-2ex]{0pt}{5.5ex} $\boldsymbol{\mathrm{Co-Ge}}$ &$4.60\pm0.03$ &12 &$0.0098\pm0.001$ &$\boldsymbol{\mathrm{Mn-Ge}}$ &$4.58\pm0.02$ &12 &$0.004\pm0.001$ &$\boldsymbol{\mathrm{Ge-Co}}$ &$4.89\pm0.05$ &12 &$0.0039\pm0.001$ \\
\hline \rule[-2ex]{0pt}{5.5ex} $\boldsymbol{\mathrm{Co-Rh}}$ &$4.60\pm0.03$& 12& $0.0077\pm0.0007$& $\boldsymbol{\mathrm{Mn-Rh}}$& $4.63\pm0.02$& 12& $0.0043\pm0.001$& $\boldsymbol{\mathrm{Ge-Mn}}$& $4.89\pm0.05$& 12& $0.0026\pm0.001$ \\
\hline
\end{tabular}
\label{tab4}
\end{table*}
\begin{figure}[b]
\centering
\includegraphics[width=1.0\linewidth]{fig9.pdf}
\caption{Magnetization curves for CRMG alloy at 5 K and 300 K.}
\label{fig:CRMG_MH}
\end{figure}
In the Fourier transformed EXAFS spectrum of Co K-edge (Fig. \ref{fig:EXAFS3}(a)), the main peak near 2.1 $\mathrm{\AA}$ has contribution from Co-Ge (2.40 $\mathrm{\AA}$), Co-Rh (2.40 $\mathrm{\AA}$) and Co-Mn (2.92 $\mathrm{\AA}$) paths. The small peak around 3.5 $\mathrm{\AA}$ at Co K-edge spectrum is the contribution of Co-Co (4.05 $\mathrm{\AA}$) coordination and a small contribution of higher bond length paths. The third peak at 4.25 $\mathrm{\AA}$ is the contribution from Co-Ge (4.60 $\mathrm{\AA}$) and Co-Rh (4.60 $\mathrm{\AA}$) coordination shells with multiple scattering paths. The use of multiple scattering paths does not include any further independent parameters as all the multiple scattering parameters are defined in terms of single scattering path parameters.\cite{PhysRevB.65.184431} In case of Ge K-edge (Fig. \ref{fig:EXAFS3}c), the first peak has the contribution from Ge-Co (2.39 $\mathrm{\AA}$), Ge-Mn (2.39 $\mathrm{\AA}$) and Ge-Rh (2.92 $\mathrm{\AA}$) paths. The peak at 3 $\mathrm{\AA}$ could not be fitted well as there is no coordination shell present at this distance. The peak at 3.6 $\mathrm{\AA}$ is the contribution from Ge-Ge (4.11 $\mathrm{\AA}$) path. The double peak between 4 $\mathrm{\AA}$ and 5 $\mathrm{\AA}$ is fitted with Ge-Co (4.89 $\mathrm{\AA}$), Ge-Mn (4.89 $\mathrm{\AA}$) and multiple scattering paths. Similarly, in case of Mn K-edge (Fig. \ref{fig:EXAFS3}(b)) the first peak has the contribution from Mn-Ge (2.40 $\mathrm{\AA}$), Mn-Rh (2.40 $\mathrm{\AA}$) and Mn-Co (2.89 $\mathrm{\AA}$). The region between 3.5 $\mathrm{\AA}$ and 4.5 $\mathrm{\AA}$ is fitted with Mn-Mn, Mn-Ge and Mn-Rh coordination shells with multiple scattering paths.
\subsection{Magnetic Properties}
\begin{figure}[b]
\centering
\includegraphics[width=1.0\linewidth]{fig10.pdf}
\caption{Temperature dependence of magnetization at 100 Oe. $T_c$ is calculated from the minima of the first order derivative of $M$ vs. $T$ curve.}
\label{fig:CRMGMT}
\end{figure}
\begin{figure*}[t]
\centering
\includegraphics[width=0.9\linewidth]{fig11.pdf}
\caption{(A) Temperature dependence of electrical resistivity($\rho$) for CRMG at zero field. The raw data is fitted with two functions, a general power law and a quadratic fit. Inset (a) shows a closer look of the $\rho$ curve between $0-100$ K. Inset (b) shows $\rho$ vs. T at $0$ and $50$ kOe. (B) (top) Comparison of measured $\rho$ (red line) with those simulated (solid circle). Temperature dependence of the carrier relaxation time ($\tau$) (bottom). }
\label{fig:Res-fits-final}
\end{figure*}
\begin{table*}
\centering
\caption{Fitting parameters for zero field electrical resistivity vs. $T$ for CRMG alloy. Two different fitting procedure involve a general power law and a quadratic fit. Residual resistivity ($\rho_0$) in both the case is same i.e. $\rho_0=48.16\ \mu\Omega$.}
\begin{tabular}{|c|c|c|c|}
\hline & & \\
$\rho(T) = \rho_0 + \rho_n\:T^n$ &
$\rho_n = 0.0047(1)$$\mu\Omega \mathrm{\:cm\:K^{-n}}$ & $n=1.535(4)$ \\
\hline & &\\
$\rho(T) = \rho_0+ \rho_1 T + \rho_2 T^2$ & $\rho_1 = 0.0496(4)$ $\mu\Omega \mathrm{\:cm\:K^{-1}}$ & $ \rho_2 = [1.742(1)]\times{10^{-4}}$$\mu\Omega\:\mathrm{cm\:K^{-2}}$ \\
\hline
\end{tabular}
\label{tab5}
\end{table*}
Figure \ref{fig:CRMG_MH} shows the isothermal magnetization curves for the CRMG alloy at 5 K and 300 K. The absence of hysteresis shows the soft magnetic nature of the alloy. The saturation magnetization values ($M_s$) at 5 K and 300 K were found to be 4.9 $\mu_B/f.u.$ and 4.7 $\mu_B/f.u.$ respectively. The obtained $M_s$ value (at 5 K) is in close agreement with those expected from Slater-Pauling rule (5 $\mu_B/f.u.$ ) which gives a strong indication of the half-metallic nature of this alloy.
Figure \ref{fig:CRMGMT} shows the temperature dependence of the magnetization at a constant field of 100 Oe. The Curie temperature ($T_c$) is found to be $\sim760$ K.
\subsection{Transport properties}
Figure \ref{fig:Res-fits-final}(A) shows the measured temperature dependence of resistivity for CRMG at $0$ kOe.
The inset (b) of Fig. \ref{fig:Res-fits-final}(A) presents the $\rho$ versus T at two different fields $0$ kOe and $50$ kOe, indicating insignificant dependence of $\rho$ on the field. To closely investigate the half-metallic nature of CRMG, we have used two different functions for fitting the zero field resistivity curve. The residual resistivity is found to be $48.16$ $\mu\Omega$ cm. In the first approach, we have fitted the resistivity data to a general power law,
\[
\rho(T) = \rho_0 + \rho_n T^n
\]
where $\rho_0$ is the residual resistivity. The value of n, when fitted in temperature range 2-400 K is found to be 1.53.
In the second approach, the resistivity curve was fitted using the equation
\begin{equation}
\rho(T) = \rho_0+ \rho_{phonon} + \rho_{magnon} = \rho_0+ \rho_1 T + \rho_2 T^2
\end{equation}
where $\rho_1$ and $\rho_2$ are arbitrary constants, $\rho_0$ is the residual resistivity which originates from the scattering of conduction electrons by the lattice defects, impurities etc., $\rho_{phonon}$ and $\rho_{magnon}$ arise due to scattering of phonons and magnons respectively.
With the quadratic temperature dependence, the curve fits well for temperature $\mathrm{T} > 35$ K as shown in the inset (a) of Fig. \ref{fig:Res-fits-final}(A). At low temperatures, however, the electrical resistivity is found to be almost independent of temperature (2-35 K). Similar behavior was also observed for other half-metallic ferromagnets like CoRuFeSi\cite{Bainsla2015631} and $\mathrm{Co_2FeSi}$ \cite{PhysRevLett.110.066601}. The results of the fitting are given in Table \ref{tab5}. It can be seen that the dominant contribution is from the electron-phonon scattering (T-dependence).
In a half-metallic ferromagnet, the states at E$_{F}$ are completely spin-polarized, and hence spin-flip scattering is not possible\cite{1989JPCM1.2351O,Kubo-1972,1989JPCM1.2351O} due to absence of electrons in the minority band gap at E$_{F}$. Thus, for a half-metallic ferromagnet, $T^2$ term is expected to be absent in the resistivity. The insignificant contribution of $T^2$ term in the resistivity data for CRMG alloy, indirectly supports the half-metallic nature.
Figure \ref{fig:Res-fits-final}(B)(top) shows a comparison of our theoretically simulated resistivity (solid circle) with those measured (red line). Within Boltztrap method, the resistivity (or the electrical conductivity) is calculated in units of the relaxation time ($\tau$). The bottom panel of Fig. \ref{fig:Res-fits-final}(B) shows the temperature dependence of these relaxation time, which varies in the range $1-2$ femtoseconds and hence lie in the typical range of $\tau$ as for most other standard compounds. The comparison of resistivity ($\rho$) is shown only between $100-400$ K, because Boltztrap method is based on a semiclassical theory and it relatively yields more accurate results at higher temperatures. This is mainly because for low temperatures, fewer bands are included in the summation for resistivity expression, and thus a much higher k-point mesh is necessary to obtain accurate estimations of the electrical properties.
\section{Summary and Conclusion}
In summary, we have studied the structural, electronic, magnetic and transport properties of equiatomic quaternary Heusler alloy, CoRhMnGe by means of both theory and experiment. Experimentally prepared alloy crystallizes in Y-type structure with almost no signature of intrinsic disorder. Ab-initio simulation confirms the stability of the measured crystal structure. It also confirms the half-metallic nature of the alloy (leading to high spin polarization) which is supported by the experimentally observed (almost) integer moment ($\sim$ 5.0 $\mu_B$) at 5K. In addition, measured electrical resistivity also indirectly supports the half-metallic behavior in the alloy. Electrical resistivity is also calculated theoretically, which compares fairly well with experiment. Carrier relaxation time ($\tau$) is shown to lie between $1-2$ fs, which is the usual range of $\tau$ in most compounds. Pressure studies reveal that half metallic nature of CRMG remains robust within a limited range of pressure (upto 30.27 GPa) beyond which it becomes metallic. There is, however, no evidence of magnetic phase transition. Electronic energy and lattice dynamics calculations show that the system is chemically as well as mechanically stable. Summarizing all these properties along with high spin polarization and relatively large value of T$_{C}$ ($\sim760$K) makes CRMG a potential candidate for spintronic applications.
\section*{Acknowledgments}
Deepika Rani would like to thank Council of Scientific and Industrial Research (CSIR), India for providing Junior Research Fellowship. Enamullah (an Institute Post Doctoral Fellow) acknowledges Indian Institute of Technology Bombay for necessary funding to support this research. AA acknowledges DST-SERB (SB/FTP/PS-153/2013) for funding
to support this research.
| 05cfd509e542d6df840c603c0c04caf522d21600 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{}
\section{Introduction \label{sec:Intro}}
Simple nonlinear dynamical systems in which trajectories may escape
through an artificial leak
\footnote{Artificial means in this context that if the leak is not present, escape cannot occur at that part of the phase space.}
placed in the phase space play an important role in recent studies.
Various fields of physics deal with either the escape dynamics of the particles or the decay rate of other physical quantities such as sound intensity, light rays, or fractal eigenstates \cite{Sch2002,Jun1993,Ern2014,Nag2005,Alt2009,Has2013,Por2008}.
It has been pointed out that the escape dynamics strongly depends on
the leak size, position, and orientation
\cite{Lai1999,Zyc1999,Afr2010,Bun2007,Det2012,Det2013,Det2011} as
well as on other pre-defined properties of the leak, for instance, the reflection coefficient \cite{Alt2013}.
Probably the most interesting question is how the escape dynamics changes if the size
of the leak varies.
Altmann et al. \cite{Alt2013R} presented numerical results about the relation between the escape rate and the leak size.
In their study, however, the measure of the leak was adjusted manually in each case.
Recently, Livorati et al. \cite{Liv2014} studied the escape in case of periodically
driven holes.
The main results of their work show parameter
(amplitude, initial phases, and period of the oscillations)
dependent fluctuations superimposed to the classical exponential decay.
Although mathematicians are interested mostly in the limiting case of vanishing small leaks \cite{Hay2005,Kel2009,Szasz2012}, in this work we present the decay dynamics through a continuously growing leak, where the size of the leak depends on a given physical property of the escaping particles.
The motivation of this study comes from the application of leaky chaotic systems
\cite{Ass2014,Kov2015,Mor2015} and crash tests \cite{Zot2015,Zot2016a}
in dynamical astronomy discussed in details below.
The model of the growing leak introduced here results in a survival probability of non-escaped trajectories that is different from the well-known classical exponential decay \cite{Lai2011,Tel2006}.
Moreover, we found a simple analytical solution describing the escape dynamics until the leak's expansion stops.
A comprehensive numerical investigation is also performed to confirm our analytic results.
The paper is organized as follows.
After the Introduction, in Section \ref{sec:model}, the motivation as an astrophysical
application is described.
Then, we give a detailed description of the model of a growing leak and its simple numerical implementation to the standard map.
The mathematical background
is presented in Sections \ref{sec:averaged_behaviour}.
Section \ref{sec:num_res} is devoted to numerical calculations
in order to compare analytic results and simulations.
Finally, we discuss our results and draw some conclusions in Section \ref{sec:sum_disc}.
\section{Model \label{sec:model}}
\subsection{Motivation \label{sec:rtbp}}
The motivation of the present study \cite{Gol1973} is the so-called planetary accretion process which is one of the two competing planet formation scenarios in these days \cite{Mat2007}.
In this process the forming planetary embryo accretes particles from its vicinity until this region -- the \textit{feeding zone}
\footnote{The planetary feeding zone is basically the basin of attraction of a given
leak where the leak in phase space can be considered as the forming planetesimal.}
-- becomes empty.
The increase of the planet depends on the mass of the particles hitting its surface.
Obviously the smaller the embryo at the beginning of this process,
the more significant the growth by the accretion.
As a very simple model of this process one might consider the gravitational planar circular restricted three body problem (RTBP).
In RTBP two point masses (star and planet) orbiting their barycenter on a circle and a
third mass-less body (test particle) moves in their gravitational potential in
the same plane.
Although the planet (and also the star) is considered as a point mass, one can define
the Hill radius ($r_{\mathrm{H}}$) in which its gravitational influence is dominant.
The particles entering the Hill radius with an appropriate velocity, i.e. slower than the escape velocity from this domain, can be removed from the dynamics and marked as escaped.
{In addition, $r_{\mathrm{H}}$ grows with the mass of the forming planet, see Eq.~(\ref{eq:rH}).}
Therefore, the growth of the planetary embryo can be considered as a growing
leak in the phase space.
Thus, from dynamical point of view, the accretion stage of the planet formation can be
described via leaky chaotic systems.
We give an estimate how the leak size depends on the mass in RTBP, see Appendix \ref{appendix:size_mass}.
\begin{figure}[b]
\includegraphics[width=0.475\textwidth]{fig1.pdf}
\caption{\label{fig:rtbp} (color online)
(a) An example of a growing leak in dynamical astronomy.
The plot shows initial conditions from the annulus around the planet's orbit.
The end-states of the particles are color coded as described in the text.
Particles have been started on circular Keplerian orbit.
The size of the planet and star are enlarged for better visualization.
The triangular Lagrangian points ($L_4\text{ and }L_5$) are also marked.
(b)-(d) Examples for individual orbits corresponding to certain initial conditions
in panel (a).
Note that the end point of the light gray (blue) trajectory is outside the plotted region.
}
\end{figure}
To illustrate the leaky RTBP,
we plot the evolution of a large number of non-interacting test particles initially placed
around the planet's orbit (see Figure \ref{fig:rtbp}).
Different colors denote different end-states of particles.
Trajectories starting from light gray (green online) points remain the part of the system during the whole
integration (1000 orbits of the planet).
Gray (red online) points represent test particles whose destination is the planet, more precisely,
the half of the Hill radius with proper velocity
\footnote{Since we consider the planet as a point mass, crash of the particles and the
planetary embryo is difficult to calculate numerically.
Consequently, half of the Hill radius is chosen as a region wherein
the particles are thought to be accreted by the forming planet.
This is, obviously, more rigorous criterion than one Hill radius.}.
Dark gray (blue online) points indicate trajectories scattered out from the system by the planet.
Although the effect of the planet's mass and size evolution in the RTBP is
dominant only in very early stages of the planet formation, the idea of a growing leak,
particularly when the size of the leak depends on a physical property of the leaving
particles, might shed light on a new kind of escape dynamics generally in leaky chaotic systems.
\subsection{Growing leak model\label{sec:leak}}
{
The discrete dynamical system we are to consider here consists a large number of
particles and a leak, where under certain conditions, the particles can
escape from the system.
}
The particles are point masses with the same mass $m$,
their initial number is $N_0$,
while after $i$ iterations we denote the number of surviving particles by $N_i$.
The leak also has an initial and an instantaneous mass, $M_0$ and $M_i$, respectively.
When a particle falls into the leak, its mass is added to that of the leak, thus
\begin{eqnarray}
M_i = M_0 + (N_0-N_i) \cdot m = \mathcal{M} - N_i \cdot m,
\label{eq:leakmass}
\end{eqnarray}
where $\mathcal{M} = M_0 + N_0 \cdot m$ is the total mass of the system.
{According to the RTBP (Appendix~\ref{appendix:size_mass}) a reasonable choice is }that the volume of the leak depends on its mass $M_i$ in the form of
\begin{eqnarray}
S_{\mathrm{leak}}(M_i) = C_{S} \cdot M_i^\gamma,
\label{eq:T_leak}
\end{eqnarray}
{where $\gamma$ is a positive constant. The coefficient $C_S$ can be written as $C_S=C_P\cdot S_{\mathrm{total}}$.
Here $C_P>0$ denotes a normalization constant while $S_{\mathrm{total}}$ is the
{volume} of the ergodic part of the phase space.}
The factor $C_P$ allows us to control the final size of the leak,
(a leak of moderate size avoids excessive restructuring of the phase space).
Let $p$ be the escape probability that a particle leaves the system
(through the leak) in the next iteration.
We {suppose} that the escape probability is proportional to the actual size of the leak compared to the whole phase space, that is,
$p=S_{\mathrm{leak}}/S_{\mathrm{total}}.$
That is, the escape probability (see Eq.~(\ref{eq:T_leak})) is given by
\begin{eqnarray}
p(M_i) = C_p \cdot M_i^\gamma. \label{eq:escape_prob_M}
\end{eqnarray}
Generally, the escape probability is changing as the mass (and size) of the leak is increasing.
At this point, it is useful to introduce some new constants and variables:
\begin{eqnarray}
\mathcal{N} = \frac{\mathcal{M}}{m},\;\;\;\;
\kappa_\infty = C_p \mathcal{M}^\gamma,
\nonumber
\label{eq:parameters}
\end{eqnarray}
\begin{eqnarray}
x_i = \frac{M_i}{\mathcal{M}},\;\;\;\;\;
y_i = \frac{N_i}{\mathcal{N}} =
N_i \frac{m}{\mathcal{M}},
\label{eq:norm_functs} \nonumber
\end{eqnarray}
where $\mathcal{N}$ is the number of particles corresponding to the total mass $\mathcal{M}$,
$\kappa_{\infty}$ {is the asymptotic escape rate} when all the mass of the system is in the leak,
{$x$ is the ratio of the mass of the leak and the total mass (mass ratio),
$y$ is the ratio of the number of the particles which are outside the leak to the total number of the particles $\mathcal{N}.$
It is obvious that
\begin{eqnarray}
x_i + y_i = 1 \nonumber
\end{eqnarray}
for all time instant.
}
We will use these dimensionless quantities through the rest of the paper.
{
The assumption of a small leak in our model corresponds to the pure exponential survival probability, i.e. when the system shows strong chaotic properties.
That is, if a static leak with size equal to the final size of the evolving leak
(set by $C_{\mathrm{P}}$) produces exponential decay,
we consider that this measure of the leak is small enough to our purposes and fits to the zero order approximation
$p=S_{\mathrm{leak}}/S_{\mathrm{total}},$
widely used in the literature, see for example \cite{Det2009}.
In addition, the exponential decay can also be observed in weakly chaotic systems for short times until the hyperbolic dynamics dominate.
}
{
Furthermore, in case of weak chaos the growing leak in the model presented should avoid the quasiperiodic domain in the phase space.
On the other hand, if the leak intersects the KAM tori during its growth, the survival probability will decay with lower different rate.
In other words, since the regular domain behaves as a forbidden region for trajectories originating outside, the leak biting into it will have an unreachable part for those trajectories resulting in a different escape probability.
However, this is no longer true when the leak originally contains islands or more precisely when the ratio of the regular islands inside and outside the leak remains constant.
}
\subsection{Simplified numerical experiment\label{sec:stmap}}
In order to analyze the escape dynamics through a continuously growing leak
defined by Eq.~(\ref{eq:escape_prob_M}), we introduce a simple test system.
Our numerical experiments are based on the standard map ($\text{mod}\; 2\pi$) which
describes the Poincar\'e map of the kicked rotator.
This choice makes it possible to check the leak's expansion in both co-ordinate
and velocity directions, respectively.
The standard map (SM) reads as follows
\begin{equation}
\begin{split}
I_{i+1}&=I_{i}+K\sin \Theta_{i},\\
\Theta_{i+1}&=\Theta_{i}+I_{i+1}.
\end{split}
\label{eq:SM}
\end{equation}
In Eq.~(\ref{eq:SM}) $K$ denotes the strength of the perturbation
and allows to study either \textit{fully hyperbolic dynamics}
($K$=5.19) or \textit{mixed phase space} structure, e.g. $K=2.7$.
{
An other reason we consider the SM is that it allows us to mimic the conservative dynamics in the RTBP where regular islands are also embedded in the chaotic sea producing the well-known structure of the phase space similar to that in Fig.~\ref{fig:SM}.
}
For simplicity, we presume that the leak grows equally in $I$ and $\Theta$ directions,
i.e. it conserves its original shape.
In order to avoid the early irregular effects in escape rate due to the location and
density of the initial conditions, a threshold time is obtained before the leak is opened.
Thus, we have a uniform distribution of the trajectories in the ergodic region of the phase space.
The threshold time is set to be $i=250$ in all simulations.
\begin{figure}[b]
\includegraphics[width=0.475\textwidth]{fig2.pdf}
\caption{\label{fig:SM}
Visualization of the numerical setup.
The invariant curves (blue online), plotted for completeness, are related to
different initial conditions than those show by dots representing the chaotic trajectories.}
\end{figure}
Figure \ref{fig:SM} shows the phase space portrait of the SM for $K$=2.7.
We place a square-shaped leak centered at point
$(I, \Theta)=$(5,5) with initial size $(\Delta I,\Delta \Theta)=(0.01,0.01)$
($S_{leak}^{(0)}=10^{-4}$)
\footnote{
Note that the mass and the size of the leak are identical parameters of the problem.
The instantaneous size can be obtained from the current mass and vice versa.
}
and store the number of escaped trajectories at every iteration step.
The semi-diagonals indicate the expansion until the leak reaches its final size
$(\Delta I,\Delta \Theta)\approx(0.316,0.316)$ ($S_{leak}^{(\infty)}=0.1$).
Initial conditions are placed uniformly in the black square
($3.2\leq \Theta\leq 3.7$, $3.2\leq I\leq 3.7$)
far from KAM islands as well as the final leak.
\begin{figure}[t]
\includegraphics[width=0.475\textwidth]{fig3.pdf}
\caption{
\label{fig:kappa_t}
Escape dynamics in SM.
{
Parameters of the simulation:
$K=2.7$, $\gamma=1$, $m=1$, $C_p = 10^{-7}/(2 \pi)^{2}$, $N_0 = 10^6$, and $M_0 = 1000$.
The leak reaches its final mass at $t\approx 6000.$}
For more details see the text.
}
\end{figure}
The result of a test run is displayed in Figure \ref{fig:kappa_t}.
It is clearly visible that the well-known exponential decay of the non-escaped trajectories starts after $\mathtt{\sim}$3000 iterations (blue squares).
Red triangles denote the instantaneous leak size, $S_{\mathrm{leak}}$, which is growing rapidly until it reaches its final ($\mathtt{\sim}$90\%) size.
One can also observe that the exponential decay starts roughly when the expansion of the leak ceases.
We can, thus, presume that the exponential behavior is a consequence of the stationary leak size with escape rate $\kappa_{\infty}$.
{
The semi-logarithmic plot of the non-escaped trajectories allows one to find the asymptotic escape rate, $\kappa_{\infty}$ as $t\to\infty$ (for strong chaotic regime).
This simulation yields $\kappa_{\infty}=0.00254$.
}
Furthermore, the numerical investigation confirms the naive idea that until the leak's expansion is present, the {instantaneous} escape rate, $\kappa(t),$ and also the escape probability is changing in time according to
$\mathrm{d}(\ln y_n)/\mathrm{d}t=-\kappa(t)$.
However, when the growth slows down significantly $\kappa(t)$ reaches the {asymptotic} escape rate $\kappa_{\infty}$ (green asterisks),
see Figure \ref{fig:kappa_t}.
This behavior can be explained as follows.
At the beginning of the simulation ($t<2000$) a very large number
of escaping trajectories feed the small leak in one iteration step
and, therefore, its mass (size) growth is accelerating.
Beyond a certain limit the mass (or equivalently the number) of escaping particles
in one iteration compared to the mass of the leak becomes small,
i.e. escape is present with moderate increase of the leak size.
In this case ($2000\leq t\leq 5000$), however, there are enough
particles in the system to observe the exponential decay.
The reason for the larger dispersion in $\kappa(t)$ and its deviation
from $\kappa_{\infty}$ beyond $t\approx 5000$ is twofold.
On the one hand, the number of non-escaped trajectories, after 5000 iterations,
becomes so small ($\sim$100) that the statistic is unreliable.
On the other hand, Figure \ref{fig:kappa_t} shows the
simulation for $K$=2.7, in which case KAM tori are responsible for
stickiness and consequently a power-law decay of trajectories for
longer escape times (not shown).
In other words, $\kappa(t)$ would follow the horizontal dashed line in
case of the fully hyperbolic dynamics, for instance, $K\ge$~5.19, with an arbitrarily large $N_0.$
\section{Results \label{sec:results}}
\subsection{Analytic solution \label{sec:averaged_behaviour}}
After having some impression about the escape dynamics from numerical simulations,
in this section, we show that a continuous approximation of the temporal
behavior of the model can be described by analytic formulae.
{
We consider the particle number $N_i$
and all the other related discrete functions $M_i$, $x_i$, and $y_i$
as being continuous functions $N(t)$, $M(t)$, $x(t)$, and $y(t)$.
}
Practically, we can do that because the particle number and the typical timescale
(number of iterations) of the process is also much higher than unity ($N_0\gg1$).
The time derivative of $N(t)$ is approximately the negative of
the average number of escaping particles $\Delta N$ during one
iteration which is $p \cdot N$, so we can write
\begin{eqnarray}
\frac{\mathrm{d}N}{\mathrm{d}t}\approx \Delta N =
- p \cdot N = - C_p \cdot M^\gamma \cdot N
\label{eq:N derivative}
\end{eqnarray}
where we used Eq.~(\ref{eq:escape_prob_M}).
As $\Delta M = - \Delta N \cdot m$, the time derivative of $M(t)$ is
\begin{eqnarray}
\frac{\mathrm{d}M}{\mathrm{d}t}\approx
C_p \cdot M^\gamma \cdot N \cdot m.
\label{eq:M derivative}
\end{eqnarray}
Combining Eq.~(\ref{eq:M derivative}), $M(t) = x(t) \cdot \mathcal{M}$, and
$N(t) = y(t) \cdot \mathcal{N}=(1-x(t)) \cdot \mathcal{M}/m$,
we get a first-order separable ordinary differential equation for $x(t)$:
\begin{eqnarray}
\frac{\mathrm{d}x}{\mathrm{d}t}
=
\kappa_{\infty} \cdot x^\gamma \cdot \left( 1 - x \right),
\label{eq:ode for x}
\end{eqnarray}
Derivation of the solution can be found in Appendix~\ref{appendix:solution_of_eq}.
{
Equation (\ref{eq:ode for x}) is a continuous approximation of the recursive difference equation
\begin{eqnarray}
x_{i+1}=x_i + \Delta x_i
\label{eq_discrete}
\end{eqnarray}
where $\Delta x_i = \kappa_\infty \cdot x_i^\gamma \cdot \left( 1 - x_i \right)$,
which gives the exact description of the discrete-time problem.
}
The implicit solution of (\ref{eq:ode for x}) can be given by
\begin{eqnarray}
t(x)=
\frac{x^{1-\gamma}}{\kappa_{\infty} \cdot (1-\gamma)}\;
{}_2F_1\left(1-\gamma,1;2-\gamma;x\right) - \tau
\label{eq:sol}
\end{eqnarray}
where the constant of integration $\tau$ follows from the initial value $x_0$ as
\begin{eqnarray}
\tau = \frac{x_0^{1-\gamma}}{\kappa_{\infty} \cdot (1-\gamma)}\;
{}_2F_1\left(1-\gamma,1;2-\gamma;x_0\right).
\label{eq:int_const}
\end{eqnarray}
\begin{figure}
\includegraphics[width=0.95\linewidth, angle=0]{fig4.pdf}
\caption{ The mass growth of the leak $x(t)$ for different $\gamma$s.
For better visibility, the constants of integration ($\tau$) are
chosen with $x_0=x_{\mathrm{PoI}}$ taken at 0 see Eq.~(\ref{eq:int_const}).
Parameter $\kappa_{\infty}$ is taken equal to $1/(2\pi)^{2} \approx 0.0253$.
\label{fig:analytical} }
\end{figure}
The solution of Eq.~(\ref{eq:ode for x}), $x(t),$ has a point of inflection (PoI) for all $\gamma>0.$
The second derivative of $x$ from (\ref{eq:ode for x})
\begin{eqnarray}
\frac{\mathrm{d}^2x}{\mathrm{d}t^2} =
\kappa_{\infty} \cdot x^{\gamma-1} \cdot \frac{\mathrm{d}x}{\mathrm{d}t}
\cdot \left[\gamma - ( 1 + \gamma )\cdot x \right],
\nonumber
\end{eqnarray}
from which the $x$ coordinate of the inflection point
($x_{\mathrm{PoI}}$)
can be obtained
\begin{eqnarray}
x_{\mathrm{PoI}} = \frac{\gamma}{1 + \gamma}.
\end{eqnarray}
We further elaborate on the error properties of the above solution in Appendix \ref{appendix:error}.
We can distinguish two parts of the leak-growing process.
The separatrix is the point of inflection of the $x(t)$ function.
Figure \ref{fig:analytical} shows the functions $x(t)$ for different $\gamma$s.
For the sake of comparison the graphs are shifted leftward, thus,
the inflection points are placed exactly above a row at $t=0.$
The mass growth $x(t)$ beyond the point $x_{\mathrm{PoI}}$ (or $t=0$) has
the same characteristic for different $\gamma$s.
The reason is that in the limit $t \rightarrow \infty$, $x \rightarrow 1,$ Eq.~(\ref{eq:ode for x})
can be written as
$\mathrm{d}x/\mathrm{d}t \approx -\kappa_{\infty}x$
which means that function $x(t)$ approximates 1 exponentially with exponent
$-\kappa_{\infty}$ and the process does not depend on $\gamma.$
This is, however, not the case to the left of the point of inflection.
In the limit of $x\rightarrow 0$, Eq.~(\ref{eq:ode for x}) can be written as
$\mathrm{d}x/\mathrm{d}t \approx \kappa_{\infty}x^\gamma$ which means that the solution
$x(t)\approx\left[\kappa_{\infty} (1-\gamma) (t +\tau)\right]^{\frac{1}{1-\gamma}}$
follows a power-law and contains both $\kappa_{\infty}$ and $\gamma.$
Furthermore, in this regime $\gamma$ defines two different behaviors.
Considering the case of $\gamma<1$ we have a point where $x(-\tau)=0.$
That is, the integration constant $\tau$ is suitable to determine a time instant
in the past when the mass of the leak was zero, i.e. when the whole growing process began.
While in the case of $\gamma \ge 1$ the function $x(t)$ approaches zero only
in the limit of $t \to -\infty.$
In summary
\begin{eqnarray}
\lim_{x\rightarrow 0} t = \left\{
\begin{array}{ll}
-\tau &\;\;\; \textrm{for $0\le\gamma<1$,}\\
-\infty &\;\;\; \textrm{for $1\le\gamma$.}
\end{array}
\right.
\label{eq_limit_of_t}
\end{eqnarray}
Nevertheless, it is obvious from Eq.~(\ref{eq:ode for x}) that $\kappa_{\infty}$ is inversely proportional to the timescale of the process.
The condition that the timescale have to be much higher than unity is equivalent
to $1/\kappa_{\infty} \gg 1$.
This fact is important to ensure that the continuous time approximation,
Eqs.~(\ref{eq:N derivative}) and (\ref{eq:M derivative}), is valid in our model.
The adopted model of growing leak defines a stochastic process,
whose complete description is possible only by using the probability theory.
The question arises naturally,
how the probability mass function of the particle number
can be calculated after the $i$th iteration if the initial one is known?
The question is important because if the standard deviations are considerable,
then we need the probability mass functions in order to have a complete description.
Otherwise, the averaged behavior, studied previously, describes the process well.
In Appendix \ref{appendix:distr} we derive the probability mass functions, and study
its properties this problem.
We should mention that during the calculation we assumed that $\gamma>0.$
However, it is obvious that solutions of Eq.~(\ref{eq:ode for x}) can also be found
for negative exponents in a similar way.
The discussion of the case $\gamma<0$ is beyond the scope of the present study.
\subsection{Numerical tests \label{sec:num_res}}
After discussing the analytic description of the survival probability,
we confirm the validity of our calculations by running several numerical simulations.
In order to demonstrate the general phenomenon of escape dynamics,
we use different $\gamma$ values in our calculations.
\begin{figure}[b]
\includegraphics[width=0.475\textwidth]{fig5.pdf}
\caption{\label{fig:gamma1}
{
(a) Survival probabilities for different parameters $K=5.19$ and $2.7$ in SM.
Parameters of the simulation:
$\gamma=1$, $m=1$, $C_p = 10^{-7}/(2 \pi)^{2}$, $N_0 = 10^6$, and $M_0 = 1000$.
The gray dashed ($K=5.19$) and dashed-dotted ($K=2.7$) lines represent the analytic formula
(\ref{eq: sol gamma 1 exp})
with
$\kappa_\infty = 0.1/(2 \pi)^{2} \approx 0.00253$, $x_0 = 10^{-3}$
and
$\kappa_\infty = 0.00285$, $x_0 = 10^{-3}$
respectively.
(b) The difference between the numerical simulation and
the analytic formula
for $K=5.19$.
}
}
\end{figure}
\begin{figure}[b]
\includegraphics[width=0.475\textwidth]{fig6.pdf}
\caption{\label{fig:gamma43}
(a) The growth of the leak's mass and the decay of particles for $K=2.7$.
The other parameters are
$\gamma=4/3$, $m=1$, $C_p = 0.2845$, $N_0 = 10^6$, and $M_0 = 5631$.
The dashed lines illustrate the analytic solutions
with
$\kappa_\infty = 0.1/(2 \pi)^{2} \approx 0.002533$, $x_0 = 0.0055$
.
(b) $S-C$ curve shows the difference in leak mass.}
\end{figure}
First, the results of the hyperbolic and mixed dynamics are compared.
In this calculation we show that for different system parameters $K=2.7\text{ and }5.19$
the analytical solution works very well.
Figure \ref{fig:gamma1}(a) shows the ratio of non-escaping trajectories $y(t)$
for the $\gamma=1$ case, i.e. the leak size depends linearly on mass.
One can easily see that the analytical solution (dashed and dotted dashed lines) fit the numerical data fairly accurately, especially for small iteration numbers, $t<2000.$
In order to be able to compare the accuracy of the results quantitatively,
we calculate the relative difference between the simulated data (S) and
the analytic solution (C).
The difference $S-C$ in percentages is plotted in Figure \ref{fig:gamma1}(b).
It shows the same tendency
what we can observe by naked eye in panel (a).
The $S-C$ diagram remains under 4\% level until $t\approx 2500.$
In addition, $S-C$ shows that in the case of $\gamma=1$ the analytic solution is more accurate for fully hyperbolic dynamics ($K=5.19$) than for mixed phase space ($K=2.7$)
for $t>2500$.
The reason of that comes form Eq.~(\ref{eq:sol gamma 1}), since it turns to be purely
exponential for $t\gg 1$, that is
$y(t)\sim exp(-\kappa_{\infty} t)$.
In addition, the decay of $y(t)$ in the latter case starts to deviate from the exponential due to the sticky effect of the KAM tori.
Physically more interesting cases are when $\gamma\neq1$ but rational.
Let us recall our motivation, the planet formation analogy in the planar RTBP.
The size of the leak in the phase space in this particular case is
proportional to $m_p^{4/3},$ see Eq.~(\ref{eq:hole2D}) in the Appendix.
Figure~\ref{fig:gamma43}(a) shows the number of surviving particles, $x(t),$
and the mass growth of the leak, $y(t),$ for $\gamma=4/3$
(squares and triangles, respectively).
The analytic solution goes together with the numerical simulation also for this value of $\gamma.$
As is well seen in panel (b) the $S-C$ diagram remains under the 5\% level until the leak reaches its final mass, $t\approx8000.$
This is not true, however, at the very beginning of the iteration,
$t<10$ after opening the leak.
In this regime sudden changes in the number of escaping trajectories appear.
Trajectories situated exactly 'above' the leak and its pre-images disappear immediately from the system.
This rapid change in the number of particles is, however, not covered
by the analytic solution and, consequently,
large differences may show up in the first phase of the $S-C$ diagram.
\begin{figure}[b]
\includegraphics[width=0.475\textwidth]{fig7.pdf}
\caption{\label{fig:logn}
(a) The leak mass vs. time for different mass distributions of particles each of them
with mean=1 and std=0.667.
Squares represent equal masses $m=1.$
(b) The differences between the leak's masses are significant only for the first ~200 iterations.
$K=2.7.$}
\end{figure}
In the previous two examples we considered particles with equal masses, $m=1.$
A more realistic scenario is when the particles in various physical problems
have different masses corresponding to a certain distribution.
The log-normal distribution is a good choice to describe the particle size (and/or mass).
We present a simulation for $\gamma=2/3$ with different kind of mass distributions, see Figure~\ref{fig:logn}.
The numerical results in panel (a) show what can also be derived directly from
Equations~(\ref{eq:N derivative}) and (\ref{eq:M derivative}):
\textit{the mass growth of the leak does not depend on the mass of the individual particles but only on the mean value of the distribution.}
Consequently, the leak's mass changes in time with the same rate
for both equal mass particles (pink squares) and log-normal distribution (red triangles),
and also for other distributions such as uniform and normal
(stars and circles in Figure~\ref{fig:logn}(a), respectively).
The statistical fluctuations in leak's mass, smaller than 15\%,
disappear after ~200 iterations, panel (b).
\section{Summary and Discussion\label{sec:sum_disc}}
The model Equations~(\ref{eq:N derivative}) and (\ref{eq:M derivative})
describe the escape dynamics in a leaky chaotic system when the size of the leak is growing in time and the expansion depends on the particles' mass.
Consequently, the escape probability is time-dependent.
The analytic solution to the problem provides a power-law behavior at the very early stage
($x\approx 0$) of the dynamical evolution. This phase depends on the exponent $\gamma$ in
Eq.~(\ref{eq:T_leak}).
However, for larger $t,$ when the feeding of the leak diminishes,
the survival decay turns to be exponential.
Between these two limits the escape rate is time dependent.
The qualitative picture is the following.
After the leak reaches roughly the 90\% of its final measure, or more precisely,
beyond the point of inflection of $x(t)$, the speed of the growth slows down.
After this point the growth of the leak is so slow that it can be thought of
as a static leak, and the decay rate turns to be exponential,
see Figure~\ref{fig:holevelo}(a).
Numerical simulations verify that the escape rate $\kappa_{\infty}$
(short thick solid line) for a static leak (red triangles) of size 0.1 is the same
as in the case of a growing leak (blue squares) when it reaches 90\% of
its final size (also 0.1), panel (a).
In addition, this behavior is in a very good agreement with the analytical solution describing
the early stage escape dynamics.
{The effect is considerable for relatively short times only as long as enough number of
particles are in the system,
therefore, the presence of the well-known power-law decay of stickiness (tail of the distribution)
in mixed phase space is not affected by the size variation of the leak. However, the crossover time, when the nonhyperbolic part of the chaotic saddle starts to dominate, can be updated.}
{The crossover time $t_{\mathrm{cross}}$ in weakly chaotic regime is written as follows (Eq.~(89) in \cite{Alt2013})}
\begin{equation*}
t_{\mathrm{cross}}\sim 1/\kappa_{\infty}
\end{equation*}
{with the assumption that the leak size is small. The growing leak model provides a simple generaliztion of this naive approximation in $\gamma\leq 1$ case}
\begin{equation}
t_{\mathrm{cross}}\sim \frac{1}{\kappa_{\infty}}\left[1+\frac{x^{1-\gamma}}{1-\gamma}+\frac{(1+\gamma)^{1+\gamma}}{\gamma^{\gamma}}\frac{\gamma}{1+\gamma}\right],
\end{equation}
{where the second and the third terms in the bracket define the shift ($t_{\mathrm{shift}}$) the crossover experiences, see the shematic view in Fig.~\ref{fig:holevelo}b. The second term is the time of the growth until the leak mass is moderate, see the approximation of Eq.~(\ref{eq:ode for x}) when $x\ll1$, while the third term can be derived from the slope of the function $x(t)$ at point $x_{\mathrm{PoI}}$, Fig.~\ref{fig:analytical}. It can be easily obtained that $t_{\mathrm{shift}}\to 0$ when $\gamma\to 0$ and $x\ll1$.}
\begin{figure}[b]
\includegraphics[width=0.475\textwidth]{fig8.pdf}
\caption{\label{fig:holevelo}
(a) Number of non-escaped particles and leak size/growth vs. time for
$K=2.7,\;\;\gamma=1/2,\;\;m_{i}=1.$ (b) Magnification during growing
process.}
\end{figure}
Equation~(\ref{eq:sol_gamma0}) properly describes also the limit case $m_{i}\to 0.$
Namely, if the mass of the particles tends to zero,
i.e. the growth of the leak is fairly slow,
one recovers the classical exponential decay for the surviving trajectories.
We note that the same effect can be seen when the initial mass of the leak $x_{0}$ is
set so large that even the massive particles ($m_{i}>0$) falling into it
do not have any effect on the leak's mass and, therefore,
it can be considered as a static leak.
{
Due to the leak expansion we can consider an instantaneous chaotic saddle in our model at every time step.
This object is reducing as the leak is growing and converges to that invariant set which corresponds to the final leak size.
This process results in a temporally changing chaotic saddle and a non-stationary exponent of the survival probability (escape rate $\kappa(t)$).
A similar phenomenon can be found in \cite{Mot2013} where the exponent is also time dependent (see Eq.~(1) in \cite{Mot2013}).
In contrast of the similarity, the temporarily changing chaotic saddle should not be confused with the transient chaotic saddle introduced in \cite{Mot2013}.}
In summary, we have presented an analytic description of the escape of
the trajectories through a continuously growing leak both in fully
hyperbolic and in mixed phase space.
We stress, however, that during the whole calculation we did not utilize explicitely the fact that $m$ is
the mass of the particle, though the basic motivation is related to the mass growth of
a planetary embryo.
Therefore, one can reformulate the model in a more general way.
{Let us write Eqs.~(\ref{eq:leakmass}), (\ref{eq:T_leak}), and (\ref{eq:escape_prob_M})
together as follows}
\begin{equation}
\begin{split}
M_i &= M_0 + (N_0-N_i) m = \mathcal{M} - N_i m,\\
S_{\mathrm{leak},i}& = C_{S} M_i^\gamma,\\
p_i &= C_p M_i^\gamma,
\label{eq:escape_prob_M_general}
\nonumber
\end{split}
\end{equation}
{where now $m$ is a physical property of the particles, $M_n\text{ and } M_0$ are
the evolved and initial additive property of the leak, and $\mathcal{M} = M_0 + N_0 m.$}
Other quantities are the same as given in the introduction of the model in Section~\ref{sec:leak}.
This means that the analytical method presented in this paper might be suitable to predict the
characteristics of the escape dynamics in different kinds of systems where the leak size
depends on some specific physical property of the particles {(charge, spin, energy level, chemical composition, etc.)}.
We also would like to draw the attention to the limitation of present model.
In fact, the dynamics in the standard map does not depend on the size of the leak.
In other words, the leak affects only the escape rate but not the individual survival trajectories themselves.
This is not the case, for instance, in the restricted three body problem,
where the growing planetary mass governs the dynamics of the surviving particles and,
therefore, should also modify the escape dynamics. {Considering such an extension in the SM, a natural choice could be the introduction of a variable nonlinearity parameter $K(M)$ whose value could also depend on the leak mass/size.}
Studying this effect is postponed to future studies.
| f7f01f112cb8d3f9eef6e6d5c6961c421cae0baf | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Deep Learning systems have surpassed other algorithms and even humans at several tasks~\cite{Bengio:2013,Krizhevsky:2012,Mnih:2013,Silver:2016}. While their applications continue to grow, deep learning systems are still considered black-box optimization methods. One of the most vital features behind their success is their ability to extract relevant yet minimal information as it progresses into deeper and deeper layers \cite{Tishby:2000}. This is an extension of \emph{Rate Coding Theory}~\cite{Berger:1971} presented as the \emph{Information Bottleneck} principle~\cite{Tishby:2000,Tishby:2015}. The information bottleneck principle has been primarily focused on systems that are (a)~feedforward, and (b)~in an open-loop, decoupled from their environment.
Neuroscientists, on the other hand, have long been studying the principles behind encoding and representation of environmental information in neural activity using principles of information theory~\cite{Borst:1999} and rate distortion theory~\cite{Simoncelli:2001}. Continuous variables from the environment are encoded as discrete spikes in the brain, which are then decoded to produce smooth continuous movement. Due to experimental limitations, an informational analysis of a closed-loop brain-body-environment behaviour system is not yet feasible.
We take a radically different approach to understanding information flow in a behaviorally-functional agent. We artificially evolve embodied agents controlled by recurrent spiking neural networks to perform a task. For this paper, we focus on a non-Markovian version of the classical pole balancing task. Pole balancing has been explored quite extensively as a benchmark for control using neural networks~\cite{Pasemann:1997,Onat:1998}. With continuous states and actions, this task serves as an ideal setting to study the transformation of the continuous input information into spikes and then back to continuous action. While the typical task is Markovian, and thus too trivial for an informational analysis, it can easily be transformed into a non-Markovian task by making the available information to the agent limited. Our approach to pole balancing incorporates an agent-centric ``visual'' interface to the pole. Therefore, information that is typically available, like the pole's angle and angular velocity, the agent's position and velocity, are not fed directly to the network. Ultimately, the minimal nature of the task makes it tractable for an investigation of a recurrent network in a closed-loop task.
The parameters of the recurrent spiking neural network that balances the pole were optimized using an evolutionary algorithm. Evolving neural networks as opposed to hand-designing them allows maximum flexibility for exploring the parameter space. While evolutionary algorithms have been very commonly used in several fields~\cite{Jong:2006}, recently, they have been proven to be efficient for optimizing deep neural networks as well~\cite{Miikkulainen:2017,Salimans:2017}. Moreover, due to the stochastic nature of the optimization, running the algorithm several times provides us with an ensemble of solutions that solve the problem. This allows the analysis of not just one solution but several to evaluate consistency of results.
The paper is organized as follows. In the first section we report on the agent, task, optimization technique, and analysis method. The section that follows presents an informational analysis for the best and top performing agents. In the last section we summarize the results.
\section{Methods}
\begin{figure}
\centering
\includegraphics[height=3.5cm]{SetupV1.pdf}
\caption{Task set up and agent design. [A]~The agent (gray circle) moves left/right along the horizontal axis (gray arrows) sensing the pole (black rod) through the seven vision rays (color dashed lines), with a range of $36$ degrees. [B]~Sensory rays have a linearly diffused receptive field from their centers and overlap at the edges with the receptive fields of adjacent rays. [C]~The agent has $7$ vision sensors (black) connected to two recurrent spiking interneurons (red) connected to two motor units (gray).}
\label{fig:setUp}
\end{figure}
{\it Agent design.}
The agent lives in a 1-dimensional world with the pole attached to its center. Seven equidistant rays of ``vision'' with overlapping receptive fields spanning 36\degree provide it with sensory stimuli (Fig.~\ref{fig:setUp}A,B). The control network of the agent has three primary components: sensory units, spiking interneurons, and motor units. There is one sensory units per ray, which merely pass on the signal received from the rays. Sensory units are fully connected to $N$ interneurons (here $N=2$), modeled by Izhikevich spiking neuron model \cite{Izhikevich:2003}. The model has 4 parameters per neuron and is governed by a two-dimensional system of ordinary differential equations~\cite{Izhikevich:2003}. Interneurons are recurrently connected (Fig.~\ref{fig:setUp}C). Therefore, each interneuron receives weighted input from each sensory unit, $S_i$, and from other spiking interneurons, $I_i$, as follows:
\begin{equation}
S_i+I_i = \sum_{j=1}^7 w_{ji}^s s_j + \sum_{j=1}^N w_{ji}^i o_i
\end{equation}
\noindent where $s_j$ is the input at the $j^{\text{\tiny th}}$ sensory unit, $w_{ji}^s$ is the strength of the connection from the $j^{\text{\tiny th}}$ sensory unit to the $i^{\text{\tiny th}}$ spiking interneuron, $w_{ji}^i$ is the strength of the recurrent connections from the $j^{\text{\tiny th}}$ to the $i^{\text{\tiny th}}$ spiking neuron, and $o_i$ is the output of the neuron. The sign of all outgoing connections from an interneuron depends on its excitatory or inhibitory nature, as identified by a binary parameter. Finally, the layer of interneurons feeds into the two motor neurons, that has the following state equation:
\begin{equation}
\tau_m \dot{m_i} = - m_i + \sum_{j=1}^N w_{ji}^m \bar{o}_j \quad
i = 1,2
\end{equation}
\begin{equation}
\bar{o}_j(t)= \frac{1}{h_j} \sum_{k=0}^{h_j} o_j(t-k)
\end{equation}
\noindent where $m_i$ represents the motor neuron state, $w_{ji}^m$ is the strength of the connection from the $j^{\text{\tiny th}}$ spiking interneuron to the $i^{\text{\tiny th}}$ motor neuron, $\bar{o}_j$ represents the firing rate code, the moving average over a window of length $h_j$ for the output of spiking interneuron $j$. Finally, the difference in output between the motor neurons results in a net force that acts upon the agent, moving it along the track. The network was simulated using Euler integration with step size 0.01.
{\it Pole Balancing Task Design.}
The agent can move left/right to balance a pole for as long as possible. The pole-balancing task was implemented based on published descriptions~\cite{Barto:1983}. The force from the agent, computed as the difference between motor unit outputs, affects the angular acceleration of the pole and acceleration of the agent. The physical parameters such as mass, length and coefficient of friction were all set as per the same reference. While typically pole-balancers receive as input the angle of the pole ($\theta$), its angular velocity ($\omega$), the agent's position ($x$) and velocity ($v$), our implementation was modified to only sense the pole through the sensory rays.
{\it Evolutionary Algorithm.}
The network was optimized using a real valued evolutionary algorithm. A solution was encoded by 38 parameters, including the intrinsic parameters of the Izhikevich neurons, the neuron-specific size of the window for estimating rate code, all connection weights (sensory-interneuron, interneuron-interneuron, interneuron-motor), and the time constant, bias and gain for the motor units. Parameters were constrained to certain ranges: connection strengths $\in$ $[-50,50]$; motor unit biases $\in$ $[-4,4]$; time-constants $\in$ $[1,2]$. The range of intrinsic parameters and the polarity of the outgoing weights from the inter-neuron depended on a binary inhibitory/excitatory neuron flag parameter in the genotype~\cite{Izhikevich:2003}. The population consisted of 100 individuals.
{\it Fitness Function.}
Performance was estimated by averaging over 16 trials, starting at pole angles $\theta_0$ between $\pm12\degree$, in increments of 3\degree, and two initial angular velocities, $\omega_0=\pm0.001$. The fitness function to be maximized was $f=(\sum_{t=1}^{T} cos(\theta_t))/T$, where $T=500s$ is the maximum duration of the run. The pole was considered balanced if it was within the sensory range of the agent. Also, the track length was fixed at $45$ units, beyond which the trial was terminated.
{\it Mutual Information.}
The amount of information contained in one variable about another was estimated using Mutual Information (MI). We quantified the information that neurons contain about pole angle ($\theta$), angular velocity ($\omega$), agent position ($x$) and agent velocity ($v$) by calculating their probability distributions (using a binning method with each bin of width 0.01):
\begin{equation}
MI(N,X) = \sum_{n \in N}\sum_{x \in X} p(x,n)log\frac{p(x,n)}{p(x) p(n)}
\end{equation}
\section{Results}
{\it Performance of evolutionary optimization.}
While pole balancing is a well-known benchmark, it was also a relatively easy task to optimize. The evolutionary algorithm found fairly good solutions (around 75\% performance) at the very first random draw of the population. Fig.~\ref{fig:bestAgent}A shows the performance of the best agent in every generation over time. Out of the $100$ evolutionary runs, $99$ converged to over $99\%$ fitness with only two spiking interneurons.
\begin{figure}
\centering
\includegraphics[height=3.5cm]{bestAgentV2.pdf}
\caption{Optimization performance. [A]~Fitness of the best individual in the population vs. generations for 100 evolutionary runs. Best run in blue, top 10 in light blue and the rest in gray. [B]~Network structure of the best agent. The width of the edges indicate the magnitude of the weights and are also color coded for polarity. The sensory units in black are identified as S1-S7, spiking interneurons N1,N2 in red and motor units M1,M2 in grey. [C]~Generalization performance over a broader and finely grained set of initial conditions that were tested during evolution (marked by cross marks).}
\label{fig:bestAgent}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=11.5cm]{bestInfoArchV2.pdf}
\caption{Behavior of the best agent and one of its trials in detail. [A]~Angle of the pole over time on 16 trials. One of the trials ($\theta_0=-6\degree$ and $\omega_0=0.001$) is highlighted and explored further. [B]~Information flowchart: from sensory input rays through the spiking inter-neuron layer, composed of 3 levels of processing (neuron potential, spiking activity and rate code), and then the motor units. [C]~A sample trace for each of the components corresponding to each box in B from the trial highlighted in A. Each color in the first figure is matched with the sensory rays in Fig.~\ref{fig:setUp}. The blue and orange colors indicate interneurons 1 and 2 respectively. The green and brown traces corresponds to the left and right motor neurons respectively. [D]~Mutual Information about the pole angle ($\theta$) for the highlighted trial for each of the components. [E]~Mutual information about $\theta$ in the neuron potential ($V$). A comparison between the trial-by-trial MI vs. the total MI across trials.}
\label{fig:bestInfoArch}
\end{figure}
\begin{figure}
\centering
\includegraphics[height=3.2cm]{miSigsV2.pdf}
\caption{MI about the four environmental variables: pole angle ($\theta$), pole angular velocity($\omega$), agent position ($x$) and agent velocity ($v$) in the network elements: neuron potentials ($V$), rate coded outputs of neurons ($R$), and motor unit states ($M$). Paired samples t-test yielded highly significant ($p\ll0.05$) differences between the information contained about $\theta, \omega$ and $v$ in $V$ versus $R$ and $R$ versus $M$.}
\label{fig:vrSignificance}
\end{figure}
{\it Network structure and performance of the best agent.}
The network structure of the best agent from the $100$ runs is shown in Fig.~\ref{fig:bestAgent}B.
The behavioral traces of this agent on the 16 trials specified in the previous section are shown in Fig.~\ref{fig:bestInfoArch}A. This agent achieved a fitness of $99.4\%$. To test for generalization, its performance was evaluated on a finer and broader range of conditions, post-optimization (Fig.~\ref{fig:bestAgent}C): initial pole angle $\theta_0$ in the range $[-45,45]$ and initial pole angular velocity $\omega_0$ in the range $[-0.01,0.01]$. As can be seen, the agent generalizes well within and outside the region it was evolved on. Note that $\theta_0$ that are beyond $18\degree$ on either side are beyond the range of the sensory rays.
{\it Encoding of environmental variables.}
The different network elements that manipulate the input are shown in Fig.~\ref{fig:bestInfoArch}B. Sensory signals first act on neurons' potential. The neuron fires based on dynamics in its potential, which is then interpreted by its rate. The motor units then convert this discrete spike rate to smooth continuous movement. Note that a single neuron has three levels of informational content - continuous valued potential, binary spikes, and discrete spike rate codes. Fig.~\ref{fig:bestInfoArch}C shows traces for the highlighted trial in Fig.~\ref{fig:bestInfoArch}A of the best agent. Although $\theta$ is the most directly available information, unlike the standard practice of directly providing it, the sensory rays provide an agent-centric perspective of $\theta$. MI between each of the network elements with $\theta$, see Fig.~\ref{fig:bestInfoArch}D, revealed that internal potential of neurons have relatively more information about $\theta$ than the spike rate and so does the motor units. Albeit only for one trial in one agent, this shows that the bottleneck does not always become narrow in control tasks. MI also reveals that indirect encoding of $\omega$, $x$ and $v$ all happen in the very first stage of the network, neuron potential (black bars in Fig.~\ref{fig:vrSignificance}). This can be attributed to the recurrent nature of connections between the interneurons and also their rather complex non-linear internal dynamics.
{\it Analysis of the Information Bottleneck.}
All available information, as shown by MI in $V$, is not necessarily used in controlling movement, as shown by relatively lower information in $R$. To further study the bottleneck, we compared the amount of information contained in neuron potentials, $V$, versus the rate coded outputs of the neurons, $R$. Note that the spikes themselves do not have any information about the environment but, in fact, encode them in its rate. For each of the environmental variables a paired samples t-test was conducted with a significance threshold of $p<0.05$. This revealed that there is a significant drop in the amount of information between $V$ and $R$ (Fig.~\ref{fig:vrSignificance}) robustly across the top ten agents. This can be attributed to the loss due to the discretization of information available in $V$ as spikes. However, the information in $R$ is sufficient to perform the behavior with great accuracy and so this is in fact an efficient encoding of information. The minimal yet relevant information that is encoded in $R$ is interpreted by the motor units. They integrate $R$ from the interneurons and their outputs directly impact $\theta$, $\omega$, $x$ and $v$ and so this is where a deviation from the IB principle is expected. Statistical analyses of the MI between $R$ and the motor units state, $M$, using the paired samples t-test yielded highly significant ($p\ll0.05$) increase in information about all environmental variables in $M$ (Fig.~\ref{fig:vrSignificance}). This shows that the IB for control tasks is not always a filtering of information but is rather filtering followed by an expansion at the control layer.
{\it Context sensitive information encoding.}
From previous analysis, we know that components of the network encode information about the environment. But what information do they encode? Typically, when a neuron is said to encode information about a feature of the environment it is thought to be a consistent, context-independent code. To explore this idea further, we compared the MI the neuron potential has about the pole's angle $I(V,\theta)$ on a trial by trial basis to the information that same neuron has across all trials about the same feature of the environment on the best agent (Fig.~\ref{fig:bestInfoArch}E) and the top 10 agents. A one-sample t-test of the combined MI with the distribution of trial-by-trial MI values yielded a highly significant difference ($p\ll0.05$). This means that the combined information is significantly lower than trial-by-trial information, and therefore that encoding is highly context-dependent across all evolved pole-balancers.
\section{Discussion}
In this paper, we have presented results from an information theoretic analysis of recurrent spiking neural networks that were evolved to perform a continuous control task: agent-centric non-Markovian pole balancing. Our results can be summarized as follows. First, networks with as few as two spiking neurons could be evolved to perform this task. Second, through the use of MI, we show that the spiking network encoded environmental variables of interest that were directly and indirectly available. Third, we show that the information bottleneck from the neuron potential to its firing rate is an efficient filtering/compression, which was followed by an increase in information at the control layer on account of their causal effect on the environment. This is a phenomenon that we expect to arise in control tasks in general, and plan to explore further with different tasks and types of networks. Perhaps, this can develop into an optimization method for neural network control. Finally, we show that the information encoded in the spiking neurons vary across trials, causing the across-trial combined information to be significantly lower. This can mean either that the same stimuli are encoded in different ways (redundancy) or that different stimuli are mapped on to the same encoding (generalization) or both. This warrants further analysis to understand the encoding in more detail and more interestingly, to understand how the context helps disambiguate the generalized representations during a trial.
| a0ca9d5358fb13aad791b1b8e7c567a9b165982d | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
A semi-quantitative relation between dynamical properties like diffusivity or relaxation times and the thermodynamics has been proposed
by Rosenfeld \cite{Rosenfeld,Rosenfeld-iop} and recently it has been extensively studied for different systems\cite{Goel_charu,murari_charu_2010_jcp,Ruchi_charu_2006,
manish_charu,Murari_charu_jcp,agarwal2009relationship,dhabal2016comparison}. The relationship suggests that the fluid should follow $X^*=C \exp [-K S_{ex}]$,
where $X^*$ is the dimensionless dynamical quantity and $S_{ex}$ is excess entropy which is the difference
between the total thermodynamic entropy ($S_{tot}$) and the corresponding ideal gas entropy ($S_{id}$) at
the same temperature ($T$) and density ($\rho$). C and K are the constants in the Rosenfeld relation.
In a liquid due to the structural correlation the total entropy is less than the corresponding ideal gas entropy. Thus, the excess entropy has a negative
value. It can also be expanded in
an infinite series, $S_{ex}= S_2+S_3+....S_n$, where $S_n$ is the entropic contribution due to n-particle spatial correlation\cite{Kirkwood}.
For a binary system, the entropic contribution due to the pair correlation ($S_2$) can be written in terms of
the partial radial distribution functions,
\begin{equation}
\frac{S_{2}}{k_{B}}=-\frac{\rho}{2} \sum_{\alpha,\beta}x_{\alpha} x_{\beta} \int_0^{\infty} \{g_{{\alpha}{ \beta}}(r) \ln g_{{\alpha} {\beta}}(r)- [g_{{\alpha}{\beta}}(r)-1]\} d {\bf r},
\label{s2_final}
\end{equation}
\noindent where $ g_{{\alpha}{ \beta}}(r)$ is the pair correlation between particles of type $\alpha$ and $\beta$, $\rho$ is the density of the system,
$x_{\alpha}$ is the mole fraction of component $\alpha$ in the mixture, and $k_B$ is the Boltzmann constant.
Thus the excess entropy can be written as a contribution from the pair, $S_2$ and higher order terms, $\Delta S$. The $\Delta S$ contains all the contribution
beyond two particles\cite{giaquinta-1,giaquinta-2,giaquinta-3,giaquinta-4} and is written as,
\begin{equation}
\Delta S= S_{ex}-S_2=\sum_{n=3}^{ \infty}S_{n},
\end{equation}
\noindent where, $S_n$ is the entropic contribution due to n-particle spatial correlation\cite{Kirkwood}.
In liquid state theory, the structure of a liquid is primarily described by radial distribution function and hence its
entropic contribution plays a crucial role in predicting the dynamical quantities. For simple liquids the pair entropy $S_{2}$ provides $80\%-90\%$ contribution to
the excess entropy \cite {Baranyai-cp, murari_charu_2010_jcp}.
Based on this microscopic picture, Dzugutov proposed a relationship between pair excess entropy $S_2$ and dynamical quantity which
is given by $X^*=C_1 \exp [-K_1 S_{2}]$\cite{Dzugutov}. Therefore, the Dzugutov scaling law is considered as a special case of the Rosenfeld scaling.
Although the Rosenfeld relation is semi-empirical in nature, Bagchi and coworkers have derived it using the connection between excess entropy and Zwanzig's rugged energy landscape model of diffusion
\cite{zwanzig,saikat_jcp,bagchi_seki}.
Samanta {\it et al.} have found that the mode coupling diffusion coefficient can be fitted to the Rosenfeld scaling under certain approximations\cite{swapanghoshprl}.
In a recent study we find that
mode coupling theory (MCT) vertex is related to pair excess entropy and higher order MCT calculation in the schematic limit can provide the phenomenological
Rosenfeld relation \cite{unravel}. Thus, the scaling law has been derived using different theories.
The Rosenfeld scaling relation is found to be valid for a wide variety of liquids including simple liquids \cite{murari_charu_2010_jcp},
water\cite{agarwal2010relationship}, ionic melts
\cite{agarwal2009relationship,Ruchi_charu_2006,manish_charu},
model polymeric melts\cite{Goel_charu} and even for the data obtained in
different experiments \cite{abramson2007viscosity,abramson2008viscosity,abramson2009viscosity}. However
the Rosenfeld behaviour is not same for all the systems.
For most of the systems like simple liquids and water
the scaling breaks down in the low temperature regime.
For ionic liquids, the Rosenfeld relationship is found to be valid up to much lower temperature and
the value of the Rosenfeld exponents are different from that found for simple liquids\cite{agarwal2009relationship}.
The ionic melts are known to be network forming liquids. Recently Coslovich and Pastore have proposed a new model of a simple liquid (NTW model)
which mimics the properties like the structure and the dynamics
of the network forming liquids \cite{coslovich_pastore_jpcm_ntw,szamel-ntw}. Since earlier studies have reported the Rosenfeld behaviour to be
different for simple and network forming liquids\cite{Murari_charu_jcp,agarwal2009relationship},
it will be interestingly to investigate the Rosenfeld scaling of the NTW model.
In this present work, using molecular dynamics simulations we perform a comparative study of the validity of the Rosenfeld relation for
a few simple liquids, the LJ system, its repulsive counterpart known as WCA (Weeks-Chandler-Andersen)
system and the NTW model. The earlier study of the Rosenfeld relationship for the LJ and WCA systems have been performed in a moderate density range ($1.05-1.25$)
\cite{Murari_charu_jcp}.
Here we study a higher density range from 1.2 to 1.6 and find that
the LJ and WCA systems
follow the Rosenfeld and Dzugutov scaling laws only up to moderately high temperature. On the other hand, similar to that found for ionic melts\cite{agarwal2009relationship}
the Rosenfeld relationship for the NTW model is valid upto a much lower temperature. However, the Rosenfeld exponent at high temperature appears
to be similar to that found for simple liquids which is different from the value obtained for the ionic melts\cite{agarwal2009relationship}. Thus our result shows that
the NTW model has mixed characteristics of simple liquids and ionic melts. We show
a connection between the validity of the Rosenfeld relation and the Arrhenius behaviour. Our study predicts that for the NTW model the validity
of the Rosenfeld relation at the low temperature is connected to it being a strong liquid. We also study the independent role of pair entropy and
residual multiparticle entropy on the dynamics as function of fragility of the system.
The paper is organized as follows:
The simulation details are given in Sec. II. In Sec. III we describe
the methods used for evaluating the various quantities of interest and provide other necessary background. Sec. IV contains a discussion of
presented results and Sec. V contains a brief conclusion.
\section{Simulation Details}
In this study, we perform extensive molecular dynamics simulations for three-dimensional binary mixtures in
the canonical ensemble. The system contains $N_A$ particles of type A and $N_B$ particles of type B under periodic boundary conditions. The total number
density is fixed at $\rho=N/V$ with the total number of particles $N = N_A + N_B$ and a system volume $V$ .
The models studied here, are the well-known models of glass-forming liquids: the binary Kob-Andersen Lennard-Jones (KALJ) liquids \cite{kob}and the
corresponding WCA version (KAWCA) \cite{chandler} and a network-forming (NTW) \cite{coslovich_pastore_jpcm_ntw}
liquid that mimics
$SiO_2$ with short-range spherical potentials. The molecular dynamics (MD) simulations have been carried out using the LAMMPS
package \cite{lammps}. For all state points, three to five independent samples with run lengths $>$ 100$\tau$ ($\tau$ is the $\alpha$-
relaxation time) are analyzed.
\subsection{ KALJ and KAWCA: binary mixture of Kob Andersen Lennard-Jones particles and corresponding WCA version}
The most frequently studied model for glass forming liquids is Kob-Andersen model which is a binary mixture (80:20) of Lennard-Jones (LJ) particles \cite{kob}.
The interatomic pair
potential between species $\alpha$ and $\beta$, with ${ \alpha,\beta}= A,B$,
$U_{\alpha\beta}(r)$ is described by a shifted and truncated Lennard-Jones (LJ) potential, as given by:
\begin{equation}
U_{\alpha\beta}(r)=
\begin{cases}
U_{\alpha\beta}^{(LJ)}(r;\sigma_{\alpha\beta},\epsilon_{\alpha\beta})- U_{\alpha\beta}^{(LJ)}(r^{(c)}_{\alpha\beta};\sigma_{\alpha\beta},\epsilon_{\alpha\beta}), & r\leq r^{(c)}_{\alpha\beta}\\
0, & r> r^{(c)}_{\alpha\beta}
\end{cases}
\label{LJ_pot}
\end{equation}
\noindent where $U_{\alpha\beta}^{(LJ)}(r;\sigma_{\alpha\beta},\epsilon_{\alpha\beta})=4\epsilon_{\alpha\beta}[({\sigma_{\alpha\beta}}/{r})^{12}-({\sigma_{\alpha\beta}}/{r})^{6}]$ and
$r^{(c)}_{\alpha\beta}=2.5\sigma_{\alpha\beta}$ for the LJ systems (KALJ) and $r^{(c)}_{\alpha\beta}$ is equal to the position of the minimum of $U_{\alpha\beta}^{(LJ)}$
for the WCA systems (KAWCA) \cite{chandler}. Length, temperature and
time are given in units of $\sigma_{AA}$, ${k_{B}T}/{\epsilon_{AA}}$ and $\tau = \surd({m_A\sigma_{AA}^2}/{\epsilon_{AA}})$,
respectively. Here we have simulated Kob Andersen Model
with the interaction parameters $\sigma_{AA}$ = 1.0, $\sigma_{AB}$ =0.8 ,$\sigma_{BB}$ =0.88, $\epsilon_{AA}$ =1, $\epsilon_{AB}$ =1.5,
$\epsilon_{BB}$ =0.5, $m_{A}$ = $m_B$=1.0. We have performed MD simulations in the canonical ensemble (NVT) using Nos\'{e}-Hoover thermostat with integration timestep 0.005$\tau$.
The time
constants for Nos\'{e}-Hoover thermostat are taken to be 100 time steps.
The sample is kept in a cubic box with periodic boundary condition.
System size is $N = 500$, $N_A = 400$ (N $=$ total number
of particles, $N_A$ $=$ number of particles of type A) and we have studied a broad range of density $\rho$ from 1.2 to
1.6 with different temperature ranges given in Table \ref{temp-ref}.
\subsection{NTW: tetrahedral network-forming liquids}
We study a model of network-forming liquids \cite{coslovich_pastore_jpcm_ntw} interacting via spherical short-ranged potentials.
This model is simple model and mimics $SiO_2$ glasses, in which tetrahedral networks strongly dominate the dynamics.
The interaction potentials are given as
\begin{equation}
U_{\alpha\beta}(r)=\epsilon_{\alpha\beta}[(\frac{\sigma_{\alpha\beta}}{r})^{12}-(1-\delta_{\alpha\beta})(\frac{\sigma_{\alpha\beta}}{r})^6].
\end{equation}
Here $\delta_{\alpha\beta}$ is the Kronecker delta function. The interaction is truncated and shifted at $r=2.5\sigma_{\alpha\beta}$. The size, mass, and energy
ratios are given as $\sigma_{AB}/\sigma_{AA}=0.49, \sigma_{BB}/\sigma_{AA}=0.85, m_B/m_A=0.57, \epsilon_{AB}/\epsilon_{AA}=24
, \epsilon_{BB}/\epsilon_{AA}=1$.
System size is $N = 500$, $N_A = 165$ (N $=$ total number
of particles, $N_A$ $=$ number of particles of type A). The number density of NTW mixtures is
$\rho = 1.655$.
These parameters are adjusted to reproduce
the radial distribution functions of the $SiO_2$ amorphous
states
which corresponds to the density $\rho = 2.37g\AA{}^{-3}$
of the so-called van Beest-Kramer-van Santen (BKS) model for
the silica glass \cite{van1990force,carre2007amorphous}.
\begin{table}
\caption{The temperature range simulated for different systems.}
\begin{center}
\begin{tabular}{ | l | l |}
\hline
system (density) & Temperature Range\\ \hline
LJ(1.2) &5.00-0.45\\ \hline
WCA(1.2) &5.00-0.30 \\ \hline
LJ(1.4) &5.00-1.00 \\ \hline
WCA(1.4) &5.00-0.95 \\ \hline
LJ(1.6) &5.00-1.89\\ \hline
WCA(1.6) &5.00-1.89\\ \hline
NTW(1.655) &1.00-0.31 \\ \hline
\end{tabular}
\end{center}
\label{temp-ref}
\end{table}
\section{Definitions and Background}
\subsection{Relaxation time}
We have calculated the relaxation times from the decay of the
overlap function q(t), from the condition $q(t = \tau_{\alpha})/N =
1/e$. $q(t)$ is defined as
\begin{eqnarray}
\langle q(t) \rangle \equiv \left \langle \int dr \rho(r, t_0 )\rho(r, t + t_0 )\right \rangle \nonumber\\
=\left \langle \sum_{i=1}^{N}\sum_{j=1}^{N} \delta({\bf{r}}_j(t_0)-{\bf{r}}_i(t+t_0)) \right \rangle \nonumber\\
=\left \langle \sum_{i=1}^{N} \delta({\bf{r}}_i(t_0)-{\bf{r}}_i(t+t_0)) \right \rangle \nonumber\\
+\left \langle \sum_{i}\sum_{j\neq i} \delta({\bf{r}}_i(t_0)-{\bf{r}}_j(t+t_0)) \right \rangle.
\end{eqnarray}
The overlap function is a two-point time correlation
function of local density $\rho(r, t)$. It has been used in
many recent studies of slow relaxation \cite{unravel,post_prl_long}.
In this
work, we consider only the self-part of the total overlap
function (i.e. neglecting the $i \neq j$ terms in the double
summation). Earlier it has been shown to be a good approximation to the full
overlap function. So,
the overlap function can be well approximated by its self part,
and written as,
\begin{eqnarray}
\langle q(t) \rangle \approx \left \langle \sum_{i=1}^{N} \delta({\bf{r}}_i(t_0)-{\bf{r}}_i(t+t_0)) \right \rangle.
\end{eqnarray}
Again, the $\delta$ function is approximated by a Heaviside step function $\Theta(x)$ which defines the
condition of βoverlapβ between two particle positions
separated by a time interval t:
\begin{eqnarray}
\langle q(t) \rangle \approx \left \langle \sum_{i=1}^{N} \Theta (\mid{\bf{r}}_i(t_0)-{\bf{r}}_i(t+t_0)\mid) \right \rangle \nonumber\\
\Theta(x) = 1, x \leq {\text{a implying βoverlapβ}} \nonumber\\
=0, \text{otherwise}.
\end{eqnarray}
The time dependent overlap function thus depends on
the choice of the cut-off parameter a, which we choose
to be 0.3. This parameter is chosen such that particle positions separated due to small amplitude vibrational motion are treated as the same, or that $a^2$ is comparable to
the value of the MSD in the plateau between the ballistic
and diffusive regimes.
Relaxation times obtained from the decay of the self intermediate scattering function $F_s (k, t)$ using the
definition $F_s (k, t = \tau_\alpha)$ = $1/e$ at a fixed $k\simeq 2\pi /r_{max}$, where $r_{max}$ is the position of the first peak of the radial
distribution function of A type particles. The self
intermediate scattering function is calculated from the
simulated trajectory as
\begin{equation}
F_s(k,t)=\frac{1}{N}\left \langle \sum_{i=1}^{N} \exp(-i{\bf{k}}.({\bf{r}}_i(t)-{\bf{r}}_i(0))) \right \rangle.
\end{equation}
For NTW model, the cut off parameter for overlap function is taken as $a=0.2$ , as the decay time is similar to the decay observed from
$F_s (k, t)$ at $k=8.2$\cite{szamel-ntw}.
Since relaxation times from $q(t)$ and $F_s (k, t)$
behave very similarly, we have used the time scale
obtained from q(t).
We find that the diffusivities (D), obtained from the mean squared
displacement (MSD) of the particles, follow a similar trend like inverse of relaxation time. However, we have used
the relaxation time data in this present work.
\subsection{Excess Entropy}
The thermodynamic excess entropy, $S_{ex}$, is defined as the difference between
the total entropy $(S_{total})$ and the ideal gas entropy $(S_{id})$ at
the same temperature $(T)$ and density $(\rho$) for all the model
glass formers\cite{murari_charu_2010_jcp,Murari_charu_jcp}. The entropy was initially evaluated at a high temperature ($T_r$)
and low reduced density ($\rho_r$)
where the system can be assumed to behave as an ideal gas.
Entropies at any other state points, relative to this reference ideal
state point, can be calculated using a combination of isochoric
and isothermal paths, ensuring that no phase transitions occur
along the path.
The entropy of the liquid has been calculated via thermodynamic integration
starting from the ideal gas binary mixture reference point, along the $T = T_r$ isotherm,
up to the studied $\rho = \rho_{t}$ density.
The entropy change along an isothermal path
is given by \cite{Murari_charu_jcp},
\begin{widetext}
\begin{equation}
S_{ex}(T_r, \rho_{t})=S(T_r,\rho_{t})- S(T_r,\rho_{r})=\frac{U(T_r,\rho_r)- U(T_r,\rho_t)}{T_r}-\int_{\rho_r}^{\rho_t}\frac{P(\rho)}{T_r}\frac{N}{\rho^2}d\rho,
\end{equation}
and along the isochoric path it is given by \cite{Murari_charu_jcp},
\begin{equation}
S_{ex}(T_t, \rho_{t})=S(T_t,\rho_t)- S(T_r,\rho_t)=\int_{T_r}^{T_t}\frac{1}{T} \left(\frac{\partial U}{\partial T}\right)_\rho dT.
\end{equation}
\end{widetext}
To calculate the integral in the above equation, we use a standard technique used for supercooled liquid \cite{Srikanth_PRL}. We fit the temperature dependence of the potential energy using the
functional form, $U=a+bT^c$, where $a,b,c$ are the fitting parameters. The low temperature trajectories have been averaged over $500\tau$ (where $\tau$
is the relaxation time defined in Sec 3A).
\section{Results}
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{fig1.eps}
\caption{The scaled relaxation time vs. excess entropy ($S_{ex}$). In the high temperature regime, the Rosenfeld scaling law follow a universal behavior with the exponent 0.8.
The $\ln \tau^*$ is shifted by $-0.2$ for NTW model to obtain the master plot. Note that, the slope remains unaltered with this shifting. The slope of the black dashed line is close to 0.8.
(inset) For the NTW model the scaled relaxation time vs. excess entropy ($S_{ex}$). The low temperature
regime is fitted to another straight line with different slope which is equal to $3.35$.}
\label{scale_tau_vs_Sex}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{fig2.eps}
\caption{The scaled relaxation time vs. pair entropy ($S_{2}$). In the high temperature regime, the Dzugutov scaling law follow a universal behavior with the exponent 0.8.
The slope of the black dashed line is close to 0.8. (inset) For the NTW model the scaled relaxation time vs. pair entropy ($S_{2}$). The low temperature
regime is also fitted to another straight line with different slope which is equal to $2.52$.}
\label{scale_tau_vs_S2}
\end{figure}
In case of simple liquids at the high temperatures, the dimensionless scaled relaxation time ($\tau^*$) follows the Rosenfeld relation which is given by,
\begin{equation}
\tau^* =C \exp (-K S_{ex}),
\end{equation}
\noindent where, $C$, $K$ are the fitting parameters and $S_{ex}$ is the excess entropy.
The scaled relaxation time $\tau^*$ can be written as $\tau^*=\tau \rho^{-1/3}(k_B T/m)^{-1/2}$, where $m$ is the mass of the particle. Note that, for NTW model we
have used the mass as $m=\sum_i x_i m_i$, where $x_i$ is the mole fraction of the i-th particle.
We plot the Rosenfeld behavior both for LJ and WCA systems at different
densities and for the NTW model at $\rho=1.655$ (Fig \ref{scale_tau_vs_Sex}). We find that the relationship is valid at high temperature regime.
Interestingly all the systems show a master plot at high temperature and
the Rosenfeld exponent ($K$) is close to $0.8$.
Note that earlier studies have reported that the Rosenfeld scaling is valid at high temperature both for LJ and WCA systems with the same value of the exponent $K$
\cite{Murari_charu_jcp}.
However the density range studied here is much higher, thus, predicting that the Rosenfeld relationship is valid over a wider range of densities.
Our study also shows that although at low temperatures the WCA system does not follow any density-temperature scaling
\cite{tarjus_epje,Pedersen-prl,unravel},
at high temperatures it shows a master plot. The NTW model on the other hand, although is a model of a simple liquid, is known to show characteristics of ionic
melts, to be specific the $SiO_2$ system \cite{coslovich_pastore_jpcm_ntw}. However the value of the Rosenfeld scaling exponent ($K$) for NTW model is similar to that of the simple liquids and
different from that reported for $SiO_2$ \cite{agarwal2009relationship}.
Our next observation is that in the low temperature regime although the LJ and WCA systems show a breakdown of the scaling law, the NTW model follows the scaling law with
higher
value of the exponent (inset of Fig \ref{scale_tau_vs_Sex}). This behavior of the NTW model is
similar to that reported for the ionic melts \cite{agarwal2009relationship}.
Thus, our study reveals that although the NTW model follows the Rosenfeld behaviour up
to a much lower temperature like other
network forming liquids, its Rosenfeld exponent is similar to that found for simple liquids ($K\sim0.8$). Therefore,
the NTW model appears to have mixed characteristics of network
forming liquids and simple liquids.
As mentioned in the introduction, the pair entropy $S_2$ provides $80\%-90\%$ contribution to
the excess entropy \cite {Baranyai-cp, murari_charu_2010_jcp,unravel} and thus, the Rosenfeld relationship can be written in terms of $S_2$ which is given by,
\begin{equation}
\tau^* =C_1 \exp (-K_1 S_{2}),
\end{equation}
where, $C_1$, $K_1$ are the fitting parameters and $S_{2}$ is the pair entropy and this scaling behavior is known as Dzugutov scaling law. In Fig \ref{scale_tau_vs_S2}
we plot the Dzugutov scaling behavior and similar to the Rosenfeld relationship we find it to be valid at high temperature regime.
The value of the exponent $K_1 \sim 0.8$, this indicates that both the Rosenfeld
and Dzugutov scaling follows an universal exponent. We also find that similar to the Rosenfeld relation,
the Dzugutov scaling behavior is valid for NTW model upto a much lower temperature (inset of Fig \ref{scale_tau_vs_S2}) with a higher value of exponent.
Next we explore the origin of validity of the Rosenfeld scaling for the network forming liquid even in the lower temperature regime.
For any system
at high temperatures, the temperature dependence of the relaxation time can be expressed in terms of Arrhenius law which is given by,
\begin{equation}
\tau = \tau_0 \exp (E/T),
\label{arh}
\end{equation}
where, E is the activation energy. Note that the ionic melts and NTW model are strong liquids which follow Arrhenius behaviour even in the lower temperature
regime.
In Fig \ref{NTW-tau-rosen-linear}a we plot the relaxation time for NTW model and find that both the high temperature and low temperature regimes can be fitted to two
independent straight lines with the activation energies $E (high)=1.50$ and $E (low)=5.83$, respectively.
The Rosenfeld scaling law can also be written in terms of unscaled relaxation time \cite{unravel} given by,
\begin{eqnarray}
\tau = C^\prime \exp (-K^\prime S_{ex}).
\label{rosenfeld-unscaled}
\end{eqnarray}
We plot the $\tau$ vs $-S_{ex}$ in Fig \ref{NTW-tau-rosen-linear}b and find that the Rosenfeld behavior is valid up to a low temperature regime with different values of the
exponent which are given in Table \ref{parameters}. Note that the break in the Rosenfeld behaviour happens exactly at the same position where the break in the Arrhenius
behavior is observed. Thus, these two plots show qualitatively similar behavior.
\begin{figure}[h]
\centering
\subfigure{
\includegraphics[width=0.4\textwidth]{fig3a.eps}}
\subfigure{
\includegraphics[width=0.4\textwidth]{fig3b.eps}}
\subfigure{
\includegraphics[width=0.4\textwidth]{fig3c.eps}}
\caption{ For the NTW model (a) the temperature dependence of the relaxation time. Both the high and low temperature regimes follow Arrhenius relationship and can be fitted to
two independent straight lines. (b) The validity of the Rosenfeld relationship. The break in both the plots (a and b) occur at the same value of the $\tau$. (c) $S_{ex}$ vs $1/T$ shows a linear
behavior over the whole temperature regime. }
\label{NTW-tau-rosen-linear}
\end{figure}
\begin{table}[h]
\caption{ The fitting parameters of the Rosenfeld scaling,$\tau = C^\prime \exp (-K^\prime S_{ex})$.}
\begin{tabular}{|c|c|c|}
\hline
NTW & $\ln C^\prime$ &$K^\prime$ \\ \hline
high T&3.07 & 0.86 \\ \hline
low T & 12.98 &3.10 \\ \hline
\end{tabular}
\label{parameters}
\end{table}
Now to establish a quantitative connection between the Arrhenius and the Rosenfeld relations, we equate Eq \ref{arh} and Eq \ref{rosenfeld-unscaled} and can write it as,
\begin{equation}
\frac{E}{T}\sim -K^\prime S_{ex}.
\label{equality}
\end{equation}
The above equation suggests that for strong liquids where E is independent of temperature if $S_{ex} \propto 1/T$,
the activation energy (E) is proportional to the Rosenfeld exponent ($K^\prime$).
Next we show, that for NTW model the $S_{ex}$ vs. $1/T$ plot indeed shows a linear behavior (Fig \ref{NTW-tau-rosen-linear}c) for the whole temperature range.
Thus, from Eq \ref{equality} we expect that
\begin{equation}
\frac{E (low)}{E (high)} \sim \frac{K^\prime(low)}{K^\prime(high)}.
\end{equation}
We calculate the ratio of the activation energies and find that $\frac{E (low)}{E (high)}=3.89$ which is close to the value of
$\frac{K^\prime(low)}{K^\prime(high)}=3.60$.
Note that for most of the systems both the Rosenfeld relationship and the Arrhenius behaviour are known to be valid only in the high temperature regime.
For the fragile systems, in the low temperature regime, the activation energy becomes temperature dependent and there is a breakdown of the Arrhenius behaviour.
However, strong liquids like the NTW model follows an Arrhenius temperature dependence even in the low temperature regime (Fig \ref{NTW-tau-rosen-linear}a)
and also follows the Rosenfeld scaling (Fig \ref{NTW-tau-rosen-linear}b).
Thus we may conclude that the breakdown of the Rosenfeld relationship is connected to the
temperature dependence of the activation energy, E.
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{fig4.eps}
\caption{The scaled relaxation time vs. RMPE ($\Delta S$). To start with the same high temperature value of $\Delta S$, for NTW model the x-axis is
shifted by $-0.2$ .}
\label{scale_tau_vs_RMPE}
\end{figure}
\begin{figure}[h]
\centering
\includegraphics[width=0.45\textwidth]{fig5.eps}
\caption{Correlation of RMPE ($\Delta S$) with pair entropy($S_2$). For NTW model the $\Delta S$ (y-axis) is
shifted by $-0.2$ .}
\label{S2_vs_RMPE}
\end{figure}
Next we explore the role of residual multi particle entropy (RMPE) in the dynamics. In Fig \ref{scale_tau_vs_RMPE} we plot
the scaled relaxation time with RMPE, $\Delta S$. Similar to that found earlier \cite{Murari_charu_jcp}, for all the systems,
there is initially a positive correlation between the relaxation time and $\Delta S$ and as the temperature is lowered the correlation becomes negative.
This role reversal of RMPE has been observed earlier where we
have reported that the small positive value of RMPE speeds up the relaxation time \cite{bssb}.
Although at moderately high temperatures, the data show a master plot, at low temperature the dependence of relaxation time on RMPE becomes system dependent.
Similar observation is made in Fig \ref{scale_tau_vs_S2} where $\tau^*$ vs. $S_2$ also shows a deviation at low temperature.
Note that these systems cover a wide range of fragility where the LJ system and WCA at high density are the most fragile systems and NTW model is the least fragile one (strongest).
Thus, the $S_2$ and $\Delta S$ dependence of the scaled relaxation time $\tau^*$ show a variation with fragility. It appears that for the strong liquid, there is a weaker
dependence of $\tau^*$ on $S_2$ and $\Delta S$.
To understand this, in Fig \ref{S2_vs_RMPE} we plot $\Delta S$ with pair entropy for all the systems and find a data collapse. The plot shows that
at high temperatures, there is a positive correlation of $S_2$ and RMPE up to a point where $S_2=-2.5K_B$ \cite{Murari_charu_jcp}.
Below $S_2=-2.5 K_B$, $S_2$ decreases with increasing $\Delta S$ and the rate of the negative correlation is
independent of fragility.
Since the $S_2$ and the $\Delta S$ act in opposite directions, their combined effect which is seen in the $S_{ex}$ and thus in the relaxation time is much weaker.
This is precisely the reason behind the spread observed in Fig \ref{scale_tau_vs_S2} and Fig \ref{scale_tau_vs_RMPE}.
Note that $\tau^*$
vs. $S_{ex}$ plot shows a less spread with fragility (Fig \ref{scale_tau_vs_Sex}).
\section{Conclusion}
In this present work we find that both the Rosenfeld and Dzugutov scaling laws are valid at the high temperature regime for LJ and WCA systems ($\rho=1.2-1.6$) and
for the NTW model ($\rho=1.655$). Interestingly, for all the systems the exponents of both the
scaling laws are universal at high temperatures\cite{Murari_charu_jcp} and different from the value predicted for the ionic melts \cite{agarwal2009relationship}.
For the LJ and the WCA systems both the scaling laws break down in the low temperature regime.
However, we show that like other network forming liquids (ionic melts)\cite{agarwal2009relationship}, the scaling laws for the NTW model
are valid even in the lower temperature regime. Thus, although in terms of validity of these
semi-empirical relationships, the NTW model behaves like other network forming liquids, the value of the exponent appears to be similar to that found for the other simple
liquids. Our study establishes that the Rosenfeld and the Arrhenius relations are correlated both qualitatively and quantitatively.
Thus, for the NTW model, the validity of the Rosenfeld relationship at lower temperatures can be connected to it being a strong liquid following Arrhenius
behaviour in this regime. We also study the independent role of pair entropy and residual multiparticle entropy on the dynamics for the systems
with different fragilities. Our study reveals that for stronger liquids the dynamics has a weaker dependence on the $S_2$ and $\Delta S$. This weaker dependence
can be traced back to the fact that the effect of $S_2$ and $\Delta S$ on the dynamics act in the opposite directions with a similar rate.
\section{Acknowledgements}
This work is dedicated to the memory of Professor Charusita Chakravarty.
A. Banerjee and M. K. Nandi contributed equally to this work.
SMB thanks Charusita Chakravarty and Srikanth Sastry for discussions.
Authors thank Department of Science and Technology (DST-SERB), India and CSIR-Multi-Scale Simulation and Modeling project for financial supports.
AB thanks CSIR and MKN thanks UGC for fellowship.
| 865223c21ef82eaea84659ca8566a2ec05c76e26 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
The absence of readily discoverable new physics (NP) at the LHC has presented the physics community with a
formidable puzzle. While the arguments for NP ``not too far" from the weak scale still loom large, there is a
distinct desire to explore wider (and wilder) theoretical options away from a simply realized weak-scale supersymmetry,
or extra space dimensions. One possible strategy to look for new physics is to abandon theoretical
preconceptions, and start looking for non-standard signatures that the NP could present.
Large classes of models offer promising avenues
for a non-standard signal in the production of new exotic particles (possibly of electroweak-scale mass)
with subsequent decay away from the interaction point (see {\em e.g.} \cite{Graham:2012th,Craig:2015pha,Izaguirre:2015pga,Batell:2016zod}). While both ATLAS and CMS have performed corresponding studies in a variety of contexts
and for different ranges of displacement \cite{CMS:2014hka,Aad:2015rba,Aad:2015uaa}, it has been recently pointed out that a dedicated and relatively inexpensive detector
\cite{Chou:2016lxi} could extend the physics reach into cases where the decay
lengths are on the order of $O(100\,{\rm m})$ and beyond.
When both the production and decay of an exotic state $S$ occur through one and the same coupling constant, the chances of detecting such
NP at the LHC experiments are not great. Indeed, a large displacement implies a very small value for the coupling, which in turn
leads to very inefficient production rates. Therefore, an ideal case for the collider studies
would be when the production and decay occur through different
coupling constants, and $\lambda_{\rm production} \gg \lambda_{\rm decay}$. For the pair-produced exotics, such a hierarchy can be made
``natural" as the $\lambda_{\rm decay}\to 0$ limit could lead to an enhanced symmetry.
If the main signal to search for is an {\em appearance} of abnormal energy deposition or exotic vertex some distance from the interaction
point, it is then very important to know how small $\lambda_{\rm decay}$ is allowed to be. In more practical terms, one would like to know if
there is an external to the LHC physics constraint on the lifetimes $\tau_S$ of such exotic particles. An obvious source for such a constraint can be
early cosmology. The big bang nucleosynthesis (BBN), and its overall agreement with observations
\cite{Cyburt:2015mya} (apart from the unclear status of $^7$Li)
can provide a limit on the lifetimes of such particles. In order to derive such limits, one would have to make a
fairly natural assumption that the Universe was indeed as hot as $T\sim m_S\sim~$electroweak~scale at some point in its history. Subsequent thermal evolution to the BBN temperatures involves self-depletion via $SS\to {\rm SM}$ due to $\lambda_{\rm production} $, in an expected WIMP-type annihilation process, and late-time decay of $S\to {\rm SM}$ where depending on lifetimes and the properties of the
decay products the BBN outcome may get affected. These mechanisms are well-understood in the BBN literature
(see {\em e.g.} \cite{Jedamzik:2009uy,Pospelov:2010hj} for reviews). We will require that the late decay of $S$ provides a small and acceptable
perturbation to the standard BBN (SBBN) outcome, which in turn will limit $\tau_S$.
In this paper, we analyze a fairly minimal model, where a new singlet scalar has predominantly a quadratic coupling to the Higgs boson that regulates both its production at colliders and the intermediate cosmological abundance at $T_{\rm BBN} \ll T \ll m_S$. Given that the
model is very predictive, it allows to place robust bounds on lifetimes of such particles with a minimum amount of model dependence.
We find that for most of the analyzed parameter space with $m_S < m_h/2$, the intermediate abundance of such particles is
large enough to affect the neutron-proton freeze out ratios at relevant temperatures. This allows us to set fairly robust bounds on
lifetimes of such particles, which come out to be remarkably strong, and shorter than $0.1$ seconds. In what follows we describe the model and
the cosmological history of $S$ (section II); derive the impact on the BBN (section III); present our results (section IV), and provide related discussion (section V).
\section{The minimal Higgs portal model}
We consider the simplest extension of the SM by a singlet scalar field $S$.
A new singlet scalar $S$ can have two interaction terms with the Standard Model (SM) at the renormalizable level, in addition to trilinear and quartic self-interactions.
In this scenario, the Lagrangian of the singlet sector (including the SM) generically takes the form
\begin{equation}
\mathcal{L}_{H/S} = \mu^2 H^\dagger H - \lambda_H\left(H^\dagger H \right)^2 - V(S) - A S H^\dagger H - \lambda_S S^2 H^\dagger H+ \mbox{kin. terms}.
\end{equation}
The Higgs expectation value $v= 246~{\rm GeV}$ is assumed to correspond to a global minimum.
The self-interaction potential $V(S) = \lambda_4 S^4 +\lambda_3 S^3 + \frac{m_{S0}^2}{2} S^2$ can be redefined in such a way that the
linear term is absent. It is important that the $A,~\lambda_3 \to 0$ and $\langle S \rangle =0$
limit would correspond to the case of stable $S$ particles. To simplify the discussion without sacrificing much generality, we
take $\lambda_{3,4} \to 0$ and assume $Av \ll m_{S0}^2, ~\lambda_S v^2$.
The physical mass of $S$ receives a contribution from the electroweak symmetry breaking, $m_S =\sqrt{m_{S0}^2 + \lambda_S v^2}$. At linear order in $A$, the mixing angle $\theta$ between physical
excitations $S$ and $h$ is
\begin{equation}
\theta = \frac{Av}{m_h^2-m_S^2}\left(1-\frac{\lambda_Sv^2}{m_S^2}\right).
\end{equation}
The $\lambda_S$ term arises because the $S$ field develops a small $A$-controlled vacuum expectation value.
The mixing parameter $\theta$ leads, via the $A$ coupling constant, to the decay of $S$ particles, which can be readily
derived from
\begin{equation}
{\cal L}_{\rm decay} = S \times \theta \sum_{\rm SM} O_{h},
\label{Oh}
\end{equation}
where $O_h$ is the set of the standard Higgs interaction terms, with the Higgs field removed: {\em e.g.} $O_h = (m_f/v) \bar ff $
for an elementary SM fermion $f$.
This Yukawa-type coupling to the SM has been tested in rare meson decays~\cite{OConnell:2006rsp,Pospelov:2007mp,Batell:2009jf,Schmidt-Hoberg:2013hba,Clarke:2013aya} and in proton fixed-target experiments~\cite{Alekhin:2015byh}. The model is mostly ruled out for large mixing angles $\theta \gtrsim 10^{-4}-10^{-2}$ over the \mbox{$m_S \sim$ MeV - 5 GeV} mass range. The proposed experiment SHiP could potential improve current sensitivity down to $\theta \sim 10^{-6}$ for $m_S \sim$~few~GeV~\cite{Alekhin:2015byh}.
In the limit of $\theta \to 0$, $S$ is stable and could be the dark matter~\cite{Silveira:1985rk,McDonald:1993ex,Burgess:2000yq}. Various limits arise from searches in direct and indirect detection if the particle is stable (see Refs.~\cite{Cline:2013gha,Athron:2017kgt} for recent reviews), but $\lambda_S$ is generically bounded from the constraints on invisible Higgs decay, independently of the direct detection limits. The Standard Model Higgs has a well-predicted decay rate into SM particles of $\Gamma_{SM} = 4.07$ MeV. So far, the properties of 125 GeV resonance are remarkably consistent with the SM Higgs, and therefore there is little doubt that its width is close to $\Gamma_{SM}$. The invisible branching ratio of Higgs decay to $SS$
final state is
\begin{align}
\label{Gh}
\Gamma_{h\to SS} &=\frac{\lambda_S^2 v^2}{8\pi m_h} \sqrt{1-\frac{4m_S^2}{m_h^2}}, \\
Br(h\to SS ) &= \frac{\Gamma_S}{\Gamma_S + \Gamma_{SM}} \simeq 10^{-2} \,\left(\frac{\lambda_S}{0.0015}\right)^2,
\label{Gh1}
\end{align}
where in the last line we assumed $Br(h\to SS) \ll 1$ and $m_S \ll m_h$. The experimental upper bound on the invisible branching ratio of a SM Higgs is 0.19 (at $2\sigma$)~\cite{Belanger:2013xza}, which translates into an upper bound on $\lambda_S$
\begin{equation}
\lambda_S \lesssim \frac{0.007}{\left(1-\frac{4m_S^2}{m_h^2}\right)^{1/4}}.
\end{equation}
If $S$ is to be stable, such small couplings would lead to an excessive abundance of $S$, which invalidates the $Z_2$ symmetric case,
and forces us to include the decay term. From now on, we will consider $\theta \neq 0$, or in other words the case of unstable $S$ particles.
Since our analysis is motivated by the
LHC physics, we will use $Br(h\to SS )$ as an input parameter, and substitute $\lambda_S$ everywhere employing
(\ref{Gh}) and (\ref{Gh1}).
\subsection{Decay products}
Since $S$ interacts with the SM in the same fashion as the Higgs with an additional $\theta$ mixing factor (\ref{Oh}), its decay properties are similar to those of a light Higgs boson. For the derivations of the actual limits on
the lifetime of $S$, we need to know its mesonic and nucleonic decay branching ratios.
The decay channels of a light Higgs have been considered in the early years of the Weinberg-Salam electroweak model~\cite{Ellis:1975ap}, with additional refinements as SM particles, hadronic resonances were being discovered and final-state interactions better understood~\cite{Ellis:1979jy,Raby:1988qf,Truong:1989my}. Hadronic decays in the mass range $2m_\pi < m_S \lesssim 4 {~\rm GeV}$ are still poorly understood, with models varying by as much as a few orders of magnitude near the di-kaon threshold~\cite{Clarke:2013aya}.
The leptonic decay channels are straightforward, with the decay rate given by
\begin{equation}
\Gamma_{S\to l \bar{l}} = \frac{\theta^2}{8\pi} \frac{m_l^2}{v^2} m_S \left(1-\frac{4m_l^2}{m_S^2}\right)^{3/2}. \label{eq:Gamff}
\end{equation}
If the decaying product is a pair of heavy quarks, there are $\mathcal{O}(1)$ corrections coming from the 1-loop QCD vertex correction~\cite{Cline:2013gha}, which yields the following correction factor~\cite{Drees:1990dq} to the fermionic expressions~(\ref{eq:Gamff})
\begin{equation}
f_q = 3\left[1 + \frac{4 \alpha_s}{3\pi} \left(\frac{9}{4}+\frac{3}{2}\ln\frac{m_q^2}{m_S^2}\right)\right]
\end{equation}
and the factor of three comes from the number of color charges. For better accuracy, we use the higher order perturbative results from the \textsc{HDecay} code~\cite{Djouadi:1997yw} for $m_S > 2.5{~\rm GeV}$.
Metastable mesons, such as $\pi^\pm$ and $K^\pm,\bar K^0, K^0$ are ``important" decay products, as they can participate in the
charge-exchange reactions with nucleons and shift the $n-p$ balance, hence affecting the whole nucleosynthetic chain.
In the mass range where the perturbative QCD calculations are no longer valid, we base our baseline calculations on Ref.~\cite{Bezrukov:2009yw}. The scalar-pion interaction can be extracted from the low-energy expansion of the trace of the QCD energy-momentum tensor~(see for {\em e.g.}.~\cite{Voloshin:1985tc,Leutwyler:1989xj}) by integrating out the three heavy quarks and using chiral perturbation theory on the remainder, yielding the effective Lagrangian~\cite{Bezrukov:2009yw}
\begin{equation}
\mathcal{L}_{S\pi\pi} = \frac{4}{9}\frac{\theta}{v} S\left(\frac{1}{2} \partial_\mu \pi^0 \partial^\mu \pi^0 + \partial_\mu \pi^+ \partial^\mu \pi^-\right)
- \frac{5}{3} \frac{\theta m_\pi^2}{v} S \left( \frac{1}{2}\pi^0\pi^0 + \pi^+ \pi^- \right),
\end{equation}
where we have inserted the SM numerical values for the number of heavy quarks and the first coefficient of the QCD beta function. This leads to
decay width to charged pions
\begin{equation}
\Gamma_{S\to \pi^+ \pi^-} = 2\Gamma_{S\to \pi^0 \pi^0} = \frac{\theta^2}{16\pi}\frac{m_S^3}{v^2} \left(\frac{2}{9} + \frac{11}{9} \frac{m_\pi^2}{m_S^2} \right)^2
\sqrt{1- \frac{4m_\pi^2}{m_S^2}}. \label{eq:LowEPi}
\end{equation}
This result is however not applicable far above the pion threshold, as final-state resonances would drastically affect this prediction. Instead, we use the pion and kaon decay width described in Ref.~\cite{Donoghue:1990xh}, where the authors matched the next-to-leading order corrections of the low-energy theorems to the dispersion results from the $\pi \pi$ phase-shift analysis above 600 MeV from the CERN-Munich group~\cite{Hyams:1973zf}. The photon decay channel is added with the prescription detailed in Ref.~\cite{Spira:1997dg}. Finally, there is a gap for $1.4~{~\rm GeV} < m_S < 2.5~{~\rm GeV}$ where no analytical treatment is entirely trustworthy, as this includes new resonances strongly coupled to $\eta \eta$ and other potential hadronic cannels. We simply follow Ref.~\cite{Bezrukov:2009yw} and interpolate between the two regimes, under the assumption that there is no order of magnitude deviation in this mass range. The branching ratios and the lifetime for $\theta = 10^{-6}$ are displayed in Fig.~\ref{fig:S_Br_tau}.
\begin{figure}
\centering
\includegraphics[width= 0.50\columnwidth]{SBr.pdf} \hspace{0.5cm}
\includegraphics[width= 0.45\columnwidth]{Slifetime.pdf}
\caption{\textit{Left}: Branching ratios of the scalar $S$ in our baseline decay model. See text for details. \textit{Right}: Scalar $S$ lifetime of our baseline model and the spectator model for the mixing angle $\theta = 10^{-6}$.}
\label{fig:S_Br_tau}
\end{figure}
As an alternative decay spectrum model, we also display the perturbative spectator approach~\cite{Gunion:1989we,McKeen:2008gd,Alekhin:2015byh}, where the relative decay width above the kaon threshold are given by
\begin{equation}
\Gamma_{\mu^+\mu^-} : \Gamma_{KK} : \Gamma_{\eta \eta} =
m_\mu^2 \beta_{\mu}^3 : 3\frac{9}{13}m_s^2\beta_K^3 : 3\frac{4}{13}m_s^2\beta_\eta^3 , \label{eq:SpectBr}
\end{equation}
with $\beta_i = \sqrt{1-4m_i^2/m_S^2}\Theta(m_S - 2m_i)$, $\Theta$ being the step-function, and we adopt the running of $s$ quark mass following Ref.~\cite{Spira:1997dg}. The pion contribution is kept as in equation~(\ref{eq:LowEPi}) and then we use the \textsc{HDecay} output at the $c$-quark threshold and above to match our baseline model.
For $m_S$ of several GeV and heavier, decays with final state nucleon-antinucleon pairs are possible. Even though the
branching to such states are generally lower than 10\%, the effect on BBN can be quite significant, and therefore these are by
far the most important channels for $\tau_S \gtrsim 1\;\sec$. On top of direct and for the most part subdominant contributions
from $S\to \bar nn,...$, we need to take into account the (anti-)nucleon states that emerge from the hadronization of the
quark decay products and heavy $B$-meson fragmentations.
\subsection{Cosmological metastable abundance}
After the temperature drops below $m_S$, the interaction of $SS$ pairs with the SM shifts towards the annihilation,
resulting in an intermediate (metastable) population of $S$ bosons.
In the mass range that we consider, the $S$ annihilation is dominated by the $s$-channel reactions $SS \to h^* \to XX$, where on the
receiving end are the pairs of the SM states $XX$ created by a Higgs-mediation process. The annihilation cross section $\sigma v$ generically takes the form
\begin{equation}
\sigma v(s) = \frac{8 \lambda_S^2 v^2}{(s-m_h^2)^2 + m_h^2 \Gamma_{{\rm SM}+S}^2}\frac{\Gamma_{{\rm SM}}^{m_h\to\sqrt{s}}}{\sqrt{s}}, \qquad \qquad
\langle \sigma v \rangle = \frac{\int_{4m_S^2}^\infty ds\; \sigma v (s)\;s \sqrt{s-4m_S^2}K_1\left(\frac{\sqrt{s}}{T}\right)}{16 T m_S^4 K_2^2\left(\frac{m_S}{T}\right)}.
\label{eq:sigmav}
\end{equation}
This formula recast the rate in terms of a Higgs width $\Gamma_{{\rm SM}}^{m_h\to\sqrt{s}}$ with a fictitious mass of $\sqrt{s}$. This form encompasses both perturbative and non-perturbative channels in the $h^*$ decay rate (with the substitution $m_h^*\to \sqrt{s}$), which we have described above. In the standard WIMP freeze out paradigm, a DM particle freezes out at $T_{\rm f.o.} \sim m_{DM}/20$, $\langle \sigma v \rangle$ is simply the nonrelativistic limit $\sigma v (\sqrt{s} =2 m_{DM})$ and the relic density can be conveniently approximated as $\Omega_{DM} h^2 \sim 0.11\times 1{\rm pb}/\langle \sigma v \rangle$. This result emerges as a solution to the Boltzmann equation\footnote{We use the standard variable definitions, where $Y = n_S/ s$ is the $S$ abundance normalized on the entropy density $s$, $x = m/T$ is the dimensionless inverse temperature, $H$ is the Hubble rate, $h_{\rm eff}$ is number of entropic relativistic degrees of freedom and $Y_{eq}$ is the normalized thermal equilibrium $S$ number density.}~\cite{Steigman:2012nb}
\begin{equation}
\frac{dY}{dx} = \frac{s\langle \sigma v \rangle}{Hx} \left [ 1 + \frac{1}{3} \frac{d(\ln h_{\rm eff})}{d (\ln T)} \right ] \left( Y_{\rm eq}^2 - Y^2\right), \label{eq:BoltzmannRelic}
\end{equation}
when the freeze out occurs in the exponentially falling region of the equilibrium density $Y_{\rm eq}(T)$. For a much smaller annihilation cross
section, $\langle \sigma v \rangle \ll 1\;{\rm pb}$, $Y$ departs from the equilibrium value earlier, possibly near the relativistic plateau $Y_{\rm eq} = n_{\rm eq}/ s \to 45 \zeta (3) /2 \pi^4 h_{\rm eff} (T)$ for $x \ll 1$. Since the nonrelativistic annihilation cross section in the minimal Higgs portal model ranges from $10^{-3}$ to $10^{-14}$ pb for $m_S \sim 1{~\rm MeV} - 60 $ GeV and \mbox{$Br(h\to SS) \sim 0.1 - 0.001$}, we numerically integrate equation~(\ref{eq:BoltzmannRelic}) to determine the metastable $S$ abundance. The results are shown in Fig.~\ref{fig:YS_abundance}, normalized to the baryon number density for a more intuitive interpretation of its impact on BBN in the following section.
For $m_S \simeq m_h/2$, the $\sigma v$ cross section evaluated at $s=4m_S^2$ is a poor approximation, as it fails to capture the strong energy dependence of the cross section near the resonance at $\sqrt{s} = m_h/2$ \cite{Griest:1990kh}. The sharp drop in the abundance above $m_S \sim 45{~\rm GeV}$ is due to the resonant contribution to the thermally averaged cross section, leading to a delayed freeze out and drastic decrease in metastable $S$ abundance. Our numerical results agree with the semi-analytic treatment of Ref.~\cite{Cline:2013gha}. For very light $m_S$, one can see that the freeze out abundances are large, and the relative spread between different input values of $Br(h\to SS)$ gets smaller, as the annihilation cross section becomes very small and the freeze out happens in the semi-relativistic regime $x_{\rm f.o.} \sim \mathcal{O}(1)$ and asymptote to the $Y_{\rm eq}$ relativistic plateau for small $m_S$. The only difference at the lightest masses is from $Y_{\rm eq}^{rel} \propto 1/h_{\rm eff}(T)$. Since $h_{\rm eff}$ is a monotonic function of temperature, weaker annihilation cross sections freeze out earlier, at a higher temperature, thus yielding smaller abundances (as seen in the $m_S = 5{~\rm MeV}$ curves in Fig.~\ref{fig:YS_abundance}). This is in contrast with the standard freeze out in the non-relativistic regime, with final abundances inversely proportional to the cross section.
We note in passing that the strong-interaction-related uncertainty ``propagates" outside the $m_S \sim 2m_\pi - 2 m_c$ window. For example, because of the relativistic freeze out, for $m_S$ smaller $ 2 m_\pi$ the hadronic channels may turn out to be important.
\begin{figure}
\centering
\includegraphics[width= 0.45\columnwidth]{Ysvsx.pdf} \hspace{1cm}
\includegraphics[width= 0.45\columnwidth]{SAbundance.pdf}
\caption{\textit{Left}: Temperature evolution ($x=m/T$) of the $Y_S$ intermediate abundance for $m_S = 5{~\rm MeV}$ and 500 MeV for the three benchmark higgs branching ratios. \textit{Right}: Metastable abundance of $S$ prior to its decay normalized over the baryon density. Values shown for $Br(h\to SS) = 10^{-1}$, $10^{-2}$ and $10^{-3}$. The dashed lines correspond to the perturbative spectator model.}
\label{fig:YS_abundance}
\end{figure}
\section{Big Bang Nucleosynthesis}
The formation of light nuclei is one of the earliest probes of NP in cosmology along with far less certain constraints imposed by the inflationary framework. BBN is well-understood within SM physics, and its outcome agrees with observational data for $^4$He and D. $^7$Li has an outstanding factor of $\sim 2-3$ discrepancy between theory and observations~\cite{Cyburt:2015mya}, with the caveat that the observed abundances
may have been affected by stellar evolution. Nevertheless, the overall success over a wide range of abundances can be used to constrain various types of NP~\cite{Pospelov:2010hj}.
The initial BBN stage is the neutron-proton ratio $n/p$ freeze out. Maintained in equilibrium by electroweak interactions at high temperatures, the neutron abundance follows $n/p \sim e^{-Q/T}$, where $Q=m_n - m_p -m_e \simeq 1.293{~\rm MeV}$, until the epoch when the weak processes decouple around temperatures of 0.7 MeV. The outcome, $n/p \simeq 1/6$, is quasi-stable, decreasing to $n/p \simeq 1/7$ at the end
of the ``deuterium bottleneck". The latter terminology is used to indicate a much delayed onset of nuclear reactions controlled by a relatively shallow $n-p$ binding energy.
Once the Universe runs out of photons that can efficiently dissociate deuterium, the bulk of the nucleosynthetic
reactions occurs at $t_{\rm deut} \sim 200 $ seconds. $^4$He has a large binding energy per nucleon, and the reactions leading to it are less Coulomb-suppressed than for heavier elements. Consequently,
most neutrons end up in the final $^4$He abundance (expressed in mass fraction from the total baryon mass) $Y_p \simeq 2 (n/p)/\left( 1 + n/p\right) \simeq 0.25$.
Traces of neutrons and incomplete nuclear burning of $A=2,~3$ nuclei light nuclei
result in the left-over abundances of $^3$He and D.
Beyond the $^4$He atomic number, the deepest bound nucleus is $^{12}$C, but its formation is completely
suppressed since it would need to be produced by a triple $^4$He collision. The $2\to2$ reactions $p+{}^4$He and $^4$He + $^4$He are also ineffective at producing heavier nuclei as the $A = 5$ and $A = 8$ elements are all unstable. The only remaining possibilities are $^4{\rm He}+{}^3{\rm He} \to {}^7{\rm Be} + \gamma$ followed by a $\beta$ decay to yield $^7{\rm Li/H} \sim \mathcal{O}(10^{-10})$ and $^6$Li formed at the $^6{\rm Li/H} \sim \mathcal{O}(10^{-14})$ via $^4$He-D fusion.
For the problem at hand - the determination of the upper limit on the $S$ lifetime - few of these details matter. This is because of relatively large
metastable abundances affecting the earliest stages of nucleosynthesis, primarily via the $n/p$ ratio.
\subsection{Neutron Enrichment}
Ample abundances of $S$ particles ($n_S \sim 10^{2}-10^{9} \times n_b$) flood the Universe with final state mesons and nucleons
that in turn could spoil the final light nuclei abundances. For example, at temperatures $T\sim 0.5 $ MeV, the protons are
$\sim6$ times more abundant than neutrons, but this ratio can be easily changed due to meson-induced charge exchange reactions.
At these temperatures, the probability of $p\to n$ conversion from charged pions is
\begin{equation}
P_{n\to p} \simeq n_p \langle \sigma v \rangle_{pn} c\tau_{\pi^+} \simeq 2\times \frac{10^{21}}{\rm cm^3}\times 1.5 \mb \times 2.6 \times 10^{-8}{~\rm sec}\times c \simeq 2.5\times 10^{-3}.
\end{equation}
It is then clear that injection of $O(10^3)$ mesons per nucleon at these temperatures can drastically increase the $n/p$ freeze out abundance.
Similarly, direct baryonic injection of $n\bar{n}$ and $p\bar{p}$ will have a similar effect on the $n/p$ ratio. On the other hand, if $S$ decays happen before the $n/p$ freeze out, the additional $p\to n$ conversions would not be as efficient, being washed out by the on-going weak interaction conversions.
The limit of the exclusion region in the $Y_S/\tau_S$ parameter space ($Y_S \equiv n_S/n_b$ from now on) is determined by solving the Boltzmann equation with the injection of charge exchange inducing particles. Given that the abundances of $S$ particles are large, the main constraints can be derived from the $n/p$ freeze out ratio.
To that effect, we would not need a complete BBN framework, but only a subset of the whole code that deals with $n\leftrightarrow p$ conversions. We follow the semi-analytic treatment by Mukhanov \cite{Mukhanov:2003xs}, that approximates
$n\leftrightarrow p$ weak conversion rates by a few integrals over thermal distributions, and assumes a ``step-like" disappearance of charged leptons below $T=m_e$,
\begin{align}
\Gamma_{n \nu_e \to p e^- } &= \frac{1+3g_a^2}{2\pi^3} G_F^2 Q^5 J(1;\infty), &\Gamma_{ p e^- \to n \nu_e } &=e^{-Q/T} \Gamma_{n \nu_e \to p e^- }, \qquad \\
\Gamma_{n e^+ \to p \bar{\nu}_e } &= \frac{1+3g_a^2}{2\pi^3} G_F^2 Q^5 J(-\infty;-\frac{m_e}{Q}), &\Gamma_{p \bar{\nu_e} \to n e^+ } &=e^{-Q/T} \Gamma_{n e^+ \to p \bar{\nu}_e }, \\
J(a,b) &\equiv \int_a^b \sqrt{1-\frac{(m_e/Q)^2}{q^2}}\frac{q^2(q-1)^2\;dq}{(1+e^{\frac{Q}{T_\nu}(q-1)})(1+e^{-\frac{Q}{T}q})} ,
\end{align}
where $g_a \simeq 1.27$ is the standard nucleon axial-vector coupling, $Q = m_n - m_p - m_e \simeq 1.293{~\rm MeV}$, and $G_F$ is the Fermi constant. The reverse reaction rates are found by detailed balance. We evaluate $J$ numerically and solve for the electron-neutrino temperature $T_\nu$ by entropy conservation, assuming a $\nu_e$ decoupling temperature of 2 MeV, which reproduces the correct entropy degrees of freedom at lower temperature~\cite{Srednicki:1988ce}. It is then straightforward to solve numerically the differential equation for $X_n = n_n/n_b$,
\begin{equation}
\frac{d X_n}{dT} = \frac{\Gamma_{n \nu_e \to p e^- }+\Gamma_{n e^+ \to p \bar{\nu}_e } }{T H(T)} \left(X_n -(1-X_n)e^{-Q/T} \right) +
\frac{\Gamma_n X_n}{T H(T)}, \label{eq:Boltzmann}
\end{equation}
where the last term represents the neutron decay with $\Gamma_n^{-1} = 880~{\rm sec}$.
This equation is approximately valid until the rapid switch-on of the nuclear reaction rates at the end of the deuterium bottleneck.
Within this approximation, one can determine the final temperature where the equation is valid by starting with $X_n = 1/2$ at early times, and solving for the deuterium bottleneck temperature by imposing $Y_p = 2 X_n (\Tdeut)= 0.25$. This results in $\Tdeut \simeq 0.068 {~\rm MeV}$ or $t_{\rm deut} \simeq 276$ sec. We take this approximation as our baseline SBBN model, which is then modified by the inclusion
of extra sources and sinks for $n,~p$, and new $n\leftrightarrow p$ reactions.
To constrain the parameter space of a species decaying into charged mesons or baryons, we proceed by solving the Boltzmann equation that includes new interactions. We will require that $Y_p$ does not deviate from SBBN by more than 4\%,
\begin{equation}
\Delta Y_p \equiv |Y_p -Y_p^{\rm SBBN}| < 0.01,
\end{equation}
which is a rather generous allowance for the errors, considering the tight observational constraints on primordial helium abundance
\cite{Cyburt:2015mya}. Consequently, it will result in conservative limits of $\tau_S$.
\subsubsection{Meson-Mediated mechanism}
Only long-lived mesons have an opportunity to interact with the baryon bath and induce proton-neutron conversions. As such, only $\pi^\pm$, $K^\pm$ and $K_L$ have lifetimes in excess of $\tau \sim 10^{-8}$ seconds, and can induce $p\leftrightarrow n$ via strong interactions. For temperatures relevant for the $n/p$ freeze out, the density of charged leptons is very high, and mesons are efficiently stopped
by the primordial plasma. We assume that they are efficiently thermalized, and take the relevant pion-induced reactions at threshold~\cite{Reno:1987qw,Pospelov:2010cw} ($c=1$),
\begin{align}
\pi^- + p &\to n + \gamma : &(\sigma v)^{\pi^-}_{pn(\gamma)} &\simeq 0.57 \mb, &Q &= 138.3{~\rm MeV},\\
\pi^- + p &\to n + \pi^0 : &(\sigma v)^{\pi^-}_{pn(\pi^0)} &\simeq 0.88 \mb, &Q &= 3.3{~\rm MeV},\\
\pi^+ + n &\to p + \gamma : &(\sigma v)^{\pi^+}_{np(\gamma)} &\simeq 0.44 \mb, &Q &= 140.9{~\rm MeV},\\
\pi^+ + n &\to p + \pi^0 : &(\sigma v)^{\pi^+}_{np(\pi^0)} &\simeq 1.26 \mb, &Q &= 5.9{~\rm MeV}.
\end{align}
The reverse reactions are irrelevant due to the short lifetime of $\pi^0$'s and the need for non-thermal $\gamma$'s of $\sim 140$ MeV energy. The $\pi^-$ reactions are to be added to the r.h.s of Boltzmann equation~(\ref{eq:Boltzmann}) via the additional term
\begin{equation}
\left.\frac{d X_n}{dT}\right|_{\pi-} = \frac{-1}{TH(T)}n_{\pi^-}^{\rm inj} \left( \langle\sigma v \rangle_{pn(\pi^0)}^{\pi^-}+ \langle\sigma v \rangle_{pn(\gamma)}^{\pi^-}\right)(1-X_n) ,
\label{eq:dXdTpi}
\end{equation}
and similarly for the $\pi^+$ reactions. The ambient population of injected pions from a $S$ decay with \mbox{$Br(S\to \pi^+\pi^-)=\xi_{\pi^\pm}$} is $n_{\pi^\pm}^{\rm inj} \simeq \xi_{\pi^\pm}\Gamma_S \tau_{\pi^\pm} Y_S n_b(T) e^{-t\Gamma_S}$, $t \simeq 2.42\sec\; ({\rm MeV}/T)^2/\sqrt{g_\star}$
and the thermal cross section are taken at their threshold value $\langle \sigma v \rangle_{np}^{\pi^+}=(\sigma v)_{np}^{\pi^+}$.
Reactions with pairs of charged particles in the initial states, such as $\pi^-p$, will be somewhat enhanced due to the
Coulomb attraction, which provides a small but non-negligible correction. We account for it following Ref.~\cite{Pospelov:2010cw}.
\begin{figure}
\centering
\includegraphics[width= 0.41\columnwidth]{Xn.pdf} \hspace{1cm}
\includegraphics[width= 0.45\columnwidth]{YSlims.pdf}
\caption{\textit{Left}: $X_n$ evolution for the SBBN and the injection of pions, kaons, baryons and muons (neutrinos) as described in the text for lifetimes of $0.05$ seconds with the initial $Y_S$ abundance tuned to yield $\Delta Y_p=0.01$ (maximum allowed shift of $Y_p$). The baryonic injection is taken at $\kappa = 0.5$ (full line), the lines for $\kappa = 1$ (dashed) and $\kappa = 0.2$ (dotted) are also displayed. \textit{Right}: Limit of injected pairs for each channel as a function of the $S$ lifetime. The upper-right dotted line for $\kappa = 0.2$ is at $Y_p = 0.26$, the upper-left dotted island yields $Y_p = 0.24$. }
\label{fig:Xn_Gamtau}
\end{figure}
The implementation of the charged kaons reactions is similar to the pion case, but the dominant reactions are rather different. The direct charge exchange between neutral and charged kaons is
\begin{equation}
\bar{K}^0 + n \to K^- + p: \qquad (\sigma v)^{\bar{K}^0}_{pn(K^-)} \simeq 10 \mb, \qquad Q = 5.3{~\rm MeV}, \label{eq:Kmp}
\end{equation}
with similar cross section for a charge-conjugated reaction, $K^0p\to K^+n$. For neutral kaons, the effects induced by
$K_L$ are the most important, and we use $\sigma (K_L n \to K^- p ) \simeq \frac{1}{2}\sigma (\bar{K}^0n \to K^-p)$ and~(\ref{eq:Kmp}) to find $(\sigma v)^{\bar{K}^0}_{pn(K^-)} \simeq 4.5 \mb$.
Additionally, efficient reactions can also proceed via $s$-quark being incorporated inside a hyperon that subsequently decays into $p/n$ + $X$.
The inclusive threshold cross section found by weighting each hyperon with their branching ratios to $p/n$ are~\cite{Pospelov:2010cw}
\begin{align}
K^- + p &\to n + X : &(\sigma v)^{K-}_{pn} &\simeq 32\mb, \\
K^- + n &\to p + X : &(\sigma v)^{K-}_{np} &\simeq 13\mb, \\
K_L + p &\to n + X : &(\sigma v)^{K-}_{pn} &\simeq 6.5\mb, \\
K_L + n &\to p + X : &(\sigma v)^{K-}_{np} &\simeq 16\mb.
\end{align}
Notice the absence of corresponding hyperon reactions initiated by $K^+$ due to the presence of anti-$s$ quark.
Representative examples of $X_n(T)$ evolutions and the sensitivity to $\xi_{\pi^\pm}Y_S/\tau_S$ parameter space are shown in Fig.~\ref{fig:Xn_Gamtau}. Left panel displays significant modifications to the evolution of neutron abundance at $\tau_S =$0.05~seconds with adjustable initial abundance, yielding $\Delta Y_p = 0.01$. The departure
from $X_n = 0.5$ at high temperatures is clearly visible. (In fact, for short $\tau_S$, the kaon injection channel at early times leads to a shift of the equilibrium value of $X_n$
to $(\sigma v)^{K^-+K_L}_{pn}/((\sigma v)^{K^-+K_L}_{pn}+(\sigma v)^{K^-+K_L}_{np}) \simeq 0.45$.)
As the temperature lowers, the Coulomb-enhanced reaction becomes stronger. For meson injection, these reactions enhance the $p\to n$ conversion, keeping $X_n$ away from the SBBN value.
Right panel gives a boundary of the exclusion regions for different
injection modes.
In addition to the already described channels, charged kaons also give rise to a population of secondary charged pions that should also be included in the analysis of $p\leftrightarrow n$ transitions. Since the constraints are already stronger than for
the charged pion case, we neglect this effect, which leads to more conservative bounds.
\subsubsection{Direct baryonic injection mechanism}
If $S$ is heavy enough, the end-products after hadronization of the primary decay products ({\em e.g.} $b$ or $c$ quarks)
may contain baryons. Since $S$ has no baryon number, one should expect an equal number of baryons and anti-baryons in the final states. Therefore, one should expect the injection of $n\bar{n}$, $p\bar{p}$, $\bar n p$ and $p \bar n$ pairs, as well as (in principle) baryonic states with higher multiplicities.
The hadronization process and decay of heavy quarks
produce much more light mesons than baryons, and a complete analysis must include a Monte Carlo study of the hadronization process (see Ref.~\cite{Kohri:2001jx} for benchmarks of heavy unstable particles decaying into 2 hadronic + 1 leptonic jets in the early BBN epoch).
Assuming that the heavy quarks inside baryons decay due to the ``main" weak decay sequence, $b\to c\to s \to u$, one should also expect a
somewhat large number of the final states with a proton or anti-proton over neutron or anti-neutron.
We will tune the branching models of $S$ to available particle data on proton production, and
take $N_{n} = \kappa N_{p}$ and $N_{\bar n} = \kappa N_{\bar p}$. Furthermore, due to a more frequent
appearance of up-quark at the end of the decay chain, we would take $\kappa \simeq 0.5$ on average.
As in the case of mesons, the thermalization of baryonic decay products is quick (see {\em e.g.}~\cite{Kawasaki:2004qu}). As a baryonic pair is created in the decay, the baryon is added to the existing population of $n$ or $p$. The anti-baryon will, however, annihilate with either
$p$ or $n$ and dissipate into lighter mesons. If it annihilates with its own antiparticle, there is no net change in $n/p$, but an annihilation with the other species induces a net $n-p$ change. The probability $P_{k\bar{l}}^{i\to j}$ of a net charge exchange $i\to j$ from a $k\bar{l}$ injection is simply given by the weighted annihilation rates
\begin{equation}
P_{p\bar{p}}^{n\to p} = \frac{X_n \langle \sigma v \rangle_{n\bar{p}}}{X_n \langle \sigma v \rangle_{n\bar{p}}+ (1-X_n)\langle \sigma v \rangle_{p\bar{p}}},
\qquad
P_{p\bar{n}}^{n\to p} = \frac{(1-X_n) \langle \sigma v \rangle_{p\bar{n}}}{X_n \langle \sigma v \rangle_{n\bar{n}}+ (1-X_n)\langle \sigma v \rangle_{p\bar{n}}}
\end{equation}
and similarly for the $n\bar{p}$ and $n\bar{n}$ injections. The baryonic annihilation rates are given by ~\cite{Reno:1987qw}
\begin{equation}
\langle \sigma v \rangle_{n\bar{n}} = \langle \sigma v \rangle_{p\bar{p}} / C = 37\mb, \qquad
\langle \sigma v \rangle_{n\bar{p}} = \langle \sigma v \rangle_{p\bar{n}} = 28\mb,
\end{equation}
where the $p\bar{p}$ has the low-$v$ Coulomb correction $C(v)$. The implementation of these processes in the Boltzmann equation then
require additional terms
\begin{equation}
\left.\frac{d X_n}{dT}\right|_{pn} = \frac{-\xi_p \Gamma_S e^{-t\Gamma_S}}{TH(T)}\left( -P_{p\bar{p}}^{n\to p} - \kappa P_{p\bar{n}}^{n\to p} + \kappa P_{p\bar{n}}^{n\to p} + \kappa^2 P_{n\bar{n}}^{n\to p} \right).
\end{equation}
As before, the outcome is displayed in Fig.~\ref{fig:Xn_Gamtau}. Again, for short $S$ lifetimes and large $Y_S$, the large numbers of injected particles completely dictates the early $X_n$ value. The constraint on $Y_S$ goes up more sharply in the short $S$ lifetime limit. There is a significant dependence on $\kappa$ for $\tau_S \gtrsim 0.1$ sec, which is washed out by the SM electroweak interactions at earlier times. If we take the extreme limit $\kappa \to 0$, no neutrons are injected and the $p\bar{p}$ pair can only further decrease the $n/p$ ratio, thus constrained by the lower $Y_p$ limit 0.24. On the other hand, a symmetric injection $\kappa =1 $ enhances the $n/p$ ratio as the anti-baryon mostly annihilates on protons, more abundant than neutrons by a factor of $\sim 6-7$ after the standard $n/p$ freeze out. For $\kappa \gtrsim 0$, the final $Y_p$ can either be increased or decreased, depending if the $S$ particles decay away before or after the displaced $X_n$ equilibrium crosses the SBBN $n/p$ freeze out curve. As shown for $\kappa = 0.2$ in Fig.~\ref{fig:Xn_Gamtau}, there is a $Y_p = 0.24$ exclusion island at low lifetimes and larger lifetimes are constrained by $Y_p = 0.26$. We use $\kappa = 0.5$ as our baryonic injection constraint benchmark.
\subsubsection{Muon-Mediated mechanism}
Muon injection physics differs from the previous scenarios of meson and baryon injection. The direct charge-exchange is through the weak force, as opposed to the strong force in the other cases, and is completely negligible over the lifetime of the muon. Instead, the reactions can happen via the energetic neutrinos emitted by the muon decays. The case for muon injection after $t \sim 100\;\sec$ has been covered in Ref.~\cite{Pospelov:2010cw}, to which we refer the reader for details. Assuming stopped muons, the authors solved for the injected neutrino energy spectrum, including redshifting and averaged over flavour oscillations, to be integrated in the $n-p$ conversion rate. At earlier times, we know background neutrinos are coupled to $e^\pm$ down to $T\simeq2{~\rm MeV}$, and energetic injected neutrinos must accordingly deplete their energy efficiently as well. Summing over the possible interactions with the background neutrinos and $e^\pm$~\cite{Dolgov:1997mb}, the collision rate of an injected electron-neutrino with the bath is given by
\begin{align}
\nonumber
\Gamma^{\nu_e}_{\rm coll} \left(E_\nu , T \right) &= \frac{7\pi}{135} G_{\rm F}^2 E_{\nu} \left[ \left(5 + g_{\rm L}^2 + g_{\rm R}^2\right) T_\nu^4 + 4 \left(g_{\rm L}^2 + g_{\rm R}^2 \right) \eta(T) \;T_\gamma^4 \right], \label{eq:Collnue}\\
&\simeq \left( \frac{E_\nu}{32{~\rm MeV}}\right) \left[ \frac{5.7}{\sec} \left(\frac{T_\nu}{1{~\rm MeV}}\right)^4 +\frac{1.3}{\sec} \eta(T_\gamma)\left(\frac{T_\gamma}{1{~\rm MeV}}\right)^4
\right],
\end{align}
$g_{\rm L} = 1/2 + \sin^2\theta_{\rm w}$, $g_{\rm R} = \sin^2\theta_{\rm w}$, while $\eta(T) = 1$ for $T\gtrsim m_e$ and exponentially falls to 0 at lower temperatures. We follow the implementation of Ref.~\cite{Pospelov:2010cw} and correct for the removal of energetic neutrinos by adding an effective collision lifetime in the neutrino energy distribution (normalized on $n_b$)
\begin{equation}
f_e\left(T,E_\nu\right) = \Gamma_S Y_S \int_T^\infty \frac{dT_1\; e^{-t_1\Gamma_S}}{H(T_1)T_1} F_e \left(E_\nu , \frac{E_0 T}{T_1}\right) e^{-\int_T^{T_1} dT_2\; \frac{\Gamma_{\rm coll}(E_\nu \frac{T_2}{T}, T_2)}{H(T_2)T_2}},
\end{equation}
where $F_e$ is the distribution at injection time $T_1$, averaged over flavour oscillations. The charge-exchange rate to be inserted in the Boltzmann equation~(\ref{eq:Boltzmann}) is
\begin{equation}
\Gamma_{pn}^\nu = n_b (T) \int_0^{E_0} \sigma_{pn}^{\bar{\nu}} f_e (T,E_\nu)\; dE_\nu
\end{equation}
and similarly for the reverse $np$ direction. The resulting constraints are shown in Fig.~\ref{fig:Xn_Gamtau}. Our results lean on the conservative side on a few assumptions. For simplicity, we assumed one collision for the neutrino thermalization, instead of following energy degradation over a shower of multiple interactions. Moreover, we took the collision time of the electron-neutrino, even though there are muon-neutrino states in the oscillations. Since $\Gamma_{\rm coll}^{\nu_e} > \Gamma_{\rm coll}^{\nu_\mu}$, we over-estimate the actual collision time and the overall conversion rate should be slightly larger.
\subsubsection{Meson injection from residual annihilations}
In addition to its decay products, $S$ can also inject particles in the cosmic medium via $SS$ annihilations to charged pions. The injected pions interact with the cosmic medium in the same way as from $S$ decays described above. The Boltzmann equation takes the addition term~(\ref{eq:dXdTpi}), with the injected pion density now given by
\begin{equation}
n_{\pi^\pm}^{\rm ann} = \tau_{\pi^\pm} n_S^2(T) \left\langle \sigma v \right \rangle_{\pi^+\pi^-} = \tau_{\pi^\pm} Y_S^2 n_b^2(T) e^{-2t\Gamma_S} \left\langle \sigma v \right \rangle_{\pi^+\pi^-},
\end{equation}
where $\langle \sigma v \rangle_{\pi^+\pi^-}$ is the non-relativistic annihilation cross section $\sigma v (2m_S)$~as per Eq.~(\ref{eq:sigmav}), rescaled by the pionic branching ratio at $\sqrt{s}= 2 m_S$. The $n_{\pi^\pm}^{\rm ann} \propto n_S^2 \propto T^6$ dependence imply a much stronger impact at high energies, enforcing the displaced initial condition $X_n^i \simeq 0.47$. As $S$ decays away, its impact on $X_n(T)$ is even more rapidly exponentially suppressed and its constraints are less stringent than decays at very short lifetimes. The bounds from annihilation are given in the $Y_S^2 \langle \sigma v \rangle_{\pi^+ \pi^-} - \tau_S$ parameter space and displayed in Fig.~\ref{fig:Y2sigann}.
\begin{figure}
\centering
\includegraphics[width= 0.48\columnwidth]{Y2sigann.pdf}
\caption{Constraints on $Y^2_S \langle \sigma v\rangle_{\pi^+\pi^-}$ from $SS$ annihilations into charged pions from the BBN $^4$He abundance at $Y_p = 0.26$.}
\label{fig:Y2sigann}
\end{figure}
\subsection{Energy density requirements}
The resultant BBN abundances depend on the nuclear reaction rates and how efficient they are as the Universe expands. One by one, the reaction rates drop out of equilibrium, as the Universe expands and cools.
If the Hubble rate is increased due to a large energy density locked in a dark sector, the active reaction time would shorten, potentially spoiling the SBBN results. For our study, the most important effect is the change of the Hubble rate during the $n/p$ freeze out, which again affects $Y_p$.
However, we can also use as a constraint a well-measured quantity in cosmology, the total energy density carried by neutrinos.
The neutrinos decouple from thermal processes at $T \sim 2 {~\rm MeV}$. If the decaying particle is heavy and does not decay into neutrinos, it will reheat electron-photon fluid with respect to the neutrinos, decrease $T_\nu/T_\gamma$ and equivalently lower $N_{\rm eff}$. The Planck collaboration measured $N_{\rm eff}= 3.04 \pm 0.33$ at $2\sigma$, including their CMB results and external cosmological data~\cite{Ade:2015xua}, which imposes $N_{\rm eff} > 2.71$ as a lower bound.
The energy densities and Hubble rate form a closed system of differential equations
\begin{equation}
\dot{\rho}_S + 3 H \rho_S = - \Gamma_S \rho_S, \qquad
\dot{\rho}_{rad} + 4 H \rho_{rad} = \Gamma_S \rho_S, \qquad
H^2 = \frac{8\pi G}{3}\left(\rho_{rad} + \rho_S\right), \label{eq:EnerDens}
\end{equation}
where we have assumed a non-relativistic $S$ and omitted the variation in relativistic degrees of freedom. Assuming step-like decoupling and changes in relativistic degrees of freedom, the $T$ evolution separates into 3 regions. For $T > T_{\nu}^{\rm decoup}$, neutrinos are in equilibrium with the electromagnetic bath and $\rho_S$ is injected equally in $e^\pm$'s, $\nu$'s and $\gamma$'s. For $T_{\nu}^{\rm decoup} > T > T_{m_e}$, the neutrinos are simply redshifted while the electron-photon bath is heated by the $S$ decays. For $T_{m_e} > T$, electrons become non-relativistic and transfer their entropy to photons, additionally heating the photon bath compared to the neutrino bath.
If $S$ does not dominate the energy density of the Universe before its decay, we can write $\rho_S = \delta_S \rho_{rad}^{SM}$, $\rho_{rad} = \rho_{rad}^{SM}(1+\delta_{rad})$ and expand~(\ref{eq:EnerDens}) around the $\delta$ perturbations to solve the system analytically. At linear order, we find the solutions
\begin{equation}
\rho_S(t) = \frac{c_S}{t^{3/2}}e^{-\Gamma_S t}, \qquad \rho_{rad}(t) = \frac{c^i_{rad}}{t^2}\left[1 +F(t)\right], \qquad
F(t) = \frac{c_S}{c^i_{rad}\sqrt{\Gamma_S}}\frac{1}{\Gamma_S t} \left[ \Gamma_{3/2}(\sqrt{\Gamma_S t})- \Gamma_{5/2}(\sqrt{\Gamma_S t}) + \frac{\sqrt{\pi}}{4}\right] \label{eq:rhos}
\end{equation}
where $\Gamma_{3/2}$, $\Gamma_{5/2}$ are incomplete Gamma functions and the integration constants $c_S$, $c_{rad}$ are set to have $\rho_S = m_S n_S$ and $\rho_{rad} = \rho_{rad}^{SM}$ at some early time $\Gamma_S t \ll 1$. After the neutrinos decouple, the injected energy is distributed to the photon-electron bath and its energy density departs for the neutrino bath
\begin{equation}
\rho^{\rm mid}_{\gamma}(t) = \tilde{g}_{\gamma+e}\frac{c^i_{rad}}{t^2}\left[1 + F(t)\right] + \tilde{g}_{\nu} \frac{c^i_{rad}}{t^2} \left[G(t) - G(t_\nu^{\rm decoup}) \right], \qquad
\rho^{\rm mid}_{\nu}(t) = \tilde{g}_{\nu}\frac{c^i_{rad}}{t^2}\left[1 + F(t)-G(t) + G(t_\nu^{\rm decoup}) \right], \qquad
\end{equation}
where $\tilde{g}_i \equiv g_{i}/(g_{\gamma+e}+ g_\nu)$ is the fraction of relativistic degrees of freedom of each bath, $t_\nu^{\rm decoup}$ the neutrino decoupling time and
\begin{equation}
G(t) = \frac{c_S}{2 c^i_{rad}}\sqrt{\frac{\pi}{\Gamma_S}}\erf\left(\sqrt{\Gamma_S t}\right)- \frac{c_S}{c^i_{rad}}\sqrt{t}e^{-\Gamma_S t}.
\end{equation}
Finally, after the electrons become non-relativistic, they effectively transfer their entropy to the photon bath. Assuming an instantaneous transition, entropy continuity implies an increase of energy density by a factor of $\delta = \left(g_e+g_\gamma\right)^{1/3}/g_\gamma^{1/3} = (11/4)^{1/3}$. Matching boundary conditions, the energy densities at late times are
\begin{align}
\rho^{\rm late}_{\gamma}(t) &= \tilde{g}_{\gamma+e}\delta\frac{c^i_{rad}}{t^2}\left[1 + \alpha F(t) + c \frac{t_e}{t}\right] + \tilde{g}_{\nu} \alpha \frac{c^i_{rad}}{t^2} \left[G(t) - G(t_e) + \delta \left(G(t_e)-G(t_\nu^{\rm decoup})\right) \right], \\
\rho^{\rm late}_{\nu}(t) &= \tilde{g}_{\nu}\frac{c^i_{rad}}{t^2}\left[1 + \alpha\left\{ F(t)-G(t) + G(t_e) +\delta( G(t_\nu^{\rm decoup})- G(t_e))\right\} + c \frac{t_e}{t} \right],
\end{align}
with $\alpha = 1/(\delta \tilde{g}_{\gamma+e} +\tilde{g}_{\nu})$ and $c$ a boundary condition that is irrelevant in the $t \to \infty$ limit.
The temperature-time dependence is found via $\rho_{rad}(t) = \pi^2 g_\star T^4/30$. Since the neutrino interaction rate scales as $\Gamma_{\nu_e}\sim T^5$, we find the neutrino decoupling time in the modified cosmology by equating $(T_\nu^{\rm decoupl})^5/H(T_\nu^{\rm decoupl}) = (T_\nu^0)^5/H_0(T_\nu^0)$, with $H$ the perturbed Hubble rate and $T_\nu^0$ the neutrino decoupling temperature in the SM. In the Maxwell-Boltzmann approximation, $T_\nu^0 = 2{~\rm MeV}$, but thermal refinements in the interaction rates and phase space tend to yield a lower value $T_\nu^0 = 1.4{~\rm MeV}$~\cite{Dolgov:2002wy}.
Then, at $\Gamma_S t \gg 1$, we can evaluate $T_\nu / T_\gamma$ and find
\begin{equation}
N_{\rm eff} = 3 \left(\frac{T_\nu}{T_\gamma}\right)^4 \left(\frac{11}{4}\right)^{4/3}
\simeq 3 \times\frac{\delta \tilde{g}_{\gamma + e} +\tilde{g}_\nu - \frac{c_S}{2c_{rad}^i}\sqrt{\frac{\pi}{\Gamma_S}}+ (1-\delta)G(t_e)+\delta G(t_\nu^{\rm decoup})}
{\delta \tilde{g}_{\gamma + e} +\tilde{g}_\nu + \frac{\tilde{g}_\nu }{\delta \tilde{g}_{\gamma + e}}\left( \frac{c_S}{2c_{rad}^i}\sqrt{\frac{\pi}{\Gamma_S}}-(1-\delta ) G(t_e) - \delta G(T_\nu^{\rm decoup})\right)}
\end{equation}
to constrain energy injection into electrons. We display in Fig.~\ref{fig:Neff} the departure from $N_{\rm eff} =3$ as a function of time for $\tau_S = 0.1 \sec$ and the two neutrino decoupling temperature benchmarks. The limits are also shown in units of stored energy density $m_S n_S/n_b$, where $m_S$ is in MeV. If the $S$ decay happens after the neutrino decoupling, all energy is deposited in the photon bath and the result is independent of our choice of $T_\nu^0$. If decays happen earlier, the photon and neutrinos are potentially still coupled and the energy emitted in $S$ decays only influences $N_{\rm eff}$ after decoupling. As such, the constraints has a $t\propto (1/T_\nu^0)^2$ dependence. We adopt the conservative side, $T_\nu^0 = 1.4{~\rm MeV}$, as our bounds on the $m_S-\tau_S$ parameter space. Notice the constraints for $SS$ annihilations to pions are much stronger and will be dominant when the pionic annihilation channel is open, \textit{i.e.} for $m_\pi < m_S < 2 m_\pi$.
\begin{figure}
\centering
\includegraphics[width= 0.48\columnwidth]{DNeff.pdf} \hspace{0.5cm}
\includegraphics[width= 0.48\columnwidth]{YmNeff.pdf}
\caption{\textit{Left}: Departure from the SM $N_{\rm eff}$ as the Universe cools down for electron injections (blue) and muon injections (orange). The extrema of the neutrino decoupling temperature ranges are shown in full lines and dashed lines as labeled in the figure. \textit{Right}: Bound of maximal stored energy decaying into electrons or muons as a function of particle lifetimes. The full line and dashed lines represent the neutrino decoupling temperatures as on the left. We also show for comparison some benchmark bounds in this parameter space from the BBN $Y_p$ results. The thin olive curves are the neutron enrichment constraint from annihilation into pions for $m_S = 140 {~\rm MeV}$ (solid) and $m_S = 275 {~\rm MeV}$ (dotted). The thin purple line is the $Y_p$ constraint for a $m_S = 250 {~\rm MeV}$ particle decaying into muons.}
\label{fig:Neff}
\end{figure}
\subsubsection{Energy injection partitioned between photon and neutrino baths (\textit{e.g. muon injection})}
The case for muon injection is somewhat interesting as its decay products, neutrinos and electrons, clearly thermalize in the two different baths, once everything is decoupled. Both $T_\gamma$ and $T_\nu$ will rise, but since the 2 neutrinos carry more energy than the electron for a muon decay, we expect a rise in $N_{\rm eff}$. More precisely, we solve a similar set of equations as~(\ref{eq:EnerDens}), except the photon bath absorbs a $\xi$ proportion of the $S$ decay energy and the neutrino bath gets the remaining $(1-\xi)$ portion. Before neutrino decoupling, the radiation bath evolves as in equation~(\ref{eq:rhos}). Each decay product carries on average the energy~\cite{Pospelov:2010cw}
\begin{equation}
\left\langle E_{e} \right \rangle = 37.0 {~\rm MeV}, \qquad \qquad \left \langle E_{\nu_e} \right \rangle = 31.7 {~\rm MeV}, \qquad \qquad \left \langle E_{\nu_\mu}\right\rangle = 37.0{~\rm MeV}.
\end{equation}
After neutrino decoupling, the energetic neutrinos can still collide with the ambient electrons until $\Gamma_{{\rm coll} -e }^{\nu_e} < H$, where $\Gamma_{{\rm coll} -e }^{\nu_e}$ is the collision rate with electrons only, the $T_\gamma$-dependent term in equation~(\ref{eq:Collnue}). Then, the energy distributed to the photon bath separates into two regimes
\begin{equation}
\xi_1 = \frac{\left \langle E_{\nu_e} \right \rangle+\frac{\Gamma_{{\rm coll} -e }^{\nu_e}}{\Gamma_{\rm coll}^{\nu_e}}\left \langle E_{\nu_e} \right \rangle + \frac{\Gamma_{{\rm coll} -e }^{\nu_\mu}}{\Gamma_{\rm coll}^{\nu_\mu}}\left \langle E_{\nu_\mu} \right \rangle}{m_\mu} \simeq 0.47, \qquad \qquad
\xi_2 = \frac{\left \langle E_{\nu_e} \right \rangle}{m_\mu} = 0.35,
\end{equation}
where the muon-neutrino collision term is given by
\begin{equation}
\Gamma^{\nu_\mu}_{\rm coll} \left(E_\nu , T \right) = \frac{7\pi}{135} G_{\rm F}^2 E_{\nu} \left[ \left(5 + \left(g_{\rm L}-1\right)^2 + g_{\rm R}^2\right) T_\nu^4 + 4 \left(\left(g_{\rm L}-1\right)^2 + g_{\rm R}^2 \right) \eta(T) \;T_\gamma^4 \right].
\end{equation}
Following the same procedure as before, we find
\begin{align}
N_{\rm eff} &= 3 \times \frac{\delta \tilde{g}_{\gamma + e +\tilde{g}_\nu }+ \frac{c_S\left(\delta \tilde{g}_{\gamma + e}(1-\xi_2)-\xi_2\tilde{g}_\nu\right)}{2c_{rad}^2 \tilde{g}_\nu}\sqrt{\frac{\pi}{\Gamma_S}}+\frac{C}{ \tilde{g}_\nu}}
{\delta \tilde{g}_{\gamma + e} +\tilde{g}_\nu - \frac{c_S\left(\delta \tilde{g}_{\gamma + e}(1-\xi_2)-\xi_2\tilde{g}_\nu\right)}{2c_{rad}^2\delta \tilde{g}_{\gamma + e}}\sqrt{\frac{\pi}{\Gamma_S}}-\frac{C}{\delta \tilde{g}_{\gamma + e}}},\\
C &= \delta (\xi_2-\xi_1)G(t_{\rm coll})+ \delta (\xi_1 - \tilde{g}_{\gamma + e}) G(t_\nu^{\rm decoup})+ \xi_2 (1-\delta) G(t_e),
\end{align}
with $t_{\rm coll}$ found by solving $\Gamma_{{\rm coll} -e }^{\nu_e} = H$. The physics is constrained by $N_{\rm eff} < 3.37$. The time-dependence of the departure from $N_{\rm eff} = 3$ is shown for $\tau_S = 0.2 \; \sec$ and the two choices of $T_\nu^0$ in Fig.~\ref{fig:Neff}. The corresponding constraints on the maximal stored energy for a given lifetime are shown on the right. For comparison with the muon-induced $Y_p$ bound, we display the curve for $m_S = 250 {~\rm MeV}$ from neutron enrichment in purple. Independently from the choice of $T_\nu^0$, the bounds from $Y_p$ $\mu$-injection are most constraining for $\tau_S \gtrsim 0.2 \;\sec$ while the annihilation to $\pi^+\pi^-$ provides the dominant constraint in the entire $2m_\mu < m_S < 2 m_\pi$ range.
\subsection{Late-time energy injection }
In the example of the $S$ particles coupled through the Higgs portal, the most stringent constraints on lifetime come from the
considerations of $n/p$ freeze out. In other models, with additional channels of annihilation that can suppress metastable abundances,
the constraints on lifetime would not be as stringent, and would mostly come from the considerations of late
energy injection. For completeness, we also discuss these constraints here.
Modification of BBN by unstable particles with lifetimes in excess of 200 seconds
has been considered in detail, both through hadronic~\cite{Reno:1987qw,Kawasaki:2004yh}, electromagnetic~\cite{Cyburt:2002uv} or combined~\cite{Kawasaki:2004qu,Jedamzik:2006xz,Pospelov:2010cw} energy cascades.
Hadronic injection after $t\gtrsim 200$ seconds is most efficient at modifying the final yields of the less abundant light nuclei D, ${}^3$He, ${}^6$Li,
and ${}^7$Li.
After most of ${}^4$He has been synthesized, the BBN enters the regime ($T \sim 50$ keV) when neutrons are rare, $O(10^{-5})$ or so, yet their
abundances are critical in determining the final abundance of deuterium. At that stage, any additional neutrons brought into
the system through external processes such as heavy particle decays lead to the increase of the deuterium abundance.
(Incidentally, it also leads to the suppression of $^7$Be and consequently of $^7$Li \cite{Reno:1987qw}.)
The increase of D production can be exacerbated by the hadro-dissociation of ${}^4$He in the process of slowing down of injected hadrons.
Additional production of $^3$He through spallation can also affect the $^3$He/D ratio \cite{Ellis:2005ii}.
Secondary and tertiary processes may also generate ${}^6$Li and $^9$Be \cite{Dimopoulos:1987fz,Pospelov:2010kq}.
Detailed studies of the ensuing constraints \cite{Jedamzik:2006xz} show strong sensitivity to hadronic (mostly nucleonic) decays of metastable
particles with lifetimes in the hundreds of seconds and longer, and initial abundances comparable or even smaller than that of baryons.
In recent years, these constraints have only got stronger, primarily due to steady observational progress in determination of primordial D/H \cite{Cooke:2013cba}.
If for some reason, hadrons and specifically nucleons are absent from the decay chains,
the abundances of light elements can be modified by the late injection of electromagnetic energy.
At early times this mechanism is inefficient, as radiation quanta with energy in excess of nuclear binding are quickly energetically degraded by
ambient plasma.
The photo-dissociation therefore sets in at late times leading to a suppression D ($t \gtrsim 10^4 $ seconds) and additional production
of $^3$He for $t \gtrsim 10^6$ seconds. Since typically 45\% of hadronic energy injection is dissipated electromagnetically in the hadronization cascade~\cite{Jedamzik:2006xz}, the late-time energy injection constraints on a heavy particle are dominated by the electromagnetic reactions in the BBN network.
\section{Results}
We are now in a position to perform a scan in parameter space of the minimal Higgs model, constrained by the consistency with BBN. In Fig.~\ref{fig:S_paramspace}, we display the parameter space, both in the lifetime and an effective decay length $L_{\rm dec} = c\tau_S\beta_S(E_S/m_S)$. We assume an average $E_S$ of 200 GeV, from a Higgs typically boosted at 400 GeV at the LHC.
The resulting constraints, along with the assumptions considered in each mass range are described below.
\begin{figure}
\centering
\includegraphics[width= 0.48\columnwidth]{taumSlim.pdf} \hspace{0.5cm}
\includegraphics[width= 0.48\columnwidth]{Ldecay.pdf}
\caption{\textit{Left}: Lifetime constraint as a function of the $S$ mass for three $h\to SS$ branching ratios. The lettered regions represent different assumptions or physics and are described in the text. The dotted lines correspond to the perturbative spectator model. \textit{Right}: Same as left, except transposed in the decay length of $S$, assuming it is boosted to $E_S = 200{~\rm GeV}$.}
\label{fig:S_paramspace}
\end{figure}
\begin{itemize}
\item \textbf{Region A $2 m_e < m_S < 2 m_\mu$} : The constraint comes from the decrease in $N_{\rm eff}$ with the entropy dump in the SM bath after neutrino decoupling. We take the neutrino decoupling temperature to be $T_\nu^0 = 1.4{~\rm MeV}$ as a conservative limit.
\item \textbf{Region B $m_\pi < m_S < 2 m_\pi$} : This region is dominated by the $SS$ annihilation to $\pi^+ \pi^-$. We also derived the same constraint as region A from $N_{\rm eff}$ up to $m_S =2 m_\mu$, in addition to the raised $N_{\rm eff}$ from decays into muons in the $2m_\mu < m_S < 2m_\pi$ and the $Y_p$ constraints from $S$ decaying into muons. They all yield weaker bounds, of $\tau_S > 0.3 \sec$ or longer.
\item \textbf{Region C $2 m_\pi < m_S < 2 m_K$} : The abundance $Y_S$ weighted by the pion branching ratio constrains the region via direct charged pion decays. We assume $2/3$ go into charged pions and $1/3$ is radiated away in $\pi^0$.
\item \textbf{Region D $2 m_K < m_S < 1.4{~\rm GeV}$} : The abundance $Y_S$ weighted by the kaon branching ratio constrains the region via direct charged kaons decays. We assume $1/2$ go into charged kaons and $1/2$ into $K^0 \bar{K}^0$. Only half of the neutral kaons
survive as $K_L$, creating similar in numbers metastable populations of $K_L$, $K^+$ and $K^-$.
\item \textbf{Region E $1.4{~\rm GeV} < m_S < 2 m_D$} : By strangeness conservation, we assume that all $s$-quarks yield a kaon, half charged and half neutral. Since we do not have model-independent branching ratios of $S$ in this mass regime, we vary the description according to the assumptions in each decay model. For the baseline model, we assume that 100\% decays to the kaons and apply our kaon injection constraints. For the perturbative spectator model, the kaon branching ratio is given by~(\ref{eq:SpectBr}), with non-negligible contributions from decays to pions, muons and eta mesons, resulting in weaker bounds until the $c$-quark threshold. At $m_S = m_c$ the hadronic modelling dependence largely goes away.
\item \textbf{Region F $2 m_D < m_S < 2 m_b$} : We utilize the branching fractions of $c\bar{c}$ from $e^+ e^-$ at $\sqrt{s} = 10.5{~\rm GeV}$ into $D$-mesons from Ref.~\cite{Lisovyi:2015uqa} and weight each channel by its inclusive $K^\pm$ branching ratios to find a hadronization yield of 0.63 $K^+K^-$ pair per $S$ decay into $c$-quarks. Rescaled by $Br(S\to c\bar{c})$, same constraints from kaon injection apply. Above the $2m_{\Lambda_c}$ threshold, a $c\bar{c}$ typically forms a $c$-baryon with a 0.06 probability~\cite{Lisovyi:2015uqa}, which then hadronizes to $p$ or $n$. We find this constraint weaker than the kaons injection and use the $K^+K^-$ result across this entire range.
\item \textbf{Region G $m_S > 2 m_b$} : The main decay channel here are pairs of $b\bar b$ quarks. The charged pion, charged kaons and proton multiplicities in the $b\bar{b}$ decay of a $Z$ boson are measured to be $18.44\pm 0.63$, $2.63\pm 0.14$ and $1.00\pm 0.08$ respectively by the ALEPH collaboration~\cite{Barate:1997ty}. We assume the ratio holds in the hadronization of lower centre-of-mass decays into $b\bar{b}$ and scale by the mean charge multiplicity fit~\cite{Sarkisyan:2015gca}
\begin{equation}
N_{ch}(s) = -0.577 + 0.394 \ln (s/s_0) + 0.213 \ln^2(s/s_0) + 0.005 (s/s_0)^{0.55},
\end{equation}
where $s_0 = 1{~\rm GeV}^2$. This fit agrees well in both $e^+e^-$ and $p\bar{p}$ collisions between $\sqrt{s} \sim 2{~\rm GeV}-2{~\rm TeV}$. This gives us an estimate for the baryon injection of the $b\bar{b}$ branching fraction of $S$. We further assume 50\% smaller injection of $n(\bar{n})$ to utilize our baryon injection constraints. The accompanying pions and kaons also independently yield comparable constraints, not shown in the figure.
\end{itemize}
\section{Discussion}
We have considered, in some detail, constraints on the lifetimes of the scalar particles, coupled to the Higgs portal via a minimal set of couplings. To stay relevant for the LHC, we have concentrated on $m_S < m_h/2$ case, that allows pair-production of $S$
states in the decay of Higgs bosons. The same coupling is responsible for the cosmological depletion of
$S$ particles, leading to their metastable abundance in the early Universe.
We find that throughout almost the whole mass range considered in this work, $2m_\mu < m_S < m_h/2$,
the constraints on the lifetime of $S$ particles are stronger than
$0.1$ seconds. Moreover, the results have a relatively mild dependence on the $Br(h\to SS)$. The reason for that is as follows:
the experimental limits on $Br(h\to SS)$ are already strong enough to limit the annihilation rate of $SS$ pairs to the SM states to be much less
than one picobarn, and consequently the metastable
abundance of $S$ particles per nucleon is quite high, $Y_p \gg 1$. This leads to a massive injection
of nucleons and mesons at early times, which raises the $n/p$ ratio, and creates larger yields of $^4$He compared to SBBN. Contributions of
very light $S$ particles to the Hubble rate during the $n/p$ freeze out also raises $Y_p$. The limits on $\tau_S$
are robust, and have rather mild dependence on the uncertainties in our treatment. This is because the initial large metastable
$Y_S$ abundance needs to be depleted prior to the $n/p$ freeze out time $t_{n/p}$, leading to the requirement $\tau_S \ll t_{n/p}$.
Consequently $O(1)$ variations in the yields of mesons and nucleons in the final states can be compensated by small variations in $\tau_S$, parametrically on the order $\log^{-1}(t_{n/p}/\tau_S)$, to produce the same influence on BBN. For the same reasons, our limits are also very insensitive to the exact observational constraint on $\Delta Y_p$, and we take a rather conservative limit of 0.01 (allowing $\pm0.01$ deviations frrom the observed/calculated mean). From the point of the LHC physics, the most promising is a scenario with a mass
$m_S$ not far below $m_h/2$. In that case, the effective decay length has to be on the order or smaller than $\sim 10^8$ meters, Fig.~\ref{fig:S_paramspace}, providing a
$10^{-6}$ {\em minimum} probability for a decay within a 100 m length purposely built detector.
Given that the high-luminosity LHC would produce copious numbers of the Higgs bosons, there is a chance to cover the entire
lifetime range for masses within $10-$to$-50$ GeV range.
It is easy to see that the above considerations can be generalized to other models of the Higgs-portal-coupled particles. For example,
consider a fermion $\chi$, coupled to the Higgs via $H^\dagger H (\bar\chi \chi)$ or $H^\dagger H(\bar \chi i \gamma_5 \chi)$ dimension-five operators, and having a small decay term such as {\em e.g.} neutrino portal $LH \chi$. The main
analysis of our work can be recast for that model, especially in the part that connects Higgs
decays with a metastable abundance of $\chi$. Evidently, for $Br(h\to \chi\bar\chi ) \sim Br(h\to SS) $ input, one will end up with $Y_\chi \sim Y_S$. The only change will be in the yields of mesons and baryons in the decays of $\chi$
compared to $S$. However, it is well known that already for $m_\chi $ above 250 MeV, the yields of pions and kaons is substantial
\cite{Alekhin:2015byh}, giving confidence that for the most parts same constraints we have derived for $\tau_S$ will translate to similar limits
on $\tau_\chi$.
The analysis performed in this paper can be easily generalized to other models of metastable particles, with different types of interactions, via $Z$, $Z'$ etc. In the limit when $Z'$ is outside of the LHC reach, one could have a set of effective operators connecting $\chi$ with
the SM fields, such as $\frac{1}{\Lambda^2}\bar \chi \gamma_\mu \chi \bar q \gamma_\mu q$, where $\Lambda$ is some energy scale.
The $\chi$ pair-production cross section in this case will scale as $\sigma_{q\bar q \to \chi \bar\chi } \propto E_q^2 \Lambda^{-4}$, where
$E_q$ is a typical (anti-)quark energy, while cosmological annihilation cross section has $\sigma_{\chi \bar\chi \to q\bar q } v \propto
m_\chi^2 \Lambda^{-4}$ scaling. Therefore, the LHC-relevant cross section can be enhanced relative to the annihilation
rate by a parametrically large ratio, $ E_q^2/m_\chi^2 $ if $m_\chi$ is parametrically smaller than the TeV scale. Therefore,
one can easily have a range of parameters with a relatively large $\chi\bar\chi$ pair-production cross section, while having very small annihilation
rates, rendering $Y_\chi \gg 1$, and resulting again in strong BBN constraints on lifetimes, $\tau_\chi < 0.1$ seconds.
Therefore, we conclude that some simple $Z'$ mediated models of metastable particles can also be strongly restricted by cosmology, making
them a perfect candidate for the searches of metastable particles at the LHC.
It is also instructive to consider models where constraints on the lifetime of metastable particles are {\em much weaker}. Clearly, one needs
an effective new mechanism for the self-annihilation in the early Universe, as the Higgs channel is too inefficient. Staying within
the Higgs portal models, consider the following potential with two real scalars,
\begin{equation}
V(H,S_1,S_2) = H^\dagger H (\lambda_1 S_1^2 + \lambda_2 S_2^2+ A_1S_1 + A_2S_2) +\lambda_{12}S_1^2S_2^2 + V(S_1) + V(S_2) + V(H^\dagger H),
\end{equation}
with the following hierarchy of couplings:
\begin{equation}
\lambda_1 \gg \lambda_2;~ A_1 \ll A_2; ~ \lambda_{12} \sim O(1);~ m_{S_1} > m_{S_2}.
\end{equation}
These choices will lead to a long-lived $S_1$, somewhat shorter-lived $S_2$, a predominant decay of
the Higgs boson to pairs of $S_1$, and cosmological
depletion of $S_1$ via $S_1S_1 \to S_2 S_2$ annihilation with potentially a large cross section due to a sizeable $\lambda_{12}$
coupling. Most importantly, in this model the Higgs decay to pairs of $S_1$ does not result in a prediction of $Y_{S1}$ abundance,
which can be quite small even for small values of $Br(H\to S_1S_1)$.
If $Y_{S1} \ll 1$, there would not be enough decay mesons and nucleons to affect early $n/p$ freeze out, and constraints on $\tau_{S1}$ will be coming only from the considerations of late decays
with hadronic or electromagnetic energy injection. Instead of $\tau_{S}<0.1$ sec, one expects to have sensitivity to $\tau_{S1} \sim 10^3$ seconds, or even worse, beyond $10^4$ seconds, if decays of $S_1$ are mostly leptonic. This example is not unique, and
there are other models where constraints on lifetimes and decay lengths are relatively lax, provided that there are extra channels that ensure
efficient cosmological annihilation of metastable particles.
\section*{Acknowledgements}
We thank D. Curtin, M. McCullough, P. Meade, M. Papucci and J. Shelton for soliciting this study, as well as D. Curtin and B. Shuve for
very helpful discussions. Research
at Perimeter Institute is supported by the Government
of Canada through Industry Canada and by the
Province of Ontario through the Ministry of Economic
Development \& Innovation.
| 22bccd1d90f799677f9544751cf615125217a839 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\subsection*{Gate dependence of superconductivity}
Figure~\ref{Fig-SC} (a) presents the temperature-dependent sheet resistance $R_S$ ($T$) at various gate voltages $V_g$. A clear gate-dependent superconducting transition is observed. We define the critical temperature $T_c$~ as the temperature at which $R_S$ reaches half of its value at 350~mK. The normal state resistance $R_S$ (350 mK) decreases monotonically with increasing $V_g$~ [Fig.~\ref{Fig-SC} (b)], which is consistent with previous reports \cite{davis2017anisotropic,rout2017six}. The monotonic increase of $R_S$ is contrasted with the nonmonotonic dependence of $T_c$~ on $V_g$. A similar dome-shaped region in the carrier density-temperature phase diagram is seen in many unconventional superconductors and in the (100) LaAlO$_3$/SrTiO$_3$~ interface.
\par
In the (100) LaAlO$_3$/SrTiO$_3$~ interface, the Hall coefficient depends nonmonotonically on the gate voltage. Surprisingly, this nonmonotonic behavior is also seen in the gate dependence of the Shubnikovβde Haas oscillations (SdH) frequency. Both the SdH frequency and low field inverse Hall coefficient follow the gate dependence of $T_c$~ for the (100) interface \cite{maniv2015strong,smink2017gate}, or the superconductivity starts appearing when the low field inverse Hall coefficient decreases from its maximum value \cite{singh2017competition}. By contrast, for the (111) interface the inverse Hall coefficient monotonically decreases with $V_g$~ [Fig.~\ref{Fig-SC} (c)] consistent with previous observations \cite{davis2017anisotropic, rout2017six}. In the case of the (111) LaAlO$_3$/SrTiO$_3$~ interface, the titanium t$_{2g}$ bands are split into low and high spin states due to the atomic spin-orbit interaction \cite{xiao2011interface, doennig2013massive}. We have shown that the lower spin state is first populated when accumulating electrons with increasing $V_g$~\cite{rout2017six}. This two-band scenario complicates the interpretation of the Hall data. We have estimated the amount of carrier density modulation due to the electric field effect similar to Refs. \cite{caviglia2008electric,biscaras2012two}. Since the $V_g$~range used is relatively small, the nonlinearities in the dielectric constant ($\epsilon$) can be neglected and thus the corresponding modulation of electron density is $\simeq$ 1.3 $\times$ 10$^{13}$ cm$^{-2}$ with $\epsilon \simeq$ 15000. This value is much smaller than the net change in 1/$\left| {eR_H} \right|$ of $\simeq$ 4.3 $\times$ 10$^{13}$ cm$^{-2}$. Moreover, the electron density due to the field effect increases with $V_g$~ in contrast to the observed behavior in Fig.~\ref{Fig-SC} (c). All these observations indicate the presence of a hole band in addition to electron band(s) in the (111) interface. We have confirmed this scenario by analyzing the normal state transport data via a simplistic noninteracting two-band model with one hole and one electron band (see Ref. \cite{Supple} for more details). Therefore, it is possible that the hole contribution to the electronic transport (and perhaps to superconductivity) becomes important in this $V_g$~range \cite{davis2017anisotropic}. This is also consistent with the polar structure of the (111) interface \cite{herranz2012high}.
\begin{figure*}
\begin{center}
\includegraphics[width=0.8\hsize]{Figure2.pdf}
\caption{ Magnetoresistance $R_S$~($H$) at $T =$ 90 mK in (a) perpendicular ($\vec{H}$ perpendicular to the current and interface) and (b) longitudinal ($\vec{H}$ parallel to the current and interface) configurations for various $V_g$. (c) $H_{c\perp}$~ and $H_{c\parallel}$~ at 90 mK as a function of $V_g$~ along with the Chandrasekhar-Clogston limit $H_P$. (d) Gate dependence of $\xi_{GL}$ (90 mK), $\xi_{GL}$(0), and $\overline{d}$.}
\label{Fig-Hc}
\end{center}
\end{figure*}
\par
The sheet resistance versus magnetic field at 90~mK for various gate voltages is plotted in Figs.~\ref{Fig-Hc} (a) and~\ref{Fig-Hc} (b) for perpendicular and parallel field configurations, where the sample is properly aligned with the field within an accuracy of 2$^{\circ}$. We define the critical field ($H_{c\perp}$) for the perpendicular magnetic field configuration such that $R_S$($H_{c\perp}$)$=R_S$ (350~mK)/2 and a similar criterion is followed for $H_{c\parallel}$ \cite{Remark}. In Fig.~\ref{Fig-Hc} (c) we plot $H_{c\parallel}$~ and $H_{c\perp}$~ as a function of $V_g$~ both exhibiting nonmonotonic behavior with the maximum at the same gate voltage as $T_c$. $H_{c\parallel}$$>$$H_{c\perp}$~ for all gate voltages reaching a maximal ratio of $\sim$16. Such strong anisotropy between two field orientations is evidence for 2D superconductivity in the (111) interface. Thus, it is expected that the superconducting layer thickness ($d$) should be smaller than the Ginzburg-Landau coherence length ($\xi_{GL}$). To check this, we extract $\xi_{GL}$ from $H_{c\perp}$~ using the relation: ${\xi _{GL}} = \sqrt {{\Phi _0}/2\pi {H_{c \bot }}} $ . It is presented in Fig.~\ref{Fig-Hc} (d) together with its extrapolation to zero temperature using $H_{c\perp}$($T$)=$H_{c\perp}$(0)$(1-T/T_c)$ valid for a 2D superconductor. Since the parallel magnetic field fully penetrates a 2D $(d \ll \xi)$ superconductor we can only estimate the upper limit for $d$ denoted as $\overline{d}$, which can be found from $\overline{d} = \sqrt 3 {\Phi _0}/\pi {\xi _{GL}}{H_{c\parallel }}$ [see Fig.~\ref{Fig-Hc} (d)]. We note that, for all $V_g$, $\overline{d}<\xi_{GL}(0)$, rendering superconductivity in the (111) SrTiO$_3$/LaAlO$_3$~ two dimensional.
\par
For a parallel field configuration in a 2D superconductor, the orbital motion and vortices can be neglected making the Zeeman energy the dominant pair-breaking effect. This leads to an upper (Chandrasekhar-Clogston) limit of $H_{c\parallel}$~ given by $H_P = 3.5k_BT_c/\sqrt{2}g\mu_B$ ($\mu_B$ is the Bohr magneton) in the BCS weak coupling limit \cite{chandrasekhar1962note,clogston1962upper}. Assuming a gyromagnetic ratio of $g\simeq$2, we observe $H_{c\parallel}$$>H_P$ for all gate voltages reaching a maximal ratio of $\sim$11 [Fig.~\ref{Fig-Hc} (c)]. In the presence of strong spin-orbit coupling the Chandrasekhar-Clogston limit can be relaxed. Other reasons for breaking this limit could be strong coupling superconductivity, many-body effects, and an anisotropic pairing mechanism.
\par
To determine the spin-orbit interaction from $H_{c\parallel}$~, we use a somewhat oversimplified picture of spin-orbit scattering that suppresses spin orientation by the Zeeman field \cite{klemm1975theory}. For a strong spin-orbit interaction, $H_{c\parallel}$~ can be expressed in terms of the spin-orbit energy ($\varepsilon_{SO}$) as ${H_{c\parallel}} = 0.602\sqrt { \varepsilon_{SO} /{k_B}{T_c}} {H_P}$ with $\varepsilon_{SO}= \hbar/\tau _{SO}$, and $\tau _{SO}$ is the spin-orbit scattering time. Remarkably, this analysis reveals a nonmonotonic dependence of $\varepsilon_{SO}$ on $V_g$~ as shown in Fig.~\ref{Fig-WAL} (b). This is the main finding of our Letter. For (110) LaAlO$_3$/SrTiO$_3$~, gate-independent spin-orbit coupling has been observed \cite{herranz2015engineering}; perhaps because of the nonpolar structure of this interface. The findings on the (110) interface are contrasted with our results of a strong and gate-tunable spin-orbit interaction for the (111) interface that follows the behavior of the superconducting dome. A weaker correlation between spin-orbit coupling and $T_c$~ in the (100) interface can be deduced by combining Refs. \cite{shalom2010tuning, caviglia2010tunable,liang2015nonmonotonically}, where $H_{c\parallel}$ is smaller.
\begin{figure*}
\begin{center}
\includegraphics[width=1\hsize]{Figure3.pdf}
\caption{(a) The normalized perpendicular magnetoconductance $\Delta \sigma (H)/\sigma _0$ for different $V_g$~ at $T =$ 1.3 K. The black solid lines are the fits according to Eq. \ref{WAL}. (b) $\varepsilon_{SO}$ as a function of $V_g$~ determined from $H_{c\parallel}$~ (see the text for more details). (c) Gate dependence of ${H_i }$ and ${H_{SO} }$ extracted from the fitting of weak antilocalization. (d) Gate dependence of $\tau_{i}$, $\tau _{SO}$, and $\tau$. The inset shows $\tau _{SO}$ as a function of $\tau^{-1}$ along with the solid line as a guide to the eye.}
\label{Fig-WAL}
\end{center}
\end{figure*}
\par
To further confirm the presence of a spin-orbit interaction, we studied the perpendicular magnetoresistance well above $T_c$~ at 1.3 K [Fig.~\ref{Fig-WAL} (a)]. For a 2D diffusive metallic system placed in a perpendicular magnetic field ($H$), the field-dependent quantum correction to conductivity $\Delta \sigma (H)$ normalized by quantum conductance ($\sigma _0=2e^2/h$) can be expressed as \cite{maekawa1981magnetoresistance,caviglia2010tunable}
\begin{equation}\label{WAL}
\begin{split}
\frac{{\Delta \sigma (H)}}{{{\sigma _0}}} = \Psi \left( {\frac{H}{{{H_i} + {H_{SO}}}}} \right)\\
+ \frac{1}{{2\sqrt {1 - {\gamma ^2}} }} \Psi \left( {\frac{H}{{{H_i} + {H_{SO}}(1 + \sqrt {1 - {\gamma ^2}} )}}} \right)\\
- \frac{1}{{2\sqrt {1 - {\gamma ^2}} }} \Psi \left( {\frac{H}{{{H_i} + {H_{SO}}(1 - \sqrt {1 - {\gamma ^2}} )}}} \right)\\
- \frac{{A{H^2}}}{{1 + C{H^2}}}.
\end{split}
\end{equation}
where $\Psi (x) = \ln (x) + \psi \left( {\frac{1}{2} + \frac{1}{x}} \right)$ [$\psi (x)$ is the digamma function] and $\gamma = g{\mu _B}H/4eD{H_{SO}}$ ($D$ is the diffusion coefficient). $H_{i}$ and $H_{SO} $ are the inelastic and spin-orbit fields, respectively. The classical orbital magnetoresistance contributes a Kohler term to Eq.~(\ref{WAL}) with the parameters $A$ and $C$. Figure~\ref{Fig-WAL} (c) shows $H_{i}$ and $H_{SO}$ for different $V_g$~ (see Supplemental Material for the gate dependence of $g$, $A$, and $C$ \cite{Supple}). Clearly, $H_{SO} > H_{i}$ for all $V_g$~, suggesting that we are in the weak antilocalization regime [see Fig.~\ref{Fig-WAL} (a)]. $H_{SO}$ from weak antilocalization [Fig.~\ref{Fig-WAL} (c)] shows nonmonotonic behavior similar to $\varepsilon_{SO}$ inferred from superconductivity [Fig.~\ref{Fig-WAL} (b)], and, furthermore, they have maximum value at the same gate voltage as $T_c$.
\par
In general, the LaAlO$_3$/SrTiO$_3$~ interface has a complicated band structure involving multiple contributions from the titanium $d$ bands \cite{shalom2010shubnikov, lerer2011low}. Therefore, the extracted parameters from weak antilocalization do not correspond to an individual band; instead an averaged value over all the bands should be considered \cite{rainer1985multiband}. We have extracted various averaged time scales, i. e. $\tau _{SO}$, $\tau _{i}$ (inelastic time), and $\tau$ (elastic scattering time) [Fig.~\ref{Fig-WAL} (d)]. The $\tau _{SO (i)}$ are related to $H_{SO (i)}$ determined from weak antilocalization as ${H_{SO (i)}} = \hbar /4eD{\tau _{SO (i)}}$. The effective diffusion coefficient ($D$) and $\tau$ are calculated using a na\"{\i}ve Drude model for a 2D electron gas (see Ref. \cite{Supple}). Using this analysis we find that $\tau _{SO}$ depends linearly on $\tau^{-1}$ for $V_g$~ $<$ -25 V [see the inset in Fig.~\ref{Fig-WAL} (d)] while for $V_g$~ $>$ -25 V both $\tau_{SO}$ and $\tau$ increase with $V_g$~[Fig.~\ref{Fig-WAL} (d)].
\par
The low $V_g$~regime ($V_g$~$<$ -25 V) is governed by a D'yakonov-Perel'-type spin-orbit relaxation mechanism for which $\tau_{SO}\propto \tau^{-1}$. In this scenario the electron precesses around the spin-orbit field, which is changing due to momentum scattering at a typical time $\tau$ \cite{vzutic2004spintronics}. The high $V_g$~ regime, on the other hand, is characterized by $\tau_{SO}\propto \tau$, suggesting that the electron spin is coupled to the crystal momentum. Interestingly these two regimes separated by the point where $\tau_{SO} \simeq \tau$ and the maximum of $T_c$~(and $H_{c\parallel}$) dome lies close to this $V_g$. All these observations suggest the mixing of multiple bands in the presence of a strong spin-orbit interaction for higher $V_g$. This scenario concurs with our recent report of crystalline sixfold anisotropic magnetoresistance in the (111) interfaces \cite{rout2017six}, where the sixfold term appears as a result of another band with higher spin state $J$ getting populated with increasing $V_g$. It is therefore possible that the crystalline spin-orbit interaction becomes important close to this avoided band crossing region due to the orbital mixing \cite{zhong2013theory,nakamura2013multi}. This interaction becomes smaller as $V_g$~is further increased away from the band crossing regime, resulting in a dome in the spin-orbit energy versus $V_g$. Such a multiband effect can also lead to dome-shaped superconductivity with maximum $T_c$~lying at this regime [as observed in Fig.~\ref{Fig-SC} (b)] similar to the case for the (100) interface \cite{maniv2015strong}. A more exotic mechanism of superconductivity in the LaAlO$_3$/SrTiO$_3$~interface involves the formation of a Fulde-Ferrell-Larkin-Ovchinikov (FFLO) state due to large spin-orbit coupling \cite{michaeli2012superconducting}. This can somewhat explain the nonmonotonic gate dependence of $H_{c\parallel}$~and $T_c$~with the maxima lying at $\tau_{SO} = \tau$. However, the $H_{c\parallel}$~for a quasi-2D superconductor in a FFLO state is estimated to be at most 2.5 times the Chandrasekhar-Clogston limit \cite{shimahara1997fulde}, which is much lower than the observed values [see Fig.~\ref{Fig-Hc} (c)]. Therefore, a full theoretical understanding of the phenomenological link observed here between the superconducting dome and the spin-orbit energy is yet to be developed.
\par
Salje \emph{et al.}~have found that for SrTiO$_3$~ below $\sim$70 K the tetragonal symmetry is lowered and the Sr atoms are displaced along the [111] direction leading to the breaking of local inversion symmetry \cite{salje2013domains}. It is therefore possible that a (111) SrTiO$_3$-based polar interface has such broken inversion symmetry in addition to conventional inversion symmetry breaking observed at polar oxide interfaces, which can result in an unconventional superconductivity. It has been recently suggested that dichalcogenide monolayers with hexagonal structure can be a realization of exotic Ising superconductivity where the spins are locked in an out-of-plane configuration due to the breaking of centrosymmetry \cite{lu2015evidence,xi2016ising,saito2016superconductivity}. We also note that the possibility for a nodeless time-reversal-symmetry-breaking superconducting order parameter has been proposed for (111) SrTiO$_3$-based interfaces from symmetry considerations \cite{scheurer2017selection}.
\par
In summary, the superconducting transition temperature $T_c$~ of the (111) LaAlO$_3$/SrTiO$_3$~ interface has a nonmonotonic dependence on the gate voltage. Maximum $T_c$~ is found at the same gate voltage where maximal values of spin-orbit field H$_{SO}$ and spin-orbit energy $\varepsilon_{SO}$ are observed. H$_{SO}$ is extracted from weak antilocalization while $\varepsilon_{SO}$ is estimated from the superconducting properties. The $H_{c\parallel}$~ exceeds the Chandrasekhar-Clogston limit by more than an order of magnitude due to a strong spin-orbit interaction. We suggest that the crystalline spin-orbit interaction becomes important close to an avoided band crossing region. In this regime orbital mixing can lead to enhanced spin-orbit interaction and superconductivity, which become weaker as $V_g$~ is tuned away from this avoided band crossing regime. This results in a dome in the spin-orbit energy (and $T_c$~) versus $V_g$. However, a deeper insight to the link between spin-orbit interaction and the superconducting dome requires further development of theoretical models for this unique hexagonal oxide interface.
\par
P.K.R. and E.M. contributed equally to this work. We are indebted to Moshe Goldstein and Alexander Palevski for useful discussions. This work has been supported by the Israel Science Foundation under Grant No. 382/17, the Israel Ministry of Science technology and space under Contract No. 3-11875 and the Bi-national science foundation under Grant No. 2014047.
| 79afd95e591ea7a6829a132a57f9f8c693c9e1ac | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{#1}\setcounter{equation}{0}}
\newcommand{\ra}{\rangle}
\newcommand{\la}{\langle}
\newcommand{\p}{\partial}
\newcommand{\hp}{{\Phi}}
\newcommand{\hq}{{Q_B}}
\newcommand{\he}{{\eta_0}}
\newcommand{\ha}{{{A}}}
\newcommand{\rrr}{\big\rangle\big\rangle}
\newcommand{\lllb}{\Bigl\langle\Bigl\langle}
\newcommand{\rrrb}{\Bigr\rangle\Bigr\rangle}
\begin{document}
\begin{center}
\large {\bf Hawking radiation of massive bosons via tunneling from black strings}
\end{center}
\begin{center}
Zhong-Wen Feng
$\footnote{E-mail:[email protected]}$
Shu-Zheng Yang
\end{center}
\begin{center}
\textit{College of Physics and Space Science, China West Normal University, Nanchong 637009, China}
\end{center}
\noindent
{\bf Abstract:} In the present paper, the Hawking radiation of massive bosons from 4-dimensional and 5-dimensional black strings are studied in quantum tunneling formalism. Firstly, we derive the Hamilton-Jacobi equation set via the Proca equation and WKB approximation. Then, the tunneling rates and Hawking temperature of the black strings are obtained. Our calculations show that the tunneling rates and Hawking temperatures are related to the properties of black strings' spacetime. When compare our results with those of scalars and fermions cases, it finds that they are the same.
\noindent
{\bf Keywords:}Massive bosons; Hawking radiation; Black strings
\section{Introduction}
\label{Int}
One common feature among black hole physics is that the existence of black holes' thermal properties. This idea was first proposed by Bekenstein, who proved that the entropy of black holes have satisfies the relation $S = {A \mathord{\left/ {\vphantom {A 4}} \right. \kern-\nulldelimiterspace} 4}$, where $A$ is the horizon area \cite{ch1}. In 1976, inspired by this entropy theory of black holes, Hawking discovered that black holes have thermal radiation (now people called this theory as Hawking radiation), and the temperature of black holes (Hawking temperature) can be expressed as $T = {\kappa \mathord{\left/ {\vphantom {\kappa {2\pi }}} \right. \kern-\nulldelimiterspace} {2\pi }}$ with the surface gravity of black holes $\kappa$ \cite{ch2}. The Hawking radiation is very for the foundations of physics, which was great influence on gravitational theory, quantum mechanism, and thermodynamic. Therefore, the Hawking radiation has received wide attention.
Based on the original theory of Hawking, the black hole radiation was considered as a quantum tunneling effect near the horizon. So, people can discuss the Hawking radiation and the temperature of black holes via the quantum tunneling method. In Ref.~\cite{ch3}, Parikh and Wilczek put forward the first kind of quantum tunneling method, which is called as the Null Geodesic method. Using this method, they calculate the tunneling behaviors of massless scalar particles from Schwarzschild black hole as well as Reissner-Nordstr\"{o}m black hole and obtained those black holes' temperature. The Hamilton-Jacobi ansatz is another kind of quantum tunneling method. In Ref.~\cite{ch3+}, the authors studied the massive scalar particles tunneling from spherically symmetric spacetimes. Subsequently, Kerner and Mann extended Hamilton-Jacobi ansatz to fermions case and studied fermion tunneling. They first applied the WKB approximation to Dirac equation. Then, neglecting the higher-order terms of $\mathcal{O(\hbar)}$, the resulting equation to leading order in $\hbar$ becomes the Hamilton-Jacobi equation. According the Hamilton-Jacobi equation, Kerner and Mann investigated the Hawking radiation of fermions via quantum tunneling from in the Rindler spacetime and a general non-rotating black hole \cite{ch4,ch5}. Soon after that, Banerjee and Majhi developed the Hamilton-Jacobi ansatz to solve the information loss paradox \cite{ch5+,ch6+}. Nowadays, the radiation behavior in many complicated spacetimes has been investigated via quantum tunneling method \cite{ch6,ch7,ch8,ch9,ch10,ch11,ch12,ch13}.
It is a well-known issue that the bosons or vector particles (e.g. $Z$, $W^ \pm$) play a very fundamental role as in the standard model for electroweak interaction and the high energy physics. Therefore, the quantum tunneling of bosons from the black hole has recently attracted people's attention \cite{ch13+,ch14+,ch14,ch15,ch18+,ch19+,ch16,ch17,ch20+,ch21+,ch22+,ch23+}. In Ref.~\cite{ch14,ch15}, Kruglov first used the Proca equation to calculate the Hawking radiation of vector from low dimensional spacetimes. Then, we extended Kruglov's work and studied the tunneling behavior of massive bosons from Kerr-de Sitter black hole and 5-dimensional Schwarzschild-Tangherlini black hole \cite{ch18+}. On the other hand, the black strings are important solutions of the Einstein-Maxwell equations. Those cylindrically symmetric solutions are helpful for us to study the AdS/CFT correspondence. Thus, in the paper we will use Proca equation to study the vector particles tunneling from 4-dimensional rotating black string and 5-dimensional black string.
The remainders of this paper are outlined as follows. Using the Proca equation and WKB approximation, the tunneling process of massive bosons from 4-dimensional rotating black string are investigated in Section~\ref{II}. In Section~\ref{III}, we extend our calculations and study massive bosons Hawing radiation from 5-dimensional static black string. In Section~\ref{Dis}, some conclusions and discussions are presented.
\section{Massive bosons tunneling from the rotating black string}
\label{II}
In Ref.~\cite{ch16}, the authors analyzed the Einstein-Maxwell equations with a cosmological constant, which in a stationary spacetime admitting an isometry group of $R\times \rm{U(1)}$. In their work, the rotating black string metric is given by
\begin{align}
\label{eq1}
ds^2 & = - \left[ {\alpha ^2 r^2 - \frac{{4M}}{{\alpha r}}\left( {1 - \frac{{a^2 \alpha ^2 }}{2}} \right)} \right]dt^2 + \left[ {\alpha ^2 r^2 - \frac{{4M}}{{\alpha r}}\left( {1 - \frac{3}{2}a^2 \alpha ^2 } \right)} \right]^{ - 1} dr^2
\nonumber \\
& + \left( {r^2 - \frac{{4Mra^2 }}{{\alpha r}}} \right)d\phi ^2 + \alpha ^2 r^2 dz^2 - \frac{{8Ma\sqrt {1 - \frac{{a^2 \alpha ^2 }}{2}} }}{{\alpha r}}dtd\phi ,
\end{align}
where $\alpha^2$ is the negative cosmological constant, which satisfy the relation $\alpha = - {\Lambda \mathord{\left/ {\vphantom {\Lambda 3}} \right.
\kern-\nulldelimiterspace} 3}$. Meanwhile, we denote that $a^2 \alpha ^2 = 1 - {\varepsilon \mathord{\left/ {\vphantom {\varepsilon M}} \right. \kern-\nulldelimiterspace} M}$, $\varepsilon = \sqrt {M^2 - {{8J^2 \alpha ^2 } \mathord{\left/ {\vphantom {{8J^2 \alpha ^2 } 9}} \right. \kern-\nulldelimiterspace} 9}}$, $J = \left( {{{3Ma} \mathord{\left/ {\vphantom {{3Ma} 2}} \right.
\kern-\nulldelimiterspace} 2}} \right)\sqrt {1 - {{a^2 \alpha ^2 } \mathord{\left/ {\vphantom {{a^2 \alpha ^2 } 2}} \right. \kern-\nulldelimiterspace} 2}} $, where $M$ and $J$ are the mass and angular momentum densities of the rotating black string, respectively. For simplicity, the metric (\ref{eq1}) becomes to
\begin{eqnarray}
\label{eq2}
ds^2 = - \Delta \left( {\gamma dt - \frac{\delta }{{\alpha ^2 }}d\phi } \right)^2 + r^2 \left( {\gamma d\phi - \delta dt} \right)^2 + \frac{1}{\Delta }dr^2 + \alpha ^2 r^2 dz^2 ,
\end{eqnarray}
with $\Delta = a^2 r^2 - {b \mathord{\left/ {\vphantom {b {\alpha r}}} \right. \kern-\nulldelimiterspace} {\alpha r}}$, $b = 4M\left( {1 - {{3a^2 \alpha ^2 } \mathord{\left/ {\vphantom {{3a^2 \alpha ^2 } 2}} \right. \kern-\nulldelimiterspace} 2}} \right)$,$\gamma = \sqrt {{{\left( {2 - a^2 \alpha ^2 } \right)} \mathord{\left/ {\vphantom {{\left( {2 - a^2 \alpha ^2 } \right)} {\left( {2 - 3a^2 \alpha ^2 } \right)}}} \right. \kern-\nulldelimiterspace} {\left( {2 - 3a^2 \alpha ^2 } \right)}}}$ ,${{a\alpha ^2 } \mathord{\left/ {\vphantom {{a\alpha ^2 } {\sqrt {1 - {{3a^2 \alpha ^2 } \mathord{\left/ {\vphantom {{3a^2 \alpha ^2 } 2}} \right. \kern-\nulldelimiterspace} 2}} }}} \right. \kern-\nulldelimiterspace} {\sqrt {1 - {{3a^2 \alpha ^2 } \mathord{\left/ {\vphantom {{3a^2 \alpha ^2 } 2}} \right. \kern-\nulldelimiterspace} 2}} }}$, respectively. The horizons are given by the null-hyper surface condition $g^{\mu \nu } \left( {\partial _\mu F} \right)\left( {\partial _\nu F} \right) = 0$, where $F$ is the hyper-surface of black string. So, the black string has three horizons when $\Delta$ vanishes, the outer event horizon is located at $r_ + = a^{ - 1} b^{{1 \mathord{\left/ {\vphantom {1 3}} \right. \kern-\nulldelimiterspace} 3}}$.
Now, applying the dragging coordinate transformation
\begin{eqnarray}
\label{eq4}
d\varphi = d\phi + \Omega dt= d\phi + \frac{{r^2 \gamma \alpha ^4 - \Delta \gamma \delta \alpha ^2 }}{{r^2 \gamma ^2 \alpha ^4 - \Delta \delta ^2 }} dt,
\end{eqnarray}
one can rewrite the metric (\ref{eq2}) as
\begin{eqnarray}
\label{eq5}
ds^2 = - A(r) dt^2 + \frac{1}{B(r)}dr^2 + C(r)d\varphi ^2 + D(r)dz^2 ,
\end{eqnarray}
where $A(r) = {{\Delta r^2 \left( {\alpha ^2 \gamma ^2 - \delta ^2 } \right)^2 } \mathord{\left/
{\vphantom {{\Delta r^2 \left( {\alpha ^2 \gamma ^2 - \delta ^2 } \right)^2 } {\left( {\Delta \delta ^2 - \alpha ^4 r^2 \gamma ^2 } \right)}}} \right.
\kern-\nulldelimiterspace} {\left( {\Delta \delta ^2 - \alpha ^4 r^2 \gamma ^2 } \right)}}$, $B (r) = \Delta ^{ - 1}$, $C(r) = r^2 \gamma ^2 - {{\Delta \delta ^4 } \mathord{\left/ {\vphantom {{\Delta \delta ^4 } {\alpha ^4 }}} \right. \kern-\nulldelimiterspace} {\alpha ^4 }}$, $D (r) = \alpha ^2 r^2$. On the event horizon, the $B(r)$ can be expended as $B(r)=B'(r_+)(r-r_+)+\mathcal{O}[(r-r_+)^2]$ \cite{ch17}. According to Ref.~\cite{ch14}, in order to study the tunneling behavior of massive bosons on the event horizon, one need to use the Proca equation
\begin{eqnarray}
\label{eq6}
D_\mu \psi ^{ \nu \mu } + \frac{{m^2 }}{{\hbar ^2 }}\psi ^\nu = 0,
\end{eqnarray}
\begin{eqnarray}
\label{eq7}
\psi _{\nu \mu} = D_\nu \psi _\mu - D_\mu \psi _\nu = \partial _\nu \psi _\mu - \partial _\mu \psi _\nu ,
\end{eqnarray}
where $D_\mu$ are covariant derivatives, $\psi _\nu$ are depended on $\psi _t ,\psi _r ,\psi _\theta$ and $\psi _\varphi$, and the mass of bosons is $m$. The anti-symmetric tensor $\psi ^{\mu \nu }$ obeys the relation $\psi _{\mu \nu } = \partial _\mu \psi _\nu - \partial _\nu \psi _\mu$. Thus, with the help of equation $D_\mu \psi _{\mu \nu } = {{\partial _\mu \left( {\sqrt { - g} \psi ^{\mu \nu } } \right)} \mathord{\left/ {\vphantom {{\partial _\mu \left( {\sqrt { - g} \psi ^{\mu \nu } } \right)} {\sqrt { - g} }}} \right. \kern-\nulldelimiterspace} {\sqrt { - g} }}$, Eq.~(\ref{eq6}) becomes
\begin{align}
\label{eq8}
\frac{1}{{\sqrt { - g} }}\partial _\mu \left( {\sqrt { - g} \psi ^{\nu \mu } } \right) + \frac{{m^2 }}{{\hbar ^2 }}\psi ^\nu = 0.
\end{align}
where $\psi ^\nu$ and $\psi ^{\mu \nu }$ are given as follows
\begin{align}
\label{eq9}
& \psi ^0 = -{A ^{-1}}\psi _0 ,\psi ^1 = B\psi _1 ,\psi ^2 = {C ^{-1}}\psi _2 ,\psi ^3 = {D ^{-1}}\psi _3 ,\psi ^{01} = - {B}{A ^{-1}}\psi _{01},
\nonumber \\
& \psi ^{02} = - {{(AC)} ^{-1}}\psi _{02} ,\psi ^{03} = - {{(AD)} ^{-1}}\psi _{03} ,\psi ^{12} = {B}{C ^{-1}}\psi _{12} ,\psi ^{13} = {B}{D ^{-1}}\psi _{13},
\nonumber \\
& \psi ^{23} = {{(CD)} ^{-1}}\psi _{23} .
\end{align}
Putting Eq.~(\ref{eq9}) into Eq.~(\ref{eq6}), and then, considering the relationship in Eq.~(\ref{eq7}), one obtains
\begin{align}
\label{eq10}
& \frac{1}{{\sqrt { - g} }}\left\{ {\partial _r \left[ {\sqrt { - g} \frac{B}{A}\left( {\partial _t \psi _1 - \partial _r \psi _{_0 } } \right)} \right] + \partial _\varphi \left[ {\sqrt { - g} \frac{1}{{AC}}\left( {\partial _t \varphi _2 - \partial _\varphi \psi _{_0 } } \right)} \right]} \right.
\nonumber \\
& \left. { + \partial _z \left[ {\sqrt { - g} \frac{1}{{AD}}\left( {\partial _t \psi _3 - \partial _z \psi _{_0 } } \right)} \right]} \right\} + \frac{{m^2 }}{{\hbar ^2 }}\frac{1}{A}\psi _0 = 0,
\end{align}
\begin{align}
\label{eq11}
& \frac{1}{{\sqrt { - g} }}\left\{ {\partial _t \left[ {\sqrt { - g} \frac{B}{A}\left( {\partial _t \psi _{_1 } - \partial _r \psi _0 } \right)} \right] + \partial _\varphi \left[ {\sqrt { - g} \frac{B}{C}\left( {\partial _r \psi _{_2 } - \partial _\varphi \psi _1 } \right)} \right]} \right.
\nonumber \\
& \left. { + \partial _z \left[ {\sqrt { - g} \frac{B}{D}\left( {\partial _r \psi _{_3 } - \partial _z \psi _1 } \right)} \right]} \right\} + \frac{{m^2 }}{{\hbar ^2 }}B\psi _1 = 0,
\end{align}
\begin{align}
\label{eq12}
& \frac{1}{{\sqrt { - g} }}\left\{ {\partial _t \left[ {\sqrt { - g} \frac{1}{{AC}}\left( {\partial _t \psi _{_2 } - \partial _\varphi \psi _0 } \right)} \right] + \partial _r \left[ {\sqrt { - g} \frac{B}{C}\left( {\partial _\varphi \psi _{_1 } - \partial _r \psi _2 } \right)} \right]} \right.
\nonumber \\
& \left. { + \partial _z \left[ {\sqrt { - g} \frac{1}{{CD}}\left( {\partial _\varphi \psi _3 - \partial _z \psi _2 } \right)} \right]} \right\} + \frac{{m^2 }}{{\hbar ^2 }}\frac{1}{C}\psi _2 = 0,
\end{align}
\begin{align}
\label{eq13}
& \frac{1}{{\sqrt { - g} }}\left\{ {\partial _t \left[ {\sqrt { - g} \frac{1}{{AD}}\left( {\partial _t \psi _3 - \partial _z \psi _0 } \right)} \right] + \partial _r \left[ {\sqrt { - g} \frac{B}{D}\left( {\partial _z \psi _1 - \partial _r \psi _3 } \right)} \right]} \right.
\nonumber \\
& \left. { + \partial _\varphi \left[ {\sqrt { - g} \frac{1}{{CD}}\left( {\partial _z \psi _2 - \partial _\varphi \psi _3 } \right)} \right]} \right\} + \frac{{m^2 }}{{\hbar ^2 }}\frac{1}{D}\psi _3 = 0.
\end{align}
In order to solve Eq.~(\ref{eq10})-Eq.~(\ref{eq13}), one can express the solutions to above equation set in the following form
\begin{eqnarray}
\label{eq14}
\psi _\nu = \left( {c_0 ,c_1 ,c_2 ,c_3 } \right)\exp \left\{ {\frac{i}{\hbar }\left[ {S_0 \left( {t,r,\varphi ,z} \right) + \sum\limits_i {\hbar ^i } S_i \left( {t,r,\varphi ,z} \right)} \right]} \right\},
\end{eqnarray}
where $i=1,2,3,\cdots$. According to the WKB approximation and Eq.~(\ref{eq14}), Eq.~(\ref{eq10})-Eq.~(\ref{eq13}) can be reexpressed as
\begin{align}
\label{eq15}
& \left\{ {B\left[ { c_0 \left( {\partial _r S_0 } \right)^2 - c_1 \left( {\partial _t S_0 } \right)\left( {\partial _r S_0 } \right) } \right] + C^{ - 1} \left[ c_0 \left( {\partial _\varphi S_0 } \right)^2 - {c_2 \left( {\partial _t S_0 } \right)\left( {\partial _\varphi S_0 } \right) } \right]} \right.
\nonumber \\
& \left. { + D^{ - 1} \left[ c_0 \left( {\partial _z S_0 } \right)^2 - {c_3 \left( {\partial _t S_0 } \right)\left( {\partial _z S_0 } \right) } \right]} \right\} + m^2 \psi _0 = 0,
\end{align}
\begin{align}
\label{eq16}
&\left\{ {A^{ - 1} \left[ {c_0 \left( {\partial _t S_0 } \right)\left( {\partial _r S_0 } \right) - c_1 \left( {\partial _t S_0 } \right)^2 } \right] + C^{ - 1} \left[ {c_1 \left( {\partial _\varphi S_0 } \right)^2 - c_2 \left( {\partial _t S_0 } \right)\left( {\partial _\varphi S_0 } \right)} \right]} \right.
\nonumber \\
&\left. { + D^{ - 1} \left[ {c_1 \left( {\partial _z S_0 } \right)^2 - c_3 \left( {\partial _r S_0 } \right)\left( {\partial _z S_0 } \right)} \right]} \right\} + m^2 \psi _1 = 0,
\end{align}
\begin{align}
\label{eq17}
& \left\{ {A^{ - 1} \left[ {c_0 \left( {\partial _t S_0 } \right)\left( {\partial _\varphi S_0 } \right) - c_2 \left( {\partial _t S_0 } \right)^2 } \right] + B\left[ {c_2 \left( {\partial _r S_0 } \right)^2 - c_1 \left( {\partial _r S_0 } \right)\left( {\partial _\varphi S_0 } \right)} \right]} \right.
\nonumber \\
& \left. { + D^{ - 1} \left[ {c_2 \left( {\partial _z S_0 } \right)^2 - c_3 \left( {\partial _z S_0 } \right)\left( {\partial _\varphi S_0 } \right)} \right]} \right\} + m^2 \psi _2 = 0,
\end{align}
\begin{align}
\label{eq18}
& \left\{ {A^{ - 1} \left[ {c_0 \left( {\partial _t S_0 } \right)\left( {\partial _z S_0 } \right) - c_3 \left( {\partial _t S_0 } \right)^2 } \right] + B\left[ {c_3 \left( {\partial _r S_0 } \right)^2 - c_1 \left( {\partial _r S_0 } \right)\left( {\partial _z S_0 } \right)} \right]} \right.
\nonumber \\
& \left. { + C^{ - 1} \left[ {c_3 \left( {\partial _\varphi S_0 } \right)^2 - c_2 \left( {\partial _\varphi S_0 } \right)\left( {\partial _z S_0 } \right)} \right]} \right\} + m^2 \psi _3 = 0.
\end{align}
By analysing the spacetime of the rotating black string, one can find 3 Killing vectors. Therefore, the action $S_0$ can be expressed as
\begin{eqnarray}
\label{eq19}
S_0 = - \left( {\omega - j\Omega } \right)t + W\left( r \right) + j\varphi + \mathcal{Z} \left( z \right),
\end{eqnarray}
with the angular momentum $\omega$ and the energy of massive bosons $j$, respectively. Now, substituting Eq.~(\ref{eq19}) into Eq.~(\ref{eq15})-Eq.~(\ref{eq18}), a $4\times4$ matrix is obtained as
\begin{eqnarray}
\label{eq20}
\Lambda \left( {c_0 ,c_1 ,c_2 ,c_3 } \right)^T = 0.
\end{eqnarray}
where $\Lambda $ is a $4\times4$ matrix, which components are given as follows
\begin{eqnarray}
\label{eq21}
\begin{array}{l}
\Lambda _{00} = - B\left( {W'} \right)^2 - {C^{-1}}\left( {\partial _z \mathcal{Z} } \right)^2 - {D^{-1}}j^2 - m^2 ,\Lambda _{01} = - BW'\left( {\omega - j\Omega } \right), \\
\\
\Lambda _{02} = - {C^{-1}}\left( {\omega - j\Omega } \right)\left( {\partial _z \mathcal{Z} } \right),\Lambda _{03} = -{D^{-1}}\left( {\omega - j\Omega } \right)j, \\
\\
\Lambda _{10} = - {A^{-1}}\left( {\omega - j\Omega } \right)W',\Lambda _{11} = m^2 - {A^{-1}}\left( {\omega - j\Omega } \right)^2 + {C^{-1}}\left( {\partial _z \mathcal{Z} } \right)^2 - {D^{-1}}j^2 , \\
\\
\Lambda _{12} = {C^{-1}}\left( {\omega - j\Omega } \right)\left( {\partial _z \mathcal{Z} } \right),\Lambda _{13} = {D^{-1}}jW',\Lambda _{20} = - {A^{-1}}\left( {\omega - j\Omega } \right)\left( {\partial _z \mathcal{Z}} \right), \\
\\
\Lambda _{21} = BW'\left( {\partial _z \mathcal{Z} } \right),\Lambda _{22} = {A^{-1}}\left( {\omega - j\Omega } \right)^2 + B\left( {W'} \right)^2 + {D^{-1}}j^2 + m^2 , \\
\\
\Lambda _{23} = - {j}{D ^{-1}}\left( {\partial _z Z} \right),\Lambda _{30} = - {A^{-1}}\left( {\omega - j\Omega } \right)j,\Lambda _{31} = - BW'j, \\
\\
\Lambda _{32} = - {C^{-1}}\left( {\partial _z \mathcal{Z} } \right)j,\Lambda _{33} = m^2 -{A^{-1}}\left( {\omega - j\Omega } \right)^2 + B\left( {W'} \right)^2 + {C^{-1}}\left( {\partial _z \mathcal{Z} } \right)^2 . \\
\end{array}
\end{eqnarray}
In above equation, we denote that $W' = \partial _r W$. For obtaining the nontrivial solution of Eq.~(\ref{eq20}), it requires the ${\rm{det}}\left( \Lambda \right) = 0$. Thus, one has
\begin{eqnarray}
\label{eq22}
W_ \pm = \pm \int {\sqrt {\frac{{CD\left( {\omega - j\Omega } \right)^2 - ACj^2 - ACDm^2 - AD\left( {\partial _z \mathcal{Z}} \right)^2 }}{{ABCD}}} dr},
\end{eqnarray}
and the above equation is computed as
\begin{eqnarray}
\label{eq23}
W_ \pm = \pm i\pi \left[ {\frac{{\left( {\omega - j\Omega _{r_ + } } \right)\gamma \alpha ^4 r_ + ^2 }}{{\left( {2\alpha ^4 r_ + ^3 + b} \right)\left( {\alpha ^2 \gamma ^2 - \delta ^2 } \right)}}} \right]+ \mathcal{O}(RealPart),
\end{eqnarray}
where $+/-$ represents the outgoing/incoming solutions on the outer event horizon. Here we only need the imaginary part since the real part of Eq.~(\ref{eq23}) does not contribute to the tunneling rate. The tunneling rate of massive bosons from 4-dimensional rotating black string is given by
\begin{eqnarray}
\label{eq24}
\Gamma = \frac{{\Gamma _{\left( {emission} \right)} }}{{\Gamma _{\left( {absorpation} \right)} }} = \frac{{\exp \left( { - 2{\mathop{\rm Im}\nolimits} W_ + - 2{\mathop{\rm Im}\nolimits} \Xi } \right)}}{{\exp \left( { - 2{\mathop{\rm Im}\nolimits} W_ - - 2{\mathop{\rm Im}\nolimits} \Xi } \right)}} = \exp \left[ { - \frac{{4\pi \left( {\omega - j\Omega _{r_ + } } \right)\gamma \alpha ^4 r_ + ^2 }}{{\left( {2\alpha ^4 r_ + ^3 + b} \right)\left( {\alpha ^2 \gamma ^2 - \delta ^2 } \right)}}} \right].
\end{eqnarray}
Comparing Eq.~(\ref{eq24}) with the expression of Boltzman factor, that is $\Gamma = \exp \left( {{E \mathord{\left/ {\vphantom {E T}} \right.
\kern-\nulldelimiterspace} T}} \right)$ with the temperature $T$ and the energy of particles $\omega$ \cite{ch21}, the Hawking temperature of 4-dimensional rotating black string is
\begin{eqnarray}
\label{eq25}
T_H = \frac{{\left( {2\alpha ^4 r_ + ^3 + b} \right)\left( {\alpha ^2 \gamma ^2 - \delta ^2 } \right)}}{{4\pi \gamma \alpha ^4 r_ + ^2 }}.
\end{eqnarray}
From Eq.~(\ref{eq24}) and Eq.~(\ref{eq25}), one can find that the tunneling rate and Hawking temperature of massive bosons from 4-dimensional rotating black string are related to the outer event horizon $r_+$, the mass $M$ and angular momentum $J$ of the black string . When comparing our results with those of scalars and fermions cases, it finds that they are the same \cite{ch20}.
\section{Massive bosons tunneling from the 5-dimensional black string}
\label{III}
In this section, the massive bosons tunneling behavior from 5-dimensional black string will be investigated by using the Proca equation. The metric of 5-dimensional black string is given by \cite{ch19}
\begin{eqnarray}
\label{eq26}
ds^2 = - A\left( r \right)dt^2 + \frac{1}{{B\left( r \right)}}dr^2 + C\left( r \right)d\theta ^2 + D\left( r \right)d\varphi ^2 + E dz^2 .
\end{eqnarray}
where $F\left( r \right) = 1 - {{r_H } \mathord{\left/ {\vphantom {{r_H } r}} \right. \kern-\nulldelimiterspace} r}$, $G\left( r \right) =( 1 - {{r_H } \mathord{\left/ {\vphantom {{r_H } r}} \right. \kern-\nulldelimiterspace} r})^{ - 1}$, $C\left( r \right) = r^2$, $D\left( r \right) = r^2 \sin ^2 \theta$, $E=1$. $r_H$ is the event horizon of the 5-dimensional black string, we can rewrite the $B(r)$ as $B(r)=B'(r_+)(r-r_+)+\mathcal{O}[(r-r_+)^2]$. According to the line element of the 5-dimensional black string, one yields the following expression
\begin{align}
\label{eq27}
& \psi ^0 = - {A ^{-1}}\psi _0 ,\psi ^1 = B\psi _1 ,\psi ^2 = {C ^{-1}}\psi _2 ,\psi ^3 = {D ^{-1}}\psi _3 ,\psi ^4 = {E ^{-1}}\psi _4 ,
\nonumber \\
& \psi ^{01} = - {B ^{-1}}{A}\psi _{01} ,\psi ^{02} = - {{(AC) ^{-1}}}\psi _{02} ,\psi ^{03} = - {{(AD)}^{-1}}\psi _{03} ,
\nonumber \\
&\psi ^{04} = - {{(AE)}^{-1}}\psi _{04}, \psi ^{12} = {B^{-1}}{C}\psi _{12} ,\psi ^{13} = {B ^{-1}}{D}\psi _{13} ,\psi ^{14} = {B ^{-1}}{E}\psi _{14} ,
\nonumber \\
&\psi ^{23} = {{(CD)}^{-1}}\psi _{23} , \psi ^{24} = {{(CE)}^{-1}}\psi _{24} , \psi ^{34} = {{(DE)}^{-1}}\psi _{34} .
\end{align}
Putting Eq. (\ref{eq27}) and Eq. (\ref{eq14}) into Eq. (\ref{eq6}), and keeping the first order term of $\hbar$, one has
\begin{align}
\label{eq28}
& B\left[ {c_1 \left( {\partial _t S_0} \right)\left( {\partial _r S_0} \right) - c_0 \left( {\partial _r S_0} \right)^2 } \right] + {C ^{-1}}\left[ {c_2 \left( {\partial _t S_0} \right)\left( {\partial _\theta S} \right) - c_0 \left( {\partial _\theta S_0} \right)^2 } \right]
\nonumber \\
& + {D ^{-1}}\left[ {c_3 \left( {\partial _t S_0} \right)\left( {\partial _\varphi S_0} \right) - c_0 \left( {\partial _\varphi S_0} \right)^2 } \right] + {E ^{-1}}\left[ {c_4 \left( {\partial _t S_0} \right)\left( {\partial _z S_0} \right) - c_0 \left( {\partial _z S_0} \right)^2 } \right]
\nonumber \\
& - m^2 \psi _0 = 0,
\end{align}
\begin{align}
\label{eq29}
& {A ^{-1}}\left[ {c_0 \left( {\partial _r S_0} \right)\left( {\partial _t S_0} \right) - c_1 \left( {\partial _t S_0} \right)^2 } \right] + {C ^{-1}}\left[ {c_1 \left( {\partial _\theta S_0} \right)^2 - c_2 \left( {\partial _r S_0} \right)\left( {\partial _\theta S_0} \right)} \right]
\nonumber \\
& + {D ^{-1}}\left[ {c_1 \left( {\partial _\varphi S_0} \right)^2 - c_3 \left( {\partial _r S_0} \right)\left( {\partial _\varphi S_0} \right)} \right] + {E ^{-1}}\left[ {c_1 \left( {\partial _z S_0} \right)^2 - c_4 \left( {\partial _r S_0} \right)\left( {\partial _z S_0} \right)} \right]
\nonumber \\
& + m^2 \psi _1 = 0,
\end{align}
\begin{align}
\label{eq30}
& {A ^{-1}}\left[ {c_0 \left( {\partial _\theta S_0} \right)\left( {\partial _t S_0} \right) - c_2 \left( {\partial _t S_0} \right)^2 } \right] + B\left[ {c_2 \left( {\partial _r S_0} \right)^2 - c_1 \left( {\partial _\theta S_0} \right)\left( {\partial _r S_0} \right)} \right]
\nonumber \\
& + {D ^{-1}}\left[ {c_2 \left( {\partial _\varphi S_0} \right)^2 - c_3 \left( {\partial _\varphi S_0} \right)\left( {\partial _\theta S_0} \right)} \right] + {E ^{-1}}\left[ {c_2 \left( {\partial _z S_0} \right)^2 - c_4 \left( {\partial _\theta S_0} \right)\left( {\partial _z S_0} \right)} \right]
\nonumber \\
& + m^2 \psi _2 = 0,
\end{align}
\begin{align}
\label{eq31}
& {A ^{-1}}\left[ {c_0 \left( {\partial _\varphi S_0} \right)\left( {\partial _t S_0} \right) - c_3 \left( {\partial _t S_0} \right)^2 } \right] + B\left[ {c_3 \left( {\partial _r S_0} \right)^2 - c_1 \left( {\partial _r S_0} \right)\left( {\partial _\varphi S_0} \right)} \right]
\nonumber \\
& + {C ^{-1}}\left[ {c_3 \left( {\partial _\theta S_0} \right)^2 - c_2 \left( {\partial _\theta S_0} \right)\left( {\partial _\varphi S_0} \right)} \right] +{E ^{-1}}\left[ {c_3 \left( {\partial _z S_0} \right)^2 - c_4 \left( {\partial _z S} \right)\left( {\partial _\varphi S} \right)} \right]
\nonumber \\
& + m^2 \psi _3 = 0,
\end{align}
\begin{align}
\label{eq32}
& {A ^{-1}}\left[ {c_0 \left( {\partial _z S_0} \right)\left( {\partial _t S_0} \right) - c_4 \left( {\partial _t S_0} \right)^2 } \right] + B\left[ {c_4 \left( {\partial _r S_0} \right)^2 - c_1 \left( {\partial _r S_0} \right)\left( {\partial _z S_0} \right)} \right]
\nonumber \\
& + {C ^{-1}}\left[ {c_4 \left( {\partial _\theta S_0} \right)^2 - c_2 \left( {\partial _\theta S_0} \right)\left( {\partial _z S_0} \right)} \right] + {D ^{-1}}\left[ {c_4 \left( {\partial _\varphi S_0} \right)^2 - c_3 \left( {\partial _\varphi S_0} \right)\left( {\partial _z S_0} \right)} \right]
\nonumber \\
& + m^2 \psi _4 = 0.
\end{align}
Considering properties of spacetime, the action for 5-dimensional black string can be expressed as
\begin{eqnarray}
\label{eq33+}
S_0 = - \omega t + W\left( r \right) + \Theta \left( \theta \right) + j\varphi + \mathcal{Z} \left( z \right),
\end{eqnarray}
where $\omega$ and $j$ are the energy and angular momentum of the massive bosons. Inserting Eq.~(\ref{eq33+}) into Eq.~(\ref{eq28})-Eq.~(\ref{eq32}), the $5\times5$ matrix equation can be expressed as
\begin{align}
\label{eq34+}
\Lambda \left( {c_0 ,c_1 ,c_2 ,c_3 ,c_4 } \right)^T = 0,
\end{align}
where d the elements of $\Lambda$ are
\begin{eqnarray}
\label{eq33}
\begin{array}{l}
\Lambda _{00} = - B\left( {W'} \right)^2 - {C ^{-1}}\left( {\partial _\theta \Theta } \right)^2 -{D ^{-1}}j^2 - m^2 + {E ^{-1}}\left( {\partial _z \mathcal{Z} } \right)^2 ,\Lambda _{01} = - B\left( {W'} \right)\omega , \\
\\
\Lambda _{02} = - {C ^{-1}}\omega \left( {\partial _\theta \Theta } \right),\Lambda _{03} = - {D ^{-1}}\omega j,\Lambda _{04} = - {D ^{-1}}\omega \left( {\partial _z \mathcal{Z} } \right),\Lambda _{10} = - {A ^{-1}}\omega W', \\
\\
\Lambda _{11} = m^2 - {A ^{-1}}\omega ^2 + {C ^{-1}}\left( {\partial _\theta \Theta } \right)^2 - {D ^{-1}}j^2 + {K ^{-1}}\left( {\partial _z \mathcal{Z} } \right)^2 ,\Lambda _{12} = {C ^{-1}}\omega \left( {\partial _\theta \Theta } \right), \\
\\
\Lambda _{13} = {D ^{-1}}jW', \Lambda _{14} = - {D ^{-1}}\left( {\partial _z \mathcal{Z} } \right)W',\Lambda _{20} = - {A ^{-1}}\omega \left( {\partial _\theta \Theta } \right),\Lambda _{21} = BW'\left( {\partial _\theta \Theta } \right), \\
\\
\Lambda _{22} = - {A ^{-1}}\omega ^2 + B\left( {W'} \right)^2 + {D ^{-1}}j^2 + m^2 + {E ^{-1}}\left( {\partial _z \mathcal{Z} } \right)^2 ,\Lambda _{23} = - {D ^{-1}}j\left( {\partial _\theta \Theta } \right), \\
\\
\Lambda _{24} = {E ^{-1}}j \left( {\partial _\theta \Theta } \right),\Lambda _{30} = - {A ^{-1}}\omega j,\Lambda _{31} = - B W'j,\Lambda _{32} = - {C ^{-1}}\left( {\partial _\theta \Theta } \right)j, \\
\\
\Lambda _{33} = m^2 - {A ^{-1}}\omega ^2 + B\left( {W'} \right)^2 + {C ^{-1}}\left( {\partial _\theta \Theta } \right)^2 + {E ^{-1}}\left( {\partial _z \mathcal{Z} } \right)^2 ,\Lambda _{34} = - {D ^{-1}}\left( {\partial _z \mathcal{Z} } \right)j, \\
\\
\Lambda _{40} = - {A ^{-1}}\omega \left( {\partial _z \mathcal{Z} } \right),\Lambda _{41} = - BW'\left( {\partial _z \mathcal{Z} } \right),\Lambda _{42} = - {C ^{-1}}\left( {\partial _\theta \Theta } \right)\left( {\partial _z \mathcal{Z} } \right), \\
\\
\Lambda _{43} = - {D ^{-1}}j\left( {\partial _z \mathcal{Z} } \right),\Lambda _{44} = m^2 - {A ^{-1}}\omega ^2 + B\left( {W'} \right)^2 + {C ^{-1}}\left( {\partial _\theta \Theta } \right)^2 + {D ^{-1}}j^2 , \\
\end{array}
\end{eqnarray}
For obtaining a nontrivial solution, the determinant of the matrix $\Lambda$ must equals to zero. Therefore, $\det \left( \Lambda \right) = 0$ leads to the following equation
\begin{align}
\label{eq34}
{\mathop{\rm Im}\nolimits} W_ \pm' & = \pm \rm{Im} \int {\sqrt {\frac{{C\left[ {DE\omega ^2 - AEj^2 - ADEm^2 - AD\left( {\partial _z \mathcal{Z} } \right)^2 } \right] - ADE\left( {\partial _\theta \Theta } \right)^2 }}{{ABCDE}}} dr}
\nonumber \\
& = \pm i\pi \omega r_H,
\end{align}
where the plus/minus denote the outgoing/incoming solutions of massive bosons. Here, we need to neglect the real part of above equation. Finally, the tunneling rate of the massive bosons is
\begin{eqnarray}
\label{eq35}
\Gamma = \frac{{\Gamma _{\left( {emission} \right)} }}{{\Gamma _{\left( {absorpation} \right)} }} = \frac{{\exp \left( { - 2{\mathop{\rm Im}\nolimits} W_ + - 2{\mathop{\rm Im}\nolimits} \Xi } \right)}}{{\exp \left( { - 2{\mathop{\rm Im}\nolimits} W_ - - 2{\mathop{\rm Im}\nolimits} \Xi } \right)}} = \exp \left( { - {{4\pi }}\pi \omega r_H } \right).
\end{eqnarray}
According to the the Boltzman factor, the Hawking temperature of 5-dimensional black string becomes
\begin{eqnarray}
\label{eq36}
T_H ' = \frac{1}{{4\pi r_H }}.
\end{eqnarray}
Obviously, the Hawking temperature of 5-dimensional black string is only depended on the event horizon $r_H$. If people investigates the scalar particles and fermions tunneling from the 5-dimensional black string, they would obtain the same results.
\section{Discussion and conclusion}
\label{Dis}
In this paper, the tunneling behaviors of massive bosons from the event horizon of 4-dimensional rotating black string and 5-dimensional black string have been investigated via the Proca equation. By analyzing Eq.~(\ref{eq24}) and Eq.~(\ref{eq25}), one can see that the tunneling rate and Hawking temperature of 4-dimensional rotating black string are related to the the event horizon, mass and angular momentum line densities of the rotating black string, and the negative cosmological constant. For the 5-dimensional black string case, it finds that $\Gamma'$ and $T_H'$ are only dependent on the the event horizon of the black string. Moreover, when comparing our results with those of scalars and fermions tunneling cases, it finds that they are the same. Meanwhile, according to the WKB approximation, we derived the Hamilton-Jacobi ansatz from Proca equation, which indicates that the Hamilton-Jacobi equation is a fundamental equation in the semiclassical theory. Therefore, Hamilton-Jacobi ansatz can help us to investigate the semiclassical dynamic behavior of emit particles on the black holes' event horizon.
\vspace*{3.0ex}
{\bf Acknowledgements}
\vspace*{1.0ex}
This work is supported by the Natural Science Foundation of China (Grant No. 11573022).
| f7b2ca9753ab340bf0eaf379695f4c60957fd1b6 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{sec:intro}
The ATLAS and CMS collaborations have observed the Higgs boson~\cite{Aad:2012tfa, Chatrchyan:2012xdj}.
The measurements of the Higgs boson couplings to the standard model (SM) particles are essential tests of the SM. The Higgs boson couplings to the weak bosons, $HZZ$ and $HWW$, can be measured by studying the production cross sections times branching fractions~\cite{Heinemeyer:2013tqa},
by studying the angular distribution in the decay processes $H\to ZZ^*_{} \to 4l$ and $H\to WW^*_{} \to l\nu l\nu$~\cite{DellAquila:1985mtb, Nelson:1986ki, Barger:1993wt, Kramer:1993jn, Choi:2002jk, Buszello:2002uu, Hagiwara:2009wt, Gao:2010qx, Bolognesi:2012mm}
and by studying kinematics of the production processes for instance the jet azimuthal angle correlation in the weak boson fusion production~\cite{Plehn:2001nj, Hankele:2006ma, Hagiwara:2009wt, Englert:2012xt, Nakamura:2016agl}.
The ATLAS and CMS collaborations have already performed measurements of $HZZ$ and $HWW$ interactions by using the above three approaches
in refs.~\cite{Aad:2013wqa, Aad:2014eva, Aad:2015gba, Aad:2015ona, Khachatryan:2016vau, Chatrchyan:2013iaa, Khachatryan:2014ira, Khachatryan:2014jba},
in refs.~\cite{Aad:2013xqa, Aad:2015rwa, Aad:2015mxa, Chatrchyan:2012jja, Chatrchyan:2013mxa, Chatrchyan:2013iaa, Khachatryan:2014kca} and
in refs.~\cite{Aad:2015tna, Aad:2016nal, Khachatryan:2016tnr}, respectively.
The production processes $pp \to ZH$ and $pp \to W^{\pm}_{}H$~\cite{Glashow:1978ab, Barger:1986jt, Kniehl:1990iva, HAN1991167, Ohnemus:1992bd, Ciccolini:2003jy, Brein:2003wg, Ferrera:2011bk,Brein:2011vx, Denner:2011id, Dawson:2012gs, Altenkamp:2012sx, Brein:2012ne} provide direct access to the $HZZ$ and $HWW$ couplings, respectively. In these processes, not only the three approaches mentioned above, but also decay properties of the $Z$ and $W$ bosons can be used to study the $HZZ$ and $HWW$ interactions~\cite{Christensen:2010pf, Desai:2011yj, Godbole:2013lna, Delaunay:2013npa, Maltoni:2013sma, Anderson:2013afp, Godbole:2014cfa, Dwivedi:2016xwm, Ferreira:2016jea, Alioli:2017ces}. \\
An arbitrary polarised state of a massive particle with spin $s$ is entirely described by a $(2s+1) \times (2s+1)$ polarisation density matrix; see e.g. refs.~\cite{landau:1965, Schiff:2015, BOURRELY198095, Craigie:1984tk}.
When a particle with spin 1 is in an either completely or partially polarised state, it has 9 independent angular distributions of its decay products at its rest frame and the coefficients of these distributions are written in terms of all the elements of the polarisation density matrix; see e.g. refs.~\cite{BOURRELY198095, Craigie:1984tk}. An experimental determination of the complete decay angular distributions of a particle's decay products is, therefore, identical to an attempt to determine the polarisation density matrix, from which detailed information in reactions which produce that particle can be extracted; see e.g. ref.~\cite{Hagiwara:1986vm} for a theoretical study and e.g. refs.~\cite{Abbiendi:2000ei, Abbiendi:2003wv, Abdallah:2008sf} for experimental studies~\footnote{A general polarisation density matrix $\rho$ for a spin 1 massive particle has 8 degrees of freedom with the normalisation condition $tr(\rho)=1$, and it is possible to parametrise it with 8 real parameters; see e.g. refs.~\cite{BOURRELY198095, Craigie:1984tk}. Recent studies which relate these 8 parameters with observables can be found in refs.~\cite{Ots:2004hk,Ots:2006dv,Boudjema:2009fz,Aguilar-Saavedra:2015yza,Rahaman:2016pqj,Aguilar-Saavedra:2017zkn,Rahaman:2017qql}}. \\
As the $Z$ boson in the process $e^+_{}e^-_{} \to ZH$ can be in a polarised state~\cite{Kelly:1980ue, Rattazzi:1988ye, Barger:1993wt, Hagiwara:1993sw}, so the $Z$ boson in the process $pp \to ZH$ can be. The $W^+_{}$ and $W^-_{}$ in the process $pp \to W^{\pm}_{}H$ can be also in polarised states. It would be, therefore, possible to study $HZZ$, $HZ\gamma$ and $HWW$ interactions in detail from a careful analysis of these states of polarisation. As we mentioned in the previous paragraph, the polarisation density matrix of the $Z$ ($W$) boson contains the complete information about a state of polarisation of the $Z$ ($W$) boson. Hence all of the elements of the polarisation density matrix should be made use of in such a careful analysis.
In this paper, we present a systematic approach to analyse the polarisation density matrices of the $Z$ boson and $W^{\pm}_{}$ in the processes $pp \to ZH$ and $pp \to W^{\pm}_{}H$. With the aim of making maximum use of the information about states of polarisation, we relate all of the elements of the polarisation density matrices with observables which we can measure at the environment of proton-proton ($pp$) collisions. We discuss consequences of non-standard $HZZ$, $HZ\gamma$ and $HWW$ interactions for these observables.
A detailed analysis of the polarisation of the $Z$ boson in the process $e^+_{}e^-_{} \to ZH$ is present in ref.~\cite{Hagiwara:1993sw}. Our approach in the process $pp \to ZH$ is an extension of that work to $pp$ collisions. \\
The paper is organised as follows.
In Section~\ref{sec:helicityamp}, we introduce all the ingredients needed for our analyses in the following sections. At first, we give non-standard $HZZ$, $HZ\gamma$ and $HWW$ couplings. We present the complete helicity amplitudes for the sub-processes $q\bar{q}\to ZH$, $u\bar{d}\to W^+_{}H$ and $d\bar{u}\to W^-_{}H$. These amplitudes are given for the non-standard couplings. We derive the relations between the helicity amplitudes imposed by the CP and $\mathrm{CP\widetilde{T}}$ symmetries.
In Section~\ref{sec:densitymatrix}, we analyse the polarisation density matrices of the $Z$ boson in the process $pp \to ZH$ in detail.
First of all, we define the polarisation density matrices consisting of the helicity amplitudes of the previous section.
Then we derive the 4 different differential cross sections for the process $pp \to ZH$ followed by $Z \to f\bar{f}$ with respect to the $Z$ decay angles. Among the $36$ coefficients ($=9 \times 4$) of these 4 different differential angular distributions, only $15$ coefficients can be non-zero. These coefficients are written in terms of the elements of the polarisation density matrices.
The restrictions on these coefficients imposed by the CP and $\mathrm{CP\widetilde{T}}$ symmetries are clarified.
We focus on the coefficients which are strictly zero in the SM due to CP invariance or small in the SM due to the smallness of re-scattering effects, and study the influences of the non-standard $HZZ$ and $HZ\gamma$ couplings.
In Section~\ref{sec:polWboson}, we analyse the polarisation density matrices of the $W^{+}_{}$ and $W^{-}_{}$ in the process $pp \to W^{\pm}_{}H$ following the same procedure as the previous section.
Section~\ref{sec:summary} gives a summary.
\section{Constituents of the polarisation density matrices}\label{sec:helicityamp}
As the base vectors for the density matrices~\footnote{Hereafter we call a polarisation density matrix a density matrix for short.}, we choose the eigenfunctions of the helicity operator. The density matrices are, therefore, constructed of helicity amplitudes.
In this section we derive the complete helicity amplitudes for the sub-processes $q\bar{q}\to ZH$, $u\bar{d}\to W^+_{}H$ and $d\bar{u}\to W^-_{}H$.
\subsection{Effective Lagrangian with CP-odd operators}
\begin{figure}[t]
\centering
\includegraphics[scale=0.45]{pp-VH-vertex.pdf}
\hspace{1.5cm}
\includegraphics[scale=0.45]{pp-VH-vertex-W.pdf}
\caption{\small
Feynman rules for the $HZV$ ($V=Z,\gamma$) vertex and the $HWW$ vertex. }
\label{figure:vertex}
\end{figure}
We derive non-standard $HZZ$, $HZ\gamma$ and $HWW$ couplings from the following effective Lagrangian~\cite{Hagiwara:1993sw, Hagiwara:2000tk}:
\begin{align}
{\cal L}_{\mathrm{eff}}^{}
= \bigl( 1 + a_Z^{} ) \frac{1}{2} g_Z^{}m_Z^{} HZ_{\mu}^{}Z^{\mu}_{}
+ \frac{g_Z^{}}{2 m_Z^{}} \sum_{V=Z,\gamma}^{} \Bigl\{
b_V^{} HZ_{\mu\nu}^{} V^{\mu\nu}_{} +
c_V^{} \bigl[ (\partial^{\mu}_{}H) Z^{\nu}_{} - (\partial^{\nu}_{}H) Z^{\mu}_{} \bigr] V_{\mu\nu}^{} \nonumber \\
+ \frac{1}{2}\widetilde{b}_V^{} \epsilon^{\mu\nu\rho\sigma}_{} H Z_{\mu\nu}^{} V_{\rho\sigma}^{}
\Bigr\} \nonumber \\
+ \bigl( 1 + a_W^{} ) g m_W^{} HW^{\dagger}_{\mu}W^{\mu}_{}
+ \frac{g}{m_W^{}} \Bigl\{
b_W^{} HW_{\mu\nu}^{\dagger} W^{\mu\nu}_{} +
c_W^{} \bigl[ (\partial_{\mu}^{}H) W_{\nu}^{\dagger} W^{\mu\nu}_{} - (\partial^{\nu}_{}H) W^{\mu}_{} W_{\mu\nu}^{\dagger} \bigr] \nonumber \\
+ \frac{1}{2}\widetilde{b}_W^{} \epsilon^{\mu\nu\rho\sigma}_{} H W_{\mu\nu}^{\dagger} W_{\rho\sigma}^{}
\Bigr\},
\label{eq:lagrangian}
\end{align}
where $V_{\mu\nu}^{} = \partial_{\mu}^{} V_{\nu}^{} - \partial_{\nu}^{} V_{\mu}^{}$, $W_{\mu\nu}^{} = \partial_{\mu}^{} W_{\nu}^{} - \partial_{\nu}^{} W_{\mu}^{}$, $W^{\mu}_{}$ is the $W^-_{}$ field, $g=e/\sin{\theta_w^{}}$, $g_Z^{}=g/\cos{\theta_w^{}}$ with $\theta_w^{}$ being the electroweak mixing angle and $e$ being the proton charge, and $\epsilon_{0123}^{}=+1$ in our convention.
All of the eleven coefficients $a_Z^{}$, $a_W^{}$, $b_V^{}$, $c_V^{}$ and $\widetilde{b}_V^{}$ ($V=Z,\gamma,W$) are real so that ${\cal L}_{\mathrm{eff}}^{} = {\cal L}_{\mathrm{eff}}^{\dagger}$ and zero at the tree level in the SM. The operators whose coefficients are $a_Z^{}$, $a_W^{}$, $b_V^{}$ and $c_V^{}$ ($V=Z,\gamma,W$) are CP-even and the operators whose coefficients are $\widetilde{b}_V^{}$ ($V=Z,\gamma,W$) are CP-odd. The Lagrangian with only the CP-even operators indicates that $H$ is a CP-even scalar field, while the Lagrangian with only the CP-odd operators indicates that $H$ is a CP-odd scalar field. In both cases, the theory is CP invariant. If both the CP-even operator(s) and the CP-odd operator(s) exist, $H$ is no longer a CP eigenstate and the theory is not CP invariant. We call $a_Z^{}$, $a_W^{}$, $b_V^{}$ and $c_V^{}$ ($V=Z,\gamma,W$) CP-even coefficients and $\widetilde{b}_V^{}$ ($V=Z,\gamma,W$) CP-odd coefficients. \\
If we assign the momenta as shown in Figure~\ref{figure:vertex}~\footnote{All pictures in this paper are drawn by using the program JaxoDraw~\cite{Binosi:2003yf}.}, Feynman rules for the $HZV$ ($V=Z,\gamma$) vertex and the $HWW$ vertex can be expressed as
\begin{subequations}
\begin{align}
\Gamma^{\mu \nu}_V (p,k) & =m_Z^{} \biggl( \delta_{ZV}^{} + h_1^V + h_2^V \frac{\hat{s}}{m_Z^2} \biggr) g^{\mu \nu}_{} + h_3^V \frac{k^{\mu}_{} p^{\nu}_{}}{m_Z^{}} +\widetilde{h}_4^V \epsilon^{\mu \nu \alpha \beta}_{} \frac{p_{\alpha}^{}k_{\beta}^{}}{m_Z^{}}, \\
\Gamma^{\mu \nu}_W (p,k) & =m_W^{} \biggl( 1 + h_1^W + h_2^W \frac{\hat{s}}{m_W^2} \biggr) g^{\mu \nu}_{} + h_3^W \frac{k^{\mu}_{} p^{\nu}_{}}{m_W^{}} +\widetilde{h}_4^W \epsilon^{\mu \nu \alpha \beta}_{} \frac{p_{\alpha}^{}k_{\beta}^{}}{m_W^{}},
\end{align}
\end{subequations}
where $\delta_{ZV}^{}=1$ for $V=Z$, $\delta_{ZV}^{}=0$ for $V=\gamma$, and $\hat{s}=(p)^2_{}$. Terms proportional to $k^{\nu}_{}$ vanish for the on-shell $Z$ boson and the on-shell $W$ boson, and terms proportional to $p^{\mu}_{}$ also vanish in our processes where the intermediate off-shell vector boson ($Z$, $\gamma$ or $W$) couples to the four-vector consisting of the two massless quarks. These terms are, therefore, not included in the above formulae. All the form factors are constant and expressed in terms of the coefficients in the effective Lagrangian in eq.~(\ref{eq:lagrangian}):
\begin{align}
h_1^Z & = a_Z^{}+b_Z^{} - (b_Z^{}-c_Z^{}) \frac{m_H^2}{m_Z^2},
& h_1^{\gamma} & = \frac{1}{2} (b_{\gamma}^{}-c_{\gamma}^{}) \frac{m_Z^2-m_H^2}{m_Z^2},
& h_1^W & = a_W^{}+b_W^{} - (b_W^{}-c_W^{}) \frac{m_H^2}{m_W^2}, \nonumber \\
h_2^Z & = b_Z^{},
& h_2^{\gamma} & = \frac{1}{2} ( b_{\gamma}^{} + c_{\gamma}^{} ),
& h_2^W & = b_W^{}, \nonumber \\
h_3^Z & = - 2(b_Z^{}-c_Z^{}),
& h_3^{\gamma} & = - (b_{\gamma}^{}-c_{\gamma}^{}),
& h_3^W & = - 2(b_W^{}-c_W^{}), \nonumber \\
\widetilde{h}_4^Z & = -2 \widetilde{b}_Z^{},
& \widetilde{h}_4^{\gamma} & = - \widetilde{b}_{\gamma}^{},
& \widetilde{h}_4^W & = -2 \widetilde{b}_W^{}.
\end{align}
The form factors $h_i^V$ $(V=Z, \gamma, W)$ $(i=1,2,3)$ consist of the CP-even coefficients and $\widetilde{h}_4^V$ $(V=Z, \gamma, W)$ consist of the CP-odd coefficients. We call $h_i^V$ $(V=Z, \gamma, W)$ $(i=1,2,3)$ CP-even form factors and $\widetilde{h}_4^V$ $(V=Z, \gamma, W)$ CP-odd form factors.
\begin{figure}[t]
\centering
\includegraphics[scale=0.5]{pp-VH-diagrams.pdf}
\caption{\small
Feynman diagrams for the sub-processes $q\bar{q}\to ZH$, $u\bar{d}\to W^+_{}H$ and $d\bar{u}\to W^-_{}H$. The circles denote the non-standard couplings derived from the effective Lagrangian in eq.~(\ref{eq:lagrangian}).}
\label{figure:diagrams}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.5]{pp-VH-frames.pdf}
\caption{\small
{\it Left}: the $q\overline{q^{(\prime)}_{}}$ c.m. frame, where the quark ($q$) moves along the positive direction of the $z$-axis. The production amplitudes ${\cal M}_{\sigma}^{\lambda}(q \overline{q^{(\prime)}_{}})$ are evaluated in this frame.
{\it Right}: the $\overline{q^{(\prime)}_{}}q$ c.m. frame, where the antiquark ($\overline{q^{(\prime)}_{}}$) moves along the positive direction of the $z$-axis. The production amplitudes ${\cal M}_{\sigma}^{\lambda}(\overline{q^{(\prime)}_{}}q)$ are evaluated in this frame. The quark helicity $\sigma(=\pm1)$, the antiquark helicity $-\sigma$ and the weak boson helicity $\lambda(=\pm1,0)$ are shown, and the polar angle $\theta$ of the weak boson $V$ from the $z$-axis is denoted in both the frames.
}
\label{figure:frames}
\end{figure}
\subsection{Helicity amplitudes}\label{sec:helamp}
Feynman diagrams for the sub-processes $q\bar{q}\to ZH$, $u\bar{d}\to W^+_{}H$ and $d\bar{u}\to W^-_{}H$ are shown in Figure~\ref{figure:diagrams}. The circles denote the non-standard couplings derived in the previous section. We evaluate production helicity amplitudes in the following two frames. Let us assume that a direction of the $z$-axis is fixed along $pp$ collisions. In one frame, which we call the $q\overline{q^{(\prime)}_{}}$ centre-of-mass (c.m.) frame and is shown in the left picture of Figure~\ref{figure:frames}, the quark ($q$) moves along the positive direction of the $z$-axis. In another frame, which we call the $\overline{q^{(\prime)}_{}}q$ c.m. frame and is shown in the right picture of Figure~\ref{figure:frames}, the antiquark ($\overline{q^{(\prime)}_{}}$) moves along the positive direction of the $z$-axis.
In Figure~\ref{figure:frames}, our notation for helicities and the polar angle $\theta$ are also shown.
We neglect the masses of $q$ and $\overline{q^{(\prime)}_{}}$, thus the helicity of $\overline{q^{(\prime)}_{}}$ is always opposite to that of $q$ in our sub-processes. Helicity amplitudes are, therefore, given for the quark helicity $\sigma(=\pm1)$ and the weak boson helicity $\lambda(=\pm1,0)$; the antiquark helicity is automatically fixed to $-\sigma$~\footnote{The fermion helicity is always normalised to $\pm1$ in this paper.}.
We denote the helicity amplitudes evaluated in the $q\overline{q^{(\prime)}_{}}$ c.m. frame by ${\cal M}_{\sigma}^{\lambda}(q \overline{q^{(\prime)}_{}})$ and those evaluated in the $\overline{q^{(\prime)}_{}}q$ c.m. frame by ${\cal M}_{\sigma}^{\lambda}(\overline{q^{(\prime)}_{}}q)$.
It should be emphasised that $\sigma$ always denotes the helicity of the quark and $\theta$ is the polar angle of the weak boson from the $z$-axis (not the quark momentum direction).
The helicity amplitudes ${\cal M}_{\sigma}^{\lambda}(q \bar{q})$ for the sub-process $q\bar{q}\to ZH$ are
\begin{subequations}\label{eq:ZampAll1}
\begin{align}
{\cal M}_{\sigma}^{\lambda=\pm}(q \bar{q}) &= \sigma \frac{1+\sigma \lambda \cos{\theta}}{\sqrt{2}} \hat{M}_{\sigma}^{\lambda=\pm}, \label{eq:Zamp1} \\
{\cal M}_{\sigma}^{\lambda=0}(q \bar{q}) &= \sin{\theta} \hat{M}_{\sigma}^{\lambda=0}, \label{eq:Zamp2}
\end{align}
\end{subequations}
where
\begin{align}
\hat{M}_{\sigma}^{\lambda=\pm} & =
g_Z^{} m_Z^{} \sqrt{\hat{s}} \Biggl[
\frac{g_{Z \sigma}^{}}{\hat{s} - m_Z^2} \biggl( 1+ h_1^Z + h_2^Z \frac{\hat{s}}{m_Z^2} + i \lambda \widetilde{h}_4^Z \frac{k\sqrt{\hat{s}}}{m_Z^2} \biggr) +
\frac{Q_q^{}e}{\hat{s}} \biggl( h_1^{\gamma} + h_2^{\gamma} \frac{\hat{s}}{m_Z^2} + i \lambda \widetilde{h}_4^{\gamma} \frac{k\sqrt{\hat{s}}}{m_Z^2} \biggr) \Biggr], \nonumber \\
\hat{M}_{\sigma}^{\lambda=0} & = -
g_Z^{} w \sqrt{\hat{s}} \Biggl[
\frac{g_{Z \sigma}^{}}{\hat{s} - m_Z^2} \biggl( 1+h_1^Z + h_2^Z \frac{\hat{s}}{m_Z^2} + h_3^Z \frac{k^2_{}\sqrt{\hat{s}}}{m_Z^2w} \biggr) +
\frac{Q_q^{}e}{\hat{s}} \biggl( h_1^{\gamma} + h_2^{\gamma} \frac{\hat{s}}{m_Z^2} + h_3^{\gamma} \frac{k^2_{}\sqrt{\hat{s}}}{m_Z^2w} \biggr) \Biggr].
\label{eq:Zsubamplitudes}
\end{align}
Here $\sqrt{\hat{s}}$ is the $q\bar{q}$ c.m. energy, $w$ is the energy of the $Z$ boson: $w = (\hat{s}+m_Z^2-m_H^2)/(2\sqrt{\hat{s}})$, $k$ is the momentum of the $Z$ boson: $k=\sqrt{w^2_{}-m_Z^2}$, $Q_q^{}$ is the electric charge of the quark in units of $e$, $g_{Z+}^{} = g_Z^{} ( - Q_q^{} \sin^2_{}{\theta_w^{}})$, and $g_{Z-}^{} = g_Z^{} (T_q^3 - Q_q^{} \sin^2_{}{\theta_w^{}})$ where $T_u^3=1/2$ and $T_d^3=-1/2$. \\
Although the actual calculation of the helicity amplitudes ${\cal M}_{\sigma}^{\lambda}(\bar{q}q)$ is easy, it is also possible to estimate them from ${\cal M}_{\sigma}^{\lambda}(q\bar{q})$ as follows.
If we denote the helicity of $\bar{q}$ by $\sigma$ and the helicity of $q$ by $-\sigma$ (this is opposite to our helicity notation), the angular part is the same as ${\cal M}_{\sigma}^{\lambda}(q\bar{q})$ and the amplitudes are given by
\begin{subequations}\label{eq:ZampAll2-temp}
\begin{align}
{\cal M}_{}^{\lambda=\pm}\bigl(\bar{q}(\sigma)q(-\sigma) \bigr) &= \sigma \frac{1+\sigma \lambda \cos{\theta}}{\sqrt{2}} \hat{M}_{-\sigma}^{\lambda=\pm}, \label{eq:Zamp3-temp} \\
{\cal M}_{}^{\lambda=0}\bigl(\bar{q}(\sigma)q(-\sigma) \bigr) &= \sin{\theta} \hat{M}_{-\sigma}^{\lambda=0}, \label{eq:Zamp4-temp}
\end{align}
\end{subequations}
where the helicity of each particle is explicitly shown in parenthesis. Because the helicity of a quark is equal to its chirality and the helicity of an antiquark is opposite to its chirality in the massless limit, $q$ with helicity $-\sigma$ has the chirality $-\sigma$.
Hence the coupling $g_{Z \sigma}^{}$ must be replaced by $g_{Z -\sigma}^{}$ in eq.~(\ref{eq:Zsubamplitudes}), and this replacement is expressed by $\hat{M}_{-\sigma}^{\lambda}$ in eq.~(\ref{eq:ZampAll2-temp}). By the simple replacement $\sigma \to -\sigma$ in eq.~(\ref{eq:ZampAll2-temp}), we obtain the helicity amplitudes ${\cal M}_{\sigma}^{\lambda}(\bar{q}q)$ in our helicity notation:
\begin{subequations}\label{eq:ZampAll2}
\begin{align}
{\cal M}_{\sigma}^{\lambda=\pm}(\bar{q}q) &= -\sigma \frac{1-\sigma \lambda \cos{\theta}}{\sqrt{2}} \hat{M}_{\sigma}^{\lambda=\pm}, \label{eq:Zamp3}\\
{\cal M}_{\sigma}^{\lambda=0}(\bar{q}q) &= \sin{\theta} \hat{M}_{\sigma}^{\lambda=0}. \label{eq:Zamp4}
\end{align}
\end{subequations}
We note that the relative sign among the amplitudes in eq.~(\ref{eq:ZampAll1}) and that among the amplitudes in eq.~(\ref{eq:ZampAll2}) are important, because they appear in the off-diagonal elements of the density matrices. \\
Similarly, the helicity amplitudes for the sub-process $u\bar{d}\to W^+_{}H$ are given by
\begin{subequations}\label{eq:WpampAll}
\begin{align}
{\cal M}_{-}^{\lambda=\pm}(u \bar{d}) &= - \frac{1 - \lambda \cos{\theta}}{\sqrt{2}} (V^{}_{ud})^*_{}\ \hat{N}_{}^{\lambda=\pm}, \label{eq:Wpamp1} \\
{\cal M}_{-}^{\lambda=0}(u \bar{d}) &= \sin{\theta}\ (V^{}_{ud})^*_{}\ \hat{N}_{}^{\lambda=0}, \label{eq:Wpamp2} \\
{\cal M}_{-}^{\lambda=\pm}(\bar{d}u) &= \frac{1 + \lambda \cos{\theta}}{\sqrt{2}} (V^{}_{ud})^*_{}\ \hat{N}_{}^{\lambda=\pm}, \label{eq:Wpamp3}\\
{\cal M}_{-}^{\lambda=0}(\bar{d}u) &= \sin{\theta}\ (V^{}_{ud})^*_{}\ \hat{N}_{}^{\lambda=0}, \label{eq:Wpmp4}
\end{align}
\end{subequations}
and those for the sub-process $d\bar{u}\to W^-_{}H$ are given by
\begin{subequations}\label{eq:WmampAll}
\begin{align}
{\cal M}_{-}^{\lambda=\pm}(d \bar{u}) &= - \frac{1 - \lambda \cos{\theta}}{\sqrt{2}} V^{}_{ud}\ \hat{N}_{}^{\lambda=\pm}, \label{eq:Wpamm1} \\
{\cal M}_{-}^{\lambda=0}(d \bar{u}) &= \sin{\theta}\ V^{}_{ud}\ \hat{N}_{}^{\lambda=0}, \label{eq:Wpamm2} \\
{\cal M}_{-}^{\lambda=\pm}(\bar{u}d) &= \frac{1 + \lambda \cos{\theta}}{\sqrt{2}} V^{}_{ud}\ \hat{N}_{}^{\lambda=\pm}, \label{eq:Wpamm3}\\
{\cal M}_{-}^{\lambda=0}(\bar{u}d) &= \sin{\theta}\ V^{}_{ud}\ \hat{N}_{}^{\lambda=0}, \label{eq:Wpmm4}
\end{align}
\end{subequations}
where
\begin{align}
\hat{N}_{}^{\lambda=\pm} & =
\frac{1}{\sqrt{2}}g^2_{} m_W^{} \sqrt{\hat{s}}
\frac{1}{\hat{s} - m_W^2} \biggl( 1 + h_1^W + h_2^W \frac{\hat{s}}{m_W^2} + i \lambda \widetilde{h}_4^W \frac{k\sqrt{\hat{s}}}{m_W^2} \biggr), \nonumber \\
\hat{N}_{}^{\lambda=0} & = -
\frac{1}{\sqrt{2}}g^2_{} w \sqrt{\hat{s}}
\frac{1}{\hat{s} - m_W^2} \biggl( 1 + h_1^W + h_2^W \frac{\hat{s}}{m_W^2} + h_3^W \frac{k^2_{}\sqrt{\hat{s}}}{m_W^2 w} \biggr).
\label{eq:Wsubamplitudes}
\end{align}
Here $w$ is the energy of the $W$ boson: $w = (\hat{s}+m_W^2-m_H^2)/(2\sqrt{\hat{s}})$, $k$ is the momentum of the $W$ boson: $k=\sqrt{w^2_{}-m_W^2}$, and $V_{ud}^{}$ is the element of the Cabibbo-Kobayashi-Maskawa (CKM) matrix.
\\
By looking at the helicity amplitudes, we can already discuss a difference between the $Z$ boson and the $W$ boson with regard to states of polarisation.
Since the $W$ boson couples to only a fermion with chirality $-1$, the initial quark $u$ or $d$ always has the helicity $\sigma=-1$.
Therefore, while the $Z$ boson is in a partially polarised state (often called a mixed state; see e.g. refs.~\cite{landau:1965, Schiff:2015, BOURRELY198095, Craigie:1984tk}), the $W$ boson (both the $W^+_{}$ and $W^-_{}$) is in a completely polarised state (often called a pure state).
\subsection{Requests from symmetries}
\begin{figure}[t]
\centering
\includegraphics[scale=0.42]{pp-VH-CP-Rot.pdf}
\includegraphics[scale=0.42]{pp-WH-CP-Rot.pdf}
\caption{\small The original states, the states after CP transformation and the states after the rotation around the $z$-axis by $\pi$ are shown for the process $q\bar{q} \to ZH$ (upper figures) and the process $u\bar{d} \to W^+_{}H$ (lower figures).}
\label{figure:CP}
\end{figure}
Conditions imposed by symmetries lead to certain relations between the helicity amplitudes. The upper figures in Figure~\ref{figure:CP} show the original states in the process $q\bar{q} \to ZH$ (left), the states after CP transformation (middle) and the states after the rotation around the $z$-axis by $\pi$ (right).
From these figures, we find that the invariance under CP leads to the relation
\begin{align}
{\cal M}_{\sigma}^{\lambda}(q \bar{q}) (\theta) = {\cal M}_{\sigma}^{-\lambda}(q \bar{q}) (\pi-\theta), \label{eq:CPinvZ}
\end{align}
where ${\cal M}_{\sigma}^{\lambda}(q \bar{q}) (\theta) = {\cal M}_{\sigma}^{\lambda}(q \bar{q})$ given in eq.~(\ref{eq:ZampAll1}), and ${\cal M}_{\sigma}^{-\lambda}(q \bar{q}) (\pi-\theta)$ is obtained by setting $\theta \to \pi - \theta$ in ${\cal M}_{\sigma}^{-\lambda}(q \bar{q})$.
It is easy to see that a non-zero value of the CP-odd form factors $\widetilde{h}_4^Z$ and/or $\widetilde{h}_4^{\gamma}$ violates this relation.
Similarly, the lower figures in Figure~\ref{figure:CP} show the original states in the process $u\bar{d} \to W^+_{}H$ (left), the states after CP transformation (middle) and the states after the rotation around the $z$-axis by $\pi$ (right). From these figures, we find that CP invariance requests the relation
\begin{align}
{\cal M}_{-}^{\lambda}(u \bar{d}) (\theta) = {\cal M}_{-}^{-\lambda}(d \bar{u}) (\pi-\theta), \label{eq:CPinvW}
\end{align}
where ${\cal M}_{-}^{\lambda}(u \bar{d}) (\theta) = {\cal M}_{-}^{\lambda}(u \bar{d})$ given in eq.~(\ref{eq:WpampAll}), and ${\cal M}_{-}^{-\lambda}(d \bar{u}) (\pi-\theta)$ is obtained by setting $\theta \to \pi - \theta$ in ${\cal M}_{-}^{-\lambda}(d \bar{u})$ given in eq.~(\ref{eq:WmampAll}). This relation is violated by the imaginary part of the element $V_{ud}^{}$ of the CKM matrix, even when the CP-odd form factor $\widetilde{h}_4^W$ is zero.
However, this CP violation phase in $V_{ud}^{}$~\cite{Kobayashi:1973fv} is always regarded as an overall common phase among the amplitudes.
For instance, the phase in $V^{}_{ud}$ is an overall common phase among ${\cal M}_{-}^{\lambda=+}(d \bar{u})$, ${\cal M}_{-}^{\lambda=-}(d \bar{u})$ and ${\cal M}_{-}^{\lambda=0}(d \bar{u})$.
Therefore, the effect of the CP violation phase in the CKM matrix never appears even in the off-diagonal elements of the density matrices. Only a non-zero value of the CP-odd form factor $\widetilde{h}_4^W$ violates the relation in eq.~(\ref{eq:CPinvW}) {\it and} affects a state of polarisation of the $W$ boson.
We can write the relations in eqs.~(\ref{eq:CPinvZ}) and (\ref{eq:CPinvW}) in the following convenient forms, respectively:
\begin{subequations}
\begin{align}
\hat{M}_{\sigma}^{\lambda} & = \hat{M}_{\sigma}^{-\lambda}, \label{eq:CPinvZ-2}\\
\hat{N}_{}^{\lambda} & = \hat{N}_{}^{-\lambda}.
\end{align}
\end{subequations}
CPT invariance {\it and} the unitarity condition conclude~\cite{Hagiwara:1986vm} that the following relations hold at the tree level approximation and violation of the relations indicates the existence of re-scattering effects:
\begin{subequations}\label{eq:CPTinv}
\begin{align}
{\cal M}_{\sigma}^{\lambda}(q \bar{q}) (\theta) &= \bigl\{ {\cal M}_{\sigma}^{-\lambda}(q \bar{q}) (\pi-\theta) \bigr\}^*_{}, \label{eq:CPTinvZ} \\
{\cal M}_{-}^{\lambda}(u \bar{d}) (\theta) &= \bigl\{ {\cal M}_{-}^{-\lambda}(d \bar{u}) (\pi-\theta) \bigr\}^*_{}. \label{eq:CPTinvW}
\end{align}
\end{subequations}
It is easy to see that the relations hold if all of the form factors are real.
Following ref.~\cite{Hagiwara:1986vm}, we call the invariance which leads to these relations $\mathrm{CP\widetilde{T}}$ invariance.
We can write the relations in eq.~(\ref{eq:CPTinv}) in the following convenient forms, respectively:
\begin{subequations}
\begin{align}
\hat{M}_{\sigma}^{\lambda} & = \bigl( \hat{M}_{\sigma}^{-\lambda} \bigr)^*_{}, \label{eq:CPTinvZ-2}\\
\hat{N}_{}^{\lambda} & = \bigl( \hat{N}_{}^{-\lambda} \bigr)^*_{}.
\end{align}
\end{subequations}
\section{Polarisation of the $Z$ boson}\label{sec:densitymatrix}
\subsection{Polarisation density matrices of the $Z$ boson}\label{sec:derivation}
In this section we define the density matrices of the $Z$ boson by using our notation of the helicity amplitudes of the previous section. We consider the sub-process
\begin{align}
q(\sigma) + \bar{q}( -\sigma ) & \to Z(\lambda) + H; \nonumber \\
Z(\lambda) & \to f(\tau) + \bar{f}(-\tau),
\end{align}
where the helicity of each particle is shown in parenthesis. We neglect the masses of the final fermion $f$ and the final antifermion $\bar{f}$, thus the helicity of $\bar{f}$ is always opposite to that of $f$. We express the full helicity amplitude as
\begin{align}
{\cal T}_{\sigma}^{\tau}(q\bar{q}) = P_Z^{} \sum_{\lambda=\pm,0}^{} {\cal M}_{\sigma}^{\lambda}(q \bar{q})\ D_{\lambda}^{\tau},\label{eq:fullampZ}
\end{align}
where the production amplitude ${\cal M}_{\sigma}^{\lambda}(q \bar{q})$ is given in eq.~(\ref{eq:ZampAll1}), $D_{\lambda}^{\tau}$ is the decay helicity amplitude, and
\begin{align}
P_Z^{} = ( Q^2_{} - m_Z^2 + i m_Z^{} \Gamma_Z^{} )^{-1}_{}
\end{align}
denotes the propagator factor of the $Z$ boson.
We evaluate the decay amplitude in the following four-momentum frame:
\begin{align}
Z:&\ \ \bigl( m_Z^{}, 0, 0, 0 \bigr) \nonumber \\
f:&\ \ \frac{m_Z^{}}{2}\bigl( 1,\ \sin{\widehat{\theta}} \cos{\widehat{\phi}},\ \sin{\widehat{\theta}} \sin{\widehat{\phi}},\ \cos{\widehat{\theta}} \bigr) \nonumber \\
\bar{f} :&\ \ \frac{m_Z^{}}{2}\bigl( 1,\ -\sin{\widehat{\theta}} \cos{\widehat{\phi}},\ -\sin{\widehat{\theta}} \sin{\widehat{\phi}},\ -\cos{\widehat{\theta}} \bigr). \label{eq:decaykinematics}
\end{align}
The decay amplitude is
\begin{align}
D_{\lambda}^{\tau} = g_{Zf\bar{f}}^{\tau}\ m_Z^{}\ d_{\lambda}^{\tau},
\end{align}
where
\begin{subequations}\label{eq:decayamp}
\begin{align}
d_{\lambda=\pm}^{\tau} & = \tau \frac{1+\lambda \tau \cos{\widehat{\theta}}}{\sqrt{2}} e^{i \lambda \widehat{\phi}}_{}, \\
d_{\lambda=0}^{\tau} & = \sin{\widehat{\theta}}.
\end{align}
\end{subequations}
The explicit forms of the coupling $g_{Zf\bar{f}}^{\tau}$ are, $g_{Zl\bar{l}}^{+} = g_Z^{} \sin^2_{}{\theta_w^{}}$ and $g_{Zl\bar{l}}^{-} = g_Z^{} ( -1/2+\sin^2_{}{\theta_w^{}} )$ for a charged lepton pair, $g_{Zu\bar{u}}^{+} = -(2/3)g_Z^{}\sin^2_{}{\theta_w^{}}$ and $g_{Zu\bar{u}}^{-} = g_Z^{} (1/2-(2/3)\sin^2_{}{\theta_w^{}})$ for a up-type quark pair, $g_{Zd\bar{d}}^{+} = (1/3)g_Z^{} \sin^2_{}{\theta_w^{}}$ and $g_{Zd\bar{d}}^{-} = g_Z^{} (-1/2+(1/3)\sin^2_{}{\theta_w^{}})$ for a down-type quark pair. A straightforward manipulation gives
\begin{align}
\sum_{\sigma}^{} \bigl| {\cal T}_{\sigma}^{\tau}(q\bar{q}) \bigr|^2_{}
& = \bigl|P_Z^{} m_Z^{} g_{Zf\bar{f}}^{\tau} \bigr|^2_{}
\sum_{\sigma}^{}\sum_{\lambda^{\prime}_{}, \lambda} \bigl( d_{\lambda^{\prime}}^{\tau} \bigr)^*_{} \rho_{\sigma}^{\lambda^{\prime} \lambda}(q\bar{q}) d_{\lambda}^{\tau} \nonumber \\
& = \bigl|P_Z^{} m_Z^{} g_{Zf\bar{f}}^{\tau} \bigr|^2_{}
d_{}^{\tau \dagger} \sum_{\sigma}^{} \rho_{\sigma}^{}(q\bar{q}) d_{}^{\tau}, \label{eq:ampsquaredZ}
\end{align}
where
\begin{align}
\sum_{\sigma} \rho_{\sigma}^{\lambda^{\prime} \lambda}(q\bar{q}) = \sum_{\sigma}
\bigl\{ {\cal M}_{\sigma}^{\lambda^{\prime}_{}}(q \bar{q}) \bigr\}^{\ast}_{}
{\cal M}_{\sigma}^{\lambda}(q \bar{q})\label{eq:densityMqqbarZ}
\end{align}
represents the elements of the density matrix in the helicity basis of the $Z$ boson in the $q\bar{q}$ c.m. frame, and at the last equality the following $3\times3$ matrix form is employed:
\begin{align}
\rho_{\sigma}^{}(q\bar{q}) =
\left(
\begin{array}{ccc}
\rho_{\sigma}^{++}(q\bar{q}) & \rho_{\sigma}^{+-}(q\bar{q}) & \rho_{\sigma}^{+0}(q\bar{q}) \\
\rho_{\sigma}^{-+}(q\bar{q}) & \rho_{\sigma}^{--}(q\bar{q}) & \rho_{\sigma}^{-0}(q\bar{q}) \\
\rho_{\sigma}^{0+}(q\bar{q}) & \rho_{\sigma}^{0-}(q\bar{q}) & \rho_{\sigma}^{00}(q\bar{q}) \\
\end{array}
\right)
,\ \ \
d^{\tau}_{} =
\left(
\begin{array}{c}
d^{\tau}_{+} \\
d^{\tau}_{-} \\
d^{\tau}_{0} \\
\end{array}
\right).\label{eq:matrixform}
\end{align}
The density matrix is a $3\times3$ Hermitian matrix: $\sum_{\sigma} \rho_{\sigma}^{}(q\bar{q}) = ( \sum_{\sigma} \rho_{\sigma}^{}(q\bar{q}) )^{\dagger}_{}$~\footnote{In general a density matrix $\rho$ has a normalisation condition $tr(\rho)=1$; see e.g. refs.~\cite{landau:1965, Schiff:2015, BOURRELY198095, Craigie:1984tk}. However, we employ the non-normalised form such as eq.~(\ref{eq:densityMqqbarZ}) and call it a density matrix in this paper. The degrees of freedom of our density matrices is, therefore, not 8 but 9.}.\\
The full helicity amplitude ${\cal T}_{\sigma}^{\tau}(\bar{q}q)$, in which the production amplitude is evaluated in the $\bar{q}q$ c.m. frame shown in the right picture of Figure~\ref{figure:frames}, can be treated in the same manner:
\begin{subequations}
\begin{align}
{\cal T}_{\sigma}^{\tau}(\bar{q}q) & = P_Z^{} \sum_{\lambda=\pm,0}^{} {\cal M}_{\sigma}^{\lambda}(\bar{q}q)\ D_{\lambda}^{\tau}, \label{eq:fullampZqbarq} \\
\sum_{\sigma}^{} \bigl| {\cal T}_{\sigma}^{\tau}(\bar{q}q) \bigr|^2_{}
& = \bigl|P_Z^{} m_Z^{} g_{Zf\bar{f}}^{\tau} \bigr|^2_{}
\sum_{\sigma}^{}\sum_{\lambda^{\prime}_{}, \lambda} \bigl( d_{\lambda^{\prime}}^{\tau} \bigr)^*_{} \rho_{\sigma}^{\lambda^{\prime} \lambda}(\bar{q}q) d_{\lambda}^{\tau} \nonumber \\
& = \bigl|P_Z^{} m_Z^{} g_{Zf\bar{f}}^{\tau} \bigr|^2_{}
d_{}^{\tau \dagger} \sum_{\sigma}^{} \rho_{\sigma}^{}(\bar{q}q) d_{}^{\tau}, \label{eq:ampsquaredZqbarq}
\end{align}
\end{subequations}
where
\begin{align}
\sum_{\sigma} \rho_{\sigma}^{\lambda^{\prime} \lambda}(\bar{q} q) = \sum_{\sigma}
\bigl\{ {\cal M}_{\sigma}^{\lambda^{\prime}_{}}(\bar{q} q) \bigr\}^{\ast}_{}
{\cal M}_{\sigma}^{\lambda}(\bar{q} q)\label{eq:densityMqbarqZ}
\end{align}
represents the elements of the density matrix in the helicity basis of the $Z$ boson in the $\bar{q}q$ c.m. frame. The production amplitude ${\cal M}_{\sigma}^{\lambda}(\bar{q} q)$ is given in eq.~(\ref{eq:ZampAll2}).
\subsection{Decay angular distributions of the polarised $Z$ boson}\label{sec:crossZboson}
In terms of the density matrices defined in eqs.~(\ref{eq:densityMqqbarZ}) and (\ref{eq:densityMqbarqZ}),
the complete differential cross section for the process $pp\to ZH$ followed by $Z\to f\bar{f}$ in the narrow width approximation can be expressed as follows:
\begin{align}
\frac{d\sigma}{d\hat{s}\ dy\ d\cos{\theta}\ d\cos{\widehat{\theta}}\ d\widehat{\phi} }
= & \frac{m_Z^{}k}{ 12288 \pi^3_{} \Gamma_Z^{} s \hat{s}^{\frac{3}{2}}_{}}
\sum_f \sum_{\tau} \bigl| g_{Zf\bar{f}}^{\tau} \bigr|^2_{} C_f^{} \nonumber \\
& \times \sum_{q} \biggl[ q(x_1^{}) \bar{q}(x_2^{})\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \rho_{\sigma}^{}(q\bar{q}) d_{}^{\tau}
+
\bar{q}(x_1^{}) q(x_2^{})\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \rho_{\sigma}^{}(\bar{q}q) d_{}^{\tau} \biggr],\label{eq:differentialZH}
\end{align}
where $s$ is the c.m. energy squared of the $pp$ collisions, $\hat{s}$ is the c.m. energy squared of the $q\bar{q}$ collisions,
$y$ is the rapidity of the $q\bar{q}$ c.m. frame and the $\bar{q}q$ c.m. frame (see Figure~\ref{figure:frames}) on the $pp$ c.m. frame (i.e. the experimental frame),
$C_f^{}$ is an effective colour factor: $C_f^{}=1$ for $Z$ decays into a charged lepton pair and $C_f^{}=3$ for $Z$ decays into a quark pair, $q(x_i^{})$ and $\bar{q}(x_i^{})$ are the quark and antiquark parton distribution functions (PDFs) with energy fraction $x_i^{}$. Summations are performed over the final fermion flavor $f$, the helicity $\tau$ of the final fermion, and the initial quark and antiquark flavor $q$. Averages are performed over the helicity and colour of the initial quark, and those of the initial antiquark. Recall that $k$ is the momentum of the $Z$ boson and $\theta$ is the polar angle of the $Z$ boson in the $q\bar{q}$ c.m. frame and the $\bar{q}q$ c.m. frame. The decay angles $\widehat{\theta}$ and $\widehat{\phi}$ are defined in eq.~(\ref{eq:decaykinematics}).
Here $x_1^{}$ ($x_2^{}$) denotes the energy fraction of the initial quark or antiquark in the proton that moves along the positive (negative) direction of the $z$-axis, once a direction of the $z$-axis is fixed along the $pp$ collisions.
The integration variables $\hat{s}$ and $y$ are related to $x_1^{}$ and $x_2^{}$:
\begin{align}
\hat{s} = s x_1^{} x_2^{},\ \ \ y = \frac{1}{2}\ln{\frac{x_1^{}}{x_2^{}}}.
\end{align}
The allowed region is
\begin{align}
(m_Z^{} + m_H^{})^2_{} < \hat{s} < s, \ \ \ - \ln{\sqrt{\frac{s}{\hat{s}}}} < y < \ln{\sqrt{\frac{s}{\hat{s}}}}.\label{eq:allowedregion}
\end{align}
\\
\renewcommand{\arraystretch}{2.5}
\begin{table}[t]
\centering
\begin{tabular}{c r r}
\toprule
\addlinespace[-3mm]
Label & Integration over $y$ & Integration over $\cos{\theta}$ \\
\addlinespace[-2mm]
\midrule
${\cal A}$ & \scalebox{1.}{$\displaystyle \int_{-y_{\mathrm{cut}}^{}}^{y_{\mathrm{cut}}^{}} dy$} & \scalebox{1.}{$\displaystyle \int^{1-\epsilon}_{-(1-\epsilon)} d \cos{\theta}$} \\
${\cal B}$ & \scalebox{1.}{$\displaystyle \biggl( \int_{0}^{y_{\mathrm{cut}}^{}} - \int_{-y_{\mathrm{cut}}^{}}^{0} \biggr) dy$} & \scalebox{1.}{$\displaystyle \int^{1-\epsilon}_{-(1-\epsilon)} d \cos{\theta}$} \\
${\cal C}$ & \scalebox{1.}{$\displaystyle \int_{-y_{\mathrm{cut}}^{}}^{y_{\mathrm{cut}}^{}} dy$} & \scalebox{1.}{$\displaystyle \biggl( \int^{1-\epsilon}_{0} - \int^{0}_{-(1-\epsilon)} \biggr) d \cos{\theta}$} \\
${\cal D}$ & \scalebox{1.}{$\displaystyle \biggl( \int_{0}^{y_{\mathrm{cut}}^{}} - \int_{-y_{\mathrm{cut}}^{}}^{0} \biggr) dy$} & \scalebox{1.}{$\displaystyle \biggl( \int^{1-\epsilon}_{0} - \int^{0}_{-(1-\epsilon)} \biggr) d \cos{\theta}$} \\
\addlinespace[1mm]
\bottomrule
\end{tabular}
\caption{\small The 4 different approaches labelled ${\cal A}$, ${\cal B}$, ${\cal C}$ and ${\cal D}$ for performing integration over the rapidity $y$ and the polar angle $\cos{\theta}$ in the differential cross section of eq.~(\ref{eq:differentialZH}). A value of $y_{\mathrm{cut}}^{}$ is $\ln{\sqrt{s/\hat{s}}}$; see eq.~(\ref{eq:allowedregion}). A smaller value can be also chosen for $y_{\mathrm{cut}}^{}$. A value of $\epsilon$ ($0 \le \epsilon < 1$) will be determined according to experimental conditions.}
\label{table:integration}
\end{table}
\renewcommand{\arraystretch}{1}
\renewcommand{\arraystretch}{1.5}
\begin{table}[t]
\centering
\begin{tabular}{c c c}
\midrule
$c_1^{}$ & $\frac{4}{3} -2\epsilon + \epsilon^2_{} - \frac{1}{3}\epsilon^3_{}$ & $\frac{4}{3}$ \\
$c_2^{}$ & $\frac{4}{3} -2\epsilon^2_{} + \frac{2}{3}\epsilon^3_{}$ & $\frac{4}{3}$ \\
$c_3^{}$ & $\frac{1}{\sqrt{2}} \bigl[(1-\epsilon)\sqrt{\epsilon(2-\epsilon)} + \sin^{-1}_{}{(1-\epsilon)} \bigr]$ & $\frac{\pi}{2\sqrt{2}}$ \\
$c_4^{}$ & $\frac{2}{3\sqrt{2}} \bigl[1-2 \epsilon \sqrt{\epsilon(2-\epsilon)} + \sqrt{\epsilon^5_{}(2-\epsilon)} \bigr]$ & $\frac{2}{3\sqrt{2}}$ \\
$c_5^{}$ & $1 -2\epsilon + \epsilon^2_{}$ & $1$ \\
\midrule
\end{tabular}
\caption{\small Constant coefficients in the density matrices after $\cos{\theta}$ integration; see eqs.~(\ref{eq:cosInt1}) and (\ref{eq:cosInt2}). The last column gives the values for $\epsilon =0$.}
\label{table:coefficients}
\end{table}
\renewcommand{\arraystretch}{1}
Now we derive the differential cross sections with respect to $\hat{s}$, $\cos{\widehat{\theta}}$ and $\widehat{\phi}$ by integrating over $y$ and $\cos{\theta}$ in eq.~(\ref{eq:differentialZH}). We consider the 4 different integration approaches summarised in Table~\ref{table:integration}.
In the complete differential cross section of eq.~(\ref{eq:differentialZH}), the polar angle $\theta$ dependence appears only in $\rho_{\sigma}^{}(q\bar{q})$ and $\rho_{\sigma}^{}(\bar{q}q)$. The $\cos{\theta}$ integration in the approaches ${\cal A}$ and ${\cal B}$ results in
\begin{subequations}\label{eq:cosInt1}
\begin{align}
\bigl\langle \rho_{\sigma}^{}(q\bar{q}) \bigr\rangle
\equiv \int^{1-\epsilon}_{-(1-\epsilon)} d \cos{\theta}\ \rho_{\sigma}^{}(q\bar{q})
=
\left(
\begin{array}{ccc}
c_1^{} \bigl| \hat{M}_{\sigma}^+ \bigr|^2_{} & \frac{1}{2} c_2^{} \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{-} & c_3^{} \sigma \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} \\
\frac{1}{2} c_2^{} \hat{M}_{\sigma}^{+} \bigl( \hat{M}_{\sigma}^{-} \bigr)^*_{} & c_1^{} \bigl| \hat{M}_{\sigma}^- \bigr|^2_{} & c_3^{} \sigma \bigl( \hat{M}_{\sigma}^{-} \bigr)^*_{} \hat{M}_{\sigma}^{0} \\
c_3^{} \sigma \hat{M}_{\sigma}^{+} \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} & c_3^{} \sigma \hat{M}_{\sigma}^{-} \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} & c_2^{} \bigl| \hat{M}_{\sigma}^0 \bigr|^2_{} \\
\end{array}
\right)\label{eq:densqqbarint1}
\end{align}
and
\begin{align}
\bigl\langle \rho_{\sigma}^{}(\bar{q}q) \bigr\rangle
\equiv \int^{1-\epsilon}_{-(1-\epsilon)} d \cos{\theta}\ \rho_{\sigma}^{}(\bar{q}q)
=
\left(
\begin{array}{ccc}
c_1^{} \bigl| \hat{M}_{\sigma}^+ \bigr|^2_{} & \frac{1}{2} c_2^{} \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{-} & - c_3^{} \sigma \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} \\
\frac{1}{2} c_2^{} \hat{M}_{\sigma}^{+} \bigl( \hat{M}_{\sigma}^{-} \bigr)^*_{} & c_1^{} \bigl| \hat{M}_{\sigma}^- \bigr|^2_{} & - c_3^{} \sigma \bigl( \hat{M}_{\sigma}^{-} \bigr)^*_{} \hat{M}_{\sigma}^{0} \\
- c_3^{} \sigma \hat{M}_{\sigma}^{+} \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} & - c_3^{} \sigma \hat{M}_{\sigma}^{-} \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} & c_2^{} \bigl| \hat{M}_{\sigma}^0 \bigr|^2_{} \\
\end{array}
\right), \label{eq:densqbarqint1}
\end{align}
\end{subequations}
where $c_i^{}$ $(i=1,2,3)$ are constant values depending on $\epsilon$ and summarised in Table~\ref{table:coefficients}, and $\hat{M}_{\sigma}^{\lambda=\pm,0}$ are defined in eq.~(\ref{eq:Zsubamplitudes}).
Notice the difference between eq.~(\ref{eq:densqqbarint1}) and eq.~(\ref{eq:densqbarqint1}) that there is a minus sign in front of the elements that have the overall $\sigma$ in eq.~(\ref{eq:densqbarqint1}). This result is actually obvious from the comparison between the amplitudes in eq.~(\ref{eq:ZampAll1}) and those in eq.~(\ref{eq:ZampAll2}).
After integration over $y$ and $\cos{\theta}$ in the approach ${\cal A}$, the differential cross section can be expressed as
\begin{align}
\frac{d\sigma}{d\hat{s}\ d\cos{\widehat{\theta}}\ d\widehat{\phi} }\biggr|_{{\cal A}}^{}
& = {\cal T} \int_{-y_{\mathrm{cut}}^{}}^{y_{\mathrm{cut}}^{}} dy \biggl[ q(x_1^{}) \bar{q}(x_2^{})\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \bigl \langle \rho_{\sigma}^{}(q\bar{q}) \bigr \rangle d_{}^{\tau}
+
\bar{q}(x_1^{}) q(x_2^{})\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \bigl \langle \rho_{\sigma}^{}(\bar{q}q) \bigr \rangle d_{}^{\tau} \biggr] \nonumber \\
& = {\cal T} \underbrace{\int_{0}^{y_{\mathrm{cut}}^{}}}_{x_1^{}>x_2^{}} dy \biggl[ \underbrace{q(x_1^{}) \bar{q}(x_2^{})}_{A}\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \bigl \langle \rho_{\sigma}^{}(q\bar{q}) \bigr \rangle d_{}^{\tau}
+
\underbrace{\bar{q}(x_1^{}) q(x_2^{})}_{B}\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \bigl \langle \rho_{\sigma}^{}(\bar{q}q) \bigr \rangle d_{}^{\tau} \biggr] \nonumber \\
& +
{\cal T} \underbrace{\int^{0}_{-y_{\mathrm{cut}}^{}}}_{x_2^{}>x_1^{}} dy \biggl[ \underbrace{\bar{q}(x_2^{}) q(x_1^{})}_B \ d_{}^{\tau \dagger} \sum_{\sigma}^{} \bigl \langle \rho_{\sigma}^{}(q\bar{q}) \bigr \rangle d_{}^{\tau}
+
\underbrace{q(x_2^{}) \bar{q}(x_1^{})}_{A}\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \bigl \langle \rho_{\sigma}^{}(\bar{q}q) \bigr \rangle d_{}^{\tau} \biggr]
\nonumber \\
=
{\cal T} \int_{0}^{y_{\mathrm{cut}}^{}} & dy\ 2 \Bigl[ q(x_1^{}) \bar{q}(x_2^{}) + \bar{q}(x_1^{}) q(x_2^{}) \Bigr]
d_{}^{\tau \dagger} \sum_{\sigma}^{}
\left(
\begin{array}{ccc}
c_1^{} \bigl| \hat{M}_{\sigma}^+ \bigr|^2_{} & \frac{1}{2} c_2^{} \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{-} & 0 \\
\frac{1}{2} c_2^{} \hat{M}_{\sigma}^{+} \bigl( \hat{M}_{\sigma}^{-} \bigr)^*_{} & c_1^{} \bigl| \hat{M}_{\sigma}^- \bigr|^2_{} & 0 \\
0 & 0 & c_2^{} \bigl| \hat{M}_{\sigma}^0 \bigr|^2_{} \\
\end{array}
\right)
d_{}^{\tau}
,\label{eq:PartdifferentialZH1}
\end{align}
where
\begin{align}
{\cal T} =\frac{m_Z^{}k}{ 12288 \pi^3_{} \Gamma_Z^{} s \hat{s}^{\frac{3}{2}}_{}}
\sum_f \sum_{\tau} \bigl| g_{Zf\bar{f}}^{\tau} \bigr|^2_{} C_f^{} \sum_q \label{eq:simplywriting}
\end{align}
is introduced to simplify our writing. The PDFs labelled $A$ in the second equality (i.e. $q(x_1^{}) \bar{q}(x_2^{})$ and $q(x_2^{}) \bar{q}(x_1^{})$) give the same numerical contribution after integration over $y$, therefore they are combined in the third (i.e. last) equality. The same is done for the PDFs labelled $B$.
In the last equality, $\bigl \langle \rho_{\sigma}^{}(q\bar{q}) \bigr \rangle$ and $\bigl \langle \rho_{\sigma}^{}(\bar{q}q) \bigr \rangle$ are added, and the elements that have the overall $\sigma$ vanish due to the sign difference in $\bigl \langle \rho_{\sigma}^{}(q\bar{q}) \bigr \rangle$ and $\bigl \langle \rho_{\sigma}^{}(\bar{q}q) \bigr \rangle$.
The vanished elements in the approach ${\cal A}$ in eq.~(\ref{eq:PartdifferentialZH1}) revive by performing integration in the approach ${\cal B}$:
\begin{align}
\frac{d\sigma}{d\hat{s}\ d\cos{\widehat{\theta}}\ d\widehat{\phi} }\biggr|_{{\cal B}}^{}
& = {\cal T} \biggl( \int_{0}^{y_{\mathrm{cut}}^{}} - \int^{0}_{-y_{\mathrm{cut}}^{}} \biggr)dy \biggl[ q(x_1^{}) \bar{q}(x_2^{})\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \bigl \langle \rho_{\sigma}^{}(q\bar{q}) \bigr \rangle d_{}^{\tau}
+
\bar{q}(x_1^{}) q(x_2^{})\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \bigl \langle \rho_{\sigma}^{}(\bar{q}q) \bigr \rangle d_{}^{\tau} \biggr] \nonumber \\
& = {\cal T} \underbrace{\int_{0}^{y_{\mathrm{cut}}^{}}}_{x_1^{}>x_2^{}} dy \biggl[ \underbrace{q(x_1^{}) \bar{q}(x_2^{})}_{A}\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \bigl \langle \rho_{\sigma}^{}(q\bar{q}) \bigr \rangle d_{}^{\tau}
+
\underbrace{\bar{q}(x_1^{}) q(x_2^{})}_{B}\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \bigl \langle \rho_{\sigma}^{}(\bar{q}q) \bigr \rangle d_{}^{\tau} \biggr] \nonumber \\
& -
{\cal T} \underbrace{\int^{0}_{-y_{\mathrm{cut}}^{}}}_{x_2^{}>x_1^{}} dy \biggl[ \underbrace{\bar{q}(x_2^{}) q(x_1^{})}_B \ d_{}^{\tau \dagger} \sum_{\sigma}^{} \bigl \langle \rho_{\sigma}^{}(q\bar{q}) \bigr \rangle d_{}^{\tau}
+
\underbrace{q(x_2^{}) \bar{q}(x_1^{})}_{A}\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \bigl \langle \rho_{\sigma}^{}(\bar{q}q) \bigr \rangle d_{}^{\tau} \biggr]
\nonumber \\
=
{\cal T} \int_{0}^{y_{\mathrm{cut}}^{}} & dy\ 2 \Bigl[ q(x_1^{}) \bar{q}(x_2^{}) - \bar{q}(x_1^{}) q(x_2^{}) \Bigr]
d_{}^{\tau \dagger} \sum_{\sigma}^{}
\left(
\begin{array}{ccc}
0 & 0 & c_3^{} \sigma \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} \\
0 & 0 & c_3^{} \sigma \bigl( \hat{M}_{\sigma}^{-} \bigr)^*_{} \hat{M}_{\sigma}^{0} \\
c_3^{} \sigma \hat{M}_{\sigma}^{+} \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} & c_3^{} \sigma \hat{M}_{\sigma}^{-} \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} & 0 \\
\end{array}
\right)
d_{}^{\tau},\label{eq:PartdifferentialZH2}
\end{align}
where, in the last equality, $\bigl \langle \rho_{\sigma}^{}(q\bar{q}) \bigr \rangle$ and $\bigl \langle \rho_{\sigma}^{}(\bar{q}q) \bigr \rangle$ are subtracted. As a result, in contrast to the approach ${\cal A}$, only the elements that have the overall $\sigma$ survive. \\
The $\cos{\theta}$ integration in the approaches ${\cal C}$ and ${\cal D}$ results in
\begin{subequations}\label{eq:cosInt2}
\begin{align}
\overline{ \bigl \langle \rho_{\sigma}^{}(q\bar{q}) \bigr \rangle } \equiv
\biggl( \int^{1-\epsilon}_{0} - \int^{0}_{-(1-\epsilon)} \biggr) d \cos{\theta}\ \rho_{\sigma}^{}(q\bar{q})
=
\left(
\begin{array}{ccc}
c_5^{} \sigma \bigl| \hat{M}_{\sigma}^+ \bigr|^2_{} & 0 & c_4^{} \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} \\
0 & - c_5^{} \sigma \bigl| \hat{M}_{\sigma}^- \bigr|^2_{} & - c_4^{} \bigl( \hat{M}_{\sigma}^{-} \bigr)^*_{} \hat{M}_{\sigma}^{0} \\
c_4^{} \hat{M}_{\sigma}^{+} \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} & - c_4^{} \hat{M}_{\sigma}^{-} \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} & 0 \\
\end{array}
\right) \label{eq:densqqbarint2}
\end{align}
and
\begin{align}
\overline{ \bigl \langle \rho_{\sigma}^{}(\bar{q}q) \bigr \rangle } \equiv
\biggl( \int^{1-\epsilon}_{0} - \int^{0}_{-(1-\epsilon)} \biggr) d \cos{\theta}\ \rho_{\sigma}^{}(\bar{q}q)
=
\left(
\begin{array}{ccc}
- c_5^{} \sigma \bigl| \hat{M}_{\sigma}^+ \bigr|^2_{} & 0 & c_4^{} \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} \\
0 & c_5^{} \sigma \bigl| \hat{M}_{\sigma}^- \bigr|^2_{} & - c_4^{} \bigl( \hat{M}_{\sigma}^{-} \bigr)^*_{} \hat{M}_{\sigma}^{0} \\
c_4^{} \hat{M}_{\sigma}^{+} \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} & - c_4^{} \hat{M}_{\sigma}^{-} \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} & 0 \\
\end{array}
\right), \label{eq:densqbarqint2}
\end{align}
\end{subequations}
where the constant values $c_i^{}$ $(i=4,5)$ are summarised in Table~\ref{table:coefficients}. Notice again that the sign in front of the elements that have the overall $\sigma$ is different between $\overline{ \langle \rho_{\sigma}^{}(q\bar{q}) \rangle }$ and $\overline{ \langle \rho_{\sigma}^{}(\bar{q}q) \rangle }$. The integration over $y$ and $\cos{\theta}$ in the approaches ${\cal C}$ and ${\cal D}$ proceeds in the same manners as eq.~(\ref{eq:PartdifferentialZH1}) and eq.~(\ref{eq:PartdifferentialZH2}), respectively:
\begin{align}
\frac{d\sigma}{d\hat{s}\ d\cos{\widehat{\theta}}\ d\widehat{\phi} }\biggr|_{{\cal C}}^{}
& = {\cal T} \int_{-y_{\mathrm{cut}}^{}}^{y_{\mathrm{cut}}^{}} dy \biggl[ q(x_1^{}) \bar{q}(x_2^{})\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \overline{ \bigl \langle \rho_{\sigma}^{}(q\bar{q}) \bigr \rangle } d_{}^{\tau}
+
\bar{q}(x_1^{}) q(x_2^{})\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \overline{ \bigl \langle \rho_{\sigma}^{}(\bar{q}q) \bigr \rangle } d_{}^{\tau} \biggr] \nonumber \\
=
{\cal T} \int_{0}^{y_{\mathrm{cut}}^{}} & dy\ 2 \Bigl[ q(x_1^{}) \bar{q}(x_2^{}) + \bar{q}(x_1^{}) q(x_2^{}) \Bigr]
d_{}^{\tau \dagger} \sum_{\sigma}^{}
\left(
\begin{array}{ccc}
0 & 0 & c_4^{} \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} \\
0 & 0 & - c_4^{} \bigl( \hat{M}_{\sigma}^{-} \bigr)^*_{} \hat{M}_{\sigma}^{0} \\
c_4^{} \hat{M}_{\sigma}^{+} \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} & - c_4^{} \hat{M}_{\sigma}^{-} \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} & 0 \\
\end{array}
\right)
d_{}^{\tau},
\label{eq:PartdifferentialZH3}
\end{align}
\begin{align}
\frac{d\sigma}{d\hat{s}\ d\cos{\widehat{\theta}}\ d\widehat{\phi} }\biggr|_{{\cal D}}^{}
& = {\cal T} \biggl( \int_{0}^{y_{\mathrm{cut}}^{}} - \int^{0}_{-y_{\mathrm{cut}}^{}} \biggr)dy \biggl[ q(x_1^{}) \bar{q}(x_2^{})\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \overline{ \bigl \langle \rho_{\sigma}^{}(q\bar{q}) \bigr \rangle } d_{}^{\tau}
+
\bar{q}(x_1^{}) q(x_2^{})\ d_{}^{\tau \dagger} \sum_{\sigma}^{} \overline{ \bigl \langle \rho_{\sigma}^{}(\bar{q}q) \bigr \rangle } d_{}^{\tau} \biggr] \nonumber \\
=
{\cal T} \int_{0}^{y_{\mathrm{cut}}^{}} & dy\ 2 \Bigl[ q(x_1^{}) \bar{q}(x_2^{}) - \bar{q}(x_1^{}) q(x_2^{}) \Bigr]
d_{}^{\tau \dagger} \sum_{\sigma}^{}
\left(
\begin{array}{ccc}
c_5^{} \sigma \bigl| \hat{M}_{\sigma}^+ \bigr|^2_{} & 0 & 0 \\
0 & - c_5^{} \sigma \bigl| \hat{M}_{\sigma}^- \bigr|^2_{} & 0 \\
0 & 0 & 0 \\
\end{array}
\right)
d_{}^{\tau}.\label{eq:PartdifferentialZH4}
\end{align}
Only the elements that do not have the overall $\sigma$ survive in eq.~(\ref{eq:PartdifferentialZH3}), while only the elements that have the overall $\sigma$ survive in eq.~(\ref{eq:PartdifferentialZH4}).\\
The differential cross sections with respect to $\hat{s}$ and the decay angles $\widehat{\theta}$ and $\widehat{\phi}$ have 9 independent angular distributions and can be expressed as
\begin{align}
\frac{d\sigma}{d\hat{s}\ d\cos{\widehat{\theta}}\ d\widehat{\phi} }\biggr|_{i(={\cal A, B, C, D})}^{}
&= F_{i1}^{} \bigl(1+\cos^2_{}{\widehat{\theta}} \bigr)
+ F_{i2}^{} \bigl(1-3\cos^2_{}{\widehat{\theta}} \bigr)
+ F_{i3}^{} \cos{\widehat{\theta}} \nonumber \\
& + F_{i4}^{} \sin{\widehat{\theta}} \cos{\widehat{\phi}}
+ F_{i5}^{} \sin{2\widehat{\theta}} \cos{\widehat{\phi}}
+ F_{i6}^{} \sin^2_{}{\widehat{\theta}} \cos{2\widehat{\phi}} \nonumber \\
& + F_{i7}^{} \sin{\widehat{\theta}} \sin{\widehat{\phi}}
+ F_{i8}^{} \sin{2\widehat{\theta}} \sin{\widehat{\phi}}
+ F_{i9}^{} \sin^2_{}{\widehat{\theta}} \sin{2\widehat{\phi}},\label{eq:diffcrossgeneral}
\end{align}
where the coefficients $F_{ia}^{}$ $(i={\cal A, B, C, D})$ $(a=1,2,\cdots,9)$ are functions of $\hat{s}$ and written in terms of the non-vanishing elements in eqs.~(\ref{eq:PartdifferentialZH1}), (\ref{eq:PartdifferentialZH2}), (\ref{eq:PartdifferentialZH3}) and (\ref{eq:PartdifferentialZH4}). Note that there are in total 36 ($=4 \times 9$) coefficients.
It is straightforward to obtain the explicit form of the coefficients $F_{ia}^{}$ $(i={\cal A, B, C, D})$ $(a=1,2,\cdots,9)$:
\begin{subequations}\label{eq:observables}
\begin{align}
F_{ {\cal A} ({\cal C}) a}^{}
& =
{\cal T} \int_{0}^{y_{\mathrm{cut}}^{}} dy\ 2 \Bigl[ q(x_1^{}) \bar{q}(x_2^{}) + \bar{q}(x_1^{}) q(x_2^{}) \Bigr] \sum_{\sigma} f_{ {\cal A} ({\cal C}) a}^{}, \\
F_{ {\cal B} ({\cal D}) a}^{}
& =
{\cal T} \int_{0}^{y_{\mathrm{cut}}^{}} dy\ 2 \Bigl[ q(x_1^{}) \bar{q}(x_2^{}) - \bar{q}(x_1^{}) q(x_2^{}) \Bigr] \sum_{\sigma} f_{ {\cal B} ({\cal D}) a}^{},
\end{align}
\end{subequations}
where
\begin{subequations}\label{eq:observables-small}
\begin{align}
f_{{\cal A}1}^{} & = \frac{1}{2}\bigl( c_1^{} \bigl| \hat{M}_{\sigma}^+ \bigr|^2_{} + c_1^{} \bigl| \hat{M}_{\sigma}^- \bigr|^2_{} + c_2^{} \bigl| \hat{M}_{\sigma}^0 \bigr|^2_{} \bigr),
& f_{{\cal B}1}^{} & = 0, \nonumber \\
f_{{\cal A}2}^{} & = \frac{1}{2} c_2^{} \bigl| \hat{M}_{\sigma}^0 \bigr|^2_{},
& f_{{\cal B}2}^{} & = 0, \nonumber \\
f_{{\cal A}3}^{} & = c_1^{} \bigl( \bigl| \hat{M}_{\sigma}^+ \bigr|^2_{} - \bigl| \hat{M}_{\sigma}^- \bigr|^2_{} \bigr)\tau,
& f_{{\cal B}3}^{} & = 0, \nonumber \\
f_{{\cal A}4}^{} & = 0,
& f_{{\cal B}4}^{} & = \sqrt{2} \sigma c_3^{} {\it Re} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} + \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr] \tau, \nonumber \\
f_{{\cal A}5}^{} & = 0,
& f_{{\cal B}5}^{} & = \frac{1}{\sqrt{2}} \sigma c_3^{} {\it Re} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} - \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr] , \nonumber \\
f_{{\cal A}6}^{} & = \frac{1}{2} c_2^{} {\it Re} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr] ,
& f_{{\cal B}6}^{} & = 0, \nonumber \\
f_{{\cal A}7}^{} & = 0,
& f_{{\cal B}7}^{} & = \sqrt{2} \sigma c_3^{} {\it Im} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} + \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr] \tau, \nonumber \\
f_{{\cal A}8}^{} & = 0,
& f_{{\cal B}8}^{} & = \frac{1}{\sqrt{2}} \sigma c_3^{} {\it Im} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} - \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr] , \nonumber \\
f_{{\cal A}9}^{} & = \frac{1}{2} c_2^{} {\it Im} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr] ,
& f_{{\cal B}9}^{} & = 0,
\end{align}
and
\begin{align}
f_{{\cal C}1}^{} & = 0,
& f_{{\cal D}1}^{} & = \frac{1}{2} \sigma c_5^{} \bigl( \bigl| \hat{M}_{\sigma}^+ \bigr|^2_{} - \bigl| \hat{M}_{\sigma}^- \bigr|^2_{} \bigr), \nonumber \\
f_{{\cal C}2}^{} & = 0,
& f_{{\cal D}2}^{} & = 0, \nonumber \\
f_{{\cal C}3}^{} & = 0,
& f_{{\cal D}3}^{} & = \sigma c_5^{} \bigl( \bigl| \hat{M}_{\sigma}^+ \bigr|^2_{} + \bigl| \hat{M}_{\sigma}^- \bigr|^2_{} \bigr) \tau, \nonumber \\
f_{{\cal C}4}^{} & = \sqrt{2} c_4^{} {\it Re} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} - \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr] \tau,
& f_{{\cal D}4}^{} & = 0, \nonumber \\
f_{{\cal C}5}^{} & = \frac{1}{\sqrt{2}} c_4^{} {\it Re} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} + \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr],
& f_{{\cal D}5}^{} & = 0, \nonumber \\
f_{{\cal C}6}^{} & = 0,
& f_{{\cal D}6}^{} & = 0, \nonumber \\
f_{{\cal C}7}^{} & = \sqrt{2} c_4^{} {\it Im} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} - \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr] \tau,
& f_{{\cal D}7}^{} & = 0, \nonumber \\
f_{{\cal C}8}^{} & = \frac{1}{\sqrt{2}} c_4^{} {\it Im} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} + \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr],
& f_{{\cal D}8}^{} & = 0, \nonumber \\
f_{{\cal C}9}^{} & = 0,
& f_{{\cal D}9}^{} & = 0.
\end{align}
\end{subequations}
Among the 36 coefficients, only the 15 coefficients can be non-zero. It is easy to notice that there are the 10 combinations of the elements of the density matrix in total. However, only 9 of them are independent~\footnote{Among the 3 combinations in $f_{{\cal A}1}^{}$, $f_{{\cal A}2}^{}$ and $f_{{\cal D}3}^{}$, only 2 of them are independent, since one of them can be constructed from the other two.}.
Some of them are strictly zero, if the amplitudes satisfy the restriction in eq.~(\ref{eq:CPinvZ-2}) from CP invariance and/or the restriction in eq.~(\ref{eq:CPTinvZ-2}) from $\mathrm{CP\widetilde{T}}$ invariance. These symmetry properties can be explicitly checked in the following way.
By applying the restrictions in eqs.~(\ref{eq:CPinvZ-2}) and (\ref{eq:CPTinvZ-2}) to the combination in $f_{{\cal A}9}^{}$, we find
\begin{align}
\mathrm{CP}\ \mathrm{invariance}:\ & {\it Im} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr]
= {\it Im} \bigl[ | \hat{M}_{\sigma}^{+} |^2_{} \bigr]
= 0, \nonumber \\
\mathrm{CP\widetilde{T}}\ \mathrm{invariance}:\ & {\it Im} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr]
= {\it Im} \bigl[ ( \hat{M}_{\sigma}^{-} )^2_{} \bigr]
\ne 0,
\end{align}
where the former means that CP invariance requires it to be zero, while the latter means that $\mathrm{CP\widetilde{T}}$ invariance does not require it to be zero. This indicates that observation of a nonzero value in $F_{ {\cal A}9}^{}$ signals CP violation. Similarly, by applying the restrictions in eqs.~(\ref{eq:CPinvZ-2}) and (\ref{eq:CPTinvZ-2}) to the combination in $f_{{\cal B}8}^{}$ and $f_{{\cal C}7}^{}$, we find
\begin{align}
\mathrm{CP}\ \mathrm{invariance}:\ & {\it Im} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} - \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr]
= {\it Im} \bigl\{ 2 i {\it Im}\bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} \bigr] \bigr\} \ne 0,\nonumber \\
\mathrm{CP\widetilde{T}}\ \mathrm{invariance}:\ & {\it Im} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} - \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr]
= {\it Im} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} - \hat{M}_{\sigma}^{0} \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \bigr] = 0,
\end{align}
where $\mathrm{CP\widetilde{T}}$ invariance requires it to be zero, while CP invariance does not. This indicates that observation of a nonzero value in $F_{ {\cal B}8}^{}$ or $F_{ {\cal C}7}^{}$ signals $\mathrm{CP\widetilde{T}}$ violation.
Finally, by applying the restrictions in eqs.~(\ref{eq:CPinvZ-2}) and (\ref{eq:CPTinvZ-2}) to the combination in $f_{{\cal B}5}^{}$ and $f_{{\cal C}4}^{}$, we find
\begin{align}
\mathrm{CP}\ \mathrm{invariance}:\ & {\it Re} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} - \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr]
= {\it Re} \bigl\{ 2 i {\it Im}\bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} \bigr] \bigr\} = 0,\nonumber \\
\mathrm{CP\widetilde{T}}\ \mathrm{invariance}:\ & {\it Re} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} - \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr]
= {\it Re} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} - \hat{M}_{\sigma}^{0} \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \bigr] = 0,
\end{align}
where both CP invariance and $\mathrm{CP\widetilde{T}}$ invariance require it to be zero. This indicates that observation of a nonzero value in $F_{ {\cal B}5}^{}$ or $F_{ {\cal C}4}^{}$ signals both CP violation {\it and} $\mathrm{CP\widetilde{T}}$ violation. These symmetry properties of the 10 combinations are summarised in Table~\ref{table:symproperty}. The symbol $+$ means that the symmetry does not require the combination to be zero, while $-$ means that the symmetry requires the combination to be zero.
We also show the coefficients $F_{ia}^{}$ by measuring which the combinations of the density matrix elements can be determined.
The symbol $\circ$ in the column "$f$ charge" means that observation of the coefficient $F_{ia}^{}$ requires distinguishing the fermion $f$ from the antifermion $\bar{f}$. In other words, it requires identification of the charge (flavor) of the final fermion $f$.
This can be confirmed by performing a translation $\widehat{\theta} \to \pi - \widehat{\theta}$ {\it and} $\widehat{\phi} \to \widehat{\phi} + \pi$ in eq.~(\ref{eq:diffcrossgeneral}) and observing the change of the sign.
Among the 15 coefficients, only 9 of them do not require the charge identification of the final fermion $f$. It should be emphasised that these 9 coefficients are necessary and sufficient to determine all of the 9 independent combinations of the density matrix elements. \\
\begin{table}
\begin{tabular}{@{}l cccccc@{}}
\toprule
\addlinespace[1mm]
Combinations of & \multicolumn{2}{l}{Symmetry properties} & Observables & $f$ charge & $HZ\gamma$ & d.o.p \\
\cmidrule(l){2-3}
the density matrix elements & CP & $\mathrm{CP\widetilde{T}}$ & & & \\
\midrule
\addlinespace[1mm]
$c_1^{} \bigl| \hat{M}_{\sigma}^+ \bigr|^2_{} + c_1^{} \bigl| \hat{M}_{\sigma}^- \bigr|^2_{} + c_2^{} \bigl| \hat{M}_{\sigma}^0 \bigr|^2_{}$ & $+$ & $+$ & $F_{ {\cal A} 1}^{}$ & - & - & -\\
$\bigl| \hat{M}_{\sigma}^0 \bigr|^2_{}$ & $+$ & $+$ & $F_{ {\cal A} 2}^{}$ & - & - & -\\
$\bigl| \hat{M}_{\sigma}^+ \bigr|^2_{} + \bigl| \hat{M}_{\sigma}^- \bigr|^2_{}$ & $+$ & $+$ & $F_{ {\cal D} 3}^{}$ & $\circ$ & $\circ$ & $\circ$ \\
$\bigl| \hat{M}_{\sigma}^+ \bigr|^2_{} - \bigl| \hat{M}_{\sigma}^- \bigr|^2_{}$ & $-$ & $-$ & $F_{ {\cal A} 3}^{}$ & $\circ$ & - & -\\
& & & $F_{ {\cal D} 1}^{}$ & - & $\circ$ & $\circ$ \\
${\it Re} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} + \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr]$ & $+$ & $+$ & $F_{ {\cal B} 4}^{}$ & $\circ$ & $\circ$ & $\circ$\\
& & & $F_{ {\cal C} 5}^{}$ & - & - & -\\
${\it Re} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} - \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr]$ & $-$ & $-$ & $F_{ {\cal B} 5}^{}$ & - & $\circ$ & $\circ$ \\
& & & $F_{ {\cal C} 4}^{}$ & $\circ$ & - & -\\
${\it Re} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr]$ & $+$ & $+$ & $F_{ {\cal A} 6}^{}$ & - & - & - \\
${\it Im} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} + \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr]$ & $-$ & $+$ & $F_{ {\cal B} 7}^{}$ & $\circ$ & $\circ$ & $\circ$\\
& & & $F_{ {\cal C} 8}^{}$ & - & - & - \\
${\it Im} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{0} - \bigl( \hat{M}_{\sigma}^{0} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr]$ & $+$ & $-$ & $F_{ {\cal B} 8}^{}$ & - & $\circ$ & $\circ$\\
& & & $F_{ {\cal C} 7}^{}$ & $\circ$ & - & - \\
${\it Im} \bigl[ \bigl( \hat{M}_{\sigma}^{+} \bigr)^*_{} \hat{M}_{\sigma}^{-} \bigr]$ & $-$ & $+$ & $F_{ {\cal A} 9}^{}$ & - & - & -\\
\addlinespace[1mm]
\bottomrule
\end{tabular}
\caption{\small Symmetry properties of the 10 combinations of the density matrix elements. The coefficients $F_{ia}^{}$ of the differential angular distributions, by measuring which the combinations of the density matrix elements can be determined, are also shown.
The symbol $-$ means that the symmetry (CP or $\mathrm{CP\widetilde{T}}$) requires the combination to be zero, while the symbol $+$ means that the symmetry does not; observation of a non-zero value in the combination with the symbol $-$ under CP, for instance, signals CP violation.
The symbol $\circ$ in the column "$f$ charge" means that observation of the coefficient $F_{ia}^{}$ requires the charge (or flavor) identification of the final fermion $f$.
The symbol $\circ$ in the column "$HZ\gamma$" indicates that the coefficient $F_{ia}^{}$ has a good sensitivity to the $HZ\gamma$ coupling; see a discussion at the paragraph of eq.~(\ref{eq:amp:simplified}).
The symbol $\circ$ in the last column indicates that the coefficient $F_{ia}^{}$ is weakened according to the degree of polarisation of the $Z$ boson; see a discussion at the last paragraph of Section~\ref{sec:angcoeff-W}.
By the simple replacements $\hat{M}^{\lambda}_{\sigma} \to \hat{N}^{\lambda}_{}$ and $F_{ia}^{} \to F_{ia}^{W}$, the table corresponding to the process $pp\to W^{\pm}_{}H$ is immediately obtained; see Section~\ref{sec:polWboson}.}
\label{table:symproperty}
\end{table}
The coefficients $F_{ia}^{}$ $(i={\cal A, B, C, D})$ $(a=1,2,\cdots,9)$ will be experimentally determined by measuring the decay angles $\widehat{\theta}$ and $\widehat{\phi}$ in the rest frame of the $Z$ boson (see eq.~(\ref{eq:decaykinematics})), since we completely know the differential angular distributions as eq.~(\ref{eq:diffcrossgeneral}). We just do not know the coefficients $F_{ia}^{}$ which uniquely depend on a state of polarisation of the $Z$ boson.
With appropriate integration over $\cos{\widehat{\theta}}$ and $\widehat{\phi}$, it is possible to isolate the angular distributions:
\begin{subequations}\label{eq:cosIntall}
\begin{align}
\int^1_{-1} d\cos{\widehat{\theta}} \frac{d\sigma}{d\hat{s} d\cos{\widehat{\theta}} d\widehat{\phi} } & = \frac{8}{3}F_{1}^{}
+ \frac{\pi}{2} F_{4}^{} \cos{\widehat{\phi}}
+ \frac{4}{3} F_{6}^{} \cos{2\widehat{\phi}}
+ \frac{\pi}{2} F_{7}^{} \sin{\widehat{\phi}}
+ \frac{4}{3} F_{9}^{} \sin{2\widehat{\phi}}, \label{eq:cosInt21}\\
\biggl( \int^1_{0} - \int_{-1}^0 \biggr) d\cos{\widehat{\theta}} \frac{d\sigma}{d\hat{s} d\cos{\widehat{\theta}} d\widehat{\phi} } & = F_{3}^{}
+ \frac{4}{3} F_{5}^{} \cos{\widehat{\phi}}
+ \frac{4}{3} F_{8}^{} \sin{\widehat{\phi}}, \label{eq:cosInt22}
\end{align}
\end{subequations}
and
\begin{subequations}\label{eq:phiIntall}
\begin{align}
\frac{1}{2\pi}\int^{2\pi}_{0} d\widehat{\phi} \frac{d\sigma}{d\hat{s} d\cos{\widehat{\theta}} d\widehat{\phi} } & = F_{1}^{} \bigl(1+\cos^2_{}{\widehat{\theta}} \bigr)
+ F_{2}^{} \bigl(1-3\cos^2_{}{\widehat{\theta}} \bigr)
+ F_{3}^{} \cos{\widehat{\theta}}, \label{eq:phiInt1} \\
\frac{1}{4} \biggl( \int^{\pi/2}_{0} - \int^{\pi}_{\pi/2} - \int^{3\pi/2}_{\pi} + \int_{3\pi/2}^{2\pi} \biggr) d\widehat{\phi} \frac{d\sigma}{d\hat{s} d\cos{\widehat{\theta}} d\widehat{\phi} } & =
F_{4}^{} \sin{\widehat{\theta}} +
F_{5}^{} \sin{2\widehat{\theta}}, \label{eq:phiInt2} \\
\frac{1}{4} \biggl( \int^{\pi/4}_{0} - \int^{\pi/2}_{\pi/4} - \int^{3\pi/4}_{\pi/2} + \int_{3\pi/4}^{\pi} + \int_{\pi}^{5\pi/4} - \int_{5\pi/4}^{3\pi/2} & - \int_{3\pi/2}^{7\pi/4} + \int_{7\pi/4}^{2\pi} \biggr) d\widehat{\phi} \frac{d\sigma}{d\hat{s} d\cos{\widehat{\theta}} d\widehat{\phi} } =
F_{6}^{} \sin^2_{}{\widehat{\theta}}, \label{eq:phiInt3}\\
\frac{1}{4} \biggl( \int^{\pi}_{0} - \int_{\pi}^{2\pi} \biggr) d\widehat{\phi} \frac{d\sigma}{d\hat{s} d\cos{\widehat{\theta}} d\widehat{\phi} } & =
F_{7}^{} \sin{\widehat{\theta}} +
F_{8}^{} \sin{2\widehat{\theta}}, \\
\frac{1}{4} \biggl( \int^{\pi/2}_{0} - \int^{\pi}_{\pi/2} + \int^{3\pi/2}_{\pi} - \int_{3\pi/2}^{2\pi} \biggr) d\widehat{\phi} \frac{d\sigma}{d\hat{s} d\cos{\widehat{\theta}} d\widehat{\phi} } & =
F_{9}^{} \sin^2_{}{\widehat{\theta}}. \label{eq:phiInt5}
\end{align}
\end{subequations}
Here the index $i(={\cal A, B, C, D})$ is omitted. By combining the 2 approaches in eq.~(\ref{eq:cosIntall}) and the 5 approaches in eq.~(\ref{eq:phiIntall}), we obtain the $10$ ($=2\times 5$) combinations. The 2 of them simply give zero (i.e. eqs.~(\ref{eq:cosInt22}) and (\ref{eq:phiInt3}), and eqs.~(\ref{eq:cosInt22}) and (\ref{eq:phiInt5})). Each of the remaining 8 combinations gives one of $F_a^{}$ $(a=1,3,4,5,6,7,8,9)$. For example, eqs.~(\ref{eq:cosInt22}) and (\ref{eq:phiInt2}) gives $F_5^{}$. Only $F_2^{}$ is not determined in this method. By a fitting procedure in the differential cross section with respect to $\hat{s}$ and $\cos{\widehat{\theta}}$ in eq.~(\ref{eq:phiInt1}), $F_2^{}$ may be determined.
\begin{figure}[t]
\centering
\includegraphics[scale=0.42]{fig1.pdf}
\caption{\small
The coefficients $-F_{{\cal A} 9}^{}$ (left panel), $F_{{\cal C} 8}^{}$ (middle panel) and $F_{{\cal B} 7}^{}$ (right panel) divided by $F_{{\cal A} 1}^{}$ are shown. These coefficients are constrained to be identically zero by CP invariance. In each panel, $\Box$ points give predictions for $\widetilde{h}_4^Z=0.1$, and $\times$ points give those for $\widetilde{h}_4^{\gamma}=0.1$. For each result, the values in the 5 bins, from left to right, are obtained after integration over $\hat{s}$ in the regions $(m_Z^{}+m_H^{}) < \hat{s}^{1/2}_{} < 300$, $300 < \hat{s}^{1/2}_{} < 400$, $400 < \hat{s}^{1/2}_{} < 600$, $600 < \hat{s}^{1/2}_{} < 1000$ and $1000 < \hat{s}^{1/2}_{} < 14000$ in units of GeV.}
\label{figure:ZHplot1}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.42]{fig2.pdf}
\caption{\small
The coefficients $F_{{\cal B} 8}^{}$ (left panel) and $F_{{\cal C} 7}^{}$ (right panel) divided by $F_{{\cal A} 1}^{}$ are shown in the same manner as Figure~\ref{figure:ZHplot1}. These coefficients are constrained to be identically zero by $\mathrm{CP\widetilde{T}}$ invariance. In each panel, $\Box$ points give predictions for $h_3^Z=0+0.1i$, and $\times$ points give those for $h_3^{\gamma}=0+0.1i$.}
\label{figure:ZHplot2}
\end{figure}
\begin{figure}[t]
\centering
\includegraphics[scale=0.35]{fig3.pdf}
\caption{\small
The angular coefficients $F_{{\cal A} 3}^{}$ (left panel), $F_{{\cal D} 1}^{}$ (middle-left panel), $-F_{{\cal B} 5}^{}$ (middle-right panel) and $-F_{{\cal C} 4}^{}$ (right panel) divided by $F_{{\cal A} 1}^{}$ are shown in the same manner as Figure~\ref{figure:ZHplot1}. These coefficients are doubly constrained to be identically zero by CP invariance {\it and} $\mathrm{CP\widetilde{T}}$ invariance. In each panel, $\Box$ points give predictions for $\widetilde{h}_4^Z=0+0.1i$ and $\times$ points give those for $\widetilde{h}_4^{\gamma}=0+0.1i$.}
\label{figure:ZHplot3}
\end{figure}
\subsection{Influences of non-standard $HZZ$ and $HZ\gamma$ interactions}\label{sec:angcoeff}
In the previous section, we have clarified the restrictions on the coefficients of the differential angular distributions imposed by the CP and $\mathrm{CP\widetilde{T}}$ symmetries; see Table~\ref{table:symproperty}.
Some of the coefficients are strictly zero in the SM due to CP invariance and some of them are small in the SM due to the smallness of re-scattering effects. These coefficients are in particular interesting as observables at the LHC, since observation of a non-zero or large value in these coefficients immediately signals the existence of physics beyond the SM. In this Section, we focus on these coefficients and study the influences of the non-standard $HZZ$ and $HZ\gamma$ couplings.\\
Our numerical results are produced for the $14$ TeV LHC. We set $m_H^{}=125.5$ GeV, and $\epsilon =0$ in the $\cos{\theta}$ integration; see eqs.~(\ref{eq:cosInt1}) and (\ref{eq:cosInt2}), and Table~\ref{table:coefficients}.
As the final fermion flavor $f$ summed in eq.~(\ref{eq:simplywriting}), only the electron and the muon are considered when calculating the coefficients $F_{ia}^{}$ $(i={\cal A, B, C, D})$ $(a=3,4,7)$ (Recall that observation of these coefficients requires the charge identification of $f$), and all the quark flavors but the top quark are additionally considered when calculating the other coefficients.
MSTW PDFs~\cite{Martin:2009iq} are used. The phase space integration is performed with the program BASES~\cite{Kawabata:1995th}. \\
In Figure~\ref{figure:ZHplot1}, the coefficients $-F_{{\cal A} 9}^{}$ (left panel), $F_{{\cal C} 8}^{}$ (middle panel) and $F_{{\cal B} 7}^{}$ (right panel) divided by $F_{{\cal A} 1}^{}$ are shown~\footnote{When calculating $F_{{\cal A} 1}^{}$, by which $F_{ia}^{}$ $(i={\cal A, B, C, D})$ $(a=3,4,7)$ are divided, we consider only the electron and the muon as the final fermion flavor $f$ summed in eq.~(\ref{eq:simplywriting}).}. For each result, the values in the 5 bins, from left to right, are obtained after integration over $\hat{s}$ in the regions $(m_Z^{}+m_H^{}) < \hat{s}^{1/2}_{} < 300$, $300 < \hat{s}^{1/2}_{} < 400$, $400 < \hat{s}^{1/2}_{} < 600$, $600 < \hat{s}^{1/2}_{} < 1000$ and $1000 < \hat{s}^{1/2}_{} < 14000$ in units of GeV.
In each panel, $\Box$ points give predictions for the CP-odd form factor $\widetilde{h}_4^Z=0.1$, and $\times$ points give those for the CP-odd form factor $\widetilde{h}_4^{\gamma}=0.1$. CP invariance requires these 4 coefficients to be identically zero, thus observation of a non-zero value signals CP violation. \\
The coefficients $F_{{\cal C} 8}^{}$ and $F_{{\cal B} 7}^{}$ measure the same combination of the density matrix elements (see Table~\ref{table:symproperty}), thus show the same dependence on $\hat{s}^{1/2}_{}$.
However, $F_{{\cal C} 8}^{}$'s sensitivity to $\widetilde{h}_4^Z$ is stronger than its sensitivity to $\widetilde{h}_4^{\gamma}$, while $F_{{\cal B} 7}^{}$'s sensitivity to $\widetilde{h}_4^Z$ is comparable to its sensitivity to $\widetilde{h}_4^{\gamma}$.
This difference originates from the overall $\sigma$ in $f_{{\cal B} 7}^{}$ and is explained as follows. Eq.~(\ref{eq:Zsubamplitudes}) shows that $\hat{M}^{\lambda}_+$ and $\hat{M}^{\lambda}_-$ can be written in the following simplified form as the sum of the $Z$ boson contribution and the $\gamma$ contribution:
\begin{subequations}\label{eq:amp:simplified}
\begin{align}
\hat{M}^{\lambda}_+ = & g_{Z+}^{} f_Z^{\lambda} + f_{\gamma}^{\lambda},\\
\hat{M}^{\lambda}_- = & g_{Z-}^{} f_Z^{\lambda} + f_{\gamma}^{\lambda}.
\end{align}
\end{subequations}
By using this expression, we derive
\begin{subequations}
\begin{align}
(\hat{M}^{\lambda^{\prime}_{}}_+)^*_{} \times \hat{M}^{\lambda}_+
=
g_{Z+}^2 (f_Z^{\lambda^{\prime}_{}})^*_{} f_Z^{\lambda}
+
g_{Z+}^{} \bigl[ (f_Z^{\lambda^{\prime}_{}})^*_{} f_{\gamma}^{\lambda} + (f_{\gamma}^{\lambda^{\prime}_{}})^*_{} f_{Z}^{\lambda} \bigr]
+
(f_{\gamma}^{\lambda^{\prime}_{}})^*_{} f_{\gamma}^{\lambda}, \label{eq:interfer1} \\
(\hat{M}^{\lambda^{\prime}_{}}_-)^*_{} \times \hat{M}^{\lambda}_-
=
g_{Z-}^2 (f_Z^{\lambda^{\prime}_{}})^*_{} f_Z^{\lambda}
+
g_{Z-}^{} \bigl[ (f_Z^{\lambda^{\prime}_{}})^*_{} f_{\gamma}^{\lambda} + (f_{\gamma}^{\lambda^{\prime}_{}})^*_{} f_{Z}^{\lambda} \bigr]
+
(f_{\gamma}^{\lambda^{\prime}_{}})^*_{} f_{\gamma}^{\lambda}. \label{eq:interfer2}
\end{align}
\end{subequations}
The couplings $g_{Z+}^{}$ and $g_{Z-}^{}$ have opposite signs to each other, both for the up-type quarks and for the down-type quarks.
The interference term between the $Z$ boson contribution and the $\gamma$ contribution in eq.~(\ref{eq:interfer1}) and that in eq.~(\ref{eq:interfer2}), therefore, have opposite signs to each other, too. Considering $f_{\gamma}^{\lambda}=0$ at the tree level approximation in the SM, the interference terms may give a larger contribution than the $(f_{\gamma}^{\lambda^{\prime}_{}})^*_{} f_{\gamma}^{\lambda}$ term.
The coefficient $F_{{\cal C} 8}^{}$ has a structure that eq.~(\ref{eq:interfer1}) and eq.~(\ref{eq:interfer2}) are added and the interference terms tend to cancel to each other. On the other hand, $F_{{\cal B} 7}^{}$ has a structure that eq.~(\ref{eq:interfer2}) and eq.~(\ref{eq:interfer1}) are subtracted due to the overall $\sigma$ in $f_{{\cal B} 7}^{}$, and the interference terms contribute in the same direction. Therefore, as long as $f_{\gamma}^{\lambda}$ (i.e the $HZ\gamma$ coupling) is small, the effects of it appear larger in $F_{{\cal B} 7}^{}$ than in $F_{{\cal C} 8}^{}$.
Notice that this discussion is independent on explicit choices for $\lambda$ and $\lambda^{\prime}_{}$. The above finding is true not only in $F_{{\cal B} 7}^{}$ but also in the other coefficients which have the overall $\sigma$ in $f_{{i}a}^{}$. These coefficients are indicated by the symbol $\circ$ in the column "$HZ\gamma$" in Table~\ref{table:symproperty}. \\
In Figure~\ref{figure:ZHplot2}, the coefficients $F_{{\cal B} 8}^{}$ (left panel) and $F_{{\cal C} 7}^{}$ (right panel) divided by $F_{{\cal A} 1}^{}$ are shown in the same manner as Figure~\ref{figure:ZHplot1}.
$\mathrm{CP\widetilde{T}}$ invariance requires these 2 coefficients to be identically zero, hence observation of a non-zero value indicates the existence of re-scattering effects.
Re-scattering effects can be approximately included by allowing imaginary parts in the form factors~\cite{Hagiwara:1986vm}.
In each panel, $\Box$ points give predictions for $h_3^Z=0+0.1i$, and $\times$ points give those for $h_3^{\gamma}=0+0.1i$. \\
In Figure~~\ref{figure:ZHplot3}, the coefficients $F_{{\cal A} 3}^{}$ (left panel), $F_{{\cal D} 1}^{}$ (middle-left panel), $-F_{{\cal B} 5}^{}$ (middle-right panel) and $-F_{{\cal C} 4}^{}$ (right panel) divided by $F_{{\cal A} 1}^{}$ are shown in the same manner as Figure~\ref{figure:ZHplot1}. These 4 coefficients are doubly constrained to be identically zero by CP invariance {\it and} $\mathrm{CP\widetilde{T}}$ invariance: they are strictly zero if CP is conserved, even if $\mathrm{CP\widetilde{T}}$ is violated, for instance. Observation of a non-zero value indicates CP violation {\it and} the existence of re-scattering effects. In each panel, $\Box$ points give predictions for the CP-odd from factor $\widetilde{h}_4^Z=0+0.1i$, and $\times$ points give those for the CP-odd form factor $\widetilde{h}_4^{\gamma}=0+0.1i$.
\section{Polarisation of the $W$ boson}\label{sec:polWboson}
In this section, we analyse the density matrices of the $W^{+}_{}$ and $W^{-}_{}$ in the process $pp \to W^{\pm}_{}H$ following the same procedure as Section~\ref{sec:densitymatrix}.
Because (1) a determination of the density matrix of the $W$ boson is difficult in $pp$ collisions when the $W$ boson decays into a charged lepton and a neutrino and (2) we cannot distinguish $W^+_{}$ from $W^-_{}$ in view of the difficulty of flavor identification of both $W^+_{}$ and $W^-_{}$ decay products when the $W$ boson decays into two quarks, we consider the process $pp \to W^{\pm}_{}H$ ($W^{\pm}_{} \to jj$) as the sum of the process $pp \to W^+_{}H$ ($W^+_{} \to jj$) and the process $pp \to W^-_{}H$ ($W^-_{} \to jj$).
Just as the process $pp\to ZH$, only the $15$ coefficients among the $36$ coefficients of the 4 different differential angular distributions of the dijets can be non-zero. However only 9 coefficients among these $15$ coefficients are actually measurable.
\subsection{Polarisation density matrices of the $W^+_{}$ and $W^-_{}$}\label{sec:densitymatrixW}
First of all, we define the density matrices of the $W^+_{}$ and $W^-_{}$. We consider the sub-processes
\begin{subequations}
\begin{align}
u(\sigma) + \bar{d}( -\sigma ) & \to W^+_{}(\lambda) + H; \nonumber \\
W^+_{}(\lambda) & \to u(\tau) + \bar{d}(-\tau),
\end{align}
and
\begin{align}
d(\sigma) + \bar{u}( -\sigma ) & \to W^-_{}(\lambda) + H; \nonumber \\
W^-_{}(\lambda) & \to d(\tau) + \bar{u}(-\tau),
\end{align}
\end{subequations}
where the helicity of each particle is shown in parenthesis. We neglect the masses of the quark and the antiquark from the $W$ decay, hence the helicity of the antiquark is always opposite to that of the quark.
We prepare the 4 full helicity amplitudes ${\cal T}(u\bar{d})$, ${\cal T}(d\bar{u})$, ${\cal T}(\bar{d}u)$ and ${\cal T}(\bar{u}d)$ in which the production amplitude is evaluated in the $u\bar{d}$ c.m. frame, the $d\bar{u}$ c.m. frame, the $\bar{d}u$ c.m. frame and the $\bar{u}d$ c.m. frame, respectively.
Recall that these c.m. frames are shown in Figure~\ref{figure:frames} and the production amplitudes are given in eqs.~(\ref{eq:WpampAll}) and (\ref{eq:WmampAll}).
These 4 full helicity amplitudes are given by
\begin{subequations}
\begin{align}
{\cal T}(u\bar{d}) & = P_W^{} \sum_{\lambda=\pm,0}^{} {\cal M}_{-}^{\lambda}(u \bar{d})\ D_{\lambda}^{},\label{eq:fullampW1} \\
{\cal T}(\bar{d}u) & = P_W^{} \sum_{\lambda=\pm,0}^{} {\cal M}_{-}^{\lambda}(\bar{d} u)\ D_{\lambda}^{},\label{eq:fullampW2} \\
{\cal T}(d\bar{u}) & = P_W^{} \sum_{\lambda=\pm,0}^{} {\cal M}_{-}^{\lambda}(d \bar{u})\ \widetilde{D}_{\lambda}^{},\label{eq:fullampW3} \\
{\cal T}(\bar{u}d) & = P_W^{} \sum_{\lambda=\pm,0}^{} {\cal M}_{-}^{\lambda}(\bar{u}d)\ \widetilde{D}_{\lambda}^{},\label{eq:fullampW4}
\end{align}
\end{subequations}
where
\begin{align}
P_W^{} = ( Q^2_{} - m_W^2 + i m_W^{} \Gamma_W^{} )^{-1}_{}
\end{align}
denotes the propagator factor of the $W$ boson, $D^{}_{\lambda}$ is the helicity amplitude for the decay process $W^+_{} \to u \bar{d}$ and $\widetilde{D}^{}_{\lambda}$ is the helicity amplitude for the decay process $W^-_{} \to d \bar{u}$:
\begin{subequations}
\begin{align}
D^{}_{\lambda} & = \frac{g}{\sqrt{2}} V_{ud}^{} m_W^{}\ d^-_{\lambda}, \\
\widetilde{D}^{}_{\lambda} & = \frac{g}{\sqrt{2}} (V_{ud}^{})^*_{} m_W^{}\ d^-_{\lambda},
\end{align}
\end{subequations}
where $d^-_{\lambda}$ is given in eq.~(\ref{eq:decayamp}).
The decay amplitudes are evaluated in the following four-momentum frame:
\begin{align}
W^+_{}\ (W^-_{}):&\ \ \bigl( m_W^{},0, 0, 0 \bigr) \nonumber \\
u \ \ (d):&\ \ \frac{m_W^{}}{2}\bigl( 1,\ \sin{\widehat{\theta}} \cos{\widehat{\phi}},\ \sin{\widehat{\theta}} \sin{\widehat{\phi}},\ \cos{\widehat{\theta}} \bigr) \nonumber \\
\bar{d}\ \ (\bar{u}) :&\ \ \frac{m_W^{}}{2}\bigl( 1,\ -\sin{\widehat{\theta}} \cos{\widehat{\phi}},\ -\sin{\widehat{\theta}} \sin{\widehat{\phi}},\ -\cos{\widehat{\theta}} \bigr). \label{eq:decaykinematics-W}
\end{align}
Then we obtain
\begin{subequations}
\begin{align}
\bigl| {\cal T}(u\bar{d}) \bigr|^2_{}
& = \bigl|P_W^{} m_W^{} g V_{ud}^{} / \sqrt{2} \bigr|^2_{}
\sum_{\lambda^{\prime}_{}, \lambda} \bigl( d_{\lambda^{\prime}}^{-} \bigr)^*_{} \rho_{}^{\lambda^{\prime} \lambda}(u\bar{d})\ d_{\lambda}^{-}, \label{eq:ampsquaredW1} \\
\bigl| {\cal T}(\bar{d}u) \bigr|^2_{}
& = \bigl|P_W^{} m_W^{} g V_{ud}^{} / \sqrt{2} \bigr|^2_{}
\sum_{\lambda^{\prime}_{}, \lambda} \bigl( d_{\lambda^{\prime}}^{-} \bigr)^*_{} \rho_{}^{\lambda^{\prime} \lambda}(\bar{d}u)\ d_{\lambda}^{-}, \label{eq:ampsquaredW2} \\
\bigl| {\cal T}(d\bar{u}) \bigr|^2_{}
& = \bigl|P_W^{} m_W^{} g V_{ud}^{} / \sqrt{2} \bigr|^2_{}
\sum_{\lambda^{\prime}_{}, \lambda} \bigl( d_{\lambda^{\prime}}^{-} \bigr)^*_{} \rho_{}^{\lambda^{\prime} \lambda}(d\bar{u})\ d_{\lambda}^{-}, \label{eq:ampsquaredW3} \\
\bigl| {\cal T}(\bar{u}d) \bigr|^2_{}
& = \bigl|P_W^{} m_W^{} g V_{ud}^{} / \sqrt{2} \bigr|^2_{}
\sum_{\lambda^{\prime}_{}, \lambda} \bigl( d_{\lambda^{\prime}}^{-} \bigr)^*_{} \rho_{}^{\lambda^{\prime} \lambda}(\bar{u}d)\ d_{\lambda}^{-}, \label{eq:ampsquaredW4}
\end{align}
\end{subequations}
where
\begin{subequations}\label{eq:densityMW12}
\begin{align}
\rho_{}^{\lambda^{\prime} \lambda}(u\bar{d}) & =
\bigl\{ {\cal M}_{-}^{\lambda^{\prime}_{}}(u \bar{d}) \bigr\}^{\ast}_{}
{\cal M}_{-}^{\lambda}(u \bar{d}),\label{eq:densityMW1} \\
\rho_{}^{\lambda^{\prime} \lambda}(\bar{d}u) & =
\bigl\{ {\cal M}_{-}^{\lambda^{\prime}_{}}(\bar{d}u) \bigr\}^{\ast}_{}
{\cal M}_{-}^{\lambda}(\bar{d}u),\label{eq:densityMW2}
\end{align}
\end{subequations}
represent the elements of the density matrix in the helicity basis of the $W^+_{}$ in the $u\bar{d}$ c.m. frame and those in the $\bar{d}u$ c.m. frame, respectively, and
\begin{subequations}\label{eq:densityMW34}
\begin{align}
\rho_{}^{\lambda^{\prime} \lambda}(d \bar{u}) & =
\bigl\{ {\cal M}_{-}^{\lambda^{\prime}_{}}(d \bar{u}) \bigr\}^{\ast}_{}
{\cal M}_{-}^{\lambda}(d \bar{u}),\label{eq:densityMW3} \\
\rho_{}^{\lambda^{\prime} \lambda}(\bar{u} d) & =
\bigl\{ {\cal M}_{-}^{\lambda^{\prime}_{}}(\bar{u} d) \bigr\}^{\ast}_{}
{\cal M}_{-}^{\lambda}(\bar{u} d),\label{eq:densityMW4}
\end{align}
\end{subequations}
represent the elements of the density matrix in the helicity basis of the $W^-_{}$ in the $d\bar{u}$ c.m. frame and those in the $\bar{u}d$ c.m. frame, respectively.
By looking at the amplitudes in eqs.~(\ref{eq:WpampAll}) and (\ref{eq:WmampAll}), it is easy to find the relations
\begin{subequations}\label{eq:relationdensitymatrix}
\begin{align}
\rho_{}^{\lambda^{\prime} \lambda}(u\bar{d}) & = \rho_{}^{\lambda^{\prime} \lambda}(d\bar{u}), \\
\rho_{}^{\lambda^{\prime} \lambda}(\bar{d}u) & = \rho_{}^{\lambda^{\prime} \lambda}(\bar{u}d).
\end{align}
\end{subequations}
The relations indicate that the $W^+_{}$ and $W^-_{}$ are always in the same state of polarisation. This fact makes our analysis in the next section simpler, because we do not have to distinguish the $W^+_{}$ from the $W^-_{}$ with regard to states of polarisation.
\subsection{Decay angular distributions of the polarised $W^+_{}$ and $W^-_{}$}\label{sec:corsspolW}
The complete differential cross section for the process $pp \to W^+_{}H$ followed by $W^+_{} \to jj$ in the narrow width approximation can be expressed in terms of the density matrices in eq.~(\ref{eq:densityMW12}) as
\begin{align}
\frac{d\sigma^{W^+_{}}_{}}{d\hat{s}\ dy\ d\cos{\theta}\ d\cos{\widehat{\theta}}\ d\widehat{\phi} }
= & \frac{m_W^{}k}{ 8192 \pi^3_{} \Gamma_W^{} s \hat{s}^{\frac{3}{2}}_{}}
g^2_{} \Bigl(\sum_{u, d} \bigl| V_{ud}^{} \bigr|^2_{} \Bigr)\nonumber \\
& \times \sum_{u,d} \biggl[ u(x_1^{}) \bar{d}(x_2^{})\ d_{}^{- \dagger} \rho_{}^{}(u\bar{d}) d_{}^{-}
+
\bar{d}(x_1^{}) u(x_2^{})\ d_{}^{- \dagger} \rho_{}^{}(\bar{d}u) d_{}^{-} \biggr],\label{eq:differentialWH1}
\end{align}
and that for the process $pp \to W^-_{}H$ followed by $W^-_{} \to jj$ can be expressed in terms of the density matrices in eq.~(\ref{eq:densityMW34}) as
\begin{align}
\frac{d\sigma^{W^-_{}}_{}}{d\hat{s}\ dy\ d\cos{\theta}\ d\cos{\widehat{\theta}}\ d\widehat{\phi} }
= & \frac{m_W^{}k}{ 8192 \pi^3_{} \Gamma_W^{} s \hat{s}^{\frac{3}{2}}_{}}
g^2_{} \Bigl(\sum_{u, d} \bigl| V_{ud}^{} \bigr|^2_{} \Bigr)\nonumber \\
& \times \sum_{u,d} \biggl[ d(x_1^{}) \bar{u}(x_2^{})\ d_{}^{- \dagger} \rho_{}^{}(d\bar{u}) d_{}^{-}
+
\bar{u}(x_1^{}) d(x_2^{})\ d_{}^{- \dagger} \rho_{}^{}(\bar{u}d) d_{}^{-} \biggr], \label{eq:differentialWH2}
\end{align}
where the $3\times 3$ matrix form in eq.~(\ref{eq:matrixform}) is used. For the definition of variables, see below eq.~(\ref{eq:differentialZH}).
As the quark flavors $u$ and $d$ summed in the above equations, we consider all the quark flavors but the top quark. As a result, the unitarity of the CKM matrix gives $\sum_{u, d} | V_{ud}^{} |^2_{} = 2$. By using the relations in eq.~(\ref{eq:relationdensitymatrix}), we obtain the complete differential cross section for the process $pp \to W^{\pm}_{}H$ followed by $W^{\pm}_{} \to jj$ as the sum of the above two cross sections in the following compact form:
\begin{align}
\frac{d\sigma^{W^+_{}+W^-_{}}_{}}{d\hat{s}\ dy\ d\cos{\theta}\ d\cos{\widehat{\theta}}\ d\widehat{\phi} }
& = \frac{m_W^{}k}{ 4096 \pi^3_{} \Gamma_W^{} s \hat{s}^{\frac{3}{2}}_{}}
g^2_{} \nonumber \\
\times \sum_{u,d} \biggl[ \Bigl\{ u(x_1^{}) & \bar{d}(x_2^{}) + d(x_1^{}) \bar{u}(x_2^{}) \Bigr\}\ d_{}^{- \dagger} \rho(u\bar{d}) d_{}^{-}
+
\Bigl\{ \bar{u}(x_1^{}) d(x_2^{}) + \bar{d}(x_1^{}) u(x_2^{}) \Bigr\}\ d_{}^{- \dagger} \rho(\bar{d}u) d_{}^{-} \biggr].\label{eq:differentialWH}
\end{align}
As in Section~\ref{sec:crossZboson}, we derive the differential cross section with respect to $\hat{s}$, $\cos{\widehat{\theta}}$ and $\widehat{\phi}$ by integrating over $y$ and $\cos{\theta}$ in eq.~(\ref{eq:differentialWH}).
It is straightforward to confirm that the integration over $y$ and $\cos{\theta}$ in the 4 different approaches summarised in Table~\ref{table:integration} can be performed in the same manner as eqs.~(\ref{eq:PartdifferentialZH1}), (\ref{eq:PartdifferentialZH2}), (\ref{eq:PartdifferentialZH3}) and (\ref{eq:PartdifferentialZH4}).
The resulting differential cross sections are
\begin{subequations}
\begin{align}
\frac{d\sigma^{W^+_{}+W^-_{}}}{d\hat{s}\ d\cos{\widehat{\theta}}\ d\widehat{\phi} }\biggr|_{{\cal A}}^{}
= &
{\cal T}^W_{} \int_{0}^{y_{\mathrm{cut}}^{}} dy\ 2 \Bigl[ u(x_1^{}) \bar{d}(x_2^{}) + d(x_1^{}) \bar{u}(x_2^{}) + \bar{u}(x_1^{}) d(x_2^{}) + \bar{d}(x_1^{}) u(x_2^{}) \Bigr] \nonumber \\
& \times
d_{}^{- \dagger}
\left(
\begin{array}{ccc}
c_1^{} | \hat{N}_{}^+ |^2_{} & \frac{1}{2} c_2^{} ( \hat{N}_{}^{+} )^*_{} \hat{N}_{}^{-} & 0 \\
\frac{1}{2} c_2^{} \hat{N}_{}^{+} ( \hat{N}_{}^{-} )^*_{} & c_1^{} | \hat{N}_{}^- |^2_{} & 0 \\
0 & 0 & c_2^{} | \hat{N}_{}^0 |^2_{} \\
\end{array}
\right)
d_{}^{-}
,\label{eq:PartdifferentialWH1}\\
\frac{d\sigma^{W^+_{}+W^-_{}}}{d\hat{s}\ d\cos{\widehat{\theta}}\ d\widehat{\phi} }\biggr|_{{\cal B}}^{}
= &
{\cal T}^W_{} \int_{0}^{y_{\mathrm{cut}}^{}} dy\ 2 \Bigl[ u(x_1^{}) \bar{d}(x_2^{}) + d(x_1^{}) \bar{u}(x_2^{}) - \bar{u}(x_1^{}) d(x_2^{}) - \bar{d}(x_1^{}) u(x_2^{}) \Bigr] \nonumber \\
& \times
d_{}^{- \dagger}
\left(
\begin{array}{ccc}
0 & 0 & - c_3^{} ( \hat{N}_{}^{+} )^*_{} \hat{N}_{}^{0} \\
0 & 0 & - c_3^{} ( \hat{N}_{}^{-} )^*_{} \hat{N}_{}^{0} \\
- c_3^{} \hat{N}_{}^{+} ( \hat{N}_{}^{0} )^*_{} & - c_3^{} \hat{N}_{}^{-} ( \hat{N}_{}^{0} )^*_{} & 0 \\
\end{array}
\right)
d_{}^{-},\label{eq:PartdifferentialWH2} \\
\frac{d\sigma^{W^+_{}+W^-_{}}}{d\hat{s}\ d\cos{\widehat{\theta}}\ d\widehat{\phi} }\biggr|_{{\cal C}}^{}
= &
{\cal T}^W_{} \int_{0}^{y_{\mathrm{cut}}^{}} dy\ 2 \Bigl[ u(x_1^{}) \bar{d}(x_2^{}) + d(x_1^{}) \bar{u}(x_2^{}) + \bar{u}(x_1^{}) d(x_2^{}) + \bar{d}(x_1^{}) u(x_2^{}) \Bigr] \nonumber \\
& \times
d_{}^{- \dagger}
\left(
\begin{array}{ccc}
0 & 0 & c_4^{} ( \hat{N}_{}^{+} )^*_{} \hat{N}_{}^{0} \\
0 & 0 & - c_4^{} ( \hat{N}_{}^{-} )^*_{} \hat{N}_{}^{0} \\
c_4^{} \hat{N}_{}^{+} ( \hat{N}_{}^{0} )^*_{} & - c_4^{} \hat{N}_{}^{-} ( \hat{N}_{}^{0} )^*_{} & 0 \\
\end{array}
\right)
d_{}^{-},
\label{eq:PartdifferentialWH3}\\
\frac{d\sigma^{W^+_{}+W^-_{}}}{d\hat{s}\ d\cos{\widehat{\theta}}\ d\widehat{\phi} }\biggr|_{{\cal D}}^{}
= &
{\cal T}^W_{} \int_{0}^{y_{\mathrm{cut}}^{}} dy\ 2 \Bigl[ u(x_1^{}) \bar{d}(x_2^{}) + d(x_1^{}) \bar{u}(x_2^{}) - \bar{u}(x_1^{}) d(x_2^{}) - \bar{d}(x_1^{}) u(x_2^{}) \Bigr] \nonumber \\
& \times
d_{}^{- \dagger}
\left(
\begin{array}{ccc}
- c_5^{} | \hat{N}_{}^+ |^2_{} & 0 & 0 \\
0 & c_5^{} | \hat{N}_{}^- |^2_{} & 0 \\
0 & 0 & 0 \\
\end{array}
\right)
d_{}^{-},\label{eq:PartdifferentialWH4}
\end{align}
\end{subequations}
where
\begin{align}
{\cal T}^W_{} =\frac{m_W^{}k}{ 4096 \pi^3_{} \Gamma_W^{} s \hat{s}^{\frac{3}{2}}_{}} g^2_{} \sum_{u, d} \bigl| V_{ud}^{} \bigr|^2_{},
\end{align}
the constant values $c_i^{}$ $(i=1,2,3,4,5)$ are summarised in Table~\ref{table:coefficients}, and $\hat{N}^{\lambda=\pm,0}_{}$ are defined in eq.~(\ref{eq:Wsubamplitudes}). By comparing the above 4 differential cross sections with eq.~(\ref{eq:diffcrossgeneral}), we obtain the 36 coefficients $F_{ia}^{W}$ $(i={\cal A, B, C, D})$ $(a=1,2,\cdots,9)$ in total:
\begin{subequations}\label{eq:observablesppWH}
\begin{align}
F_{ {\cal A} ({\cal C}) a}^{W}
& =
{\cal T}^W_{} \int_{0}^{y_{\mathrm{cut}}^{}} dy\ 2 \Bigl[ u(x_1^{}) \bar{d}(x_2^{}) + d(x_1^{}) \bar{u}(x_2^{}) + \bar{u}(x_1^{}) d(x_2^{}) + \bar{d}(x_1^{}) u(x_2^{}) \Bigr] f_{ {\cal A} ({\cal C}) a}^{W}, \\
F_{ {\cal B} ({\cal D}) a}^{W}
& =
{\cal T}^W_{} \int_{0}^{y_{\mathrm{cut}}^{}} dy\ 2 \Bigl[ u(x_1^{}) \bar{d}(x_2^{}) + d(x_1^{}) \bar{u}(x_2^{}) - \bar{u}(x_1^{}) d(x_2^{}) - \bar{d}(x_1^{}) u(x_2^{}) \Bigr] f_{ {\cal B} ({\cal D}) a}^{W},
\end{align}
\end{subequations}
where $f_{ia}^{W}$ $(i={\cal A, B, C, D})$ $(a=1,2,\cdots,9)$ are obtained by the following replacements in those coefficients of the process $pp \to ZH$ in eq.~(\ref{eq:observables-small}):
\begin{align}
\hat{M}_{\sigma}^{\lambda} \to \hat{N}_{}^{\lambda}, \ \
\sigma \to -1,\ \
\tau \to -1.
\end{align}
Table~\ref{table:symproperty} corresponding to the process $pp \to W^{\pm}_{}H$ is obtained by the following simple replacements in Table~\ref{table:symproperty}:
\begin{align}
\hat{M}_{\sigma}^{\lambda} \to \hat{N}_{}^{\lambda}, \ \
F_{ia}^{} \to F_{ia}^{W}.
\end{align}
Just as the process $pp\to ZH$, there are in total $15$ coefficients which can be non-zero.
Observation of the 6 coefficients $F_{{\cal D} 3}^{W}$, $F_{{\cal A} 3}^{W}$, $F_{{\cal B} 4}^{W}$, $F_{{\cal C} 4}^{W}$, $F_{{\cal B} 7}^{W}$ and $F_{{\cal C} 7}^{W}$ has difficulty, since it requires the charge (or flavor) identification of the parent quark of the jet. Only the remaining 9 coefficients $F_{{\cal A} 1}^{W}$, $F_{{\cal A} 2}^{W}$, $F_{{\cal D} 1}^{W}$, $F_{{\cal C} 5}^{W}$, $F_{{\cal B} 5}^{W}$, $F_{{\cal A} 6}^{W}$, $F_{{\cal C} 8}^{W}$, $F_{{\cal B} 8}^{W}$ and $F_{{\cal A} 9}^{W}$ are, therefore, actually measurable. We emphasise that these 9 coefficients are necessary and sufficient to determine all of the 9 independent combinations of the density matrix elements.
\subsection{Influences of non-standard $HWW$ interaction}\label{sec:angcoeff-W}
\begin{figure}[t]
\centering
\includegraphics[scale=0.45]{fig4.pdf}
\caption{\small
In the left panel, the 2 coefficients $-F_{{\cal A} 9}^{W}$ ($\Box$) and $F_{{\cal C} 8}^{W}$ ($\times$) divided by $F_{{\cal A} 1}^{W}$ are shown for $\widetilde{h}_4^W=0.1$. These 2 coefficients are constrained to be identically zero by CP invariance.
In the middle panel, the coefficient $F_{{\cal B} 8}^{W}$ divided by $F_{{\cal A} 1}^{W}$ are shown for $h_3^W=0+0.1i$. This coefficient is constrained to be identically zero by $\mathrm{CP\widetilde{T}}$ invariance.
In the right panel, the 2 coefficients $F_{{\cal D} 1}^{W}$ ($\Box$) and $-F_{{\cal B} 5}^{W}$ ($\times$) divided by $F_{{\cal A} 1}^{W}$ are shown for $\widetilde{h}_4^W=0+0.1i$.
These 2 coefficients are doubly constrained to be identically zero by CP invariance {\it and} $\mathrm{CP\widetilde{T}}$ invariance.
For each result, the values in the 5 bins, from left to right, are obtained after integration over $\hat{s}$ in the regions $(m_W^{}+m_H^{}) < \hat{s}^{1/2}_{} < 300$, $300 < \hat{s}^{1/2}_{} < 400$, $400 < \hat{s}^{1/2}_{} < 600$, $600 < \hat{s}^{1/2}_{} < 1000$ and $1000 < \hat{s}^{1/2}_{} < 14000$ in units of GeV. }
\label{figure:WHplot}
\end{figure}
In this Section, with the same motivation as Section~\ref{sec:angcoeff}, we focus on the measurable coefficients which are strictly zero or small in the SM, and study the influences of the non-standard $HWW$ coupling.
In the left panel of Figure~\ref{figure:WHplot}, the 2 coefficients $-F_{{\cal A} 9}^{W}$ ($\Box$) and $F_{{\cal C} 8}^{W}$ ($\times$) divided by $F_{{\cal A} 1}^{W}$ are shown for the CP-odd form factor $\widetilde{h}_4^W=0.1$. CP invariance requires these 2 coefficients to be identically zero; observation of a non-zero value signals CP violation.
In the middle panel of Figure~\ref{figure:WHplot}, the coefficient $F_{{\cal B} 8}^{W}$ divided by $F_{{\cal A} 1}^{W}$ is shown for the CP-even form factor $h_3^W=0+0.1i$. $\mathrm{CP\widetilde{T}}$ invariance requires this coefficient to be identically zero; observation of a non-zero value indicates the existence of re-scattering effects.
In the right panel of Figure~\ref{figure:WHplot}, the 2 coefficients $F_{{\cal D} 1}^{W}$ ($\Box$) and $-F_{{\cal B} 5}^{W}$ ($\times$) divided by $F_{{\cal A} 1}^{W}$ are shown for the CP-odd form factor $\widetilde{h}_4^W=0+0.1i$.
These 2 coefficients are doubly constrained to be identically zero by CP invariance {\it and} $\mathrm{CP\widetilde{T}}$ invariance; observation of a non-zero value indicates CP violation {\it and} the existence of re-scattering effects.
For each result, the values in the 5 bins, from left to right, are obtained after integration over $\hat{s}$ in the regions $(m_W^{}+m_H^{}) < \hat{s}^{1/2}_{} < 300$, $300 < \hat{s}^{1/2}_{} < 400$, $400 < \hat{s}^{1/2}_{} < 600$, $600 < \hat{s}^{1/2}_{} < 1000$ and $1000 < \hat{s}^{1/2}_{} < 14000$ in units of GeV. \\
We discuss the differences between the coefficients $F_{ia}^{}$ in the process $pp\to ZH$ and those $F_{ia}^{W}$ in the process $pp\to W^{\pm}_{}H$.
From the comparisons between the results for the non-standard $HZZ$ coupling in Figures~\ref{figure:ZHplot1}, \ref{figure:ZHplot2} and \ref{figure:ZHplot3} and the results for the non-standard $HWW$ coupling in Figure~\ref{figure:WHplot}, we notice that $F_{{\cal A} 9}^{W}$ and $F_{{\cal C} 8}^{W}$ are comparable with $F_{{\cal A} 9}^{}$ and $F_{{\cal C} 8}^{}$, respectively,
while $F_{{\cal B} 8}^{W}$, $F_{{\cal D} 1}^{W}$ and $F_{{\cal B} 5}^{W}$ are consistently larger than $F_{{\cal B} 8}^{}$, $F_{{\cal D} 1}^{}$ and $F_{{\cal B} 5}^{}$, respectively.
As we have discussed at the last paragraph of Section~\ref{sec:helamp}, the $Z$ boson is in a partially polarised state, while the $W^+_{}$ and $W^-_{}$ are in a completely polarised state. The degree of polarisation affects the magnitudes of the coefficients, hence it is expected that the coefficients $F_{ia}^{W}$ are always equal or larger than the coefficients $F_{ia}^{}$. The existence of the overall $\sigma$ in $F_{{\cal B} 8}^{}$, $F_{{\cal D} 1}^{}$ and $F_{{\cal B} 5}^{}$ (see eq.~(\ref{eq:observables-small})) indicates that these coefficients are weakened according to the degree of polarisation of the $Z$ boson. These coefficients are denoted by the symbol $\circ$ in the last column in Table~\ref{table:symproperty}.
\section{Summary}\label{sec:summary}
The measurements of the Higgs boson couplings to the SM particles are essential tests of the SM.
The $Z$ boson in the process $pp\to ZH$ and the $W^+_{}$ and $W^-_{}$ in the process $pp \to W^{\pm}_{}H$ can be in polarised states, and it would be possible to study the Higgs boson couplings to the weak bosons ($HZZ$, $HZ\gamma$ and $HWW$) in detail from a careful analysis of these states of polarisation.
A density matrix contains the complete information about a state of polarisation, and all of the elements of the density matrix should be made use of in such a careful analysis.
In this paper, such an analysis approach has been presented. \\
Determination of the density matrix of the $Z$ ($W$) boson requires measurements of the angular distributions of the $Z$ ($W$) decay products. In $pp$ collisions, this is difficult when the $W$ boson decays into a charged lepton and a neutrino. When the $W$ boson decays into two quarks, we cannot distinguish $W^+_{}$ from $W^-_{}$ in view of the difficulty of flavor identification of both $W^+_{}$ and $W^-_{}$ decay products.
Therefore, we have considered the process $pp \to W^{\pm}_{}H$ ($W^{\pm}_{} \to jj$) as the sum of the process $pp \to W^+_{}H$ ($W^+_{} \to jj$) and the process $pp \to W^-_{}H$ ($W^-_{} \to jj$).
We have found that the $W^+_{}$ and $W^-_{}$ are always in the same state of polarisation (i.e. the density matrix of the $W^+_{}$ is always the same as that of $W^-_{}$).
By using this fact, we have developed an analysis approach which can be applied both to $pp \to ZH$ ($Z \to f\bar{f}$) and to $pp \to W^{\pm}_{}H$ ($W^{\pm}_{} \to jj$) in the same manner. \\
We have derived the 4 different differential cross sections with respect to $\hat{s}$ (the invariant mass squared of the $Z$ ($W$) boson and the Higgs boson) and $\cos{\widehat{\theta}}$ and $\widehat{\phi}$ (the $Z$ ($W$) decay angles), by integrating the complete differential cross section over the other phase space variables in the 4 different approaches.
Among the 36 coefficients ($=9 \times 4$) of these 4 differential angular distributions, only 15 coefficients can be non-zero (these coefficients are summarised in Table~\ref{table:symproperty}).
These 15 coefficients are written in terms of the elements of the density matrices, and there exist 9 independent combinations of the elements of the density matrices.
Observation of the 9 coefficients among the 15 coefficients does not require the charge (or flavor) identification of the $Z$ ($W$) decay products.
In the analysis of the $W^+_{}$ and $W^-_{}$, only these 9 coefficients are measurable. We have found that these 9 coefficients are necessary and sufficient to determine all of the 9 independent combinations of the density matrix elements (i.e. one coefficient corresponds to one combination).
We have clarified the restrictions on the 15 coefficients imposed by the CP and $\mathrm{CP\widetilde{T}}$ symmetries.
Some of the coefficients are required to be identically zero by CP invariance; observation of a non-zero value in these coefficients signals CP violation. Similarly, some of the coefficients are required to be identically zero by $\mathrm{CP\widetilde{T}}$ invariance; observation of a non-zero value in these coefficients indicates the existence of re-scattering effects.
These coefficients are in particular interesting as observables at the LHC, since observation of a non-zero or large value in these coefficients immediately signals the existence of non-standard $HZZ$, $HZ\gamma$ and/or $HWW$ interactions.
\section*{Acknowledgments}
I would like to thank Kaoru Hagiwara for many discussions. I am also grateful to members of CP3, Universit\'e Catholique de Louvain for their hospitality.
My work is supported by the Alexander von Humboldt Foundation.
\small
\bibliographystyle{JHEP}
\nocite{*}
| 383cf6592c4367b00fa48b3fa27cde61710c3ba4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Spin superfluidity was suggested in the 1970s \cite{ES-78b,ES-82} (see recent reviews \cite{Adv,Mac}). Manifestation of spin superfluidity is a stable spin supercurrent proportional to the gradient of the phase $\varphi$ (spin rotation angle in a plane) and not accompanied by dissipation, in contrast to a dissipative spin diffusion current proportional to the gradient of spin density.
In general the spin current proportional to the gradient of the phase $\varphi$ is ubiquitous and exists in any spin wave, although in these cases variation of the phase $\varphi$ is small \cite{HalHoh}. Analogy with mass and charge persistent currents (supercurrents) arises when at long (macroscopical) spatial intervals along streamlines the phase variation is many times larger than $2\pi$.
The supercurrent state is a helical spin structure, but in contrast to equilibrium helical structures is metastable.
An elementary process of relaxation of the supercurrent is phase slip. In this process a vortex with $2\pi$ phase variation around it crosses current streamlines decreasing the total phase variation along streamlines by $2\pi$. Phase slips are suppressed by energetic barriers for vortex creation, which disappear when phase gradients reach critical values determined by the Landau criterion.
Emergence of superfluidity is conditioned by special topology of the order parameter space. This requires the easy-plane anisotropy for the spontaneous magnetization in ferromagnets or the antiferromagnetic vector in antiferromagnets \cite{Adv,Mac}.
Experimentally evidence of spin superfluidity was obtained in the $B$-phase of superfluid $^3$He \cite{Bun}, in which spin degrees of freedom of the order parameter make it analogous to antiferromagnets. Recently Bozhko {\em et al.} \cite{spinY} declared detection of spin supercurrent at room temperature in a coherent magnon condensate created in yttrium-iron-garnet (YIG) magnetic films by strong parametric pumping \cite{Dem6}. In YIG the equilibrium order parameter in the spin space is not confined to some easy plane as required for spin superfluidity. Nevertheless, Sun {\em et al.} \cite{Pokr} argued that spin superfluidity is still possible.
Indeed, one cannot rule out that metastability of supercurrent states is provided by barriers not connected with topology of the equilibrium order parameter, if the system is far from the equilibrium. In this work we confirm that it is possible indeed and determine critical values of possible supercurrents at which metastability is lost (the analogue of the Landau criterion).
In the end we overview experiments on experimental detection of spin superfluidity in various materials. We define the term {\em superfluidity} in its original meaning known from the times of Kamerlingh Ohnes and Kapitza: transport of some physical quantity (mass, charge, or spin) on macroscopical distances without essential dissipation. This definitely requires a large number of full $2\pi$ rotations along current streamlines as pointed out above.
On the basis of this we conclude that recent claims of observation spin superfluidity in YIG films \cite{spinY} and in antiferromagnets \cite{BunLvo} were not justified, and further experiments are needed for observation of spin superfluidity in magnetically ordered solids.
\section{Topology and superfluid spin currents } \label{TopCur}
In a superfluid the order parameter is a complex wave function $\psi = \psi_0 e^{i\varphi}$, where the modulus $\psi_0$ of the wave function is a positive constant determined by minimization of the energy. The
phase $\varphi$ is a degeneracy parameter since the energy does not depend on $\varphi$ because of gauge invariance. Any of degenerate ground states in a closed annular channel (torus) maps on some point at a circumference $|\psi|=\psi_0$ in the complex plane $\psi$, while a
current state with the phase change $2\pi n$ around the torus maps onto a circumference (Fig.~\ref{Fig02}a) winding around the circumference $n$ times. It is evident that it is impossible to change $n$ keeping the path on the circumference $|\psi|=\psi_0$ all the time. In the language of topology states with different $n$ belong to different classes, and
$n$ is a {\em topological charge}. Only a phase slip can change it when the path in the complex plane leaves the circumference. This should cost energy, which is spent on creation of a vortex crossing the cross-section of the torus channel and changing $n$ to $n-1$.
\begin{figure}
\begin{center}
\includegraphics[width=20pc]{Topology.pdf}
\end{center}
\caption{\label{Fig02}Mapping of current states on the order parameter space.\newline
a) Mass currents in superfluids. The current state in torus maps on a circumference of radius $|\psi|$ on the complex plane $\psi$. \newline
b) Spin currents in an isotropic ferromagnet. The current state in torus maps on an equatorial circumference on the sphere of radius $M$ (top). Continuous shift of the path on the sphere (middle) reduces it to a point (bottom), which is the ground state without currents. \newline
c) Spin currents in an easy-plane ferromagnet. Easy-plane anisotropy contracts the order parameter space to an equatorial circumference in the $xy$ plane topologically equivalent to the order parameter space in superfluids. \newline
d) Spin currents in an isotropic ferromagnet in a magnetic field $\bm H$ parallel to the axis $z$ with nonequilibrium magnetization $M_z$ supported by magnon pumping. Spin is confined in the plane normal to $\bm H$. This is an ``easy plane'' of dynamical origin.}
\end{figure}
If we consider transport of spin parallel to the axis $z$ the analog of the phase of the superfluid order parameter is the rotation angle of the spin component in the plane $xy$, which we note also as $\varphi$. In isotropic ferromagnets the order parameter space is a sphere of radius equal to the absolute value of the magnetization vector $\bm M$ (Fig.~\ref{Fig02}b). All points on this sphere correspond to the same energy of the ground state. Suppose we created the spin current state with monotonously varying phase $\varphi$ in a torus. This state maps on the equatorial circumference on the order parameter sphere. Topology allows to continuously shift the circumference and to reduce it to the point of the northern pole. During this process shown in Fig.~\ref{Fig02}b the path remains on the sphere all the time and therefore no energetic barrier resists to the transformation. Thus metastability of the current state is not expected.
In a ferromagnet with easy-plane anisotropy the order parameter space contracts from the sphere to an equatorial circumference in the $xy$ plane. This makes the order parameter space topologically equivalent to that in superfluids (Fig.~\ref{Fig02}c). Now transformation of the equatorial circumference to the point shown in Fig.~\ref{Fig02}b costs anisotropy energy. This allows to expect metastable spin currents (supercurrents). They relax to the ground state via phase slips, in which magnetic vortices cross spin current streamlines. States with vortices maps on a hemisphere of radius $M$ either above or below the equator.
Up to now we considered states close to the equilibrium (ground) state. In a ferromagnet in a magnetic field the equilibrium magnetization is parallel to the field. However, by pumping magnons into the sample it is possible to tilt the magnetization with respect to the magnetic field. This creates a nonstationary state, in which the magnetization precesses around the magnetic field. Although the state is far from the true equilibrium, but it, nevertheless, is a state of minimal energy at fixed magnetization $M_z$. Because of inevitable spin relaxation the state of uniform precession requires permanent pumping of spin and energy. However, if processes violating the spin conservation law are weak, one can ignore them and treat the state as a quasi-equilibrium state. The state of uniform precession maps on a circumference parallel to the $xy$ plane, but in contrast to the easy-plane ferromagnet (Fig.~\ref{Fig02}c) the plane confining the precessing magnetization is much above the equator and not far the northern pole (Fig.~\ref{Fig02}d). One can consider also a current state, in which the phase (the rotation angle in the $xy$ plane) varies not only in time but also in space with a constant gradient. The current state will be metastable due to the same reason as in an easy-plane ferromagnet: in order to relax via phase slips the magnetization should go away from the circumference on which the state of uniform precession maps, and this increases the energy. Then the plane, in which the magnetization precesses, can be considered as an effective ``easy plane'' originating not from the equilibrium order parameter topology but created dynamically. The concept of dynamical easy plane is applicable for a magnon condensate in YIG films with some modifications. They take into account that the spin conservation law is not exact due to magnetostatic energy and the precession is not uniform since spin waves in YIG films have the energy minima at non-zero wave vectors.
In our discussion we assumed that phase gradients were small and ignored the gradient-dependent (kinetic) energy. At growing gradient we reach the critical gradient at which barriers making the supercurrent stable vanish. The critical gradient must be determined from the. criterion analogous to the famous Landau criterion \cite{Adv,Son17}.
\section{ Spin waves in YIG films} \label{LLT}
The coherent state of magnons is nothing else but a classical spin wave, and one can use the classical equations of the Landau--Lifshitz--Gilbert (LLG) theory.
YIG is a ferrimagnet with complicated magnetic structure consisting of numerous sublattices. However at slow degrees of freedom relevant for our analysis one can treat it simply as an isotropic ferromagnet \cite{Melk} with the spontaneous magnetization $\bm M$ described by the hamiltonian
\begin{eqnarray}
{\cal H }=\int\left[-\bm H\cdot \bm M +D{\nabla_i \bm M \cdot \nabla_i \bm M\over 2} \right]d\bm r
+ \int {\bm \nabla \cdot \bm M(\bm r) \bm \nabla \cdot \bm M(\bm r_1)\over 2|\bm r-\bm r_1| } d\bm r\,d\bm r_1 .
\eem{}
Here the first term is the Zeeman energy in the magnetic field $\bm H$, the second term $\propto D$ is the inhomogeneous exchange energy, and the last one is the magnetostatic (dipolar) energy.
In a weak spin wave propagating in the plane $xz$ in a magnetic field $\bm H$ parallel to the axis $z$
$M_z \approx M-{M_\perp^2/2M}$ and $\bm \nabla \cdot \bm M \approx \nabla_x M_x$,
where $M_\perp=\sqrt{M_x^2+M_y^2}$. In the LLG theory the absolute value of the magnetization vector $\bm M$ does not vary in space and time, and the linearized LLG equations are reduced to two equations for only two independent magnetization components:
\begin{eqnarray}
\dot M_x=-\gamma M{\delta {\cal H} \over \delta M_y}=-\gamma H M_y + \gamma DM (\nabla_x^2 M_y+\nabla_z^2 M_y),
\nonumber \\
\dot M_y=\gamma M_z {\delta {\cal H} \over \delta M_x} =\gamma H M_x - \gamma DM (\nabla_x^2 M_x+\nabla_z^2 M_x)
- \gamma M\nabla_x \left(\int {\nabla_x M_x(\bm r_1) \over |\bm r-\bm r_1| }d\bm r_1 \right),
\eem{em}
where $\gamma$ is the gyromagnetic ratio and $\delta {\cal H} / \delta M_x$ and $\delta {\cal H} / \delta M_y$ are functional derivatives. For the plane wave with the frequency $\omega$ and the wave vector $\bm k(k_x,0,k_z)$ Eqs.~(\ref{em}) yield the spectrum
\begin{equation}
\omega(k)= \gamma \sqrt{(H + DM k^2)\left(H + DM k^2 +{4\pi M k_x^2\over k^2}\right)}.
\ee{disp}
\begin{figure}
\includegraphics[width=15pc]{Fig1.pdf}\hspace{2pc}
\begin{minipage}[b]{14pc}\caption{\label{Fig1}The YIG film of thickness $d$ in a magnetic field $\bm H$ parallel to the axis $z$.}
\end{minipage}
\end{figure}
A spin wave propagating in the film of thickness $d$ parallel to the plane $yz$ (Fig.~\ref{Fig1}) must satisfy the boundary conditions at two film surfaces $x=\pm d/2$. In the situation when both the exchange and the dipole energies are of importance, the boundary problem is not trivial and was discussed in details in Ref.~\cite{Son17}. According to this analysis, the spectrum of the spin wave propagating along the magnetic field in the film is
\begin{equation}
\omega(k_z) \approx \gamma \left(H + DM k_z^2 +{2\pi^3 M \over k_z^2d^2}\right).
\ee{disp1}
This follows from the spectrum \eq{disp} at $k_x= \pi /d$ and corresponds to
a plane wave with weakly elliptical polarization propagating along the magnetic field (the axis $z$).
The energy $E = \omega (M-\langle M_z\rangle)/ \gamma$
and the frequency $\omega(k_z)$ given by \eq{disp1} have two degenerate minima \cite{Melk} at finite $k_z =\pm k_0$ where magnons can condense (Fig.~\ref{Fig2}). Here
\begin{equation}
k_0=\left({2\pi^3\over Dd^2}\right)^{1/4}=\left({2\pi^2\over l_d^2 d^2}\right)^{1/4},
\ee{}
where $l_d=\sqrt{D/\pi}$ is a small scale determined by the exchange energy.
\section{Spin supercurrent in the YIG magnon condensate}
In the linear theory the distribution of magnons between two condensates is arbitrary and does not affect the total energy at fixed magnetization $\langle M_z\rangle$. But non-linear corrections lift this degeneracy. The most important nonlinear term arises from the magnetostatic energy, which is maximal
for the standing wave (two energy minima are equally populated by magnons):
\begin{eqnarray}
M_x =2\sqrt{2M(M-\langle M_z\rangle)}\cos {\pi x\over d} \cos k_zz\cos \omega t,
\nonumber \\
M_y = 2\left(1+{2\pi^3 M\over Hk_z^2d^2}\right)\sqrt{2M(M-\langle M_z\rangle)}\cos {\pi x\over d}\cos k_zz \sin\omega t,
\eem{sw}
and is equal to
\begin{equation}
E_{ms} =\int {\nabla _z M_z(\bm r) \nabla_z M_z(\bm r_1))\over 2|\bm r-\bm r_1| } d\bm r\,d\bm r_1={3 \pi (M-\langle M_z\rangle)^2 \over 2}.
\ee{mse}
The spin current appears if the wave numbers $k_z$ of two condensates differ from $\pm k_0$ (Fig.~\ref{Fig2}). The current is proportional to the phase gradient $\nabla_z\varphi=K=k_z-k_0 \ll k_0$. Keeping the magnetization $\langle M_z \rangle$ fixed as before and taking into account the nonlinear magnetostatic term (\ref{mse}) the energy in the spin-current state apart from some constant terms is
\begin{equation}
\Delta E = {d^2\omega (k_z)\over dk_z^2} {M-\langle M_z\rangle\over \gamma}{(\nabla_z\varphi)^2\over 2}+{3 \pi (M-\langle M_z\rangle)^2 \over 2}.
\ee{hMz}
\begin{figure}[b]
\includegraphics[width=15pc]{Fig2.pdf}
\begin{minipage}[b]{14pc}\caption[]{\label{Fig2}The spin-wave spectrum in a YIG film. In the ground state the magnon condensate occupies two minima in the $k$ space with $k_z=\pm k_0$ (large circles). In the current state two parts of the condensate are shifted to $k = \pm k_0 +K$ (small circles). }
\end{minipage}
\end{figure}
Stability of the spin-current state can be checked following the principal idea of the Landau criterion of superfluidity \cite{Adv,Son17}.
Let us consider slowly varying in space weak perturbations $m_z=M_z-\langle M_z \rangle $ and $\nabla_z\varphi'=\nabla_z \varphi-K $. For stability of the supercurrent the quadratic form obtained from expansion of the energy (\ref{hMz}) in perturbations $m_z$ and $\nabla_z \varphi'$ must be always positive. This takes place as far as $\nabla_z\varphi=K$ is less than the critical value
\begin{equation}
(\nabla_z\varphi)_{cr} =\sqrt{3\pi \gamma (M-\langle M_z \rangle)\over {d^2\omega (k_0)\over dk_z^2} }= \sqrt{{ 3(M-\langle M_z \rangle)\over M}}{ k_0^2 d\over 4\pi }.
\ee{}
Note that applying our course of derivation to superfluid hydrodynamics one obtains exactly the Landau critical velocity equal to the sound velocity (see Sec.~2.1 in Ref.~\cite{Adv}). One can find numerical estimation of the critical gradient and comparison with other theories in Ref.~\cite{Son17}.
\section{Experimental detection of spin superfluidity} \label{DC}
A smoking gun of spin superfluidity in the $B$-phase of superfluid $^3$He was an experiment with a spin current through a long channel connecting two cells filled by the fluid with coherently precessing spins \cite{BRJ}. A small difference of the precession frequencies in two cells leads to a linear growth of
difference of the precession phases in the cells and a phase gradient in the channel. When the gradient reaches the critical value phase slips must occur. Phase slips were detected experimentally at the total phase difference exceeding $2\pi$.
A sharp $2\pi$ phase slip is reliable evidence of non-trivial spin supercurrents at phase gradients restricted by finite critical values.
Some time ago Bunkov {\em et al.} \cite{BunLvo} declared that their observation of coherent spin precession in antiferromagnets demonstrated ``high-$T_c$ spin superfluidity'' since random inhomogeneity of their samples inevitably generates local spin currents. The very idea that random spin currents
fluctuating at the disorder scale are the same as persistent currents, which are constant at macroscopic scales, is bizarre at least. Moreover, no shred of evidence, direct or indirect, of presence of any spin current, random or non-random, was presented. The spin current, or any quantity connected with it, is totally absent in any experimental data or formula. The only ``evidence'' was the statement that spin currents {\em must} be present in their experiment. Apparently Bunkov {\em et al.} do not see any difference between a theoretical prediction ({\em must} be present) and experimental confirmation of the prediction.
Recently Bozhko {\em et al.} \cite{spinY} declared detection of spin superfluidity at high temperatures in a decaying magnon condensate in a YIG film. In their experiment the phase gradient emerged from spin precession difference produced by a temperature gradient. But the estimate made in Ref.~\cite{Son17} showed that the total phase difference across the magnon cloud in the experiment did not exceed about 1/3 of the full $2\pi$ rotation. Spin current relaxation via discrete $2\pi$ phase slips is impossible in this case independently from phase gradient. Therefore Bozhko {\em et al.} dealt with trivially stable spin currents present in any spin wave. Their existence does not require new confirmations after a half-a-century experiments with spin waves at {\em all} temperatures.
In summary, spin superfluidity, which was revealed in superfluid $^3$He-$B$,
still waits its experimental confirmation in magnetically ordered solids.
\section*{References}
\providecommand{\newblock}{}
| e7760e17807f7821158ea080cf25f7dd80e986d4 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section*{Abstract}
{\bf
We propose using smeared boundary states $e^{-\tau H}|\cal B\rangle$ as variational approximations to the ground state of a conformal field theory deformed by relevant bulk operators. This is motivated by recent studies of quantum quenches in CFTs and of the entanglement spectrum in massive theories. It gives a simple criterion for choosing which boundary state should correspond to which combination of bulk operators, and leads to a rudimentary phase diagram of the theory in the vicinity of the RG fixed point corresponding to the CFT, as well as rigorous upper bounds on the universal amplitude of the free energy. In the case of the 2d minimal models explicit formulae are available. As a side result we show that the matrix elements of bulk operators between smeared Ishibashi states are simply given by the fusion rules of the CFT.
}
\vspace{10pt}
\noindent\rule{\textwidth}{1pt}
\tableofcontents\thispagestyle{fancy}
\noindent\rule{\textwidth}{1pt}
\vspace{10pt}
.
\section{Introduction}\label{sec1}
Conformal field theories (CFTs) are supposed to correspond to the non-trivial renormalization group (RG) fixed points of relativistic quantum field theories (QFTs). Such theories typically contain a number of scaling operators of dimension $\Delta<d$ (where $d$ is the space-time dimension), which, if added to the action, are relevant and drive the theory to what is, generically, a trivial fixed point. The points along this trajectory then correspond to a massive QFT. In general there is a multiplicity of such basins of attraction of the RG flows, but enumerating them and determining which combinations of relevant operators lead to which basins, and therefore to what kind of massive QFT, in general requires non-perturbative methods.
This problem is equivalent to mapping out the phase diagram in the vicinity of the critical point corresponding to the CFT.
Another way of characterizing these massive theories is through the analysis of the possible boundary states of the CFT. Imagine the scenario in which the relevant operators are switched on in only a half-space, say $x_0<0$. This will then appear as some boundary condition on the CFT in $x_0>0$. However the boundary conditions themselves undergo RG flows, with fixed points corresponding to so-called conformal boundary conditions. Therefore on scales $\sim M^{-1}$, where $M$ is the mass scale of the perturbed theory, the correlations near the boundary should be those of a conformal boundary condition, deformed by \em irrelevant \em boundary operators.
A similar question is raised through recent work on the spectrum of the entanglement hamiltonian in massive QFTs \cite{LR}. If the theory is defined in ${\bf R}^D$ and is in its ground state, and we study the entanglement between the degrees of freedom in the half-space $A: x_1>0$ and its complement, then the entanglement, or modular, hamiltonian $K_A=-(1/2\pi)\log\rho_A$, where $\rho_A$ is the reduced density matrix of $A$, takes the form \cite{HSK1,HSK2}
\be
K_A=\int_{x_1>0}x_1T_{00}(x)d^Dx\,,
\ee
which is nothing but the generator of rotations in euclidean space, or of boosts in lorentzian signature.
In 1+1 dimensions we may consider a conformal transformation $z=x_1+ix_0=\epsilon e^w$ which sends the euclidean $z$-plane, punctured at the origin by a disc of radius $\epsilon$ representing the UV cutoff, to an semi-infinite cylinder of circumference $2\pi$. $K_A$ is then simply the generator of translations around this cylinder. However, if the QFT corresponds to a perturbed CFT, it is not conformally invariant, but rather the couplings transform as
\be
\lambda\to\lambda\,\epsilon^{2-\Delta}e^{(2-\Delta){\rm Re}\,w}\,,
\ee
where $\Delta<2$ is the scaling dimension of the perturbing operator. Thus the dimensionless coupling
$g=\lambda\epsilon^{2-\Delta}$ is effectively switched on over a length scale $O(1)$ near ${\rm Re}\,w\sim\log(1/g) $. If we are interested the low-lying spectrum of $K_A$, corresponding to the R\'enyi entropies ${\rm Tr}\,\rho_A^n$ with $n\gg1$, the effective circumference of the cylinder is $2\pi n$ and we are then in a similar situation to the above, where the massive theory for ${\rm Re}\,w>\log(1/g)$ acts as an effective boundary condition on the CFT in ${\rm Re}\,w<\log(1/g)$. As concluded in \cite{LR}, the low-lying spectrum of $K_A$ should therefore be that of the (boundary) CFT, with an appropriate boundary condition depending on the bulk perturbation.
The same question arises in the context of quantum quenches \cite{CCQQ}. In this case we are interested in the real time evolution of an initial state $|\Psi_0\rangle$ under a hamiltonian $H$ of which it is not an eigenstate. An example is the case where $H=H_{CFT}$ and $|\Psi_0\rangle$ is the ground state of the massive perturbed CFT. This is a difficult problem, and in \cite{CCQQ,JCQQ} the step was taken of replacing this ground state by a conformal boundary state perturbed by irrelevant operators.\footnote{In \cite{CCQQ} only the smeared states of the form (\ref{smeared}) were considered, which happen to lead to subsystem thermalization, while in \cite{JCQQ} it was argued that more general states should lead to a generalized Gibbs ensemble.} This allows the explicit computation of the imaginary time evolution and the continuation to real time, which would be very difficult for the exact ground state of the massive theory.
Thus an important problem in all these cases is to determine to which conformal boundary condition a particular combination of bulk operators should correspond. For simple examples this is apparent by physical inspection. For example, the CFT corresponding to the critical Ising model has two relevant operators, coupling to the magnetic field $h$ and the deviation $t$ of the reduced temperature from its critical value. There are three stable RG fixed points at
$(h=0,t\to+\infty)$ and $h\to\pm\infty$, respectively the sinks for the disordered and the two ordered phases, and corresponding to the three conformal boundary conditions when the Ising spins are respectively free and fixed, either up or down.
One way to make this identification is to think of the boundary condition as defining a state $|\cal B\rangle$ when the theory is quantized on a time slice $x_0=$ constant. In that language we may regard the perturbed CFT as described by a hamiltonian operator
\be
\hat H=\hat H_{CFT}+\sum_j\lambda_j\int\hat\Phi_j(x)d^{d-1}x\,,
\ee
where the $\{\hat\Phi_j\}$ are relevant operators. We then ask which $|\cal B\rangle$, suitably deformed by boundary irrelevant operators, is closest in some sense to the ground state of $\hat H$ at strong coupling.
Conformal boundary states by themselves contain no scale, and therefore cannot be good candidates for the ground state of $\hat H$. Indeed, they must have infinite energy compared to this state. In known examples in 2d (and, for example, for free theories in higher dimensions) they are also non-normalizable. We must therefore deform them by irrelevant boundary operators in order to give them a scale. The simplest such operator is the stress tensor $\hat T_{00}$, which has scaling dimension $d$ and therefore boundary RG eigenvalue $(d-1)-d=-1$. Since its space integral is the CFT hamiltonian, including only this operator is tantamount to considering boundary states `smeared' by evolution in imaginary time:
\be\label{smeared}
e^{-\tau \hat H_{CFT}}|\cal B\rangle\,,
\ee
where $\tau>0$ is parameter with the dimensions of length. Such states have finite energy and correlation length $\propto\tau^{-1}$, and also finite norm.
Such smeared boundary states may be thought of as a continuum version of matrix product states (MPS). Indeed, a lattice discretization of the euclidean path integral, illustrated in Fig.~\ref{MPS}, suggests that such states correspond to matrices with internal dimension $\sim N^{\tau/\delta\tau}$, where $N$ is the number of states on each lattice edge and $\delta\tau$ is the time step. However, unlike discrete MPS states, the smeared boundary state (\ref{smeared}) automatically has the correct short-distance behavior of the CFT.
\begin{figure}[h]\label{MPS}
\centering
\includegraphics[width=0.9\textwidth]{smeared}
\caption{Path integral for smeared boundary state (left) and its lattice discretization (right) as a matrix product state. On the right, each vertical column of lattice sites represents a matrix. The horizontal lines represent contractions between these in the internal space, and the vertical dangling bonds label the physical degrees of freedom.}
\end{figure}
From this point of view it is therefore natural to regard
(\ref{smeared}) as a variational \em ansatz\em, with $\tau$ and the choice of boundary state $|\cal B\rangle$ as variational parameters.
In this paper we explore this idea further and show that this program can be carried through explicitly for the $A_m$ (diagonal) series of unitary minimal 2d CFTs. It should be extendable to the other non-diagonal minimal models, and in principle to other rational 2d CFTs, and indeed to higher dimensional theories if enough information is available about the CFT.
More specifically, given a set of physical conformal boundary states $\{|a\rangle\}$, (whose definition is recalled in Sec.~\ref{sec2}) we take as a variational ground state
\be
|\{\alpha_a\},\{\tau_a\}\rangle=\sum_a\alpha_a\,e^{-\tau_a\hat H_{CFT}}|a\rangle\,,
\ee
and compute the variational energy per unit volume
\be
\lim_{L\to\infty}\frac1{L^D}\frac{\langle\{\alpha_a\},\{\tau_a\}|\hat H_{CFT}+\sum_j\lambda_j\int\hat\Phi_j(x)d^{d-1}x|\{\alpha_a\},\{\tau_a\}\rangle}{\langle\{\alpha_a\},\{\tau_a\}|\{\alpha_a\},\{\tau_a\}\rangle}\,,
\ee
minimizing this with respect to the $\{\alpha_a\}$ and $\{\tau_a\}$.
An important general consequence of the analysis is that, in the limit $L\to\infty$, the minimizing states are always purely physical, that is all but one $\{\alpha_a\}$ vanishes. This is because both $H_{CFT}$ and the perturbing operators turn out to be diagonal in this basis. This is reassuring, as in principle the minimizers could be non-physical linear combinations of these, for example the Ishibashi states in 1+1 dimensions.
Specializing now to the case of 1+1 dimensions,
for the minimal models, the precise values of these diagonal matrix elements are related to the elements of the modular $S$-matrix of the CFT, and, with these in hand, it is straightforward, for a fixed set of couplings $\{\lambda_j\}$ to determine which values of $\tau_a$ and $a$ minimize the variational energy, and thus to map out a rudimentary phase diagram of the theory in the vicinity of the CFT.
It then turns out that although this approach yields correct results in some aspects, for example in determining which combination of bulk couplings $\{\lambda_j\}$ best matches a given boundary state $a$, it is not capable of reproducing some of the finer details of the phase boundaries between different states $a$. In this approximation these are always first-order, and `massless' RG flows to other non-trivial CFTs are
not properly accounted for. This can be seen as a limitation of the particular trial state, which could be remedied by including other operators acting on the boundary state, but at the cost of the loss of analytic tractability.
However, an amusing side result of the analysis is that matrix elements of primary bulk operators $\hat\Phi_j$ between Ishibashi states $\langle\langle i|$, $|k\rangle\rangle$ (which are boundary states within a single Virasoro module) are simply proportional to the fusion rule coefficients:
\be
\langle\langle i|e^{-\tau H}\,\hat\Phi_j\,e^{-\tau H}|k\rangle\rangle\propto N_{ijk}\,.
\ee
This is a consequence of the Verlinde formula \cite{Verlinde}, and to our knowledge has not been previously observed. Although this matrix element should be proportional to the OPE coefficient $c_{ijk}$ which governs the matrix element $\langle i|\hat\Phi_j|k\rangle$ between highest weight states, it is rather surprising that the contributions of all the descendent states should conspire to give the integer-valued fusion rule coefficient.
The outline of this paper is as follows. In Sec.~\ref{sec2} we set up the formalism and prove some general results. In Sec.~\ref{sec3} we apply this to the case of the diagonal minimal models, with the $A_3$ and $A_4$ cases as specific examples, and finally in Sec.~\ref{sec4} give a summary and some further remarks.
After this paper was completed, I was made aware of Ref.~\cite{Kon}, in which similar ideas are explored. However that paper is based on comparing ratios of overlaps between different boundary states and numerical approximations to the exact ground state of the deformed theory, rather than the variational method adopted here. The overlap method is shown to work well for the case of the perturbed Ising model, but is computationally more intensive.
\section{General formalism.}\label{sec2}
As described in the Introduction, we consider a $d$ $(=D+1)$-dimensional CFT perturbed by its bulk primary operators $\{\Phi_j\}$ with coupling constants
$\{\lambda_j\}$, so the hamiltonian is:
\be\label{Hp}
\hat H=\hat H_{CFT}+\sum_j\lambda_j\int\hat\Phi_j(x)d^{D}x\,.
\ee
The theory is quantized on a spatial torus of volume $L^{D}$, where $L$ is much larger than any other scale in the theory.
We assume for simplicity that the $\{\Phi_j\}$ are all scalars and that they have their CFT normalization
\be\label{norm}
\langle\hat\Phi_j(x)\hat\Phi_j(0)\rangle_{CFT}=|x|^{-2\Delta_j}\,,
\ee
where $\Delta_j$ is the scaling dimension of $\Phi_j$.
Although we are interested in relevant perturbations with $\Delta_j<d$, these will in general lead to a finite number of primitive UV divergences up to some finite order in the couplings (as for a super-renormalizable deformation of a free theory), in particular in the ground state energy which we are trying to approximate. These divergences may be subtracted by adding a finite number of counter-terms to $\hat H$ determined the OPEs of the $\{\Phi_j\}$. We assume this has been done.
For example, if $2\Delta_j\geq d$ there is a UV divergence in the ground state energy at $O(\lambda_j^2)$. This is subtracted by a term in $\hat H$ proportional to the unit operator. This does not affect the variational procedure in general.
The case $2\Delta_j=d$ is special and leads to a logarithmic anomaly in the energy. This will be discussed for the 2d Ising model in Sec.~\ref{seclog}.
Conformal boundary states $|\cal B\rangle$ are defined by the condition
\be\label{T0k}
\hat T_{0k}(x)|{\cal B}\rangle=0\,,\quad(k=1,\ldots,D)\,,
\ee
where $\hat T_{ij}$ is the energy-momentum tensor of the CFT. That is, they are annihilated by the momentum density operator, and so are invariant under local time reparametrizations. (For boundaries with a space-like normal there is no energy flow across the boundary.)
Although in higher dimensions these states, and their classification, are poorly understood except for free or weakly coupled CFTs, in 2d much more is known \cite{JC89,PZ}. The Hilbert space is acted on by two copies $({\cal V}\otimes\overline{\cal V})$ of the Virasoro algebra, generated by
\be
\hat L_n=\frac L{2\pi}\int e^{2\pi nix/L}\,\hat T(x)dx\,,\quad\hat{\overline L}_n=\frac L{2\pi}\int e^{-2\pi nix/L}\,\hat{\overline T}(x)dx\,,
\ee
where, as usual, $T\equiv T_{zz}=-T_{00}+T_{11}-2iT_{01}$ and
$\overline T\equiv T_{\bar z\bar z}=-T_{00}+T_{01}+2iT_{01}$, in euclidean signature.
It is spanned by states $|i,N\rangle\otimes\overline{|i',N'\rangle}$, where $N$ labels the states of a module of $\cal V$ with highest weight state labelled by $i$, and similarly for $\overline{\cal V}$. For CFTs with central charge $c\geq1$, this is a Virasoro module, while for the minimal models with $c<1$ it is a Kac module with the null states projected out.
The condition (\ref{T0k}) then corresponds to
\be
\big(\hat T(x)-\hat{\overline T}(x)\big)|{\cal B}\rangle=0\,.
\ee
In terms of the Virasoro generators this becomes
\be
\big(\hat L_n-\hat{\overline L}_{-n}\big)|{\cal B}\rangle=0\,,
\ee
whose solution is the span of the Ishibashi states
\be\label{Ishi}
|i\rangle\rangle=\sum_N|i,N\rangle\otimes\overline{|i,N\rangle}\,,
\ee
where the sum is over all the orthonormalized states in the module.
However, the Ishibashi states are not physical, in the sense that, when they are chosen as boundary states on the opposite edges $x_0=\pm\tau$ of an annulus, the partition function $Z={\rm Tr}\,e^{-L\hat H'}$ evaluated in terms of the generator $\hat H'$ of translations around the annulus does not have the form of a sum over eigenstates with non-negative integer coefficients, as it must if periodic spatial boundary conditions are imposed. For the diagonal minimal $A_m$ models, the physical states which do have this property are linear combinations of the Ishibashi states
\be\label{phys}
|a\rangle=\sum_j\frac{S^i_a}{(S^i_0)^{1/2}}\,|i\rangle\rangle\,,
\ee
where $S^i_k$ is the matrix by which the Virasoro characters transform under modular transformations.
The multiplicities of the eigenstates $j$ of $\hat H'$ which do propagate are then given by the fusion rule coefficients $N^j_{ab}$. In particular the vacuum state propagates only if $a=b$, that is $N^0_{ab}=\delta_{ab}$.
While similar results are available for the non-diagonal minimal models, wider results for general CFTs are not available, which is why we mainly restrict to the $A_m$ models in explicit calculations.
In higher dimensions, the boundary states satisfying (\ref{T0k}) also form a linear space, and we assume that the physical states may be identified analogously. Consider the partition function in the slab ${\mathbb T}^D\times\{-\tau,\tau\}$ (where ${\mathbb T}^D$ is a $D$-dimensional torus of volume $L^D$) with boundary states $|a\rangle$, $|b\rangle$ at $x_0=\pm\tau$:
\be
Z_{ab}=\langle b|e^{-2\tau\hat H_{CFT}}|a\rangle\,,
\ee
(where $\hat H_{CFT}$ is the generator of translations in $x_0$)
and, similarly to the 2d case, demand that when evaluated by quantizing in one of the spatial directions, it has the form of a trace over intermediate states whose energies all scale like $\tau^{-1}$. However this is difficult to implement since this spectrum on the torus is not related to the conformal spectrum for $d>2$.
In fact, we shall need only a weaker condition: that physical states $\{a,b,\ldots\}$ should satisfy
\be\label{gap}
Z_{ab}/(Z_{aa}Z_{bb})^{1/2}=O\big(e^{-{\rm const.}(L/2\tau)^D}\big)\qquad\mbox{for $L/\tau\to\infty$}\,.
\ee
This may be understood as follows: when the boundary conditions are the same, we expect that
\be
Z_{aa}\sim e^{\sigma_a(L/2\tau)^D}\,,
\ee
where the exponent is (minus) the Casimir energy of a system between two identical plates. In this geometry this is always attractive, thus $\sigma_a>0$. It must scale as $L^D$, and, since the boundary conditions and the bulk theory are scale-invariant, also as $\tau_a^{-D}$. The quantity $-\sigma_a L^{D-1}/(2\tau)^D$ is the ground state energy of the generator of translations around one of the spatial cycles of the torus. On the other hand the exponent on the right hand side of (\ref{gap}) is the gap to the lowest-energy state in the sector with $ab$ boundary conditions. It is the interfacial energy from the point of view of $d$-dimensional classical statistical mechanics.
Thus the physical boundary states may in principle be determined by diagonalizing the partition functions in the slab in the limit $L\gg\tau$. We assume that this has been done.
As discussed in the introduction, we use as variational states for the ground state of the perturbed hamiltonian (\ref{Hp}), the ansatz
\be
|\{\alpha_a\},\{\tau_a\}\rangle=\sum_a\alpha_a\,e^{-\tau_a\hat H_{CFT}}|a\rangle\,.
\ee
We first discuss the inner product of these states
\be
\langle a|e^{-\tau_aH_{CFT}}e^{-\tau_bH_{CFT}}|b\rangle\,.
\ee
This is the partition function $Z_{ab}$ in slab of width $\tau_a+\tau_b$ with boundary conditions $a,b$ on opposite faces. As discussed above, for physical boundary states in the limit $L\gg\tau_a+\tau_b$
\be\label{Zab}
Z_{ab}\sim\delta_{ab}\, e^{\sigma_a(L/(2\tau_a))^D}\,.
\ee
The matrix elements of the unperturbed hamiltonian $\hat H_{CFT}$ are, for the same reason, diagonal in this basis as long as
$L\gg\tau_{a,b}$, and may be found by differentiating (\ref{Zab})
\be
\langle a|\hat H_{CFT}\,e^{-2\tau_a\hat H_{CFT}}|a\rangle\sim\delta_{ab}\frac{D\sigma_aL^D}{(2\tau_a)^{D+1}} e^{\sigma_a(L/(2\tau_a))^D}\,.
\ee
Finally we need the matrix elements of the perturbation
\be
\langle a|e^{-\tau_a\hat H_{CFT}}\,\hat\Phi_j(x)\,e^{-\tau_b\hat H_{CFT}}|b\rangle\,,
\ee
which is a one-point function in the slab. Once again, if we evaluate this by inserting a complete set of eigenstates of a generator of translations around the torus, this is dominated by its ground state if $L\gg\tau_{a,b}$, but this contributes only if $a=b$. So the perturbation is also diagonal in the basis of physical states (but not in the Ishibashi basis: see Sec.~\ref{sec33}).
When $a=b$ the one-point functions in the mid-plane of the slab have the form
\be
\langle\Phi_j(x)\rangle=\frac{A_a^j}{(2\tau_a)^{\Delta_j}}\,,
\ee
where the amplitudes $A_a^j$ are universal given the normalization (\ref{norm}) of the operator.
Since the perturbed hamiltonian is diagonal in the physical basis of variational states, the problem becomes much simpler:
for each $a$ we should minimize the variational energy per unit volume
\be\label{Ea}
E_a=\frac{D\sigma_a}{(2\tau_a)^{D+1}}+\sum_j\lambda_j\frac{A_a^j}{(2\tau_a)^{\Delta_j}}\,,
\ee
with respect to $\tau_a$, and then choose the $a$ which gives the absolute minimum.
Note that having found this minimum $a$ for a particular set of couplings $\{\lambda_j\}$, since $E_a$ transforms multiplicatively under
\be
\lambda_j\to e^{(D+1-\Delta_j)\ell}\lambda_j\,,\quad \tau_a\to e^{-\ell}\tau_a\,,\quad
E_a\to e^{(D+1)\ell}E_a\,,
\ee
the absolute minimum will occur for the same value of $a$ along an RG trajectory. This is reassuring, since each point on the trajectory should be described by the same massive QFT up to a rescaling of the mass, which is proportional to $1/\tau_a^{\rm min}$.
Since the $\{\Phi_j\}$ are relevant,
$\Delta_j<D+1$, so that the behavior of $E_a$ as $\tau_a\to0$ (but still $\gg\epsilon$) is dominated by the first term and is positive if $\sigma_a>0$
(which corresponds to the physically reasonable case of an attractive Casimir force.) As $\tau_a\to\infty$ it approaches zero, dominated by the term with smallest $\Delta_j$ and $\lambda_j\not=0$. If the sign is negative this implies that $E_a$ has a negative minimum at some finite value of $\tau_a$. At least for the 2d minimal models we can show that there is always some $a$ for which $\lambda_jA_a^j<0$, so this minimum always exists.
\subsection{Trace of the energy-momentum tensor.}
We may infer a general result about the trace $\langle\Theta\rangle=\langle T^i_i\rangle$ of the energy-momentum tensor in the perturbed theory as approximated by this method. For given set of relevant perturbations $\{\lambda_j\}$ this is given by the response of the action to a scale transformation
\be
\Theta(x)=-\sum_j(D+1-\Delta_j)\lambda_j\Phi_j(x)\,.
\ee
Differentiating (\ref{Ea}) we see that at the minimum
\be
\frac{(D+1)D\sigma_a}{(2\tau_a)^{D+1}}+\sum_j\Delta_j\lambda_j\langle\Phi_j\rangle=0\,,
\ee
and so
\be
E=-(D+1)^{-1}\sum_j\Delta_j\lambda_j\langle\Phi_j\rangle+\sum_j\lambda_j\langle\Phi_j\rangle=-(D+1)^{-1}\langle\Theta\rangle\,.
\ee
Once again, this is reassuring, as we expect that in the ground state of a relativistic theory $\langle T_{00}\rangle=-\langle T_{kk}\rangle$ for $k\not=0$, and so
\be
\langle\Theta\rangle =-\langle T_{00}\rangle+\sum_{k=1}^D\langle T_{kk}\rangle=-(D+1)\langle T_{00}\rangle=-(D+1)E\,.
\ee
The variational method therefore gives a lower bound on $\langle\Theta\rangle$.
\section{2d minimal models}\label{sec3}
We now specialize to the case of the 2d $A_m$ minimal models.
In 2d, the Casimir amplitude $\sigma_a=\pi c/24$, independent of $a$, where $c$ is the central charge.
When $a=b$ the expectation values of the one-point functions in a long strip of width $2\tau_a$ may be found by a conformal mapping from the half-plane to have the form
\be\label{strip}
\langle\Phi_j(x,\tau)\rangle_{\text{strip}}
=\frac{\widetilde A_a^j}{\left((2\tau_a/\pi)\sin(\pi\tau/2\tau_a)\right)^{\Delta_j}}\,,
\ee
where the amplitude governs the behavior of the one-point function in the upper half-plane $y>0$ with boundary condition $a$ on the real axis:
\be
\langle\Phi_j(y)\rangle_{\text{half-plane}}=\frac{\widetilde A_a^j}{y^{\Delta_j}}\,.
\ee
In (\ref{strip}) we should set $\tau=\tau_a$, whence we read off that $A_a^j=\pi^{\Delta_j}\widetilde A_a^j$.
If the operator $\Phi_j$ has its standard CFT normalization (\ref{norm}), the amplitudes $\widetilde A_a^j$ are universal. In \cite{CL} they were computed in terms of the overlap between the boundary state $|a\rangle$ and the highest weight state $|j\rangle$ corresponding to the primary operator $\Phi_j$:
\be
\widetilde A^j_a=\frac{\langle j|a\rangle}{\langle 0|a\rangle}\,.
\ee
This follows by conformally mapping the upper half plane to a semi-infinite cylinder $x>0$ with a boundary condition $a$ at $x=0$, and comparing the result for $x\to\infty$ with the result of inserting a complete set of eigenstates of the generator of translations along the cylinder.
For the $A_m$ minimal models, inserting the expression (\ref{phys}) for $|a\rangle$ we then find \cite{CL}
\be\label{At}
\widetilde A^j_a=\frac{S_a^j}{S_a^0}\left(\frac{S_0^0}{S_0^j}\right)^{1/2}\,.
\ee
To summarize, the variational energy (\ref{Ea}) in this case is given by
\be
E_a=\frac{\pi c}{24(2\tau_a)^2}+\sum_j\lambda_j\frac{S_a^j}{S_a^0}\left(\frac{S_0^0}{S_0^j}\right)^{1/2}\frac{\pi^{\Delta_j}}{(2\tau_a)^{\Delta_j}}\,.
\ee
It is useful to rescale the couplings by positive constants $\tilde\lambda_j=\pi^{\Delta_j}(S_0^0/S^j_0)^{1/2}\lambda_j$
so that this simplifies to
\be
E_a=\frac{\pi c}{24(2\tau_a)^2}+\sum_{j\not=0}\frac{S_a^j}{S_a^0}\frac{\tilde\lambda_j}{(2\tau_a)^{\Delta_j}}\,.
\ee
Note that that sum over $j$ excludes $j=0$ which corresponds to adding the unit operator and therefore a constant shift in the energy.
There are two general statements which follow from the fact that $S$ is a symmetric orthogonal matrix, and that the elements $S_0^j$ are all positive.
First, since all its rows are orthogonal and non-zero it follows that, for $j\not=0$, some of the elements $S_a^j$ are positive and some negative. Therefore, if ${j^*}$ corresponds to the smallest value of $\Delta_j$ such that $\lambda_j\not=0$, and therefore dominates the behavior of $E_a$ as $\tau_a\to\infty$, no matter what the sign of $\lambda_{j^*}$ we may always find at least one $a$ such that
$\lambda_{j^*}S_a^j<0$, and so $E_a$ approaches zero from below. Since $E_a\to+\infty$ as $\tau_a\to0$, this implies that, for these $a$, $E_a$ has a negative minimum at finite $\tau_a$, corresponding to a finite correlation length. This rules out the possibility
that this variational ansatz can describe massless flows to another non-trivial CFT.
Second, we may ask whether there is a combination of couplings $\{\lambda_j\}$ which will lead to a prescribed $b$ as overall minimum. The answer is affirmative. For, suppose we choose
\be\label{combo}
\tilde\lambda_j=-g(2\mu)^{\Delta_j-2}\,S^j_b\,,
\ee
where $g$ is a positive constant and $\mu$ is some fixed scale $>\epsilon$. Then the second term in (\ref{Ea}) is, when $\tau_a=\mu$,
\be
-\frac g{S^0_a\mu^2}\sum_{j\not=0}S^j_aS^j_b=-\frac g{S^0_a\mu^2}\left(\delta_{ab}-S_a^0S_b^0\right)\,.
\ee
Since $0<S^0_a<1$, this is $<0$ if $a=b$ and $>0$ otherwise. Thus, at this scale, the boundary state $b$ will correspond to the lowest trial energy\footnote{This does not rule out the possibility that some other $E_a$ might come lower than this at some other scale, but in practice this does not seem to happen.}. Note that (\ref{combo}) implies including some irrelevant couplings in the mix of deformations.
Further results depend on the detailed form of the modular $S$-matrix for the $A_m$ models.
In particular, we may ask what happens if a single $\lambda_j$ is non-zero.
Depending on whether $\lambda_j>0$ or $<0$, we have to determine which value of $a$ minimizes (maximizes) the ratio $S^j_a/S^0_a$.
Label the positions of the bulk operators in the Kac table by $j=(r,s)$, with $1\leq r\leq m-1$ and $1\leq s\leq m$, and $(r,s)$ identified with $(m-r,m+1-s)$. The label $j=0$ corresponds to $(r,s)=(1,1)$). Similarly label the boundary states $a$ by $(\alpha,\beta)$.
The $A_m$ minimal series of CFTs is conjectured to be the scaling limit of the critical lattice RSOS $A_m$ models \cite{RSOS1,RSOS2}. These models are defined on a square lattice. At each node $R$ there is an integer-valued height variable
$h(R)$ satisfying $1\leq h(R)\leq m$, with the RSOS constraint that $|h(R)-h(R')|=1$ if $R$ and $R'$ are nearest neighbors. The heights may be thought of as living at the nodes of the Dynkin diagram $A_m$, so each configuration is a many-to-one embedding of the diagram into the square lattice. The critical Boltzmann weights of the lattice model are specified in terms of the elements $s_h^0(m)$ of the Perron-Frobenius eigenvector corresponding to the largest eigenvalue of the adjacency matrix of $A_m$. The general eigenvector has the form
\be
s_h^j(m)\propto\sin\frac{\pi jh}{m+1}\,.
\ee
The microscopic interpretation of the conformal boundary states (\ref{phys}) for these models has been given in
\cite{SB,PB}.
The simplest boundary states are when the boundary lies at 45$^{\circ}$ to the principal lattice vectors, and the heights on the boundary are all fixed to the same particular value $h$, say. These have been identified with the conformal boundary conditions with the Kac labels $(\alpha,\beta)=(1,h)$. The second simplest type of microscopic boundary condition is when the boundary heights are fixed to $h$ and on the neighboring diagonal they are fixed to $h+1$. These have been identified with $(\alpha,\beta)=(h,1)$. In \cite{PB} a complete set of microscopic boundary conditions was identified for each Kac label $(\alpha,\beta)$ but these become increasingly complicated. In general the microscopic boundary states corresponding to labels near the center of the Kac table are increasingly disordered.
The ratios of elements of the modular $S$-matrix for the diagonal $A_m$ models are
\be\label{ratio}
\frac{S^j_a}{S^0_a}=\frac{S^{r,s}_{\alpha,\beta}}{S^{1,1}_{\alpha,\beta}}
=(-1)^{(r+s)(\alpha+\beta)}\,\frac{\sin\frac{\pi r\alpha}m}{\sin\frac{\pi \alpha}m}
\,\frac{\sin\frac{\pi s\beta}{m+1}}{\sin\frac{\pi \beta}{m+1}}=(-1)^{(r+s)(\alpha+\beta)}\,\frac{s_\alpha^r(m-1)}{s_\alpha^1(m-1)}\frac{s_\beta^s(m)}{s_\beta^1(m)}\,.
\ee
Locating the global maximum and minimum of this expression for general $(r,s)$ is simplified by the fact that, for fixed $(r,s)$, it factorizes into expressions depending only on $\alpha$ and $\beta$ respectively. Thus we can restrict to the four possible products of the maximum and minimum of each factor, and compare these values.
In each factor the numerator is an oscillating function which is modulated by the positive denominator, which itself has minima
at $\alpha=1,m-1$ (and $\beta=1,m$), which for the lattice $A_m$ models correspond to the most ordered states.
The most relevant bulk operator corresponds to $(r,s)=(2,2)$, when (\ref{ratio}) becomes
\be
4\cos\frac{\pi\alpha}m\cos\frac{\pi\beta}{m+1}\,.
\ee
The extrema of each factor are at $\alpha=1, m-1$ and $\beta=1,m$. Thus for $\lambda_{2,2}>0$ the minimum energy corresponds to $(\alpha,\beta)=(1,m)=(m-1,1)$, and, for $\lambda_{2,2}<0$, $(\alpha,\beta)=(1,1)=(m-1,m)$.
These correspond to the most ordered states, at the ends of the Dynkin diagram. This is to be expected as, in the Landau-Ginzburg correspondence, $\Phi_{2,2}$ is the most relevant Z$_2$ symmetry breaking operator.
Similarly, the most relevant Z$_2$ even operator is $\Phi_{3,3}$, when (\ref{ratio}) becomes
\be
(2\cos\frac{2\pi\alpha}m+1)(2\cos\frac{2\pi\beta}{m+1}+1)\,.
\ee
If $m$ is even, the first factor varies between $2\cos\frac{2\pi}m+1$ at $\alpha=1,m-1$, and $-1$ at $\alpha=\frac12m$,
and the second factor varies between $2\cos\frac{2\pi}{m+1}+1$ at $\beta=1,m$, and $2\cos\frac{\pi m}{m+1}+1$ at $\beta=\frac12m, \frac12m+1$. Thus for $\lambda_{3,3}<0$ there are degenerate minima for $(\alpha,\beta)=(1,1)=(m-1,m)$ and $(\alpha,\beta)=(1,m)=(m-1,1)$ (the most ordered states, which break the Z$_2$ symmetry.).
On the other hand for $\lambda_{3,3}>0$ we need to compare the quantities
\be
(-1)(2\cos\frac{2\pi}{m+1}+1)\,, \quad (2\cos\frac{\pi m}{m+1}+1)(2\cos\frac{2\pi}m+1)\,.
\ee
Numerically, the first is more negative, so the minimum energy in this case corresponds to $\alpha=\frac12m$, $\beta=1,m$. These are Z$_2$-symmetric states.
For odd $m$ the same story holds, with $\alpha$ and $\beta$ interchanged.
Another interesting special case is $\Phi_{1,3}$. This is a perturbation which, with the correct sign, is supposed to flow to the $A_{m-1}$ minimal CFT, and for the other sign to a state with large degeneracy \cite{RSOS1}. As we have seen, such massless flows cannot be accounted for within this set of trial states. In this case (\ref{ratio}) simplifies to
\be
2\cos\frac{2\pi\beta}{m+1}+1\,,
\ee
independent of $\alpha$. Depending on the sign of the coupling, this picks out the boundary states either with $\beta=1,m$ or with $\beta\approx\frac12m$. In both cases, however, there is an $(m-1)$-fold degeneracy of candidate ground states. This reflects a flow towards a true first-order transition, as expected for one sign of the coupling \cite{RSOS1}, or the best attempt of this approximation to reproduce the critical point of the $A_{m-1}$ model, as expected for the other sign. This is an important check on the effectiveness of our approach.
These somewhat cryptic general remarks are best illustrated with some simple examples.
\subsection{The Ising model.}
This corresponds to $A_3$. The perturbed hamiltonian is
\be\label{HIsing}
\widehat H=\widehat H_{CFT}+t\int\hat\varepsilon dx+h\int\hat\sigma dx\,,
\ee
where $\varepsilon=\Phi_{2,1}=\Phi_{1,3}$ and $\sigma=\Phi_{1,2}=\Phi_{2,2}$ are the energy density and magnetization operators respectively.
In this case, the bulk operators are $\{\Phi_j\}=(1,\epsilon,\sigma)$, and the boundary states in the same labeling are $(+,-,f)$, corresponding to fixed$(+)$, fixed$(-)$ and free boundary conditions on the Ising spins. The $S$-matrix in this ordering of the basis is
\be
S=\left(\begin{array}{ccc}\ffrac12&\ffrac12&\ffrac1{\sqrt2}\\
\ffrac12&\ffrac12&-\ffrac1{\sqrt2}\\
\ffrac1{\sqrt2}&-\ffrac1{\sqrt2}&0\end{array}\right)\,.
\ee
After rescaling the couplings as above, we find that
\begin{eqnarray}
E_+&=&\frac\pi{48(2\tau_+)^2}+\frac t{2\tau_+}+\sqrt2\frac h{(2\tau_+)^{1/8}}\,,\label{Eaa}\\
E_-&=&\frac\pi{48(2\tau_-)^2}+\frac t{2\tau_-}-\sqrt2\frac h{(2\tau_-)^{1/8}}\,,\label{E2a}\\
E_f&=&\frac\pi{48(2\tau_f)^2}-\frac t{2\tau_f}\,.\label{Ef}
\end{eqnarray}
For $h=0$, $t>0$, corresponding to the disordered state, it is clear that the minimizer is $E_f$. For the opposite sign of $t$ with $h>0$, the minimizer is $E_-$, corresponding to negative magnetization (recall the definition of the sign of the couplings in (\ref{HIsing})), and vice versa. As $h\to0$ from either side with $t<0$, we remain in one or the other of these states, corresponding to spontaneous symmetry breaking.
However, for $t>0$ and $0<h\ll t^{15/8}$ there is a problem. The minimum of $E_-$ is found by balancing the last two terms in (\ref{E2a}), and therefore occurs when $\tau_-=O((t/h)^{7/8})$ at a value
$E_-=-O(h^{8/7}/t^{1/7})$. On the other hand the minimum of $E_f=-O(t^2)$ is much lower in this limit. This would suggest, incorrectly, that the magnetization is zero in the ground state. As we increase the ratio
$h/t^{15/8}$, eventually these levels cross, but there is no reason for $\tau_-$ and $\tau_f$ to be equal at this point.
This appears to be an inherent problem of using a variational ansatz which is not sufficiently complex. It could presumably be overcome by using a trial state of the form
\be
e^{-\tau\hat H_{CFT}}\,e^{-h_s\int\hat\sigma_s dx}\,|f\rangle\,,
\ee
where $\hat\sigma_s$ is the boundary magnetization coupling to a boundary magnetic field $h_s$, at the cost of loss of analytic tractability.
\subsubsection{Logarithmic anomaly.}\label{seclog}
When $h=0$ it follows from (\ref{Eaa},\ref{E2a},\ref{Ef}) that the minimum energy scales like $t^2$. Yet it has been known since Onsager that the correct behavior is $t^2\log t$. The origin of this logarithmic anomaly is a cancellation between the scaling term $t^{2/(2-\Delta_\varepsilon)}$ and the analytic background $\propto t^2$, both of which occur with amplitudes which diverge as $\Delta_\varepsilon\to1$. This may be accounted for within the variational approach by adding a counter-term as before, proportional to the space-time integral of the 2-point function, which will now also have a logarithmic dependence on the IR cutoff $\tau$. Thus, for example, (\ref{Ef}) becomes
\be
E_f=\frac\pi{48(2\tau_f)^2}-\frac t{2\tau_f}-At^2\log(\tau_f/\epsilon)\,,
\ee
where $\epsilon$ is the short-distance cutoff and $A$ is a (calculable) $O(1)$ constant. The minimum still occurs at
$\tau_f\sim t^{-1}$, but the last term now contributes the desired logarithm at the minimum.
\subsection{The tricritical Ising model.}
This corresponds to the $A_4$ lattice model with heights
$h(R)\in\{1,2,3,4\}$. The RSOS condition means that $h$ is even on even sites odd $s$ on odd sites, or vice versa.
In the Landau-Ginzburg picture it corresponds to a scalar field $\phi$ with a $\phi^6$ interaction, and a Z$_2$ symmetry under $\phi\to-\phi$.
Note that in the lattice model this Z$_2$ symmetry is implemented by reflecting the Dynkin diagram \em and \em a sublattice shift. The Kac table with bulk operators labelled by Landau-Ginzburg is shown in Fig.~2.
Note that odd $r$ is Z$_2$ odd and vice versa.
However another model in the same universality class is the spin-1 (Blume-Capel) Ising model, which may be thought of as an Ising model with vacancies.
\begin{figure}[h]\label{LG}
\centering
\includegraphics[width=0.25\textwidth]{A4LG}
\caption{Landau-Ginzburg assignment of bulk operators in the $A_4$ Kac table.}
\end{figure}
\begin{figure}[h]\label{A4}
\centering
\includegraphics[width=0.4\textwidth]{A4bcv2}
\caption{Correspondence between boundary conditions in lattice models and Kac labels of conformal boundary states: in the $A_4$ model according to Ref.~\cite{PB} (upper labels), and in the Blume-Capel model, according to Ref.~\cite{Affleck} (lower labels).}
\end{figure}
The usually accepted phase diagram and RG flows of the tricritical Ising model near the tricritical fixed point are quite complex. (See for example Fig.~4.2 of \cite{JCbook}.) In the Z$_2$-even sector, turning on the most relevant operator $\Phi_{3,3}\sim\phi^2$ gives flows either to the high-temperature disordered phase, or to the 2 coexisting low-temperature ordered phases. Turning on the $\Phi_{1,3}\sim\phi^4$ operator gives flows either to a first-order transition between these ordered phases and
a disordered phase with vacancies, or to the $A_3$ Ising fixed point.
As for the Ising model, turning on the $\Phi_{2,2}\sim\phi$ operator leads to broken-symmetry phases. However, at low temperatures there may be coexistence between two such phases with different densities of vacancies. These persist to finite temperature, giving `wings' in the phase diagram which end in lines of Ising-like transitions. These lines meet in the tricritical point and correspond to flows generated by the non-leading but relevant Z$_2$-odd operator $\Phi_{2,1}\sim\phi^3$.
According to Behrend and Pearce \cite{PB}, the labelling of the boundary states in the $A_4$ lattice model is as shown in Fig.~3. On the same diagram we have indicated their interpretation in the Blume-Capel model, due to Chim \cite{Chim} and Affleck \cite{Affleck}, which is perhaps more intuitive. Here $(\pm)$ label totally ordered states, $(0)$ is a vacancy-rich state, and $(0\pm)$ are partially ordered states. $(d)$ is a multicritical point separating these in the boundary RG flows \cite{Affleck}.
Note that the $\alpha=2$ states are Z$_2$ even while the Z$_2$ symmetry interchanges $\alpha=1$ and $\alpha=3$ (keeping $\beta$ the same.)
Let us see how well the variational approach reproduces this picture. According to the earlier analysis, turning on
the $\Phi_{2,2}$ operator corresponds to the boundary states at the corners of the Kac table in Fig.~\ref{A4}. These are the most ordered states.
Again, turning on the $\Phi_{3,3}$ perturbation corresponds to the boundary states
which extremize $\big(2\cos(\pi\alpha/2)+1\big)\big(2\cos(2\pi\beta/5)+1\big)$. This gives $\beta=1,4$, and, depending on the sign of the coupling, either $\alpha=2$ or $\alpha=1,4$. These correspond to the disordered and ordered Ising-like phases, respectively, as expected.
Turning on $\Phi_{1,3}$ corresponds to maximizing only the second factor $\big(2\cos(2\pi\beta/5)+1\big)$, and so, depending on the sign of the coupling gives either $\beta=1,4$, corresponding to coexistence between these Ising-like phases (instead of a second order critical point as it should), or $\beta=2,3$ coexistence between partially ordered phases and a disordered, vacancy-rich phase.
Turning on $\Phi_{2,1}$, on the other hand, corresponds to extremizing
$(-1)^{\alpha+\beta}\cos(\pi\alpha/4)$. For one sign of the coupling we get coexistence between the strongly ordered phase $(-)=(1,2,1,2)$ and the partially ordered phase $(0+)=(3,2/4,3,2/4)$, and for the other sign we get coexistence between their Z$_2$ partners. This is once again in general agreement with the wings of the phase diagram, except that the approximation suggests a first-order rather than an Ising-like continuous transition.
We conclude that for this model the boundary states roughly reproduce the expected RG flows when a single relevant operator is turned on, with the exception that flows to non-trivial CFTs are approximated by first-order rather than continuous transitions.
\subsection{Matrix elements between Ishibashi states and the fusion rules.}\label{sec33}
As an aside, we mention a curiosity which follows from the result (\ref{At}) for the matrix element of a bulk primary field between physical states in the limit $L\to\infty$:
\be
\langle a|e^{-\tau\hat H}\hat\Phi_je^{-\tau\hat H}|b\rangle=\delta_{ab}\left(\frac\pi{2\tau}\right)^{\Delta_j}\,
\frac{S_a^j}{S_a^0}\left(\frac{S_0^0}{S_0^j}\right)^{1/2}
\ee
and the definition of these states in terms of the Ishibashi states (\ref{phys}), which, on inverting becomes:
\be
|i\rangle\rangle=\sum_aS^i_a(S^i_0)^{1/2}\,|a\rangle\,.
\ee
Hence
\be
\langle\langle i|e^{-\tau\hat H}\hat\Phi_je^{-\tau\hat H}|k\rangle\rangle=
\left(\frac\pi{2\tau}\right)^{\Delta_j}\left(\frac{S_0^0}{S_0^j}\right)^{1/2}(S^i_0)^{1/2}(S^k_0)^{1/2}
\,\sum_a\frac{S_a^iS_a^jS_a^k}{S^0_a}\,.
\ee
We recognize the sum over $a$ as the Verlinde formula \cite{Verlinde} for the fusion rule coefficient $N_{ijk}$. Taking into account the normalization of the states, we have
\be\label{fusion}
\frac{\langle\langle i|e^{-\tau\hat H}\hat\Phi_je^{-\tau\hat H}|k\rangle\rangle}
{\big(\langle\langle i|e^{-2\tau\hat H}|i\rangle\rangle
\langle\langle k|e^{-2\tau\hat H}|k\rangle\rangle\big)^{1/2}}
=\left(\frac\pi{2\tau}\right)^{\Delta_j}\left(\frac{S_0^0}{S_0^j}\right)^{1/2}\,\,N_{ijk}\,.
\ee
Note that the first factor could be absorbed into a redefinition of the normalization of $\hat\Phi_j$.
This result is somewhat surprising, and, to our knowledge, has not been noticed before. If we insert the definition (\ref{Ishi}) of the Ishibashi states into the numerator, the leading term for $\tau\gg L$ is proportional to the OPE coefficient $c_{ijk}$, which certainly vanishes whenever $N_{ijk}$ does. The contributions of all the descendent states are all proportional to $c_{ijk}$, with coefficients which could, in principle, be computed from the Virasoro algebra. However, it is remarkable that they all conspire to sum to the integer-valued fusion rule coefficient $N_{ijk}$.
If there were an independent way of establishing (\ref{fusion}) this would give an alternative derivation of the Verlinde formula. It would be interesting to see whether this result extends to non-diagonal minimal models and to other rational CFTs. Although it has been derived here only for the Virasoro minimal models, there seems to be no obstacle in principle to its generalization to other rational CFTs, and it then suggests that the 1-point functions of bulk fields between suitable boundary states are determined by purely topological data of the fusion algebra. It also gives a possible way to \em define \em (at least ratios of) the fusion rules for non-rational CFTs.
\section{Conclusion.}\label{sec4}
We have proposed using smeared boundary states as trial variational states for massive deformations of CFTs. This is motivated by the uses of these states in quantum quenches and entanglement studies. In the case of the 2d minimal models we can perform explicit calculations which show this method gives a qualitative picture of the phase diagram in the vicinity of the CFT. Its main failing is that it cannot correctly predict a flow to a non-trivial CFT. In this case it appears to suggest phase coexistence rather than a continuous transition. In addition, the boundaries between different states corresponding to different renormalization group sinks are always first-order transitions. This is a necessary consequence of the variational method.
However the method, by its nature, always gives the correct scaling of the energy with the coupling constants. From a numerical point of view it cannot be competitive with earlier methods such as the truncated conformal space approach \cite{TCFT1,TCFT2}, but it is much simpler and moreover gives new insight into the physical relationship between conformal boundary states and ground states of gapped theories. Since it gives a bound on the universal term in the free energy, it would be interesting to make a detailed comparison with exact results available for integrable perturbations \cite{Fateev}.
\section*{Acknowledgements}
The author thanks V.~Pasquier, H.~Saleur and G.~Vidal for helpful discussions, and A.~Konechny for pointing out Ref.~\cite{Kon}.
\paragraph{Funding information.}
This work was supported in part by funds from the Simons Foundation, and by the Perimeter Institute for Theoretical Physics. Research at the Perimeter Institute is supported by the Government of Canada through the Department of Innovation, Science and Economic
Development and by the Province of Ontario through the Ministry of Research and Innovation.
| 04e770e0a29d3a1324ffedae3420a4e611a9e048 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}\label{s:intro}
Hexahedral meshes are often preferred to tetrahedral meshes as they offer
excellent numerical properties: faster assembly\,\oldcite{remacle2016gpu}, orthogonal
grids in the wall-normal direction for wall-bounded
flows, high accuracy in solid mechanics,
both for statics\,\oldcite{wang2004back} and dynamics,
or for quasi-incompressible materials\,\oldcite{benzley1995comparison}
\footnote{In many references, the
accuracy of linear hexahedra is shown to be equivalent to the
accuracy of quadratic tetrahedra with the same mesh density. Note
that quadratic tetrahedra have one extra node per edge of the mesh,
which multiplies the number of degrees of freedom by $7$.}.
Generating hex-meshes is however still an open problem for general 3D domains.
Finite element meshes should ideally fill the 3D domain in a conformal
fashion but should also respect some size and quality constraints in
order to be suitable for finite element formulations.
The validity of elements is usually the most important constraint and can be checked
by verifying the local injectivity of their mapping; in the usual finite
element language, one should check the positivity of the Jacobian determinant.
While checking the validity of a linear tetrahedron just consists in ensuring
its volume positivity, checking the validity of a linear hexahedron is not trivial.
Testing hexahedron validity is of particular interest when generating hex meshes
with an indirect method\,\oldcite{baudouin2014frontal,botella2016indirect,%
sokolov2016hexahedral}. In these methods, a huge set of hexahedral elements whose
cardinality can be as high as 40 times the number of vertices of the mesh is
computed\,\oldcite{pellerin2017}. Computing the validity robustly and
rapidly is then essential for the efficiency of these methods.
Many algorithms have been proposed in the literature for checking the validity of
hexahedra, however they do not provide
any strong guarantees except method of\,\oldcite{johnen2013geometrical}.
In this paper, we particularize this method for the linear hexahedron and propose
an efficient and simple implementation.
\paragraph{Previous works}
\citet{knupp1990invertibility} has shown that the positivity of the
Jacobian determinant at the $8$ corners of a linear hexahedron,
as well as on its edges, is not sufficient to ensure its validity.
He conjectured that any hexahedra having a positive Jacobian determinant on its
boundary is valid.
However, the Jacobian determinant on the faces are biquadratic functions;
verifying their positivity is complex and, to our knowledge, no practical
algorithm has been presented.
Some authors have proposed to check the validity by ensuring the
positivity of sets of tetrahedra constructed from the $8$ nodes of the
hexahedron\,\oldcite{ivanenko1999harmonic,grandy1999conservative,ushakova2001conditions,vavasis2003bernstein,shangyou2005subtetrabedral}.
The number of tetrahedra ranges from $8$ to $64$.
\citet{ushakova2011nondegeneracy} compiled and empirically studied these tests.
It is known that the positivity of the $8$ corner tetrahedra is a necessary
condition\,\oldcite{knupp1990invertibility,ivanenko1999harmonic}.
\citet{ushakova2011nondegeneracy} showed that none of the tests that consider less
than $58$ tetrahedral volumes constitute a sufficient condition.
The volume of the hexahedron is sometimes used in commercial
packages\,\oldcite{ushakova2001conditions}. It can be expressed from the volume of $10$
tetrahedra. It is a poor test alone but gives a sharper necessary condition when
combined with the $8$ corner tetrahedra.
Another original method for checking the validity of linear hexahedra has been
proposed by \citet{knabner2003conditions}.
The Jacobian determinant of the hexahedron is expanded into the monomial basis.
Positivity conditions are derived from the monomial coefficients of respectively
a quadratic one-dimensional polynomial, a biquadratic polynomial and a triquadratic
polynomial. The latter enables to check the positivity of the Jacobian determinant
of the hexahedron.
However, it is needed to linearize inequalities containing a square root which
implies this approach to be only a sufficient condition.
A parameter provided by the user allows to determine the precision of this linearization.
A method for checking the validity of curved finite element of any type has
been proposed by \citet{johnen2013geometrical}.
This method consists in expanding the Jacobian determinant into the B\'ezier basis
of order 2. Thanks to the convex hull property of B\'ezier expansion, the minimum of
these coefficients gives a lower bound of the Jacobian determinant. Moreover, the
minimum of specific coefficients gives an upper bound of the minimum of the
Jacobian determinant. These bounds are subsequently sharpened by ``subdividing''
in a recursive and adaptive manner which allows to compute the minimum of the
Jacobian determinant with any prescribed tolerance.
This method can be employed for the validity of the linear hexahedron since
it is a particular case of the curved hexahedron.
\paragraph{Contribution}
To the best of our knowledge, the method\,\oldcite{johnen2013geometrical} is the
only method to robustly check the validity of linear hexahedra. However,
the general framework used for curved elements is not well-adapted for an efficient
computation of the validity of one specific type of element. In this work, this
method is optimized for to the specific case of the linear hexahedron.
We start by introducing the validity of the linear quadrangle and hexahedron
(\S\ref{s:validity}), and the B\'ezier expansion of the Jacobian determinant
(\S\ref{s:bezier}). Then, two substantial improvements are presented:
we show that only $20$ quantities have to be computed
instead of $27$ (\S\ref{s:only20}) and that those quantities can be computed as the
volume of tetrahedra (\S\ref{s:tet2lag}).
Finally, we present the complete algorithm (\S\ref{s:algo}) and demonstrate that
this new algorithm is robust and efficient (\S\ref{s:results}).
The C++ code implementing the algorithm will be available in
Gmsh\,\oldcite{geuzaine2009gmsh} (\url{www.gmsh.info}).
\section{Validity of finite elements}\label{s:validity}
Let us consider a $d$-dimensional physical linear finite element which is
geometrically defined by a set of
$N$ points $\n k\in\R^d,\ k = 1, \dots , \Np$, called nodes, and a set of Lagrange shape
functions $\L{}k(\vec\xi): \Omega_{ref}\subset\R^d\to\R,\ k = 1, \dots , \Np$. These polynomial
functions allow to map a reference unit element, represented by the domain of
definition $\Omega_{ref}$, to the physical element (see Figure~\ref{f:mapping}):
\beqn{e:mapping}
\vec x(\vec\xi) = \sum_{k=1}^{\Np}\L{}{k}(\vec\xi)\,\n k.
\end{equation}
\begin{figure}[ht!]
\begin{center}
\setlength\figwidth{.5\textwidth}
\setlength\figheight{.3\textwidth}
\input{figures/tikz/mapping.tex}
\caption{Mapping between the reference and the physical hexahedron.}
\label{f:mapping}
\end{center}
\end{figure}
The Jacobian matrix of this mapping, denoted
$\jac: \Omega_{ref} \to \R^{d\times d} : \vec\xi \mapsto \jac(\vec\xi)$,
is by definition the matrix of the first-order partial derivatives of $\vec x$,
\emph{i.e.}\xspace $\Par{\jac}_{ij} = \frac{\partial\sx_i}{\partial\sxr_j}$.
Since the mapping is polynomial, each element of $\jac$ is polynomial.
To be well-defined, finite element formulations require the mapping between
the reference and any physical element to be injective\,\oldcite{frey1978some}. This
imposes to the determinant of the Jacobian matrix (the Jacobian determinant)
to be non-zero for every point of $\Omega_{ref}$\,\oldcite{zhang2005bijectivity} and
we conventionally impose it to be strictly positive.
A physical element is valid if its
Jacobian determinant is positive everywhere on the reference domain,
otherwise it is invalid. The validity of linear simplices
(\emph{i.e.}\xspace linear triangles and tetrahedra) is easy to check: since the Jacobian
determinant is constant for these elements, it is sufficient to compute it
at any point $\vec\xi\in\Omega_{ref}$ and verify that it is positive.
In practice, it is equivalent to compute the signed area of linear triangle since it is
equal to the Jacobian determinant divided by 2. Similarly, the signed volume of linear
tetrahedra is equal to the Jacobian determinant divided by 6 and can equivalently be
computed to check their validity.
The Jacobian determinant of linear quadrangles and hexahedra, on the
other hand,
is not constant over their reference domain. It is necessary to compute the
minimum of their Jacobian determinant in order to check their validity.
The two following sections are dedicated to explaining how to achieve it.
\subsection{Validity control of a linear quadrangle}
In finite element codes, the domain of definition $\Omega_{ref}$ of the quadrangular
element is taken as the domain $[-1,1]\times[-1,1]$ due to better numerical properties.
This choice has no impact on the validity criterion and we will
consider $\Omega_{ref}\equiv[0,1]\times[0,1]$ in this paper for clarity reasons.
Consequently, the Lagrange shape functions for a linear quadrangle reads:
\[
\left\{
\begin{array}{lcc}
\L{}1(\xra, \xrb) = &(1-\xra)&(1-\xrb)\\
\L{}2(\xra, \xrb) = &\xra&(1-\xrb)\\
\L{}3(\xra, \xrb) = &\xra&\xrb\\
\L{}4(\xra, \xrb) = &(1-\xra)&\phantom.\xrb.
\end{array}
\right.
\]
This implies that the mapping of a quadrangle (cf. equation~\eqref{e:mapping}),
is bilinear. Let $(x_k, y_k)$
denotes the coordinates of the node $\n k$, and let us write shortly any difference
$(x_j-x_i)$ as $\xx ij$ (and similarly for the $y$ coordinate). The partial derivative
of $x$ with respect to $\xra$ is noted $\da x$. The Jacobian matrix is given by:
\
\jac(\xra, \xrb)=\Par{
\begin{array}{cc}
\da x & \db x\\
\da y & \db y
\end{array}
}=\Par{
\begin{array}{c@{\ \ \ }c}
\xx12\,(1-\xrb)+\xx43\,\xrb & \xx14\,(1-\xra)+\xx23\,\xra\\
\yy12\,(1-\xrb)+\yy43\,\xrb & \yy14\,(1-\xra)+\yy23\,\xra
\end{array}
}
\
and the Jacobian determinant is given by:
\begin{align}\label{e:determinant2D}
\begin{split}
\detJ(\xra, \xrb) = \det\jac = &\phantom{{}+{}}\L{}1(\xra, \xrb)\,\Cro{\xx12\,\yy14-\yy12\,\xx14}
+\L{}2(\xra, \xrb)\,\Cro{\xx12\,\yy23-\yy12\,\xx23}\\
&+\L{}3(\xra, \xrb)\,\Cro{\xx43\,\yy14-\yy43\,\xx14}
+\L{}4(\xra, \xrb)\,\Cro{\xx43\,\yy23-\yy43\,\xx23}\\
= &\phantom{{}+{}}\sum_{k=1}^4\L{}k(\xra, \xrb)\,\detJ_k
\end{split}
\end{align}
where the coefficient $\detJ_k$ is the value taken by the Jacobian determinant
at corner $k$. As a consequence, the Jacobian determinant is also bilinear and its minimum
is reached at one of the four corners. The validity control of linear quadrangle thus
consists in computing the Jacobian determinant at each corner and in verifying that none
is negative. An equivalent, but computationally more expensive test would be
to compute the angles of the four corners and to check if they lie between $0\degree$
and $180\degree$.
The four quantities to compute (either the angles or the coefficients $\detJ_k$) are not
linearly independent. Indeed, concerning the angles, the existing linear relation is
that the four angles of a quadrangle sum up to $360\degree$. Now, from
equation~\eqref{e:determinant2D}, we can deduce that the Jacobian
determinant at \emph{e.g.}\xspace the first corner is equal to the third component of the vector
$\v12\times\v14$, where $\v ij=\n j-\n i=(\xx ij, \yy ij)$ is the vector that goes
from node $i$ to node $j$. But, for two vectors $\vec a$
and $\vec b$ of the $xy$-plane, it is well-known that the value of the third component
of their cross product $\vec a\times\vec b$ is equal to the signed area of the
parallelogram they span. In consequence, the Jacobian determinant at corner 1 is equal
to two times the signed area of the triangle defined by $\n 1$, $\n 2$ and $\n 4$. Let us
note $A_k$ the signed area of the triangle of corner $k$. Since the total area of the
quadrangle is equal to $A_1+A_3$ or $A_2+A_4$, we have the following relation
concerning the Jacobian determinant: $\detJ_1+\detJ_3 = \detJ_2+\detJ_4$
(see Figure~\ref{f:areas}).
\begin{figure}[ht!]
\begin{center}
\setlength\figwidth{.4\textwidth}
\input{figures/tikz/areas.tex}
\caption{The linear relationship between the areas of the triangles in a quadrangle
($A_1+A_3=A_2+A_4$) implies an equivalent linear relationship between the four
coefficients of the Jacobian determinant $\detJ_k$: $\detJ_1+\detJ_3 = \detJ_2+\detJ_4$,
where $\detJ_k$ is the value taken by the Jacobian determinant
at corner $k$.}
\label{f:areas}
\end{center}
\end{figure}
\subsection{Validity control of a linear hexahedron}
Let $(x_1(\vec\xi),\,x_2(\vec\xi),\,x_3(\vec\xi))$ be the trilinear mapping of the hexahedron.
The 3D Jacobian determinant is by definition:
\beqn{e:det3d}
\detJ(\xra,\xrb,\xrc) = \sum_{i,\,j,\,k\,=1}^3 \varepsilon_{i,\,j,\,k}
\ \ \da{\big(x_i\big)}\ \db{\big(x_j\big)}\ \dc{\big(x_k\big)}
\end{equation}
where $\varepsilon_{i,\,j,\,k}$ is the permutation symbol. We have that $\Da{x_i}$ is a
bilinear function in $\xrb$ and $\xrc$, and similarly for $\Db{x_i}$ and $\Dc{x_i}$.
This means that each term of the sum in equation~\eqref{e:det3d} is triquadratic and so
is the Jacobian determinant of the linear hexahedron. As a
consequence, the minimum of the Jacobian determinant is not
necessarily located at one of the eight corners. A more sophisticated validity test for
hexahedra would be to compute the minimum of $\detJ$ on the edges. This can be easily
implemented since the Jacobian determinant restricted to an edge is a quadratic function in
one of the reference variables. However, it has been proved in\,\oldcite{knupp1990invertibility}
that this test is not sufficient. One step further would be the ``face test'' that would
consist in computing the global minimum of a biquadratic function
(defined on a square domain) for the 6 faces of the hexahedron. However, there is,
to the best of our knowledge, no proof that it would be sufficient,
\emph{i.e.}\xspace that the global minimum cannot be exclusively located in the volume.
Currently, the only existing technique to robustly compute the validity of linear
hexahedra is the method proposed in\,\oldcite{johnen2013geometrical}. This method
computes bounds on the minimum of the Jacobian determinant that can be sharpened as
much as desired. The main drawback of the proposed algorithm is the general framework
used for curved elements that is not well-adapted for an efficient computation for
the linear hexahedron. We thus propose to adapt this method to the particular case
that concerns us.
In the next section, we introduce the B\'ezier formulation that allows to compute the
bounds and subsequently accurately compute the minimum of $\detJ$.
\section{B\'ezier expansion of hexahedra Jacobian determinant}\label{s:bezier}
Polynomial quantities can be expanded into the so-called B\'ezier basis in order to make
use of B\'ezier expansion properties. In this section, we first
introduce the B\'ezier expansion, then we derive the transformation matrix that
computes the B\'ezier coefficients from the Lagrange coefficients.
\subsection{Definition of B\'ezier expansion}
Let $\B\o k$ be the Bernstein polynomial function whose expression is:
\[
\B\o k(t) = \nchoosek\o k\ t^k\,(1-t)^{\o-k}\qquad t\in[0,1],\ \ k=0,\dots,n
\]
where $\nchoosek\o k=\frac{\o!}{k!(\o-k)!}$ is the binomial coefficient. These functions
allow to construct the hexahedral B\'ezier functions in term of the tensor product
of three Bernstein polynomials:
\beqn{e:bezDefinition}
\B\o{ijk}(\xra,\xrb,\xrc) = \B\o i(\xra)\ \B\o j(\xrb)\ \B\o k(\xrc).
\end{equation}
These functions, $\{\B\o{ijk}\}_{(0\leq i,j,k\leq\o)}$, defines the
\emph{B\'ezier basis} of the hexahedral polynomial space of order $\o$. Since the Jacobian
determinant of the linear hexahedron is a triquadratic function, it is included
in the hexahedral polynomial space of order $2$ and it can be expanded into
the B\'ezier basis of order 2. There exists thus a unique set of coefficients $b_{ijk}$
(also known as control values) such that we have:
\beqn{e:bezExpansion}
\detJ(\vec\xi) = \sum_{i,j,k=0}^2\ b_{ijk}\,\B2{ijk}(\vec\xi)
\end{equation}
where the right member of the above expression is the \emph{B\'ezier expansion} of the
Jacobian determinant. The number of coefficients is $27$ since every index can take three
values.
B\'ezier bases have the property that the basis functions are positive over
their domain of definition and sum up to~1.
This implies the well-known convex hull property which, in our case, gives
that $\min_{ijk}b_{ijk}\leq\min_{\vec\xi}\detJ$. In addition to that, some
B\'ezier coefficients are actual values of the Jacobian determinant. Those are the one
``located'' at the corners of the element. For example, we have: $b_{000}=\detJ(0,0,0)$ and
$b_{200}=\detJ(1,0,0)$.
The minimum of these corner coefficients constitutes an upper bound for
$\min_{\vec\xi}\detJ$.
In other words, the control values allow to bound the minimum of the Jacobian
determinant from below and above.
A positive lower bound implies the positivity of the Jacobian determinant and the
validity of the element. On the other hand, a negative upper bound implies that the
element is invalid. In the third and last case, when the lower bound is negative and the
upper bound is positive, nothing can be told concerning the validity of the element.
Those bounds are not necessarily sharp. However, they can
be sharpened as much as desired by ``subdividing'', \emph{i.e.}\xspace by expanding the same
function defined on a smaller domain, called a subdomain\,\oldcite{johnen2013geometrical}.
It is proven in\,\oldcite{leroy2008certificats,leroy2011certificats} that such
subdivision algorithm always stops and
that it can be used to check the positivity of a multivariate polynomials.
Moreover, the bounds converge quadratically with the size of the
subdomains\,\oldcite{cohen1985rates}. The subdivision algorithm can be implemented in a
recursive and adaptive manner making the validity check very
efficient\,\oldcite{johnen2013geometrical}.
In the following section, we explain how to compute the 27 coefficients $b_{ijk}$ of the B\'ezier expansion~\eqref{e:bezExpansion}.
\subsection{Computation of the B\'ezier coefficients}
In order to compute the 27 B\'ezier coefficients we have to write a linear
system of equations.
Let us consider a different indexing for B\'ezier coefficients and B\'ezier
functions for which the order is given in Figure~\ref{f:ordercoeff}. This permits
to gather the $27$ B\'ezier coefficients into a vector $\vec b$ for which we
have, for example, $b_1\!=\!b_{000}$, $b_2\!=\!b_{200}$ and $b_9\!=\!b_{100}$.
We will use a greek letter to refer to this new indexing.
\begin{figure}[ht!]
\begin{center}
\setlength\figwidth{.3\textwidth}
\input{figures/tikz/ordering3.tex}
\caption{Ordering of the nodes. Low order nodes are in black while high order nodes are in
gray.}
\label{f:ordercoeff}
\end{center}
\end{figure}
In the same way, $B_\i$ will refer to a certain function $B_{ijk}$ such
that to respect the order defined in Figure~\ref{f:ordercoeff}.
Let $\vec\xi_\i, \i=1,\dots,27$ be different points of the reference domain.
In practice, these points are taken as the uniformly spaced nodes of the order 2
hexahedron, which limits numerical errors.
We order them in the same way, such that we have $\vec\xi_1=(0,0,0)$, $\vec\xi_2=(1,0,0)$ and $
\vec\xi_9=(1/2,0,0)$ for example. Let $\vec c$ be the vector of the Jacobian determinant
computed at those points, \emph{i.e.}\xspace $c_\i = \detJ(\vec\xi_\i)$.
From the definition of the
B\'ezier expansion~\eqref{e:bezExpansion}, we can write the following linear system:
\begin{align*}
\detJ(\vec\xi_\i) &= \sum_{\j=1}^{27} b_\j\,B_\j(\vec\xi_\i)\qquad\forall \i\in\{1,\dots,27\}\\
\Leftrightarrow\qquad \vec c &= \vec A \vec b
\end{align*}
where $\vec A$ is a transformation matrix\footnote{
Indeed, we can expand the Jacobian determinant into the traditional Lagrange
functions of order 2 for the hexahedral element, in which case we have: $
\detJ(\vec\xi)=\sum_{j=1}^{27}c_j\,L_j(\vec\xi)$. The sets $\{L_j\}$ and $\{B_j\}$ are two
different bases of the same functional space for which $\vec c$ and $\vec b$ are the
respective coefficients of the Jacobian determinant.} in which each element
$A_{\i\j}$ is equal to $B_\j(\vec\xi_\i)$. The inverse of $\vec A$, denoted $\Tb$,
is the matrix that computes the B\'ezier coefficients from the computed values
of the Jacobian determinant, \emph{i.e.}\xspace $\vec b =\Tb\,\vec c$.
Matrix $\Tb$ is given in Table~\ref{t:Tb}.
To calculate the vector $\vec c$, one may derive
the analytical expression of the Jacobian determinant, as we did in 2D (see equation~
\eqref{e:determinant2D}). But we will see in Section~\ref{s:tet2lag} that it can
be performed by computing the volume of tetrahedra. Moreover, we show in the next
section that only a small part of $\vec c$ has to be computed.
\begin{table}
\[
\Tb = \Par{\begin{array}{cccc:cccc|cccc:cccc:cccc|cccccc|c}
\multicolumn{8}{c|}{\phantom{\Big(}\identity88\phantom{\Big(}} & \multicolumn{19}{c}{\zeros8{19}}\\\hline
-\mf12&-\mf12&0&0&0&0&0&0&\multicolumn{12}{c|}{}\\
0&-\mf12&-\mf12&0&0&0&0&0&\multicolumn{12}{c|}{}\\
0&0&-\mf12&-\mf12&0&0&0&0&\multicolumn{12}{c|}{}\\
-\mf12&0&0&-\mf12&0&0&0&0&\multicolumn{12}{c|}{}\\\cdashline{1-8}[4pt/4pt]
-\mf12&0&0&0&-\mf12&0&0&0&\multicolumn{12}{c|}{}\\
0&-\mf12&0&0&0&-\mf12&0&0&\multicolumn{12}{c|}{\multirow{2}{*}{$2\,\identity{12}{12}$}}&\multicolumn{7}{c}{\multirow{2}{*}{$\zeros{12}7$}}\\
0&0&-\mf12&0&0&0&-\mf12&0&\multicolumn{12}{c|}{}\\
0&0&0&-\mf12&0&0&0&-\mf12&\multicolumn{12}{c|}{}\\\cdashline{1-8}[4pt/4pt]
0&0&0&0&-\mf12&-\mf12&0&0&\multicolumn{12}{c|}{}\\
0&0&0&0&0&-\mf12&-\mf12&0&\multicolumn{12}{c|}{}\\
0&0&0&0&0&0&-\mf12&-\mf12&\multicolumn{12}{c|}{}\\
0&0&0&0&-\mf12&0&0&-\mf12&\multicolumn{12}{c|}{}\\\hline
\mf14&\mf14&\mf14&\mf14&0&0&0&0&-1&-1&-1&-1&0&0&0&0&0&0&0&0&&&&&&&0\\\cdashline{1-20}[4pt/4pt]
\mf14&\mf14&0&0&\mf14&\mf14&0&0&-1&0&0&0&-1&-1&0&0&-1&0&0&0&&&&&&&0\\
0&\mf14&\mf14&0&0&\mf14&\mf14&0&0&-1&0&0&0&-1&-1&0&0&-1&0&0&\multicolumn{6}{c|}{\multirow{2}{*}{$4\,\identity66$}}&0\\
0&0&\mf14&\mf14&0&0&\mf14&\mf14&0&0&-1&0&0&0&-1&-1&0&0&-1&0&\multicolumn{6}{c|}{}&0\\
\mf14&0&0&\mf14&\mf14&0&0&\mf14&0&0&0&-1&-1&0&0&-1&0&0&0&-1&&&&&&&0\\\cdashline{1-20}[4pt/4pt]
0&0&0&0&\mf14&\mf14&\mf14&\mf14&0&0&0&0&0&0&0&0&-1&-1&-1&-1&&&&&&&0\\\hline
\multicolumn{8}{c|}{\phantom{\Big(}-\mf18\,\ones18\phantom{\Big(}}&\multicolumn{12}{c|}{\mf12\,\ones1{12}}&\multicolumn{6}{c|}{~-2\,\ones16~}&8
\end{array}}
\]
\caption{Transformation matrix that computes the B\'ezier coefficients from the
sampling of the Jacobian determinant. Submatrix $\identity mm$ designate the identity
matrix of dimension $m$, submatrix $\zeros mn$ is a $m$ by $n$ matrix with only $0$ and $
\ones mn$ is a $m$ by $n$ matrix containing only $1$.}
\label{t:Tb}
\end{table}
\section{Linear dependency of the coefficients}\label{s:only20}
Like for quadrangles, B\'ezier coefficients of the hexahedral elements are not all
linearly independent. This is linked to the fact that the Taylor series expansion of the
Jacobian determinant contains only $20$ non-zero coefficients, as demonstrated
in\,\oldcite{knupp1990invertibility}. In this section, we formulate the dependency between the
coefficients through a similar reasoning. We then construct a transformation matrix between
the 20 linearly independent Jacobian determinant values and the $27$ B\'ezier coefficients.
The Jacobian determinant can be written as the triple scalar product:
\[
\detJ=\triple{\da\vec x}{\db\vec x}{\dc\vec x}.
\]
This permits to compute the derivatives of the Jacobian determinant in terms of
derivatives of the mapping. Given that the mapping is trilinear, the only
non-zero derivatives of $\vec x$ are $\da\vec x$, $\db\vec x$, $\dc\vec x$, $\d\vec x{\xra\xrb}$,
$\d\vec x{\xra\xrc}$, $\d\vec x{\xrb\xrc}$ and $\d\vec x{\xra\xrb\xrc}$.
The derivatives of $J$ can be found in\,\oldcite{knupp1990invertibility}
and result in the following observation:
\begin{observation}\label{p:derivative}
The following non-trivial high-order derivatives of the Jacobian determinant
are equal to zero: $\d\detJ{\xra\xra\xrb\xrb} = \d\detJ{\xra\xra\xrc\xrc}
= \d\detJ{\xrb\xrb\xrc\xrc} = 0$.
\end{observation}
Let us consider the monomial basis $\{M_{ijk}\}_{(0\leq i,j,k\leq2)}$, where $M_{ijk}
=M_{ijk}(\xra,\xrb,\xrc)=\xra^i\xrb^j\xrc^k$ and let us expand the Jacobian
determinant into this basis.
Let $m_{ijk}$ be the coefficients of this expansion.
Observation~\ref{p:derivative} admits the following corollary:
\begin{corollary}\label{p:monomial}
7 monomial coefficients of the Jacobian determinant are always equal to zero:
$m_{220}=m_{202}=m_{022}=m_{221}=m_{212}=m_{122}=m_{222}=0$.
\end{corollary}
Corollary~\ref{p:monomial} implies that the Jacobian determinant
space is of dimension 20 and that it is possible to obtain 7 linear relations
between the 27 B\'ezier/Lagrange coefficients. We will obtain them by writing the
expression of the monomimial coefficients in function of the B\'ezier coefficients.
Let $a_{\i\j}$ be the coefficient of monomial $\i$ in the expression of the
B\'ezier function $\j$ (whose definition is given at equation~\eqref{e:bezDefinition}).
Mathematically, we have $B_\j(\vec\xi) = \sum_{\i=1}^{27}a_{\i\j}\,M_\i(\vec\xi)$.
We can thus write:
\[
\detJ(\vec\xi) = \sum_{\j=1}^{27} b_\j\,B_\j(\vec\xi) = \sum_{\i=1}^{27}\underbrace{\Cro{\sum_{\j=1}^{27}b_\j\,a_{\i\j}}}_{\displaystyle m_\i}M_\i(\vec\xi)
\]
The linear relations between the B\'ezier coefficients are found by considering
the equations $m_\i=\sum_{\j=1}^{27}a_{\i\j}\,b_\j$ for the 7 monomial
coefficients of Corollary~\ref{p:monomial}.
This leads to the matrix given in Table~\ref{t:7LinRel} that
computes the last 7 B\'ezier coefficients in function of the first ones.
\begin{table}
\[
\vec b_{21\rightarrow27} = \Par{\begin{array}{cccc:cccc|cccc:cccc:cccc}
\mf14&\mf14&\mf14&\mf14&0&0&0&0&-\mf12&-\mf12&-\mf12&-\mf12&0&0&0&0&0&0&0&0\\\cdashline{1-20}[4pt/4pt]
\mf14&\mf14&0&0&\mf14&\mf14&0&0&-\mf12&0&0&0&-\mf12&-\mf12&0&0&-\mf12&0&0&0\\
0&\mf14&\mf14&0&0&\mf14&\mf14&0&0&-\mf12&0&0&0&-\mf12&-\mf12&0&0&-\mf12&0&0\\
0&0&\mf14&\mf14&0&0&\mf14&\mf14&0&0&-\mf12&0&0&0&-\mf12&-\mf12&0&0&-\mf12&0\\
\mf14&0&0&\mf14&\mf14&0&0&\mf14&0&0&0&-\mf12&-\mf12&0&0&-\mf12&0&0&0&-\mf12\\\cdashline{1-20}[4pt/4pt]
0&0&0&0&\mf14&\mf14&\mf14&\mf14&0&0&0&0&0&0&0&0&-\mf12&-\mf12&-\mf12&-\mf12\\\cline{1-20}
\multicolumn{8}{c|}{\phantom{\Big(}\mf14\,\ones18\phantom{\Big(}}&\multicolumn{12}{c}{-\mf14\,\ones1{12}}
\end{array}}\cdot\vec b_{20}
\]
\caption{Computation of the last 7 B\'ezier coefficients in function of the 20 first.
}
\label{t:7LinRel}
\end{table}
Let us write $\vec D$ the matrix that computes the 27 B\'ezier coefficients from
the first 20 B\'ezier coefficients. Matrix $\vec D$ is constructed by extending
the matrix given in Table~\ref{t:7LinRel} with an identity matrix of size 20.
We have:
\[
\vec b = \vec D\,\vec b_{20}
\]
where $\vec b_{20}$ is the vector containing the first 20 components of $\vec b$.
Constructing the matrix that computes the 27 B\'ezier coefficients in function of
20 Lagrange coefficients is now straightforward. Matrix $\Tb$ (see Table~\ref{t:Tb})
is such that the first 20 B\'ezier coefficients depends only on the first 20 Lagrange
coefficients. Let $\Tsub$ be the $20\times20$ upper left submatrix of $\Tb$ and $\vec
c_{20}$ the first 20 components of $\vec c$. We have that:
\[
\vec b_{20}=\Tsub\,\vec c_{20}\quad\Leftrightarrow\quad\vec b =
\vec D\,\Tsub\vec c_{20} = \Tsmall\vec c_{20}
\]
where $\Tsmall$, the matrix that computes all the B\'ezier coefficients from
the first 20 Lagrange coefficients, is given in Table~\ref{t:Tsmall}.
\begin{table}
\[
\Tsmall = \Par{\begin{array}{cccc:cccc|cccc:cccc:cccc}
\multicolumn{8}{c|}{\phantom{\Big(}\identity88\phantom{\Big(}} & \multicolumn{12}{c}{\zeros8{12}}\\\hline
-\mf12&-\mf12&0&0&0&0&0&0&\\
0&-\mf12&-\mf12&0&0&0&0&0&\\
0&0&-\mf12&-\mf12&0&0&0&0&\\
-\mf12&0&0&-\mf12&0&0&0&0&\\\cdashline{1-8}[4pt/4pt]
-\mf12&0&0&0&-\mf12&0&0&0&\\
0&-\mf12&0&0&0&-\mf12&0&0&\multicolumn{12}{c}{\multirow{2}{*}{$2\,\identity{12}{12}$}}\\
0&0&-\mf12&0&0&0&-\mf12&0&\\
0&0&0&-\mf12&0&0&0&-\mf12&\\\cdashline{1-8}[4pt/4pt]
0&0&0&0&-\mf12&-\mf12&0&0&\\
0&0&0&0&0&-\mf12&-\mf12&0&\\
0&0&0&0&0&0&-\mf12&-\mf12&\\
0&0&0&0&-\mf12&0&0&-\mf12&\\\hline
-\mf34&-\mf34&-\mf34&-\mf34&0&0&0&0&~1~&~1~&~1~&~1~&~0~&~0~&~0~&~0~&~0~&~0~&~0~&~0~\\\cdashline{1-20}[4pt/4pt]
-\mf34&-\mf34&0&0&-\mf34&-\mf34&0&0&1&0&0&0&1&1&0&0&1&0&0&0\\
0&-\mf34&-\mf34&0&0&-\mf34&-\mf34&0&0&1&0&0&0&1&1&0&0&1&0&0\\
0&0&-\mf34&-\mf34&0&0&-\mf34&-\mf34&0&0&1&0&0&0&1&1&0&0&1&0\\
-\mf34&0&0&-\mf34&-\mf34&0&0&-\mf34&0&0&0&1&1&0&0&1&0&0&0&1\\\cdashline{1-20}[4pt/4pt]
0&0&0&0&-\mf34&-\mf34&-\mf34&-\mf34&0&0&0&0&0&0&0&0&1&1&1&1\\\hline
\multicolumn{8}{c|}{\phantom{\Big(}-\mf58\,\ones18\phantom{\Big(}}&\multicolumn{12}{c}{\mf12\,\ones1{12}}
\end{array}}
\]
\caption{Transformation matrix that allows to compute the B\'ezier coefficients from 20
samplings of the Jacobian determinant.
matrix of dimension $m$, submatrix $\zeros mn$ is a $m$ by $n$ matrix with only $0$ and $
\ones mn$ is a $m$ by $n$ matrix containing only $1$.
}
\label{t:Tsmall}
\end{table}
\section{Expression of the 20 Lagrange coefficients in function of 20 tetrahedral volumes}
\label{s:tet2lag}
In this section we show that the 20 Lagrange coefficients that has to be computed are
equal to the volume of tetrahedra.
Recalling that $\vec x$ is the column vector $(x, y, z)^{\text T}$, the Jacobian matrix can be
written as:
\[
\jac(\xra,\xrb,\xrc)=\Par{
\,\Cro{\begin{array}{c}\\\da\vec x\\~\end{array}}\,
\Cro{\begin{array}{c}\\\db\vec x\\~\end{array}}\,
\Cro{\begin{array}{c}\\\dc\vec x\\~\end{array}}\,
},
\]
Let us recall that $\v\i\j$ denotes the difference $(\n\j-\n\i)$.
We can express the derivatives of $\vec x$ from the definition of the
mapping~\eqref{e:mapping} and the Lagrange functions given in~\ref{s:lagrangeFunction3D}:
\[
\left\{
\begin{array}{l}
\da\vec x = \v12\ (1-\xrb)\,(1-\xrc) + \v43\ \xrb\,(1-\xrc) + \v56\ (1-\xrb)\,\xrc + \v87\ \xrb\,\xrc\\
\db\vec x = \v14\ (1-\xra)\,(1-\xrc) + \v23\ \xra\,(1-\xrc) + \v58\ (1-\xra)\,\xrc + \v67\ \xra\,\xrc\\
\dc\vec x = \v15\ (1-\xra)\,(1-\xrb) + \v26\ \xra\,(1-\xrb) + \v48\ (1-\xra)\,\xrb + \v37\ \xra\,\xrb.
\end{array}
\right.
\]
In the following, $\det{\vec a, \vec b, \vec c}$ will denote the determinant of the matrix
made up of columns $\vec a$, $\vec b$ and $\vec c$. Note that $\det{\vec a, \vec b, \vec c}
$ equals $\triple{\vec a}{\vec b}{\vec c}$ and is a trilinear function. Moreover, if the
three vectors have the same origin, then the determinant is also 6 times the volume of the
tetrahedron that the vectors define. Lastly, if the three vectors are not
linearly independent, then the determinant is zero.
There are two types of Lagrange coefficients we are interested in: the coefficients that
correspond to the corners and the coefficients that correspond to the edges of the
hexahedron. By symmetry of the problem, there must be also two types of tetrahedra to
identify. It is already well-known that the Jacobian determinant computed at a corner
corresponds to 6 times the volume of the tetrahedron constructed from the 3 edges of the
corner. Let us formulate it mathematically for the first corner:
\[
\detJ_1 = \det{\jac(0,0,0)} = \det{\v12,\,\v14,\,\v15} = 6\ \text{vol}\,(\n1,\,\n2,\,\n4,\,\n5)
\]
where $\vol(\cdot)$ refer to the volume of the tetrahedron defined by the four nodes.
In a similar manner, we can express the 9th value of the Jacobian determinant as the volume
of a tetrahedron:
\begin{align*}
\detJ_9 = \det{\jac(1/2,0,0)} &= \det{\v12,\ \frac{\v14+\v23}2,\ \frac{\v15+\v26}2} \\
&= \det{\v12,\ \frac{\n4+\n3}2-\frac{\n1+\n2}2,\ \frac{\n5+\n6}2-\frac{\n1+\n2}2} \\
&= \det{\v12,\ \Cro{\frac{\n4+\n3}2-\n1}+\Cro{\n1-\frac{\n1+\n2}2},\ \Cro{\frac{\n5+\n6}2-\n1}+\Cro{\n1-\frac{\n1+\n2}2}}
\end{align*}
where the terms $\Cro{\n1-\frac{\n1+\n2}2}$ are equal to $-\frac{\v12}2$. By trilinearity
of the determinant and dependency with respect to the first vector ($\v12$), the terms
$-\frac{\v12}2$ vanish and we obtain:
\[
\detJ_9 = 6\ \text{vol}\,\Par{\n1,\ \n2,\ \frac{\n4+\n3}2,\ \frac{\n5+\n6}2}
\]
Figure~\ref{f:tet} shows the tetrahedra that correspond to four value of the Jacobian
determinant.
\begin{figure}[ht!]
\begin{center}
\setlength\figwidth{.2\textwidth}
\input{figures/tikz/tetrahedron1.tex}
\input{figures/tikz/tetrahedron9.tex}
\input{figures/tikz/tetrahedron2.tex}
\input{figures/tikz/tetrahedron10.tex}
\caption{Different tetrahedra whose volume corresponds to the value of the respective
coefficients $\detJ_1$, $\detJ_9$, $\detJ_2$ and $\detJ_{10}$ divided by $6$.}
\label{f:tet}
\end{center}
\end{figure}
\section{The algorithm}\label{s:algo}
The algorithm that computes the validity of a linear hexahedron takes as input
the 8 nodes coordinates of the element.
It returns true if the element is valid and return false if the element is invalid.
The execution is the following:
\begin{enumerate}
\item Compute the 20 volumes of Section~\ref{s:tet2lag} and put them in vector $\vec v$
(ordering them as in Figure~\ref{f:ordercoeff}).
\item If at least one volume is negative, return False.
\item Compute the B\'ezier coefficients $\vec b = \Tsmall\vec v$ where $\Tsmall$
is the matrix given in Table~\ref{t:Tsmall}.
\item If all the B\'ezier coefficients in $\vec b_{9\to27}$ are positive, return True.
\item Return \texttt{recursive\_subdivision}($\vec b$).
\end{enumerate}
In Step~4, the 8 first B\'ezier coefficients are equal to the volume of the corner
tetrahedra and must be positive otherwise the algorithm would have stop at Step~2.
The subdivision algorithm, \texttt{recursive\_subdivision}($\vec b$), is identical
to the subdivision algorithm presented in paper\,\oldcite{johnen2013geometrical}
(although implemented in a more efficient manner in our new implementation).
It takes a vector of 27 B\'ezier coefficients as input and return true if the
Jacobian determinant is strictly positive on the subdomain, otherwise it returns
false. The algorithm is:
\begin{enumerate}
\item Subdivide: Compute the subcoefficients $\vec b^i,\ i=1,\dots,8$ as described
in paper\,\oldcite{johnen2013geometrical}.
\item For each $\vec b^i$:
\item \ \ If at least one of the coefficients in $\vec b^i_{1\to8}$ is negative,
return False.
\item \ \ If all the coefficients in $\vec b^i_{9\to27}$ are positive, continue the loop.
\item \ \ If \texttt{recursive\_subdivision}($\vec b^i$) is false, return False.
\item Return True.
\end{enumerate}
In Step~3 of this algorithm, it is checked if the 8 first B\'ezier coefficients are
not negative since they are actual values of the Jacobian determinant. In Step~4,
the positivity of the 19 other coefficients ensures that the Jacobian determinant is
positive on the corresponding subdomain in which case the algorithm skip Step~5 and
continue the loop. While there is no negative real value of the Jacobian
determinant but at least one negative B\'ezier coefficients, the algorithm
subdivide (Step~5).
\section{Results}\label{s:results}
We begin the results with unitary tests. The Jacobian determinant of the hexahedron
defined in Figure~\ref{f:invalidButUshakovaValid} is positive at the 8 corners, the
center of the edges, the center of the faces and the center of the volume.
Moreover, the hexahedron passes the Ushakova's\,\oldcite{ushakova2011nondegeneracy} test6
that requires the computation of $24$ tetrahedral volumes.
Our algorithm detects that this hexahedron is invalid.
\begin{figure}[htp]
\begin{center}
\begin{tikzpicture}
\node[anchor=east] at (0, 0) {
\begin{tabular}{c|ccc}
i & x & y & z\\
\hline
0 & 0 & 0 & 0 \\
1 & 1 & 0 & 0 \\
2 & 1.7615170641459 & 0.594764272968121 & 0.15552188663289 \\
3 & 0.438888411629833 & 1.53098020041072 & 0.185631029838277 \\
4 & 1.3859049651391 & 0.0755794018509022 & 1.77483024073906 \\
5 & 1.22129676447071 & 0.271876165350328 & 0.630922158503566 \\
6 & 1.77365642274365 & 1.25103990471942 & 1.83300604452892 \\
7 & 0.0769922201302364 & 0.940424880836765 & 1.45521546591891 \\
\end{tabular}
};
\node[anchor=west] at (0, 0) {
\includegraphics[width=.3\textwidth]{hex8}
};
\end{tikzpicture}
\end{center}
\caption{Invalid hexahedron for which the Jacobian determinant is positive at the
$27$ nodes of the second-order hexahedron and for which the 24 tetrahedral volumes of
Ushakova's\,\oldcite{ushakova2011nondegeneracy} test6 are all positive.}
\label{f:invalidButUshakovaValid}
\end{figure}
Figure~\ref{f:validNegativeOshakova} presents a hexahedron that does not pass
Ushakova's\,\oldcite{ushakova2011nondegeneracy} test6, despite the fact that the
element is valid.
\begin{figure}[htp]
\begin{center}
\begin{tikzpicture}
\node[anchor=east] at (0, 0) {
\begin{tabular}{c|ccc}
i & x & y & z\\
\hline
0 & 0 &0& 0\\
1 &1 &0& 0 \\
2 &1.539& 0.704696 &1.84011\\
3 & 0.166589 &1.08208 &0.162539\\
4 & 0.0501127& 1.96347 &1.56559\\
5 & 0.422336 &0.00419138& 1.43038\\
6 & 0.509917 & 0.0214216 &1.55322\\
7 & 0.40783 & 1.73452 & 1.93234\\
\end{tabular}
};
\node[anchor=west] at (0, 0) {
\includegraphics[width=.3\textwidth]{hex6}
};
\end{tikzpicture}
\end{center}
\caption{Valid hexahedron that does not pass Ushakova's\,\oldcite{ushakova2011nondegeneracy} test6.}
\label{f:validNegativeOshakova}
\end{figure}
In hexahedral mesh community, it is common to measure the quality of hexahedra by
computing the minimum of the ``scaled Jacobian'' on the 8
corners\,\oldcite{knupp2000achievingB,yamakawa2003fully}.
For the hexahedron of Figure~\ref{f:invalidHighCornerQual}, this quality measure
is equal to 0.64 although the element is invalid. This demonstrates that even
invalid hexahedra can have a good quality at the corners.
\begin{figure}[htp]
\begin{center}
\begin{tikzpicture}
\node[anchor=east] at (0, 0) {
\begin{tabular}{c|ccc}
i & x & y & z\\
\hline
1 & 0.464949491866817 & 0.358989226966155 & 0.0133365886410108\\
2 & 0.481795709097567 & 0.358745078890347 & 0.0163884395886105\\
3 & 0.482406079287087 & 0.351664784691916 & 0.0235297708059938\\
4 & 0.466719565416425 & 0.339945677053133 & 0.0278023621326335\\
5 & 0.465498825037385 & 0.320291756950591 & -0.00277718436231578\\
6 & 0.465987121189001 & 0.321085238196966 & -0.0042420728171636\\
7 & 0.501998962370677 & 0.322367015594958 & -0.0116275521103549\\
8 & 0.487166966765343 & 0.308816797387616 & 0.0115054780724508
\end{tabular}
};
\node[anchor=west] at (0, 0) {
\includegraphics[width=.3\textwidth]{hex1}
};
\end{tikzpicture}
\end{center}
\caption{Invalid hexahedron for which the minimum of the scaled Jacobian computed at the
corners is equal to 0.64.}
\label{f:invalidHighCornerQual}
\end{figure}
For the next experimentation, we compare our method with some previous methods
on different datasets. The results are given in Table~\ref{t:comparison}.
The datasets have been generated by the algorithm
described in\,\oldcite{pellerin2017} which takes a tetrahedral mesh as input and computes
hexahedra that can be created by combining tetrahedra. This algorithm can generate a
large amount of hexahedra of different qualtity.
We have considered two models. The first one, ``Fusee\_1'',
contains $71,947$ vertices and $349,893$ tetrahedra. The second one is ``FT47'' and
contains $370,401$ vertices and $2,085,394$ tetrahedra. Both of them are available
on the website \url{www.hextreme.eu}.
We have disabled the validity check during the hexahedra creation and, for each model,
we have generated three datasets of hexahedra by varying the desired minimal quality $q$
(computed at the corners). Datasets that correspond to $q=-1$ contain a large proportion
of invalid hexahedra while datasets that correspond to $q=0.5$ contain only valid
hexahedra. We have compared our new implementation with the previous
one\,\oldcite{johnen2013geometrical}, as well as the 5 first validity tests presented
in\,\oldcite{ushakova2011nondegeneracy}. These tests consist in computing the volume of
respectively 8, 10, 24, 32 and 58 tetrahedra and returning False as soon as a
negative volume is found or returning True if no negative volume is obtained.
For each algorithm we store the execution time as well as the
number of false valid (the number of invalid hexahedra that pass the test) and the
number of false invalid (the number of valid hexahedra that do not pass the test).
The experimentation has been conducted in serial on a MacBook Pro 2016 @ 2.9 GHz.
Our new implementation detects the same invalid hexahedra than our previous
implementation. We have taken this result as the reference for computing the false
invalid and false valid elements of the methods from\,\oldcite{ushakova2011nondegeneracy}.
Test~1 computes the volume of corner tetrahedra, which corresponds to a necessary
condition. As expected, Test~1 misses invalid elements but never finds false invalid.
Test~2 to Test~4 are neither sufficient nor necessary. Test~4 misses very few
invalid hexahedra, however.
Test~5 corresponds to a sufficient condition and can miss as much as 80\% of
valid elements (see dataset Fusee\_1, $q=-1$).
Our new implementation is about $15$ to $30$ time faster than the algorithm
designed for curvilinear elements and runs at similar speed than Test~5
of\,\oldcite{ushakova2011nondegeneracy} which consists in computing $58$ tetrahedral
volumes. Our new algorithm can check the validity of hexahedra at a rate of
between $6$ million and $12$ million hexahedra per second on a single core.
The speed is higher when there is a large proportion of invalid
hexahedra since the algorithm can stop at an early stage if a negative
Jacobian determinant is obtained.
\begin{table}
\renewcommand{\arraystretch}{1.}
\setlength{\tabcolsep}{3pt}
\centering
\scriptsize
\caption{Comparison of our new implementation with some previous methods. The datasets
differ in the number of hexahedra and proportion of invalid element amongst them.
For each method, the computation time, the number of false valid and the number
of false invalid are given. Numbers are given with 3 significant digits.}
\begin{tabular}{lrr@{\qquad}l@{\quad}rrrrrrr}
\toprule
Dataset & \# hex & \# invalid & & Ours & \citet{johnen2013geometrical}
& Test 1\,\oldcite{ushakova2011nondegeneracy} & Test 2\,\oldcite{ushakova2011nondegeneracy}
& Test 3\,\oldcite{ushakova2011nondegeneracy} & Test 4\,\oldcite{ushakova2011nondegeneracy}
& Test 5\,\oldcite{ushakova2011nondegeneracy} \\
\midrule
Fusee\_1, $q=0.5$ & $334,000$ & $0$ & \# false valid & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$\\
& & & \# false invalid & $0$ & $0$ & $0$ & $529$ & $10,000$ & $10,000$ & $111,000$\\[.5ex]
& & & time [s] & $0.0565$ & $0.812$ & $0.0180$ & $0.0168$ & $0.0349$ & $0.0549$ & $0.0866$\\[2ex]
Fusee\_1, $q=0$ & $2,040,000$ & $79,900$ & \# false valid & $0$ & $0$ & $79,900$ & $73,900$ & $0$ & $0$ & $0$\\
& & & \# false invalid & $0$ & $0$ & $0$ & $10,200$ & $814,000$ & $814,000$ & $1,590,000$\\[.5ex]
& & & time [s] & $0.339$ & $6.60$ & $0.0945$ & $0.0977$ & $0.161$ & $0.220$ & $0.358$\\[2ex]
Fusee\_1, $q=-1$ & $6,060,000$ & $4,110,000$ & \# false valid & $0$ & $0$ & $80,000$ & $74,000$ & $48,400$ & $0$ & $0$\\
& & & \# false invalid & $0$ & $0$ & $0$ & $10,200$ & $814,000$ & $814,000$ & $1,590,000$\\[.5ex]
& & & time [s] & $0.488$ & $15.5$ & $0.202$ & $0.212$ & $0.301$ & $0.347$ & $0.418$\\[2ex]
FT47, $q=0.5$ & $3,000,000$ & $0$ & \# false valid & $0$ & $0$ & $0$ & $0$ & $0$ & $0$ & $0$\\
& & & \# false invalid & $0$ & $0$ & $0$ & $1,890$ & $115,000$ & $115,000$ & $1,060,000$\\[.5ex]
& & & time [s] & $0.459$ & $6.93$ & $0.181$ & $0.203$ & $0.341$ & $0.463$ & $0.725$\\[2ex]
FT47, $q=0$ & $14,700,000$ & $366,000$ & \# false valid & $0$ & $0$ & $366,000$ & $342,000$ & $7$ & $7$ & $0$\\
& & & \# false invalid & $0$ & $0$ & $0$ & $38,900$ & $4,880,000$ & $4,880,000$ & $11,100,000$\\[.5ex]
& & & time [s] & $2.43$ & $42.5$ & $0.712$ & $0.872$ & $1.30$ & $1.73$ & $2.33$\\[2ex]
FT47, $q=-1$ & $40,500,000$ & $26,100,000$ & \# false valid & $0$ & $0$ & $370,000$ & $346,000$ & $247,000$ & $7$ & $0$\\
& & & \# false invalid & $0$ & $0$ & $0$ & $38,900$ & $4,880,000$ & $4,880,000$ & $11,100,000$\\[.5ex]
& & & time [s] & $3.17$ & $102$ & $1.55$ & $1.54$ & $2.08$ & $2.49$ & $3.09$\\
\bottomrule
\end{tabular}%
\label{t:comparison}%
\end{table}%
\section{Conclusion}\label{s:conclusion}
Our implementation is able to check the validity of linear hexahedral elements
in a very efficient manner. The algorithm benefit from the robustness of the
previous method for checking the validity of curvilinear
elements\,\oldcite{johnen2013geometrical} on which it is based.
The novelty consists of two improvements: (1) a reduced number of quantities
to be computed at the beginning of the algorithm and (2) the computation of those
quantities as tetrahedral volumes instead of the Jacobian determinant. The
particularization to hexahedra also permits a fine-tuned implementation.
Our new code runs more than 15 time faster than the previous code for curvilinear
elements and runs at similar speed than the sufficient but not necessary method
presented in\,\oldcite{ushakova2011nondegeneracy}.
More than $6$ million hexahedra per second can be analyzed on a single core of
a personal computer.
The algorithm is simple and can readily be implemented from the information
given in this paper. The C++ code will be available in Gmsh
(\url{www.gmsh.info}).
\section*{Acknowledgements}
This research project was funded by the European Research Council
(project HEXTREME, ERC-2015-AdG-694020) and the TILDA project. The TILDA (Towards
Industrial LES/DNS in Aeronautics - Paving the Way for Future Accurate CFD)
project has received funding from the European Unionβs Horizon 2020 research
and innovation program under grant agreement No 635962.
The project is a collaboration between NUMECA, DLR, ONERA, DASSAULT, SAFRAN, CERFACS,
CENAERO, UCL, UNIBG, ICL and TsAGI.
| ab0fb62247e7bcc8f470a7906f52fcb1cf0253fa | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\IEEEPARstart{M}{achine} learning techniques play a central role in data analysis, data fusion and visualization. As geophysical acquisition tools become more sophisticated and gather more information, data analysts relay more on machine learning techniques for
generating meaningful representations of the data. A coherent representation of complex data often includes a feature extraction step followed by a dimensionality reduction step, which results in a compact and visual model.
Analysis tasks such as clustering, classification, anomaly detection or regression may be carried out in the constructed low-dimensional space. Common dimensionality reduction methods such as Principal Component Analysis (PCA) \cite{Pca1} and Linear Discriminant Analysis (LDA) \cite{LDA_book} project the feature space into a low dimensional space by constructing meaningful coordinated that are linear combinations of the original feature vectors.
PCA is widely used for low-dimensional modeling of geoscience datasets. Jones \& Christopher \cite{Jones} applied PCA to infer aerosol specification for research of oceans or more complex land surfaces. Griparis and Faur \cite{Griparis} applied a linear dimensionality reduction tool, Linear Discriminant Analysis (LDA) for a projection of earth observations into a low-dimensional space. Their low-dimensional representation resulted in a cluster organization of the image data by land types. PCA and Self organization maps \cite{Kohonen} were applied for pattern recognition in volcano seismic spectra by Unglert et. al. \cite{Unglert} and for geologic pattern recognition by Roden et. al. \cite{Roden}.
Another key issue in processing large amounts of data is the ability to fuse data from different sensors. Typical seismometers record data using three channels. These three channels capture the motion in the horizontal and perpendicular directions to the earth. Each channel may be processed separately and the results can be combined. Alternatively, a fused representation may be formed for common analysis. Recent advances in machine learning and in particular the use of non-linear kernel-based algorithm enable to construct data-driven fusions and to compute geometry-preserving low-dimensional embeddings. Such kernel-based embedding techniques are known as manifold learning methods, among them Local Linear Embedding \cite{LLE}, Lapacian Eigenmaps \cite{Belkin1} and Diffusion Maps (DM) \cite{Lafon}. Manifold learning methods overcome limitations of linear dimensionality reduction tools such as PCA and LDA \cite{manifold_book}. When the relationship between the original high-dimensional points is complex and non-linear, linear projections may fail to organize the data in a way that is loyal to the intrinsic physical parameters that drives the observed phenomena.
This work focuses on extending manifold learning techniques for low-dimensional modeling and kernel based data-driven fusion of seismic data. Identifying the characteristic of seismic events is a challenging and important task. This includes the discrimination between earthquakes and explosions which is not only an essential component of nuclear test monitoring but it is also important for the maintaining the quality of earthquake catalogs. For example, wrong classification of explosions as earthquakes may cause the erroneous estimation of seismicity hazard. The discrimination task is typically performed based on some extracted seismic parameters. Among such parameters is the focal depth, the ratio between surface wave magnitude and body wave magnitude and the spectral ratio between different seismic phases \cite{Blandford}, \cite{rodgers1997comparison}.
Discrimination methods based on seismic parameters give only a partial solution to the problem. For instance, a larger half of seismic events reported by the Comprehensive Nuclear-Test-Ban Treaty Organization (CTBTO) are not βscreened outβ as natural events or even are not considered for the discrimination at all although most of those events are typically earthquakes \cite{BenHorin}.
Recently, this problem and other geophysical challenges have been approached using machine
learning frameworks. Hidden Markov model were proposed in \cite{Ohrnberger}, \cite{Beyreuther}, \cite{Hammer} and modeled the data in an unsupervised manner. Artificial neural networks \cite{Tiira}, \cite{DelPezzo}, \cite{esposito2006automatic} or support vector machines \cite{Kortstrom,ruano2014seismic} were also used to construct a classifier in a supervised manner. The study in \cite{Kuyuk} utilizes Self Organization Maps to distinguish micro-earthquakes from quarry blasts in the vicinity of Istanbul, Turkey. Manifold learning is used in \cite{ramirez2011machine} for seismic phase classification. In \cite{mishne2015graph} a graph is used to detect sea mines in side-scan sonar images. The DM method is used in \cite{fernandez2015diffusion} for visualization of meteorological data. A non-linear dimensionality reduction is proposed in \cite{rabin2016earthquake} to discriminate between earthquakes and explosions.
In this study, the manifold learning approach that was presented in \cite{rabin2016earthquake} is extended by using a kernel-based fusion method for identification of seismic events. The method is model-free, and it is based on signal processing for feature extraction followed by manifold learning techniques for embedding the data. Furthermore, the method reviles the underlying intrinsic physical properties of the data, which results in a natural organization of the events by type. Since seismic data is recorded at multiple channels, we suggest fusing the information to extract a more reliable representation for the seismic recordings. The fusion framework is based on a recent work by \cite{lindenbaum,lindenbaum2015learning}. The study extends Diffusion Maps (DM) \cite{Lafon}, which has been successfully applied for phase classification \cite{ramirez2011machine}, for estimation of arrival times \cite{taylor2011estimation} and for events discrimination \cite{Poster}. Other constructions for fusing kernels were proposed in \cite{salhov2016multi,lederman,michaeli}.
The proposed framework begins with a preprocessing stage in which a time-frequency representation is extracted from each seismic event. The training phase includes the construction of a normalized graph that holds the local connections between the seismic events. A low dimensional map is then obtained by the eigen-decomposition of the graph. The constructed embedding is distance preserving. Thus the geometry of the dataset is kept in the new embedding coordinates. By utilizing the low dimensional embedding, we demonstrate capabilities of classification, location estimation and anomaly detection of seismic events.
The paper is organized as follows: Sections \ref{sec:Manifold} and \ref{SecMulti} present the machine learning frameworks for manifold learning and data fusion. In Section \ref{sec:Data} the data set is described. The mathematical methods required for analysis of seismic data are provided in Section \ref{sec:Pre}. The proposed framework and experimental results are presented in Section \ref{sec:Exp}. We conclude this work in Section \ref{sec:Future}.
\section{Manifold Learning }
\label{sec:Manifold}
This section reviews the manifold learning method that is applied in this work for non-linear dimensionality reduction, diffusion maps. The method's main ingredient is a kernel function. Here, radial basis kernel functions are used, their construction is described in detail.
\subsection{Radial Basis Kernel Function}
Kernel functions are vastly utilized in machine learning. Classification, clustering and manifold learning use some affinity measure to learn the relations among data points. A kernel is a pre-defined similarity function designed to capture the fundamental structure of a high dimensional data set.
Given a high dimensional data set \begin{math} {\myvec{X} = \lbrace
{ { \myvec{x}_1,\myvec{x}_2, \myvec{x}_3,...,\myvec{x}_M }} \rbrace
},\myvec{x}_i \in {\mathbb{R}^{D}},
\end{math} a kernel ${{\cal{K}} : \mymat{X}\times{\mymat{X}}\longrightarrow{\mathbb{R}} }$ is an affinity function over all pairs of points in $\mymat{X}$. The discrete kernel is represented by a matrix $\mymat{K}$ with following properties
\begin{itemize}
\item {Symmetry \begin{math}{K_{i,j}={\cal{K}}(\myvec{x}_i,\myvec{x}_j)={\cal{K}}(\myvec{x}_j,\myvec{x}_i) }
\end{math}}
\item {Positive semi-definiteness: \begin{math}{ \myvec{v}_i^T \mymat{K} \myvec{v}_i \geq 0 }\end{math} for all $\myvec{v}_i \in
\mathbb{R}^M$ and \begin{math}{{\cal{K}}(\myvec{x}_i,\myvec{x}_j)
\geq 0. }
\end{math}}
\end{itemize}
These properties guarantee that the matrix $\mymat{K}$ has
real eigenvectors and non-negative real eigenvalues.
In this study radial basis functions (RBF) are used for constructing the kernel. The RBF kernel function is defined by
\begin{equation}\label{eq:RBF}
{ K_{i,j}=exp\{
{-\frac{||\myvec{x}_i-\myvec{x}_j||^2}{2 \sigma^2}\}} }.\end{equation} Applying the Euclidean distance to high dimensional pairs of distant vectors could somewhat be misleading, as data is typically sparse in the high-dimensional space. For this reason the decaying property of the Gaussian kernel is beneficial. The Gaussian tends to zero for distant points, whereas its value is close to one for adjacent points.
\subsection{Setting the Kernel's Bandwidth}
The kernel's bandwidth $\sigma$ controls the number of points taken into consideration by the kernel.
A simple choice for \begin{math} \sigma \end{math} is based on the standard deviation of the data. This approach is good when the data is sampled from a uniform distribution. In this study, we use a max-min measure. The method was proposed in \cite{Keller} and aims to find a small scale to maintain local connectivities. The scale is set to
\begin{equation} \label{eq:MaxMin}
\sigma^2_{\text{MaxMin}}={\cal{C}}\cdot \underset{j}{\max} [ \underset{i,i\neq j}{\min} (||\myvec{x}_i-\myvec{x}_j||^2)],
\end{equation}
where ${\cal{C}} \in [2,3]$. Alternative methods such as \cite{Singer,lindenbaum2015musical} have demonstrated similar results in our experiments.
\subsection{Non-Linear Dimensionality Reduction}
\label{SecDiff}
Most dimensionality reduction methods are unsupervised frameworks that seek for a low dimensional representation of complex, high dimensional data sets. Each method preserves a certain criteria while reducing the dimension of the data. Principal component analysis (PCA) \cite{PCA}, reduces the dimension of the data while preserving most of the variance. Non linear methods such as Local Linear Embedding \cite{LLE}, Laplacian Eigenmaps
\cite{Luo}, Diffusion Maps (DM) \cite{Lafon} preserve the local structure of the high-dimensional data. In particular, in DM \cite{Lafon}, a metric that describes the intrinsic connectivity between the data points is defined. This metric is preserved in the low-dimensional space, resulting in a distance-preserving embedding. The metric is refereed to as diffusion distance, it is defined later in this subsection.
The DM framework enforces a fictitious random walk on the graph of a high dimensional data set $\myvec{X}=\{\myvec{x}_1,..,\myvec{x}_M \},\myvec{x}_i \in \mathbb{R}^D$. This results in a Markovian process that travels in the high-dimensional space only in areas where the sampled data exists. The method has been demonstrated useful when applied to audio signals \cite{lindenbaum2015musical}, image editing \cite{farbman2010diffusion}, medical data analysis \cite{haghverdi2015diffusion} and other types of data sets.
Reducing the dimension of a data set by construction of DM coordinates is performed using the following steps
\begin{enumerate}
\item Given a data set $\myvec{X}$ compute an RBF kernel $\myvec{K}$ based on Eq. \ref{eq:RBF}.
\item Normalize the kernel using $\mymat{D}$ where \begin{math} D_{i,i}=\underset{j}{\sum}{K_{ij}} \end{math}.
Construct the row stochastic matrix $\mymat{P}$ by \begin{equation}
{P_{i,j}\defeq{\cal{P}}(\myvec{x}_i,\myvec{x}_j)\defeq[{{\mymat{D}}^{-1}{\mymat{K}} }}]_{i,j}
\label{EquationPDM}
.\end{equation}
\item Compute the spectral decomposition of the matrix \begin{math} \mymat{P} \end{math} to obtain a sequence of eigenvalues \begin{math}{\lbrace {\lambda_m}\rbrace }
\end{math} and normalized right eigenvectors \begin{math}{\lbrace{{\mbox{\boldmath${\psi}$}}_m}\rbrace }
\end{math} that satisfy ${ {\mymat{P}} {\mbox{\boldmath${\psi}$}_m} =\lambda_m{\mbox{\boldmath${\psi}$}}_m, m=0,...,M-1}
$;
\item Define the $d$-dimensional ($d \ll D$) DM representation as
\begin{equation}\label{EQPSI}{ \myvec{\Psi}{(\myvec{x}_i)}: \myvec{x}_i
\longmapsto \begin{bmatrix} { \lambda_1\psi_1(i)} , {.} {.} {.}
,
{\lambda_{d}\psi_{d}(i)}\\
\end{bmatrix}^T \in{\mathbb{R}^{d}} },
\end{equation}
where $\psi_m(i)$ denotes the $i^{\rm{th}}$ element of ${\mbox{\boldmath${\psi}$}_m}$.
\end{enumerate}
The power of the DM framework stems from the Diffusion Distance (Eq. \ref{EqDist}). It was shown in \cite{Lafon} that the Euclidean distance in the embedded space $ \myvec{\Psi}{(\myvec{x}_i)}$ is equal to a weighted distance between rows of the probability matrix $\myvec{P}$.
This distance is defined as the Diffusion Distance
\begin{equation}{ \label{EqDist} { {\cal{D}}^2_t( \myvec{x}_i,\myvec{x}_j)=||{\mymat{\Psi}_t{(\myvec{x}_i)}}-{\mymat{\Psi}_t{(\myvec{x}_j)}}||^2=||\myvec{P}_{i,:}-\myvec{P}_{j,:}||^2_{\tiny\mymat{W}^{-1}}},}
\end{equation}
where $\mymat{W}$ is a diagonal matrix with elements
$W_{i,i}=\frac{D_{i,i}}{\sum_{i=1}^M D_{i,i}}$. Thus, the DM embedding is distance preserving, meaning that neighboring points in the high-dimensional space are embedded close to each other in the diffusion coordinates.
\section{Data Fusion}
\label{SecMulti}
Many physical phenomena are sampled using multiple types of sensing devices. Each sensor provides a noisy measurement of a latent parameter of interest. Data fusion is the process of incorporating multiple observation of the same data points to find a more coherent and accurate representation. \\
{\bf{Problem Formulation:}} Given multiple sets of data points $\mymat{X}^l\text{ }, l=1,...,L$. Each
view is a high dimensional dataset ${\mymat{X}^l = \lbrace{ {
\myvec{x}_1^l,\myvec{x}_2^l, \myvec{x}_3^l,...,\myvec{x}_M^l }}
\rbrace , \myvec{x}_i^l\in {\mathbb{R}^{D}} }$. Find a reliable low dimensional representation $\myvec{\Psi}(\myvec{X}^1,...,\myvec{X}^L)\in \mathbb{R}^d$.
\subsection{Multi-View Diffusion Maps (Multi-View DM)}
\label{Sec2} An approach for fusion kernel matrices in the spirit of DM framework was presented in \cite{lindenbaum}. The idea is to enforce a random walk model based on the kernels that model each view by restraining
the random walker to ``hop'' between views in each time step.
The construction requires to compute a Gaussian kernel for each view
\begin{equation} \label{EQK}{ K^l_{i,j}=exp\{
{-\frac{||\myvec{x}^l_i-\myvec{x}^l_j||^2}{2 \sigma_l^2}\}},\text{ } l=1,...,L },\end{equation}
then the multi-view kernel is formed by the following matrix
\begin{equation} \label{EQKMAT}
\mymat{\widehat{K}}= \begin{bmatrix} \mymat{0}_{M \times M} & {\mymat{K}^1\mymat{K}^2}& {\mymat{K}^1\mymat{K}^3}&...& {\mymat{K}^1\mymat{K}^p} \\
\mymat{K}^2\mymat{K}^1 & \mymat{0}_{M \times M} & {\mymat{K}^2\mymat{K}^3}&...& {\mymat{K}^2\mymat{K}^p}\\ \mymat{K}^3\mymat{K}^1 & {\mymat{K}^3\mymat{K}^2} &
\mymat{0}_{M \times M} &...& {\mymat{K}^3\mymat{K}^p}\\:&:&:&...&:\\\mymat{K}^p\mymat{K}^1 & {\mymat{K}^p\mymat{K}^2} & {\mymat{K}^p\mymat{K}^3} &...&
{\mymat{0}_{M \times M}}.
\end{bmatrix}. \end{equation}
Next, re-normalizing using the diagonal matrix $\mymat{\widehat{D}}$ where
\begin{math}
{\widehat{D}}_{i,i}=\underset{j}{\sum}{{\widehat{K}}_{i,j}}
\end{math}, the normalized row-stochastic matrix is defined as
\begin{equation}
\label{phat}
\mymat{\widehat{P}}={\mymat{\widehat{D}}}^{-1}\mymat{\widehat{K}}, ~~~
{\widehat{P}}_{i,j}=\frac{{{\widehat{K}}_{i,j}} }{\widehat{D}_{i,i} },
\end{equation}
where the $m,l$ block is a square $M\times M$ matrix located at\\
$[1+(m-1)M,1+(l-1)M], l=1,...,L$. This block describes the probability of
transition between view $\mymat{X}^m$ and $\mymat{X}^l$.
The multi-view DM representation for $\mymat{X}^l$ is computed
by
\begin{equation}
\label{Map1}
{ \myvec{\widehat{\Psi}}_t{(\myvec{x}^l_i)}: \myvec{x}^l_i
\longmapsto
\begin{bmatrix} { \lambda_1^{t}\psi_1(i+\bar{l})} , {.} {.} {.} ,
{\lambda_{d}^{t}\psi_{d}(i+\bar{l})}
\end{bmatrix}^T \in{\mathbb{R}^{d}} },
\end{equation} where $\bar{l}=(l-1)\cdot M$.
The final low dimensional representation is defined by a concatenation of all low dimensional multi-view mappings
\begin{equation} \label{eq:MVDMrep} \vec{\myvec{\Psi}}(\vec{\myvec{X}})=
{[ \myvec{\widehat{\Psi}}{(\myvec{X}^1)}},\myvec{\widehat{\Psi}}{(\myvec{X}^2)},...,\myvec{\widehat{\Psi}}{(\myvec{X}^L)}].
\end{equation}
\subsection{Alternative Methods}
\label{SECALT}
Here we provide a brief description of several methods for fusing the views before the application of a spectral decomposition. \\
{\bf{Kernel Product (KP)}}: Multiplying the kernel matrices element wise $\mymat{K}^{{\circ}}\defeq \mymat{K}^1
\circ \mymat{K}^2 \circ ...\circ \mymat{K}^L $, ${K}_{ij}^{\circ}\defeq {K}_{ij}^1
\cdot {K}_{ij}^2\cdot...\cdot {K}_{ij}^L$, then normalizing by the sum of rows. The resulting row stochastic matrix is denoted as $\mymat{P}^{{\circ}}$.
This kernel corresponds to the approach in \cite{Lafon}.\\
{\bf{Kernel Sum (KS)}}: Defining the sum
kernel $\mymat{K}^{+}\defeq\sum^L_{l=1}\mymat{K}^l$. Normalizing the sum kernel by the sum
of rows, to compute $\mymat{P}^{+}$. This random walk sums the step probabilities from each view.
This approach is proposed in \cite{Zhou}.\\
{\bf{Kernel Canonical Correlation Analysis (KCCA)}}: This method detailed in \cite{lai2000kernel,akaho2006kernel} extend the well know Canonical Correlation Analysis (CCA).Two kernels $\myvec{K}^1 \text{and }\myvec{K}^2$ are constructed in each view as in Eq. (\ref{EQK}) and the canonical vectors $\myvec{v}_1 \text{and }\myvec{v}_2$ are computed by solving the following generalized eigenvalue problem
\begin{equation} \label{eq:KCCA}
\begin{bmatrix} \mymat{0}_{M \times M} & {\mymat{K}^1\cdot \mymat{K}^2} \\ {{\mymat{K}^2\cdot \mymat{K}^1}} & \mymat{0}_{M \times M} \end{bmatrix} \begin{pmatrix}
{\myvec{v}_1} \\ {\myvec{v}_2}
\end{pmatrix}= \rho \cdot \begin{bmatrix} (\mymat{K}^1+\gamma\myvec{I})^2 & {\mymat{0}_{M \times M}} \\ {\mymat{0}_{M \times M}} &(\mymat{K}^2+\gamma\myvec{I})^2 \end{bmatrix} \begin{pmatrix}
{\myvec{v}_1} \\ {\myvec{v}_2}
\end{pmatrix},
\end{equation} where $\gamma \myvec{I}$ are regularization terms which guarantee that the matrices $(\mymat{K}^1+\gamma\myvec{I})^2$ and $(\mymat{K}^2+\gamma\myvec{I})^2 $ are invertible.
\section{Seismic Data Set}
\label{sec:Data}
The data set that is used for demonstrating the proposed kernel based approaches includes $2023$ explosions and $105$ earthquakes. $1654$ of the explosions occurred at the Shidiya phosphate quarry in the Southern Jordan between the years $2005$-$2015$ (see a map of the region in Figure \ref{fig:QuarryLocationsA}). These events were reported by the Israel National Data Center at the Soreq Nuclear Research Center with magnitudes $2\leq \text{ML}\leq3$ seismic. The rest of the events were taken from the seismic catalog of the Geophysical Institute of Israel between the years $2004$-$2014$ . All events were reported in Israel between latitudes $29^{\circ}$N-$32.5^{\circ}$N and longitudes $34.2^{\circ}$E-$35.7^{\circ}$E with duration magnitudes Md $\ge 2.5$.
Most of the earthquakes in the dataset occurred in the Dead Sea transform \cite{Garfunkel}.
The dataset includes the February 11, 2004 earthquake with the duration magnitude of Md = 5.1. This was the strongest event in this area since 1927 \cite{Hofstetter}. Twelve aftershocks that are included in the dataset are associated with this main shock.
The majority of the explosions in the dataset are ripple-fire query blasts. Moreover, the dataset consists of several one shot explosions, for instance, two experimental underwater explosions in the Dead Sea \cite{Hofstetter} and surface and near-surface experimental explosions at the Oron quarry \cite{gitterman2009source} and at the Sayarim Military Range \cite{Fee} in the Negev desert.
The dataset consists of seismogram recordings from the HRFI (Harif) station in Israel. The station is part of the Israel National Seismic Network \cite{Hofstetter}. It is equipped with a three component broad band STS-2 seismometer and a Quanterra data logger. The seismograms are sampled at a frequency of $40$ Hz. Waveform segments of $2.5$-minutes (6000 samples) have been selected for every event. In each waveform, the first P phase onsets reside $30$ seconds after the beginning of each waveform.
Figure \ref{fig:QuarryLocationsA} displays the events on the regional map.
\begin{figure}
\centering
\includegraphics[scale = 0.45]{MULTI-VIEW_KERNELS_Events_FigY1.jpg}
\caption{Seismic events in the data set and the HRFI station.}\label{fig:QuarryLocationsA}
\end{figure}
\section{Seismic Preprocessing and Feature Extraction Methods}
\label{sec:Pre}
This section provides background on typical methods that are used for seismic signal processing as well as the description of the feature extraction method that was applied here. First, the STA/LTA detector is reviewed. Next, we describe how the alignment between the different waveforms was implemented. Last, the feature extraction step, which results in a time-frequency representation of the seismic signal, is described.
\subsection{Short and Long Time Average (STA/LTA)}
\label{STA}
Detection of seismic signal embedded in the background noise is a classical problem in the signal processing theory. In the context of statistical decision theory it may be formulated as a choice between two alternatives: a waveform contains solely the noise or it contains a signal of interest superimposed on the noise. The STA/LTA trigger is a most widely accepted detection algorithm in seismology \cite{trnkoczytopic}. It relies on the assumption that a signal is characterize by a concentration of higher energy level compared with the energy level of the noise. This is done by comparing short-time energy average to a long-time energy average using a Short Time Average/ Long Time Average (STA/LTA) detector. Usually a band-pass filter is applied before the STA/LTA test.
Given a time signal $\myvec{y}(n)$ the ratio $R(i)$ is computed at each time instance $i$ is computed as follows
\begin{equation} \label{eq:STALTA}
R(i)=\frac{L\cdot[\overset{i+S}{\underset{{j=i}}{\sum}}y^2(j)]}{S\cdot(\overset{i+L}{\underset{{j=i}}{\sum}}y^2(j))},
\end{equation} where $L \gg S$ are the number of samples used for the long and short average correspondingly. The ratio $R(i)$ is compared to a threshold $\delta$ to identify time windows suspected as seismic events.
\subsection{Seismic Event Alignment}
All waveform segments in the dataset were extracted according to the first P phase onset time. Those onset times were manually picked by the analysts. However, our selective waveform inspection showed that the P onsets often have actual offsets of several seconds, sometimes even of ten seconds.
In order to increase the accuracy of the alignment, Algorithm \ref{alg:Trigger} is proposed to detect the first P onsets.
\begin{algorithm}[h]
\caption{Seismic trigger alignment} \textbf{Input:} Input time signals $\myvec{y}[n]$.\\ \textbf{Output:} Estimated time sample $\hat{n}_P$ for P onset of seismic event.
\begin{algorithmic}[1]
\STATE Apply a finite impulse response band pass filter to $\myvec{y}[n]$. The filter $\myvec{h}_1$ is designed to pass the signal between $f^{(1)}_L=2[Hz] \text{ and } f^{(1)}_H=4[Hz]$. The filtered signal is denoted as $\myvec{\tilde{y}}^{(1)}[n]$
\STATE Compute the STA/LTA ratio based on Eq. (\ref{eq:STALTA}).
\STATE Set $n^{(1)} \defeq \min (n),\text{s.t. } R(n)> \delta$. The threshold $\delta$ is computed based on the following formula $\delta= \min (4,0.3 \cdot \max (R(n)))$.
\STATE Repeat steps 1-3 using $f^{(2)}_L=4[Hz],f^{(2)}_H=8[Hz]$ and $f^{(3)}_L=8[Hz],f^{(3)}_H=12[Hz]$. Denote the trigger indexes as $n^{(2)}$ and $n^{(3)}$.
\STATE Set the estimated trigger as $\hat{n}\defeq \min (n^{(1)},n^{(2)},n^{(3)})$.
\end{algorithmic}
\label{alg:Trigger}
\end{algorithm}
Algorithm \ref{alg:Trigger} aligns the seismic events based on the STA/LTA ratios which are computed using three filtered versions of the input signal. We assume that most of the energy of the seismic signature is between $2[Hz]$ and $12[Hz]$. Figure \ref{fig:STALTA} presents a visual example for the application of Algorithm \ref{alg:Trigger}.
\begin{figure}
\centering
\includegraphics[scale = 0.2]{STALTA.png}
\caption{The STA/LTA ratios (Eq. \ref{eq:STALTA}) computed for an earthquake. Each ratio $R(i)$ is computed using one of three filtered signals $\myvec{\tilde{y}}^{(1)}(n),\myvec{\tilde{y}}^{(2)}(n),\myvec{\tilde{y}}^{(3)}(n)$. The filters are designed as explained in step 2 of Algorithm \ref{alg:Trigger}. The constant black line is an example of a threshold $\delta=4$. The onset $\hat{n}_P$ is defined as the first cross point of the threshold $\delta$.}\label{fig:STALTA}
\end{figure}
\subsection{Feature Extraction}
In this study a time-frequency representations, named sonograms \cite{joswig1990pattern}, is used, with some modification. The sonogram is a normalized short time Fourier transform (STFT) rearranged to be equal tempered on a logarithmic scale.
Each raw single-trace seismic waveform input is denoted by $\myvec{y}(n)\in \mathbb{R}^{\bar{N}}$. The length of the signals in this study is $N=6,000$ with a sampling rate of $F_s=40 Hz$. An example of seismic signals recorded using three channels is presented in Figure \ref{fig:Sig}.
\begin{figure}
\centering
\includegraphics[scale = 0.25]{SignalZ.png}
\includegraphics[scale = 0.25]{SignalE.png}
\includegraphics[scale = 0.25]{SignalN.png}
\caption{4500 samples from a recording of an explosion. Top - Z channel. Middle - E channel. Bottom - N channel.}\label{fig:Sig}
\end{figure}
The sonogram is extracted from $\myvec{y}(n)$ based on the following steps:
\begin{itemize}
\item {Given a recorded signal $\myvec{y}(n)\in \mathbb{R}^N$ the short time Fourier transform (STFT) is computed by
\begin{equation}
\label{eq:STFT}
\myvec{{{STFT}}}(f,t) = \sum\limits_{n = 1}^{N} {w(n-\ell) \cdot {y}(n)} \cdot {e^{ - j2\pi f}} ,
\end{equation}
where $w(n-t)$ is a Hann window function with a length of $N_0=256$ and a $s=0.8$ overlap. The time indexes are $\ell=(1-s)\cdot N_O\cdot t ,t=1,...,T$. The number of time bins is computed using the following equation
\begin{equation}
\label{eq:TimeWin}
T=\lceil{\frac{N-N_0}{(1-s)\cdot N_0}}\rceil+1
\end{equation}}
\item{The Spectrogram is the normalized energy of $\myvec{STFT}(f,t)$
\begin{equation}
\label{eq:NSPEC}
\myvec{R}(f,t) = \frac{\myvec{STFT}(f,t)^2 }{N_0}.
\end{equation} The Spectrogram $\myvec{R}(f,t)$ contains $T$ time bins and $F=N_0$ frequency bins. }
\item{The frequency scale is then rearrange to be equally tempered on a logarithmic scale, such that the final spectrogram contains $11$ frequency bands. The frequency bands are presented in Table \ref{table0}.
\begin{figure}
\centering
\includegraphics[scale = 0.26]{Sono.png}
\caption{A sonogram extracted from the E channel of an explosion seismogram.}\label{fig:Sono}
\end{figure}
\begin{table}
\centering
\caption{The list of frequency bands used for the sonogram computation.}
\label{table0}
\begin{tabular}[t!]{|c|c|c|}
\hline
{ Band Number} & f-start & f-end \\
\hline
{\#1} & ${0}$ [Hz] & ${0}$ [Hz] \\
{\#2} & ${0.157}$ [Hz] & ${0.315}$ [Hz] \\
{\#3} & ${0.315}$ [Hz] & ${0.630}$ [Hz] \\
{\#4 } & ${0.630}$ [Hz] & ${1.102}$ [Hz] \\
{\#5} & ${1.102}$ [Hz] & ${1.889}$ [Hz] \\
{\#6} & ${1.889}$ [Hz] & ${2.992}$ [Hz] \\
{\#7} & ${2.992}$ [Hz] & ${4.567}$ [Hz] \\
{\#8} & ${4.567}$ [Hz] & ${6.772}$ [Hz] \\
{\#9} & ${6.772}$ [Hz] & ${9.921}$ [Hz] \\
{\#10} & ${9.921}$ [Hz] & ${14.331}$ [Hz] \\
{\#11} & ${14.331}$ [Hz] & ${20}$ [Hz] \\
\hline
\end{tabular}
\end{table}
\item The bins are normalized such that the sum of energy in every frequency band is equal to $1$. The resulted sonogram is denoted by $\myvec{S}(k,t)$, where $k$ is the frequency band number, and $t$ is the time window number. Finally, we transpose the sonogram matrix into a Sonovector $\myvec{x}$ by concatenating the columns such that
\begin{equation} \label{eq:Sonovec}
\myvec{x}=\myvec{S}(:).
\end{equation}
}
\end{itemize}
An example of a sonogram extracted from an explosion is presented in Figure \ref{fig:Sono}.
\section{Case Studies}
\label{sec:Exp}
To evaluate the strength of multi-view DM for identifying the properties of seismic events we perform the following experiments.
\subsection{Discrimination Between Earthquakes and Explosions}
We consider the earthquake-explosion discrimination problem as a supervised binary classification task. A homogeneous evaluation data set is constructed by using data from 105 earthquakes and a random sample of 210 explosions. The sampling is repeated 200 times, and the results are the average of all trials. Algorithm \ref{alg:SeismicDM} is applied to extract a low dimensional representation of the seismic data. The number of data samples used for each events is 6000, where $N_1=1199$ (samples before onset) and $N_2=3800$ (samples after onset). An example of a 3-dimensional single view DM mapping is presented in Figure \ref{fig:MapE}. In this example, the explosions seem geometrically concentrated, while the earthquakes are spread out. This spread out structure may be associated with the diversity of the time-spectral information describing earthquakes, as oppose to the explosions that were mostly generated in specific quarries. The separation is clearly visible in this example. An evaluation of the separation is performed using a 1-fold cross-validation procedure. Test points are classified by using a simple K-NN classifier in a $d=4$ dimensional representation. The optimal dimension ($d=4$) for classification was found empirically based on our data set. The average accuracy of classification for various values of $K$ are presented in Figure \ref{fig:ClassifyExpEq}. Thus, the multi-view approach shows better performance with 95\% of correct discrimination.
\begin{algorithm}
\caption{Mapping of seismic data} \textbf{Input:} Three sets of time signals $\myvec{Y}_E, \myvec{Y}_N,\myvec{Y}_Z$. One for each seismic channel.\\ \textbf{Output:} A low dimensional mapping $\myvec{\Psi}(\myvec{Y}_E,\myvec{Y}_N,\myvec{Y}_Z)$.
\begin{algorithmic}[1]
\STATE Apply Algorithm \ref{alg:Trigger} to each time signal $\myvec{y}_Z^{(i)}$ and estimate the P onset $\hat{n}^{(i)}$.
\STATE Define the aligned truncated signal as $\myvec{\bar{y}}_Z^{(i)}(n)\defeq[y_Z^{(i)}(\hat{n}^{(i)}-N_1),...,y_Z^{(i)}(\hat{n}^{(i)}+N_2)]$.
\STATE Compute $\myvec{\bar{y}}_E^{(i)}(n) \text{ and } \myvec{\bar{y}}_N^{(i)}(n)$ in a similar manner.
\STATE Compute the Sonovecs based on Eqs. (\ref{eq:STFT}), (\ref{eq:NSPEC}) and (\ref{eq:Sonovec}).
\STATE Compute the DM mappings $\myvec{\Psi}_E,\myvec{\Psi}_N,\myvec{\Psi}_Z$ (Eq. (\ref{EQPSI})).
\STATE Compute the multi-view DM mapping $\vec {\myvec{\Psi}} $ (Eq. (\ref{eq:MVDMrep})).
\end{algorithmic}
\label{alg:SeismicDM}
\end{algorithm}
\begin{figure}
\centering
\includegraphics[scale = 0.3]{EXPvsEQmapMV.png}
\caption{A 3-dimensional DM mapping extracted from recordings of the E channel. Blue points represent man-made explosions from a variety of sources. Yellow points represent recordings of earthquakes most of which were originated in southern part of Israel.}\label{fig:MapE}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale = 0.2]{EQvsExpClassify200runs.png}
\caption{The classification accuracy for two classes, 105 earthquakes and 210 explosions.}\label{fig:ClassifyExpEq}
\end{figure}
\subsection{Quarry Classification}
Identification and separation of quarries by attributing the explosions to the known sources is a challenging task in observational seismology \cite{Harris1}, \cite{Harris2}. Here we demonstrate how the DM representation can be utilized to identify the origin of an explosion.
\begin{table} \label{Table:Loc}\caption{Description of quarry clusters.}
\begin{tabular}[h!]{|l|c|c|c|c|}
\hline
{Quarry Clusters} & \# of events & Center Lat & Center Lon & Distances to HRFI \\
\hline
{Shidiya, Jordan } & ${250}$ & ${29.91}^{\circ}$ & ${36.32}^{\circ}$ & $125\text{[Km]}$ \\
{Oron, Israel } & ${222}$ & ${30.82}^{\circ}$ & ${35.04}^{\circ}$ & $86.7\text{[Km]}$ \\
{Rotem, Israel } & ${115}$ & ${31.09}^{\circ}$ & ${35.19}^{\circ}$ & $117.7\text{[Km]}$ \\
{M. Ramon, Israel } & ${8}$ & ${30.46}^{\circ}$ & ${34.95}^{\circ}$ & $47.3\text{[Km]}$ \\
{Har Tuv, Israel } & ${7}$ & ${31.68}^{\circ}$ & ${35.05}^{\circ}$ & $128.2\text{[Km]}$ \\
\hline
\end{tabular}
\end{table}
For this study 602 seismograms of explosions are used. The explosions occurred in 5 quarry clusters (see Table \ref{Table:Loc} and Figure \ref{fig:QuarryLocations}) and the label data was taken from seismic catalogs. It should be noted that the quarry clusters may include several neighboring quarries and the quarry area may be of several kilometers (like Rotem) or more than ten kilometers (like Shidiya). Moreover, the precise (βground truthβ) location for most of explosions inside a quarry are not known. We estimate that the hypo-center accuracy in the used seismic catalogs is about a few kilometers for the explosions in Israel and it is more than ten kilometers for the explosions in Jordan, which are located outside the Israeli seismic network. The mean latitude and longitude are computed for the explosions belonging to each cluster and referred them to the nearby quarry (see Table \ref{Table:Loc}).
The application of Algorithm \ref{alg:SeismicDM} yields a low dimensional representation of the seismic recordings. An example of a 3-dimensional single view DM mapping is presented in Figure \ref{fig:QuarriesMapZ}.
\begin{figure}
\centering
\includegraphics[scale = 0.4]{MULTI-VIEW_KERNELS_QuarryClaccif_Fig6.jpg}
\caption{Map of quarry clusters.}\label{fig:QuarryLocations}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale = 0.24]{QuarryClustersDM.png}
\caption{A 3-dimensional diffusion mapping of 602 explosions.}\label{fig:QuarriesMapZ}
\end{figure}
The mapping is followed by a classification step that is performed based on a 1-fold cross validation using K-NN with $K=3$. The accuracy of the classification is presented in Figure \ref{fig:ClassifyQuarry}. The multi-view approach shows a peak performance of 85\% of correct classification rate.
\begin{figure}
\centering
\includegraphics[scale = 0.25]{ExpClassify50runsNew.png}
\caption{The classification accuracy for 5 source locations.}\label{fig:ClassifyQuarry}
\end{figure}
\subsection{Location Estimation}
The following case study demonstrates how the diffusion coordinates extract underlying physical properties of the sampled signal. In particular we show that the low dimensional representation that is generated by diffusion maps organizes the events with respect to their source location, even though this was not an input parameter of the algorithm. The original high-dimensional space holds the sonogram of each event. Nearly co-located events with the similar source mechanisms and magnitudes should have a similar time-frequency content and, consequently, have similar sonograms. Therefore, we expect them to lie close to each other in the high dimensional space. The diffusion distance, which is the metric that is preserved in DM, embeds the data while keeping its geometrical structure. Thus, physical properties (such as the source location) that characterize the sonogram and therefore define the geometric structure of the points in the high-dimensional space, are preserved in low-dimensional DM embedding. Note that such a geometry preserving metric does not exist in linear dimensionality reduction methods like PCA.
The dataset for this study includes 352 explosions that occurred in 4 quarry clustering Israel out 5 clusters above. The explosions in Jordan were removed since they are located at a large distance from the HRFI station. We show that the location of seismic events can be evaluated from the DM embedding coordinates. A similar evaluation based on a linear projection that was calculated with PCA yields a less accurate correlation to the events' true location.
Figure \ref{fig:MapLoc} (top image) displays the longitude and latitude coordinates of catalog locations of the events. These are the source locations of the seismic events. The points are colored by distance in kilometers from HRFI station. The middle and bottom images of Figure \ref{fig:MapLoc} present the two-dimensional PCA and DM embeddings of the dataset, respectively. It is clearly evident that the DM (bottom image in Figure \ref{fig:MapLoc}) representation has captured the location variability, while in the PCA representation this intrinsic factor is less obvious (middle image in Figure \ref{fig:MapLoc}). In the DM embedding, the clusters are well separated with respect to the event's location. In PCA the separation is not as clear, meaning that the low-dimensional PCA representation does not reveal this property. The Pearson correlation coefficients between first two diffusion coordinates and relative latitude and longitude are 0.82 and 0.77 for both dimensions respectively. The Pearson correlation coefficients between first two principle and relative latitude and longitude are 0.56 and 0.39 respectively.
\begin{figure}
\includegraphics[width=3.7in]{MapLocTruth1.png}
\includegraphics[width=3.7in]{PCAloc.png}
\includegraphics[width=3.7in]{MapLocN.png}
\caption{Top- the manually estimated location of events. Middle- the first two principle components of the N-channel. Bottom- the first two diffusion coordinated of the N-channel. Color represents the distance from HRFI station.}
\label{fig:MapLoc}
\end{figure}
\subsection{Detecting Anomalous Events}
This case study demonstrates the diffusion representation's ability to detect anomalous events among set of events at specific site. When two events are nearly co-located, have close magnitudes but with different source mechanisms, then their sonograms should be quite different as well.
Ripple-fire explosions are part of routine mining production cycles at the Oron phosphate quarry in Israel. In July 2006, three experimental one shot explosions were conducted by the Geophysical Institute of Israel at the Oron quarry \cite{gitterman2009source}. Our goal is to distinguish between the one shot explosions and the ripple-fire quarry blasts. This is not a trivial task, as all the events were conducted at very close distances.
\begin{algorithm}
\caption{K-NN based anomaly detection} \textbf{Input:} Low dimensional mapping $\myvec{\Psi}$.\\ \textbf{Output:} A set of indexes $\cal{I}$ of suspected anomalies.
\begin{algorithmic}[1]
\STATE Find $K$ nearest neighbors for all data points $\myvec{\bar{\Psi}}(\myvec{y}_i),i=1,...,M$, denote the set as $\cal{J}$.
\STATE Define the K-NN average distance as $\myvec{\hat{D}}_i\defeq \sum^K_{l=2} \frac{||\myvec{\Psi}(\myvec{y}_i)-\myvec{\Psi}(\myvec{y}_{j_l})||^2}{K} $.
\STATE Find all points with average distance $\myvec{\hat{D}}_i$ larger then a threshold $\delta$.
\end{algorithmic}
\label{alg:KNNanomaly}
\end{algorithm}
To remove the variability created by the location of the events, 98 blast from a small region surrounding the ground truth location of the experimental explosions as reported in \cite{gitterman2009source} are used. Algorithm \ref{alg:SeismicDM} is applied and a mapping extracted from the Z-channel is used to find the suspected anomalies. The diffusion maps embedding is presented in Figure \ref{fig:Anomalies}. The three anomaly points are colored in blue, they are clearly separated from the main cluster. The anomalies are automatically identified using Algorithm \ref{alg:KNNanomaly} with $K=4$ and a threshold set as four times the median of all distances $\hat{D}_i,i=1,...,M$. The average K-NN distance for the 98 blasts is presented in Figure \ref{fig:AverageKNN}. The four events that were suspected as anomalies include the three experimental explosions (which are described in \cite{gitterman2009source}).
\begin{figure}
\centering
\includegraphics[scale = 0.25]{Anomalies3.png}
\caption{Diffusion representation of 98 explosions recorded using the Z-channel. The suspected anomalies are colored in blue.}\label{fig:Anomalies}
\end{figure}
\begin{figure}
\centering
\includegraphics[scale = 1]{AverageKNN.png}
\caption{Average K-NN distance for each explosion. The distance is computed using $d=3$ coordinates and $K=5$ nearest neighbors.}\label{fig:AverageKNN}
\end{figure}
\section{Conclusion}
\label{sec:Future}
In this paper, we have adapted a multi-view manifold learning framework for fusion of seismic recordings and for low-dimensional modeling. The abilities of kernel fusion methods for extracting meaningful seismic parameters were demonstrated on various case studies. Various algorithms for classification of seismic events type, location estimation and anomaly detection were presented. These algorithms can be used as decision support tools for analysts who need to determine the source, location and type of recorded seismic events. Correct classification of events results in improved and more accurate seismic bulletins.
The proposed method is model free, thus it does not require knowledge of physical parameters. The underlying physical parameters are revealed by the diffusion maps and multi-view constructions. This type of kernel based sensor fusion is new in seismic signal processing and it overcomes some of the limitation of traditional model based fusion methods.
\section*{Acknowledgments}
This research was supported by the research grant of Pazy Foundation. We would like to thank Yochai Ben Horin for his advice and suggestions. We are grateful to Dov Zakosky and Batia Reich for providing us with the seismic catalog of Geophysical Institute of Israel.
\ifCLASSOPTIONcaptionsoff
\newpage
\fi
| 867083fc8e92ee9e5588e0687cf33547f6c5e151 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Over the past few decades, many efforts have been devoted to establishing
thermodynamics for general nonequilibrium
systems~\cite{Evans:1993tm,Gallavotti:1995wv,Jarzynski:1997uj,Oono:1998uj,
Sekimoto:1998uf,Lebowitz:1999tv,Crooks:1999ta,Hatano:2001uc,Seifert:2005vb,
Speck:2005wp}.
Among them, stochastic thermodynamics is one of the most widely used
approaches~\cite{Seifert:2008gv,Seifert:2012es}.
In stochastic thermodynamics, dynamics of a system surrounded by a thermal
environment is described as a stochastic process governed by
the Langevin equation or the master equation.
Thermodynamic quantities such as heat, work, and entropy production are
defined at the stochastic trajectory level in the way consistent with classical
thermodynamics~\cite{Sekimoto:1998uf,Hatano:2001uc,Seifert:2005vb,
Esposito:2012jt}.
Suppose that a system, whose configuration is denoted by $\bm{s}$, evolves
along a stochastic path $ {\mathsf{s}[\tau]} = \{\bm{s}(t)|0\leq t\leq\tau\}$
in contact with a thermal environment. A time evolution is accompanied by
the entropy production, which is decomposed into the
sum $\Delta S_{\rm tot}({\mathsf{s}[\tau]}) = \Delta S_{\rm sys}({\mathsf{s}[\tau]}) + \Delta S_{\rm
env}({\mathsf{s}[\tau]})$. In stochastic thermodynamics, the system entropy change
$\Delta S_{\rm sys}$ is taken as the difference of the Shannon entropy of
the system while the environment entropy change is taken as
\begin{equation}\label{eq:defEnvEP}
\Delta S_{\text{env}}=
\ln\frac{\mathcal{P}({\mathsf{s}[\tau]} |\bm{s}(0))}
{\mathcal{P}^\dagger({\mathsf{s}^\dagger[\tau]}|\bm{s}^\dagger(0))} ,
\end{equation}
where $\mathcal{P}({\mathsf{s}[\tau]}|\bm{s}(0))$ denotes the conditional path
probability of a system following the path ${\mathsf{s}[\tau]}$ to a given initial
configuration $\bm{s}(0)$ and $\mathcal{P}^\dagger$ denotes the conditional
path probability of a system following the time reversed path ${\mathsf{s}^\dagger[\tau]}$ to a
given initial configuration $\bm{s}^\dagger(0)$ in the reverse
process~\cite{Lebowitz:1999tv,Seifert:2005vb,Esposito:2010bd,Lee:2013to,
Kwon:2016kc}~(detailed notations will be explained later).
The Boltzmann constant $k_B$ is set to unity throughout the paper.
From the definition of the entropy production, stochastic thermodynamics
predicts several fluctuation theorems~\cite{Lebowitz:1999tv,Crooks:1999ta,
Hatano:2001uc,Seifert:2005vb,Speck:2005wp} for the statistical properties of
the entropy production and related quantities, which have been
examined experimentally~\cite{Wang:2002hw,Carberry:2004fk,Trepagnier:2004wr,
Wang:2005fe, Douarche:2005tk,Douarche:2005un}.
The entropy production in \eqref{eq:defEnvEP} is written in terms of the
time irreversibility of the system. It is interesting to note that $\Delta
S_{\rm env}$ is determined by the irreversibility of the system only.
There have been several attempts to show the consistency of
the entropy production of stochastic thermodynamics
with that of classical thermodynamics.
The consistency was first suggested for a stochastic system by
invoking an analogy to a chemical reaction system~\cite{Schnakenberg:1976wb}.
For master equation systems, the entropy production in \eqref{eq:defEnvEP}
is shown to be consistent with the second law of
thermodynamics~\cite{Lebowitz:1999tv}.
For Langevin equation systems, the expression in \eqref{eq:defEnvEP}
leads to the Clausius relation $\Delta S_{\rm env} = \frac{\Delta Q}{T}$
where $\Delta Q$ is the heat dissipated into the thermal environment of
temperature $T$~\cite{Seifert:2005vb}.
Despite the consistency at the phenomenological level,
the entropy production in terms of the path irreversibility
still remains to be verified microscopically.
Maes and Neto\v{c}n\'{y} tried to establish the relation
\eqref{eq:defEnvEP} for a thermal equilibrium case
by considering Hamiltonian dynamics for a coupled
system consisting of a physical system and a surrounding
environment~\cite{Maes:2003tc}. Under the Markov approximation
that the degrees of freedom of the environment should equilibrate
instantaneously, they showed that the irreversibility of the physical
system is equal to the change in the entropy of the environment.
More recently, the similar approach is applied to discrete systems
described by the master equation~\cite{Hinrichsen:2011tz,Ziener:2015kl}.
In this paper, we extend the approach of Ref.~\cite{Maes:2003tc} to
a system which is driven by an arbitrary force and surrounded
by a thermal environment. We obtain the expression for the entropy
production starting from the deterministic equations of motion and using the
Markov approximation. The
expression is shown to be the same as the one obtained from the Langevin
equation formalism. The entropy production in
\eqref{eq:defEnvEP} depends crucially on the choice the reverse
process. Especially, when the driving force
depends on the velocity as in the Lorentz force, different choices lead to
different expressions for the entropy production.
Our approach provides a systematic way for the proper choice of a reverse
process. We apply our approach to a charged particle in the presence of the
time-varying magnetic field.
This paper is organized as follows.
In Sec.~\ref{sec2}, we introduce the setting of the problem. We
consider deterministic Newtonian dynamics for a total that consists of
a physical system of interest and a surrounding environment.
The physical system is driven by a nonconservative force.
We coarse-grain the environmental degrees of freedom to derive the effective
dynamics of the system by adopting the Markov approximation.
In Sec.~\ref{sec3}, we derive the expression for the
irreversibility. We will show that the
irreversibility is the same as that obtained from the Langevin equation
approach.
In order to calculate the irreversibility, one needs to introduce a reverse
process. We suggest a rule for the
choice of a proper reverse process. The dependence on the choice of a reverse
process is significant when the driving force depends on the velocity.
We explain the rule for the Lorentz force system in Sec.~\ref{sec4}.
We summarize our results in Sec.~\ref{sec:Discussion}.
\section{Coarse graining}\label{sec2}
We consider a classical system $\mathcal{S}$
described by $N$ Cartesian coordinates
$x_{1\leq i\leq N}$ for position and $v_{1\leq i \leq N}$ for velocity.
The system interacts with an environment $\mathcal{E}$, which is described by
$(M-N)$ Cartesian coordinates $x_{N<i\leq M}$ and
$v_{N<i\leq M}$ for position and velocity, respectively.
The configuration of the total system $\mathcal{U}$
corresponds to a point in the
$2M$-dimensional phase space $\Omega$. The phase space point is denoted by
$\bm{c} = (\bm{X},\bm{V})$ where
$\bm{X} \equiv (x_1,\cdots,x_N, x_{N+1},\cdots,x_M)$ and
$\bm{V} \equiv (v_1,\cdots,v_N,v_{N+1},\cdots,v_M)$.
Similarly, the configuration of the system $\mathcal{S}$ corresponds to
a point $\bm{s} =(\bm{x},\bm{v})$ in
the $2N$-dimensional phase space with $\bm{x}=(x_1,\cdots,x_N)$ and
$\bm{v}=(v_1,\cdots,v_N)$.
The total system evolves in time following the deterministic
Newtonian equations of motion:
\begin{equation}\label{eq:EqofMotion}
\begin{aligned}
\dot{x}_i &= v_i \\
\dot{v}_i &=
\begin{cases}
-\frac{\partial \Phi(\bm{X})}{\partial x_i} +
f_i(\bm{s},\bm{\lambda}) & (1\leq i \leq N), \vspace{2mm} \\
-\frac{\partial \Phi(\bm{X})}{\partial x_i} & (N<i\leq M),
\end{cases}
\end{aligned}
\end{equation}
where $\Phi(\bm{X})$ is a potential energy function of the total system
and $\bm{f}(\bm{s},\bm{\lambda}) =
(f_1(\bm{s},\bm{\lambda}),\cdots,f_N(\bm{s},\bm{\lambda}))$
is an additional nonconservative driving force applied to the system.
It may include $L$ control parameters denoted by
$\bm{\lambda} = \bm{\lambda}(t) = (\lambda_1(t),\cdots,\lambda_L(t))$, each
of which may depend on time.
We set all masses to be unity without loss of generality.
If the total system starts with a configuration $\bm{c}$ at time $t$,
its subsequent state is determined uniquely by the equations of motion. Let
$\mathcal{T}_{\Delta t}(\bm{c};t)$ be the configuration after the time
interval $\Delta t$, which will be referred to as a trajectory function.
\begin{figure}
\includegraphics*[width=\columnwidth]{fig1.eps}
\caption{Dynamics in the $2M$-dimensional configuration space of the
whole system $\mathcal{U}$ and the coarse-grained $2N$-dimensional
configuration space of the system $\mathcal{S}$.
The constant energy surface $\Omega_E$ is divided into the subsets
$V(\bm{s};E)$. The diagram in the left hand side represents the
deterministic time evolution of $\mathcal{U}$ followed by the equilibration
according to the Markov approximation.
The gray scale of the shading reflects the probability density.
The darker the area is, the higher the probability density is.
}
\label{fig1}
\end{figure}
The total energy of $\mathcal{U}$ is given by $H(\bm{c}) =
\frac{1}{2}\sum_{i=1}^M v_i^2 + \Phi(\bm{X})$.
All the states of same energy $E$ constitute a constant energy
surface $\Omega_E\equiv\{\bm{c}|H(\bm{c})=E\} \subset \Omega$.
The total energy is not conserved in the presence of the driving force.
If $\bm{c}\in
\Omega_E$, then the configuration $\bm{c}' = \mathcal{T}_{dt}(\bm{c};t)$
belongs to another energy surface $\Omega_{E+dE}$ where
\begin{equation}\label{delE}
dE = H(\bm{c}') - H(\bm{c}) = \sum_i f_i(\bm{s},\bm{\lambda}(t)) v_i dt .
\end{equation}
Figure~\ref{fig1} illustrates
the jump between energy surfaces.
The aim of this section is to derive the effective dynamics of the system
out of the
deterministic dynamics of the whole system. This can be done by
coarse-graining the degrees of freedom of the environment. The most
successful method is to introduce the Markovian approximation that the degrees
of freedom of the environment equilibrate instantaneously to a given system
configuration~\cite{Maes:2003tc,Hinrichsen:2011tz}. The assumption is valid
in the limiting case where the environment relaxes
infinitely faster than the
system~\cite{Pigolotti:2008wd,Puglisi:2010ud,Hinrichsen:2011tz,Santillan:2011gu,Esposito:2012jt,Bo:2014bq,Ziener:2015kl,Wang:2016ho}.
We adopt the Markov approximation to obtain the effective dynamics.
The coarse-graining is done by the mapping
\begin{equation}
\pi(\bm{c}) = \bm{s},
\end{equation}
which decimates the degrees of the freedom of the environment.
For a given $\bm{c}\in \Omega_E$, the corresponding system configuration
$\bm{s} = \pi(\bm{c})$ is unique. On the other hand, there are many states in
$\Omega_E$ that are coarse-grained to the same state $\bm{s}$. The set of
all such states are denoted by
\begin{equation}
V(\bm{s};E) \equiv \{\bm{c}|\pi(\bm{c})=\bm{s} \mbox{ and } H(\bm{c})=E\} \
.
\end{equation}
These subsets are represented as the rectangular regions in
Fig.~\ref{fig1}.
We are interested in the transition probability that the system
configuration jumps from $\bm{s}$ to $\bm{s}'$ in the infinitesimal time
interval $dt$ given that the
whole system is distributed according to the probability distribution
$P(\bm{c})$ in the energy surface $\Omega_E$ initially.
Such a transition is accompanied
with the energy change $dE = \sum_i f_i v_i dt$.
It can be written as
\begin{widetext}
\begin{equation}\label{eq:TransitionProb}
W_{dt}(\bm{s}\rightarrow \bm{s}';E,t)
= \frac{\int_{V(\bm{s};E)}d\bm{c}
\int_{V(\bm{s}';E+dE)}d\bm{c}'
P(\bm{c}) \delta(\bm{c}'- \mathcal{T}_{dt}(\bm{c};t))}
{\int_{V(\bm{s};E)}d\bm{c} P(\bm{c})}
\end{equation}
\end{widetext}
where $\delta(\cdot)$ is the Dirac delta function, and
$\int_{V(\bm{s};E)}d\bm{c}$ represents the
integration over the space $V(\bm{s};E)$.
The denominator is the probability that the
system $\mathcal{S}$ is in the configuration $\bm{s}$, while the numerator
is the joint probability that the system is at $\bm{s}$
initially and at $\bm{s}'$ after the time interval $dt$.
The Markov approximation simplifies the transition probability greatly.
Since the environment is assumed to be in the equilibrium state, $P(\bm{c})$
is uniform within each $V(\bm{s};E)$ sector~\cite{Maes:2003tc}.
Thus the factors $P(\bm{c})$ in the denominator and the numerator
cancel each other.
The remaining factor in the numerator is equal to the volume of
$V_{dt}(\bm{s}\to\bm{s}';E,t)$ that is defined as
\begin{equation*}
V_{dt}(\bm{s}\rightarrow \bm{s}';E,t)
=\{ \bm{c} | \bm{c}\in V(\bm{s};E) \mbox{ and }
\pi(\mathcal{T}_{dt}(\bm{c};t))=\bm{s}' \}.
\end{equation*}
It is the subset of $V(\bm{s};E)$ consisting of configurations $\bm{c} \in
V(\bm{s};E)$ that are coarse-grained to $\bm{s}'$ after
time $dt$.
Therefore, the transition probability is given by
\begin{equation}\label{eq:TransitionProb2}
W_{dt}(\bm{s}\rightarrow \bm{s}';E,t)
=\frac{|V_{dt}(\bm{s}\rightarrow \bm{s}';E,t)|}{|V(\bm{s};E)|},
\end{equation}
where $|(\cdot)|$ denotes the volume of the set $(\cdot)$ in the phase space.
The time evolution under the Markov approximation is illustrated in
Fig.~\ref{fig1}. The transition probability depends on $t$
explicitly because of the $t$ dependence of the trajectory function
$\mathcal{T}_{dt}(\bm{c};t)$.
\section{Irreversibility}\label{sec3}
In this section, we quantify the time irreversibility by comparing the
transition probability of a trajectory ${\mathsf{s}[\tau]}$ in a given
dynamical process, called the forward process, with the that of
a time-reversed trajectory denoted by
${\mathsf{s}^\dagger[\tau]} = \{\bm{\epsilon s}(\tau-t) | 0 \leq t \leq \tau\}$
in the corresponding reverse process.
Here, $\bm{\epsilon}$ is the time-reversal operator that
changes the sign of all the velocity coordinates. That is, $\bm{\epsilon
s} = (\bm{x},-\bm{v})$ for $\bm{s} = (\bm{x},\bm{v})$.
We first remark on the issue in defining the reverse process to a given
forward process. Consider, for example, a charged particle in the presence
of the uniform magnetic field $\bm{B}$.
Many literatures take it granted that the
magnetic field should be flipped~($\bm{B}\to -\bm{B}$) in the reverse
process because they are the time-reversal counterpart to each
other~\cite{Kampen:2011vs,Risken:1996vl}.
On the other hand, some studies claim that one should use the same
field $\bm{B}$ on the ground that the irreversibility is meaningful when a
trajectory and its time-reversed trajectory are compared in the
setting~\cite{Ganguly:2013vk,Chaudhuri:2014vf, Kwon:2016kc,Chaudhuri:2016ke}.
Such a difficulty arises when the driving force $\bm{f}$ depends explicitly
on the velocity so that it breaks the time-reversal symmetry.
We will provide an argument that guides us to choose the appropriate reverse
process for a general driving force $\bm{f}$.
Consider a forward process with a driving force
$\bm{f}(\bm{s},\bm{\lambda})$ for a time interval $0\leq t \leq \tau$.
Suppose that the system evolves along a trajectory
${\mathsf{s}[\tau]} : \bm{s}(t_0=0) \to \cdots \to \bm{s}(t_l) \to \cdots \to
\bm{s}(t_n=\tau)$ with $t_l = l dt$.
The forward trajectory is to be compared with the time-reversed one
${\mathsf{s}^\dagger[\tau]}: \bm{s}^\dagger(t_0) \to \cdots \to \bm{s}^\dagger(t_l) \to \cdots \to
\bm{s}^\dagger(t_n)$ with $\bm{s}^\dagger(t_l) = \bm{\epsilon
s}(t_{n-l}=\tau-t_l)$ in the reverse process.
Since the driving force $\bm{f}$ works on the system,
the whole system $\mathcal{U}$ jumps from one energy surface $\Omega_E$
to the other $\Omega_{E+dE}$ with $dE$ in \eqref{delE} in each
step~[see also Fig.~\ref{fig1}].
In defining the reverse process with the
choice of the driving force $\bm{f}^\dagger(\bm{s},\bm{\lambda}^\dagger)$,
we require that not only the system $\mathcal{S}$ should return back from
$\bm{\epsilon s}(t_{l+1})$ to $\bm{\epsilon s}(t_l)$ and but also the
whole system $\mathcal{U}$ from $\Omega_{E+dE}$ to $\Omega_{E}$ for each $l$
in the reverse process.
The energy surface requirement constraints the possible form of
$\bm{f}^\dagger(\bm{s},\bm{\lambda}^\dagger)$.
The work $dE^\dagger$ done by $\bm{f}^\dagger$ in the reverse process should
cancel $dE$, which yields
\begin{equation}
\sum_i f_i^\dagger (\bm{\epsilon s},\bm{\lambda}^{\dagger}(t)) (-dx_i)
= -\sum_i f_i (\bm{s},\bm{\lambda}(\tau-t)) dx_i
\end{equation}
up to the leading order in $dt$. It suggests that the driving
force in the reverse process should be chosen as
\begin{equation}\label{eq:ReverseForce2}
\bm{f}^\dagger(\bm{s},\bm{\lambda}^\dagger(t)) =
\bm{f}(\bm{\epsilon} \bm{s},\bm{\lambda}(\tau-t)).
\end{equation}
The meaning of this choice is clear.
The forces acting on the system at each time step constitute a sequence
$\{\bm{F}_0,\ldots,\bm{F}_l,\ldots,\bm{F}_n\}$ with $\bm{F}_l =
\bm{f}(\bm{s}(t_l),\bm{\lambda}(t_l))$.
The choice in \eqref{eq:ReverseForce2} implies that
the forces in the reverse process constitute the sequence
$\{\bm{F}^\dagger_0,\ldots,\bm{F}^\dagger_l,\ldots,\bm{F}^\dagger_n\}$
with $\bm{F}^\dagger_l = \bm{f}^\dagger(\bm{s}^\dagger(t_{l}),
\bm{\lambda}^\dagger(t_{l})) = \bm{f}(\bm{s}(t_{n-l}),\bm{\lambda}(t_{n-l})) =
\bm{F}_{n-l}$.
The system is acted on by the {\em same force values} in the time-reversed
order.
Note that $\bm{f}^\dagger$ has a {different function form} from $\bm{f}$
when $\bm{f}$ depends on the velocity $\bm{v}$. An
explicit example involving a charged particle in the presence of the
magnetic field will be discussed in Sec.~\ref{sec4}.
Another important property of the choice \eqref{eq:ReverseForce2} is that
every trajectory ${\mathsf{c}[\tau]}=\{{\bm{c}} (t)|0\leq t\leq\tau\}$ of the whole system
$\mathcal{U}$ in the forward process is traced back in the reverse process.
Formally we have
\begin{equation}\label{eq:Property}
\mathcal{T}_{t}^\dagger ( \bm{\epsilon} \mathcal{T}_{t}(\bm{c};0);\tau-t)
= \bm{\epsilon} \bm{c}
\end{equation}
with the trajectory function $\mathcal{T}^\dagger$
of the reverse process.
Once the reverse process is defined, the transition probability during the
infinitesimal time interval is given by
\begin{equation}
W_{dt}^\dagger(\bm{s}\to\bm{s}';E,t) =
\frac{ | V_{dt}^\dagger(\bm{s}\to\bm{s}';E,t) | }{ | V(s;E) | } ,
\end{equation}
where
\begin{equation*} V_{dt}^\dagger(\bm{s}\to \bm{s}';E,t) =
\{ \bm{c} | \bm{c}~\in V(\bm{s};E)
\mbox{ and } \pi(\mathcal{T}^\dagger_{dt}(\bm{c},t)) = \bm{s}'\} \ .
\end{equation*}
Thus, the irreversibility, given by the log ratio of the path probabilities
as appeared in the right hand side of \eqref{eq:defEnvEP},
is given by the sum of
\begin{equation}\label{eq:Irreversibility1}
dI = \ln\frac{W_{dt}(\bm{s}\to\bm{s}';E,t)}
{W_{dt}^\dagger(\bm{\epsilon}\bm{s}'\to\bm{\epsilon}\bm{s};E+dE,\tau-t)}
= dI_1 + dI_2 \ ,
\end{equation}
where
\begin{equation}
\begin{aligned}
dI_1 = &\ln\frac{|V(\bm{\epsilon} \bm{s}';E+dE)|}{|V(\bm{s};E)|} \\
dI_2 = &\ln\frac{|V_{dt}(\bm{s}\rightarrow \bm{s}';E,t)|}
{|V_{dt}^\dagger(\bm{\epsilon} \bm{s}'\rightarrow \bm{\epsilon} \bm{s};
E+dE,\tau-t)|}.
\end{aligned}
\end{equation}
Using the property in \eqref{eq:Property}, one finds that
$V_{dt}^\dagger(\bm{\epsilon} \bm{s}'\to\bm{\epsilon}\bm{s};E+dE,\tau-t)
=\bm{\epsilon} \mathcal{T}_{dt}(V(\bm{s}\to\bm{s}';E,t))$.
One also finds that
$V(\bm{\epsilon s};E) = \bm{\epsilon} V(\bm{s};E)$ and that
the phase space volume is invariant under the operation of $\bm{\epsilon}$.
Therefore, the irreversibility is given by
\begin{equation}\label{eq:Irreversibility2}
\begin{aligned}
dI_1 = & \ln\frac{|V(\bm{s}';E+dE)|}{|V(\bm{s};E)|} \\
dI_2 = & \ln\frac{|V(\bm{s}\to\bm{s}';E,t)|}
{|\mathcal{T}_{dt}(V(\bm{s}\to\bm{s}';E,t))|}.
\end{aligned}
\end{equation}
We stress that $dI$ in \eqref{eq:Irreversibility1} measures the time
irreversibility of the whole system including the physical system and the
environment. The choice in \eqref{eq:ReverseForce2} guarantees that the
environment returns to the original energy surface in the reverse process.
The subspace $V(\bm{s};E)$ comprises the accessible states of the
environment to a given system state $\bm{s}$ in the energy surface $\Omega_E$.
Thus, $\ln |V(\bm{s},E)|$ is the Boltzmann entropy of the
environment and $dI_1$ in Eq.~(\ref{eq:Irreversibility2}) is equal
to the change in the entropy of the environment.
It can also be written in the Clausius form in the weak coupling limit.
The energy $E$ of the total system $\mathcal{U}$ is decomposed into
the sum $E = E_{\rm sys} + E_{\rm env} + E_{\rm int}$, where $E_{\rm
sys}~(E_{\rm env})$ is the energy of the
system~(environment) and $E_{\rm int}$ is the interaction energy between
them.
In the weak coupling limit, $E_{\rm int}$ is negligible so that
$E \simeq E_{\rm sys} + E_{\rm env}$.
Hence, we have $\ln | V(\bm{s};E) | = S_{\rm env}(E_{\rm env} =
E-E_{\rm sys}(\bm{s}))$ and
$\ln |V(\bm{s'};E+dE)| = S_{\rm env} (E_{\rm env} = E+dE - E_{\rm sys}(\bm{s'}))$, where
$S_{\rm env}(E_{\rm env})$ denotes the entropy of the environment
as a function of the energy.
We note that $dE$ is the work done by the driving force on the
system. The first law of thermodynamics implies that $E_{\rm sys}(\bm{s}') -
E_{\rm sys}(\bm{s}) = dE - dQ$ where $dQ$ denotes the heat dissipated to the
environment. Consequently, we obtain that
\begin{equation}\label{dI1res}
dI_1 = \frac{dQ}{T} ,
\end{equation}
where $T = \left( \partial S_{\rm env}/\partial E_{\rm env} \right)^{-1}$
is the temperature of the environment. Extension to systems at strong
coupling with the environment would be interesting~\cite{Seifert:2016ik},
which we do not pursue in this work.
The quantity $dI_2$ involves the expansion rate of the phase space volume
during the time evolution. It is determined by the determinant of
the Jacobian matrix
$\mathsf{J} = \partial \bm{c}'/\partial \bm{c}$ with $\bm{c}' =
\mathcal{T}_{dt}(\bm{c};t)$ for $\bm{c} \in V(\bm{s};E)$.
The Jacobian matrix $\mathsf{J}$ is a block matrix of size $2M\times 2M$ in
the form of
\begin{equation}
\mathsf{J}=\begin{pmatrix}
\mathsf{A} & \mathsf{B} \\
\mathsf{C} & \mathsf{D}
\end{pmatrix}
\end{equation}
where $A_{mn}=(\partial x'_m/\partial x_n) = \delta_{mn}$,
$B_{mn}=(\partial x'_m/\partial v_n)=\delta_{mn} dt$,
\begin{equation*}
C_{mn}=\frac{\partial v_m'}{\partial x_n} =
\left(-\frac{\partial^2\Phi}{\partial x_m \partial x_n}
+\sum_{i,j=1}^N\delta_{im}\delta_{jn}\frac{\partial f_i}{\partial
x_j}\right)dt ,
\end{equation*}
and
\begin{equation}\label{eq:MatrixD}
D_{mn}=\frac{\partial v'_m}{\partial v_n} = \delta_{mn}
+\sum_{i,j=1}^N\delta_{im}\delta_{jn}\frac{\partial f_i}{\partial v_j}dt\\
\end{equation}
are the submatrices of size $M\times M$~($m,n=1,\cdots,M$) up to the first
order in $dt$, where $\delta_{mn}$ is the Kronecker delta symbol.
The determinant of the block matrix is given by
$\det(\mathsf{J})=\det(\mathsf{D})\det(\mathsf{A}-\mathsf{B}
\mathsf{D}^{-1}\mathsf{C})$ \cite{Silvester:2000wl}.
Note that $\mathsf{A}=\mathsf{I}$, $\mathsf{B} = (dt) \mathsf{I}$,
$\mathsf{C} = O(dt)$, and $\mathsf{D} = \mathsf{I} + \mathcal{O}(dt)$.
Thus, we obtain that $\det(\mathsf{J})=\det(\mathsf{D})
=\prod_{m=1}^M D_{mm} = 1+dt\sum_{i=1}^N
\partial f_i/ \partial v_i$ up to $\mathcal{O}(dt)$, which yields that
\begin{equation}\label{dI2res}
dI_2 = \ln \det{\mathsf{J}}^{-1} =
-dt \left(\nabla_{\bm{v}} \cdot \bm{f}\right)
\end{equation}
with the shorthand notation $(\nabla_{\bm{v}} \cdot \bm{f}) \equiv
\sum_{i=1}^N \partial f_i / \partial v_i$.
Combining \eqref{dI1res} and \eqref{dI2res}, we finally obtain
\begin{equation}\label{eq:Irreversibility3}
dI = \frac{dQ}{T}
-dt \left[\nabla_{\bm{v}} \cdot \bm{f}(\bm{s},\bm{\lambda})\right] \ .
\end{equation}
When the driving force does not depend on the velocity, then the
irreversibility in \eqref{eq:Irreversibility3} is equal to the change in the
entropy of the environment $dS_{\rm env}$. The same is true even in the
presence of the velocity-dependent force as long as it has the vanishing
divergence with respect to the velocity~($\nabla_{\bm{v}} \cdot \bm{f}=0$).
The additional contribution becomes nonzero when
$\nabla_{\bm{v}} \cdot \bm{f} \neq 0$.
The thermodynamic meaning of the additional term remains unknown yet.
We now show that the irreversibility in \eqref{eq:Irreversibility3} based on
the deterministic dynamics incorporated with the Markovian approximation and
the weak coupling limit is reproduced in the phenomenological Langevin
equation approach.
Consider the Langevin equations
\begin{equation}\label{eq:ForwardLangevin}
\begin{aligned}
\dot{x}_i &= v_i\\
\dot{v}_i &= f_{{\rm c},i}(\bm{x}_s)
+f_i(\bm{s},\bm{\lambda})-\gamma v_i +\xi_i(t) \ .
\end{aligned}
\end{equation}
In comparison with \eqref{eq:EqofMotion}, interactions with the environment are
treated with the damping force and the thermal white noise satisfying
$\langle\xi_i(t)\rangle=0$ and
$\langle\xi_i(t)\xi_j(t')\rangle=2\gamma T \delta_{ij}\delta(t-t')$. The
system is driven by the conservative force denoted by
$\bm{f}_{\rm c}(\bm{s})$ and the nonequililbrium driving force $\bm{f}$.
The Langevin equations for the reverse process are given by
\begin{equation}\label{eq:ReverseLangvin}
\begin{aligned}
\dot{x}_i &= v_i\\
\dot{v}_i &= f_{{\rm c},i}
+f_i^\dagger(\bm{s},\bm{\lambda}^\dagger)-\gamma v_i +\xi_i(t) \ .
\end{aligned}
\end{equation}
The Onsager-Machlup formalism allows one to write down the path probability
for the Langevin equation system~\cite{Onsager:1953uv}.
Using the formalism, we obtain the logarithm of the path probability ratio
of the forward and reverse processes during the infinitesimal time interval
$dt$. It is given by
\begin{equation}
\begin{aligned}\label{eq:EP_stochastic}
dI = &\frac{dQ}{T}-dt\nabla_{\bm{v}}\cdot
\left[\bm{f}(\bm{s},\bm{\lambda})-\delta{\bm{f}}\right] +
\frac{\delta\bm{f}}{\gamma T} \circ d\bm{v} \\
& + \frac{dt}{\gamma T} \delta\bm{f} \cdot \left[-\bm{f}_{\rm c}(\bm{s}) -\bm{f}(\bm{s},\bm{\lambda})
+\delta{\bm{f}}-\gamma\bm{v}\right],
\end{aligned}
\end{equation}
where $\delta\bm{f} \equiv [\bm{f}(\bm{s},\bm{\lambda})
-\bm{f}^\dagger(\bm{\epsilon}\bm{s},
\bm{\lambda}^\dagger)]/2$ and the notation $()\circ d\bm{v}$
stands for the stochastic integral in the Stratonovich
sense~\cite{Gardiner:2010tp} (see Appendix~\ref{sec:appendixA} for
derivation).
When we choose the driving force $\bm{f}^\dagger$ in the reverse process
according to \eqref{eq:ReverseForce2},
$\delta\bm{f}$ is identically zero and the
two irreversibilities in \eqref{eq:Irreversibility3} and
\eqref{eq:EP_stochastic} become the same. Our theory substantiates the
Langevin equation approach under the choice of \eqref{eq:ReverseForce2}.
\section{Charged particle under the Lorentz force}\label{sec4}
The irreversibility in \eqref{eq:Irreversibility1} depends crucially on the
definition of the reverse process to a given forward process. We have
proposed that the force $\bm{f}^\dagger$ should be chosen as in
\eqref{eq:ReverseForce2} on the ground that the whole system should move
back to the original energy surface in the reverse process. This choice is
characterized by the fact that the sequence of the force values
in the reversed process is
the same as that in the forward process in the time-reversed order.
In order to stress that the force {\em values} are the same, we refer to this
choice as the V rule.
There is an alternative choice where the
function {\em form} of the force is taken to be the
same~\cite{Ganguly:2013vk, Chaudhuri:2014vf, Kwon:2016kc, Chaudhuri:2016ke}.
It is formulated as
\begin{equation}\label{Frule}
\bm{f}^\ddagger(\bm{s},\bm{\lambda}^\ddagger(t)) = \bm{f}(\bm{s},
\bm{\lambda}(\tau-t))
\end{equation}
In order to distinguish it from $\bm{f}^\dagger$ according to the V rule,
we use the superscript ${}^\ddagger$.
This choice will be referred to as the F rule. The merit
of the F rule is that the forward and the reverse processes are compared in
the same physical system characterized by the driving force of same form.
When the force depends on the velocity, the forces in the
reverse processes $\bm{f}^\dagger$ and $\bm{f}^\ddagger$ are different, so
are the irreversibility. In this
section, we compare the two choices for a charged particle under the Lorentz
force.
Consider a charged particle of mass $m$ and of charge $q$ in the
three-dimensional space with cylindrical symmetry around the $\hat{z}$
direction. The time-dependent magnetic field $\bm{B}(t) =
bt\hat{z}$ is applied to the $z$ direction with a constant $b>0$.
According to the Maxwell equation
$\nabla_{\bm{x}} \times \bm{E} = -\frac{\partial}{\partial t}\bm{B}$,
the time-varying magnetic field induces the electric field
$\bm{E}(\bm{x}) = \frac{1}{2}b (y \hat{x} - x \hat{y}) = -\frac{1}{2}b
r\hat{\theta}$ with $r=\sqrt{x^2+y^2}$ and the unit vector $\hat{\theta}$
in the azimuthal direction. The electric field line circulates
around the origin in the clockwise direction.
The particle is then applied to the Lorentz force
\begin{equation}
\bm{f}(\bm{x},\bm{v},\bm{\lambda}(t)) =
q \bm{v} \times \bm{B}(t) + q \bm{E}(\bm{x}) \ .
\end{equation}
The field strengths are regarded as the parameters $\bm{\lambda}$.
According to the V rule the force $\bm{f}^\dagger$ in the reverse process
is given by
\begin{equation}
\begin{aligned}
\bm{f}^\dagger(\bm{x},\bm{v},\bm{\lambda}^\dagger(t)) & =
\bm{f}(\bm{x},-\bm{v},\bm{\lambda}(\tau-t)) \\
& = -q \bm{v}\times \bm{B}(\tau-t) + q \bm{E}(\bm{x}) \ .
\end{aligned}
\end{equation}
It amounts to the situation that the particle is subject to the Lorentz force
under the fields
\begin{equation}
\bm{B}^\dagger(t) = -\bm{B}(\tau-t) , \
\bm{E}^\dagger(\bm{x}) = \bm{E}(\bm{x}) .
\end{equation}
Note that the magnetic field is flipped to the opposite direction.
We compare the field configurations in the forward and the reverse
processes in Fig.~\ref{fig2}.
The electro-magnetic fields in the reverse process
also satisfy the Maxwell's equation,
$\nabla_{\bm x}\times \bm{E}^\dagger =
-\frac{\partial}{\partial t} \bm{B}^\dagger$.
On the other hand, the reverse process force according to the F rule,
denoted by $\bm{f}^\ddagger$, is given by
\begin{equation}
\begin{aligned}
\bm{f}^\ddagger(\bm{x},\bm{v},\bm{\lambda}^\ddagger(t)) & =
\bm{f}(\bm{x},\bm{v},\bm{\lambda}(\tau-t)) \\
& = q \bm{v}\times \bm{B}(\tau-t) + q \bm{E}(\bm{x}) \ .
\end{aligned}
\end{equation}
It corresponds to a Lorentz-like force under the fields
\begin{equation}
\bm{B}^\ddagger(t) = \bm{B}(\tau-t) , \
\bm{E}^\ddagger(\bm{x}) = \bm{E}(\bm{x}) .
\end{equation}
These fields do not satisfy the Maxwell's equation,
$\nabla_{\bm x}\times \bm{E}^\ddagger \neq
-\frac{\partial}{\partial t} \bm{B}^\ddagger$.
Namely, the reverse process in the F rule is an artificial process with
non-physical electro-magnetic fields.
\begin{figure}
\includegraphics*[width=\columnwidth]{fig2.eps}
\caption{Magnetic and electric field configurations in the forward
process~(left) and in the reverse process according to the V rule~(middle)
and the F rule~(right).
The varying width of an arrow stands for the change of the magnetic field
strength in time.}
\label{fig2}
\end{figure}
The consistency with electromagnetism suggests that the V rule be
the proper way to define the reverse process for systems driven by a
velocity-dependent force. Under the V rule, the irreversibility consists of
the Clausius entropy change of the environment and the additional term $-dt
[\nabla_{\bm{v}} \cdot \bm{f}]$. We do not know whether the additional
term can be related to any thermodynamic quantity.
In nature, the magnetic Lorentz force is the unique example of
a velocity-dependent force among the fundamental forces.
If we restrict ourselves to the fundamental Lorentz force, the additional
term vanishes because the magnetic Lorentz force is divergence-free.
Then, the irreversibility reduces to the conventional entropy production
of the environment.
One may consider velocity-dependent forces. However, they are not the
fundamental forces but the phenomenological forces~\cite{Cerino:2015cc}.
\section{Summary}\label{sec:Discussion}
In stochastic thermodynamics, the entropy production is
given by the logarithm of the ratio of the path probabilities of the system.
In this work, we derived
the connection between the irreversibility and the entropy production
starting from the microscopic deterministic equations of motion of the whole
system $\mathcal{U}$ consisting of a physical system $\mathcal{S}$ and an
environment $\mathcal{E}$. The key
assumption behind the connection is the Markovian approximation that the
environmental degrees of freedom equilibrates so fast that they are always
in the equilibrium state to a given configuration of $\mathcal{S}$.
Our approach is
an extension of those in Refs.~\cite{{Maes:2003tc},Hinrichsen:2011tz,
Ziener:2015kl} to systems having the continuous degrees of freedom and being
driven by an external force.
We have shown that the irreversibility derived from the microscopic point of
view has the same expression as the entropy production of the corresponding
Langevin equation system.
It is crucial to consider a proper reverse process to a given
forward process in characterizing the time irreversibility. In this work, we
suggest the V rule that the sequence of the force values in the reverse
process should be the same as that in the forward process in the
time-reversed order. It is formulated in \eqref{eq:ReverseForce2}. This rule
is favored because it guarantees that the whole system returns to the
original energy surface in the reverse process. This choice is contrasted to
the F rule in \eqref{Frule}, where the force in the reverse process has the
same function form as the force in the forward process. The two choices are
compared for a charged particle in the presence of time-varying magnetic
field and the induced electric field.
\begin{acknowledgments}
This work was supported by the the National Research Foundation of
Korea~(NRF) grant funded by the Korea government~(MSIP)
(No.~2016R1A2B2013972). We thank Prof. Hyunggyu Park and Prof. Chulan Kwon
for helpful discussions.
\end{acknowledgments}
| 9cf24b6d1ffe6d4623ef16c517001173517451d2 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Blind deconvolution is a fundamental problem in low level vision, and is always drawing research attentions \cite{perrone2014total,pan2016robust,krishnan2011blind,krishnan2009fast,pan2016blind}.
Given a blurry image $\mathbf y$, blind deconvolution aims to recover a clear version $\bf x$, in which it is crucial to first estimate blur kernel $\bf k$ successfully.
Formally, the degradation of image blur is modeled as
\begin{equation}\label{eq1}
\mathbf y= \mathbf x \otimes \mathbf k+ \mathbf n,
\end{equation}
where $\mathbf x$ and $\mathbf y$ are with size $M\times N$, $\mathbf k$ is with size $L\times K$,
$\otimes$ is the 2D convolution operator and $\mathbf n$ is usually assumed as random Gaussian noises.
Blind deconvolution needs to jointly estimate blur kernel $\bf k$ and recover clear image $\bf x$.
The most successful blind deconvolution methods are based on the maximum-a-posterior (MAP) framework.
MAP tries to jointly estimate $\mathbf k$ and $\mathbf x$ by maximizing the posterior $p(\mathbf k, \mathbf x|\mathbf y)$, which can be further reformulated as an optimization on regularized least squares~\cite{chan1998total},
\begin{equation}\label{eq2}
\mathbf{\hat x}, \mathbf{\hat k}=\arg\min_{\mathbf x, \mathbf k}{\left(\|
\mathbf x\otimes \mathbf k-\mathbf y\|^2+\lambda g\left(\mathbf x\right)+\sigma h\left(\mathbf k\right)\right)}
\end{equation}
where $g$ and $h$ are prior functions designed to prefer a sharp image and an ideal kernel, respectively.
It is not trivial to solve the optimization problem in Eqn. \eqref{eq2}, and instead it is usually addressed as alternate steps,
\begin{equation}\label{eq3}
\mathbf {\hat{x}}^{(i+1)} = \arg\min_{\mathbf x}{\left(\|\mathbf x \otimes \mathbf {\hat{k}}^{(i)}-\mathbf y\|^2+\lambda g\left(\mathbf x\right)\right)}
\end{equation}
and
\begin{equation}\label{eq4}
\mathbf{\hat{k}}^{(i+1)} = \arg\min_{\mathbf{k}}{\left(\|\mathbf{\hat{x}}^{(i+1)}\otimes \mathbf k- \mathbf y\|^2+\sigma h\left(\mathbf k\right)\right)}.
\end{equation}
In the most blind deconvolution methods, kernel size $(L, K)$ is hyper-parameters that should be manually set.
An ideal choice is the ground truth size to constrain the support domain, which however is not available in practical applications, requiring hand-crafted tuning.
On one hand, a smaller kernel size than ground truth cannot provide enough support domain for estimated blur kernel.
Therefore, kernel size in the existing methods is usually pre-defined as a large value to guarantee support domain.
\begin{figure*}[t]
\centering
\setlength{\tabcolsep}{2pt}
\small
\begin{tabular}{ccccc}
\multirow{1}{*}[50pt]{\includegraphics[width=.3\linewidth]{tend_final.eps}} &
\includegraphics[width=.155\linewidth]{yk} &
\includegraphics[width=.155\linewidth]{xk1} &
\includegraphics[width=.155\linewidth]{xk2} &
\includegraphics[width=.155\linewidth]{xk3} \vspace{-2pt} \\
&\includegraphics[width=.155\linewidth]{kd.pdf} &
\includegraphics[width=.155\linewidth, trim={86.5pt 87.5pt 18pt 16pt}, clip]{k1d.pdf} &
\includegraphics[width=.155\linewidth]{k2d.pdf} &
\includegraphics[width=.155\linewidth]{k3d.pdf} \vspace{-2pt} \\
& truth size = 23 & size = 23, err = 1.9 & size = 47, err = 5.9& size = 69, err = 69.9 \vspace{2pt}\\
(a) & (b) & (c) & (d) & (e) \vspace{10pt} \\
\end{tabular}\\
\caption{Large kernels produce inferior results. (a) Numerical errors with kernel size.
(b) Blurry image and ground truth kernel. (c-e) Deblurred results. In the first row of (c-e) are restored images and corresponding estimated kernels; in the second row are support domains ($k>0$), where adjacent positive pixels are colored identically and zeros are white.
In this experiment, we omitted regularization $h$;
hence, k-step equals to a bare least squares optimization.
We also avoided using multi-scaling scheme and threshold in this experiment. Parameters that performed well on the truth size were kept identical for larger sizes during the experiment.}\label{fig1}
\end{figure*}
On the other hand, as shown in Figure~\ref{fig1}, oversized kernels are very likely to introduce estimation errors, and hence lead to unreasonable results.
Hereby, we name this phenomenon \textit{larger-kernel effect}.
This interesting fact was first mentioned by Fergus~\textit{et~al.}~\cite{fergus2006removing}.
Then Cho and Lee~\cite{cho2016convergence} showed a similar result that the residual cost of (\ref{eq2}) increases with over-estimated kernel size.
However, such annoying phenomenon was not well analyzed and studied yet.
Note that most MAP-based blind deconvolution algorithms adopt the trial-and-error strategy to tune kernel size, so the larger-kernel effect is a very common problem.
In this paper, we first explore the mechanism of larger-kernel effect and then propose a novel low rank-based regularization to relieve this adverse effect.
Theoretically, we analyze the mechanism to introduce kernel estimation error in oversized kernel size.
Specifically, we reformulate convolution of (\ref{eq3}) and (\ref{eq4}) to affine transformations and analyze their properties on kernel size.
We show that for $\mathbf x$ in sparse distributions, this larger-kernel effect remains with probability one.
We also conduct simulation experiments to show that kernel error is expected to increase with kernel size even without noise $\mathbf n$.
Furthermore, we attempt to find out a proper regularization to suppress noise in large kernels.
By exploiting the low rank property of blur kernels, we propose a low-rank regularization to reduce noises in $\mathbf{\hat k}$, suppressing larger-kernel effect.
Experimental results on both synthetic and real blurry images validate the effectiveness of the proposed method, and show its robustness to against over-estimated kernel size. Our contributions are two-folds:
\begin{itemize}
\item
We give a thorough analysis to mechanism of the phenomenon that over-estimated kernel size yields inferior results in blind deconvolution, on which little research attention has been paid.
\item
We propose a low rank-based regularization to effectively suppress \textit{larger-kernel effect} along with efficient optimization algorithm, and performs favorably on oversized blur kernel than state-of-the-arts.
\end{itemize}
\section{Larger-kernel effect}
In this section, we describe the \textit{larger-kernel effect} in detail and provide a mathematical explanation.
\subsection{Phenomenon}
In Figure ~\ref{fig1}(b-c), it has shown that the larger the kernel size would lead to more inferior deblurring results, since the estimated blur kernel with larger support domain is very likely to introduce noises and estimation errors.
Figure ~\ref{fig1}(a) shows both the error ratio (err)~\cite{levin2009understanding} of restored images and the Summed Squared Difference (SSD) of estimated kernels reach the lowest at the truth size and increase afterwards.
\subsection{Mechanism}
To analyze the source of larger-kernel effect, we firstly introduce an interesting fact that we call \textit{inflating effect}.
\newtheorem{thm}{Claim}
\begin{thm}(Inflating Effect)
Let $ A = [\bm{v_1} \ldots \bm{v_n}]$, where $\bm{v_i} \in \mathbb{R}^m$ $(m \ge n + 1)$.
Let $ B = [\bm{w_1} \: A \: \bm{w_2}]$, where $ \bm{w_1}, \bm{w_2} \in \mathbb{R}^m $ and
$\mathrm{rank}(B)>\mathrm{rank}(A)$. Given an m-D random vector $\bm{b}$ whose elements
are i.i.d. with the continuous probability density function p, for $ \bm{u} \in \mathbb{R}^m$
$$ \Pr\big(\inf\{\|B\bm{u}-\bm{b}\|^2\}<\inf\{\|A\bm{u}-\bm{b}\|^2\}\big) = 1.$$
\end{thm}
\begin{proof}
\begin{equation*}
\begin{aligned}
&&&\Pr\big(\inf\{\|B\bm{u}-\bm{b}\|^2\}\ge \inf\{\|A\bm{u}-\bm{b}\|^2\}\big) \\
&=&&\Pr\big(\bm{b}\in \mathbb{R}^m \setminus (\spn\{B\}\setminus \spn\{A\})\big) \\
&=&&\int_\Omega {d {p(\bm{b})}}
\end{aligned}
\end{equation*}
where $\Omega = \mathbb{R}^m \setminus (\spn\{B\}\setminus \spn\{A\})$.
For $\mathrm{rank}(B)>\mathrm{rank}(A)$, we have $\mathrm{dim}(\spn\{B\}\setminus \spn\{A\}) > 0$. Hence, the Lebesgue measure of
$\Omega$ is zero, and the probability is zero.
\end{proof}
Claim 1 shows that padding linear independent columns to a thin matrix leads to a different least squares solution with lower residue squared cost.
The convolution part in (\ref{eq1}) is equivalent to linear transforms:
\begin{equation}\label{eq6}
\bm y = \mathbf T_{\mathbf{k}}\bm{x} + \bm n = \mathbf T_{\mathbf{x}}\bm{k} + \bm n.
\end{equation}
where italic letters $\bm y, \bm x$, $ \bm k $ and $\bm n $ represent column-wise expanded vectors of 2D $\mathbf y, \mathbf x$, $ \mathbf k $ and $\mathbf n $, respectively; $\mathbf T_{\mathbf k}\in\mathbb{R}^{MN\times MN}$ and $\mathbf{T_{\mathbf x}}\in \mathbb{R}^{MN\times LK}$ are blocked banded Toeplitz matrices~\cite{andrews1977digital, gray2006toeplitz}; $L$ and $K$ are required to be odd.
We attribute the larger-kernel effect to either substep \eqref{eq3} or \eqref{eq4}. On one hand, $\mathbf T_{\mathbf k}$ remains identical when $L$ and $K$ increase by wrapping a layer of zeros around $\mathbf k$ and the result of x-step keeps the same. Hence, x-step should not be blamed as the source of the larger-kernel effect. On the other hand, when $\mathbf k$ is larger, $\mathbf T_{\mathbf x}$ will become inflated for the same $\mathbf x$. In 1D cases, where $N = K = 1$, assume $L = 2l + 1$, then
\begin{equation}\label{eq_Tx}
\begin{aligned}
&\mathbf T_{\bm{x}}(L) \\= &\left[\begin{matrix}
x_{l+1}& \cdots& x_2& x_1& 0& \cdots& 0 \\
\vdots & & \vdots& x_2& x_1& \ddots& \vdots \\
x_{M-1}& & \vdots& \vdots& x_2& \ddots& 0 \\
x_M& \ddots& \vdots& \vdots& \vdots& \ddots& x_1 \\
0& \ddots& x_{M-1}& \vdots& \vdots& & x_2 \\
\vdots& \ddots& x_M& x_{M-1}& \vdots& & \vdots \\
0& \cdots& 0& x_M& x_{M-1}& \cdots& x_{M-l}
\end{matrix}\right]
\end{aligned}
\end{equation}
During blind deconvolution iterations, for identical values of $\bm{\hat x}^{(i)}$, a larger $L$ introduces more columns onto both sizes of $\mathbf T_{\bm{\hat x}^{(i)}}$ and results in different
solutions. To illustrate this point, we tested a 1D version of blind deconvolution without kernel regularization and took different values of $L$ (truth and double and four times the truth size) for the 50th
k-step optimization after 49 truth-size iterations (see Figure ~\ref{fig2}). Figure ~\ref{fig2}(a-c) show that the optimal solutions in different sizes differ slightly on the main body
that lies within the ground truth size (colored in red), but greatly outside this range (colored in green) where zeros are expected. Figure ~\ref{fig2}(d-f) compare
ground truth to estimated kernels in (a-c) after non-negativity and sum-to-one projections.
Larger sizes yield more positive noises; hence, they lower the
weight of the main body after projections and change the outlook of estimated kernel.
\begin{figure}[t]
\begin{center}
\setlength{\tabcolsep}{5pt}
\small
\begin{tabular}{cccc}
& \includegraphics[width=.4\linewidth]{sz.eps} &
\includegraphics[width=.4\linewidth]{psz} \vspace{0pt} \\
& (a) & (d)\\
\rotatebox{90}{Estimated kernels} & \includegraphics[width=.4\linewidth]{sz2} &
\includegraphics[width=.4\linewidth]{psz2} \vspace{0pt} \\
& (b) & (e) \\
&
\includegraphics[width=.4\linewidth]{sz4} &
\includegraphics[width=.4\linewidth]{psz4} \\
& (c) & (f) \\
& index & index \\
\end{tabular}\\
\end{center}
\caption{Estimated kernels in 1D Blind deconvolution simulations. Left column are optimized kernels of different sizes after 50\small{th} iteration. Right column are corresponding normalized kernels of left after non-negativity and sum-to-one projections. In this experiment, $\bm x$ is a 255$\times$1 vector extracted from a real image and the truth $\bm k$ is generated by marginalizing a $23\times23$ truth kernel from Levin's dataset~\cite{levin2009understanding}. The signal prior is $\frac{\ell_1}{\ell_2}$. This figure is recommended to view in color.}
\label{fig2}
\end{figure}
\subsection{Probability of larger-kernel effect}
Even if $\bm{\hat x}^{(i)}$ successfully iterates to truth $\bm{x}$, Claim 1 implicates
the larger-kernel effect remains under the existence of random noise $\bm n$.
We show
\begin{equation}\label{eq_pro_equals_1}
\Pr \left(\mathrm{rank}\left(\mathbf T_{\bm{x}}\left(L+2\right)\right) > \mathrm{rank} \left(\mathbf T_{\bm{x}} \left(L\right)\right)\right) = 1,
\end{equation}
under which, the \textit{inflating effect} holds for probability one in blind deconvolution.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=\linewidth]{siyao3_h.pdf}
\end{center}
\caption{Quantitative simulations show that error increases with kernel size.
%
(a) The extracted row from a clear image.
%
(b) Singular boundaries of $\mathbf T_{\bm{x}}^\dagger$ and $\|\mathbf T_{\bm{x}}^\dagger \bm{n}\|$ with sampled $\bm n$. (c) Synthetic sparse signals. (d) The smallest, greatest and mean singular values of $(\mathbf T_{\bm x_{gt}}+\mathbf T_{\Delta \bm x})^\dagger \mathbf T_{\bm x_{gt}} - \mathbf I$.}
\label{fig3}
\end{figure*}
Above all, we have
\begin{equation}\label{eq_prob_trans}
\begin{aligned}
&\Pr \left(\mathrm{rank}\left(\mathbf T_{\bm{x}}\left(L+2\right)\right) > \mathrm{rank} \left(\mathbf T_{\bm{x}} \left(L\right)\right)\right)
\\ \ge &\Pr\left(\mathrm{rank}\left(\mathbf T_{\bm{x}}\left(M\right)\right)=M\right).
\end{aligned}
\end{equation}
Kaltofen and Lobo~\cite{kaltofen1996rank} proved
that for an M-by-M Toeplitz matrix composed of finite filed of $q$ elements,
\begin{equation}\label{eq_finite_prob}
\Pr\big(\mathrm{rank}(\mathbf T_{M\times M}) = M\big) = 1-1/q.
\end{equation}
Herein, clear images are statistically sparse on derivative fields~\cite{olshausen1996emergence, weiss2007makes}, and elements of $\bm x$ are modeled to be continuous in hyper-Laplacian distributions~\cite{krishnan2009fast}:
\begin{equation}\label{eq_pdf}
p({x}) = \begin{cases}
\beta \exp\left(-\gamma |x|^{\alpha}\right) & ,x \in [-1, 1] \\
0 &, otherwise.
\end{cases}
\end{equation}
Then we get the following claim:
\begin{thm}
$$ \Pr\big(\mathrm{rank}(\mathbf T_{\bm{x}}(M)) = M\big) = 1.$$
\end{thm}
\begin{proof}
See supplementary file.
\end{proof}
To now, we have shown that for $\mathbf{x}$ in sparse distribution, the \textit{inflating effect} happens almost surely.
\subsection{Quantification of error increment}
Assume $\mathbf{\hat x}$ iterates to ground truth $\mathbf{x}_{gt}$ during iterations. Then, for estimated kernel $\mathbf{\hat k}$, we have
\begin{equation}\label{eq_first_noise_k}
\begin{aligned}
\hat{\bm{k}} = \arg\min_{\bm k}{\|\mathbf T_{\bm x_{gt}}\bm k - \bm y\|^2}
= \bm k_{gt} + \mathbf T_{\bm x_{gt}}^\dagger \bm n
\end{aligned}
\end{equation}
where $\dagger$ represents Moore-Penrose pseudo-inverse. Then,
\begin{equation}\label{eq_first_noise_ssd}
SSD=\|\bm {\hat k} - \bm k_{gt} \|^2 = \|\mathbf T_{x_{gt}}^\dagger \bm n\|^2.
\end{equation}
Assume $\|\bm n\| = 1$, then
\begin{equation}\label{eq_first_bound}
s_1(\mathbf T_{\bm x_{gt}}^\dagger) \le \|\mathbf T_{\bm x_{gt}}^\dagger \bm n\| \le s_n(\mathbf T_{\bm x_{gt}}^\dagger),
\end{equation}
where $s_1$ and $s_n$ represents the smallest and the greatest singular values, respectively.
The \textit{inflating effect} implicates that a larger kernel size amplifies the error in $\mathbf {\hat k}$ due to noise $\mathbf n$. To quantify this increment, we extracted a line $\bm x$ from a clear image in Levin's set~\cite{levin2009understanding} as shown in Figure ~\ref{fig3}(a), and plotted $s_1(\mathbf T_{\bm{x}}^\dagger)$ and $s_n(\mathbf T_{\bm{x}}^\dagger)$ with increasing kernel size $L$. We also generated normalized random Gaussian vectors $\bm n$ and compared $\|\mathbf T_{\bm x} \bm n\|$ to simulated boundaries of singular values (see Figure ~\ref{fig3}(b)). The error in $\mathbf{\hat k}$ increases hyper-linearly with kernel size.
In practice, nuances are expected between $\mathbf{\hat{x}}$ and $\mathbf{x}_{gt}$. Cho and Lee~\cite{cho2016convergence} indicated that $\mathbf{\hat x}$ should be regarded as a sparse approximation to $\mathbf{x}_{gt}$, not the ground truth. Hence,
\begin{equation}\label{eq_not_same}
\mathbf{\hat x} = \mathbf{x}_{gt} + \Delta \mathbf x,
\end{equation}
which yields implicit noise~\cite{shan2008high}.
Assume $\mathbf n = \mathbf 0$, then,
\begin{equation}\label{eq_second_noise_conv}
\begin{aligned}
\mathbf{\hat x} \otimes \mathbf{k}_{gt} = \mathbf x_{gt} \otimes \mathbf k_{gt} + \Delta \mathbf x \otimes \mathbf k_{gt}
\end{aligned}
\end{equation}
and
\begin{equation}\label{eq_second_k}
\begin{aligned}
\bm{\hat k} = & \arg\min_{\bm k}\|\mathbf T_{\bm {\hat x}} \bm k - \bm y\|^2 \\
= & (\mathbf T_{\bm x_{gt}} + \mathbf T_{\Delta \bm x})^\dagger \mathbf T_{\bm x_{gt}}\bm{k}_{gt}.
\end{aligned}
\end{equation}
Then,
\begin{equation}\label{eq_second_ssd}
SSD=\|\left((\mathbf T_{\bm x_{gt}}+\mathbf T_{\Delta \bm x})^\dagger \mathbf T_{\bm x_{gt}}-\mathbf I\right)\bm{k}_{gt}\|^2.
\end{equation}
To quantify how singular values of $(\mathbf T_{\bm x_{gt}}+\mathbf T_{\Delta \bm x})^\dagger \mathbf T_{\bm x_{gt}} - \mathbf I$ changes with kernel size, we simulated 100 times, in each of which we generated a stochastic sparse signal $\bm x_{gt}$ with length 254 under PDF in (\ref{eq_pdf}) with $M = 254$, $\gamma = 10$ and $\alpha = 0.5$, and generated random Gaussian vector $\Delta \bm{x}$ where $ \|\Delta \bm x\| = \|\bm x_{gt}\| / 100$. Figure ~\ref{fig3}(c) shows one example of generated $\bm x_{gt}$ and $\Delta \bm{x}$. Figure ~\ref{fig3}(d) shows means and standard deviations of $s_1$, $s_n$ and $\bar s$, which is the average of singular values, of simulated $(\mathbf T_{\bm x}+\mathbf T_{\Delta \bm x})^\dagger \mathbf T_{\bm x} - \mathbf I$ on $L$. The error of $\mathbf{\hat k}$ is expected to grow with kernel size even $\mathbf n =0$.
\begin{figure*}[t]
\begin{center}
\includegraphics[width=0.9\linewidth]{siyao6m.pdf}
\end{center}
\caption{Singular values of clean kernels and noisy matrices. (a) The support domain (black) of a $47\times47$ random positive half Gaussian noise matrix. (b) The distribution of singular values of (a). (c) $\log\det$ costs of random Gaussian noise matrix (black), the truth kernel from~\cite{levin2009understanding} after zero-padding (red), and Gaussian PSF with $\text{size}/6$ standard deviation (blue) on kernel size. (d) Scaled (maximum to 1) singular value distributions of clean, impure and regularized kernels. }
\label{fig6}
\end{figure*}
\section{Low-rank regularization}
Blind deconvolution is an ill-posed problem for lacking sufficient information. Without regularization, MAP degrades to Maximum Likelihood (ML), which yields infinite solutions~\cite{levin2009understanding}. As prior information, kernel regularization should be designed to compensate the shortage of ML and to guide the optimization to expected results.
Great amount of studies focus on image regularization to describe natural images, e.g., Total Variation (TV-$\ell_1$)~\cite{levin2007image,wang2008new,ren2015fast}, hyper-Laplacian~\cite{krishnan2009fast}, dictionary sparsity~\cite{zhang2010bregmanized,hu2010single}, patch-based low rank prior~\cite{ren2016image}, non-local similarity~\cite{dong2013nonlocal} and deep discriminative prior~\cite{li2018learning}.
Unfortunately, kernel optimization doesn't attract much attention of the literature. Previous works adopted various kernel regularizations, e.g., $\ell_2$-norm~\cite{xu2010two, gong2016blind, cho2009fast, xu2013unnatural,pan2016blind}, $\ell_1$-norm~\cite{krishnan2011blind, shan2008high, pan2016robust} and $\ell_\alpha$-norm $(0<\alpha<1)$~\cite{zuo2015discriminative}, which, however, generally treated kernel regularization as an accessory and lacked a detailed discussion.
The larger-kernel effect is yielded by noise in ultra-sized kernels. Figure ~\ref{fig1} and Figure ~\ref{fig2} show that without kernel regularization, the main bodies of estimated kernels can emerge clearly, but increasing noises take greater amounts when $k$ is larger. To constrain $\hat k$ to be clean, regularization $h$ is expected to distinguish noise from ideal kernels efficiently.
To suppress the noise in estimated kernels, we take low-rank regularization on $k$ such that k-step \eqref{eq4} becomes
\begin{equation}\label{eq_low_rank_opt}
\mathbf{\hat{k}}^{(i+1)} = \arg\min_{\mathbf k}{\left(\|\mathbf{\hat{x}}^{(i+1)}\otimes \mathbf k- \mathbf y\|^2+\sigma \mathrm{rank}\left(\mathbf k\right)\right)}.
\end{equation}
Because the direct rank optimization is an NP-hard problem, continuous proximal functions are required. Fazel~\textit{et~al.}~\cite{fazel2003log} proposed
\begin{equation}\label{heuristic_proxy}
\log\det\left(\mathbf X + \delta \mathbf I\right)
\end{equation}
as a heuristic proxy for
$\mathbf X\in \mathbb{S}^N_+ $ where $\mathbf I$ is the N-by-N identity matrix and $\delta$
is a small positive number.
To allow this approximation to play a role in general matrices, the low-rank object is substituted to $(\mathbf X \mathbf X^T)^{1/2}$~\cite{dong2014compressive}. The regularization function then becomes
\begin{equation}\label{heuristic_proxy2}
h(\mathbf X)=\log{\det ((\mathbf X \mathbf X^T)^{\frac{1}{2}}+\delta \mathbf I)} = \sum_j \log(s_i + \delta),
\end{equation}
where $s_i$ is the $i$-th singular value of $X$.
Taking low-rank regularization on kernels is motivated by a generic phenomenon of noise matrices~\cite{andrews1977digital}.
Figure ~\ref{fig6}(a-b) shows a non-negative Gaussian noise matrix and its singular values in decreasing order. For a noise matrix, where light and darkness alternate irregularly, the distribution of singular values decays sharply at lower indices; then, it breaks and drag a relatively long and flat tail to the last.
In contrast, ideal kernels respond much lower to $\log\det$ regularization (see Figure ~\ref{fig6}(c)).
Based on this fact, noise matrices are distinguished by high $\log\det$ cost from real kernels. Figure ~\ref{fig6}(d) shows that singular values of a low-rank regularized kernel are distributed similarly as the ground truth, compared with the impure one.
One intelligible explanation on the low-rank property of ideal kernels is the continuity of blur motions. Rank of a matrix equals the number of independent rows or columns; it reversely reflects how similar these rows or columns are. Speed of a camera motion is deemed to be continuous~\cite{fang2014separable}. Hence, the local trajectory of a blur kernel emerges similar to neighbor pixels, which is measured in a low value by the continuous proxy of rank.
\begin{figure}[t]
\begin{center}
\includegraphics[width=0.8\linewidth]{insight.pdf}
\end{center}
\caption{Comparison on respond to noise. The cost ratio is calculated as $1+\frac{\text{cost}(\epsilon)-\text{cost}(0)}{|\text{cost}(0)|}$. This figure is recommended to view in color.}
\label{fig7}
\end{figure}
Compared to previous $\ell_{\alpha}$ norms, low-rank regularization responds more efficiently to noise. To illustrate this point,
we generated a noisy kernel by adding a small percentage ($\epsilon$) of non-negative Gaussian
noise and $1-\epsilon$ of the real kernel. Figure ~\ref{fig7} shows that the low-rank cost rapidly adjust favorably to the noise but $\ell_{\alpha}$ norms fail. That is because $\ell_\alpha$ only takes statistical information. An extreme example consists of disrupting a truth kernel and randomly reorganizing its elements, with $\ell_\alpha$ cost unchanged. In contrast, rank (singular values) corresponds to structural information.
\section{Optimization}
Function $\log\det$ is non-convex (and it is actually concave on $\mathbb{S}_+$). To solve the low-rank regularized least
squares ~\eqref{eq4},
we introduce an auxiliary variable $\Psi = k$ and reformulate the optimization into
\begin{equation}\label{2_optimiz}
\begin{aligned}
\min_{\mathbf k, \mathbf \Psi} & \left(\mathbf{\|\hat x}^{(i+1)}\otimes \mathbf \Psi-\mathbf y\|^2 +
\sigma\log\det\left(\left(\mathbf k \mathbf k\right)^{\frac{1}{2}}+\delta\mathbf{I}\right)\right) \\
\text{s.t. } & \mathbf \Psi = \mathbf k
\end{aligned}
\end{equation}
Using the Lagrange method, (\ref{2_optimiz}) is solved by two alternate sub-optimizations
\begin{equation}\label{eq:two_step}
\left\{
\begin{aligned}
\mathbf{\hat \Psi}^{(j+1)} &= \arg\min_{\mathbf \Psi} {\|\mathbf{\hat x}^{(i+1)}\otimes \mathbf \Psi-\mathbf y\|^2 + \mu \|\mathbf \Psi -
\mathbf {\hat k}^{(j)}\|^2 }\\
\mathbf{\hat k}^{(j+1)} &= \arg\min_{\mathbf k}{
\frac{1}{2\tau}\|\mathbf k-\mathbf{\hat\Psi}^{(j+1)}\|^2+\sigma h(\mathbf k)}
\end{aligned}
\right.
\end{equation}
where $j$ is the iteration number while $\mu$ and $\tau$ are trade-off parameters.
The $\Psi$-substep is convex and accomplished using the Conjugate Gradient (CG) method. For $k$-substep,
low rank is adopted with limit; otherwise, the regularization may change the
main body of kernel---an extreme result is $\mathbf{\hat k} = 0$. Thus, our strategy is to lower
the rank at $\mathbf{\hat \Psi}$ locally. Using the first-order Taylor expansion of $h$ at fixed matrix $\mathbf Z$:
\begin{equation}\label{taylor}
\begin{aligned}
h_{\mathbf Z}(\mathbf X)=h\left(\mathbf Z\right) + \sum_i\frac{s_i-\hat{s}_i}{\hat{s}_i+\delta},
\end{aligned}
\end{equation}
where $\hat{s}_i$ is the $i$-th eigenvalue of $\mathbf Z$, the k-substep in \eqref{eq:two_step} is transformed into an iterative optimization
\begin{equation}\label{final_eq}
\mathbf k^{(t+1)} = \arg\min_{\mathbf k}
\left(\frac{1}{2\tau}\|\mathbf k-\mathbf{\hat\Psi}^{(j+1)}\|^2+\sigma h_{\mathbf k^{(t)}}\left(\mathbf k\right) \right)
\end{equation}
where $t$ is the inner iteration number. For convenience, we set $\sigma$ as a flag (if $\sigma=0$, the k-substep will be skipped) and only tuned $\tau$ as the trade-off parameter.
Define the proximal mapping of function $\phi$ as follows:
\begin{equation}\label{prox1}
\prox_\phi(v) = \arg\min_u\left(\frac{1}{2}\|u-v\|^2 + \phi(u)\right).
\end{equation}
Dong~\textit{et~al.}~\cite{dong2014compressive} proved that one solution to the proximal mapping of $\tau h_{\mathbf Z}$ is
\begin{equation}\label{prox2}
\prox_{\tau
h_{\mathbf Z}}\left(\mathbf X\right)=\mathbf U\left(\mathbf \Sigma-\tau\mathrm{diag}\left(\bm{w}\right)\right)_+\mathbf V^T
\end{equation}
where
$\mathbf U\mathbf \Sigma \mathbf V^T$ is SVD of $\mathbf X$, $w_i = 1/\left(\hat{s}_i+\delta \right) $
and
${(\cdot)}_+=\max{\{ \cdot,0 \}}$.
Local low-rank optimization is implemented as iterations via the given parameter $\tau$ (see Algorithm~\ref{algk}).
In our implementation, $\mu$ is designed to exponentially grow with $j$ to allow more freedom of $\mathbf{\hat \Psi}$ for early iterations.
\emph{Overall Implementation.} We took deconvolution sechme in~\cite{krishnan2011blind} where $g = \ell_1/\ell_2$ (but with small modification) and applied non-blind deconvolution method proposed in~\cite{krishnan2009fast}.
\begin{algorithm}[t]
\caption{Updating k with low-rank regularization}\label{algk}
\begin{algorithmic}[1]
\Require{$\mathbf x$, $\mathbf y$, $\mu$, $\tau$, $OuterIterMax$, $CGIterMax$, $innerIterMax$}
\Ensure{$\mathbf{\hat k}$}
\For{$j\gets0$ to $OuterIterMax - 1$}
\If{$j=0$}
\State{$\mathbf{\hat \Psi}^{(j+1)}\gets \min_{\mathbf\Psi} \|\mathbf x\otimes\mathbf \Psi-\mathbf y\|^2 $ using CG
with maximum $CGIterMax$ iterations}
\Else
\State{$\mu^{(j)} \gets \mu \left(e^j/e^{OuterIterMax}\right)$}
\State{\small{$\mathbf{\hat \Psi}^{(j+1)}\gets \min_{\mathbf \Psi}{ \|\mathbf x \otimes \mathbf \Psi - \mathbf y\|^2 + \mu^{(j)} \|\mathbf \Psi - \mathbf k^{(j)}\|^2 }$} using CG with $CGIterMax$ iterations}
\EndIf
\State{Initializing $\mathbf k^{(0)}$ with all singular values equal to 1}
\For{$t\gets 0$ to $innerIterMax - 1$}
\State{$\mathbf k^{(t+1)}\gets \prox_{\tau h_{\mathbf k^{(t)}}}\left(\mathbf{\hat \Psi}^{(j)}\right)$}
\EndFor{}
\State{$\mathbf{\hat k}^{(j+1)}\gets \max\{\mathbf k^{(innerIterMax)}, 0\}$}
\State{$\mathbf{\hat k}^{(j+1)}\gets \mathbf{\hat k}^{(j+1)}/\sum \mathbf{\hat k}^{(j+1)}$}
\EndFor
\State{$\mathbf{\hat k}\gets \mathbf{\hat k}^{(outerIterMax)}$}
\end{algorithmic}
\end{algorithm}
\begin{algorithm}[t]
\caption{Blind Deconvolution (single-scaling version)}\label{algall}
\begin{algorithmic}[1]
\Require{blurry image $\mathbf y$, kernel size $L$, $\lambda$, $\eta$, $\tau$, $IterMax$}
\Ensure{clear image $x$ and degradation kernel $k$}
\State{$\mathbf y\gets\left[\mathbf \nabla_h{\mathbf y}, \mathbf \nabla_v{\mathbf y}\right]$}
\State{Initialize $\mathbf x \gets \mathbf y$}
\State{Initialize $\mathbf k$ with an $L\times L$ zero matrix adding [0.5 0.5] in the center}
\For{$t\gets 1$ to $IterMax$}
\State{Update $\mathbf x$ using Algorithm 3 in~\cite{krishnan2011blind}}
\State{Update $\mathbf k$ using Algorithm~\ref{algk}}
\EndFor{}
\State{$x\gets$ Non-blind deconvolution $(\mathbf k, \mathbf y)$}
\end{algorithmic}
\end{algorithm}
\section{Experimental Results}
In this section, we first discuss the effects of low rank-based regularization, then evaluate the proposed method on benchmark datasets, and finally demonstrate its effectiveness on real-world blurry images.
The source code is available at \url{https://github.com/lisiyaoATbnu/low_rank_kernel}.
\begin{figure}[b]
\centering
\setlength{\tabcolsep}{3pt}
\small
\begin{tabular}{ccc}
\includegraphics[width=.29\linewidth]{cxk1} &
\includegraphics[width=.29\linewidth]{cxk2} &
\includegraphics[width=.29\linewidth]{cxk3} \vspace{-4pt} \\
size=23, err=1.55 & size=47, err=1.56 & size=69, err=2.14 \vspace{1pt}\\
\end{tabular}\\
\caption{Deblurring results using low-rank regularization.}\label{fig8}
\end{figure}
\begin{figure*}[t]
\centering
\setlength{\tabcolsep}{1pt}
\small
\begin{tabular}{ccccccccccccccccccccc}
\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/blur11}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/blur11_31_2}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/blur11_61_rank2}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/blur11_61_non2}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/blur11_61_l22}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/blur11_61_l12}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/blur11_61_la2}} \vspace{-1pt}\\
\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/hanzi}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/hanzi_31}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/hanzi_61_rank}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/hanzi_61_non}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/hanzi_61_l2}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/hanzi_61_l1}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/hanzi_61_la}} \vspace{-1pt}\\
\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/nv}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/nv_31}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/nv_61_rank}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/nv_61_non}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/nv_61_l2}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/nv_61_l1}}
&\multicolumn{3}{c}{\includegraphics[width=0.13\linewidth]{final_fig8/nv_61_la}} \vspace{-1pt}\\
\multicolumn{3}{c}{Blurry}
& \multicolumn{3}{c}{$31\times31$}
& \multicolumn{3}{c}{Low rank}
& \multicolumn{3}{c}{None}
& \multicolumn{3}{c}{$\ell_2$}
& \multicolumn{3}{c}{$\ell_1$}
& \multicolumn{3}{c}{$\ell_\alpha$} \vspace{10pt}\\
\end{tabular}\\
\caption{Comparison of different kernel priors on real-world images. Large kernel size is $61\times61$. It's recommended to zoom in.}
\label{fig9}
\end{figure*}
\begin{figure}[t]
\centering
\setlength{\tabcolsep}{0.2pt}
\small
\begin{tabular}{cc}
\includegraphics[width=0.49\linewidth]{compare1}
& \includegraphics[width=0.49\linewidth]{compare2}\\
\small{(a)} & \small{(b)}\\
\end{tabular}\\
\caption{Success rates on truth (a) and double (b) sizes. }
\label{fig10}
\end{figure}
\subsection{Effects of low rank-based regularization}
Corresponding to high error ratios of large kernels in Figure ~\ref{fig1},
we repeat the experiment using same parameters except $\mu$ and $\sigma$. Figure ~\ref{fig8} shows
low-rank regularized kernels are much more robust to kernel size. Noises in kernels are efficiently reduced and qualities of restored images are enhanced.
We further verify it on real-world images by imposing different regularization terms. As in Figure ~\ref{fig9}, blur kernels with low-rank regularization have less noises, while the others suffer from strong noises, yielding artifacts in the deblurring images.
We note that in experiments of Figure ~\ref{fig8} and Figure ~\ref{fig9}, we deliberately omitted multi-scaling scheme to expose the effectiveness of low-rank regularization itself.
\subsection{Evaluation on synthetic dataset}
The proposed method is quantitatively evaluated on dataset from~\cite{levin2009understanding}. Figure ~\ref{fig10} shows the success rates of state-of-the-art methods versus our implementations with and without (set $\mu$ and $\sigma$ zero) low-rank regularization. The average PSNRs in Figure ~\ref{fig10} with different sizes are compared in Table~\ref{table:psnr}. Parameters are fixed during the whole experiment: $\sigma=1$, $\mu=1$, $\tau=5\times10^{-5}$, $OuterIterMax=20$, $CGIterMax=3$ and $innerIterMax=10$; a 7-layer multi-scaling pyramid is taken. Kernel elements smaller than 1/20 of the maximum are cut to zero, which is also taken in \cite{cho2009fast,krishnan2009fast}. Low-rank regularization works more effectively than the regularization-free implementation and the state-of-art.
\begin{table}[t]
\small
\centering
\begin{tabular}{cccc}
\hline
Method & prior & truth size & double size
\\
\hline
\cite{perrone2014total} &-- & 27.34 &23.29
\\\cite{cho2009fast} & $\ell_2$ &26.85 & 25.74
\\ \cite{xu2010two} & $\ell_2$ &26.91 & 26.71
\\ \cite{shan2008high} & $\ell_1$ & 26.54 & 26.44
\\\cite{krishnan2011blind} & $\ell_1$ &25.34 & 23.95
\\\cite{zuo2015discriminative} & $\ell_\alpha$ &26.58 & 26.83
\\\hline
$\sigma=0$ & -- & 26.68 & 23.85
\\$\sigma=1$ &$\log\det$ & \textbf{27.36} & \textbf{27.47} \\
\hline
\end{tabular} \vspace{10pt} \\
\caption{Average PSNRs (dB) with truth and double sizes in experiments of Figure ~\ref{fig10}.}
\label{table:psnr}
\end{table}
\begin{figure*}[t!]
\begin{center}
\setlength{\tabcolsep}{1pt}
\small
\begin{tabular}{cccc}
\includegraphics[width=0.21\linewidth]{roma_blur}
&\includegraphics[width=0.21\linewidth]{cho_roma_xkk.jpg}
&\includegraphics[width=0.21\linewidth]{cho_roma_xkk2.jpg}
&\includegraphics[width=0.21\linewidth]{xu_roma_xkk.jpg}
\vspace{-4pt}\\
Blurry & $\ell_2$~\cite{cho2009fast} $85\times85$ & $\ell_2$~\cite{cho2009fast} $185\times185$ & $\ell_2$~\cite{xu2010two} $85\times85$\\
\includegraphics[width=0.21\linewidth]{xu_roma_xkk2.jpg}
&\includegraphics[width=0.21\linewidth]{krishnan_roma_xkk.jpg}
&\includegraphics[width=0.21\linewidth]{krishnan_roma_xkk2.jpg}
&\includegraphics[width=0.21\linewidth]{zuo_roma_xkk.jpg} \vspace{-4pt}\\
$\ell_2$~\cite{xu2010two} $185\time185$ & $\ell_1$~\cite{krishnan2011blind} $85\times85$ & $\ell_1$~\cite{krishnan2011blind} $185\time185$ & $\ell_\alpha$~\cite{zuo2015discriminative} $85\times85$ \\
\includegraphics[width=0.21\linewidth]{zuo_roma_xkk2.jpg}
&\includegraphics[width=0.21\linewidth]{ours_roma_xkk.jpg}
&\includegraphics[width=0.21\linewidth]{ours_roma_non_xkk2.jpg}
&\includegraphics[width=0.21\linewidth]{ours_roma_xkk2.jpg} \vspace{-4pt}\\
$\ell_\alpha$~\cite{zuo2015discriminative} $185\times185$ & Low rank (ours) $85\times85$ &None (ours) &Low rank (ours) $185\times185$\\
\end{tabular}\\
\end{center}
\caption{Test on real-world image \texttt{roma}. Each domain (positive parts) of estimated kernel is displayed at the bottom right corner of corresponding restored image.}
\label{fig11}
\end{figure*}
\begin{figure*}[t!]
\begin{center}
\setlength{\tabcolsep}{2pt}
\small
\begin{tabular}{ccc}
\includegraphics[width=0.26\linewidth]{postcard.jpg}
&\includegraphics[width=0.26\linewidth]{cho_postcard_xkk.jpg}
&\includegraphics[width=0.26\linewidth]{xu_postcard_xkk.jpg} \vspace{-2pt}\\
Blurry & $\ell_2$, 1/20 max threshold~\cite{cho2009fast} & $\ell_2$, heuristic domain detector~\cite{xu2010two} \\
\includegraphics[width=0.26\linewidth]{krishnan_postcard_xkk.jpg}
&\includegraphics[width=0.26\linewidth]{zuo_postcard_xkk.jpg}
&\includegraphics[width=0.26\linewidth]{ours_last_tune.png}
\vspace{-2pt}\\
$\ell_1$~\cite{krishnan2011blind}, 1/20 max threshold & $\ell_\alpha$, none~\cite{zuo2015discriminative} & Low rank, 1/20 max threshold\\
\end{tabular}\\
\end{center}
\caption{Test on real-world image \texttt{postcard}. Kernel regularizations are listed under restored images.}
\label{fig12}
\end{figure*}
\subsection{Evaluation on real-world blurry images}
We compared our implementation to state-of-the-art methods on real-world images to reveal the robustness of low rank regularization on large kernel size. Specifically, \cite{xu2010two} takes a heuristic iterative support domain detector based on the differences of elements of $\hat k$, which is regarded to be more effective than 1/20 threshold. Figure ~\ref{fig11} shows that $185\times185$ size yields strong noises in estimated kernels of previous works~\cite{cho2009fast,xu2010two}, and even changes main bodies of kernels~\cite{krishnan2011blind,zuo2015discriminative}. In contrast, low rank regularization can keep the kernel relatively stable for the larger size. One more comparison of different regularizations and refinement methods on large kernel size are shown in Figure~\ref{fig12}.
As for computational efficiency of our method, it takes about 85s on a Lenovo ThinkCentre computer with Core i7 processor to process images with size $255\times 255$.
\section{Conclusion}
In this paper, we demonstrate that over-estimated kernel sizes produce increased noises in estimated kernel. We attribute the larger-kernel effect to the \textit{inflating effect}. To reduce this effect, we propose a low-rank based regularization on kernel, which could suppress noise while remaining restored main body of optimized kernel.
The success of blind deconvolution is contributed by many aspects. In practical implementations, even for noise-free $\mathbf y$, the intermediate $\mathbf{\hat x}^{(i)}$ is unlikely to iterate to ground truth, hence some parts of $\mathbf y$ will be treated as implicit noises, which may intensify the effect even more than expected and require future researches.
\section*{Acknowledgement}
This work is supported by the grants from the National Natural Science Foundation of China (61472043) and the National Key R\&D program of China (2017YFC1502505).
We thank Ping Guo for constructive conversation.
Qian Yin is the corresponding author.
\newpage
{\small
\bibliographystyle{ieee}
| e4b122552cded5ecbae19dbbc59de647f26a2fb7 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
\label{intro}
The central goal of the beam energy scan (BES) program at RHIC is to seek
evidence for the existence of a critical point in the phase
diagram of strong-interaction matter. The hope is to detect this
postulated second order
phase transition point through the analysis of higher order cumulants
of net charge fluctuations. Maxima of cumulants of net charge fluctuations,
e.g. the $2^{nd}$ and $4^{th}$ order cumulants, trace the chiral
crossover transition line at small values of the baryon chemical potential
and
diverge at a critical point.
In heavy ion experiments the observed
net charge fluctuations are expected to reflect thermal conditions
at the time of chemical-freeze out of various hadron species.
If this freeze-out happens close to the pseudo-critical
line for the chiral transition of QCD, where thermal fluctuations are
large, the measured fluctuations
have a chance to be indicative for the divergent fluctuations that will
appear at a critical point in the QCD phase diagram.
A crucial anchor point for this scenario is to establish the relation
between freeze-out and the QCD chiral transition at small or even
vanishing net baryon chemical potential. In this case reliable theoretical
calculations, based on Taylor series expansions in lattice QCD, exist
and can be confronted with experimental findings at the LHC as well as
the highest beam energies at RHIC. In experiments at the LHC one
can analyze moments of charge fluctuations at almost vanishing
baryon chemical potential ($\mu_B$) which allows a direct comparison with
lattice QCD calculations performed at $\mu_B=0$.
\section{Thermal conditions at vanishing net baryon number}
The basic bulk thermodynamic observables,
pressure ($P$), energy ($\epsilon$) and entropy ($s$) density
of strong-interaction matter
at vanishing baryon chemical potential, have been calculated in lattice
QCD (for a recent review see: \cite{Ding:2015ona}). These calculations have
recently been extended to non-vanishing
baryon number densities using analytic continuation of calculations
performed at imaginary values of $\mu_B$ \cite{Gunther:2016vcp}
and Taylor expansions in $\mu_B$ \cite{Bazavov:2017dus}.
In Fig.~\ref{fig:phasediagram}~(left) we show results for lines of
constant $P$, $\epsilon$ and $s$, in the $T$-$\mu_B$ plane
(phase diagram) obtained from a Taylor series up to
${\cal O}(\mu_B^4)$ \cite{Bazavov:2017dus}. Lines are drawn for three values
of these observables in the crossover region for the QCD chiral transition,
which is well characterized by the current uncertainty on the chiral
transition temperature, $T_c=154(9)$~MeV, at $\mu_B=0$ \cite{Bazavov:2011nk}.
As can be seen in Fig.~\ref{fig:phasediagram}~(right) in this temperature
interval the energy density
changes by about a factor three,
$\epsilon_c = (0.34\pm 0.16)$~GeV/fm$^3$ \cite{Bazavov:2014pvz}.
Also shown in Fig.~\ref{fig:phasediagram}~(left) are experimental
results for freeze-out parameters determined by the ALICE Collaboration
at the LHC \cite{Floris:2014pta} and the STAR Collaboration from the BES
at RHIC \cite{Das:2014qca} by comparing measured particle
yields with predictions from a statistical hadronization model, which utilizes
the thermodynamics of a hadron resonance gas. Obviously, there is a
significant difference in the determination of the freeze-out
temperature at $\mu_B\simeq 0$.
While the ALICE result for the freeze-out temperature ($T_f$) agrees well with
the central value of the pseudo-critical temperature ($T_c$), the STAR results
favor a larger value, $T_f\sim 165$~MeV, which is close to the
hadronization temperature obtained by Becattini et al.
\cite{Becattini:2016xct}.
\begin{figure}[t]
\begin{center}
\hspace*{-0.1cm}\includegraphics[width=84mm]{pressure4_wide_LCP.pdf}\hspace*{-0.4cm}
\includegraphics[width=72mm]{energy_crit.pdf}
\caption{{\it Left:} Lines of constant pressure, energy and entropy density,
as given in Table II of Ref.~\cite{Bazavov:2017dus},
as function of baryon chemical potential. Solid black lines indicate the
current uncertainty on the variation of the pseudo-critical temperature of
the chiral transition, $T_c(\mu_B)$
with $\mu_B$. For a discussion of the data points see text.
{\it Right:}
The energy density at $\mu_B=0$ as function of temperature. The box
reflects current errors on the crossover transition temperature,
$T_c=154(9)$~MeV.
\vspace*{-0.5cm}
}
\label{fig:phasediagram}
\end{center}
\end{figure}
A 10~MeV accuracy for the determination of the freeze-out
temperature, which anyhow is not considered to be a
temperature uniquely defined for all particle species, but rather a statistical
average, may be considered to be appropriate for many purposes. However,
in the search for evidence for a critical point such a
difference has substantial consequences for expected properties of
net charge fluctuations as the size of the critical region, in which charge
fluctuations may become large, may well be only of that order
\cite{Schaefer:2006ds}. A 10~MeV difference between $T_c$ and $T_f$ thus may
decide whether or not freeze-out happens in the critical region.
Cumulants of net charge fluctuations and correlations among
fluctuations of different conserved charges, i.e. baryon number ($B$),
electric charge ($Q$) and strangeness ($S$) can be obtained as
derivatives of the logarithm of the QCD partition function \cite{Bazavov:2017dus},
\begin{equation}
\chi_{n}^{X} = \left. \frac{1}{VT^3}
\frac{\partial^{n} \ln Z(V,T,\vec{\mu})}{\partial X^n}
\right|_{\vec{\mu}=0}\;\; ,\;\;
\chi_{nm}^{XY} = \left. \frac{1}{VT^3}
\frac{\partial^{(n+m)} \ln Z(V,T,\vec{\mu})}{\partial X^n\ \partial Y^m}
\right|_{\vec{\mu}=0}\;\; ,\;\; X,\ Y=B,\ Q,\ S \;\; ,
\label{cumulants}
\end{equation}
with $\vec{\mu} \equiv (\mu_B,\ \mu_Q,\ \mu_S)$ denoting the three
chemical potentials connected with the conserved charges.
As can be seen in Fig.~\ref{fig:chiratios} ratios of cumulants of
net charge fluctuations change rapidly in the crossover region.
The ratio of $4^{th}$ and $2^{nd}$ order net baryon-number fluctuations,
$\chi_4^B/\chi_2^B$,
changes from almost unity at $T=145$~MeV to about 0.5 at $T=165$~MeV
(see Fig.~\ref{fig:chiratios}~(left)). In the temperature range identified
by ALICE as the freeze-out region, $T_f=156(2)$~MeV, the ratio is
$\chi_4^B/\chi_2^B\simeq 0.75$. The situation is similar for ratios of
conserved charge correlations. For the ratio of $4^{th}$ and $2^{nd}$ order
cumulants characterizing correlations between
net strangeness and net baryon-number fluctuations one finds
$\chi_{31}^{BS}/\chi_{11}^{BS}\simeq 0.74$ at $T=155$~MeV, while
this ratio drops to $\sim 0.36$ at $T=165$~MeV. This is in contrast to
HRG model calculations with point-like non-interacting hadrons, where
these ratios are unity irrespective of the particle content in the hadron
spectrum.
Also on the level of second order cumulants, which currently get
measured by the ALICE Collaboration \cite{Rustamov:2017lio},
differences between HRG model
calculations and QCD results are already significant. In
Fig.~\ref{fig:chiratios}~(right) we show the ratio, $\chi_{11}^{BQ}/\chi_2^B$,
which in resonance gas models for non-interacting, point-like hadrons has the
interpretation of the relative contribution of charged baryons to the
total baryon contribution of the pressure. As can be seen at $T=165$~MeV
the ratio $\chi_{11}^{BQ}/\chi_2^B$ is about 50\% smaller than predicted
by HRG model calculations based on the experimentally known hadron spectrum
(PDG-HRG). Model calculations that include additional strange baryon
resonances predicted in QCD motivated quark model calculations (QM-HRG)
provide a better approximation to QCD calculations but still over-predict
the ratio $\chi_{11}^{BQ}/\chi_2^B$ for $T\raise0.3ex\hbox{$>$\kern-0.75em\raise-1.1ex\hbox{$\sim$}} 150$~MeV. A similar
behavior has been found for the correlations between net baryon-number
and net strangeness fluctuations, $\chi_{11}^{BS}$, normalized to either
$\chi_2^B$ or $\chi_2^S$. This may be taken as evidence for the importance
of additional strange baryon resonance contributing significantly
to the thermodynamics close to the chiral transition.
There is, of course,
also the possibility that additional interactions among hadrons need to
be taken into account in order to improve agreement between QCD and hadron
gas calculations at low temperature. However, as is also evident from
Fig.~\ref{fig:chiratios}, it seems that neither
taking into account
repulsive interactions through
the finite volume of baryons \cite{Andronic:2012ut}
nor an additional
attractive contribution through
a van der Waals interaction (vdW-HRG) can improve the
validity range of HRG model based calculations \cite{Vovchenko:2016rkn}.
\begin{figure}[t]
\begin{center}
\hspace*{-0.1cm}\includegraphics[width=72mm]{chiB4_B2_HRG_fK1-eps-converted-to.pdf}\hspace*{-0.1cm}
\includegraphics[width=72mm]{chiB1Q1_B2_HRG_large-eps-converted-to.pdf}
\caption{
The ratio of fourth and second order cumulants of net baryon-number
fluctuations (left) and the correlations between
net electric charge and net baryon-number fluctuations
normalized to the second order cumulant of net baryon-number fluctuations
(right). The lines show results obtained in various hadron resonance gas
model calculations (see text).
\vspace*{-0.5cm}
}
\label{fig:chiratios}
\end{center}
\end{figure}
\section{Cumulants of conserved charge fluctuations at non-zero net
baryon number}
Ratios of $4^{th}$ and $2^{nd}$ order cumulants such as
$\chi_4^B/\chi_2^B$ (Fig.~\ref{fig:chiratios}~(left))
or $\chi_{31}^{BX}/\chi_{11}^{BX}$, $X=Q,\ S$,
also give the leading ${\cal O}(\mu_B^2)$
corrections to the $2^{nd}$ order cumulants themselves. For
$\mu_Q=\mu_S=0$ one has
\begin{equation}
\chi_{2}^{B}(T,\mu_B) = \chi_{2}^{B} + \frac{1}{2}\chi_{4}^{B}
\left( \frac{\mu_B}{T} \right)^2+ {\cal O}(\mu_B^4)\;\;\;\; ,\;\;\;\;
\chi_{11}^{BX}(T,\mu_B) = \chi_{11}^{BX} + \frac{1}{2}\chi_{31}^{BX}
\left( \frac{\mu_B}{T} \right)^2+ {\cal O}(\mu_B^4) \;\;,\;\; X=Q, \; S \; .
\label{expansion}
\end{equation}
As these ratios are smaller than the corresponding HRG
values, this also means that differences between lattice QCD results for
these cumulants and HRG model calculations increase with $\mu_B$. Some
results for $\chi_{11}^{BS}$ and $\chi_{11}^{BQ}$ at three values
of the temperature are shown in Fig.~\ref{fig:chi42}. Obviously
deviations from HRG model calculations become large for $\mu_B/T \raise0.3ex\hbox{$>$\kern-0.75em\raise-1.1ex\hbox{$\sim$}} 1.5$,
which in the BES at RHIC corresponds to $\sqrt{s_{NN}} \raise0.3ex\hbox{$<$\kern-0.75em\raise-1.1ex\hbox{$\sim$}} 15$~GeV.
The situation is similar for ratios of cumulants
of net baryon-number fluctuations.
For instance,
\begin{equation}
\frac{\chi_4^B(T,\mu_B)}{\chi_2^B(T,\mu_B)}
= \frac{\chi_4^B}{\chi_2^B}\left( 1 + \frac{1}{2} \left(
\frac{\chi_6^B}{\chi_4^B} - \frac{\chi_4^B}{\chi_2^B} \right)
\left( \frac{\mu_B}{T} \right)^2 + {\cal O}(\mu_B^4) \right)
\;\;\; {\rm for} \;\;\; \mu_Q=\mu_S=0\; .
\label{B42}
\end{equation}
is unity in HRG model calculations with non-interacting, point-like hadrons
for all values of $\mu_B$. In QCD, however, the ${\cal O}(\mu_B^2)$
expansion coefficient in Eq.~\ref{B42} is negative
for $T\raise0.3ex\hbox{$>$\kern-0.75em\raise-1.1ex\hbox{$\sim$}} 150$~MeV \cite{Karsch:2015nqx}. This indicates that
the ratio $\chi_4^B(T,\mu_B) / \chi_2^B(T,\mu_B)$ becomes smaller
with increasing $\mu_B$. Even if the ratio is close to HRG model
results at $\mu_B=0$, it thus will further deviate from this with
increasing $\mu_B$. This is consistent with the trend found by
STAR for the corresponding kurtosis ratio of net proton-number
fluctuations, $\kappa_P\sigma^2_P=\chi_4^P/\chi_2^P$ \cite{Thader:2016gpa}.
\begin{figure}[t]
\begin{center}
\hspace*{-0.1cm}\includegraphics[width=72mm]{B1S1_O5_muB-eps-converted-to.pdf}\hspace*{-0.1cm}
\includegraphics[width=72mm]{B1Q1_O5_muB-eps-converted-to.pdf}
\caption{ Correlation between net baryon-number fluctuations
and net strangeness fluctuations (left) as well as net electric charge
fluctuations (right) for three values of the temperature. Shown are results
from a Taylor expansion up to ${\cal O}(\mu_B^2)$ with $\mu_Q=\mu_S=0$.
Also shown are results for a hadron resonance gas in the Boltzmann
approximation.
\vspace*{-0.5cm}
}
\label{fig:chi42}
\end{center}
\end{figure}
\section{Conclusions}
Many ratios of $4^{th}$ and $2^{nd}$ order cumulants of conserved charge
fluctuations calculated in QCD
agree with HRG model calculations within 25\% or better at temperatures
below $T=155$~MeV. For observables including strangeness fluctuations this
often requires to take into account contributions from additional strange
hadrons not listed in the Particle Data Tables. At $T=165$~MeV
HRG model calculations often deviate from QCD results by more than 50\%
even for observables that only involve quadratic charge fluctuations.
\vspace*{0.2cm}
\noindent
{\it Acknowledgements:}
This work was supported through Contracts No. DE-SC001270 with the
U.S. Department of Energy and No. 05P15PBCAA with the
German Bundesministerium f\"ur Bildung und Forschung.
\bibliographystyle{elsarticle-num}
| 804d208b8369292be74df4bb754d39129cb6d9b1 | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
\section{Introduction}
Establishing causal relations between variables from observation of their behaviour in time is central to scientific investigation and it is at the core of data-science where these causal relations are the basis for the construction of useful models and tools capable of prediction.
The capability to predict (future) outcomes from the analytics of (past) input data is crucial in modeling and it should be the main property to take into consideration in model selection, when the validity and meaningfulness of a model is assessed.
From an high-level perspective, we can say that the whole scientific method is constructed around a circular procedure consisting in observation, modelling, prediction and testing. In such a procedure, the accuracy of prediction is used as a selection tool between models.
In addition, the principle of parsimony favours the simplest model when two models have similar predictive power.
The scientific method is the rational process that, for the last 400 years, has mostly contributed to scientific discoveries, technological progresses and the advancement of human knowledge.
Machine learning and data-science are nowadays pursuing the ambition to mechanize this discovery process by feeding machines with data and using different methodologies to build systems able to make models and predictions by themselves.
However, the automatisation of this process requires to identify, without the help of human intuition, the relevant variables and the relations between these variables out of a large quantity of data.
Predictive models are methodologies, systems or equations which identify and make use of such relations between sets of variables in a way that the knowledge about a set of variables provides information about the values of the other set of variables.
This problem is intrinsically high-dimensional with many input and output data.
Any model that aims to explain the underlying system will involve a number of elements which must be of the order of magnitude of the number of relevant relations between the system's variables.
In complex systems, such as financial markets or the brain, prediction is probabilistic in nature and modeling concerns inferring the probability of the values of a set of variables given the values of another set.
This requires the estimation of the joint probability of all variables in the system and, in complex systems, the number of variables with potential macroscopic effects on the whole system is very large.
This poses a great challenge for the model construction/selection and its parameter estimation because the number of relations between variables scales with -at least- the square of the number of variables but, \CR{for a given fix} observation window, the amount of information gathered from such variables scales -at most- linearly with the number of variables \cite{bruckstein2009sparse,theodoridis2012sparsity}.
For instance, a linear model for a system with $p$ variables requires the estimation from observation of $p(p+1)/2$ parameters (the distinct elements of the covariance matrix).
In order to estimate $\mathcal O(p^2)$ parameters one needs a comparable number of observations requiring time series of length $q \sim p$ or larger to gather a sufficient information content from a number of observations which scales as $p \times q \sim \mathcal O(p^2)$.
However, the number of parameters in the model can be reduced by considering only $\mathcal O(p)$ out of the $\mathcal O(p^2)$ relations between the variables reducing in this way the required time series length to $\mathcal O(p)$.
Such models with reduced numbers of parameters are referred in the literature as sparse models.
In this paper we consider two instances of linear sparse modelling: Glasso \cite{tibshirani1996} which penalizes non-zero parameters by introducing a $\ell_1$ norm penalization and LoGo \cite{LoGo16} which reduces the inference network to an $\mathcal O(p)$ number of links selected by using information filtering networks \cite{asteetal2005,tumminelloetal2005,TMFG}.
The results from these two sparse models are compared with the $\ell_2$ norm penalization (non-sparse) ridge model \cite{tikhonov1963solution,hoerl1970ridge}.
{
This paper is an exploratory attempt to map the parameter-regions of time series length, number of variables, penalization parameters and kinds of models to define the boundaries where probabilistic models can be reasonably constructed from the analytics of observation data.
In particular, we investigate empirically, \CR{by means of a linear autoregressive model with sparse inference structure,} the true \CR{causality} link retrieval performances in the region of short time-series and large number of variables which is the most critical region -- and the most interesting -- in many practical cases.
\CR{Causality is defined in information theoretic sense as a significant reduction on uncertainty over the present values of a given variable provided by the knowledge of the past values of another variable obtained in excess to the knowledge provided by the past of the variable itself and --in the conditional case-- the past of all other variables \cite{zaremba2014measures}. We measure such information by using transfer entropy and, within the present linear modelling, this coincides with the concept of Granger causality and conditional Granger causality \cite{schreiber2000measuring}.
The use of transfer entropy has the advantage of being a concept directly extensible to non-linear modelling.
However, non-linearity is not tackled within the present paper.
Linear models with multivariate normal distributions have the unique advantage that causality and partial correlations are directly linked, largely simplifying the computation of transfer entropy and directly mapping the problem into the sparse inverse covariance problem \cite{tibshirani1996,LoGo16}.
}
Results are reported for artificially generated time series from an autoregressive model of $p=100$ variables and time series lengths $q$ between 10 and 20,000 data points.
Robustness of the results has been verified over a wider range of $p$ from 20 to 200 variables.
Our results demonstrate that sparse models are superior in retrieving the true causality structure for short time series.
Interestingly, this is despite considerable inaccuracies in the inference network of these sparse models.
We indeed observe that statistical validation of causality is crucial in identifying the true causal links, and this identification is highly enhanced in sparse models.
}
The paper is structured as follows.
In section \ref{s.definitions} we briefly review the basic concepts of mutual information and conditional transfer entropy and their estimation from data that will then be used in the rest of the paper.
We also introduce the concepts of sparse inverse covariance, inference network and causality networks.
Section \ref{s.causalNet} concerns the retrieval of causality network from the computation and statistical validation of conditional transfer entropy.
Results are reported in Section \ref{s.results} where the retrieval of the true causality network from the analytics of time series from an autoregressive process of $p=100$ variables is discussed.
Conclusions and perspectives are given in Section \ref{s.conclusions}.
\section{Estimation of conditional transfer entropy from data}
\label{s.conditionslSigma}\label{s.definitions}
In this paper causality is quantified by means of statistically-validated transfer entropy.
Transfer entropy $T(\mathbf{ Z_i} \rightarrow \mathbf{ Z_j})$ quantifies the amount of uncertainty on a random variable, $\mathbf{ Z_j}$, explained by {the past} of another variable, $\mathbf{Z_i}$ conditioned to the knowledge about the past of $\mathbf{ Z_j}$ itself.
Conditional transfer entropy, $T(\mathbf{ Z_i} \rightarrow \mathbf{ Z_j} | \mathbf W)$, includes an extra condition also to a set variables $\mathbf W$.
These quantities are introduced in details in Appendix \ref{cTE} (see also \cite{shannon2001mathematical,schreiber2000measuring,anderson1984multivariate}).
Let us here just report the main expression for the conditional transfer entropy that we shall use in this paper:
\begin{align}\label{General_cTE}
T(\mathbf{ Z_i} \rightarrow \mathbf{ Z_j}|\mathbf{ W}) & {}=
H(\mathbf Z_{\mathbf j,t} | \{\mathbf Z_{\mathbf j,t}^{lag},\mathbf{ W}_{t}\})
-
H(\mathbf Z_{\mathbf j,t} | \{\mathbf Z_{\mathbf i,t}^{lag},\mathbf Z_{\mathbf j,t}^{lag},\mathbf{ W}_{t}\}) \;\;.
\end{align}
Where $H(.|.)$ is the conditional entropy, $\mathbf Z_{\mathbf j,t}$ is a random variable at time $t$, whereas $\mathbf Z_{\mathbf i,t}^{lag} =\{ \mathbf Z_{\mathbf i,t-1},...,\mathbf Z_{\mathbf i,t-\tau}\}$ is the lagged set of random variable `$\mathbf i$' considering previous times $t-1...t-\tau$ and $\mathbf{ W}_{t}$ are all other variables and their lags (see Appendix \ref{cTE}, Eq.\ref{General_cTE1}).
In this paper we use Shannon entropy and restrict to linear modeling with multivariate normal setting (see Appendix \ref{Shannon}).
In this context the conditional transfer entropy can be expressed in terms of the determinants of conditional covariances $\mbox{det}( \mathbf \Sigma(.|.))$ (see Eq.\ref{H2} in Appendix \ref{Shannon}):
\begin{align}\label{General_cTE_Sigma}
T(\mathbf{ Z_i} \rightarrow \mathbf{ Z_j}|\mathbf{ W}) & {}=
\frac12 \log \mbox{det}\!\left( \mathbf \Sigma(\mathbf Z_{\mathbf j,t} | \{\mathbf Z_{\mathbf j,t}^{lag},\mathbf{ W}_{t}\})\right)
-
\frac12 \log \mbox{det}\!\left( \mathbf \Sigma(\mathbf Z_{\mathbf j,t} | \{\mathbf Z_{\mathbf j,t}^{lag},\mathbf Z_{\mathbf i,t}^{lag},\mathbf{ W}_{t}\})\right) \;\;.
\end{align}
Conditional covariances can be conveniently computed in terms of the inverse covariance of the whole set of variables
$\mathbf Z_t = \{ \mathbf Z_{k,t},\mathbf Z_{k,t-1},...\mathbf Z_{k,t-\tau}\}_{k=1}^p \in \mathbb R^{p\times(\tau+1)}$
(see Appendix \ref{InverseCovJandSigma}).
Such inverse covariance matrix, $\mathbf J$, represents the structure of conditional dependencies among all couples of variables in the system and their lags.
Each sub-part of $\mathbf J$ is associated with the conditional covariances of the variables in that part with respect to all others.
In terms of $\mathbf J$, the expression for the conditional transfer entropy becomes:
\begin{align}\label{TE20}
T(\mathbf{ Z_i} \rightarrow \mathbf{ Z_j}|\mathbf{ W}) = - \frac12 \log\mbox{det}\! \left(\mathbf {J_{1,1}} - \mathbf {J_{1,2}} (\mathbf {J_{2,2}})^{-1} \mathbf {J_{2,1}})\right)+ \frac12 \log \mbox{det}\! \left (\mathbf{ J_{1,1} }\right ) \;\;.
\end{align}
where the indices `$\mathbf 1$' and `$\mathbf 2$' refer to sub-matrices of $\mathbf J$ respectively associated with the variables $ \mathbf Z_{\mathbf j,t}$ and $\mathbf Z_{\mathbf i,t}^{lag}$.
\CB{
\subsection{Causality and inference networks}
The inverse covariance $\mathbf J$, also known as precision matrix, represents the structure of conditional dependencies.
If we interpret the structure of $\mathbf J$ as a network, where nodes are the variables and non-zero entries correspond to edges of the network, then we shall see that any two sub-sets of nodes that are not directly connected by one or more edges are conditionally-independent.
Condition is with respect to all other variables.
Links between variables at different lags are associated with causality with direction going from larger to smaller lags.
The network becomes therefore a directed graph.
In such a network entropies can be associated with nodes, conditional mutual information can be associated with edges between variables with the same lag and conditional transfer entropy can be associated to edges between variables with different lags.
A nice property of this mapping of information measures with directed networks is that there is a simple way to aggregate information which is directly associated with topological properties of the network.
Entropy, mutual information and transfer entropies can be defined for any aggregated sub set of nodes with their values directly associated to the presence, direction and weight of network edges between these sub-parts.
Non-zero transfer entropies indicating --for instance-- variable $\mathbf i$ causing variable $\mathbf j$ are associated with some non-zero entries in the inverse covariance matrix $\mathbf J$ between lagged variables $\mathbf i$ (i.e. $\mathbf{ Z_{i}}_{,t-\tau}$, with $\tau >0$) and variable $\mathbf j$ (i.e. $\mathbf{ Z_{j}}_{,t}$)).
\CR{In linear models, }these non-zero entries define the \CR{estimated} {\it inference network}.
However, not all edges in this inference network correspond to transfer entropies that are significantly different from zero.
To extract the structure of the {\it causality network} we shall retain only the edges in the inference network which correspond to statistically validated transfer entropies.
}
Conditioning eliminates the effect of the other variables retaining only the exclusive contribution from the two variables in consideration.
This should provide estimations of transfer entropy that are less affected by spurious effects from other variables.
On the other hand, conditioning in itself can introduce spurious effects, indeed two independent variables can become dependent due to conditioning \cite{anderson1984multivariate}.
In this paper we explore two extreme conditioning cases: i) condition to all other variables and their lags; ii) unconditioned.
{
In principle, one would like to identify the maximal value of $T(\mathbf{ Z_i} \rightarrow \mathbf{ Z_j} | \mathbf W)$ over all lags and all possible conditionings $\mathbf W$.
However, the use of multiple lags and conditionings increases the dimensionality of the problem making estimation of transfer entropy very hard especially when only a limited amount of measurements is available (i.e. short time-series).
This is because the calculation of the conditional covariance requires the estimation of the inverse covariance of the whole set of variables and such an estimation is strongly affected by noise and uncertainty.
Therefore, a standard approach is to reduce the number of variables and lags to keep dimensionality low and estimate conditional covariances with appropriate penalizers \cite{tikhonov1963solution,hoerl1970ridge,tibshirani1996,friedmanetal2008}.
An alternative approach is to invert the covariance matrix only locally on low dimensional sub-sets of variables selected by using information filtering networks \cite{asteetal2005,tumminelloetal2005,TMFG} and then reconstruct the global inversion by means of the LoGo approach \cite{LoGo16}.
Let us here briefly account for these two approaches. }
\subsection{Penalized inversions}
The estimate of the inverse covariance is a challenging task to which a large body of literature has been dedicated \cite{friedman2008sparse}.
From an intuitive perspective, one can say that the problem lies in the fact that uncertainty is associated with nearly zero eigenvalues of the covariance matrix.
Variations in these small eigenvalues have relatively small effects on the entries of the covariance matrix itself but have major effects on the estimation of its inverse.
Indeed small fluctuations of small values can yield to unbounded contributions to the inverse.
A way to cure such near-singular matrices is by adding finite positive terms to the diagonal which move the eigenvalues away from zero:
$\hat {\mathbf J} = \left((1-\gamma) \mathbf S +\gamma \mathbf I_N \right)^{-1}$, where $\mathbf S = \mbox{Cov}(\mathbf Z)$ is the covariance matrix of the set of variables $\mathbf Z \in \mathbb R^N$ estimated from data and $\mathbf I_{N}\in \mathbb R^{N\times N}$ is the identity matrix (where $N=p\times(\tau+1)$, see later).
This is what is performed in the so-called ridge regression \cite{hoerl1970ridge}, also known as shrinkage mean-square-error estimator \cite{gruber1998improving} or Tikhonov regularization \cite{tikhonov1963solution}.
The effect of the additional positive diagonal elements is quivalent to compute the inverse covariance which maximizes the log-likelihood: $\log \mbox{det}( \hat {\mathbf J}) - \mbox{tr}(\mathbf S\hat{\mathbf J}) - \gamma ||\hat {\mathbf J} ||_2$, where the last term penalizes large off-diagonal coefficients in the inverse covariance with a $\ell_2$ norm penalization \cite{witten2009covariance}.
The regularizer parameter $\gamma$ tunes the strength of this penalization.
This regularization is very simple and effective.
However, with this method insignificant elements in the precision matrix are penalized toward small values but they are never set to zero.
By using instead $\ell_1$ norm penalization: $\log \mbox{det}( \hat {\mathbf J}) - \mbox{tr}(\mathbf S\hat{\mathbf J}) - \gamma ||\hat{\mathbf J} ||_1$, insignificant elements are forced to zero leading to a sparse inverse covariance.
This is the so-called lasso regularization \cite{tibshirani1996,meinshausenbuehlmann2006,friedmanetal2008}.
The advantage of a sparse inverse covariance consists in the provision of a network \CR{representing a conditional dependency structure.} Indeed, let us recall that \CR{in linear models} zero entries in the inverse covariance are associated with couples of non-conditionally dependent variables.
\subsection{Information filtering network approach: LoGo}
An alternative approach to obtain sparse inverse covariance is by using information filtering networks generated by keeping the elements that contribute most to the covariance by means of a greedy process.
\CR{This approach, named LoGo, proceeds by first constructing a chordal information-filtering graph such as a Maximum Spanning Tree (MST) \cite{kruskal1956,mantegna1999} or a Triangulated Maximally Filtered Graph (TMFG) \cite{TMFG}.
These graphs are build by retaining edges that maximally contribute to a given gain function which, in this case, is the log-likelihood or --more simply-- the sum of the squared correlation coefficients \cite{asteetal2005,tumminelloetal2005,TMFG}.
Then, this chordal structure is interpreted as the inference structure of the joint probability distribution function with non-zero conditional dependency only between variables that are directly connected by an edge.
On this structure the sparse inverse covariance is computed in such a way to preserve the values of the correlation coefficients between couples of variables that are directly connected with an information-filtering graph edge.
The main advantage} of this approach is that inversion is performed at local level on a small subsets of variables and then the global inverse is reconstructed by joining the local parts through the information filtering network.
Because of this Local-Global construction this method is named LoGo.
It has been shown that LoGo method yields to statistically significant sparse precision matrices that outperform the ones with the same sparsity computed with lasso method \cite{LoGo16}.
\section{Causality network reterival} \label{s.causalNet}
\subsection{Simulated multivariate autoregressive linear process}
In order to be able to test if causality measures can retrieve the true causality network in the underlying process, we generated artificial multivariate normal time series with known sparse causality structure by using the following autoregressive multivariate linear process \cite{hamilton1994time}:
\begin{align}\label{Process}
\mathbf Z_t = \sum_{\lambda=1}^\tau \mathbf A_\lambda \mathbf Z_{t-\lambda} + \mathbf U_t
\end{align}
where $\mathbf A_\lambda \in \mathbb R^{p \times p}$ are matrices with random entries drawn from a normal distribution.
The matrices are made upper diagonal (diagonal included) by putting to zero all lower diagonal coefficients and made sparse by keeping only a $\mathcal O(p)$ total number of entries different from zero in the upper and diagonal part.
$\mathbf U_t \in \mathbb R^p$ are random normally distributed uncorrelated variables.
This process produces autocorrelated, cross-correlated and causally dependent time series.
We chose it because it is among the simplest processes that can generate this kind of structured datasets.
The dependency and causality structure is determined by the non-zero entries of the matrices $\mathbf A_\lambda$.
The upper-triangular structure of these matrices simplify the causality structure eliminating causality cycles.
Their sparsity reduces dependency and causality interactions among variables.
The process is made autoregressive and stationary by keeping the eigenvalues of $\mathbf A_\lambda$ all smaller than one in absolute value.
For the tests we used $\tau=5$, $p=100$ and sparsity is enforced to have a number of links approximately equal to $p$.
We reconstructed the network from time series of different lengths $q$ between 5 to 20,000 points.
To test statistical reliability the process was repeated 100 times with every time a different set of randomly generated matrices $\mathbf A_\lambda$.
We verify that the results are robust and consistent by varying sample sizes from $p=20$ to $200$, by changing sparsity with number of links from $0.5 p$ to $5 p$ and for $\tau$ from 1 to 10.
\CR{We verified that the presence of isolated nodes or highly connected hub nodes does not affect results significantly.}
\subsection{Causality and inference network retrieval}
We tested the agreement between the causality structure of the underlying process and the one inferred from the analysis of $p$ time-series of different lengths $q$, $\mathbf Z_t \in \mathbb R^p$ with $t=1..q$, generated by using Eq.\ref{Process}.
We have $p$ different variables and $\tau$ lags.
The dimensionality of the problem is therefore $N = p \times (\tau+1)$ variables at all lags including zero.
To estimate the inference and causality networks we started by computing the inverse covariance, $\mathbf J \in \mathbb R^{ N \times N}$, for all variables at all lags $\mathbf Z \in \mathbb R^{ N \times q}$ by using the following three different estimation methods:
\begin{itemize}
\item[1)] $\ell_1$ norm penalization (Glasso \cite{friedmanetal2008});
\item[2)] $\ell_2$ norm penalization (ridge \cite{tikhonov1963solution});
\item[3)] information filtering network (LoGo \cite{LoGo16}).
\end{itemize}
We retrieved the inference network by looking at all couples of variables, with indices ${\mathbf i} \in[1,..,p]$ and ${\mathbf j} \in[1,..,p]$, which have non-zero entries in the inverse covariance matrix $\mathbf J$ between the lagged set of $\mathbf j$ and the non-lagged $\mathbf i$.
Clearly, for the ridge method the result is a complete graph but for the Glasso and LoGo the results are sparse networks with edges corresponding to non-zero conditional transfer entropies between variables $\mathbf i$ and $\mathbf j$.
For the LoGo calculation we make use of the regularizer parameter as a local shrinkage factor to improve the local inversion of the covariance of the 4-cliques and triangular separators (see \cite{LoGo16}).
We then estimated transfer entropy between couples of variables, ${\mathbf i} \rightarrow {\mathbf j}$ conditioned to all other variables in the system.
This is obtained by estimating of the inverse covariance matrix (indicated with an `hat' symbol) by using Eq.\ref{TE20b} (see Appendix \ref{s.CTE}) with:
\begin{align}
\mathbf{Z_1} &= \mathbf{ Z_j}_{,t}\\ \nonumber
\mathbf{Z_2} &=\{\mathbf{ Z_i}_{,t-1}...\mathbf{ Z_i}_{,t-\tau}\}\\ \nonumber
\mathbf{Z_3} &= \{\mathbf{ Z_j}_{,t-1}...\mathbf{ Z_j}_{,t-\tau},\mathbf{ W}\} \;.
\end{align}
With $\mathbf{ W}$ a conditioning to all variables $\mathbf{Z}$ except $\mathbf{Z_1},\mathbf{Z_2}$ and $\{\mathbf{ Z_j}_{,t-1}...\mathbf{ Z_j}_{,t-\tau}\}$.
The result is a $p \times p$ matrix of conditional transfer entropies $T(\mathbf{ Z_i}_{,t} \rightarrow \mathbf{ Z_j}_{,t})$.
Finally, to retrieve the causality network we retained the network of statistically validated conditional transfer entropies only.
Statistical validation was performed as follows.
\subsection{Statistical validation of causality}
{Statistical validation has been performed from likelihood ratio statistical test. Indeed, entropy and likelihood are intimately related: entropy measures uncertainty and likelihood measures the reduction in uncertainty provided by the model.
Specifically, the Shannon entropy associated with a set of random variables, $\mathbf{ Z_i}$, with probability distribution $p(\mathbf{ Z_i})$ is $H(\mathbf{ Z_i}) = - \mathbb E [ \log p(\mathbf{ Z_i}) ]$ (Eq.\ref{entropy}) whereas the log-likelihood for the model $\hat p(\mathbf{ Z_i})$ associated with a set of independent observations $\mathbf{ \hat Z}_{i,t}$ with $t=1..q$ is $\log \mathcal L(\mathbf{\hat Z}_{\mathbf i}) = \sum_{t=1}^q \log \hat p(\mathbf{\hat Z}_{i,t})$ which can be written as $\log \mathcal L(\mathbf{\hat Z}_{\mathbf i}) = q \mathbb E_{\hat p} [ \log \hat p(\mathbf{ Z_i}) ]$.
Note that $q$ is the total available number of observations which, in practice, is the length of the time-series minus the maximum number of lags.
It is evident from these expressions that entropy and the log-likelihood are strictly related trough this link might be non-trivial.
In the case of linear modeling this connection is quite evident because the entropy estimate is $H = \frac 12 (- \log |\mathbf{\hat J} | + p \log (2\pi) + p)$ and the log-likelihood is $\log \mathcal L = \frac q2 ( \log |\mathbf{\hat J} | - Tr( \mathbf{\hat \Sigma} \mathbf{\hat J} ) - p \log (2\pi))$.
For the three models we study in this paper we have $Tr( \mathbf{\hat \Sigma} \mathbf{\hat J} ) =p$ and therefore the log-likelihood is equal to $q$ times the opposite of the entropy estimate.
Transfer entropy, or conditional transfer entropy, are differences between two entropies: the one of a set of variables conditioned to their own past minus the one conditioned also to {the past} of another variable.
This, in turns, is the difference of the unitary log-likelihood of two models and therefore it is the logarithm of a likelihood ratio.
As Wilks pointed out \cite{wilks1938large,vuong1989likelihood} the null distribution of such model is asymptotically quite universal.
Following the likelihood ratio formalism, we have $\lambda=q T$ and the probability of observing a transfer entropy larger than $T$, estimated under null hypothesis, is given by $p_v \sim 1-\chi_c^2(r qT,d)$ with $r\simeq 2$ and $\chi_c^2$ the chi-square the cumulative distribution function with $d$ degrees of freedom which are the difference between the number of parameters in the two models.
In our case the two models have respectively $\tau (p_j^2+1)$ and $\tau(p_j^2+1) + \tau (p_j \, p_i)$ parameters.
}
\subsection{Statistical validation of the network}
The procedures described in the previous two subsections produce the inference network and causality network.
Such networks are then compared with the known underlying network of true causalities in the underlying process which is defined by the non-zero elements in the matrices $A_\lambda$ (see Eq.\ref{Process}).
The overlapping between the retrieved links in the inference or causality networks with the ones in the true network underlying the process is an indication of a discovery of a true causality relation.
However some discoveries can be obtained just by chance or some methodologies might discover more links only because they produce denser networks.
We therefore tested the hypothesis that the matching links in the retrieved networks are not obtained just by chances by computing the null-hypothesis probability to obtain the same or a larger number of matches randomly.
Such probability is given by the conjugate cumulative hypergeometric distribution for a number equal or larger than $\mathrm{TP}$ of `true positive' matching causality links between an inferred network of $n$ links and a process network of $K$ true causality links, from a population of $p^2-p$ possible links:
\begin{align}\label{HyperTest}
P(X \ge \mathrm{TP} | n , K , p)= 1 - \sum_{k=0}^{\mathrm{TP}-1} {\frac {{\binom {K}{k}}{\binom {p^2-p -K}{n-k}}}{\binom {p^2-p}{n}}} \;\;\;.
\end{align}
Small values of $P$ indicate that the retrieved $\mathrm{TP}$ links out of $K$ are unlikely to be found by randomly picking $n$ edges from $p^2-p$ possibilities.
Note that in the confusion matrix notation \cite{swets2014signal} we have \CR{ $n=\mathrm{TP}+\mathrm{FP}$ and $K = \mathrm{TP}+\mathrm{FN}$ }with $\mathrm{TP}$ number of true positives, $\mathrm{FP}$ number of false positives, $\mathrm{FN}$ number of false negatives and $\mathrm{TN}$ number of true negatives.
\CR{The total number of `negative' (unlinked couples of vertices) in the true model is instead $m=\mathrm{FP}+\mathrm{TN}$.}
\section{Results}\label{s.results}
\subsection{Computation and validation of conditional transfer entropies}
By using Eq.\ref{Process} we generated 100 multivariate autoregressive processes with known causality structures.
We here report results for $p=100$ but analogous outcomes were observed for dimensionalities between $p=20$ and $200$ variables.
Conditional transfer entropies between all couples of variables, conditioned to all other variables in the system, were computed by estimating the inverse covariances by using tree methodologies, ridge, lasso and LoGo and applying Eq.\ref{TE20}.
Conditional transfer entropies were statistically validated with respect to null hypothesis (no causality) at $p_v=1\%$ p-value. Results for Bonferroni adjusted p-value at 1\% (i.e. $p_v = 0.01/(p^2-p) \sim 10^{-6}$ for $p=100$) are reported in Appendix \ref{BonferroniValidated}.
We also tested other values of $p_v$ from $10^{-8}$ to 0.1 obtaining consistent results.
We observe that small $p_v$ reduce the number of validated causality links but increase the chance that these links match with the true network in the process.
Conversely large values of $p_v$ increase the numbers of mismatched links but also of the true links discoveries.
Let us note that here we use $p_v$ as a thresholding criteria and we are not claiming any evidence of statistical significance of the causality.
We assess the goodness of this choice a-posteriori by comparing the resulting causality network with the known causality network of the process.
\begin{center}
\begin{figure}[t]
\includegraphics[width=0.80\textwidth]{SignificantRetreivals.pdf}
\caption{
{\bf Regions in the $p/q$-$\gamma$ space where causality networks for the three models are statistically significant.}
The significance regions are all at the left of the corresponding lines.
Tick line reports the boundary $P< 0.05$ (Eq.\ref{HyperTest}) and dotted lines indicate $P< 10^{-8}$ significance levels ($P$ is averaged over 100 processes).
The plots refer to $p=100$ and report the region where the causality network are all significant for 100 processes. }
\label{f.significance}
\end{figure}
\end{center}
\begin{center}
\begin{figure}[t]
\includegraphics[width=0.80\textwidth]{FractionRetreivedLinks_Colromap.pdf}
\caption{
{\bf True positive rate: fraction of retrieved true causality links ($\mathrm{TP}$) with respect to the total number of links in the process (n).}
The three panels refer to ridges, Glasso and LoGo (top, central and bottom).
Data are average fractions over 100 processes.
}
\label{f.RegionFractionRetreived}
\end{figure}
\end{center}
\subsection{Statistical significance of the recovered causality network.}
Results for the contour frontiers of significant causality links for the three models are reported in Fig.\ref{f.significance} for a range of time series with lengths $q$ between 10 and 20,000 and regularizer parameters $\gamma$ between $10^{-8}$ and $0.5$.
Statistical significance is computed by using Eq.\ref{HyperTest} and results are reported for both $P < 0.05$ and $P < 10^{-8}$ (continuous and dotted lines respectively).
As one can see, the overall behaviours for the three methodologies are little affected by the threshold on $P$.
We observe that LoGo significance region extends well beyond the Glasso and ridge regions.
The value of the regularizer parameter $\gamma$ affects the results for the three models in a different way.
Glasso has a region in the plane $\gamma-p/q$ where it has best performances (in this case it appears to be around $\gamma \simeq 0.1$ and $p/q \simeq 2.5$).
Ridge appears instead to be little affected with mostly constant performances across the range of $\gamma$.
LoGo has best performances for small, even infinitesimal, values of $\gamma$.
Indeed, differently from Glasso in this case $\gamma$ does not control sparsity but instead acts as local shrinkage parameter.
Very small values can be useful in some particular cases to reduce the effect of noise but large values have only the effect to reduce information.
\begin{table}[!h]
\caption{
{\bf Causality network validation.} Comparison between fraction of true positive (${TP/n}$) and fraction of false positive ($FP/n$), statistically validated, causality links for the three models and different time-series lengths.
The table reports only the case for the parameter $\gamma=0.1$.
Statistical validation of conditional transfer entropy is at $p_v=1\%$ p-value.
Note that LoGo can perform better than reported in this table for smaller values of $\gamma$ (see Figs.\ref{f.significance} and \ref{f.RegionFractionRetreived}).
}
\begin{center}
\begin{tabular}{@{\extracolsep{4pt}}llllllllll@{}}
\hline
q & 10 & 20 & 30 & 50 & 200 & 300 & 1000 & 20000 \\ \hline
ridge TP/n & 0.00 & 0.00 & 0.00 & 0.00 & 0.23$^{**}$ & 0.49$^{**}$ & 0.76$^{**}$ & 0.93$^{**}$ \\
ridge FP/n & 0.00 & 0.00 & 0.00 & 0.00 & 0.00 & 0.10 & 0.65 & 1.06 \\ \hline
Glasso TP/n & 0.00 & 0.00 & 0.00 & 0.13$^{**}$ & 0.48$^{**}$ & 0.53$^{**}$ & 0.62$^{**}$ & 0.74$^{**}$ \\
Glasso FP/n & 0.00 & 0.00 & 0.00 & 0.00 & 0.06 & 0.10 & 0.23 & 0.54 \\ \hline
LoGo TP/n & 0.00 & 0.08$^*$ & 0.21$^{**}$ & 0.37$^{**}$ & 0.61$^{**}$ & 0.65$^{**}$ & 0.75$^{**}$ & 0.90$^{**}$ \\
LoGo FP/n & 0.00 & 0.00 & 0.00 & 0.01 & 0.06 & 0.08 & 0.15 & 0.34 \\ \hline
&& $^{*}$ $P < 0.05$;& $^{**}$ $P < 10^{-8}$ &&&\\
\end{tabular}\end{center}
\label{tab:fraction}
\end{table}
\subsection{Causality links retrieval}
Once identified the parameter-regions where the retrieved causality links are statistically significant, we also measured the fraction of true links retrieved.
\CR{Indeed, given that the true underlying causality network is sparse, one could do significantly better than random by discovering only a few true positives.
Instead, from any practical perspective we aim to discover a significant fraction of the edges.
Figure \ref{f.RegionFractionRetreived} shows that the fraction of causality links correctly discovered (true positive, $\mathrm{TP}$) with respect to the total number of causality links in the process ($n$) is indeed large reaching values above 50\%.}
This is the so-called true positive rate or sensitivity, which takes values between 0 (no links discovered) and 1 (all links discovered).
Reported values are averages over 100 processes.
We observe that the region with discovering of 10\% or more true causality links greatly overlaps with the statistical validity region of Fig.\ref{f.significance}.
We note that when the observation time becomes long, $p/q \lessapprox 0.25$, ridge discovery rate becomes larger than LoGo.
However, statistical significance is still inferior to LoGo, indeed the ridge network becomes dense when $q$ increases and the larger discovery rate of true causality links is also accompanied by a larger rate of false links incorrectly identified (false positive $\mathrm{FP}$).
The fraction of false positives with respect to the total number of causality links in the process ($n$) are reported in Table~\ref{tab:fraction} together with the true positive rate for comparison.
This number can reach values larger than one because the process is sparse and there are much more possibilities to randomly chose false links than true links.
\CR{Note this is not the false positive rate, which instead is $\mathrm{FP}/m$, and cannot be larger than one.}
Consistently with Fig.\ref{f.significance} we observe that, for short time series, up to $p/q \sim 0.5$ the sparse models have better capability to identify true causality links and to discard the false ones with LoGo being superior to Glasso.
Remarkably, LoGo can identify a significant fraction of causality links already from time-series with lengths of 30 data-points only.
P-value significances, reported in the table with one or two stars indicate when all values of $P(X \ge \mathrm{TP} | n , K , p)$ from Eq.\ref{HyperTest} for all 100 processes have respectively $P<0.05$ or $P<10^{-8}$.
Again we observe that LoGo discovery rate region extends well beyond the Glasso and ridge regions.
\begin{table}[!h]
\caption{
{\bf Inference network validation: comparison between fraction of true positive (TP/n) and fraction of false positive (FP/n).}
Data for ridge are only for comparison because it is a complete graph with all links present.
The table reports only the case for the parameter $\gamma=0.1$.
}
\begin{center}
\begin{tabular}{@{\extracolsep{4pt}}llllllllll@{}}
\hline
q & 10 & 20 & 30 & 50 & 200 & 300 & 1000 & 20000 \\ \hline
ridge TP/n & 1.00 & 1.00 & 1.00 & 1.00 & 1.00 & 1.00 & 1.00 & 1.00 \\
ridge FP/n & 97.84 & 97.84 & 97.84 & 97.84 & 97.84 & 97.84 & 97.84 & 97.84 \\ \hline
Glasso TP/n & 0.61$^*$ & 0.74$^*$ & 0.79$^*$ & 0.85$^*$ & 0.87$^{**}$ & 0.84$^{**}$ & 0.80$^{**}$ & 0.80$^{**}$ \\
Glasso FP/n & 28.39 & 38.11 & 45.79 & 53.58 & 40.61 & 26.60 & 1.54 & 0.92 \\ \hline
LoGo TP/n & 0.31$^*$ & 0.50$^{**}$ & 0.58$^{**}$ & 0.63$^{**}$ & 0.75$^{**}$ & 0.78$^{**}$ & 0.85$^{**}$ & 0.93$^{**}$ \\
LoGo FP/n & 4.53 & 4.27 & 4.18 & 4.03 & 3.72 & 3.63 & 3.44 & 3.21 \\ \hline
&& $^{*}$ $P < 0.05$; & $^{**}$ $P < 10^{-8}$ &&&\\
\end{tabular}\end{center}
\label{tab:fraction_Inference}
\end{table}
\begin{center}
\begin{figure}[t]
\includegraphics[width=0.45\textwidth]{ROC_cond.pdf}
\includegraphics[width=0.45\textwidth]{ROC_All_detail.pdf}
\caption{
{\bf ROC values, for each model and each parameter combination.} X-axis false positive rates ($\mathrm{FP}/m$), y-axis true positive rates ($\mathrm{TP}/n$).
Left and right figures are the same with X-axis expanded on the low values only for the right figure to better visualise the differences between the various models.
Large symbols refer to $\gamma = 0.1$ and validation at p-value $p_v=0.01$.
Color intensity is proportional to time series length.
Inference network results are all outside the range of the plot.
Reported values are averages over 100 processes. }
\label{f.ROC}
\end{figure}
\end{center}
\subsection{Inference network}
We have so far empirically demonstrated that a significant part of the true causality network can be retrieved from the statistically validated network of conditional transfer entropies.
Results depend on the choice of the threshold value of the $p_v$ at which null hypothesis is rejected.
We observed that lower $p_v$ are associated with network with fewer true positives but also fewer false positives and conversely larger $p_v$ yield to causality networks with larger true positives but also larger false positives.
Let us here report on the extreme case of the inference network which contains all causality channels with no validation.
For the ridge model this network is the complete graph with all variables connected to each-other.
Instead, for Glasso and LoGo the inference network is sparse.
Results are summarized in Table.\ref{tab:fraction_Inference}.
In terms of true positive rate we first notice that they are all larger than the ones in Table.\ref{tab:fraction}.
Indeed, the network of statistically validated conditional transfer entropies is a sub-network of the inference network.
On the other hand we notice that the false positive fraction is much larger than the ones in Table.\ref{tab:fraction_Inference}.
Ridge network has a fraction of 1 because, in this case, the inference network is the complete graph.
Galsso also contains a very large number of false positives reaching even 55 times the number of links in the true network and getting to lower fractions only from long time-series with $q >1000$.
These numbers also indicate that Galsso networks are not sparse.
LoGo has a sparser and more significant inference network with smaller fractions of false positives which stay below $5n$, which is anyway a large number of misclassification.
Nonetheless, we observe that, despite such large fractions of $\mathrm{FP}$, the discovered true positives are statistically significant.
\begin{table}[!h]
\caption{
{\bf Unconditioned transfer entropy network: comparison between fraction of true positive (TP/n) and fraction of false positive (FP/n).}
Statistical validation of transfer entropy is at $p_v=1\%$ p-value.
The table reports only the case for the parameter $\gamma=0.1$.
}
\begin{center}
\begin{tabular}{@{\extracolsep{4pt}}lllllllll@{}}
\hline
q & 10 & 20 & 30 & 50 & 200 & 300 & 1000 & 20000 \\ \hline
ridge TP/n & 0.02 & 0.39$^{**}$ & 0.45$^{**}$ & 0.51$^{**}$ & 0.65$^{**}$ & 0.69$^{**}$ & 0.78$^{**}$ & 0.92$^{**}$ \\
ridge FP/n & 0.07 & 1.06 & 0.95 & 0.85 & 0.93 & 0.99 & 1.20 & 1.73 \\ \hline
Glasso TP/n & 0.00 & 0.24$^{**}$ & 0.35$^{**}$ & 0.43$^{**}$ & 0.57$^{**}$ & 0.60$^{**}$ & 0.67$^{**}$ & 0.77$^{**}$ \\
Glasso FP/n & 0.00 & 0.10 & 0.20 & 0.29 & 0.51 & 0.56 & 0.73 & 1.66 \\ \hline
LoGo TP/n & 0.11 & 0.34$^{**}$ & 0.41$^{**}$ & 0.47$^{**}$ & 0.63$^{**}$ & 0.66$^{**}$ & 0.76$^{**}$ & 0.89$^{**}$ \\
LoGo FP/n & 0.02 & 0.16 & 0.25 & 0.34 & 0.59 & 0.66 & 0.87 & 1.49 \\ \hline
&& & $^{**}$ $P < 10^{-8}$ &&&\\
\end{tabular}\end{center}
\label{tab:fraction_Unconditional}
\end{table}
\subsection{Unconditioned transfer entropy network}
We last tested whether conditioning to {the past} of all other variables gives better causality network retrievals than the unconditioned case.
Here, transfer entropy, $T(\mathbf{ Z_i} \rightarrow \mathbf{ Z_j})$, is computed by using Eq.\ref{TE20} with $\mathbf W=\emptyset$, the empty set.
For the ridge case this unconditional transfer entropy depends only from the time-series, $\mathbf{ Z}_{i,t}$, $\{\mathbf{ Z}_{i,t-1},...,\mathbf{ Z}_{i,t-\tau}\}$ and $\{\mathbf{ Z}_{j,t-1},...,\mathbf{ Z}_{j,t-\tau}\}$ (with $\tau = 5$ in this case).
Glasso and LoGo cases are instead hybrid because a conditional dependency has been already introduced in the sparse structure of the inverse covariance $\mathbf J$ (the inference network).
Results are reported in Tab.\ref{tab:fraction_Unconditional} where we observe that these networks retrieve a larger quantity of true positives than the ones constructed from conditional entropy.
However, the fraction of false positive is also larger than the ones in Tab.\ref{tab:fraction} although it is smaller than what observed in the inference network in Tab.\ref{tab:fraction_Inference}.
{Overall, these results indicate that conditioning is effective in discarding false positives. }
\subsection{Summary of all results in a single ROC plot}
In summary, we have investigated the networks associated with conditional transfer entropy, unconditional transfer entropy and inference for three models under a range of different parameters.
In the previous sub-sections we have provided some comparisons between the performances of the three models in different ranges of parameters.
Let us here provide a summary of all results within a single ROC plot \cite{swets2014signal}.
Figure \ref{f.ROC} reports the ROC values, for each model and each parameter combination, x-axis are false positive rates ($\mathrm{FP}/m$) and y-axis true positive rates ($\mathrm{TP}/n$).
Each point is an average over 100 processes.
Points above the diagonal line are associated to relatively well performing models with the upper left corner representing the point where models correctly discover all true causality links without any false positive.
The plot reports with large symbols the cases for $\gamma = 0.1$ and validation at p-value $p_v=0.01$, which can be compared with the data reported in the tables.
\CR{We note that, by construction, LoGo models are sparse (with a number of edges $\sim 3p$ \cite{LoGo16}). This restrains the ROC results to the left-hand side of the plot.
For this reason an expanded view of the figure is also proposed with the x-axis scaled.
Note that this ROC curve is provided as a visual tool for intuitive comparison between models.}
Overall from Tables \ref{tab:fraction}, \ref{tab:fraction_Inference}, \ref{tab:fraction_Unconditional} and Fig. \ref{f.ROC} we conclude that all models obtain better results for longer time series and that conditional transfer entropy over-performs the unconditional counterparts (see, Tables \ref{tab:fraction} and \ref{tab:fraction_Unconditional} and the two separated ROC figures for conditional and unconditional transfer entropies reported in Fig.\ref{f.ROCCoUc} in appendix \ref{A.Co.UC}).
In the range of short time series, when $q\le p$, which is of interest for this paper, LoGo is the best performing model with better performances achieved for small $\gamma \lesssim 10^{-4}$ and validation with small p-values $p_v \lesssim 10^{-4}$.
LoGo is consistently the best performing model also for longer time-series up to lengths of $q \sim1000$.
Instead, above $q=2000$ ridge begins to provide better results.
For long time series, at $q=20,000$, the best performing model is ridge with parameters $\gamma =10^{-5}$, p-value $p_v = 5\;10^{-6}$.
LoGo is also performing well when time series are long with best performance obtained at $q=20,000$ for parameters $\gamma =10^{-10}$, p-value $p_v = 5\;10^{-6}$.
We note that LoGo instead performs poorly in the region of parameters with $\gamma \le 0.1$ and $p_v \le 0.01$ for short time-series $q\le p/2$.
\section{Conclusions and perspectives}\label{s.conclusions}
In this paper we have undertaken the challenging task to explore models and parameter regions where the analytics of time series can retrieve significant fractions of true causality links from linear multivariate autoregressive process with known causality structure.
Results demonstrate that sparse models with conditional transfer entropy are the ones who achieve best results with significant causality link retrievals already for very short time series even with $q \le p/5 = 20$.
\CR{This region is very critical and general considerations would suggest that no solutions can be discovered.
Indeed, this result is in apparent contradiction with a general analytical results in \cite{varga2016replica,papp2016fluctuation} who find that no significant solutions should be retrieved for $q \le N/2 = 150$.
However, we notice that the problem we are addressing here is different from the one in \cite{varga2016replica,papp2016fluctuation}.
In this paper we have been considering an underlying sparse true causality structure and such a sparsity changes considerably the condition of the problem yielding to significant solutions even well below the theoretical limit from \cite{varga2016replica,papp2016fluctuation} which is instead associated to non-sparse models.
}
Unexpectedly, we observed that the structure of the inference networks in the two sparse models, Glasso and LoGo, have excessive numbers of false positives yielding to rather poor performances.
However, in these models false positive can be efficiently filtered out by imposing statistical significance of the transfer entropies.
Results are affected by the choice of the parameters and the fact that the models depend on various parameters ($q$, $p$, $\gamma$, $p_v$, $P$) make the navigation in this space quite complex.
We observed that the choice of p-values, $p_v$, for valid transfer entropies affects results.
Within our setting we obtained best results with the smaller p-values especially in the regions of short time-series.
We note that the regularizer parameter $\gamma$ also plays an important role and best performances are obtained by combination of the two parameters $\gamma$ and $p_v$.
Not surprizingly, longer time-series yield to better results.
We observe that conditioning to all other variables or unconditioning is affecting the transfer entropy estimation with better performing causality network retrieval obtained for conditioned transfer entropies.
However, qualitatively, results are comparable. Other intermediate cases, such as conditioning to past of all other variables only, have been explored again with qualitatively comparable results.
It must be said that in the present system results are expected to be robust to different conditionings because the underlying network of the investigated processes is sparse. For denser inference structures, conditioning could affect more the results.
Consistently with the findings in \cite{LoGo16} we find that LoGo outperforms the other methods.
This is encouraging because the present settings of LoGo is using a simple class of information filtering networks, namely the TMFG \cite{TMFG}, obtained by retaining largest correlations.
There are a number of alternative information filtering networks which should be explored.
In particular, given the importance of statistical validation emerged from the present work, it would be interesting to explore statistical validation within the process of construction of the information filtering networks themselves.
In this paper we investigate a simple case with a linear autoregressive multivariate normal process analysed by means of linear models.
Both LoGo and Glasso can be extended to the non-linear case with LoGo being particularly suitable for non-parametric approaches as well \cite{LoGo16}.
There are Alternative methods to extract causality networks from short time series, in particular Multispatial CCM \cite{sugihara2012detecting,clark2015spatial} appears to perform well for short time series.
A comparison between different approaches and the application of these methods to real data will be extremely interesting.
However this should be the object of future works.
\subsection*{Acknowledgement}
T.A. acknowledges support of the UK Economic and Social Research Council (ESRC) in funding the Systemic Risk Centre (ES/K002309/1).
TDM wishes to thank the COST Action TD1210 for partially supporting this work and Complexity Science Hub Vienna.
\subsection*{Conflict of interest disclosure}
The authors declare that there is no conflict of interest regarding the publication of this paper.
| 537d9f761c8e273eae8aa233420441477cad08de | {
"file_path": "/home/ubuntu/dolma-v1_7/arxiv-0000.json.gz"
} |
Subsets and Splits